TFW Changelog

All notable changes to the Trace-First Workflow framework. Format: Keep a Changelog. Versioning: Semantic Versioning.

[Unreleased]

Documentation

• FLC (TFW_20261011-192605_FLC): independently reviewed, bounded diagnosis of workflow completion, coordination and knowledge continuity. Preserves observed violations, contract ambiguities and evidence limits; changes no framework code, role rule or official operating mode. Owner accepted the diagnostic result for closure.

[3.9.2] — 2026-10-12

Decision-oriented economics (TFW_20261012-135034_ECOV) connects accepted results, comparable work and approximate resources to management/product decisions and retrospectives. See Updating to TFW 3.9.2 for compatible adoption and the route from earlier supported versions.

Changed

• Full roles distinguish unavailable binding before extraction, actual extraction failure, optional unavailable metrics and missing returns. An attributable nonmeasurement outcome can finish ordinary collection accountability; required measurement and product-proof gates remain.
• Final role returns and Coordinator integration define a common cycle checkpoint. Interim estimates stay interim; a disclosed short delivery/cleanup tail does not trigger endless recapture. Complete selected bytes do not imply a complete task or complete measured coverage.
• Reports show expected/received units separately from selected measurements and relate resources to actual output, scope, quality and uncertainty. Routine reports can be concise. Practical approximation tolerance is not a guaranteed error band, a billing claim or a reason to hide gaps.

Fixed

• Research no longer permits null for unavailable required source fields in numeric manifests; pre-extraction gaps belong in the existing role artifact. Coordinator, Executor, Researcher and Reviewer clauses now agree with the common Full outcomes and their installed command copies.

Numeric v1 records, the collector/reconciliation helper, prior measurements, independent review, owner acceptance and the 3.9.1 communication contract are preserved. No measured saving or provider-wide accuracy claim is made.

[3.9.1] — 2026-10-12

Gate and Signal Contract (TFW_20261012-120804_GSC) keeps substantive inter-role results in their owned files and makes the checkpoint action explicit. See Updating to TFW 3.9.1 for the exact source and safe adoption order, including the route from earlier supported versions.

Changed

• One closed message catalogue owns selected tfw-gates-only edges. Commands and completion signals carry exact record references; findings, verdicts, questions and grounds remain in those records. Native wait/readback and a bounded status fallback preserve continuation.
• ONB without blocking questions notifies the Coordinator and continues without acknowledgement. A real blocker waits for the responsible decision. RF and REVIEW return control to their Coordinator and end the current role act, without declaring the whole task complete.
• Gate, Check, Decision gate, WAIT and Role return have distinct actions. The Mandatory Saint-Exupéry check still precedes presenting HL/TS; material failure requires revision and necessary value, proof and Executor freedom must survive simplification.
• Canonical workflows, templates and selected adapter copies agree. Research's required human-facing stage interaction and substantive owner–Coordinator discussion retain their scope.

Fixed

• Selected glossary links resolve to renamed check headings. The project release contract now states the existing guard's 5 MiB Git-blob limit; executable policy is unchanged.

[3.9.0] — 2026-10-06

This minor release keeps every TFW project to its result and its selected trace: working material leaves the project, evidence/ becomes a registry, Daily keeps its record folder clean, and comments stop being a channel between agents. See Updating to TFW 3.9.0; it explains the harm of comment correspondence and offers an optional cleanup task, never applied silently.

Changed

Lean Project Footprint (TFW_20261004-174815_LPF), independently reviewed; Candidate 0863d299d86a1ca1776ffb966d54a1b848e7c348:

• Working material — raw output, logs, exports, screenshots and scratch scripts — lives in the system temporary directory under tfw/<ID>/, private to the role that made it and removed when its work ends. evidence/ holds only EV registry rows (verified · how · observed · result); the Attachments section is gone.
• A comment in a value file carries value for its reader or is read by a program; comments are not a channel between agents. Every adapter root states both rules in one line.
• Daily records list every product with its project path; the record folder holds only record and economics files before each orderly turn return.

Fixed

• LPF (TFW_20261004-174815_LPF): the Cursor and Antigravity rule templates list the Coordinator commands in one row.

[3.8.1] — 2026-10-02

This patch release makes product economics a standard command, keeps Daily optional and adds shared official quote history with reproducible task-owned monetary bases. See Updating to TFW 3.8.1.

Added

• ECON (TFW_20261001-001035_ECON): independently reviewed standard /tfw-economics product analysis across four default adapter installations/updates, with optional cumulative Daily capture and complete executive/detail text-table reports. Shared public monthly quotes reuse exact model/profile prices for less than 30 days; immutable task-owned price bases preserve offline recorded estimates and distinguish historical reconstruction from current revaluation. Last-good/failure, custom receivers and dated history are preserved. Replacement Candidate 1fafca12ebe912f30115335c5931aade482659f2; native host limits remain as recorded in its independent review.

[3.8.0] — 2026-09-29

This minor release adds portable task economics and light framework delivery: install and update fetch only the framework, release archives carry only framework material, and a new task shows a lagging version. See Updating to TFW 3.8.0.

Added

• TEQM (TFW_20260928-181352_TEQM): portable compact role economics, shared validation and task/phase reports with selected project/date/tag/model/role views and CSV export. Three bounded numeric-source recipes distinguish tokens, native time and conditional API prices. Full role returns now supply their own files; final economics presentation precedes safe disposable cleanup. The optional standard-library helper preserves the ordinary lifecycle's runtime independence. Each child agent is its own source named by the stream it writes — a Claude Code subagent is <sessionId>/<agentId>, a Codex child its rollout's session_meta.id — an Antigravity counter absent beside reported ones decodes as a proto3 zero while a row reporting none is left out and disclosed, and reports name measured units that reconciliation omitted (Daily 20260929-104929_claude-subagent-economics). Project comparisons use selected sources; the project-specific demonstration flag was removed.

Changed

• The Claude Code and Antigravity Coordinator profiles again allow separate role agents for small one-phase work, as 3.6.x did, now as an offered owner choice: in the single-phase agent mode the task Coordinator stays the owner's only conversation and launches each needed role as a distinct agent with only /tfw-* <task>; the Reviewer is a new agent examining a fixed Candidate. Outside this mode the full-chat route and its subagent limits are unchanged. The Claude profile also drops its dated-source section, which repeated rules stated elsewhere and still argued for the mode that 3.7.0 removed (Daily 20260929-130629_single-phase-agent-mode).
• LFD (TFW_20260929-003444_LFD): install and update fetch only the framework. update.md Step 0 names one Git-only method — a blobless, depth-1, cone-mode sparse clone of the pinned tag (or an authorized untagged commit) into .tfw/.upstream/.clone with core.autocrlf=false and an empty core.attributesFile, a count-objects size check that discloses a full download, one fallback, a raw-byte check of every file and deletion of the clone — about 0.70 MiB for this release instead of about 114 MiB (the old fetch, measured at v3.7.1), with the pin and provenance guarantees unchanged. quickstart.md and the root README prompts fetch only .tfw/ and editions/ (about 0.79 MiB instead of a ≈123 MiB full clone). A new root .gitattributes keeps release archives framework-only (≈0.50 MiB instead of 178.9 MiB at v3.7.1). The new-task startup card compares the installed version with the newest exact release through one ls-remote bounded by about 13 s and never updates by itself. The update drops a repeated adapter-sync run and a third payload check, replaces "only when safe" with a pre-seal check that staging never entered Git, and asks receipts for commands and results instead of pasted output. Projects on 3.7.1 or earlier still fetch once by their installed text; README "Updating TFW" shows the one-time light-fetch request.

[3.7.1] — 2026-09-28

This patch restores strategic depth before HL approval and permits checked, owner-accepted Daily results to follow the project's release route without artificial Full artifacts. See Updating to TFW 3.7.1.

Fixed

• Plan preserves its Strategic Architect mindset and explicitly requires helping the owner decide what is worth doing and why before HL approval. Grounded project exploration precedes a direction; planning compares at least two meaningful approaches and the consequence of leaving things unchanged, and challenges both the owner's proposal and the Coordinator's recommendation.
• Plan and the HL template require three views in chat and HL: the concrete before/after outcome, the flow that creates its value, and the alternatives comparison. Template criteria are read before the preview; the existing HL approval gate requires one coherent strategic proposal before phases. Coordination, role authority and model/reasoning selection rules remain unchanged.
• Release selects evidence by the result's actual Full or Daily route. Daily workers may perform explicitly authorized release effects using their selected record, checks and owner acceptance; project-required review and separate publication permissions remain in force. The Daily skill, release entries and project release-contract template carry the same distinction.

[3.7.0] — 2026-09-28

This release combines the independently reviewed optional Daily route (PTW) and the owner-accepted task/phase coordination model (ATC). New-work topology changes from 3.6.1; active historical carriers remain readable under their actual authority. See Updating to TFW 3.7.0.

Added

• PTW optional Daily (TFW_20260923-223151_PTW): tfw-daily-task completes a bounded request after current-source, purpose and authority checks, retaining a separate selected-Trace form. Opt-in Codex/Claude entries and ordered init/update preservation remain outside the ten Full role commands; existing local forms/history and formal Full duties are preserved.

Fixed

• ATC gate-only launches and continuations exclude briefings, solution hints and copied reasoning. Required messages name a real workflow checkpoint; advice cannot become a gate by relabelling it.
• ATC (TFW_20260928-015408_ATC) keeps the owner-facing task Coordinator responsible for strategy, phase advancement and bounded owner updates. Phase Coordinators keep local role traffic and return only phase-level decisions, blockers and results. Required vertical gates and strategic owner discussion remain available without a separate mandatory GATEWAY.
• Codex task Coordinators use addressed role/phase returns and do not treat incidental wait commentary as a formal gate. Task-only Sections still use the resolved project's task prefix.

Changed

• ATC makes the direct task Coordinator the proportionate one-phase route and gives long work separate bounded phase Coordinators, with a safe acknowledged successor handover. New task/phase statuses issue upstream_route: owner:{human} / coordinator:{native-address}; historical owner_gateway carriers, GATEWAY and LEAD remain readable under their original authority.
• Claude and Antigravity state their owner-assisted full-role limits; subagents no longer substitute for TFW Researcher, Executor or independent Reviewer units. A named principal, host or optional read-only route checker grants no mandate, and no mandatory MCP server or new permanent test ships.

[3.6.1] — 2026-09-23

PCUX correction (TFW_20260922-192606_PCUX): choose autonomy by its practical consequences, protect the owner's conversation, and make each model/effort choice deliberate.

Fixed

• The startup card explains who starts agents, what the platform can do or needs a click for, which decisions remain with the owner, and whether the current interface is a GATEWAY or direct Coordinator. It states the benefit and briefly confirms the chosen arrangement; discussing the intent and using TFW roles are no longer presented as alternatives.
• Brief vertical TFW gate returns are explicitly enabled and required by default, including with manual agent creation. This is neither silence nor permission for peer chatter, solution hints or unrestricted dialogue; an unclear mode answer grants nothing.
• Full-session delegation requires a separate GATEWAY and a Coordinator per ready phase unless the owner explicitly chooses direct coordination. Inherited routing is not that choice; active work is preserved through a safe handover. The admitted small-task compact route is unchanged.
• Every role launch requires a justified model choice or the owner's explicit model policy, plus separate supported reasoning effort. An unexplained default, maximum effort or omitted comparison cannot replace judgment. Apply the selection when authorized or resolve the exact owner action.
• Codex offers configured-default model with per-launch effort selection, or owner-named models reused within the selected scope. Its current creation restriction does not require repeated model-name questions and is not imposed on other providers.
• Codex tasks use one task-only Section when available, with placement verified. Other platforms use supported grouping and disclose actual limits rather than silently omitting the convenience.

Compatibility and updating

Read Updating to TFW 3.6.1. Refresh the connected Plan/conventions/HL/profile group and installed Plan copies. No task-state schema, automatic authority migration, native-tool override or new provider reliability claim.

[3.6.0] — 2026-09-23

PCUX — Provider-specific Coordinator UX and Autonomy Guidance (TFW_20260922-192606_PCUX). The owner sees one practical startup arrangement and chooses the operating mode before substantive planning. Provider mechanics stay in selected profiles; the common TFW role, authority and review contracts remain shared.

Added

• Three selected Coordinator profiles for Codex, Claude Code and Antigravity. The active persistent adapter names its exact profile; runtime roles do not discover profiles through the tooling manifest.
• One mandatory owner-facing startup card: current capabilities and limits, mode, role topology, titles/grouping, workspaces, model/effort responsibility, exact activation commands, vertical returns, remaining owner actions and safe close. Unknown values remain unknown.
• Current-selection fields reporting and selection_ref, plus immutable coordination_selected events for actual owner choices, including future-effective switches and revocation.

Changed

• Plan asks the initial mode at new-task entry. Its later Coordination Selection gate validates and records that answer without routine repeated approval; existing continuations retain settled choices.
• A new role starts with only /tfw-* <task[/phase]>. Task files supply its context; the creation receipt or first normal gate can supply its native address without a second activation ceremony.
• GATEWAY topology and two-peer dialogue permission are independent. A selected persistent GATEWAY receives Coordinator-level gates/results, not worker traffic, and never runs Plan or monitors workers.
• Manual creation still requires vertical native gate reporting; only an explicit fully manual owner-transfer choice disables inter-agent sends. Authorized gates-only work advances without per-role mechanical approval, while reserved owner decisions remain gates.
• Claude guidance distinguishes click-created full chats in one serialized local workspace from a conditional compact role chain. Antigravity full chats remain owner-manual under the dated report; title and compact-tool availability must be checked on the actual surface.
• Plan preserves strategic architecture, Working Backwards, visible consequential hypotheses and Saint-Exupery judgment, alongside the complete per-launch model and reasoning-effort selection rule.
• Task closure explicitly requires docs/knowledge dispositions, a truthful task changelog and safe, attributable resource cleanup. A changelog or completed task grants no release authority.

Compatibility and updating

Read Updating to TFW 3.6.0, including from 3.5.2 and earlier. Install the connected coordination, state-reader, profile and adapter changes together. Complete older five-field routing remains readable at its actual authority; new writes use the complete seven-field form after source verification. Never bulk-rewrite task state, immutable journals or historical grants. This release does not claim a completed current Claude/Antigravity native cycle or browser trial; those offers retain their dated, conditional evidence limits.

[3.5.2] — 2026-09-22

CMTR — Per-Launch Model and Reasoning Selection. Coordinators now choose model and reasoning effort at each actual role launch from verified native options, protecting the required quality floor while avoiding unnecessary token/session spend. This patch also hardens gate-only communication and removes duplicated framework instructions.

Changed

• Before every role launch, Plan checks prerequisites, sets the quality floor, selects model and reasoning effort separately, and records one lower-resource alternative with its material risk.
• Native authorized launches receive the selected parameters. Other environments receive the exact two-line Launch / Why advice instead of an unsupported automation claim.
• Launch evidence separates requested settings, observable effective settings, delivery, outcome and material rework; completion alone does not prove that a model choice was sufficient or minimal.
• Simplified session identity specification in .tfw/conventions.md (§4), replacing cryptic BNF pseudo-code and unicode character references with clear formatting rules and representative examples.
• Normalized historical task grammar classification in .tfw/conventions.md (§4) from 2.0.0-dirty to legacy un-prefixed timestamp.
• Streamlined trace discipline in .tfw/conventions.md (§13) by removing references to the 2.1.0-retired debt snapshot.
• Cleaned closed internal bug references (TD-144, TD-178) from anti-pattern items in .tfw/conventions.md (§14).
• Hardened tfw-gates-only monitoring across persistent adapters: roles use addressed gate/status messages, bounded provider waits and durable returns, never foreign transcripts, session logs, unrelated task history or relayed unreturned reasoning.
• Compressed duplicated artifact inventories, template-owned semantic tables and the flat anti-pattern catalogue in .tfw/conventions.md while retaining one canonical coordination rule.

Fixed

• Stale Antigravity command routing in .tfw/conventions.md (§9 Tool Adapter Pattern) and .tfw/adapters/README.md: replaced deprecated .agents/workflows/ paths with .agents/skills/tfw-{command}/SKILL.md to match the adapter manifest and installed skills.
• Restored the canonical Workflow activation and routing contract after over-aggressive deduplication, including journal-first context, complete routing-spine validation, legacy-absence refusal, non-activating inputs, exact coordination authority and continuation provenance.
• Restored Coordinator ownership of research/iterations.yaml: Plan prepares each pending iteration before dispatch and closes it on return; Researcher refuses incomplete pipeline control state and never creates or edits the control file.
• Restored .claude/commands/tfw-plan.md as an exact copy of its canonical workflow.

Removed

• Pre-release historical polemic regarding 2.0.0-dirty.3 template naming from .tfw/conventions.md (§10.4).
• Internal maintainer-only morphological analysis terminology note (§14.1) from agent-facing .tfw/conventions.md.

Compatibility and updating

Read Updating to TFW 3.5.2, including from 3.5.1 and earlier. Refresh Plan, Research, conventions, glossary and the affected persistent adapters as one connected coordination group; install the canonical Claude Plan and Research copies. Existing task state and routing fields remain compatible and must not be rewritten.

[3.5.1] — 2026-09-22

RVAG — Reviewer Value-Assurance Gate. Review now judges the accepted result before process polish: purpose and domain behavior, material assurance, then authority/result identity and continuation. Administrative record defects no longer restart unchanged product execution unless they invalidate a material claim or authorized action.

Changed

• Each candidate finding is classified by subject, affected claim or authority, fact and oracle, concrete harm, material consequence, owner, completion condition, route and Candidate effect before one aggregate verdict is derived.
• Mixed rounds keep per-item product, assurance, authority and finite record-only routes. The highest required authority sequences the next act but does not reclassify lower items or move unchanged VALUE.
• Verification depth follows material claims, risk/criticality, dependencies, environment, oracle/authority, evidence gaps and stated limits instead of a changed-file population ratio.
• Permanent guards require a named protected behavior/invariant, failure consequence and relevant counterfactual detection. Positive controls, temporary diagnostics, counts and governance checks retain their narrower labels.
• A completed Reviewer sends one compact logical envelope after durable REVIEW and authorized trace: REVIEW · <reviewer-unit> · <task-or-phase> · <verdict> · <review-artifact@ref>. Delivery is provider-bounded and semantically idempotent; no universal exactly-once guarantee is claimed.

Fixed

• Prevented a breached citation, count, label or other process-shaped criterion from forcing REVISE without a named material consequence for the accepted result or authority.
• Prevented the highest rung in a mixed review from dragging record-only or assurance-only items into product re-execution.
• Preserved strong blocking behavior for wrong purpose/behavior, safety/security, human authority, accepted-result identity and insufficient material proof.

Removed

• Active tfw.review.min_verify_ratio from project and receiver-default configuration.
• The blanket any discrepancy → 100% verification escalation and historical firing-rate labels as current verdict inputs.

Compatibility and updating

Read Updating to TFW 3.5.1, including from 3.5.0 and earlier. Refresh the canonical Review, Review stage/artifact templates, conventions, glossary and Config workflow as one connected semantic group; update Claude full copies from their canonical workflows. Codex and Antigravity thin skills and accepted AGSK transport remain unchanged. Customized or ambiguous affected files refuse replacement instead of being silently overwritten.

[3.5.0] — 2026-09-22

FRATS & AGSK — Framework Refactoring, Explicit Coordination, and Antigravity Skill Migration. Coordination is now an explicit, task-local contract rather than an inferred autonomy mode or a chain of conversational agents, with transcript isolation and restored Strategic Architect planning. The active workflow corpus is compressed to ≤1,400 words per workflow, and the Antigravity adapter is unified with Codex around the shared .agents/skills/ surface, retiring legacy .agents/workflows/ and .agent/.

Changed

• Current task state carries one five-field routing spine: coordinator_route, owner_gateway, dialogue, activation, and coordination_authority. A /tfw-* invocation activates a role only when it names the exact skill, task/phase, and owner-direct, delegated, or continuation source.
• tfw-gates-only permits only vertical role-to-own-Coordinator status, gate, and durable-return traffic. Iterative dialogue is a separate, exact grant and requires an isolated GATEWAY that is neither the root Coordinator nor a workflow role.
• Plan again acts as a Strategic Architect: it works backwards from the stakeholder-visible finish, renders the future state, challenges consequential assumptions, applies the Saint-Exupery removal gate, selects coordination explicitly, and gives Executors functional freedom inside goal, value, boundaries, DoD, and DoF.
• Provider entry points disclose native provisioning, addressed send, wait/readback, and title capabilities before coordination is selected. Capability never grants authority or proves reliability; unavailable orchestration is reported instead of simulated.
• Ten canonical workflows own their rules once, stay within the 1,400-word safety ceiling, and regenerate twenty exact Claude/Antigravity projections. Repeated activation, return, knowledge, and handover instructions were consolidated at their actual owners.
• Current artifact issuance uses one deterministic grammar, including TS__{ID}.md for a single-phase TS, TS__phase-{x}__{phase_slug}.md for a phase TS, and __rev{N} only for a formal revision. Historical artifact names remain readable and are not renamed.
• Antigravity commands route through .agents/skills/tfw-{command}/SKILL.md (shared with Codex) instead of the deprecated .agents/workflows/tfw-{command}.md.
• The persistent Antigravity rule (.agents/rules/tfw.md) includes explicit coordination messaging instructions for vertical send_message communication under tfw-gates-only.
• Adapter documentation (.tfw/adapters/antigravity/README.md) documents cross-session addressed messaging mechanics.
• KNOWLEDGE.md Adapters row updated: 10 routes, .agents/skills/ surface, .agent fully retired.
• .tfw/glossary.md Tool Adapter definition updated to reference .agents/skills/.
• Adapter entry points in README.md, README.ru.md, README.kk.md updated to .agents/rules/tfw.md plus .agents/skills/tfw-*/SKILL.md.

Fixed

• Added Transcript isolation: no role may read, tail, resume, search, or reconstruct another active role's transcript, reasoning, tool output, terminal, or unreturned worktree for monitoring, review, trust validation, context recovery, or pre-solving. Use bounded wait/status, one addressed status request, and returned artifacts instead.
• Restored Coordinator purpose that earlier compression had reduced to routing mechanics, including Working Backwards, press-release preview, future-state rendering, uncomfortable questions, two-sided research hypotheses, and explicit checks for unnecessary scope.
• Repaired contradictory Reviewer ownership, stale configuration wording, a dead glossary anchor, inconsistent Phase/Step/Stage/Gate headings, and ambiguous single-phase artifact naming.
• Corrected legacy decision D75 from trajectory 310,485→112,536 (−63.8%) to the independently reproduced 310,485→112,206 (−63.9%); the active-corpus result and cited sources are unchanged.

Removed

• The .agent/ directory (singular) and its contents (rules/agents.md).
• The .agents/workflows/ directory and all 10 deprecated workflow files (tfw-plan.md through tfw-init.md).

Compatibility and updating

Read Updating to TFW 3.5.0, including from 3.4.1 and earlier. Downstream projects must classify .agent/ contents before deletion: framework-owned files are safe to remove; custom rules move to .agents/rules/; custom workflows convert to skills in .agents/skills/. Customized or ambiguous affected files refuse replacement instead of being silently deleted. Active tasks continuing under the new contract require the five routing spine fields (coordinator_route, owner_gateway, dialogue, activation, coordination_authority).

[3.4.1] — 2026-09-14

RWNR — Resume Without Resume. Returning and interrupted work now uses the existing /tfw-plan entry, which resolves task and phase state before selecting the next canonical workflow.

Changed

• Planning is the single public entry for both new and existing work. It preserves the selected task's authority, phase state, Role Lock, named AT principal/unit mandate and direct return route.
• Init, update and adapter inventories expose the exact ten-command surface and propagate retirement of an owned Resume installation without inventing a substitute command.

Removed

• The public /tfw-resume command, canonical Resume workflow, installed Codex skill and generated Claude/agent copies.
• Two transient maintainer tests whose assertions were coupled to the temporary pre-landing position of master rather than a stable framework contract.

Compatibility and updating

Read Updating to TFW 3.4.1, including from 3.4.0 and at equal-version retry. Refresh compatible planning/update readers before removing owned Resume files and registrations. Existing task state, phase state, journals, approvals and historical references remain readable; customized or ambiguous affected files refuse replacement instead of being silently deleted.

[3.4.0] — 2026-09-14

TKL — Team Knowledge Lifecycle. Material knowledge follows the work that produced it and is qualified where it is needed. Planning no longer stops for an unrelated global pending-task count.

Changed

• Producing roles preserve material insight in their existing handovers, including interrupted and stopped work. Missing context, justified-none, retain-only and owed publication remain distinct.
• Selected human knowledge and technical decisions use source-bound qualification and independent records. Current use checks scope, authority and incoming correction/conflict relations.
• The stable knowledge entry retains architecture/reference links, legacy topic and decision meaning. Actual capture and independent acceptance of changed final claims remain part of the existing PTTC close.
• Clean Full setup and bounded adoption use compatible readers before retiring the old gate and settings. Existing custom choices, history and accumulated knowledge remain preserved.

