Skip to content

TS — TFW-60 / Phase AC: Update Without Guesswork

Date: 2026-08-30 Author: Claude Code (Coordinator), on_behalf_of: saubakirov Status: ✅ APPROVED — owner, 2026-08-30, at revision 1. Execution authorized via /tfw-handoff Revision 2: after onboarding, 2026-08-30. Two blocking questions answered, nine rulings on the executor's recommendations, risks and inconsistencies; changes carry R2. Census 23 → 26 counted. One question touched frozen text (the briefing's inputs); filed as amendment A7 and approved by the owner the same day — the briefing reads Changed as a fourth block. Phase HL: HL Phase AC — approved by owner 2026-08-30 Master freeze: 56c3d70 — baseline after amendment A6 Origin: fourth and fifth field reports, both on v2.0.0-dirty.4 Predecessor read through the Pre-TS Gate: Phase AB RF revision 2 and REVIEW revision 2 (APPROVE, 4846f27). AB delivered the whole-or-refuse dispatcher for identifiers, computed manifest guarantees, the current grammar, the test split and the source pin in update.md. It did not touch the status cell — classify_status() still takes the first [A-Z_]+ token — and its pin is written from HEAD. Both are this phase's work, not AB's failure.

1. Objective

Remove every place where the update path guesses or decides for the owner. After this phase: the pin is derived from the tag the operator names; the update asks who is acting, where tasks go and how the project verifies before it writes anything, and ends by telling the owner what changed in their own language; the migration refuses a status cell it cannot read whole and names every phase directory it left without state; the payload copy cannot overwrite the receiver's own configuration; every instruction in the path can be executed as written by a receiver on any earlier tag. A task's abbreviation is the initials of a title a person can read.

2. Scope

In scope: update.md Steps −1, 0, 3, 5, 6, 7 and a new briefing step; the abbreviation rule in plan.md, init.md, conventions.md, glossary.md and the HL template; the status-cell classifier and phase reporting in migrate_board.py; the missing-phase-state report in gen_index.py --check tasks; the CHANGELOG's reach across skipped tags, its retracted-wording quote and its dead references; the RELEASE.md rules that make those permanent; the Claude rules markers; the Antigravity template; the event, profile, status and project-config templates; the migration guide; tests for every behaviour change.

Out of scope: the identifier grammar; status.md keys, the event schema, the lifecycle vocabulary, the index format; renaming or rewriting any task directory, event or already-migrated status.md; hand-authoring phase state for other tasks in this repository; the four consumer projects; the actor field (TFW-54); transport mode (TFW-61); Phases B and C; claiming 2.0.0; the fifth report's §6 minor items — since semantics, created seconds provenance, a README route template, a stale index in a non-first container, --check project before migration — to be recorded in RF §6 Observations for the reviewer to file.

Decisions the owner has already settled

Decision Ruling
Phase AC exists, after AB and before B; fourth-report defects 3–6, 8 and the filed debt are its scope 2026-08-30, in chat; amendment A6
Fifth-report items 6–8 and 10 stay in the phase 2026-08-30, HL gate: approved as written
2.0.0 is not ready; this phase ships as the next 2.0.0-dirty tag 2026-08-30, superseding 2026-08-29
The abbreviation is the initials of the approved title, proposed with the title 2026-08-30

3. Principles Check

Principle (master HL §7) Enforced by
P1 Pain before mechanism AC-6, AC-7, AC-8 — each fixes a shape measured on a real project, reproduced as a fixture
P4 Stable paths over status moves AC-8 — nothing is renamed; UNDECLARED is a value a person resolves with an event
P5 Local truth, derived views AC-8 — the manifest and the gate name what they did not write instead of implying it
P9 No trace deletion during simplification AC-2 — the CHANGELOG quotes retired wording; it never deletes the entry it retracts
P10 Every phase pays for its release surface AC-11 — the next dirty tag with a consumer run, 2.0.0 unclaimed

4. Affected Files

Byte-identical adapter copies are excluded from the count by S32; work artifacts by S46. The executor measures this table at onboarding and reports the measured census in the RF.