Added

• Upstream maintainers enforce a 10 MiB Git-blob limit on staged objects and complete release history. Five exact pre-policy evidence objects are retained; two excluded attachment identities are recorded in the policy manifest. This guard adds no receiver runtime.

Removed

• The live global pending/digest Knowledge Gate, its interval/mode and maintained fact/inventory limits, processed-marker/state writes and fresh knowledge-state scaffold. Existing state and markers remain inert historical evidence; no new counter, registry or receiver runtime replaces them.

Compatibility and updating

Read migrations/3.4.0.md, including from 3.3.0 and at equal-version retry. This release changes the live knowledge/configuration contract. Its entry reaches the earlier supported routes from the same pinned target and then migrations/knowledge-lifecycle.md. Complete any applicable SLC group before TKL effects; missing or conflicting preservation refuses the affected update. Do not reopen sealed tasks or manufacture completed consolidation. Full TFW is the selected scope; Light, Assisted, retrospective reporting and provider admission are unchanged.

[3.3.0] — 2026-09-09

SLC — Sealed Legacy Container. New Full projects start in workspace; existing projects retain their chosen working paths and can keep established history readable outside current-work checks.

Changed

• Clean Full acquisition and template initialization exclude upstream project state and history. New projects need one workspace folder, with no container interview or historical scaffold.
• Optional tfw.historical_containers keeps exact citations, history-only adapter repair and documentation reachable. Current discovery and the Knowledge Gate use active containers only.
• Exact temporary mixed layouts migrate only under determinate state and established disposition. Existing single/custom choices and settled keep-active decisions remain valid.
• Historical reads preserve original task paths, identifiers, bytes and lifecycle truth. Compiler references use the same whole-ID collision checks and stable documentation destinations.

Compatibility and updating

Read Updating to TFW 3.3.0, including from 3.2.0 and at equal-version retry. Earlier supported tags follow that guide's ordered predecessor route from the same pinned target. It quotes the retired mixed-list rule and the bounded successor to blanket state preservation: preserve exact historical digest pairs before removing only those keys, then narrow active config. Expected partial states recover; conflicting affected inputs refuse without erasing later work. No history is moved or marked complete; no runtime, archive controller or general gate reset is added.

[3.2.0] — 2026-09-09

PTTC — Proportionate Testing and Task Closure. The Saint-Exupery principle guides this release: remove unnecessary dependencies and obligations while retaining meaningful protection and honest completion. Existing records, settings, authority and independent review remain compatible.

Changed

• The existing Coordinator owns Full task closure. After independent approval, it resolves dispositions and applicable docs/knowledge effects, obtains independent assessment of changed final claims, checks actual selected landing, and only then writes valid terminal records.
• Evidence can be reused when relevant inputs, output, oracle, authority and environment assumptions remain applicable. A changed enclosing commit or an unrelated task alone is not a reason to repeat product work; a material change or insufficient evidence still requires affected checks.
• An already selected, authorized close can resume directly. It does not require a new planning selection, phase matrix, role holder or capture cycle merely because the Coordinator returned.
• Upstream source/Git/temp-tree contracts are separated from generated-output tests. Pure checks no longer acquire the MkDocs fixture. Maintainer guidance selects checks by the consequence of a change, including broad checks when release, dependency or output risk requires them.

Removed

• Duplicated Reviewer capture/closing duties, marker-only DONE, and the requirement to run a new build/test command solely to fill a review checkpoint.
• Overwritten duplicate test definitions, redundant aliases and orphan helpers in the affected maintainer family. Historical wording checks remain at their immutable acceptance sources; current structural, semantic and provenance checks remain active.

Fixed

• Reconstructable closing-record errors can be repaired without replaying accepted product work. Original events remain immutable. Unknown acceptance, changed authority or oracle, and real defects cannot use this administrative recovery route.
• Terminal writes validate the whole status/event pair and its actual references before writing, rather than treating one changed field or two capture markers as sufficient proof of completion.

Compatibility and updating

Updating agents must read Updating to TFW 3.2.0, including from 3.1.0. Pin the published v3.2.0 object and use its update workflow. Earlier receivers follow the guide's ordered routes and all still-needed migration obligations from the same target. No new configuration key, lifecycle ID, event kind, runtime, global quota or registry is introduced. Existing project checks and Assisted/Light contracts are unchanged; upstream Python/pytest/MkDocs and Git history are not receiver prerequisites. Sync the selected canonical instructions and installed adapters together, preserving project-owned state and historical approval epochs.

Bounded native observations and independent review support the selected changes; they do not prove universal agent compliance, broader provider admission, or universal time/token savings. Review, landing and publication remain distinct; task closure grants no release or deployment authority.

[3.1.0] — 2026-09-08

CRUE — Clear Release and Update Experience. This release makes the update instructions responsible for discovering technical facts, preserving the receiving project and explaining the actual result. It builds on 3.0.0; it does not automatically select autonomous mode or extend provider admission.

Changed

• Ordinary updates discover existing identity, task containers, checks and customizations before asking a question. Connected framework changes no longer require a fixed interview or one approval per file; unresolved consequential project choices still reach the owner.
• Current framework values in .tfw/README.md ship with the framework. A customized or historically cited receiver copy is preserved byte-for-byte before replacement; the project's root README, purpose, settings, knowledge, profiles and task history remain project-owned.
• Update explanations lead with the actual outcome, useful project-facing benefits, preservation, limitations and the next action. They use release facts and observed run facts without inventing successful checks, savings, reliability or owner comprehension.
• Release preparation follows the receiving project's own output, audience, versioning and checks. A project without releases remains valid. This repository's concrete release procedure lives in its own RELEASE.md, not in a universal requirement imposed on receiving projects.
• Selected, inspected, stable sibling traces may accompany an authorized exact-path commit without requiring their task to be DONE. Operational, private, authority-bearing and mixed changes retain their own gates; an unrelated arrival does not silently change the selected release composition.

Added

• A project-owned immutable update-attempt record with exact source, applied/skipped effects, preservation references, check results, cleanup disposition and recovery/next action. It is history, not an installation lock or a replacement for observing current files.
• A version-addressed 3.1.0 migration guide, including earlier-version ordering, explicit replacements for retired instructions and useful user onboarding.

Fixed

• Minor and patch migration guides named by intervening changelog entries are read from the same pinned target, not skipped merely because no major boundary was crossed.
• Equal version numbers no longer end an interrupted update without checking actual source and state. An explicitly authorized untagged source records its full commit instead of inventing a release tag.
• Cleanup is resolved after checks; the immutable receipt is sealed next; the final message is rendered last. A receipt cannot claim future message delivery or understanding.
• Adapter repair preserves an existing project instead of restarting initialization. Canonical and installed instructions are synchronized; Antigravity uses the manifest's plural .agents targets without treating legacy rule compatibility as proof of legacy workflow discovery.
• Maintainer release checks no longer require the current files to stay identical to 3.0.0 forever. Historical package integrity and current version/changelog/migration consistency are checked separately.
• The optional doctor's controlled-fixture tests remain; an unrelated finding anywhere in the live task corpus is no longer turned into a blanket release failure by a global zero-findings test.

Compatibility and updating

Updating agents must read Updating to TFW 3.1.0, even from 3.0.0. Pin v3.1.0, read its .tfw/workflows/update.md, and use the guide's explicit successor rules when older guidance differs. From below stable 2.0.0 first read the 2.0.0 migration. The 3.1.0 guide orders all still-applicable 2.0/2.1/2.2/3.0 obligations, including dirty tags; no intermediate payload installation is required. Unknown/custom and 0.x receivers need a project-specific migration plan. Do not copy task traces, knowledge state, team profiles or another project's release rules.

No additional configuration keys are retired by 3.1.0. Preserve receiver-owned values and execute older key migrations only where still needed. Historical migrations and approved task artifacts retain their recorded meaning; the new instructions apply prospectively, not by rewriting history.

Verification is bounded: source/adapter and synthetic checks do not prove universal native behavior or human comprehension. The one pre-correction field campaign had four environment stops and two changed Claude copies with unverified/unavailable outcomes; it was not six successful updates. Final corrections were locally verified and independently reviewed, without a second native campaign.

[3.0.0] — 2026-09-07

Changed

• Agent Team is now a provider-neutral, explicitly owner-selected execution mode: one stable LEAD principal is attribution, while Coordinator, Researcher, Executor, and Reviewer remain distinct directly addressable working units. Human-rooted initiation, immutable mandate/grant ceilings, preserved proposal origin, bounded dispatch, exact return channels, and independent review remain separate from profile, binding, title, provider, role, and writer metadata.
• The methodology is provider-neutral, but the first 3.0.0 implementation is Codex-first and admits only provider-homogeneous long-lived role chains. A complete Claude-only chain requires its own native proof gate before admission; fresh cross-provider runs are bounded helpers only and cannot route into another provider's existing long-lived task.
• Concurrent mutation uses isolated worktrees, exact-path staging, producer-attributed landing commits, and retained Candidate reachability. Session identity is navigation-only; only the exact LEAD root Coordinator renders the LEAD title, and same-principal children retain ordinary role titles.

• Handoff, Research, and Review set optional event writer only when Who Is Acting resolves a declared principal; otherwise they omit it. They never create one profile per session, and a shared principal grants a child no unit identity or authority.

• Task-local status.md, journals, and artifacts are now the only ordinary Full discovery inputs. The shared portfolio cache, generator, freshness check, default public Tasks navigation, and root-guide catalogue routes are removed. Generated documentation still copies every task Markdown source and creates an unlisted landing per recognized task so direct and cited links remain resolvable, including tasks without an HL.

• The Knowledge Gate is complete in .tfw/workflows/knowledge.md and linked from Plan. It semantically defines configuration/state reads, whole-ID discovery, ambiguity handling, selected section extraction, tuple hashing, pending/removed reconciliation, threshold outcomes, and the hard stop on uncertainty without requiring a repository helper, Python, or PyYAML.
• Event summary and status title, goal, value, outcome, and lifecycle_verbatim retain complete one-line prose. Numeric ceilings and migration truncation are removed; “brief, normally no more than 120 code points” remains summary-writing advice only. Structural carrier, identity, accountability, reference, time, lifecycle, and transition checks remain material.

Added

• Canonical glossary routers now name Principal, Initiation Chain, Worktree Protocol, Landing Commit, and AT while leaving each normative procedure at its single owner.

• The reviewed Phase E release package fixes the six destination contents and SHA-256 digests before release application. The release commit is separate from implementation/closure and precedes any independently authorized tag, push, publication, or deployment.

• Upstream-only tools/tfw_state.py provides side-effect-free semantic readers and tools/tfw_doctor.py provides exactly four optional read-only operations: status, check tasks, check project, and knowledge-pending, with deterministic human/JSON output and exits 0/1/2 for clean/material/indeterminate results. These tools and their Python/PyYAML environment are outside the copied Full payload and Assisted and are not ordinary gates.

• The pre-2.0 conversion remains available as the self-contained tools/migrations/2.0.0/ bundle acquired from the same immutable upstream release ref as the target payload. Its help works without PyYAML; a missing dependency exits 2 before project reads or writes. The root documentation workflow may upload an explicitly requested disposable status artifact for seven days; default push and dispatch publish none and diagnostic exit never gates.

Removed

• .tfw/scripts/gen_index.py, .tfw/scripts/migrate_board.py, their payload tests, and tracked workspace/00-INDEX.md. Remove stale receiver copies after preserving project-owned state.
• tfw.journal.max_summary_length, tfw.knowledge.max_index_lines, and tfw.knowledge.max_index_facts_lines. Do not map them to another validity rule.

Compatibility and updating

Read the target's .tfw/workflows/update.md, not the installed one. Pin v3.0.0, verify its version, and follow .tfw/migrations/3.0.0.md. Receivers on 2.2.0 use that guide directly. Receivers on 2.1.0, 2.0.0, a 2.0.0-dirty* tag, or 1.x first follow the ordered prior routes named there; 0.x and unknown/custom receivers require a project-specific migration plan. Intermediate tags need not be installed, but their still-applicable migration actions may not be skipped.

The retired current-workflow instruction is: “A writer is not named yet — that is TFW-54 — so do not create a profile per session.” Replace current payload copies through the canonical update route; preserve historical true-at-the-time artifacts. If a receiver created per-session profiles, preserve the trace and use a separately reviewed owner-authorized migration to stable principals and external machine bindings.

Pin the target release first, acquire tools/migrations/2.0.0/ from that same immutable ref when pre-2.0 conversion is still needed, and follow the target update.md. Remove obsolete payload files only after exact-path comparison. Preserve receiver config, knowledge state, North Star, team, task history, journals, snapshots, and unrelated build commands. Review project-owned build.*: replace stale .tfw/scripts/ or gen_index.py commands with real project checks, and do not add the optional doctor as a default live-corpus gate. Remove the three retired config keys from both active config and local templates.

Verify VERSION/config agreement, adapter parity, absence of retired paths/keys/current references, configured task-container resolution, selected status/event structural postconditions, receiver build commands, the target release's test procedure, and zero Assisted/runtime dependency growth. Rollback by restoring the pre-update payload/config copies and exact removed paths from the clean backup or prior commit; never rewrite task history or immutable events. These semantics begin with the eventual release containing this Unreleased entry and with each task's applicable TS approval epoch; historical approved work retains its recorded rules. This entry deliberately chooses no release number.

[2.2.0] — 2026-09-06

Additive supersession note (2026-09-06): when a later pinned target contains the RTBO payload-tool removal, use that target's Unreleased/released mapping and the self-contained tools/migrations/2.0.0/ bundle at the same immutable ref. The 2.2.0 commands below remain the historical procedure for 2.2.0 and are not rewritten retroactively.

Preparatory release for field use before complete CRATM. It includes the completed RCFR, VBSA and RTPSN tasks and completed CRATM phases A–C. CRATM remains open; D/E, RTBO and APD are not completed or advanced by this release. Internal research/task traces are not features of the payload.

Version decision: on 2026-09-06 the owner reserved 3.0.0 for complete CRATM and authorized 2.2.0 for these accumulated changes. This is an explicit exception to the usual MAJOR classification of removed/renamed configuration keys and prospective contract changes, not a claim that migration is unnecessary. /tfw-release records the exception in RELEASE.md.

Added

• One adapter manifest owns persistent roots, command routes, workflow roles and copy strategies. Installation/update checks cover clean receivers, repeatability, exact copies and preservation of project-owned text. Codex uses thin local skills; full-copy adapters keep canonical workflow bytes.
• Stable project-local principals can be human or agent. Optional descriptive organization/project roles do not grant permission. An agent names a human in accountable_to and one immutable Boolean may_rule_amendments; grant changes require a new handle, not a changed old grant. Optional event writer identifies the principal separately from human on_behalf_of, tool via and filename token.
• Session identity is resolved from authoritative task/phase state before applicable role work. The title uses {WORK} · {TASK}[ · {PHASE}], adds a stable disambiguator only on collision, and claims success only after readback. Lack of host support is reported once and does not block the workflow.

Changed

• RCFR — selective context and knowledge reconciliation. Canonical workflows own ordered Read Contracts; entry points no longer preload the common libraries. Required rules, artifact lineage, dynamic reads and hard stops remain. Knowledge reconciliation uses content digests per full task identity rather than a sequence/date/task cursor; migration is approved and writes state last.
• Scope accounting now governs only the declared value-bearing surface. Every planned path has one semantic class (VALUE, ASSURANCE, TRACE, or DERIVED), while the universal delivery measures are exactly logical touched VALUE files and touched text LOC between one immutable Baseline and Candidate. Ordinary assurance and lifecycle trace remain mandatory without spending that budget.
• Planning declares the selector, immutable denominator, soft-trigger disposition, prospective authority, and any M1–M6 hard constraint. Execution fixes Candidate after required VALUE/ASSURANCE work and tests but before EV/RF. RF and exactly one EV accounting row bind the result; REVIEW independently replays it.
• The local .tfw/README.md North Star adds the Saint-Exupéry Principle. Plan uses its judgment universally; init and update preserve receiver North Stars and never inject this starter repository's quotation.
• CRATM A — isolation and attribution. Concurrent mutating runs use separate Git worktrees. Shared-tree commits select exact paths. Reviewed cross-session landing identifies the producing task/phase and retains the exact Candidate through cleanup; worktree isolation is not a lock.
• CRATM C — bounded amendment authority. Ordinary CL remains owner-only. Claimed delegation requires a declared human owner, a separately authorized root Coordinator and an unambiguous child-only initiation chain. Preserve the originating proposer across transcription and sessions; ordinary amendments reach the nearest eligible non-proposer with an immutable grant, otherwise the human owner. Malformed provenance refuses; reserved decisions and self-grant changes remain human.
• RTPSN — command-entry evidence is explicit. Availability, dispatch and observed workflow behavior are different claims. Thin Codex entries remain the baseline: the comparison did not establish that another entry architecture was more reliable. Session naming is navigation, not authority.

Fixed

• Rung-specific REVISE routing. Rung 1 keeps the approved TS and records the Coordinator's bound in the live REVIEW; any rung 2 produces a governing TS revision for the mixed round; rung 3 waits for a valid amendment verdict. A review round alone no longer forces a new TS. Reviewers propose dispositions; Coordinators rule them; Executors do not authorize their own corrections.
• Release verification survives publication. The scope-migration test now resolves the introducing changelog entry before or after its move out of [Unreleased]. Missing migration text, ambiguous duplicate entries and a future introducing version still fail. It no longer mistakes release bookkeeping for loss of the migration contract.
• New-project configuration names the installed release. The starter config template now matches VERSION and the framework's own config; a release regression checks all version carriers together, preventing initialization from inheriting an obsolete version number.

Removed

• Separate max_new_files and max_modified_files scope-budget sublimits. The new contract measures touched VALUE files and LOC instead of reconciling overlapping file buckets; planned limits, prospective authority and required assurance remain. Preserve the receiver's applicable numbers through the mapping below rather than substituting starter defaults.

Compatibility and updating

Read the target's .tfw/workflows/update.md, not the installed one. Pin v2.2.0 and follow Updating to TFW 2.2.0: the updating agent must read this guide even for the minor-version transition from 2.1.0. It includes parameter migration, preservation, verification and required positive user onboarding. If not yet on stable 2.0.0, first read the 2.0.0 migration and account for its applicable steps without redoing completed migration. From 2.0.0 also perform 2.1.0 Updating from 2.0.0 and its /tfw-task removal section. From a dirty tag also follow 2.0.0 Updating from a 2.0.0-dirty tag and the applicable .5/.4/.3/.2 sections named by the guide. From 1.x first use the 2.0.0 migration guide. From 0.x or unknown/custom provenance, stop for a project-specific migration plan; no verified automatic route is claimed. Keep the final target pinned throughout; historical examples naming older tags do not change it.

This change is prospective by release and TS approval epoch. Existing approved TS files and historical results retain their recorded semantics; do not rewrite them or infer their meaning from a newer install.

For a receiver whose tfw.scope_budgets still contains only the retired block, apply this mapping in one approved config merge while preserving the receiver's numbers:

Retired key	New treatment
max_files_per_phase	Rename to decomposition_trigger_files; preserve its value
max_loc	Rename to decomposition_trigger_loc; preserve its value
max_new_files	Remove; do not reinterpret it as total files
max_modified_files	Remove; do not reinterpret it as total files
—	Add owner_escalation_multiplier: 2

The retired configuration wording, quoted verbatim for discovery, is:

max_files_per_phase: 50
max_new_files: 50
max_loc: 5000
max_modified_files: 50

After updating, declare class/reason per path in each newly approved TS, use only the two VALUE measures, record a terminal decomposition disposition, and obtain prospective Owner authority at the configured multiplier or from an applicable planned zero. A mixed old/new block is not auto-merged; stop for an owner ruling. New-only blocks keep their configured values. The complete preservation/removal/default rules also appear in the version-named migration guide; the owner's 2.2.0 numbering decision does not waive them.

Preserve the receiver's North Star, knowledge, task history, team and machine-local bindings. Do not copy this starter's .tfw/knowledge_state.yaml: its old cursor is reconciled by the receiver's next approved /tfw-knowledge run, not silently reset during the payload update. Existing human profiles, legacy actor events and old session titles need no retrospective rewrite. Synchronize workflows and selected adapter roots/commands together, preserving local customizations through the approved merge.

The board/task reporter remains report-only. Recheck an inherited build.verify that invokes gen_index.py --check tasks, even when arriving from 2.1.0: that version's updating section removes the inherited gate, while unrelated project verification stays. The unchanged reader can report unknown keys: writer for the current optional field. Validate new writes against the canonical template; do not rewrite history to satisfy that reporter. A project deliberately using it as a gate must agree its treatment before adopting the new form. This is a known limitation, not a new gate.

Retired wording and the receiver's action

Search only live local rules/copies with these verbatim strings. Historical records, dated retractions, and the explicitly report-only reader may retain the old text; do not rewrite them.

Retired wording, verbatim search string	Replacement/action
On a new session, load	Remove the old unconditional common-library preload from the managed TFW entry block; use the target workflow's ordered Read Contract. Keep unrelated project startup instructions
The round's order is a TS revision	Replace the universal rule with the canonical rung table; do not rewrite prior rounds
A proposal is ruled only by an explicit owner verdict	Ordinary CL still reaches the owner; an explicitly authorized delegated chain uses the resolved-ruler rule. Do not infer a grant from a profile/title
The four keys below are the whole schema	Existing human profiles remain valid; optional descriptive fields and the explicitly defined agent fields are now permitted
`type: agent` IS ADMITTED BY THE SCHEMA AND USABLE BY NOTHING.	A stable agent principal is now usable under the current profile/authority rules, never a profile per session
A writer is not named yet	Current events may name a stable principal in optional writer; legacy actor and filename tokens remain unchanged
last_consolidation_seq / last_consolidation_task	Preserve legacy knowledge state until the approved full digest reconciliation succeeds; remove cursors only in its state-last write

Verification and release boundary

Maintainer checks: python -m pytest .tfw/scripts/ docs/scripts/ -q --collect-only, python -m pytest .tfw/scripts/ docs/scripts/ -q, python .tfw/scripts/gen_index.py --check project, and git diff --check. The suite includes receiver/config preservation, adapter parity, source-sensitive routing and the before/after-publication regression; receiving projects run their own approved build commands.

Full Role Assignment, declared autonomy-start semantics and provider-specific autonomous team profiles are not shipped. CRATM D/E and their remaining glossary/reference sweep stay pending. These foundations do not certify Claude-native or mixed-provider long-lived teams. Pilot ordinary work first; the release does not claim a downstream project has already completed its update or a measured live token saving.

[2.1.0] — 2026-09-03

The project debt registry is retired. Root TECH_DEBT.md leaves the project root and is sealed verbatim, unexamined, as tasks/DEBT-SNAPSHOT.md beside the Task Board retired at 2.0.0. Debt found in a review is written once, in the REVIEW that found it, and carries a disposition before its task can close. Nothing was created to replace the registry: the net count of maintained project artifacts falls by one and rises by none.

Not breaking, and nothing is forbidden. The obligation to maintain a project-level registry is withdrawn. Your file is your own — a project that keeps one, and reads it, is not doing anything wrong.

And a review's output becomes decisions by rule. A finding got three statements and no criterion, and a 🔄 REVISE got one destination on a lifecycle diagram that nothing implemented — the word revision appeared in the canon zero times. Now: one axis quoted from NS1 names which harms count, a named consequence decides the case, three rungs each reach a role that can discharge the item, the coordinator rules dispositions once at the close of review, the round itself is ordered in a TS revision — the coordinator's own artifact, where an executor already looks — and the loop ends on a citation bar: a round may order only items naming the condition each breaches, and when nothing can be cited the verdict is ✅ APPROVE with the remainder disposed.

Nothing was created for it, and one thing was given back. Every site already shipped — the pending — {role} state, TS_DRAFT, the amendment_escalated event kind, status.md owner, and the rev{N} sibling this corpus already used for REVIEW files. A configured budget on revisions was built inside this entry and withdrawn inside it, with its one configuration key, before release: a budget ends the loop in exhaustion, which is the ending this protocol's own principles forbid, and the research behind it found the cause it was built for in none of eleven measured deep rounds. So the release adds no configuration key, and review.md carries the whole protocol and is seven words shorter than the version it replaced.

Why this release exists

A registry every review was obliged to feed, and no workflow was obliged to act on, accumulates by construction. Measured across 25 projects on 2026-09-01: not one registry empty; 23 of 25 running the canonical flat table with zero task traces scoping any work from it; the two that do consume one had first rebuilt the artifact themselves and paid for the rebuild. This repository's own file grew from 1 463 to 12 352 words in six weeks — +247% in 19 days — reaching 121 rows, while a quality filter demanding "only promote items that would cause real problems if left unfixed" was in force the whole time. Of those 121, 74 still carry an open marker and 44 a closed one, and three carry neither — the Status column was free prose, so the split depends on the test that reads it, and that is itself part of what a registry costs. (Re-measured 2026-09-02; the row count is grep -c '^| TD-' over the sealed content, the split an awk on the last column. An earlier draft of this entry said "77 open" and no test reproduces it.) A filter in front of an open channel had already been tried.

So the channel is closed rather than filtered harder, and the outlet is replaced by a decision taken at the one moment it was ever going to be taken — while the task is still open.

⚠️ Changed

• Debt is task-local and disposed of before closing. review.md Step 5 loses its second write and gains the gate: every captured item carries paid as a phase of this task, promoted to a task created then and there, or not material ruled on the record. Three outcomes and no fourth, and a disposition must name something that already exists — → backlog, "someone should open a task" and "the next scripts pass" name nothing and are not dispositions. pending — owner is a legal waiting state that keeps the task open. A task does not reach DONE with an item undisposed.
• Discovery replaces maintenance. templates/REVIEW.md §5 states, literally, the one search that lists every captured item across REVIEW files. Its answer is re-run, never inherited, which is why no row count is quoted here: the search is the figure, and a number beside it would be one more thing to keep true. It shipped in review.md Step 5 and moved on the day of release: a template is not an installed adapter copy, a workflow is, and an adapter harness rewrites positional parameters in a workflow before the agent reads it while cmp stays green. The command belongs in a file no harness touches.
• REVIEW.md §5 is renamed Tech Debt Collected and Disposed and its Action column becomes Disposition. The words Tech Debt Collected are kept in the heading on purpose: every REVIEW file already written carries them, and the search matches on them.
• judge.md row 3 changes from Tech debt documented to Debt disposed, and states on its face why it is kept — on consequence, not rate: the undisposed item is what filled the retired registry.
• docs.md no longer reads or writes debt; its checklist loses row 4 and renumbers 5→4, 6→5.
• resume.md loses step 9 and both anti-patterns that obliged acting on what step 9 read; steps 10–15 renumber to 9–14. The workflow itself is unchanged otherwise.
• init.md no longer creates the file in a new project; Phase 4 renumbers 4→3, 5→4, 6→5, 7→6.
• compilable_contract.md deletes manifest row 9 and renumbers 10–14 → 9–13. TD-{N} resolves to tasks/DEBT-SNAPSHOT.md, which the task-container glob already compiles — one source, one page, exactly as for BOARD-SNAPSHOT.md. No citation renumbered; no redirect layer.
• .tfw/migrations/2.0.0.md gains step 6, Retire the debt registry; old 6 and 7 become 7 and 8.
• review.md Steps 4–6 carry the whole protocol at 477 words, against a 483-word baseline, and Anti-patterns at 160 against 163. Both re-measured before the tag — awk '/^## Step 4/{f=1} f&&/^## Step 7/{exit} f' .tfw/workflows/review.md | wc -w, and the same shape for the Anti-patterns block. Step 5 becomes Findings — locate, test, route, propose: an entry filter that asks only whether an observation is real, the NS1 axis, the named-consequence test, rung routing, and a proposed disposition per item. Step 4 gains the citation bar. Step 6 becomes Rule, then update traces, where the coordinator rules every proposal in one act and — on a 🔄 REVISE — returns the work to the coordinator without ordering anything: no bound, no dispatch.
• conventions.md §5 gains The 🔄 REVISE route. Rung 1 returns to execution and costs no escalation; rung 2 waits as pending — coordinator beside its item, with lifecycle: TS_DRAFT moving only when the TS actually changes, once per round; rung 3 reaches the owner through an amendment_escalated event and an HL §12 row. A rung belongs to the item and lifecycle to the task, which is why one REVISE can carry both at once — the ordinary case, and the one the old wording could not represent.
• A revision is defined, and it is the artifact a round is ordered in. A revision is repair of what was already specified: a new TS for an approved phase, or a correction to the existing one — not a review round, and not new work, because only a change of the task's declared outcome is that. The test is did the declared outcome change, never can the existing TS accept it: a rung-2 finding fails the second by construction. glossary.md gains Revision, Citation bar and Rung.

Superseded by .tfw/CHANGELOG.md [2.2.0], Fixed / Rung-specific REVISE routing (2026-09-06). The historical paragraph below is preserved; its universal TS-revision rule no longer governs new rounds.

• The round's order is a TS revision — the coordinator's own artifact. A 🔄 REVISE returns the work to the coordinator, who writes TS__{ID}__rev{N}.md naming the round, who ordered it, each item's basis, what is not re-done, and its approval. The executor is already obliged to read the TS, so the order cannot be missed and no detector exists to find it. Before this, the coordinator wrote the round into REVIEW §4 — the reviewer's file — and every consequence followed from that one misplacement: the executor needed a detector, status.md could not say a round was open, and the round was invisible in a directory listing.
• The revision grammar is stated once, in conventions.md §4, and one line generates it: sibling where exactly one must govern; appended where the record is cumulative. The TS and the REVIEW take siblings, …__rev{N}.md, and the highest ordinal governs; the RF and the ONB are appended to, one new numbered subsection per round in every section the round touches. The unsuffixed file is revision 1 and is never renamed. An ordinal is the only suffix the grammar admits — a title suffix duplicates what status.md holds, and the standing bar on those is left exactly as it was. Measured cause: PROPOSAL__TFW-58__revise_protocol — "the TS was overwritten in place; revisions 2 → 3 → 4 recorded only as header prose."
• The loop ends on a citation bar, not on a count. A round may order only items that name the condition each breaches — an acceptance criterion of the approved TS, or a frozen HL claim; everything else is disposed of. When nothing can be cited the verdict is ✅ APPROVE with the remainder disposed. A reviewer who can neither cite nor approve stops the work and returns it to the owner handle in the task's status.md, as a transition to ❌ BLOCKED naming no basis can be stated as the blocker. owner may be a human or an agent, and an agent applies the same rule upward; unassigned is a hard stop naming itself. No kind was invented — the journal vocabulary stays closed. The bar's enforcement site is structural rather than prose: the basis cell in the round's order, where an item citing nothing has nowhere to sit.

Superseded by .tfw/CHANGELOG.md [2.2.0], Fixed / Rung-specific REVISE routing (2026-09-06). Lifecycle follows the governing rung, not an unconditional TS_DRAFT step for every review round.

• A 🔄 REVISE names its lifecycle states, and there are two. conventions.md §5's verdict list now reads 🟡 TS_DRAFT while the coordinator writes the round's order, then 🟠 ONB when the executor takes it. One value could not describe both legs, and the earlier wording — back to execution — was written before the coordinator's leg existed. (develop) leaves the §5 flow diagram: it was never a state, only the activity inside ONB, and drawing it as one is what made a missing status look real. ONB's description in both config files now covers the executor's whole leg — the report written, the work under way, and re-entry after a REVISE. No status was added and none renamed, so the vocabulary enumerated in seventeen files is untouched.
• A configured budget on revisions was introduced and withdrawn inside this same unreleased entry, and so was its one configuration key. Nothing outside this repository ever carried either — 2.1.0 is not cut and v2.0.0-dirty.5 is the last tag — so there is no migration step and nothing for a receiving project to remove. Why it went: a budget ends the loop in exhaustion, which this protocol's own principle 4 forbids in those words; the research behind it found cause specification in zero of eleven measured deep rounds, so the count never had the basis it was given; it fired on the legitimate case within hours, spending revision 2 of 2 on an owner-ordered correction and leaving the task unable to finish itself; and it cost no new entities its absolute for one key. The withdrawal is recorded rather than quietly dropped, because a version that hides a reversal is worse than one that reports it.
• Acceptance authority is named where the role is defined. conventions.md §15's review.md row now reads Reviewer — marks and proposes, with disposition rulings in its forbidden column and the Coordinator holding acceptance authority. A reviewer that ruled dispositions would be deciding the consequences of its own findings — the argument that split review from handoff, one level on.
• handoff.md gains Returning after a 🔄 REVISE — what to read, what is not re-done, what is not the executor's, and the round's artifacts. A returning executor is sent to the current TS revision, the highest ordinal, with the prior REVIEW behind it as the reasoning. Its Context Loading gains the prior REVIEW as item 8 and runs to ten; the Codex tfw-handoff skill's load list gains it too, so the two adapters instruct alike.
• plan.md gains a numbered step for the coordinator's post-review act, with its two writes inline: promote what needs its own contract, each with a PROPOSAL, then write the round's TS revision, each item carrying its basis. conventions.md §15 keeps the boundary — receive, write, hand off, do not execute — and the reference runs one way, so the two writes are stated in one place. Without the numbered step, §15 and a pointer line are not something a coordinator running /tfw-plan walks past, and the no-guess guarantee held for three roles and failed for the one the change is about.
• templates/REVIEW.md §5 carries the ruling's grammar, pending — coordinator beside pending — owner, and the relocated project-wide search. judge.md row 3 checks the grammar and names the coordinator as the ruler — no row was added; the firing-rate rule applies to a new row, not to a sharper one.
• Known staleness, named rather than shipped silently: resume.md step 8 still reads "If a REVISE verdict exists for any phase — flag it as needing re-execution". A flag is not the route this release defines. The owner ruled resume.md obsolete on 2026-09-01 and its deletion a separate task, so the file is untouched and the contradiction is recorded here instead.
• conventions.md §11 gains one design rule, and §14 the matching anti-pattern: a command written into a workflow may not use $0–$9 or $ARGUMENTS, because an adapter harness substitutes those in command text before the agent reads it. Found by this release's own first review — the debt search reached the reviewer with both $0 occurrences replaced by the invocation argument while cmp stayed green, because nothing had been copied wrong. The search is now written without any $N.
• .tfw/workflows/handoff.md loses a step, and the executor's states renumber. The step that set lifecycle: RF at the head of Phase 2: Execution is deleted, not rewritten: RF means execution complete, RF written, and setting it before the work begins wrote a state the canon defines as false. ONB is set once, at the end of onboarding, and covers the executor's whole leg; the RF transition is now the last step of Phase 3, once the RF file exists. Phase 2 and Phase 3 steps renumber 8–13 → 7–12, with the new state step at 13, and the four live citations of those numbers — glossary.md's Evidence Collection, Pre-RF Gate and Execution Loop, and this repository's KNOWLEDGE.md D52 — now name the step by its title instead, because a step number is a locator into a file that renumbers. What you do: nothing, unless your project quotes handoff.md step numbers of its own — search for them and name the step instead. Search rather than recall: the first pass here named the two it remembered, and the other two were in a second file and a second phrasing.
• .tfw/templates/REVIEW.md §4 stops inviting an order. ### If REVISE — items to fix: becomes ### If REVISE — items proposed to the coordinator:, and its placeholder now asks for the condition each item breaches. The heading was the site the defect used: under it, a coordinator wrote a round's bound into the reviewer's own artifact. review.md Step 6 and conventions.md §15 already forbid that; the template a reviewer opens now says the same thing. Nothing was added — one heading and one placeholder, same section count. What you do: nothing. Existing REVIEW files keep their heading and stay searchable; the change is to the template new reviews start from.
• conventions.md §4's revision grammar classifies the EV file. The table covered TS, REVIEW, RF and ONB — four artifacts, where a round produces five. The EV file is appended, by the same generating line as the rest: nothing about it governs, and an earlier round's verification does not stop being true, so a later round has nothing to supersede. What you do: nothing. It states a rule for what was previously decided by analogy, and the analogy pointed the same way.
• glossary.md's Disposition says what a paid ruling requires. A paid ruling names the phase that pays it, and where the payment has not happened yet the same act must order it — in a round, citing the item's condition. Unordered, paid accepts an item without a decision, which is deferral under a new name and is what the three-outcome rule exists to prevent. What you do: nothing, unless you hold open paid rows with no ordered work behind them — those are the rows this sentence names.

Fixed

• 860 of 990 generated pages rendered their own YAML frontmatter as body text — every task page, every knowledge topic, every workflow and template reference page. Three independent causes with one symptom, and each is fixed by removing what made it possible rather than by escaping harder:
• the source path reached source: OS-native, because gen_docs.py hand-rolled str(p).replace("\\", "/") at thirteen sites and at a fourteenth did not — all thirteen are gone, replaced by Path.as_posix(), net one line shorter. That same string is the key path_map is built with and the basis of the source directory in rewrite_markdown_links(), so the self-lookup and every relative-link rewrite on those pages were wrong too;
• the block was assembled by string interpolation, so a title holding a double quote — and titles in this corpus do — closed the scalar early. It is now serialized with yaml.safe_dump, already a dependency; a colon, a leading % or # are handled by the same change;
• resolve_references() ran over the header as well as the body and turned a bare task id inside a title into a markdown link. Frontmatter is now added last, so the four transforms see the body only and no future transform can reach the header either. Invisible on the POSIX CI that gates merges for the first cause, which is why it survived. Three tests now cover it, each firing where the others cannot: source-level and platform-independent, unit-level over hostile titles, and an end-to-end scan of the built site.

Removed

• TECH_DEBT.md from the project root. Its 132 lines and 12 352 words are byte-identical in tasks/DEBT-SNAPSHOT.md, under a header stating the counts, the source revision c153895, and that the rows were sealed unexamined and age is not evidence of importance.
• Every canonical instruction to append to a project-level debt list, in the canon, four templates, the compilation contract, three adapter sets and the TFW:CLAUDE block. No new file, template, script, check or configuration key replaces any of it.
• build.verify — the board's checker, out of every blocking position. Gone from .tfw/project_config.yaml and from .tfw/templates/project_config.yaml, so no new project inherits it; with it, the Verify reporting line from .tfw/templates/RF.md, whose only reader was that key, and one assertion in .tfw/scripts/test_gen_index.py that read the live corpus's exit code — what that test owns, that a stateless phase is informational, is still proven on synthetic input where the exit code is asserted against inputs the test controls. Why: the board is a derived view, which conventions.md's own Discovery rules call "never authoritative" and whose absence "changes no task state" — and a never-authoritative view must not stand in a blocking check, because one malformed line in one task then stops every unrelated run. Nothing was relaxed: a check was moved out of a position it should never have held, and the thing it checks was never the authority. gen_index.py is untouched and still reports — run python .tfw/scripts/gen_index.py --check tasks when you want the report; reporting is now all it does. What you do: if your own project_config.yaml carries a build.verify that names this check, remove it. build.* is a PROJECT section an update preserves, so nobody rewrites your key for you and this entry is the only thing that can tell you. If your build.verify names something of your own, keep it — nothing here argues against having a verify command, only against gating on a derived view.

Retired wording, verbatim — for your grep over workflows, templates and rule files

• Append to project-level `TECH_DEBT.md` → REVIEW §5, once, with a disposition.
• **Update TECH_DEBT.md** — append any new items from Tech Debt Collected → the item was already written in §5; there is no second write.
• Read `TECH_DEBT.md` — current entries and | 4 | New tech debt discovered? | `TECH_DEBT.md` | _(append)_ | → docs.md has no debt step.
• Read `TECH_DEBT.md` — extract accumulated items across phases → gone; a previous phase's items were disposed of before that phase closed.
• Ignore TECH_DEBT.md items from previous phases · Ignore accumulated tech debt when planning next phase scope → replaced by the opposite anti-pattern: treating a closed phase's §5 as a queue to inherit.
• **TECH_DEBT.md** — empty or with initial entries if found → init.md creates no such file.
• - [ ] TECH_DEBT.md reviewed — no critical items blocking release → removed from both release checklists.
• Coordinator ignores executor Observations — must triage to TECH_DEBT.md → recorded in REVIEW §5 and disposed of there.
• - `TECH_DEBT.md` — accumulated tech debt from reviews (observations → triage → registry). → the file is no longer a required project-root artifact.
• | 9 | `TECH_DEBT.md` | `reference/tech-debt.md` | Copy + frontmatter | and | `TD-{N}` | `TD-59` | TECH_DEBT.md row | → manifest row deleted; TD-{N} → tasks/DEBT-SNAPSHOT.md.
• Update KNOWLEDGE.md and TECH_DEBT.md after REVIEW → Update KNOWLEDGE.md after REVIEW, in the TFW:CLAUDE block and its template source. Editing the installed copy alone lets the next update restore the retired wording.
• → backlog / → next phase → not material — {the ruling} and the two other outcomes.
• | 3 | Tech debt documented | → | 3 | Debt disposed |.
• - 🔄 **REVISE** — specific issues → back to execution (same task) → routed per item by rung; the route table is conventions.md §5.
• ## Step 5: Debt — capture once, dispose before closing → ## Step 5: Findings — locate, test, route, propose.
• 2. **Quality filter** — reject filler. Only items that would cause real problems if left unfixed. → the entry filter asks only whether the observation is real; materiality is decided once, at disposition.
• **`not material` is a first-class answer** and for most items the right one → it now states which question it answers, not owed or owed and forbidden to pay, with the barring clause cited.
• Every captured item across the project, with its disposition — one search, no maintained file: and the grep/awk block under it → moved verbatim from review.md Step 5 to templates/REVIEW.md §5.
• description: TFW Review — reviewer checks RF against TS, writes REVIEW, disposes of debt → …, writes REVIEW, proposes a disposition per finding. The reviewer no longer disposes.
• | `review.md` | Reviewer | review stage files (map.md, verify.md, judge.md), REVIEW | ONB, RF, HL, TS, code | → the same row now names acceptance authority and forbids disposition rulings.
• (back to dev) → (routed by rung), in the §5 lifecycle diagram.
• - Executor writes REVIEW file — **🔒 Role Lock violation** (start `/tfw-review` instead) and - Reviewer approves without opening any files — Step 2 (Verify) requires spot-checking RF claims against actual artifacts → removed from review.md's Anti-patterns; that block's own header already points to conventions.md §14. The second is a verbatim duplicate of its §14 counterpart. The first is not: §14 reads - Executor writes REVIEW file → **Role Lock violation** and carries no (start `/tfw-review` instead), so that clause is now in neither list. What it told the executor is stated once, in §15's Hard Stop Rule: a finished executor is instructed to start /tfw-review.

Updating from 2.0.0

1. Pin the tag — target_ref=v2.1.0 — and read this payload's .tfw/workflows/update.md from .tfw/.upstream/, not the copy installed in your project; it is what this update replaces.

Which earlier entries you also have to perform — measured, not assumed. Read your own .tfw/VERSION, then:

• on a 2.0.0-dirty tag — perform the 2.0.0 entry's Updating from a 2.0.0-dirty tag first, and Updating from 2.0.0-dirty.2, .3 or .4 as well if it names your tag, then this entry.
• on 1.x — the 2.0.0 entry's Updating from 1.x first, then this entry.
• on 2.0.0 — this entry is the whole of it.

(Corrected before the tag, and the correction is the point. An earlier draft of this section said the dirty tags "were never pushed and are not releases, so there is no intervening section to perform" — true of the tags and false of the projects. Measured across this repository's siblings on 2026-09-03 by reading their .tfw/VERSION files rather than reasoning about the tag list: two on 2.0.0-dirty.4, one on 2.0.0-dirty.5, and not one on 2.0.0 itself. They were installed from a local checkout — installed_from: [email protected] — so an unpushed tag is exactly what a receiver can be sitting on. Fifteen further siblings are on the 0.x line, for which no entry carries an updating section; that is stated here as a measurement and is not answered by a route this release can verify.)

Three things reach a receiving project, and only the first has a procedure. Sealing the debt registry is below. The other two are one edit each and they are named after it: build.verify under B, and /tfw-task under C.

A. Seal the debt registry

The full procedure with its reasons is .tfw/migrations/2.0.0.md step 6. If you are already on 2.x and are not opening that guide, this is the whole of it:

1. Seal. Move TECH_DEBT.md to tasks/DEBT-SNAPSHOT.md — whole, unsorted, unpruned, unrenumbered, and unread. One procedure whatever its size; a full registry is not asked to be cleaned up first. Move it however your project moves files; under version control, keep the file's history.
2. Warn in the header. History, never edited, never appended to. State lines and words — never a row count, which means something different in every project — and the source revision if your project has one, the date if it does not. Say plainly that the rows were sealed unexamined and age is not evidence of importance.
3. Stop. Your own text can stop naming the file whenever you next update it.

Take the payload's canon first if you can. Sealing before your installed review.md and docs.md stop naming the registry leaves a window in which the next review recreates the file you just retired.

Before sealing, read your own section headings — not the rows. If one names safety rules written after an incident, ask the person you are working for. On silence or a no, seal it with the rest and say so. On a yes, create one task with the goal lift the safety rules to where they apply, and touch nothing else, then seal and let it wait. That is the only exception: a release gate, an open defect and everything else are sealed with the rest.

Rollback: move the file back and delete the header. Nothing else in your tree was touched.

B. Remove build.verify if it names the board check

Look at your own project_config.yaml. If build.verify reads python .tfw/scripts/gen_index.py --check tasks — the value this framework shipped until now — delete the key. The reason is under Removed: 00-INDEX.md is a derived view that conventions.md declares never authoritative, and a never-authoritative view standing in a blocking gate means one malformed line in one task stops every unrelated build. That is not hypothetical; it is what closed this release's own suite for an evening.

build.* is a PROJECT section, so an update preserves it and never overwrites it. Nobody will remove that key for you, and this paragraph is the only thing that can tell you it is there.

The tool is not gone and is not deprecated. Run python .tfw/scripts/gen_index.py --check tasks whenever you want the report — it still reads every task's own state and reports every malformed, legacy or unresolved input. What changed is that nothing blocks on it. If you want it in your gate, put it back: a project that chooses to is doing nothing wrong.

Rollback: restore the key.

C. Delete /tfw-task if your adapter folders carry it

One command file in each adapter folder, with no source in .tfw/workflows/. Full reasoning, the file list and the drift-check line that must go with it: Also in 2.1.0 — /tfw-task is retired, at the end of this entry.

---

Nothing to run. No script and no check a receiver must execute. Two edits by hand: the build.verify key under B, and the two /tfw-task copies under C if you have them. Both are deletions.

Verification

• Framework and documentation suite: 322 passed, 1 skipped (315 before this release; +7 guarding the frontmatter fix), from python -m pytest .tfw/scripts/ docs/scripts/ -q. python .tfw/scripts/gen_index.py --check tasks exits 1, on the over-length journal summary recorded under Known open at this tag — it reports, and nothing gates on it. --check project: consistent. Documentation site builds; a TD-N citation opens tasks/DEBT-SNAPSHOT.md in a browser.
• The registry's byte-identity is provable from the move: git mv preserved the file, and the sealed region diffs to zero against git show c153895:TECH_DEBT.md — 132 lines and 12 352 words, re-measured against both the sealed region and the pinned revision.
• Every quantitative claim in this entry was re-measured before the tag, per the row added to RELEASE.md §5 by this release. Six figures: the two sealed-registry counts above, stable by construction; 121 rows, from grep -c '^| TD-' over the sealed content; 477 and 160, with their commands written beside them; 1 699, likewise. Three were wrong and are gone, and one could not be reproduced by any test and was replaced by one that can — each named where it stood, none quietly corrected. The 25 projects survey is dated rather than re-run: it measured other repositories on 2026-09-01 and is a snapshot, not a live count.
• Delivered by TFW_20260830-194027_TLD, single phase, one amendment (A1 — migration triages nothing).
• Read-only dry runs of the receiving instruction against three real sibling registries of opposite shapes. No file in any other project was created, modified, moved or deleted.
• RELEASE.md §5 run in full at the tag, all rows pass. Three tasks closed into this version — TFW_20260830-194027_TLD, TFW_20260902-153617_RTMW, TFW_20260902-112841_RDP — each DONE with a filled outcome; TFW_20260902-222456_RTBO is TODO and explicitly excluded. Twenty adapter copies compared, zero differing. KNOWLEDGE.md updated through /tfw-docs and /tfw-knowledge; 41 fact candidates processed, 14 admitted, 139 → 153 facts. Suite re-run immediately before the tag: 322 passed, 1 skipped.

Known open at this tag

• review.md is 1 699 words — wc -w .tfw/workflows/review.md, run 2026-09-02 — against the ≤1 200-word design rule in conventions.md §11. It was 1 407 before this release and the disposition gate is the load-bearing content of it. Recorded, not hidden.
• The sealed row region is not one Markdown table. The registry carried blank lines inside its rows, so the file renders as six blocks. Reformatting it would have edited the record, which the seal forbids.
• resume.md is obsolete by the owner's ruling of 2026-09-01 and its deletion is a separate task. This release only stops it reading a file that no longer exists.
• One journal event in this repository's own corpus is over the summary ceiling, and it was not repaired. Its summary is 123 code points against tfw.journal.max_summary_length = 120 — measured, and the only event over the ceiling in this corpus — so gen_index.py --check tasks run by hand still exits 1 and names it, alongside anything else the corpus holds when you run it. No denominator is quoted, deliberately: the claim is about one event, not a proportion, and a count of how many events exist is a number someone would have to maintain. The owner ruled that the event is not edited and the ceiling is not moved: conventions.md §4 says a written event is never edited and never deleted, and raising a limit to fit the line that broke it is the instrument this release's own amendment A13 rejected. The gate was removed instead — see build.verify under Removed — so the report is now a report. Recorded, not hidden, and not fixed: the general form, a bound on an immutable artifact whose only reader runs after the write, is carried as a fact candidate rather than as a new check.