Group Paths Counted
Update workflow .tfw/workflows/update.md — Step −1 (new), Step 0 pin at :17–20 and :86–87, Step 3 gate at :64–72, Step 5 exclusions, Step 6 table at :99–107 and allowlist at :111–115, Step 7 at :119, Step 8a briefing (new) 1
Abbreviation rule .tfw/workflows/plan.md:38–42, :55–68 · .tfw/workflows/init.md:113–114, :136–143 · .tfw/conventions.md:229–262 (Identifier), :369–398 (Artifact file naming) · .tfw/glossary.md:120 · .tfw/templates/HL.md:3–5 5
Migration honesty .tfw/scripts/migrate_board.py — classify_status():162–181, plan():865 terminal skip, render_manifest():750–752 · .tfw/scripts/gen_index.py — check_tasks():1186–1221, read_phase_status():474 2
Tests .tfw/scripts/test_migrate_board.py · .tfw/scripts/test_gen_index.py · docs/scripts/test_integration.py (markers and exclusion list, if the executor places them there) 3
Release surface .tfw/CHANGELOG.md — .3 entry (TD-198), :297 (TD-191), :200–265 fence (TD-190), updating sections · RELEASE.md §5–6 2
Adapters .tfw/adapters/claude-code/CLAUDE.md.template (markers; block brought current) · .tfw/adapters/claude-code/README.md · .tfw/adapters/antigravity/tfw-rules.md.template:5–7 · .agent/rules/tfw.md:5–8 (TD-204) · R2 .tfw/adapters/cursor/tfw.mdc.template:2, :6, :8 (same {version} defect) · R2 .tfw/adapters/codex/README.md:104 (first-run sentence → pointer) 6
R2 Marker rule, canon .tfw/conventions.md §9 Tool Adapter Pattern — the one place the marker-bounded block rule is stated (already in the census for §4; no new path) 0
R2 The framework as its own consumer CLAUDE.md (this repository's root) — carries the template's managed block between markers, byte-identical, checked by the installed-copies test 1
Templates .tfw/templates/journal/event.md:49, :70 (TD-200) · .tfw/templates/team/profile.md:12–14, :27–33 (TD-203, role location) · .tfw/templates/status.md (phase paragraph) · .tfw/templates/project_config.yaml:17 (installed_from form) 4
Briefing .tfw/templates/briefing.md — new 1
Migration guide .tfw/migrations/2.0.0.md — manifest location at :94, cd .tfw at :61, --working-tree, phase state hand-authored, three grammars unchanged 1
Copies, excluded .claude/commands/tfw-{update,plan,init}.md · .agent/workflows/tfw-{update,plan,init}.md — re-synced byte-identical in the same commit as their source 0

Census at TS: 29 physical paths, 23 counted, 1 new. R2, measured at onboarding and extended by the rulings below: 32 physical paths, 26 counted, 1 new. Configured budget since f3eb986 (owner, 2026-08-30): 50 / 50 / 5000 / 50. The phase is also inside the 30 / 15 / 3000 / 30 that governed A–AB; no ruling is needed under either. Any group appearing that this table does not name returns to the coordinator before the work proceeds.

5. Acceptance Criteria

AC-1: The pin is derived from the tag, not from HEAD

Fourth report §6. update.md:17–20 reads source_head=$(git rev-parse HEAD) and tests tag_commit = source_head; on a source whose development has moved past its release the test can never pass, and the fifth report passed it only because the source happened to stand on the tag.

  • [ ] the operator names the target — a tag, or a commit when the owner deliberately takes an untagged payload and says so in the checklist; source_head is derived from it (git rev-parse "$target_ref^{commit}"), never from HEAD
  • [ ] VERSION is read from that commit and compared with the tag's name; a mismatch stops
  • [ ] the Step 5 recheck compares against the same derived commit, so a source that moved elsewhere during the update is still detected and a source that was ahead all along is not reported as movement
  • [ ] the local-source cleanliness rule (status --porcelain -- .tfw/) stays as written

Gate: on this repository with HEAD ahead of v2.0.0-dirty.4, Step 0 as written passes for that tag; with a fabricated tag whose VERSION disagrees with its name, it stops. Evidence: both runs in the gates file, commands verbatim.

AC-2: A receiver on any earlier tag finds its path

Fourth report §2–3; TD-190, TD-191, TD-198.

  • [ ] the new entry's updating section names or points to every intervening entry's updating section ("if you are on .2, also perform the .3 section" is sufficient); RELEASE.md §6 makes this a step of every release
  • [ ] RELEASE.md §5 adds: when a release reverses a normative statement, the CHANGELOG quotes the retired wording verbatim as a search string. The .3 entry gains the quote it lacks — "Commands never duplicate workflow content — they reference it" — and what a project that already went thin does about it (TD-198)
  • [ ] the 2.0.0-dirty.2 entry's [TD-11](../../../reference/tech-debt.md) reference (:297) is corrected or removed (TD-191)
  • [ ] the 2.0.0-dirty migration fence (:200–265) carries a superseded by line pointing at the guide, and RELEASE.md names that marker as the form for any instruction a later release replaces (TD-190)
  • [ ] update.md opens with Step −1: read the target's .tfw/workflows/update.md from the pinned payload and follow it, not the installed one; the migration guide's "Updating from 1.x" position repeats it as the first instruction, since an installed 1.x workflow cannot carry it. R2: the new entry's updating section opens with the same line — a .2–.4 receiver's installed workflow lacks Step −1 too, and the entry is the first file the release tells it to open
  • [ ] no CHANGELOG entry is rewritten in substance; additions are appended to the entries they concern, dated

Gate: a reader on .2 follows the new entry to every instruction the .3 and .4 sections give, without being told which of them to open by anyone but the CHANGELOG. Evidence: the read-through recorded step by step in the gates file.

AC-3: The retired-vocabulary gate can be literally green

Third and fourth reports; six hits in update.md:68, init.md:123 and their copies.

  • [ ] the allowlist at update.md:111–115 admits text whose purpose is to retire the term — a retirement instruction anywhere in the payload — and says so in one sentence
  • [ ] the six known hits fall inside the allowlist under that wording; the wording does not admit a live use of a retired term

Gate: the check on this repository and on one consumer returns zero outside the allowlist, with the allowlist printed. Evidence: both runs.

AC-4: Every Step 6 row is executable the same way

Fourth report §4; TD-204.

  • [ ] .tfw/adapters/claude-code/CLAUDE.md.template gains <!-- TFW:CLAUDE:START --> / <!-- TFW:CLAUDE:END --> markers bounding the managed content, on the Codex pattern
  • [ ] the Step 6 table states per row whether the target is a whole copy verified by cmp or a marker-bounded block in a project-owned file verified on the region between the markers; the Claude rules and Codex routing rows are blocks, the rest are copies
  • [ ] R2 — one marker rule, stated once. conventions.md §9 Tool Adapter Pattern carries the rule for every marker-bounded row, three cases: markers present → replace the text between them; file absent → create it from the template; file present without markers → report and leave it, the operator inserts the block once and it is mechanical from then on. Step 6 and both adapter READMEs point to §9. The Codex README's "If it has no markers, append the complete block" (:104) becomes that pointer: appending to a file that already has an unmarked hand-written TFW section produces two sections — the fourth report's CLAUDE.md
  • [ ] R2 — the block is brought current with the canon it describes. The template's Context Loading names status.md as the task's only authority (conventions §10 order) and the command table carries every canonical workflow, /tfw-knowledge and /tfw-config included; {version} inside the block reads see .tfw/VERSION. Bounds: the ## TFW section — context loading, commands, templates, key references. Project identity, mandatory rules, execution mode and code standards stay outside as project-owned text
  • [ ] R2 — the framework is its own first consumer. This repository's root CLAUDE.md carries the block between markers, byte-identical to the template's block, and the installed-copies test checks it as it checks every other installed copy. Conduct and Execution Modes stay outside the block as this project's text
  • [ ] R2 — Cursor. .tfw/adapters/cursor/tfw.mdc.template carries {version} three times, the TD-204 defect one adapter over; same fix, and the Cursor row is a whole copy
  • [ ] .tfw/adapters/antigravity/tfw-rules.md.template stops requiring {version}: it reads .tfw/VERSION the way the rendered .agent/rules/tfw.md already does, and the two agree (TD-204)
  • [ ] .tfw/adapters/claude-code/README.md describes the block and the first-run rule

Gate: cmp of the marker-bounded region on a fixture CLAUDE.md carrying project text above and below the block, before and after sync; project text unchanged. Evidence: the fixture and both regions.

AC-5: installed_from has one form

Fourth report §8.

  • [ ] update.md Step 7 and templates/project_config.yaml:17 state the form: {upstream}@{verified-tag}, where {upstream} is the value of tfw.upstream as configured — a URL or a symbolic name — and never a machine-local path; when the source is a local checkout, tfw.upstream names it symbolically and the checklist records the path
  • [ ] --check project reports an installed_from that carries a drive letter, a leading / or a backslash as machine-local; record the upstream reference; it does not rewrite it
  • [ ] self remains valid for this repository

Gate: --check project on a fixture config with D:/…@v2.0.0-dirty.4; then with [email protected]. Evidence: both outputs.