Also in 2.1.0 — /tfw-task is retired

Folded in at the tag, 2026-09-02, unedited. This work sat under ## [Unreleased] while 2.1.0 itself was untagged — two unreleased sections stacked, the upper one reading as if the lower had shipped. It is the same release: TFW_20260902-153617_RTMW closed DONE before the tag was cut. The text below is moved verbatim, its headings demoted one level and one heading renamed — its Updating from 2.1.0 said from the version it now sits in, which was true under [Unreleased] and false after the fold. Nothing else changed, because a CHANGELOG entry is never rewritten in substance.

/tfw-task is retired. It shipped as two installed copies with no source in .tfw/workflows/, and the two differed from each other. Every drift instrument the project owns compares a copy against its source, so with no source none of them could see the divergence or repair it: a project on the Codex adapter and one on Claude Code received different lifecycle instructions from the same command, and the gap widened with each release. conventions.md §9 says an adapter installs "whole copies or marker-bounded blocks, and nothing of a third kind." This was the third kind.

Deleted, not given a source. The command was obsolete: plan.md and handoff.md already carry the lifecycle, and a meta-workflow spanning them crosses a mandatory role boundary. Nothing replaces it — the count of installed commands falls from 12 to 11 in both adapter folders, and the two folders now agree with .tfw/ file for file.

Removed

• .claude/commands/tfw-task.md (1 625 bytes) and .agent/workflows/tfw-task.md (1 430 bytes). They were the only two copies in either adapter folder without a source; after the deletion an orphan sweep over both folders reports zero.
• The by-name exception [ "$b" = "tfw-task" ] && continue from config.md's drift check, and the "not copied, and why" row that justified it. Deleting the files without the line would have left a check that silently excludes a file that no longer exists — a dead exception teaching the next reader a rule that was never true.
• /tfw-task from the TFW:CLAUDE command table — in the template as well as the installed CLAUDE.md. The template is what /tfw-init and /tfw-update install from, so repairing only the installed copy would have reinstated the dead command in the next receiving project.
• From .tfw/adapters/codex/README.md: "/tfw-task is intentionally absent. It is not a canonical workflow and duplicates the plan/handoff logic across a mandatory role boundary." The reasoning was sound and is preserved above; the sentence asserted a distinction between adapters that no longer exists once the command exists in neither.

Fixed

• .tfw/adapters/claude-code/README.md promised nine commands in both its file tree and its command table while eleven were installed. /tfw-config, /tfw-init and /tfw-knowledge were missing from both blocks — stale since they shipped, and unrelated to /tfw-task. Both blocks now list eleven, matching what an install actually produces.

Retired wording, verbatim — for your grep over adapter folders and rule files

/tfw-update Step 6 requires zero hits outside an allowlist for every term a release retires. The term is tfw-task. On a receiving project the allowlist is this changelog entry and your own task traces — nothing else. Every hit below is a live use and is removed:

• .claude/commands/tfw-task.md and .agent/workflows/tfw-task.md, whole files → deleted.
• | `/tfw-task` | Meta-workflow | Coordinator | Full lifecycle: plan + handoff with hard stop between them | → gone from the TFW:CLAUDE command table, in the template and in every installed CLAUDE.md.
• | `/tfw-task` | `plan.md` + `handoff.md` | Coordinator | `/tfw-task` | and ├── tfw-task.md # /tfw-task — full lifecycle meta-workflow → gone from the Claude Code adapter README's command table and file tree.
• [ "$b" = "tfw-task" ] && continue → gone from config.md's drift check.
• | `.claude/commands/tfw-task.md`, `.agent/workflows/tfw-task.md` | Adapter-only meta-workflow; it has no `.tfw/` source to copy from | → gone from the not copied, and why table.
• "/tfw-task is intentionally absent. It is not a canonical workflow and duplicates the plan/handoff logic across a mandatory role boundary." → gone from the Codex adapter README.

Updating — the deletion a receiver performs

Nothing breaks when you update, and that is the problem. /tfw-update Step 6 re-syncs .tfw/workflows/*.md into .claude/commands/tfw-*.md and .agent/workflows/tfw-*.md as copies — and a copy whose source is absent is never visited. There is no .tfw/workflows/task.md and never was, so your two tfw-task.md files survive the update byte for byte. Meanwhile the TFW:CLAUDE block replace does remove the command's row from your CLAUDE.md. Update and stop there, and you hold a working /tfw-task that neither the canon nor your own rules file mentions.

One manual step, and it is the whole procedure:

1. Delete whichever of these two your project installed — .claude/commands/tfw-task.md (Claude Code) and .agent/workflows/tfw-task.md (Antigravity). A project running Codex or Cursor alone has neither and skips this release entirely. Remove them however your project removes files; under version control, commit the removal on its own so the trace says what happened.
2. Nothing else to do. No script, no counter, no configuration key. The TFW:CLAUDE block, the drift check and both adapter READMEs arrive already correct in the payload.

Which instrument catches you if you skip step 1. The retired-vocabulary allowlist in Step 6 — it searches the installed adapter layers and your two files are hits. The drift check does not: see Not changed, deliberately below. Run the allowlist search, not the drift check, to confirm.

If your team genuinely used /tfw-task, it was /tfw-plan followed by /tfw-handoff with a hard stop between them. Run those two. Nothing is lost — the stop is a role boundary between coordinator and executor, and a single command spanning it is what made the meta-workflow wrong rather than merely redundant.

Rollback: keep the two files. Nothing in your tree reads them and nothing depends on them, so the command keeps working. You then carry, knowingly, a command absent from your rules file and invisible to your drift check — which is the state this release exists to end.

Verification

In this repository, before and after:

• Drift check over both adapter folders: silent before the change and after it.
• Orphan sweep — any copy whose source is absent: 2 before, 0 after. Both were tfw-task.
• Installed command count: 11 in .claude/commands/, 11 in .agent/workflows/, against 11 sources (10 in .tfw/workflows/ plus the nested research/base.md), and 11 Codex skills.
• grep -rn "tfw-task" over the whole tree with no extension filter: the only remaining hit outside tasks/, workspace/ and the gitignored site/ build is the historical TFW-53/D entry below, which records the state of the world in 2026 and is correct as history. The filter matters — an earlier sweep restricted to *.md missed CLAUDE.md.template entirely.
• --check tasks, --check index, --check project: clean. Test suite: 322 passed, 1 skipped.

On a receiving project, after step 1:

• The retired-vocabulary search finds tfw-task nowhere outside this changelog entry and your own task traces.
• Your installed command count falls by one in each adapter folder you use, and equals the number of sources in .tfw/workflows/ plus one for the nested research/base.md.
• /tfw-task typed at the prompt resolves to nothing. That is the expected result.

Not changed, deliberately

• The [ -f "$s" ] guard in the drift check still lets a copy with no source pass silently. This is why the defect survived as long as it did, and it is why a receiver who skips the deletion above gets a clean drift check while holding an orphan. No orphan exists in this repository today, so nothing here is being masked. Closing the guard is a separate decision with its own blast radius — every project whose adapter folder holds a locally added command would start reporting — and it is not smuggled in under a removal.

[2.0.0] — 2026-08-30

The release of the 2.0.0 line. Five pre-release tags — 2.0.0-dirty through 2.0.0-dirty.5, 2026-08-27 to 2026-08-30 — were cut to exercise the update path on four real projects before this claim was made. They were never pushed and are not releases; their entries below are history. This is the entry a receiving project reads.

From 1.x this is a major migration. Read .tfw/migrations/2.0.0.md after this entry. Nothing in your task corpus is renamed, moved or rewritten by it.

Why this release exists

Every lifecycle transition in TFW 1.x edited the root README's Task Board — the highest-frequency shared write in a project, and the reason two people could not advance two tasks without meeting in one file. 2.0.0 makes a task's live state task-local: one status.md per task and per phase, one immutable journal file per event, declared participants in team/, and a portfolio index that is derived and never authoritative. The migration that gets an existing project there refuses what it cannot read whole and computes every guarantee it prints; the update that delivers it can be followed as written by a receiver on any earlier version, asks the owner before it writes, and ends by telling the owner what changed in their own language.

Delivered by TFW-60 in four phases — A (task state and coordination), AA (portable delivery), AB (honest migration), AC (update without guesswork) — each approved at review and each run on real projects: four consumers and their field reports (stated as six; seven — see the dated correction under Verification below). The task is closed with this release; its two remaining phases (task-local debt, task-local knowledge staging) were dropped by owner ruling and are not carried forward.

⚠️ Breaking

• The root Task Board is retired. Live state lives in {task}/status.md (closed key set, one declared lifecycle value or UNDECLARED carrying the source verbatim); history in {task}/journal/{stamp}__{kind}__{token}.md, one immutable file per event, every event carrying on_behalf_of (a human handle) and via (the tool). A multi-phase task carries a status.md and a journal/ inside each phase directory; the task-level file never summarizes them. The board is captured verbatim into tasks/BOARD-SNAPSHOT.md before it is removed; every row is accounted for by name. The portfolio view is {container}/00-INDEX.md, derived, regenerated deliberately, never a gate.
• Participants are declared. team/{handle}.md, one file per person, exists before the first durable write; with several profiles the acting handle comes from a binding on the participant's own machine (~/.tfw/bindings.yaml · %LOCALAPPDATA%\tfw\bindings.yaml). Identity is asked, never inferred from a Git identity, an OS username or a folder name. There is no actor field and no profile per agent session.
• New tasks take one identifier grammar: PREFIX_YYYYMMDD-HHMMSS_ABBR. PREFIX is tfw.task_prefix; the stamp is read once from the clock; ABBR is the initials of the approved full title — Conflict Resistant Shared Workspace → CRSW — proposed with the title and approved with it; the HL header carries Title and Abbreviation side by side. No counter is read; a collision refuses and asks for another abbreviation. Legacy PREFIX-N identifiers stay readable forever and are never renamed or issued again.
• Configuration. tfw.task_containers is an ordered list — create in the first, resolve across all; tfw.installed_from records {upstream}@{verified-tag}; initial_seq, id_max_retries and review.default_mode are retired and --check project names them if present. build.verify is python .tfw/scripts/gen_index.py --check tasks; --validate is gone.
• Adapters are copies, re-synced by the update. .claude/commands/tfw-*.md and .agent/workflows/tfw-*.md are byte copies of .tfw/workflows/*.md; the Claude Code rules (CLAUDE.md) and the Codex routing (AGENTS.md) are marker-bounded blocks — <!-- TFW:CLAUDE:START/END -->, <!-- TFW:CODEX:START/END --> — replaced between the markers; a file without markers is reported and left untouched until the operator inserts the block once. Rule files read .tfw/VERSION; no {version} substitution exists.
• Scope-budget defaults are 50 / 50 / 5000 / 50 (files, new files, LOC, modified files); they were 30 / 15 / 3000 / 30. The key is project-owned and preserved by an update; only the default a new project starts from changed (owner, 2026-08-30).

Added

• templates/status.md (task and phase paragraphs, a validated worked example), templates/journal/event.md, templates/team/profile.md, templates/bindings.yaml, templates/briefing.md.
• .tfw/scripts/gen_index.py — the derived index and the three checks: --check index (is the view current), --check tasks (is each task's own state legal; names a phase directory without state), --check project (is this project consistent with the release it declares; names retired keys, a machine-local installed_from, what it did not check).
• .tfw/scripts/migrate_board.py — the one-time board migration: reads a committed board by default (--working-tree deliberate and logged), parses identifiers and status cells whole or refuses them, stops before any write on a duplicate, prints every guarantee it checked with its arithmetic under Guarantees checked and names what it did not check, lists every phase directory and says phase state is authored by hand.
• .tfw/migrations/2.0.0.md — the migration guide, routed to by update.md Step 2.
• update.md Step −1 (read the target's update.md), Step 3 🛑 (three questions before the first durable write: who is acting, where new tasks are created, what build.* is), Step 5 copy with declared exclusions (project_config.yaml, knowledge_state.yaml — printed as skipped), Step 8a briefing (four blocks from the intervening entries' Added, Changed, Fixed, Removed).
• RELEASE.md §5: an updating section reaches every earlier tag still in use; a reversed normative statement is quoted verbatim; a replaced instruction keeps its text under Superseded by.
• A HL freeze baseline is a commit with the reserved freeze scope word; an approved amendment re-freezes.

Changed

• update.md is a pinned procedure. The source is pinned from the tag the operator names — source_head=$(git rev-parse --verify "$target_ref^{commit}"), VERSION read from that commit — never from HEAD; the payload is materialized from git archive of that commit; every local file is classified against the installed baseline so upstream drift is not reported as customization; the retired-vocabulary allowlist admits text whose purpose is to retire the term. 1174 words.
• The framework's own test suite (.tfw/scripts/test_*.py) ships with the payload for the framework's maintainers. A receiving project is not asked to run it; the migration guide no longer instructs it and no gate a receiver is told to run depends on it (owner ruling, 2026-08-30).
• plan.md and init.md create a task as the identifier grammar says, name the session after the full identifier, and ask for the abbreviation with the title.
• conventions.md §4 (identity, identifier, control files, discovery, artifact naming under the current grammar), §5 (task and phase state, UNDECLARED — migration never normalizes, an accountable owner resolves with a transition event), §9 (one marker rule for every adapter block), §10.4 (a template carries the name of the artifact it produces; everything else is lower_snake_case).

Removed

• The root Task Board and every workflow step that edited it.
• The event's actor field (returns with TFW-54, not before); per-session agent profiles.
• initial_seq, id_max_retries, review.default_mode; --validate; the --doctor synonym.
• The HEAD-based source pin; cp -r as the payload copy step; the {version} substitution in adapter rule templates; the Codex adapter's append the complete block first-run behaviour.

Retired wording, verbatim — for your grep over CLAUDE.md, AGENTS.md and rule files

• Commands never duplicate workflow content — they reference it → copies are the model.
• If it has no markers, append the complete block. → report and leave; the operator inserts once.
• source_head=$(git -C {source} rev-parse HEAD) → derived from the named tag.
• __{kind}__{actor} → __{kind}__{token}; the carries actor wording → on_behalf_of and via.
• --validate → --check tasks; initial_seq → identifiers come from the clock.
• Normalizing such a value to a declared one is prohibited → an accountable owner may resolve it with a recorded transition event.

Updating from 1.x

1. Pin the tag — target_ref=v2.0.0 — and read this payload's .tfw/workflows/update.md from .tfw/.upstream/, not the copy installed in your project; it is what this update replaces.
2. Answer the three questions before the update's first write: who is acting, where new tasks are created, what build.* is. If team/ is absent, the profile is created from your answer.
3. Follow .tfw/migrations/2.0.0.md for your task corpus: dry run with a manifest, read it, apply, generate the index, remove the board, author phase state by hand where the manifest says so.
4. Merge project_config.yaml key by key: keep every ← PROJECT key, take every ← FRAMEWORK key, delete the three retired keys, set tfw.task_containers deliberately, set tfw.installed_from: {upstream}@v2.0.0.
5. Insert the TFW:CLAUDE and TFW:CODEX blocks once where those adapters are installed; re-copy the command and workflow copies; the update reports a file without markers and leaves it.
6. python .tfw/scripts/gen_index.py --check tasks and --check project. Then read the briefing the update delivers last.

Clarified 2026-09-03, from a field report on the .5 → 2.1.0 path — the text above is not rewritten, this note is the correction. Step 3's tfw.installed_from: {upstream}@v2.0.0 is written for an update that stops at 2.0.0. If you are transiting to a later tag, set it to the tag you actually land on. Executed literally in section order it writes a value the next entry's Step 0 immediately overwrites — or, worse, leaves a false one standing if you judge that step already done. The receiver who found this was following the section exactly as written, which is the only reason it was found.

Updating from a 2.0.0-dirty tag

• From .5: payload text only — the migration guide no longer instructs the framework tests and the adapters README no longer instructs a {version} substitution. Set installed_from to {upstream}@v2.0.0. Nothing else changes; nothing is renamed.
• From .2, .3 or .4: perform the .5 entry's updating section first, then the line above.
• Your installed_from names a pre-release tag the public upstream does not carry; Step 3 states the fallback baseline and its uncertainty once. That is expected and ends with this update.

Verification

• TFW-60 closed DONE 2026-08-30: phases A, AA, AB, AC each APPROVE at review; amendments A1–A8.
• Framework and documentation suite at 2.0.0-dirty.5: 315 passed, 1 skipped; --check tasks 54 validate; --check project consistent. This release changes documentation and version files only.
• Field: KZ-IT-telegram-list 1.3.0 → dirty; innoforce-ai-first 1.3.0 → .2 → .4; helpdesk 0.8.7 → .3 → .4 → .5; kaznpu-ai-lab 1.0.0 → .4. Six reports filed under TFW-60.

Corrected 2026-09-03: seven, not six. find tasks -iname '*FIELD-REPORT*' | wc -l returns 7, all committed. The count was taken before a report overwritten during a .3 → .4 run was restored, so it was true of what was visible and false of what exists. Reported by the receiver whose report it was. And it is the same class this line's own release line later spent a whole round removing — a count written beside a fact, which a neighbour moves. The fact needs no number: reports were filed by every consumer named above.

Known open at this tag

• The legacy corpus is frozen. Every task under tasks/ (PREFIX-N) is read forever and never edited again, whatever lifecycle its state file or the snapshot shows; new work opens new tasks under the current grammar and cites the old ones. TD-207 (TFW-55 without state) is moot by that ruling.
• TD-211: the payload carries this repository's own project_config.yaml and knowledge_state.yaml; the copy step excludes them, the payload boundary is a later ruling.
• TD-206: update.md Step 0 admits a commit target in prose; its block checks tag equality only.
• One test in the shipped suite (test_the_repository_stateless_phases_are_all_informational) is written against this repository's task corpus and is red elsewhere. Receivers are not asked to run the suite; it is filed for the maintainers.
• editions/02-assisted at this tag is mid-review (task TFW_20260830-114238_ASSISTED15) and is not released as Assisted 1.5 by this entry; the .tfw/ payload does not include editions/.

[2.0.0-dirty.5] — 2026-08-30

Pre-release, tagged locally and not pushed. Cut after TFW-60 Phase AC was approved at review, so a consumer already on the 2.0.0-dirty line can update to it and the owner can check the update path as a receiver. By semver 2.0.0-dirty.5 < 2.0.0, so the 2.0.0 claim stays unmade.

A project updating from .2, .3 or .4 renames nothing: no task directory, no journal event, no already-migrated status.md. Only newly created tasks are affected by the abbreviation rule.

Why this release exists

Two more real projects ran v2.0.0-dirty.4 — one updating within the line, one migrating onto it for the first time with the owner absent. The procedure held where it had been tested and guessed where it had not: the pin step could not pass on a source whose development had moved past its release; an unattended update inferred the owner's handle from a Git identity, chose the task containers and build.* alone, and reported the change in the procedure's words; cp -r of the payload overwrote the receiver's own project_config.yaml; the migration read a status cell ✅ DONE (A/V/B/C) · 🔄 Phase D … by its first token, closed a live task and wrote nothing while --check tasks answered 4 tasks validate over four phase directories without state. This release removes the guessing: where the procedure cannot know, it asks; where a tool cannot read the whole, it refuses and says so; where an instruction could not be executed as written, it is rewritten.

⚠️ Changed

• The source pin is derived from the tag the operator names, never from the source's HEAD: source_head=$(git rev-parse --verify "$target_ref^{commit}"), VERSION read from that commit and compared with the tag's name. A live source — one whose HEAD has moved on — is a valid source. The Step 5 recheck compares against the same derived commit. Retired wording, for your grep: test "$tag_commit" = "$source_head" with source_head=$(git -C {source} rev-parse HEAD).
• update.md opens with Step −1: read the target's update.md from the pinned payload and follow it, not the installed copy. The installed workflow is what the update replaces.
• A task's abbreviation is the initials of its approved full title — Conflict Resistant Shared Workspace → CRSW — proposed together with the title in one exchange and approved with it; the HL header carries Title and Abbreviation side by side. Never derived silently means never invented apart from a title and never created without approval. The grammar PREFIX_YYYYMMDD-HHMMSS_ABBR is unchanged.
• Every Step 6 row is a whole copy verified by cmp or a marker-bounded block verified on the region between its markers, and the table says which. The Claude Code rules gain <!-- TFW:CLAUDE:START --> / <!-- TFW:CLAUDE:END --> around the managed ## TFW section, on the Codex pattern; the block's content was brought current (status.md in the context-loading order, /tfw-knowledge and /tfw-config in the command table, Version: see .tfw/VERSION).
• One marker rule, stated once in conventions.md §9, for every block in every adapter: markers present → replace the text between them; file absent → create it from the template; file present without markers → report it and leave it untouched — the operator inserts the block once. Retired wording, for your grep: the Codex README's If it has no markers, append the complete block. Appending to a file that already carries an unmarked hand-written TFW section produced two sections that disagreed.
• tfw.installed_from has one form: {upstream}@{verified-tag}, where {upstream} is tfw.upstream as configured — a URL or a symbolic name — and never a machine-local path. Retired wording: {resolved-source}@{verified-tag-or-commit}. --check project reports a drive letter, a leading / or a backslash there and rewrites nothing.
• The retired-vocabulary allowlist admits text whose purpose is to retire the term — a deletion instruction, a migration step, a changelog line and their byte-identical copies. The gate is literally green on a correct project; a live use of a retired term is never admitted.
• update.md is 1174 words with Steps −1, 3, 5, 8a added; the 1200-word ceiling holds again.

Added

• 🛑 Step 3 asks the owner before the first durable project write — exactly three questions: who is acting (the handle — asked, never inferred from git config user.name, an OS username or the upstream's profiles), where new tasks are created (tfw.task_containers), how the project verifies (build.*). In AG mode the three go out as one message; the run proceeds through the read-only steps and stops at the write. The answers are recorded in the update checklist.
• Step 8a — the briefing, the update's last message, from .tfw/templates/briefing.md in content_language: four blocks built from the intervening entries' Added, Changed, Fixed and Removed sections — what is now possible, what you now do differently, what stopped breaking, what no longer has to be done. An absent section reads nothing in this release; no free text.
• Step 5 copies with declared exclusions and prints what it skipped: .tfw/project_config.yaml is merged key by key and .tfw/knowledge_state.yaml is never touched; a copy that reports nothing skipped on a project that has both files is a failed step. A payload test derives the project-owned set from the payload itself, so a new project-owned file fails until it is excluded.
• The migration manifest gains two sections: Rows carrying more than one lifecycle signal — each refused row with the signals seen — and Phase directories — every phase-* of every matched task, present or absent, with the sentence phase state is not written by migration; author {phase}/status.md by hand.
• --check tasks names a phase directory that carries no status.md: a failure when the task's own state is live; one informational line per task when the task is terminal, has no state of its own, or has malformed state (already the failure). Exit code unaffected by informational lines. On this repository: six lines over seventeen directories under tasks closed before phase state existed.
• templates/status.md carries the phase paragraph — "…this phase's live state. The task-level status.md never summarizes it." — and says when to use which.
• RELEASE.md §5 gains three rules: a release's updating section reaches every earlier tag still in use and opens with read the target's update.md; a reversed normative statement is quoted verbatim as a search string with what a project that acted on it does; a replaced instruction keeps its text under a > **Superseded by** line. Entries are appended to, dated, never rewritten.

Fixed

• The status cell is parsed whole or refused. A cell is one declared lifecycle token followed by free text carrying no second declared token and no second status symbol (Unicode category So, where every emoji marker lives — ✅ ❌ 🔄 🟢 🟡 ⬜); anything else is UNDECLARED with lifecycle_verbatim, never terminal by its first token, and receives a status.md. Measured on four real boards (114 rows) before the rule was fixed: eight rows change, every one carrying a second emoji or a second declared token; +, →, = are prose and refuse nothing.
• The payload copy can no longer overwrite project_config.yaml or knowledge_state.yaml.
• A receiver that skipped a tag finds its path: this entry's updating section names the .3 and .4 sections it must also perform. The .3 entry gained a dated addendum quoting the sentence it retired — "Commands never duplicate workflow content — they reference it" — and what a project that went thin does (TD-198); the .2 entry's TD-11 paragraph is closed with a dated line (TD-191); the 2.0.0-dirty migration fence carries > **Superseded by** .tfw/migrations/2.0.0.md (TD-190).
• The Antigravity and Cursor rule templates no longer require a {version} substitution; they read .tfw/VERSION as the rendered rule does, and the rendered rule and its template are byte- identical (TD-204). A test refuses {version} in any adapter template.
• The framework is its own first consumer: this repository's root CLAUDE.md carries the TFW:CLAUDE block byte-identical to the template's, and the installed-copies test checks it.
• Payload carriers agree with the canon: templates/journal/event.md describes via as free-form non-empty provider or tool text (TD-200); templates/team/profile.md says one file per person, that a participant's role and context go in team/README.md, and what since means (TD-203); conventions.md §4 Artifact file naming states the current grammar with examples — HL-TFW_20260829-172110_ABT.md, no title appended (TD-201).
• The migration guide opens step 1 with the target's update.md, names one manifest location (tasks/MIGRATION-2.0.0.md, beside the snapshot), says when --working-tree is the right choice, gives every command from the project root (the cd .tfw && … command is gone), lists the two manifest sections to read, and adds step 3a — phase state by hand.

Removed

• The HEAD-based pin in update.md Step 0 and its Step 5 recheck against HEAD.
• cp -r .tfw/.upstream/.tfw/. .tfw/ as the copy step; the copy is a loop with exclusions.
• The {version} substitution step from rendering the Antigravity and Cursor rule templates.
• The Codex adapter's append the complete block first-run behaviour; every block row now reports and leaves a file without markers.

Updating from 2.0.0-dirty.2, .3 or .4

1. Pin from the tag you name — target_ref=v2.0.0-dirty.5 — and read this payload's .tfw/workflows/update.md from .tfw/.upstream/, not the copy installed in your project (Step −1). Your installed workflow pins from HEAD and has no Step 3 gate; it is what this update replaces.
2. If you are on .2: also perform the .3 entry's instructions — re-sync .claude/commands/tfw-*.md, add tfw.installed_from, delete any per-session agent profiles — and read the .3 entry's dated addendum: grep your own CLAUDE.md, AGENTS.md and rule files for "Commands never duplicate workflow content" and remove it.
3. If you are on .2 or .3: also perform the .4 section — remove tfw.id_max_retries and review.default_mode if present; rename no task directory or journal file.
4. Answer the three questions the update asks before its first write: who is acting, where new tasks are created, what build.* is. Record the answers in the update checklist.
5. tfw.installed_from takes the form {upstream}@{verified-tag} — for a local checkout, a symbolic name such as [email protected], with the path in the checklist. --check project reports the D:/… form every current consumer carries; record the upstream reference instead. Nothing is rewritten for you.
6. CLAUDE.md: insert the <!-- TFW:CLAUDE:START --> … <!-- TFW:CLAUDE:END --> block once, from .tfw/adapters/claude-code/CLAUDE.md.template; until you do, Step 6 reports the file and leaves it. The same rule now governs the TFW:CODEX block in AGENTS.md: a file without markers is reported, never appended to.
7. Run python .tfw/scripts/gen_index.py --check tasks. A phase directory without status.md under a live task is now a failure — author the file from .tfw/templates/status.md with its phase paragraph. Under a terminal task it is an informational line.
8. A new task's abbreviation is the initials of its approved title, proposed and approved together with the title; existing identifiers are untouched.
9. Read the briefing the update delivers last — four blocks, in your project's language.

Verification

• TFW-60 Phase AC: APPROVE at review, 2026-08-30; verify ratio 1.0; two Low findings filed (TD-206, TD-213), none returned.
• Full framework and documentation suite: 315 passed, 1 skipped (283 + 1 at .4; +32 tests, none removed).
• Task and project consistency checks: clean — 54 tasks validate, six informational lines over seventeen stateless phase directories, exit 0; project consistent.
• Workflow adapter copies: byte-identical; Antigravity rule = template; root CLAUDE.md block = template block.
• Evidence: 34 of 37 rows verified; the three deferred are this entry, this tag, and the consumer run this tag exists for.

Known open at this tag

• The consumer run. Phase AC's last acceptance criterion is a consumer already on the line updating to this tag from Step −1, with the derived pin, the three questions and the briefing on record. That is what this tag is for; the phase stays open at KNW until it is on record and the knowledge pass has run.
• TD-206: Step 0 admits a commit target in prose; its block checks tag equality only.
• TD-211: the payload still carries this repository's own project_config.yaml and knowledge_state.yaml, which conventions §10.3 says are never sourced from upstream; the copy excludes them, the payload boundary is a later ruling.
• TD-207: in this repository TFW-55 is live in the working tree with no state file at any level and reads as history to the gate.

These are filed debts, not hidden acceptance claims. This tag is deliberately a test pre-release and is not pushed by the release operation.

[2.0.0-dirty.4] — 2026-08-30

Pre-release, tagged locally and not pushed. Cut for the next real-project update test after TFW-60 Phase AB was approved at review revision 2. By semver 2.0.0-dirty.4 < 2.0.0, so the 2.0.0 claim stays unmade.

A project updating from 2.0.0-dirty.3 does not rename an existing task or journal event. Historical identifiers remain readable; only newly created tasks use the grammar below.

Why this release exists

The third external update completed and found a failure more serious than an awkward instruction: the migration path accepted HD-30b as HD-30, wrote state onto the shipped task and printed a success guarantee over a manifest that listed one identifier twice. The parser had found a plausible prefix instead of proving the whole identifier, and the manifest printed a literal conclusion instead of deriving it from the partition it showed.

This dirty tag exists to exercise the corrected path in another real project before 2.0.0 is claimed.

⚠️ Changed — one grammar is issued for new tasks

New task directories use PREFIX_YYYYMMDD-HHMMSS_ABBR:

• PREFIX comes from tfw.task_prefix;
• the clock stamp is read once;
• ABBR is uppercase alphanumeric, approved by the owner during planning and recorded in the HL; and
• a collision refuses creation and asks for another abbreviation. It never appends a suffix or silently recomputes the stamp.

The YYYYMMDD-HHMMSS__slug grammar issued earlier in the 2.0.0-dirty line and legacy PREFIX-N identifiers are read-only compatibility forms. They are not renamed, rewritten or re-issued.

Fixed

• Identifier parsing is whole-or-refuse. One dispatcher recognizes the three named grammars. Anything else is malformed; it is reported and never receives inferred state.
• Ambiguity stops before writes. Duplicate board rows and two directories resolving to one identifier name both sources and stop before an output path is opened.
• Every printed migration guarantee is computed. The manifest shows the arithmetic for each checked partition and separately names guarantees the current run did not check. The literal Unaccounted: 0 conclusion and its tests are gone.
• Migrated prose preserves identifier bytes. Markdown cleanup inserts word boundaries without splitting or deleting underscores inside identifiers.
• Framework tests and repository-state tests are separable. Receiving projects can run the payload suite without inheriting checks against this repository's own task corpus.
• /tfw-update pins and verifies its source. The source revision and tag are checked before its VERSION is trusted, then rechecked before installation; provenance drift and local customization are reported separately. The workflow is 840 words and its two adapter copies are byte-synchronized.
• Reference validation covers the current identifier grammar. Documentation links resolve all three identifier forms instead of treating only the legacy and dirty-clock names as task paths.
• via is free-form non-empty provider/tool text. Validation now enforces the rule that the conventions state; there is no provider registry.
• Phase AA's corrective pass has its RF. The release no longer carries the execution-to-RF trace gap named at 2.0.0-dirty.3.

Updating from 2.0.0-dirty.3

1. Follow .tfw/workflows/update.md and pin this exact tag before trusting its version.
2. Re-sync only installed adapter paths as Step 6 directs.
3. Remove retired live configuration keys if present: tfw.id_max_retries and review.default_mode. Historical text may still name them.
4. Run the documented repository and payload checks. Do not rename existing task directories or journal files.

Verification

• TFW-60 Phase AB: APPROVE, review revision 2.
• Full framework and documentation suite: 283 passed, 1 skipped.
• Task and project consistency checks: clean.
• Workflow adapter copies: byte-identical.

Known open at this tag

• TD-200: the event template still describes via as a provider family although the rule and validator now accept free-form provider/tool text.
• TD-201: artifact naming conventions do not yet state the current task grammar explicitly.
• TD-203: the team-profile template still contradicts itself about agent profiles.
• TD-204: the Antigravity adapter source still asks an updater to substitute a version that the rendered rule reads directly from .tfw/VERSION.

These are filed payload debts, not hidden acceptance claims. This tag is deliberately a test pre-release and is not pushed by the release operation.

[2.0.0-dirty.3] — 2026-08-28

Pre-release, tagged locally and not pushed. Cut so a third real project can run the update path before 2.0.0 is claimed. By semver 2.0.0-dirty.3 < 2.0.0, so the claim stays unmade.

What a project updating from 2.0.0-dirty or .2 has to do about the change below: nothing. Not one journal event needs editing, in any project. That is the point of it.

Why this release exists

A second external project ran the 2.0.0-dirty.2 update and reported back. The release worked — the first consumer spent a session reconstructing the order of operations, this one spent nothing — and the run surfaced one design error and four instructions naming things the reader does not have.

⚠️ Changed — the event's third field is removed until TFW-54

actor carried two unrelated jobs at once: say who wrote this, and make the filename unique. They contradict each other. A distinct writer needs a distinct value; a declared handle needs a profile in team/. Two external projects resolved that the only way that lets work proceed — a profile per agent session — and one later deleted those profiles and left its validation gate red permanently, because events are immutable and profiles are not. The operators followed the design; the design contradicted itself.

Two identity fields remain, and the filename's third component becomes a short opaque token whose only job is that two writes in one second differ:

20260828-210350__transition__8cfe.md      on_behalf_of: who is accountable, always a human
                                          via:          what produced it

• An actor already written is tolerated, never required, and never rewritten. A reader treats it as a pre-2.0.0-dirty.3 record: no error, no comparison against team/, no dangling handle. Both name shapes match the same pattern, so nothing has to tell them apart.
• Do not create a profile per agent session. team/ holds people. Existing per-session profiles can be deleted whenever you like — nothing compares them to anything any more.
• Measured on the two consumers: one was red on two events naming deleted profiles and now validates, with nothing in that project changed.
• A writer becomes nameable at TFW-54, which is the task that will have a principal to name.

Naming the token's one job removed machinery. Uniqueness no longer comes from the second, so the clock is read once and a collision is re-drawn rather than waited out — the retry-and-sleep path is gone. The prohibition it enforced is now its own test: nothing adds to, rounds or composes a stamp.

Added

• .tfw/templates/bindings.yaml — seven workflows tell a session to read ~/.tfw/bindings.yaml when a project declares more than one participant, and until now the payload shipped nothing saying what it contains. Defined in conventions.md §4, where the glossary already pointed.
• tfw.installed_from — <source>@<tag>, written by update.md Step 7. tfw.upstream is where updates are fetched from; this is what the project actually runs. A local unpushed tag is unreachable from a remote URL, so without it the next update clones the remote, finds an older payload and reports that all is well.
• A phase carries its own journal/, read exactly like its own status.md. If your project already created phase journals by assuming this — one did — their events are now validated, and problems that were invisible will appear. That is the gate working, not a regression.

Fixed

• update.md Step 6 had no row for .claude/commands/. Both adapters are byte copies of the same workflows; only one was listed, so only one was maintained. In a project with no .agent/ directory the uncovered adapter rotted — six files still instructed agents to update a board removed the day before. Re-sync .claude/commands/tfw-*.md as part of this update, and check the layer afterwards for every term this release retires.
• conventions.md §10.4 stated a naming rule that nine of its own twenty subjects contradicted, illustrated by a template a move had deleted. The rule is corrected to the one that holds: a template carries the name of the artifact it produces; everything else in .tfw/ is lower_snake_case. Read back against all twenty templates, it contradicts none.
• Nothing checked the bare templates/… reference form, and nothing checked a bare filename at all. That was the gap that let two dead paths survive four releases behind two checks that could not see them. Every path any payload file names is now checked in all three forms, with its exemptions annotated.
• The adapter layer is checked against the retired-term register, rather than by a command someone remembers to run.
• .tfw/adapters/claude-code/README.md claimed commands never duplicate workflow content, beside twelve byte-identical copies. Copies are the model: full copies, where each tool expects them, re-synced by Step 6, in the same commit as their source.

Known open at this tag

• AC-13's acceptance half is unmet. No external project has yet completed an update to 2.0.0-dirty.3 by an operator who is not the author of this code. That is what this tag is for.
• update.md is 1380 words against a design ceiling of 1200. Twice reduced by deleting duplication; the duplication is gone and what remains is instruction. A decision on whether to accept the overrun or split the file is open.
• The corrective pass's own RF and review are not written at this tag.

Added 2026-08-30 — the wording this release retired, verbatim (TD-198)

Until this tag .tfw/adapters/claude-code/README.md stated: "Commands never duplicate workflow content — they reference it." This release retracts that sentence — copies are the model — and the entry above did not quote it, so a project that had acted on it could not find what to undo. One consumer had rewritten twelve .claude/commands/tfw-*.md into thin adapters on its strength and recorded the principle in its own CLAUDE.md; Step 6 re-copied the commands silently, and the project's rule file kept asserting the retired principle until a person reread it. The retired-vocabulary gates read terms, not principles: the quoted string is the only thing grep finds.

A project that went thin: re-copy .claude/commands/tfw-*.md from .tfw/workflows/ (Step 6 does it), then search your own CLAUDE.md, AGENTS.md and rule files for the sentence above and remove it. The adapter README now carries the correction in a blockquote so the retraction stays visible where the rule was.

[2.0.0-dirty.2] — 2026-08-27

Pre-release, tagged locally and not pushed, so the update path can be exercised against a second real project before 2.0.0 is claimed. By semver 2.0.0-dirty.2 < 2.0.0, so the 2.0.0 claim stays unmade. Not 2.0.1-dirty: 2.0.0 was never pushed, and there is nothing to patch in a release that never shipped.

Tag v2.0.0-dirty.2, cut 2026-08-28 after TFW-60 Phase AA was approved at review revision 2. Three debt items are open inside this payload and named on the tag: TD-192 (conventions.md §10.4 illustrates the template naming rule with a file this release deleted — deliberately not fixed by swapping the filename, because nine of twenty templates already contradict the rule), TD-193 (nothing checks the bare templates/… reference form, which is the mechanism gap that let TD-192 and TD-194 survive four releases), and TD-186 (KNOWLEDGE.md still names the old tooling path — owned by /tfw-docs, not by the payload).

Why this release exists

2.0.0-dirty was cut so the update path could be exercised against a real project before 2.0.0 was claimed. It was, once — and the exercise reported its own ratio: "the file copying took minutes; the rest of the session was reconstructing what to do and in what order."

The update completed. It completed because a person hand-carried two scripts and a directory the payload does not contain, and reconstructed an order this file states once, in a code fence, 150 lines down.

The framework could not deliver its own tooling. gen_index.py and migrate_board.py lived outside .tfw/, so a project reading "run python docs/scripts/migrate_board.py" did not have that file and had no instruction telling it how to get one. Worse, the tools resolved the project root by counting directories upward, which made their location load-bearing: a project that put them anywhere else had to edit .tfw/ and forfeit clean updates. The rules named docs/scripts/ as a literal in three normative files.

Every finding below came from one real external project. None of them was found in four review rounds here, because every round ran here — where the tooling already existed.

Added

• .tfw/scripts/ — the tooling ships inside the payload. /tfw-update copies .tfw/, so anything the rules require has to be in it.
• .tfw/migrations/2.0.0.md — the migration procedure, written for a project that is not this one. update.md Step 3 routes here when an update crosses a major version. A major release without a migration guide is now incomplete by rule.
• --check project — one command for is this project consistent with the release it declares: the payload, team/, the container configuration, retired keys, the version marker, carrier validity. It reports and exits; it repairs nothing and is authority over nothing, and its output names what it did not check. The best signal a consumer previously had was two framework tests it was never told to run.
• --board / --board-heading — migration finds a board wherever a project keeps it. The first consumer's board was at tasks/README.md under ## Board, because its root README is fully regenerated and a board there is destroyed.
• update.md Step 3a — diff every local .tfw/ file against the pristine previous tag before merging anything. This single check turned three declared manual merges into zero, conventions.md with its 212 changed lines included. The text says whose tag: the source's, not the receiving project's, which had no TFW tags at all.
• update.md now carries tfw.task_containers as a decision with its two real options, names initial_seq as a key to delete, accepts a local working tree as an upstream source, and creates team/ together with its first profile before the first durable write.

Changed

• Root resolution is a marker search, not depth arithmetic. The tools walk upward for .tfw/, so a project may place them anywhere; every run prints the root it resolved. The staging clone at .tfw/.upstream/ is skipped by name so it cannot capture the search.
• --check takes a subject: index, tasks, project. --validate is gone. It and --check differed by a five-line comment in project_config.yaml explaining which one the build gate wanted — and when prose is needed to tell your own names apart, the names are wrong. The comment is deleted rather than rewritten, and each subject's output states what it does not answer. build.verify becomes --check tasks.
• The board is read from a committed revision by default; the working tree is --working-tree, deliberate and logged. During the first real migration the source board was rewritten three times while being read. With no committed board the run refuses and names the opt-in — a printed notice is the thing nobody reads.
• A directory the identifier grammar does not match is reported as unresolved, never as backlog. The failing run classified TFW-01_awesome_list_restructure and TFW-02_enhanced_validation — both holding completed HL, TS and RF traces — under a heading reading "ideas, not work in progress", and printed "backlog idea, never started" in the manifest. The grammar is not widened to admit the single-underscore form: that would be an identifier-rule change. A person may rename the directory, which leaves a trace.
• --check tasks names the offending key. It reported unparseable front matter: ScannerError with no key named, and a person hand-writing five state files had to find the cause by inspection. The cause was mechanical every time: a colon followed by a space ends a YAML plain scalar.
• .tfw/templates/status.md quotes its prose values, says why, and carries a complete worked example that is checked against the real validator.
• Three templates move into directories mirroring their output — templates/team/profile.md, templates/journal/event.md, templates/knowledge/topic.md. templates/research/, evidence/ and review/ had settled that rule long ago; an underscore standing in for a directory separator is what overlooking it looks like on disk.
• plan.md names the session after the identifier exists (step 3 of Step 4, repeated on a slug change), adding the phase when the agent was given one. Step 0 demanded an identifier that does not exist until Step 4, so the instruction was unsatisfiable and what happened instead was a name carrying a role and a guess. The question-first order is kept deliberately: understanding the task before creating a folder is the right sequence.
• Runtime messages are ASCII, enforced by a test. An em dash in a refusal renders as a replacement character on a console whose codepage nobody chose.

Fixed

• A route unfixed across two releases. /tfw-research pointed at .tfw/workflows/research.md, which has never existed — the workflow became a directory. Three adapter sources carried it, not one. A test now fails if any path in any adapter source or installed copy does not resolve, with a short annotated allowlist for the two paths that intentionally live outside the tree.

The field report called this defect TD-11. No such row is in TECH_DEBT.md, and the historical TD-11 was an unrelated defect purged in an earlier sweep — so the label is dropped here rather than left citing nothing. In a release whose subject is that every instruction must name something the reader actually has, a dead identifier is the same defect one layer in. (2026-08-30, TD-191: the route defect was fixed in this release and has no debt row; the label above is history and nothing in the repository cites it.) - build.* is preserved by an update, and preserved is not the same as correct. A project that updates across a release which moved a tool keeps a command naming a path that is gone, silently and forever. --check project reports it, update.md says so, and the shipped config template now sets verify to a real command instead of a placeholder.

Canon

• UNDECLARED: migration never normalizes, an accountable owner may resolve — by setting the correct value and recording a transition event carrying from: UNDECLARED. The prohibition read as absolute, which left projects with two bad options: strand the task, or fix it with no trace. Stated in conventions.md §5, glossary.md and templates/status.md — the third copy was missed on the first pass and caught in review, which is why a test now checks a registry of retired wordings against every payload file that instructs. A rule corrected in two of its three shipped copies is not corrected: the two a reviewer reads were right, and the one a receiving project hand-authors from still said the opposite.
• Some artifacts legitimately have no journal event. The kind vocabulary is closed and stays closed; an artifact no kind covers is filed without an event rather than given an invented one. The worked example is an inbound advisory record from another project: it escalates nothing and requests no verdict, and amendment_escalated would misreport it as awaiting a ruling.

Migration from 2.0.0-dirty

Two commands change. Everything else is a payload copy.

# in .tfw/project_config.yaml
build.verify: python .tfw/scripts/gen_index.py --check tasks   # was --validate
build.lint / build.test: point at .tfw/scripts/ as well as your own test directory

If you migrated your board at 2.0.0-dirty, each generated status.md ends with a comment naming docs/scripts/migrate_board.py. Leave it. It is a true statement about a past act at the path that was correct then.

The 2.0.0-dirty entry below still names docs/scripts/ in its migration commands. That was where the tooling was at that tag, and the entry is a record rather than a live instruction. Follow .tfw/migrations/2.0.0.md, not the code fence below.

[2.0.0-dirty] — 2026-08-27

Pre-release. Tagged locally and not pushed. Cut so the update path can be exercised against real projects before 2.0.0 is claimed. Two things are deliberately open at this tag: TD-182 — the shipped Assisted edition still changes status by moving a task folder, which contradicts what this release declares, and is deferred to its own task; and TFW-60 is at PHASES, not DONE, because Phase B and Phase C remain. Releasing on a phase boundary is what the frozen contract intends — master HL §4 calls each phase "a vertical, independently releasable slice" and DoD 13 requires a phase to be releasable "without waiting for a later phase". The -dirty suffix says the framework is usable and the claim is not yet final.

Two people can now advance two tasks without meeting in the same file. TFW-60 Phase A. Until this release every lifecycle transition — create, plan, research, hand off, review, close — rewrote one table in the root README.md. Task separation did not produce file separation: three participants working on three unrelated tasks still queued behind one Markdown table, and its schema had already drifted (TD-177) while the documentation build regex-read its columns as an implicit API (TD-81).

⚠️ Breaking

The root Task Board is removed. It was a required artifact; it no longer exists. Anything that parsed it will find nothing to parse.

Live state moved into each task. {task}/status.md is now the only authority for a task's lifecycle, owner, goal, value and terminal outcome. A transition is one write, inside one task directory.

The status flow changed shape. The lifecycle ids are unchanged, but UNDECLARED is added for a value a migration source carried that the vocabulary does not contain. It is never selected by a person, and normalizing it away is prohibited.

Added

• {task}/status.md — the task state carrier. Closed key set, bounded fields, no free-text body, and every field has a named reader. Template: .tfw/templates/status.md.
• {task}/journal/ — one immutable file per coordination event, named <YYYYMMDD-HHMMSS>__<kind>__<actor>.md. The filename is the event identifier, so nothing allocates one and nothing counts. The actor is part of the name because it is the only field that separates two concurrent writers — on_behalf_of names the same accountable person for both, and via names the same provider for two sessions of one tool. Two participants recording the same kind of event in the same second therefore produce two files rather than one; one actor writing twice in a second takes the next actual second. The time is read from the system clock and never typed. A written event is never edited; a correction is a new event. Entries carry references rather than copied artifact prose, under a 120 code point summary ceiling — measured against 272 commit summaries and 63 review verdicts in this repository, where p95 is 83 and p99 is 110. Template: .tfw/templates/journal_event.md.
• Three identity fields on every event — actor (who performed it), on_behalf_of (who is accountable, always a human handle) and via (which tool produced it). An event without on_behalf_of is refused: there is no such thing as a record nobody answers for. A provider name is never an actor.
• team/{handle}.md — one profile per participant. Declared attribution, never authentication. The machine-to-handle binding lives outside the project tree, because a per-user file that is gitignored is still not sync-ignored. Template: .tfw/templates/team_profile.md. No agent profile ships in 2.0.0. The schema admits type: agent and the slot is deliberately empty: a provider family is not an actor, and what would make an agent profile meaningful — a named principal that delegates and answers to someone — is a separate task. Until then there is one accountable participant, and which tool produced a record survives in the event's via field.
• 🧩 PHASES, one new lifecycle id — a multi-phase task sits here while its phases run. Each phase directory carries its own status.md on the same closed schema, written by that phase's owner, so two phases under two owners are two files and never contend. The task file never summarizes phase state: a rollup is a second fact that has to agree with the phases, which is the synchronization problem the carrier exists to avoid. The index renders phase rows beneath their task row — what the retired board's per-phase columns showed.
• Time is recorded to the second. created and updated use YYYYMMDD-HHMMSS, the same grammar as the identifier, and are read from the system clock rather than composed. At day resolution the two fields are routinely identical on a corpus taking several transitions a day, and updated stops answering the question it exists for. A legacy source that carried only a date migrates to that date with a declared zero time — 20260819-000000 means "this day, time unknown" and is never second-accurate history.
• A validation gate that reads task-local truth — python docs/scripts/gen_index.py --validate checks every task's own state and journal against the closed schema. It is deliberately not a check that the shared index is current: requiring that would make every task-local transition fail until somebody rewrote the aggregate, which is the bottleneck this release removes.
• {container}/00-INDEX.md — a derived portfolio view, rebuilt by python docs/scripts/gen_index.py. It declares that it is derived, names its source count and freshness, and reports every unresolved input instead of dropping it. It is never authoritative: a workflow acting on a task re-reads that task's status.md first, and an absent or stale index degrades discovery without changing any task.
• tfw.task_containers — an ordered list. A task is created in the first entry and resolved by searching every entry in order.
• Clock-derived identifiers — YYYYMMDD-HHMMSS__slug, and the whole directory name is the identifier. The timestamp alone is not one: two participants offline from each other can reach the same second, and only the slug tells them apart. Same second and same slug means they created the same task — a signal, not a collision. Creating a task reads no counter and no other task directory; if the directory already exists, the writer takes a new actual timestamp under a bounded retry, never a reuse, and a clock that will not advance fails visibly instead of spinning.
• docs/scripts/gen_index.py and docs/scripts/migrate_board.py, with tests.