AC-6: The owner is asked, not guessed

Fifth report §2 and the owner's account quoted there.

  • [ ] update.md Step 3 gains a 🛑 gate before the first durable project write, with 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 same three go out as one message and the run proceeds through read-only steps and stops at the write
  • [ ] the answers are recorded in the update checklist
  • [ ] after Step 8 a new Step 8a writes the briefing to the owner in content_language, from .tfw/templates/briefing.md: three blocks — what is now possible (from Added), what stopped breaking (from Fixed), what no longer has to be done (from Removed) — derived from the intervening CHANGELOG entries, no free text of its own. R2: an absent section yields nothing in this release for its block, never invented content. A7 approved by the owner 2026-08-30: Changed is read too, as a fourth block — what you now do differently — bound to the entry's own bullets, no free text. The sequencing constraint is lifted; AC-6 is built like any other criterion
  • [ ] the briefing is the update's last message, and the checklist records that it was delivered
  • [ ] update.md stays under 1200 words with Steps −1, 3, 5, 8a added; duplication is removed before content is

Gate: an AG-mode dry run against a fixture project stops with the three questions before any write; the word count. Evidence: the transcript of that stop and the rendered briefing for this repository's .3 → .5 delta.

AC-7: The payload copy cannot overwrite project-owned files

Fifth report §4.

  • [ ] Step 5 names the exclusions: .tfw/project_config.yaml and .tfw/knowledge_state.yaml are never overwritten by the copy; the config is merged key by key as Step 3 already describes; the state file is never touched
  • [ ] the step that copies prints what it skipped; a copy that reports nothing skipped on a project that has both files is a failed step
  • [ ] executor decision, recorded in the RF: whether the payload stops carrying knowledge_state.yaml altogether (it is ⚫ project state by Step 3's own classification) or the exclusion list carries it. Either satisfies the criterion; the reason is what the RF records
  • [ ] the payload path test (test_integration.py:510) covers the exclusion list, so a new project-owned file added to the payload without an exclusion fails a test
  • [ ] R2 — decided at onboarding: the exclusion list carries both files; the payload keeps carrying them. A root .gitattributes export-ignore would govern git archive alone and add a root file outside the census. The RF records the root cause as an observation for the payload boundary (Phase AA's surface): the payload carries this repository's own project_config.yaml and knowledge_state.yaml, files conventions §10.3 says are never sourced from upstream

Gate: a fixture with a customized project_config.yaml and a non-framework knowledge_state.yaml; after the copy both are byte-identical and both are printed as skipped. Evidence: the fixture, the copy output, cmp on both files.

AC-8: The status cell is parsed whole or refused, and no phase is left without a named state

Fifth report §1. classify_status():162–181 takes the first [A-Z_]+ token; ✅ DONE (A/V/B/C) · 🔄 Phase D … R8 🟢 RF … classified DONE, and plan():865 wrote nothing for a live task.

  • [ ] a status cell is one declared lifecycle token followed by free text that contains no further declared token and no further status symbol. Anything else is UNDECLARED with lifecycle_verbatim carrying the cell, bounded as today. R2 — a status symbol is a character of Unicode category So (other symbols: ✅ ❌ 🔄 🟢 🟡 ⬜ …), measured on four corpora at onboarding: 8 rows change, every one carrying a second emoji or a second declared token. Sm, Sc, Sk (+ → = <) are prose punctuation on every board measured and are not signals — under category S entire, three single-signal rows ([TFW-52](../../TFW-52__tfw_light_v1/HL-TFW-52__tfw_light_v1.md), HD-30, live HD-31) would be refused on a plus sign or an arrow. U+FE0F and U+200D are skipped explicitly. The fixture tests a second emoji alone, a second token alone, → and + as non-signals, and a bare variation selector after the token
  • [ ] an UNDECLARED row is never classified terminal and never skipped by plan(); it receives a status.md at UNDECLARED, exactly as a single unknown token does today
  • [ ] the manifest lists such rows under their own heading — Rows carrying more than one lifecycle signal — and the Task state written note stops implying the skipped rows are all terminal
  • [ ] the manifest names every phase-* directory of every matched task and states in one sentence that phase state is not written by migration; author {phase}/status.md by hand
  • [ ] --check tasks reports a phase directory that carries no status.md: a failure when the task's own status.md is non-terminal; an informational line when the task is terminal or carries no status.md at all — a terminal legacy task migrated without state, by design, and a phase closed before phase state existed is history, not a defect. On this repository the measured set is 17 directories under TFW-42, 46, 47, 52, 53 and 55; all six tasks are terminal by the board snapshot and carry no task-level state, so all 17 are informational. It writes nothing. R2: informational lines are grouped one line per task, naming that task's stateless phase directories — six lines on this repository, not seventeen; exit code unchanged. When the task's own status.md is malformed, its stateless phases are informational — the malformed state is already the failure — and the line says so
  • [ ] templates/status.md carries the phase paragraph a phase file already uses in this repository ("…this phase's live state. The task-level status.md never summarizes it.") and says when to use which
  • [ ] the migration guide says all of the above in the order a receiver meets it
  • [ ] the AILAB-2 row shape is a committed fixture, failing before the change and passing after