Changed

• The root README.md carries a permanent route to the index and no live task table.
• Lifecycle workflows — plan, research, handoff, review, resume, release, init — read and write task state instead of the board.
• The status legend moved from the README to .tfw/glossary.md § Status Flow, where the vocabulary already lived.
• Templates use {ID} where they used {PREFIX}-{N}: both identifier grammars are readable everywhere.

Fixed

• TD-81 — the documentation generator no longer regex-reads board columns. A test now fails if a board-shaped table regex is reintroduced into docs/scripts/.
• TD-177 — the board's schema cannot drift, because there is no board.

Migration

One setting decides the layout. tfw.task_containers is a list. A new project sets one container. A project with an existing corpus lists its old container second — that is one value with two entries, not two supported layouts, and nothing else in the method changes.

tfw:
  task_containers: [workspace, tasks]   # create in the first; resolve across all

Superseded by .tfw/migrations/2.0.0.md (2026-08-30). The commands below are the record of what 2.0.0-dirty shipped — the tooling lived at docs/scripts/ then — and do not run on a project that receives the payload. Follow the guide; read this as history.

Nothing existing is renamed, moved or byte-changed. Run python docs/scripts/migrate_board.py for a dry run and read the accounting; run it with --apply to write. It adds a status.md to each task still in flight and captures the board verbatim as tasks/BOARD-SNAPSHOT.md. It opens no existing artifact in write mode and refuses to overwrite anything.

Renaming the old corpus into the new grammar was measured and refused: at this project's own migration the old identifiers were carried by 7,505 references across 666 files and 271 commit subjects. A trace that needs a translation table to be read has already lost the property the framework exists to provide.

Then generate the view, and only then remove the board:

python docs/scripts/migrate_board.py --apply
python docs/scripts/gen_index.py
# now delete the Task Board section from README.md and put the route in its place

That order matters. The project must never be without a portfolio view.

What is not in this release. Transport — whether a project collaborates through Git or through file synchronization — is a declared project mode owned by a separate task. Nothing here requires a daemon, database, lock server, vendor API or MCP host, and nothing here is required for a task to be read or advanced: with the generator deleted, tasks stay readable and workable and only discovery degrades.

Also in this release — TFW-55, Foundations

Phase A is the headline, but 2.0.0 is the first tag since 1.3.0 and TFW-55 closed in between. Its changes reach every project that updates, so they are listed rather than left to be discovered:

• .tfw/README.md is rewritten as the Philosophy of Trace. It now argues from what a trace is and what makes work continuable, instead of describing a process. 187 lines changed.
• The Project North Star is designated and populated. NS1 — Purpose, NS2 — Principles and NS3 — Non-goals are real sections with citation anchors. TFW-53 shipped the concept with nothing in it; a reviewer's Purpose Check now has something to read. Non-goals exist for the first time.
• .tfw/glossary.md gains 50 lines, including the PV priority entries the North Star introduced. This is a 🟡 merge file on update — a project with its own glossary terms must reconcile, not overwrite.
• The root README is rebuilt and localized. README.ru.md and README.kk.md are new, 240 lines each.

On update: .tfw/README.md is replaced wholesale, .tfw/glossary.md needs a merge, and the localized root READMEs are this repository's own content rather than framework files — an updating project does not receive them and does not need them.

[1.3.0] — 2026-08-18

A failed task can finally be closed as failed. TFW-53 Phase E, the last phase of the contract work. The status set could record success, work in flight and waiting — and nothing could record failure, so the only ways to close a failed task were to lie with ✅ DONE, misuse ❌ BLOCKED, or delete the folder. This project did the third: a whole-tree restore took README.md back to a state that had never contained two rejected tasks' rows, and the failure status they carried disappeared as a side effect of the method. Nobody decided it, and no rule was broken — which is why this release ships a rule rather than a reminder.

Why REJECTED and not BLOCKED. ❌ BLOCKED was defined, listed and available at the moment the rejected task closed, and the coordinator declined it and hand-wrote a token the framework did not have. BLOCKED has 0 occurrences across 46 Task Board rows. The two states are different: blocked is waiting and resumes; rejected is closed and keeps its trace. REJECTED is also the only candidate name carrying a collision, and it was kept anyway — every alternative (FAILED, CANCELLED, ABANDONED, DROPPED) presupposes the reason, and the reason belongs in the board row's description.

Added

• ❌ REJECTED as a terminal task status — the status set could record success, waiting and work in flight, but never failure, so closing a failed task meant misusing ✅ DONE, misusing ❌ BLOCKED (which means waiting) or deleting the folder. Present in conventions.md §5 (table and transition diagram, drawn as a side node reachable from any status), project_config.yaml, templates/project_config.yaml, glossary.md ## Status Flow and the README legend. Terminal — no status follows it. It is a task status, distinct from the review verdict ❌ REJECT and from the HL §12 amendment verdict of the same name; neither of those is terminal (TFW-53/E)
• conventions.md §13 — reverting a result does not revert its trace. A rejected task's folder and its board row are never deleted: the work may leave the working tree, the record that the work happened stays (TFW-53/E)
• conventions.md §14 anti-pattern — a whole-tree restore reverts the Task Board past a task's failure status. Restoring every file to an older tree also restores rows to a state that never contained the newer ones, so the loss happens silently and nobody decides it (TFW-53/E)

Changed

• glossary.md ### Amendment Log — one clause separating the HL §12 amendment verdict ❌ REJECTED from the new terminal task status of the same name. The collision is stated at both ends, so whichever file an agent opens first it learns there are two (TFW-53/E)

Notes for upgrading projects

• Additive only. /tfw-update brings one new entry in tfw.statuses, one row and one diagram node in conventions.md §5, one sentence in §13, one anti-pattern in §14, and two glossary.md edits. No status was renamed, no transition redrawn, no template field changed, no file removed. A board that never used ❌ BLOCKED and never needs ❌ REJECTED is unaffected.
• The status also lands in templates/project_config.yaml, so a newly initialised project is born with it rather than acquiring it by upgrade.
• Version note. RELEASE.md §3 lists "status flow changed" under MAJOR. This change adds a state and alters no existing one; docs/scripts/gen_docs.py parses board rows by regex with no hardcoded status set, and 68 tests pass unchanged. Released as MINOR on the impact test, by owner decision — the same standard applied at 1.1.0, where §3's "required file removed" clause also over-classified an additive-in-effect change. Second occurrence of the same misfire; recorded as tech debt against §3.

[1.2.0] — 2026-08-14

An approved HL becomes a contract, and the reviewer becomes its defender. Phases A–D of TFW-53 in one entry. Phase E (rejected-task trace restoration) is independent and not in this release.

Why it exists. An inviolable contract with a defender is the precondition for delegation: releasing a coordinator to run a team of agent sessions is only safe once the goals cannot move and something checks the result against them. That delegation mode is TFW-54, deliberately a separate task — building both at once splits the coordinator's focus, which is the failure this work exists to prevent.

Added

• HL Contract — on owner approval, templates/HL.md §1, §3, §4, §5, §6 and §7 freeze; §2, §7.2 and §8–§11 stay free; §12 becomes append-only. Carried by a header Contract field with two states, which tracks the artifact where task status tracks the pipeline. The frozen unit is the declarative claim, not the section text. conventions.md §3 owns the 21 rules, including the requirement that the approved HL be committed before the first research iteration — an uncommitted baseline makes "frozen" permanently unverifiable (TFW-53/A, TFW-48 precedent)
• §12 Amendment Log — the only channel for changing a frozen claim: a dated, evidenced proposal carrying cost and a considered alternative, ruled by an explicit owner verdict. Append-only, so a refused proposal stays visible as an attempt. Type states relation to the baseline (EXTEND / SUPERSEDE / RESTRICT), never disposition. templates/RES.md splits its recommendations to match: Refinements the coordinator applies, Amendment Proposals it may not (TFW-53/A)
• Purpose Check — templates/review/judge.md row 2 clause (a) asks is this what we set out to do? against the contract baseline plus the Project North Star, never the TS and never a Phase HL. One field quotes the clause served and names the concrete harm; a citation that resolves but is irrelevant fails the row, and so does a harm asserted with no citation. Three tests — excess and adjacency, deferral confession, materiality. Three outcomes: aligned, not fit for purpose (grounds ❌ REJECT with every other check passing, routed to the owner), and reference set internally inconsistent (a contract defect, also to the owner). "The TS scoped it this way" and "tests are green" are named as insufficient grounds to approve (TFW-53/C)
• PV Index priority 0 — Project North Star — an anchor above the task HL, answering what we are building, why, and what we are deliberately not building. Locus: designated section(s) of a README, never a task HL. Optional, with a declared fallback to master HL §1 at its contract baseline, so a review is never blocked on its absence. templates/HL.md gains a Project North Star header field. Citation namespace NS{n}; PP{n} for a project principle registry; HL §7 P{n} unchanged (TFW-53/C)
• Ten glossary articles — HL Contract, Contract Baseline, Frozen Section, Amendment, Amendment Log, Project North Star, Purpose Check, not fit for purpose and Deferral confession under a new ## Contract and Purpose Defence grouping, plus the long-missing Result Visualization beside Value Flow and Findings Map. Eight of the ten had zero definitions while five files used them (TFW-53/D)
• Anti-patterns in conventions.md §14 — silent frozen-section edits, unclassified research recommendations, applying an amendment before its verdict, research on an uncommitted baseline, a research-thread remark treated as a verdict, an agent citing its own delegation to accept an overrun, a Phase HL authoring its own acceptance criteria or principles, a reviewer approving work that satisfies the TS but not the contract, and alignment asserted without citing the clause it serves (TFW-53/A, /C)

Changed

• plan.md Step 6c is inverted — from "Update HL with research findings (present diff to user)" to classify, apply, log, escalate: refinements applied silently, amendments transcribed into §12 as PROPOSED with the section left untouched, and one batched escalation per iteration carrying evidence, cost and alternative. A coordinator may not apply a proposal it filed. Both verdict paths are specified, including a re-freeze commit at the new baseline after every approved amendment (TFW-53/B)
• research/base.md — the researcher classifies every recommendation by target section and never edits the HL (TFW-53/B)
• Reviewer Identity — "Quality guardian, not rubber stamp", extended to name the third defended object: goals, values and the north star, with authority to block work that is verified, complete and beside the point. D46 recorded the "not rubber stamp" half in April and only the first half ever shipped (TFW-53/C, D46)
• review.md:28 loads the master HL at its contract baseline, not the current file — without this the reference-set rule has nothing to bind to and reviewers keep reading the drifted version (TFW-53/C)
• conventions.md §5 REJECT branch (a) — "rework HL" now means file an amendment against the frozen sections; re-entry to 📝 HL_DRAFT reopens the free sections only. It was the one documented path that reopened a frozen contract with no proposal and no log (TFW-53/A)
• conventions.md §3 — a Phase HL is derivation-only. It may restate master content and add execution context; it may not carry its own §1, §5, §6 or §7. TFW-48's Phase A HL was a complete second contract that silently dropped three master principles (TFW-53/A)
• templates/HL.md §3.1 — Working Backwards and visualization are mandatory, not format options: written from the finished state, rendered visually with prose alone insufficient, showing the value and not only the artifact, and complete enough for a multi-phase task to be held at once (TFW-53/A)
• One name per concept — frozen baseline and committed frozen baseline retire in favour of Contract Baseline; the templates/HL.md north-star field label becomes Project North Star, so the form teaches the name the glossary defines (TFW-53/D, D28)
• Scope budgets raised to the owner's working values — max_files_per_phase 14 → 30, max_new_files 8 → 15, max_loc 1200 → 3000, max_modified_files 12 → 30. Changed in both places that carry the numbers: templates/project_config.yaml (what a new project is born with) and conventions.md §6 (the defaults table agents read inline). Rationale: the standard is set by observed practice, not by the shipped template — the owner's project had run at these values for months while the template still claimed the old ones, so an upgrade would have silently reverted them. Owner instruction, 2026-08-13. See D62
• Adapter Sync section completed — workflows/config.md. It documented 4 of 11 workflow files and 1 of 3 adapter folders; it now carries the full source → copy mapping for both full-copy folders (.claude/commands/, .agent/workflows/), an explicit not copied, and why table (research mode files, the adapter-only tfw-task, Codex thin routers), and a runnable drift check that prints every copy no longer matching its source. Two anti-patterns added: copying to one adapter folder only, and reporting a sync as done without running the check. At Phase D the check printed 14 drifted copies (7 workflows × 2 folders); all fourteen were re-synced and the check now runs silent (TFW-53/D)
• CLAUDE.md — the /tfw-plan and /tfw-review purpose cells describe what those workflows now do, and the table gains the two rows it was missing, /tfw-knowledge and /tfw-config (TFW-53/D)

Removed

• The Judge mapping-integrity check — "did the AC each HL §7 principle was mapped to pass?" It is structurally unable to detect a principle violated by the mapping itself: in a reconnaissance corpus it returned ✅ on the very acceptance criterion that carried the violation, and the reviewer later retracted his own APPROVE. Replaced by the Purpose Check above; row 2 clause (b) design soundness is unaffected (TFW-53/C)
• templates/RES.md:32 — <!-- List what should change in HL based on research. Coordinator applies these. --> The template-side twin of plan.md Step 6c. Fixing one and not the other would have reproduced the drift through the second channel (TFW-53/A)
• The inline baseline recovery command in templates/HL.md — a fourth copy of git log --format=… in the template every HL is born from, replaced by a pointer to conventions.md §3 rule 15 so it can be corrected in one place. Upgrading projects: existing HL headers keep whatever they carry — history is not rewritten; the change affects HLs created from this version onward (TFW-53/D, TD-164)
• KNOWLEDGE.md §0 from compilable_contract.md's "Where references appear" list — a section D37 removed in April (TFW-53/D, TD-167)

Fixed

• compilable_contract.md — NS{N} and PP{N} were declared in the §2 pattern table with no resolution behaviour; they now sit in the Resolution rules beside D{N}, P{N}, F{N} and TD-{N} (TD-165)
• glossary.md — the Knowledge Gate is in Step 2 of plan.md, not a "Phase 0" that does not exist (TD-163)
• Adapter parity restored — both tfw-plan copies still carried the retired "Update HL with research findings", so two of three surfaces instructed the coordinator to do what the core now forbids (TD-157). Six further workflows were repaired in the same pass: tfw-init, tfw-handoff, tfw-update, tfw-knowledge, tfw-review, tfw-research

[1.1.0] — 2026-08-13

Added

• Three promoted universal Judge rows — templates/review/judge.md grows from 7 to 10 rows, each promoted row carrying its measured non-✅ rate from a 637-row / 203-review / 3-install corpus: Evidence sufficiency (16.1% — the highest-firing check in TFW review; four gated rows turned out to be one check in three genre costumes), Backward compatibility (8.5%), Safety (4.0%, retained on consequence rather than frequency). Row 2 Philosophy aligned is sharpened into two separately answered clauses — mapping integrity and design soundness (4.5%) (TFW-56, D42 revoked)
• Structural explicit-N/A grammar in the Judge checklist — status vocabulary is ✅ / ❌ / ⚪ N/A, and ⚪ N/A requires a stated reason. A row skipped as a bare ✅ leaves the stage incomplete. Rows 7 and 8 carry an explicit contrast note (does the evidence exist vs does it establish the claim) plus a Checkpoint item requiring they be answered separately (TFW-56, F21)
• Claim & Source Checks in templates/review/verify.md — the three docs/spec verify actions promoted to unconditional: spot-check 2-3 key claims or sources, confirm every citation traces to a real artifact, verify data claims against a primary source. Table + Checkpoint item; feeds Judge row 8 (TFW-56)
• Anti-pattern in conventions.md §14 — a review checklist row added without an evidenced firing rate. Retention on consequence rather than frequency is permitted and must be written into the row (TFW-56)

Changed

• review.md steps renumbered 0-7, contiguous — Step 0 is Session Naming, the TFW standard this file never followed. Map 2→1, Verify 3→2, Judge 4→3, Decide 5→4, Tech Debt 6→5, Update Traces 7→6, Knowledge Capture 8→7. The Verify step now states that every action in verify.md is unconditional and that depth is set by min_verify_ratio, never by the kind of work under review (TFW-56, TD-106 closed by deletion)
• templates/REVIEW.md §3 realigned row-for-row with judge.md — ten rows in the same order. This also repairs a pre-existing gap: the Evidence completeness row added to judge.md in 0.8.8 had never reached REVIEW.md §3 (TFW-56)
• glossary.md — Reviewer heading is "coordinator under the reviewer Role Lock", so the phrase "review mode" no longer carries two meanings (D28); entry describes one universal 10-row checklist; Principles Check pointer corrected to review.md Step 3 (TFW-56)
• workflows/config.md — the review propagation section keeps only min_verify_ratio, whose step pointer is correct in the renumbered workflow (TFW-56)
• All adapter copies re-synced: .claude/commands/tfw-{review,config}.md, .agent/workflows/tfw-{review,config}.md, .tfw/adapters/codex/skills/tfw-review/SKILL.md, .agents/skills/tfw-review/SKILL.md (TFW-56, D54)

Removed

• Config key tfw.review.default_mode — removed from .tfw/project_config.yaml and .tfw/templates/project_config.yaml. Upgrading projects: this key is now inert. /tfw-update triages files, not keys, so a leftover default_mode: code line will not be flagged and will not break anything — delete it from your tfw.review block. tfw.review.min_verify_ratio and its 0.42 default are unchanged (TFW-56)
• Review mode files — .tfw/workflows/review/{code,docs,spec}.md and the folder itself. Byte-identical across three installs, two framework versions and two product domains; never used as an extension point. Their first verify action duplicated verify.md's Checkpoint (TFW-56, D42 revoked)
• The review mode selection step and its 🛑 WAIT gate — review.md Step 1. In 203 mode-carrying reviews no mode row was ever the sole non-✅ driving a verdict: the rows carried signal at ~8%, the selection in front of them flipped nothing. What to check is declared once, by the TS — acceptance criteria (D49) and Evidence: fields (D52) (TFW-56)
• Mode: / Review Mode template fields — from templates/review/{map,verify,judge}.md and templates/REVIEW.md, together with the mode-specific placeholder comment. Existing REVIEW files keep their headers; history is not rewritten (TFW-56)
• docs Content quality checklist row — dropped rather than promoted; the one true duplicate of universal row 4 Style & standards (TFW-56)

[1.0.0] — 2026-08-06

Added

• Minimal Commit Attribution — AI-authored commits use the searchable [agent/task/scope/role] summary subject format with explicit field meanings and a clear separation from Git author/committer metadata and actor authentication (TFW-50, D55)

Changed

• Commit subject contract (breaking) — Coordinator, Researcher, Executor, and Reviewer use the conventions-owned format when they create commits; the rule formats existing commit actions and creates no commit cadence
• Handoff and release publication boundary — ONB and release commits use Commit Attribution, while push and remote tag publication remain unavailable until explicit user approval

[0.9.0] — 2026-07-22

Added

• Evidence Enforcement — evidence/ folder mandatory in every task directory. EV template (.tfw/templates/evidence/EV.md) with Environment header, per-AC evidence table (4-status vocabulary), Verdict line, optional Attachments index. Naming: EV__{PREFIX}-{N}__{title}.md. D16 (optional folder) revoked; D53 (TFW-47/A)
• Codex first-class adapter — 11 handwritten shortcut skills in .tfw/adapters/codex/skills/tfw-*/SKILL.md, installed to .agents/skills/tfw-*/SKILL.md. Two-layer architecture: root AGENTS.md always-on recognition (TFW:CODEX marker block) + repo-local skills for discovery and progressive loading. D54 (TFW-47/B)
• Codex adapter README — executable install/repair contract: detect state → install copies → merge AGENTS block → remove legacy → verify → runtime contract (TFW-47/B)
• Codex in init.md — Phase 0 full-init vs existing-project attach/repair detection, Codex skill install and verification steps (TFW-47/B)
• Codex in update.md — safe command/routing re-sync with marker ownership, legacy cleanup guard, literal slash smoke test (TFW-47/B)
• Evidence subfolder section in conventions.md — §4 documents evidence/ as mandatory subfolder alongside research/ and review/ (TFW-47/A)

Changed

• RF template §5 — inline evidence table replaced with pointer to EV file + verdict summary. Cognitive mode clarified: executor's observational work lives in EV file (TFW-47/A)
• TS template — added ### Evidence Artifacts subsection after AC items with guidance and example table (TFW-47/A)
• handoff.md Step 11 — rewritten with 6 numbered substeps for evidence folder creation, template copy, and population. Skip condition removed — evidence always required (TFW-47/A)
• conventions.md §3 — evidence pipeline table updated: EV file row added, RF row clarified as "Summary / Reference" (TFW-47/A)
• conventions.md §14 — evidence anti-pattern wording strengthened: "evidence/ folder" language, "VERIFIED without artifact" prohibition (TFW-47/A)
• glossary.md — Adapter Command entry includes Codex (.agents/skills/tfw-*/SKILL.md), Tool Adapter definition updated with two-layer Codex architecture (TFW-47/B)
• conventions.md — added Codex two-layer adapter pattern and cross-tool /tfw-* command contract (TFW-47/B)
• quickstart.md — Codex install/repair handoff, corrected lifecycle/four-role summary (TFW-47/B)

Removed

• Legacy source-command-tfw-* skills — stale full-workflow imports that duplicated canonical workflows and created a second source of truth (TFW-47/B)

[0.8.8] — 2026-07-07

Added

• Evidence Layer — real-world verification as first-class TFW concept, separate from synthetic Verification (§4). Three-role pipeline: coordinator designs Evidence Plan (TS), executor collects evidence (RF §5), reviewer audits evidence (REVIEW). Fixed 4-status vocabulary: VERIFIED / DEFERRED / BLOCKED / N/A (TFW-46, D52)
• Evidence concept in conventions.md — §3 Evidence Sections table (4 per-template entries with cognitive modes), §12 evidence honesty rule, §14 five anti-self-deception anti-patterns (TFW-46/A)
• Evidence field in TS template — Evidence: field after Gate: in AC items with MAY-deviate instruction. Grammar: full spec, minimal, N/A, DEFERRED, or empty (TFW-46/A)
• §5 Evidence section in RF template — table (AC, What, Environment, Result, Artifact) + evidence verdict line. §5-8 renumbered to §6-9 (TFW-46/A)
• Evidence Audit in review stage files — judge.md check #7 (Evidence completeness), verify.md Evidence Verification section with table and N/A fallback (TFW-46/A)
• Step 11 (Collect evidence) in handoff.md — between build gate (Step 10) and Pre-RF Gate (Step 12). Proportionality clause, DEFERRED/BLOCKED guidance, proactive tooling note. §5 Evidence in mandatory sections (TFW-46/B)
• Trust Protocol evidence entries in review.md — 2 new entries: "Evidence: VERIFIED" (Verify level), "Evidence: N/A or no evidence" (Challenge level) (TFW-46/B)
• Evidence reminder in plan.md — Step 7 sub-step 3: coordinator considers Evidence fields when writing TS AC items (TFW-46/B)
• 5 Evidence glossary terms — Evidence, Evidence Plan, Evidence Collection, Evidence Audit, Evidence Status Vocabulary (TFW-46/C)

Changed

• RF template renumbering — §5 Observations → §6, §6 Fact Candidates → §7, §7 Strategic Insights → §8, §8 Diagrams → §9. All cross-references updated across templates, workflows, conventions (TFW-46/A)
• glossary.md — Strategic Insight entry: RF §7 → §8 (stale ref fix from Phase A renumbering) (TFW-46/C)
• All adapter copies synced: .agent/workflows/tfw-{handoff,review,plan}.md, .claude/commands/tfw-{handoff,review,plan}.md (TFW-46/C)

[0.8.7] — 2026-05-01

Added

• Mindset blocks in research templates — per-stage cognitive anchoring: Strategist (Briefing), Explorer (Gather), Analyst (Extract), Critic (Challenge). Each template has > **Mindset:** + > **Test:** blockquote between h1 and > Parent: line. Matches review template pattern (D41) (TFW-43)
• Briefing h1 guiding question — # Briefing — "What should we investigate?" added for consistency with other 3 stages (TFW-43)

Changed

• Copy-on-enter protocol in research/base.md — Step 3 creates folder only (no template copy). Step 4 copies briefing template before writing. Step 5 restructured as FOR EACH loop: copy template → read Mindset → OODA → checkpoint → 🛑 STOP per stage. Restores D31 (file existence = stage completion) (TFW-43)
• All adapter copies synced: .agent/workflows/tfw-research.md, .claude/commands/tfw-research.md (TFW-43)

Removed

• Batch template copy from research/base.md Step 3 — all 4 templates were copied at once, breaking D31 (file existence = stage completion) and observable progress. Replaced by copy-on-enter (TFW-43)