Gate: the fixture; --check tasks on a fixture task with a phase directory lacking state; and on this repository — green, with 17 informational lines and no failure. Evidence: the three runs, plus the four pinned corpora compared identifier by identifier before and after — no classification other than the multi-signal rows changes.

AC-9: The abbreviation is the initials of the approved title

Owner ruling 2026-08-30; TD-201.

  • [ ] conventions.md §4 Identifier: ABBR is the acronym of the approved full title — the initials of its significant words, uppercase alphanumeric — proposed together with the title in one exchange and approved by the owner with it. Never derived silently is rewritten to say both things it means: never invented apart from the title, never created without approval
  • [ ] plan.md Step 3.5 and Step 4.2, init.md Batch 1 and Mini-Setup 6, glossary.md Task Naming say the same in their own words; the worked examples read as title → initials (Conflict Resistant Shared Workspace → CRSW; Assisted 1.5 core and synchronization → ASSISTED15 is admissible, since digits are alphanumeric)
  • [ ] the HL template header carries Title and Abbreviation as adjacent fields, in that order
  • [ ] conventions.md §4 Artifact file naming gains the current-grammar rows with an example: HL-TFW_20260829-172110_ABT.md, RES__TFW_20260829-172110_ABT.md — no title appended, for the reason the clock-task rule already gives (TD-201)

Gate: read plan.md, init.md, conventions.md, glossary.md, templates/HL.md: one rule, four carriers, no contradiction; gen_docs.py resolves the example filename. Evidence: the five excerpts side by side in the EV.

AC-10: Payload carriers agree with the canon and with each other

TD-200, TD-203; fifth report §5–6.

  • [ ] templates/journal/event.md:49 and :70 describe via as conventions.md §4 does — free-form non-empty provider or tool text (TD-200)
  • [ ] templates/team/profile.md stops saying humans and agents alike while saying agents are unusable; it says: one file per person, and where a participant's role and context are recorded — team/README.md, which the parser skips — so no one invents a fifth key (TD-203)
  • [ ] .tfw/migrations/2.0.0.md: one manifest location (tasks/MIGRATION-2.0.0.md, beside the snapshot, or the root — one, stated); the sentence when the board in the working tree is newer than HEAD and committing is not within your authority, --working-tree is the flag for exactly this case; record the choice; every command written from the project root — cd .tfw && … at :61 rewritten
  • [ ] a grep for the four retired wordings over the templates and the guide returns nothing

Gate: the grep; the guide's commands executed from the root as written. Evidence: the grep output and one command run.

AC-11: The release describes what shipped

  • [ ] executor: the RF states what a project on 2.0.0-dirty.2, .3 or .4 must know, so the entry's updating section and the briefing can be written from it
  • [ ] executor: every fifth-report §6 item and fourth-report defect 7 is either fixed here with a one-line reason or recorded in RF §6 Observations for the reviewer to file
  • [ ] /tfw-release, after review: version bump, CHANGELOG entry and tag as one act, the tag verified to exist; the entry's updating section satisfies AC-2 for a receiver on .2, .3 or .4
  • [ ] field: one consumer already on the line updates to the new tag following the target's update.md from Step −1, with the derived pin, the three questions and the briefing on record — a field report is welcome but the checklist alone is the evidence
  • [ ] 2.0.0 is not claimed

Gate: read the entry as a receiver on .2 and follow every instruction it gives. Evidence: the tag exists; the consumer's checklist.