[0.8.6] — 2026-04-30

Changed

• Research folder structure — researchN/ flat folders at task root replaced by single research/ container with iterN/ subfolders. RES files co-located with stage files (research/iterN/RES.md). iterations.yaml moved inside research/ subfolder (TFW-42/A)
• Stage file numbering — briefing.md, gather.md, extract.md, challenge.md renamed to 1_briefing.md, 2_gather.md, 3_extract.md, 4_challenge.md. Sort order = execution order (TFW-42/A)
• Phase folder naming — PhaseA/, PhaseB/ → phase-a/, phase-b/ (kebab-case, consistent with D48) in conventions.md §4 and artifact filename table (TFW-42/A)
• iterations.yaml schema — added optional agent (free-text, traceability) and sources (list, source categories) fields. Backward compatible (TFW-42/A)
• research/base.md — Steps 0, 3, 4, 5, 6 updated: all paths use research/iterN/, numbered stage file names, co-located RES.md (TFW-42/B)
• plan.md — Step 6b: research/iterations.yaml path + agent/sources fields + multi-agent reference. Step 6c: updated RES paths. Step 7: phase-a/ naming (TFW-42/B)
• glossary.md — "Iteration (Research)": research/iterN/ paths, co-located RES. "iterations.yaml": research/ location, agent/sources fields (TFW-42/C)
• All adapter copies synced: .agent/workflows/tfw-{research,plan}.md, .claude/commands/tfw-{research,plan}.md (TFW-42/C)
• D50 in KNOWLEDGE.md §1, TFW-42 in §2 (TFW-42)
• domain F4, philosophy F25, process F21-F22, stakeholder F2 in knowledge/ topic files (TFW-42)

[0.8.5] — 2026-04-20

Added

• Requirements-first TS template — §4 Detailed Steps replaced by §5 Acceptance Criteria (verifiable gates with [depends: AC-X] dependency annotations), §3 Principles Check (HL §7 → AC mapping table), §6 Technical Guidance (reference, not instructions), §7 Definition of Failure (hard reject conditions), §9 Cross-Phase Modifications (multi-phase conflict tracker) (TFW-41/A, D49)
• Pre-TS Gate in plan.md Step 7 (3b) — coordinator reads RF of latest completed phase before writing next TS. Ensures planning against actual output, not prior plan (TFW-41/B, D49)
• Pre-RF Gate in handoff.md Step 11 — executor opens RF template and reads section headings before writing RF (TFW-41/B, D49)
• Execution Loops in handoff.md Phase 2 — when TS ACs have [depends: AC-X], executor verifies prerequisite AC gate before implementing dependent AC (TFW-41/B, D49)
• Session Naming Step 0 in handoff.md, plan.md, review.md — Role | Task-ID | Phase at session start (TFW-41/B, D49)
• ONB answer protocol in handoff.md — coordinator presents options with tradeoffs, does not decide for stakeholder (TFW-41/B, D49)
• HL §7 Principles check in review.md Judge stage — reviewer traces HL §7 → TS §3 → RF §3 for each principle (TFW-41/B, D49)
• Phase Dependencies section in HL template §4 — mermaid graph + dependency matrix for multi-phase tasks (TFW-41/A, D49)
• Embedded dimensional analysis in research templates — ## Dimensions in gather.md, ## Configuration Space in extract.md, ## Consistency Check in challenge.md. Cross-stage structural dependency as natural enforcement (TFW-41/C, D49)
• Dimensional analysis thread in research/base.md Step 5 — 3-sentence connecting logic with graceful degradation for <3 dimensions (TFW-41/C, D49)
• §14.1 Terminology Origin in conventions.md — maintainer-facing note mapping TFW terms to Zwicky GMA equivalents (TFW-41/C)
• 4 anti-patterns in conventions.md §14 — code-in-TS, coordinator planning drift, RF-from-memory, ONB source-less answers (TFW-41/A)
• 15 glossary terms — 10 execution gate terms (Acceptance Criteria, Technical Guidance, Definition of Failure, Principles Check, AC Dependency Annotation, Execution Loop, Pre-TS Gate, Pre-RF Gate, Session Naming, Phase Dependencies) + 5 dimensional analysis terms (Dimension, Alternative, Configuration Space, Consistency Check, Surviving Configuration) (TFW-41/D)
• D49 in KNOWLEDGE.md §1, TFW-41 in §2, 2 legacy entries in §3 (TFW-41)
• philosophy F24 (instructions→compliance, heuristics→competence), process F18-F20 in knowledge/ topic files (TFW-41)

Changed

• TS.md template — complete structural rewrite: §3 Principles Check, §4 Affected Files (with budget), §5 Acceptance Criteria (from §4 Detailed Steps), §6 Technical Guidance, §7 Definition of Failure, §8 Phase Risks, §9 Cross-Phase Modifications. Line count 52→84 (TFW-41/A)
• handoff.md — Step 0 (Session Naming), Execution Loops in Phase 2, ONB answer protocol, Pre-RF Gate in Phase 3. Line count 148→161 (TFW-41/B)
• plan.md — Step 0 (Session Naming), Pre-TS Gate in Step 7 (3b). Line count 145→153 (TFW-41/B)
• review.md — Step 0 (Session Naming), step renumbering (Select Review Mode = Step 1), HL §7 Principles check in Judge. Line count 145→153 (TFW-41/B)
• research/base.md — dimensional analysis thread in Step 5. Line count 129→131 (TFW-41/C)
• gather.md — ## Dimensions section before Findings. Line count 25→40 (TFW-41/C)
• extract.md — ## Configuration Space section before Findings. Line count 25→42 (TFW-41/C)
• challenge.md — ## Consistency Check section before Findings. Line count 25→47 (TFW-41/C)
• glossary.md — 2 new sections: ## Execution Gates (10 terms), ## Research — Dimensional Analysis (5 terms). Line count 197→246 (TFW-41/D)
• All Antigravity adapters synced: .agent/workflows/tfw-{handoff,plan,review,research}.md (TFW-41/D)
• philosophy F13 upgraded to ✅ verified (3 sources) with TFW-41 user quote on domain-agnosticism (TFW-41)

Removed

• TS §4 Detailed Steps — procedural implementation instructions replaced by requirements-first Acceptance Criteria (TFW-41/A, D49)

[0.8.4] — 2026-04-15

Added

• State/framework file classification — §10.3 in conventions.md: 3-category model (Framework, State, Config) with lifecycle rules. State files NEVER overwritten from upstream (TFW-40/A, D47)
• YAML naming convention — §10.4 in conventions.md: lower_snake_case for all .tfw/ YAML and template files. Uppercase reserved for root docs and .tfw/ framework docs (TFW-40/B)
• Templates for state/config files — .tfw/templates/knowledge_state.yaml (clean seq=0), .tfw/templates/project_config.yaml (annotated ← PROJECT / ← FRAMEWORK markers) (TFW-40/A)
• ⚫ STATE category in update.md — files never overwritten during tfw-update (knowledge_state.yaml, knowledge/, KNOWLEDGE.md, TECH_DEBT.md) (TFW-40/A)

Changed

• PROJECT_CONFIG.yaml → project_config.yaml ⚠️ BREAKING — all references updated across workflows, templates, adapters, conventions, glossary, compilable_contract, README, KNOWLEDGE.md, gen_docs.py (TFW-40/B)
• TOPIC_FILE.md → topic_file.md — template renamed, references updated in conventions, glossary, knowledge.md workflow (TFW-40/B)
• init.md — Phase 2 Mini-Setup now copies from templates (not upstream files), preventing state contamination (TFW-40/A)
• update.md — added ⚫ STATE category, explicit merge rules for project_config.yaml (preserve project sections, update framework sections) (TFW-40/A)
• gen_docs.py — config path updated to project_config.yaml (L165-166, L530) (TFW-40/B)
• All Claude Code adapters synced: .claude/commands/tfw-*.md — full sync from canonical .tfw/workflows/ (11 files). Fixes stale PROJECT_CONFIG.yaml references and accumulated drift from TFW-38+TFW-40 (TFW-40/B)

Migration Notes (⚠️ BREAKING)

Projects upgrading from ≤0.8.3 must: 1. Rename .tfw/PROJECT_CONFIG.yaml → .tfw/project_config.yaml 2. Rename .tfw/templates/TOPIC_FILE.md → .tfw/templates/topic_file.md (if exists) 3. Update any custom adapter files referencing PROJECT_CONFIG.yaml 4. Update docs/scripts/gen_docs.py if customized (config path changed)

[0.8.3] — 2026-04-15

Added

• 4-stage review flow — Map → Verify → Judge → Decide. Each stage = separate template file in .tfw/templates/review/ with mindset-based identity (Student/Auditor/Judge/Decision-maker) and self-check gate. Mode selection (code/docs/spec) with 🛑 WAIT gate (TFW-38/A, D41)
• Review mode files — .tfw/workflows/review/{code,docs,spec}.md. Mode-specific checklists (2-4 items) loaded at Step 2. Progressive Disclosure — agent loads only needed mode. 6 universal + mode-specific = hybrid (TFW-38/A, D42)
• Knowledge Citation Table — cascade model: Coordinator does full PV scan → HL §7.2, Executor reads HL §7.2 → ONB §7 (confirms/extends), Reviewer verifies links → verify.md (anti-hallucination gate). Unified name "Knowledge Citations" (TFW-38/B, D43)
• Project Values (PV) term — unified term for all accumulated project context. PV Index = 7 sources with scan priority in glossary.md. Replaces ambiguous "check values/knowledge/experience" (TFW-38/B, D44)
• Reviewer Identity — overall identity statement + per-stage mindsets. Trust Protocol table (7 rows). 🛑 WAIT gate on mode selection (TFW-38/A.2, D46)
• Knowledge Input Sections table in conventions.md §3 — §7.2 HL, §7 ONB, verify.md Citations Verified (TFW-38/B)
• D41-D46 in KNOWLEDGE.md §1 (TFW-38)
• TFW-38/A, A.2, B in KNOWLEDGE.md §2 Key Artifacts (TFW-38)
• 4 legacy entries in KNOWLEDGE.md §3 (TFW-38)
• philosophy F20 (investigative vs procedural workflow classes), F21 (explicit N/A as universal design principle) in knowledge/philosophy.md (TFW-38)

Changed

• review.md — rewritten: Role Lock updated, Reviewer Identity + Trust Protocol added, Steps 0-4 file-based (create stage files → synthesize into REVIEW), Steps 5-7 traces + knowledge capture (TFW-38/A+A.2)
• REVIEW.md template — restructured §1-§7: Map/Verify/Judge/Verdict/Tech Debt/Traces/Fact Candidates. Stage files listed in header. Synthesis instruction (TFW-38/A)
• HL.md template — §7.2 Knowledge Citations added (PV scan instruction, 4-column table, bootstrap note) (TFW-38/B)
• ONB.md template — §7 Knowledge Citations added (executor read-confirm, 5-column table, NEW row support) (TFW-38/B)
• verify.md template — Knowledge Citations Verified section + citation count in self-check checkpoint (TFW-38/B)
• plan.md Step 3 item 4 — "Check KNOWLEDGE.md" replaced with full PV scan instruction referencing glossary.md PV Index (TFW-38/B)
• handoff.md Phase 1 step 2 — citation-reading sub-bullet added before inconsistency check. Phase 3 step renumbered 12→11 (TFW-38/B, TD-94)
• conventions.md §15 Role Lock — review.md row updated with stage files (TFW-38/A)
• glossary.md — Reviewer updated (mode-aware + stage files), RESEARCH updated (pros/cons), Pass updated (OODA + sufficiency verdict) (TD-35, TD-36, TD-98)
• README.md — docs site link added to Links section (TD-92)
• TECH_DEBT.md — purged 41 closed items, 11 remaining. TD-33/TD-59 closed with rationale (tech debt audit)
• All adapters synced: .agent/workflows/ (TFW-38/B)

Removed

• TFW-37 (Source Audit gate) — absorbed into TFW-38 (4-stage review + Trust Protocol + docs mode source verification)
• Single-pass REVIEW workflow — replaced by 4-stage flow with file-based evidence (TFW-38/A, D41)
• 9-point monolithic review checklist — replaced by 6 universal + mode-specific (TFW-38/A, D42)
• Silent "I checked KNOWLEDGE.md" pattern — replaced by Knowledge Citation Table with verifiable links (TFW-38/B, D43)

[0.8.2] — 2026-04-10

Added

• Multi-iteration research — iterations.yaml control file, min_iterations config (default: 2), coordinator hard gate in plan.md Step 6c, researchN/ subfolder accumulation (never delete/overwrite), Iteration Status block in RES template, iter2+ briefing protocol in research/base.md (TFW-32/C, D38)
• Per-template visual sections — HL §3.1 Value Flow, RF §8 Diagrams, RES Findings Map. Convention cross-ref table in conventions.md §6. Per-template criterion: "what would THIS artifact's reader draw on a whiteboard?" (TFW-32/B, D39)
• 4-part template instruction structure — Cognitive mode → Scope → Human-Only Test → Before writing. Applied to HL §6/§11, RF §6/§7, RES FC/SI, REVIEW §5 (TFW-32/B)
• 📚 KNW pipeline status — 9th status between REV and DONE. Optional (reviewer can pre-close with N/A). REVIEW markers for tfw-docs/tfw-knowledge orchestration (TFW-32/A, D37)
• docs/knowledge exclusive write territories — tfw-docs owns KNOWLEDGE.md §1-§3, tfw-knowledge owns knowledge/ + §4. Explicit ⚠️ warnings in both workflows. Resolves collision (TFW-32/A, D37)
• README "How TFW Compares" section in .tfw/README.md — TFW vs Confluence/Notion vs AI assistants vs no methodology (TFW-32/D, D40)
• Positioning specs — audience_personas.md (3-tier hierarchy), positioning_spec.md (generates-vs-stores), translation_table.md (20 terms), philosophy_improvement.md (TFW-32/D)
• D37-D40, TFW-32/A-D in KNOWLEDGE.md §1/§2 (TFW-32/E)
• 13 new facts in knowledge/ topic files: philosophy F15-F18, convention F11-F14, process F11-F15 (TFW-32/E)
• 5 legacy entries in KNOWLEDGE.md §3 for Phase B/C/D changes (TFW-32/E)
• TD-88..92 in TECH_DEBT.md (TFW-32 reviews)
• TFW-33/34/35 future tasks in Task Board (TFW-32/E)

Changed

• README.md opening — interleave variant: imagine→reality→imagine→TFW. 3-tier audience hierarchy (product leaders > analysts > engineers) with qualifying questions. "Generates vs stores" in How It Works. AI-agents-as-team-members frame. 2 new FAQ entries. Expanded Links section (TFW-32/D, D40)
• .tfw/README.md — team dimension in The Problem, SECI generates-vs-stores in The Thesis, team memory table row, role breadth in How TFW Works, team-centric Success Criteria rewrite (TFW-32/D, D40)
• plan.md — Step 6b creates iterations.yaml, Step 6c iteration gate with min_iterations enforcement, phased subfolder diagram in Step 7. Growth: 108→140 lines (TFW-32/C)
• research/base.md — iter2+ briefing protocol (read all previous RES + iterations.yaml), Iteration Status block instruction (TFW-32/C)
• RES template — Iteration Status block, Fact Candidates sharpened with Cognitive mode + scope + Human-Only Test, Strategic Insights (Research) with Human-Only Test, Findings Map section (TFW-32/B+C)
• HL template — §3.1 renamed Value Flow (from Result Visualization), §3.2 added, §11 renamed Strategic Insights (Planning) with Cognitive mode instruction (TFW-32/B)
• RF template — §6 FC sharpened, §7 Strategic Insights (Execution) with Human-Only Test + fallback, §8 Diagrams section (TFW-32/B)
• REVIEW template — §5 FC sharpened with Cognitive mode + reviewer scope, tfw-knowledge marker in §4 (TFW-32/A+B)
• conventions.md — §6 Visual Sections cross-ref table (5 rows), §6 Knowledge Capture Sections table, KNW in pipeline diagram + status table (TFW-32/A+B)
• glossary.md — Strategic Insight updated, Value Flow + Findings Map + Per-template Naming added, KNW definition, pipeline diagram updated (TFW-32/A+B)
• knowledge_state.yaml — seq 31→32, 42→55 total facts (TFW-32/E)
• All adapters synced: .agent/workflows/, .claude/commands/ (TFW-32/C)

Removed

• KNOWLEDGE.md §0 (Philosophy & Principles, 8 entries) — all principles verified in knowledge/philosophy.md or conventions.md. §0 had no updater workflow (TFW-32/A, D37)
• tfw-knowledge Phase 4 writes to KNOWLEDGE.md §1/§2 — caused collision with tfw-docs (TFW-32/A, D37)

[0.8.1] — 2026-04-09

Added

• .tfw/quickstart.md — strict reading list for AI agents (clone → philosophy → glossary → conventions → init.md). Separates learning from execution to resolve bootstrap paradox (TFW-31)
• 3 self-contained README Quick Start prompts — New Project, Existing Project, Already Set Up. Each prompt is fully self-contained with repo URL, TFW description, and slash command references (TFW-31)
• Tutorial Mode mini-examples in init.md — task prefix examples, task board visualization with realistic entries (TFW-31)
• Star CTA in init.md Phase 5 — after value delivery, not during onboarding (TFW-31)
• Slash command listing in "Already set up" prompt — /tfw-plan, /tfw-handoff, /tfw-review, /tfw-resume (TFW-31)
• D36 (agent-first onboarding), TFW-29/31 in KNOWLEDGE.md §1/§2 (TFW-31)
• 6 new facts in knowledge/ topic files: philosophy F12-F14, process F9-F10, convention F10 (TFW-31)
• TD-87 (init.md code-specific interview question) in TECH_DEBT.md (TFW-31)

Changed

• init.md Phase 1 Discover — rewritten domain-agnostic: purpose/goals, documentation, structure, processes, people first; code-specific items last (TFW-31)
• compilable_contract.md — source manifest and nav diagram updated: .tfw/init.md → .tfw/quickstart.md (TFW-31)
• conventions.md §9 — adapter setup reference updated to quickstart.md (TFW-31)
• update.md — merge checklist updated: init.md → quickstart.md (TFW-31)
• KNOWLEDGE.md — §1 Architecture Map Init row updated, §3 Legacy entry added, §4 fact counts updated (36→42) (TFW-31)
• README.md — Quick Start section rewrite, file index and adapter table references updated (TFW-31)
• glossary.md, conventions.md, compilable_contract.md — TFW-29 consistency fixes (redundancy, numbering, reading flows) (TFW-29)

Removed

• .tfw/init.md pointer file (21 LOC) — redundant after quickstart.md became the "Getting Started" entry. All references migrated to quickstart.md (TFW-31)
• Phase 0 Bootstrap from init.md — wrong approach (injected learning into execution workflow). Replaced by quickstart.md (TFW-31)

[0.8.0] — 2026-04-08

Added

• Compilable Contract — §16 in conventions.md (Source Manifest, Reference Format, Resolution Rules, Frontmatter Convention, Output Nav Structure). Agents write text refs (RF TFW-18), build-time resolves to hyperlinks (TFW-26/A)
• Documentation Pipeline — docs/scripts/gen_docs.py (681 LOC, 68 tests), 10 reference resolvers (artifact, phase, HL-dash, TD, D, backtick-path, bare task ID, markdown link rewriter, table anchors, literate-nav), structured tasks index, section indexes, YAML frontmatter injection (TFW-26/A+B, TFW-27/B)
• docs/ infrastructure — mkdocs.yml, requirements.txt (7 packages incl. mkdocs-literate-nav, mkdocs-section-index), .github/workflows/docs.yml (TFW-26/A, TFW-27/C)
• Brand Identity — two-color discipline (charcoal #1a1a2e + teal #0d9488), Inter/JetBrains Mono typography, TFW monogram logo, docs/brand/identity.md (TFW-27/A)
• GitHub Pages Deploy — live site at tfw.saubakirov.kz, auto-deploy on push to master via GitHub Actions (TFW-27/C, absorbs TFW-28)
• Coordinator Fact Capture — philosophy in §10.1 categories, §11 Strategic Session Insights in HL template, Step 4b (fact capture) in plan.md, fact capture reminder in resume.md, "Strategic Insight" glossary term (TFW-26/FC)
• §16 Reference Format reminder in HL, TS, ONB template footers — ensures all artifact authors use resolvable cross-references (TFW-27 post-review)
• HL §11 and RF §7 scan in knowledge.md Phase 2 — explicit scan targets for Strategic/Execution Session Insights (TFW-26 post-review)
• Category coverage check in knowledge.md Phase 2 Step 3 — check §10.1 for unrepresented categories (TFW-26 post-review)
• KNOWLEDGE.md §1/§2 update step in knowledge.md Phase 4 — Architecture Decisions and Key Artifacts entries for closed tasks (TFW-26 post-review)
• D34 (Compilable Contract), D35 (Brand + Wiki + Deploy) in KNOWLEDGE.md §1 (TFW-26, TFW-27)
• TFW-26, TFW-27 in KNOWLEDGE.md §2 Key Artifacts
• 17 new facts in knowledge/ topic files: philosophy F5-F11, process F6-F8, convention F8-F9, constraint F4, stakeholder F1, environment F1-F2 (TFW-26, TFW-27)
• 2 new topic files: knowledge/stakeholder.md, knowledge/environment.md (TFW-27)
• Compilable Contract, Reference Format, Source Manifest glossary terms (TFW-26/A)
• TD-75 (knowledge quality design), TD-76 (terminology unification), TD-79..82 (gen_docs.py debt) in TECH_DEBT.md

Changed

• knowledge.md workflow rewrite — 128→95 lines (-26%). Anti-patterns merged into Behavior Rules, Limits table replaced with config ref, Phase 4 renamed "Update" with 🛑 WAIT gate (TFW-26 post-review)
• .tfw/README.md stripped — 353→138 lines. Pure philosophy paper. Removed: project structure tree, artifact types, lifecycle, scope budgets, workflows table, execution modes, roles, Getting Started. All → conventions.md/glossary.md refs (TFW-27)
• knowledge.md Phase 2 — ⚠️ block with YES/NO examples of strategic vs technical knowledge (TFW-26 post-review)
• knowledge.md Behavior Rules — "DO NOT default all facts to existing categories" (TFW-26 post-review)
• KNOWLEDGE.md §4 — fact counts updated: 27→36, 4→6 topic files
• TECH_DEBT.md — TD-52, TD-69..74, TD-77, TD-78 resolved
• All adapters synced: .agent/workflows/ — config, resume, plan, init (TFW-27)

Removed

• Anti-patterns section from knowledge.md — merged into Behavior Rules (TFW-26)
• Limits table from knowledge.md — replaced with inline ref to PROJECT_CONFIG.yaml (TFW-26)
• .tfw/README.md §Evolution — replaced with CHANGELOG link (TFW-27)
• .tfw/README.md technical reference sections — duplicated from conventions/glossary (TFW-27)
• TFW-28 as standalone task — absorbed into TFW-27/C (TFW-27)

[0.7.1] — 2026-04-04

Added

• 3 new README Values — "Honesty Over Convincingness" (renamed from "Determinism and Safety"), "Structural Enforcement" (filesystem = state machine), "Naming Creates Behavior" (terminology > explanation). Total: 5→8 values (TFW-25)
• Design Rules subsection in conventions.md §11 — P10-P13 content compressed into 4 rules: token density, inline enforcement, DNA/library, progressive disclosure (TFW-25)
• philosophy category in RF.md and REVIEW.md templates' FC category list (TFW-25 post-review)
• D34 (Values consolidation) in KNOWLEDGE.md §1 (TFW-25)
• F7 (framework value count norms) in knowledge/convention.md (TFW-25 knowledge consolidation)

Changed

• KNOWLEDGE.md §0 — pruned 14→7 principles (P4/P6 obvious, P10-P13 → conventions, P14 → README Values) (TFW-25)
• KNOWLEDGE.md §3 Legacy — pruned 35→13 items (removed all pre-TFW-22 resolved entries) (TFW-25)
• KNOWLEDGE.md §4 Tech Stack — removed entirely (obvious from repo) (TFW-25)
• KNOWLEDGE.md — §5 Project Facts renumbered to §4 after Tech Stack removal (TFW-25)
• knowledge/convention.md — pruned 12→7 facts (6 self-evident facts removed, 1 added) (TFW-25)
• knowledge/process.md — pruned 10→5 facts (5 self-evident facts removed) (TFW-25)
• TECH_DEBT.md — pruned 64→19 items (removed all resolved/accepted/obsolete entries) (TFW-25 post-review)
• KNOWLEDGE.md template — §4 Tech Stack removed, §5→§4 renumbered (TFW-25 post-review)
• knowledge.md workflow — 3 stale §5 references updated to §4 (TFW-25 post-review)

Fixed

• TD-64: KNOWLEDGE.md template referenced ## 5. Project Facts instead of ## 4.

[0.7.0] — 2026-04-04

Added

• Researcher role — 4th standalone role (after Coordinator, Executor, Reviewer), extracted from Coordinator following TFW-8 pattern. Own 🔒 ROLE LOCK: RESEARCHER. Permitted: RES, research/ stage files. Forbidden: HL, TS, ONB, RF, REVIEW, code (TFW-24)
• Research subfolder state machine — research/ subfolder with stage files (briefing.md, gather.md, extract.md, challenge.md). File existence = stage completion. Crash-resilient, zero-parsing (TFW-24)
• Resume Protocol (Step 0) in research/base.md — check filesystem state → resume from first missing file. No chat history dependency (TFW-24)
• 4 research stage templates in .tfw/templates/research/ — briefing, gather, extract, challenge. Each with Parent HL link, Goal from §1 Vision, D28 guiding question subtitle, Checkpoint with Stage complete: YES/NO, Sufficiency checklist (TFW-24/B)
• HL §1 Working Backwards — Vision narrative ("write as if done"), Impact field, stakeholder-perspective Quote (Amazon press release pattern) (TFW-24)
• HL §10 "Why Not Just...?" — internal FAQ section forcing alternatives consideration before research (TFW-24)
• tfw.content_language config — controls artifact content language (default: en). Template structure always English (TFW-23)
• P14 (Filesystem = state machine) in KNOWLEDGE.md (TFW-24)
• D29 (English-only templates), D30-D33 (Researcher role, subfolder state machine, RES synthesis, Working Backwards) in KNOWLEDGE.md (TFW-23, TFW-24)

Changed

• BREAKING: All 5 core templates (HL, TS, RF, ONB, REVIEW) — pure English headings and field labels. 32 terms translated per D28. content_language note added (TFW-23)
• BREAKING: HL template §1 restructured — generic "Vision" → Vision narrative + Impact + Quote (TFW-24)
• BREAKING: HL template §2 "Current State" — domain-agnostic ("system/process/environment" not code-specific) (TFW-23/24)
• BREAKING: HL template §5 "Definition of Done" — domain-agnostic checklist items (TFW-23)
• BREAKING: RES template — stage sections removed. RES = synthesis format (Decisions, Hypotheses, HL Recommendations, Conclusion). Stages live in research/ subfolder (TFW-24)
• BREAKING: Coordinator no longer conducts research — hands off to Researcher via /tfw-research (TFW-24)
• research/base.md Steps 3/4/5 — reference templates/research/ for stage files (TFW-24/B)
• conventions.md §4 — inline stage format replaced with templates reference (TFW-24/B)
• conventions.md §8 — Researcher role in workflows table (TFW-24)
• conventions.md §15 — Researcher row in Role Lock table, research/base.md row updated (TFW-24)
• glossary.md — Researcher role definition, Coordinator updated (research duties removed) (TFW-24)
• plan.md Step 6 — Researcher handoff with STOP instruction (TFW-24)
• PROJECT_CONFIG.yaml — RES status role = researcher (TFW-24)
• init.md Step 5 — content_language in config generation (TFW-23)
• All adapters synced: .agent/workflows/, .claude/commands/ (TFW-23, TFW-24)

Removed

• "Coordinator (Research Mode)" overlay — replaced by standalone Researcher role (TFW-24)
• Stage sections in RES template (Gather/Extract/Challenge) — moved to research/ subfolder files (TFW-24)
• Inline stage file format in conventions.md §4 — replaced by template reference (TFW-24/B)
• Mixed RU/EN headings from all 5 templates (TFW-23)

[0.6.6] — 2026-04-04

Added

• Modular research architecture — research/{base,focused,deep}.md replaces monolithic research.md (TFW-22)
• base.md: core algorithm with OODA Stage Loop, Trust Protocol, Sufficiency Verdict (504 words)
• focused.md: single-pass mode, generic criteria only (106 words)
• deep.md: multi-loop hypothesis-driven mode with metacognitive check (171 words)
• OODA Stage Loop in research — Observe→Orient→Decide→Act with YAML-configurable loops_per_stage hard limit (TFW-22)
• Sufficiency Verdict — 2-level checkpoint criteria: generic (always) + mode-specific (from mode file). Criteria = SOFT (report, not block) (TFW-22)
• Trust Protocol — 4-tier trust levels for user input (business→trust, tech→verify, numbers→empirical, experience→trust outcome) (TFW-22)
• HL template §3.1 — Визуализация результата: ASCII mandatory, mermaid for complex flows, before→after tables (TFW-22)
• HL template §10 — Обоснование RESEARCH: hypotheses table with filter, blind spots, risks of not researching, proposed focus (TFW-22)
• RES template — Hypotheses table in Briefing (from HL §10), Sufficiency Verdict format in every stage checkpoint (TFW-22)
• Step 5: Hypothesis Iteration in plan.md — FOR EACH loop presenting §10 hypotheses to user before RESEARCH decision (TFW-22)
• tfw.research.default_mode and tfw.research.modes.{focused,deep} in PROJECT_CONFIG.yaml (TFW-22)
• 3 new Config Sync Registry entries for research mode settings (TFW-22)
• P12 (DNA/Library split), P13 (Progressive Disclosure) in KNOWLEDGE.md (TFW-22)
• D25-D28 (modular research, OODA loop, Trust Protocol, Naming > Explanation) in KNOWLEDGE.md (TFW-22)

Changed

• plan.md algorithm refactor — 1213→795 words (-34%). Inline bloat (prerequisites, scope budget table, status transitions, anti-patterns) replaced with ref-inside-step pattern. DNA layer inline (Role Lock + Mindset). RESEARCH Gate strengthened (TFW-22)
• PROJECT_CONFIG.yaml workflow path: research.md → research/base.md (TFW-22)
• config.md Adapter Sync — copy command updated to research/base.md (TFW-22)
• conventions.md — 3 stale research.md references updated to research/base.md (TD-54)
• CLAUDE.md, KNOWLEDGE.md — research workflow path references updated (TFW-22)
• All 4 adapters synced: tfw-plan.md (×2), tfw-research.md (×2) (TFW-22)

Removed

• Monolithic research.md (1165 words) — replaced by research/ directory (TFW-22)
• Inline bloat in plan.md: prerequisites list, scope budget table, status transitions diagram, anti-patterns block (~400 words) (TFW-22)

Fixed

• TD-54: conventions.md L29, 181, 276 — stale research.md paths updated to research/base.md
• TD-55: conventions.md L277 — handoff.md Role Lock table: code moved from Forbidden to Permitted Artifacts (executor writes code via handoff)

[0.6.5] — 2026-04-03

Added

• Human-Only Test in RF.md, REVIEW.md, RES.md templates — FC quality gate: "would this fact be unknown without the human saying it?" Rejects agent-discoverable facts (TFW-18B)
• Human-Only Test in knowledge.md Phase 3 Step 1 — consolidation-time reject criterion for agent-discoverable facts (TFW-18B)
• Quality bar in RF.md §5 Observations + handoff.md §Observations — "report only issues that would bite the next developer" (TFW-18B)
• Quality filter in review.md Step 3 — reject filler observations before promoting to TECH_DEBT.md (TFW-18B)
• Knowledge consolidation bullet in .tfw/README.md §v3 additions (TFW-18B)
• knowledge and config rows in .tfw/README.md §Canonical Workflows table (TFW-18B)

Changed

• FC prompt reframed from "next agent's behavior" to "strategic knowledge — domain patterns, stakeholder priorities, business context" in RF.md, REVIEW.md, RES.md templates (TFW-18B)
• FC prompt reframed in research.md §Closure and handoff.md §FC guidance (TFW-18B)
• conventions.md §10.1 category examples expanded: domain → revenue patterns/client segments, stakeholder → priorities/pain points/quotes, constraint → contractual obligations, context → market conditions/competitive landscape, risk → client concentration/knowledge silos (TFW-18B)
• knowledge.md Phase 2 gather guidance: "strategic knowledge" emphasis, redirects technical details to tfw-docs (TFW-18B)
• handoff.md FC guidance reordered: leads with "stakeholder priorities, domain patterns" instead of "environment, constraints" (TFW-18B)
• All adapters synced: .agent/workflows/ (4 files) + .claude/commands/ (3 files) (TFW-18B)

[0.6.4] — 2026-04-03

Added

• /tfw-config workflow — interactive config sync with edit/verify modes and Config Sync Registry (16 mapped entries across 3 categories) (TFW-19)
• Inline budget table (Pattern A) restored in plan.md §Scope Budget per Phase — 4-row compact table with defaults + config key (TFW-19)
• Inline budget table with Rationale column restored in conventions.md §6 (TFW-19)
• Inline limits table in knowledge.md §Limits — 4-row compact table (interval, gate_mode, max_facts, max_topics) (TFW-19)
• Budget Check enforcement hook in plan.md Phase 5 — mandatory check before writing TS (TFW-19)
• Multi-phase subfolder convention in conventions.md §4 — master artifacts at root, phase artifacts in PhaseA/, PhaseB/ subfolders (TFW-19)
• Config Sync Registry term in glossary.md (TFW-19)
• config.md listed in conventions.md §8 Workflows and §15 Role Lock (TFW-19)
• Antigravity adapter tfw-config.md (TFW-19)

Changed

• TS.md template L27 — budget line now shows inline defaults format instead of «see config» (TFW-19)
• research.md §Limits — restored standard 2-line defaults header (TFW-19)
• All adapters synced: tfw-plan.md, tfw-research.md, tfw-knowledge.md, tfw-config.md (TFW-19)

Deprecated

• D17 (Pattern B pure reference) superseded by D24 (Pattern A + Config Sync Registry) (TFW-19)

Removed

• Naming Rules table from plan.md (~100 words) — already in conventions.md §4 (TD-48 resolved) (TFW-19)

Fixed

• Agent enforcement of scope budgets — Pattern B «see config» broke compliance, restored inline values (TFW-19)

[0.6.3] — 2026-04-03

Added

• Conversation history scan instruction in knowledge.md Phase 2: Gather — consolidator MUST review chat history, not just artifact Fact Candidates (was never present — root cause of missed chat facts)
• Conversation history scan instruction in RF.md, REVIEW.md, RES.md templates — agents see templates during writing, not workflow files

Fixed

• research.md — restored conversation history scan instruction lost during TFW-21 compression (v0.6.2)

[0.6.2] — 2026-04-03

Changed

• research.md — compressed from 2397→1145 words (-52%), 319→160 lines (-50%) (TFW-21)
• Removed: Example Flow (45 lines), "Good/Bad research" + "Operational" sections, duplicate Anti-patterns block
• Removed: Inline checkpoint/sufficiency templates → reference templates/RES.md
• Preserved: Research Mindset, 3 stages with mindset reminders, Briefing Protocol, Closure Protocol, all 8 Hard Rules
• Merged: Hard Rules + Anti-patterns → single Rules section (MUST/NEVER format)
• RES.md template — enhanced stage checkpoints with Agent assessment, Depth check, Recommendation fields; added external research line to Sufficiency Check (TFW-21)
• Adapter copy synced: .agent/workflows/tfw-research.md (TFW-21)

[0.6.1] — 2026-04-03

Added

• Chat history scan instruction in handoff.md, research.md, review.md — agents MUST review conversation history before writing Fact Candidates. Human messages are the primary source of project knowledge (TFW-18 post-release finding)

Changed

• All 3 adapter copies synced (tfw-handoff, tfw-research, tfw-review)

[0.6.0] — 2026-04-03

Added

• /tfw-knowledge workflow — 4-phase consolidation (Orient → Gather → Consolidate → Prune) with role lock, behavior rules, anti-patterns (TFW-18)
• TOPIC_FILE.md template — per-category knowledge files in knowledge/ folder (TFW-18)
• .tfw/knowledge_state.yaml — consolidation state tracking (seq, date, stats) (TFW-18)
• tfw.knowledge section in PROJECT_CONFIG.yaml — 6 configurable parameters: interval, gate_mode, max_index_lines, max_index_facts_lines, max_facts_per_topic, max_topic_files (TFW-18)
• §6 Fact Candidates in RF.md template — mandatory section with quality filter and anti-patterns (TFW-18)
• §5 Fact Candidates in REVIEW.md template — mandatory section with quality filter (TFW-18)
• Fact Candidates in RES.md template Closure section (TFW-18)
• Phase 0: Knowledge Gate Check in plan.md — configurable (hard/soft/off) enforcement before Phase 1 (TFW-18)
• 💡 Mindset reminders in handoff.md, research.md, review.md — capture project facts (TFW-18)
• §5 Project Facts compact index in KNOWLEDGE.md template — category/count/link table (TFW-18)
• Item 6 in docs.md checklist — Fact Candidates marker (TFW-18)
• §10.1 Fact Categories and §10.2 Knowledge Infrastructure in conventions.md (TFW-18)
• 4 glossary terms: Fact Candidate, Topic File, Knowledge Gate, Consolidation (TFW-18)
• D22 in KNOWLEDGE.md — knowledge consolidation decision (TFW-18)
• /tfw-knowledge Antigravity adapter (TFW-18)
• .user_preferences.md guidance in init.md Step 5 (TFW-18)

Changed

• All 5 existing adapter copies synced with canonical workflows (TFW-18)

[0.5.5] — 2026-04-03

Added

• Coordinator Mindset section in plan.md — quality of planning > speed of pipeline, anti-rush guidance, RESEARCH as default (TFW-17)
• Hard Rule #8 in research.md — every stage MUST include at least one external action (web search, URL read, docs) (TFW-17)
• Stage-level mindset reminders in research.md — 1-line blockquote at the start of Gather, Extract, Challenge (TFW-17)
• Depth self-check in research.md checkpoint template — "Did I use external sources, or only project files?" (TFW-17)
• External research bullet in Sufficiency Check — "Did every stage include external research?" (TFW-17)
• D21 in KNOWLEDGE.md — dual-lever fix for coordinator rush-bias + research depth (TFW-17)
• P9 in KNOWLEDGE.md — Coordinator Mindset principle (TFW-17)

Changed

• plan.md Phase 1 — "Understand the problem" → "Understand the problem deeply" with anti-rush guidance (TFW-17)
• plan.md RESEARCH Gate — coordinator must be specific about what RESEARCH could reveal, frame as risk reduction (TFW-17)
• research.md Gather stage — "Autonomous search" replaced with "Search externally: how is this problem solved elsewhere?" (TFW-17)
• P8 in KNOWLEDGE.md — updated to include external tool mandate reference (TFW-17)
• All 4 adapter copies synced — .agent/workflows/tfw-plan.md, .agent/workflows/tfw-research.md, .claude/commands/tfw-plan.md, .claude/commands/tfw-research.md (TFW-17)

Fixed

• TD-34: research.md L26 no longer references TS as primary output (confirmed resolved by TFW-14, verified TFW-17)
• Adapter desync: .agent/workflows/tfw-plan.md and .claude/commands/tfw-plan.md had stale 🔵 HL statuses, Phase 3.5 numbering, old pipeline diagram — all fixed via full copy from canonical

[0.5.4] — 2026-04-01

Added

• tfw.statuses registry in PROJECT_CONFIG.yaml — 9 status entries with role field (TFW-15)
• Concept Taxonomy in glossary.md — 5 formal definitions: Document Type, Template, Workflow, Adapter Command, Status (TFW-15)
• REJECT branching in conventions.md — user decides: HL_DRAFT / RES / TS_DRAFT (TFW-15)
• D20 in KNOWLEDGE.md — pipeline status decoupling decision (TFW-15)

Changed

• BREAKING: Pipeline statuses renamed: 🔵 HL → 📝 HL_DRAFT, 🟡 TS → 🟡 TS_DRAFT across all .tfw/ files (TFW-15)
• BREAKING: HL template status label: 🔵 HL — Ожидает ревью → 📝 HL_DRAFT — Ожидает ревью (TFW-15)
• BREAKING: TS template status label: 🟡 TS — Ожидает апрува → 🟡 TS_DRAFT — Ожидает апрува (TFW-15)
• plan.md — Phase 3.5 → Phase 4 (RESEARCH Gate), Phase 4 → Phase 5 (Decide Scope & Write TS), step numbering gap fixed (TFW-15)
• research.md — Status Transitions section updated to HL_DRAFT/TS_DRAFT (TFW-15)
• conventions.md — status table, pipeline diagram, REJECT verdict updated (TFW-15)
• glossary.md — Status Flow diagram updated (TFW-15)
• .tfw/README.md — Task Lifecycle pipeline diagram and REJECT wording updated (TFW-15)

Deprecated

• 🔵 HL and 🟡 TS status names — replaced by 📝 HL_DRAFT and 🟡 TS_DRAFT
• Phase 3.5 numbering in plan.md — replaced by clean Phase 4/5 numbering

[0.5.3] — 2026-04-01

Added

• Briefing Protocol in research.md — mandatory entry with research plan, scope intent, guiding questions before stages (TFW-14)
• Closure Protocol in research.md — mandatory exit with HL update recommendations after sufficiency check (TFW-14)
• Briefing and Closure sections in RES.md template — structural anchors for agent behavior (TFW-14)
• 3 new Hard Rules in research.md — briefing mandatory, closure mandatory, sufficiency check with specifics (TFW-14)
• 4 new Anti-patterns — skip-briefing, rush-bias, silent closure, skip-bias (TFW-14)
• HL update gate in plan.md Phase 3.5 — coordinator reads RES → updates HL → user confirms → TS (TFW-14)
• D19 in KNOWLEDGE.md — HL update = mandatory RESEARCH output (TFW-14)

Changed

• Checkpoint in research.md — extended with Stage Handoff (plan for next stage + question) (TFW-14)
• Final Checkpoint — Complexity Check replaced by Sufficiency Check ("sufficient for HL finalization?") (TFW-14)
• Turn-based rhythm — questions limit changed from "per stage" to "per turn" (≤3) across research.md, Limits table, Hard Rules, Anti-patterns (TFW-14)
• plan.md Phase 3.5 — skip-bias fix: pros/cons format, default=recommend research, user decides (TFW-14)
• Both adapters (.claude/commands/tfw-research.md, .agent/workflows/tfw-research.md) — synced with Briefing→Stages→Closure structure (TFW-14)
• Research Mindset L26 — reworded from "details needed for TS" to "refines the HL" (TFW-14 REVISE)

Fixed

• TD-34: research.md L26 referenced TS as primary output after Closure Protocol addition — now references HL

[0.5.2] — 2026-03-31

Added

• init.md workflow — AI-first project initialization (Discover → Interview → Knowledge → Setup → Verify) (TFW-13)
• /tfw-init slash command (Claude Code + Antigravity) (TFW-13)
• .tfw/adapters/README.md — adapter index + "How to Write a New Adapter" (moved from old init.md) (TFW-13)
• docs.md, release.md, update.md in conventions §15 Role Lock table (consistency fix)
• research.md in conventions §8 Workflows table (consistency fix)
• VERSION, CHANGELOG.md in conventions §2 Required Artifacts (consistency fix)

Changed

• .tfw/init.md — replaced 232-line manual guide with 20-line pointer to workflow (TFW-13)
• Antigravity README — all 9 workflows in copy/sync instructions (was 5) (consistency fix, TD-27)
• plan.md Role Lock — removed REVIEW from permitted artifacts (was inconsistent with §15 table)
• conventions §2 — all 9 workflows now listed (was 5)
• conventions §8 — reordered: init first, added research

Fixed

• TD-27: Antigravity README missing 4 workflows in copy commands
• TD-29: conventions §2 missing review, docs, release, update workflows + VERSION, CHANGELOG
• TD-30: conventions §8 missing research.md
• TD-31: conventions §15 missing docs, release, update in Role Lock
• TD-32: Antigravity README copy/sync missing research, docs, release, update
• plan.md declared "Permitted: HL, TS, REVIEW" but §15 table said "HL, TS" — fixed to match table

[0.5.1] — 2026-03-30

Added

• tfw.scope_budgets section in PROJECT_CONFIG.yaml — 4 configurable budget values (TFW-12)
• tfw.workflows section in PROJECT_CONFIG.yaml — 8 workflow entries (TFW-12)
• tfw.research section in PROJECT_CONFIG.yaml — 4 research limit entries (TFW-12)
• Config component row in KNOWLEDGE.md Architecture Map (TFW-12)

Changed

• tfw.templates in PROJECT_CONFIG.yaml — completed to 8 entries (+res, +knowledge, +release) (TFW-12)
• Scope budget values removed from docs — pure reference to tfw.scope_budgets config (TFW-12, Pattern B)
• Version strings removed from core file titles (conventions.md, glossary.md) — avoids drift on bump (TFW-12)
• Adapter templates use {version} placeholder instead of hardcoded version (TFW-12)
• CLAUDE.md, .agent/rules/tfw.md — version and template/workflow references centralized (TFW-12)
• init.md — full config example with all 4 sections, {version} replacement instructions (TFW-12)

Fixed

• CHANGELOG.md — restored missing [0.4.2] section header
• TD-25: conventions.md/glossary.md title headers fixed (no more stale version)
• TD-26: .agent/rules/tfw.md — added version reference and RES template

[0.5.0] — 2026-03-30

Added

• RESEARCH stage — optional structured investigation between HL and TS (TFW-11)
• RES.md template — Research Report artifact
• research.md workflow — standalone and pipeline research
• Phase 3.5 RESEARCH gate in plan.md
• 🔬 RES status — pipeline now 8-status (RES optional)
• Read-only AG mode definition in glossary
• RES in Role Lock Protocol (conventions §15)
• Claude Code adapter: CLAUDE.md, 9 slash commands in .claude/commands/
• Claude Code adapter: README.md setup guide
• /tfw-research slash command (Claude Code + Antigravity)
• /tfw-review slash command (Claude Code)
• /tfw-release slash command (Claude Code)
• /tfw-update slash command (Claude Code)

Changed

• Pipeline diagrams updated in all core files (8-status, RES optional)
• Coordinator role updated: conducts RESEARCH, writes RES files
• All 3 adapter templates updated (RES, full workflow/command lists)
• CLAUDE.md.template expanded with slash command table and full context loading
• Antigravity adapter copies synced (plan, research, handoff)
• init.md — RES template in config, research.md in workflow copy commands
• .tfw/README.md — project structure tree updated

[0.4.2] — 2026-03-12

Added

• tfw.upstream field in PROJECT_CONFIG.yaml — configurable source URL for tfw-update (TFW-9)
• Step 0 (Fetch Upstream) and Step 9 (Cleanup) in update.md — concrete fetch mechanism with cross-platform commands (TFW-9)
• .tfw/.upstream/ staging directory pattern — OS-independent, gitignored (TFW-9)

Changed

• update.md — all vague "upstream" references replaced with concrete .tfw/.upstream/.tfw/ paths (TFW-9)
• conventions.md §8, .tfw/README.md — update workflow description includes "Fetch upstream" step (TD-17, TD-18)
• init.md — tfw.upstream in config example, .tfw/.upstream/ gitignore note (TFW-9)
• glossary.md — tfw-update entry expanded with source resolution details (TFW-9)

[0.4.1] — 2026-03-12

Added

• review.md workflow — standalone review process with 🔒 ROLE LOCK: REVIEWER (TFW-8)
• Reviewer role — coordinator in review-locked mode (glossary, conventions) (TFW-8)
• Executor Hard Stop Rule in conventions §15 (TFW-8)

Changed

• handoff.md — removed Phase 4 (review), added executor STOP block (TFW-8)
• conventions.md — Role Lock table updated, "any role" for REVIEW removed, review.md row added (TFW-8)
• glossary.md — Coordinator role updated (review duties moved to Reviewer) (TFW-8)
• AGENTS.md — workflow list updated with review.md (TFW-8)
• README.md (.tfw/) — workflows table, roles section, evolution updated (TFW-8)
• plan.md, resume.md — review workflow references added (TFW-8)
• init.md, adapter README — review workflow in setup instructions (TFW-8)

Removed

• Review phase from handoff.md — moved to standalone review.md (TFW-8)
• "REVIEW files can be written by any role" from conventions §15 (TFW-8)

Fixed

• conventions.md §8 — docs.md workflow now listed in Workflows table (TFW-7)
• .tfw/README.md — workflow count corrected, docs workflow included (TFW-7)
• Cross-references between conventions, glossary, and README aligned (TFW-7)

[0.4.0] — 2026-03-12

Added

• VERSION file — machine-readable framework version
• CHANGELOG.md — version history (this file)
• RELEASE.md template — optional release context artifact
• tfw-release workflow — canonical release process
• tfw-update workflow — structured upgrade process for downstream projects

Changed

• PROJECT_CONFIG.yaml — added tfw.version field
• init.md — added version tracking and RELEASE.md guidance

[0.3.0] — 2026-03-02

Added

• KNOWLEDGE.md template and tfw-docs workflow (TFW-5)

Changed

• Framework cleanup: removed STEPS.md, TASK.md, Summary Discipline (TFW-4)
• Root README restructured for public readership (TFW-3)

[0.2.0] — 2026-02-25

Added

• .tfw/ directory — tool-agnostic core (conventions, templates, workflows, adapters)
• ONB and REVIEW artifact types
• 7-status lifecycle with quality gates
• 3 canonical workflows (plan, handoff, resume)
• Scope budgets per phase
• TECH_DEBT.md pipeline
• Tool adapter pattern (Claude Code, Cursor, Antigravity)
• PROJECT_CONFIG.yaml
• Anti-patterns list

Removed

• AI_ENTRY_POINT.md, SUCCESS_CRITERIA.md, 00_meta/ directory

[0.1.0] — 2024

Added

• Core concept: traces are more valuable than code
• 4-file structure (AGENTS, README, TASK, STEPS)
• Summary Discipline
• Chat→project conversion pattern
• CL/AG execution modes (informal)