Evidence Artifacts

File Description
evidence/EV__phase-ac__update_without_guesswork.md Environment header, per-AC table, verdict (required)
evidence/pin_on_live_source.txt AC-1 — both Step 0 runs, commands verbatim
evidence/skipped_tag_read_through.txt AC-2 — the .2 reader's path through the entries
evidence/gates_and_word_count.txt AC-3, AC-6 — allowlist runs, word count, AG-mode stop transcript
evidence/markers_and_exclusions.txt AC-4, AC-7 — cmp regions, copy output, both skipped files
evidence/status_cell_before_after.txt AC-8 — the AILAB-2 fixture, --check tasks runs, four-corpus comparison
evidence/carriers.txt AC-9, AC-10 — the five excerpts, the grep, the root-run command
evidence/briefing_dirty3_to_dirty5.md AC-6 — the rendered briefing for this repository's own delta

6. Technical Guidance

  • Whole-or-refuse is one rule, now applied to the second cell. parse_identifier() is the model: named forms, one dispatcher, malformed/UNDECLARED as the only fallback. Do not add a second heuristic for "closed phases then a live one" — that is the guess, restated.
  • The AILAB-2 shape has two signals: a second declared token (RF) and a second status symbol (🔄). Either alone must be enough; the fixture should test them separately.
  • iter_phase_dirs() already exists in gen_index.py:464; migrate_board.py imports from gen_index today. Reuse it for the manifest's phase listing.
  • Terminal tasks with stateless phase directories are history. Seventeen of them in this repository, under six tasks that closed before phase state existed and carry no task-level state either — migration writes state only for non-terminal tasks, by design. An informational line, not a failure.
  • The word ceiling is the pressure that keeps update.md a procedure. Steps −1, 3, 5 and 8a add roughly 250–300 words to 840. Remove duplication first; the briefing's structure lives in the template, not in the workflow.
  • Markers, first run — R2. Revision 1 called report and leave the Codex rule; the Codex README actually appends. The TS rules report and leave for every block row and puts the rule once in conventions §9. Do not invent detection of "where the TFW content probably is".
  • Do not touch other tasks' files. check_tasks naming a stateless phase directory is the deliverable; writing its status.md is not. The same for consumer projects.
  • Evidence at a pinned commit, never HEAD — the rule this task has met from both sides.
  • The RF is not optional (TD-199). The reviewer reads it against this TS and the master HL at 56c3d70.

7. Definition of Failure

  • ❌ A pin, anywhere in update.md, that is computed from the source's HEAD
  • ❌ A status cell carrying two lifecycle signals classified by either of them
  • ❌ A row skipped by plan() for a reason the manifest does not print
  • ❌ A phase directory without state that --check tasks does not name
  • ❌ A status.md written by this phase into any task other than its own fixtures
  • ❌ A copy step that can overwrite project_config.yaml or knowledge_state.yaml
  • ❌ A handle inferred from a Git identity, OS username or upstream profile — in the workflow text or in the fixture run
  • ❌ An abbreviation rule that can be read as "invent a code" or as "derive without approval"
  • ❌ update.md at or above 1200 words
  • ❌ A CHANGELOG entry rewritten in substance rather than appended to
  • ❌ A form that parsed before this change no longer parses, or any identifier classified differently except the multi-signal rows
  • ❌ Any budget limit crossed without returning to the coordinator
  • ❌ A check reported as passing that never ran

8. Phase Risks

Risk Control
UNDECLARED for multi-signal cells strands real tasks The rule conventions §5 already gives: a person resolves with a transition event from: UNDECLARED; the manifest names each row
The three questions make unattended updates impossible One message, before the first durable write; read-only steps proceed; AG already stops at approvals
update.md crosses the ceiling Duplication removed first; briefing structure in a template; word count is an AC gate
Markers break a consumer whose CLAUDE.md has none Reported and left untouched on first sync — the marker rule in conventions §9, one rule for both adapters
The exclusion list rots Covered by the payload path test that already enumerates every payload file
The phase reopens the parser wider than one cell AC-8's before/after comparison on four corpora: only multi-signal rows change
A stateless phase directory in a live consumer task turns its gate red That is the report the fifth report asked for; the owner authors the file and records the event, the shape conventions §5 already gives for UNDECLARED

9. Cross-Phase Modifications

File Also modified in Coordination note
.tfw/templates/status.md Phase B may add debt-related guidance AC adds only the phase paragraph; no key changes
.tfw/workflows/update.md Phase B (debt view in the update path, if any) AC leaves it under 1200 words; B inherits the ceiling
.tfw/scripts/gen_index.py Phase B (debt discovery) AC touches check_tasks() only

TS — TFW-60 / Phase AC: Update Without Guesswork | 2026-08-30