Skip to content

title: "Gather — "What do we NOT know?"" source: "tasks/TFW-60__conflict_resistant_shared_workspace/research/iter1/2_gather.md"

Gather — "What do we NOT know?"

Mindset: Explorer. You're mapping unknown territory. Widen before you narrow. Every assumption is a question. Test: "Can I name every dimension and its alternatives without checking my sources?" Parent: HL-TFW-60 Goal: Make concurrent human and agent work in a synchronized TFW workspace conflict-resistant by moving normal lifecycle state and coordination to stable task-local, single-writer surfaces while retaining discoverability and Git provenance.

Dimensions

No alternative is selected in Gather. Service/database alternatives remain as counterfactuals even though Phase A excludes them; they show which guarantees plain file sync cannot provide.

Dimension Alt A Alt B Alt C Alt D
D1. Catalogue persistence manually maintained Task Board generated on demand low-churn persistent router plus generated live catalogue no project catalogue
D2. Catalogue authority catalogue is authority task-local files are authority; catalogue is disposable field-split authority external tracker/database authority
D3. Status carrier marker filenames strict status.yaml bounded STATUS.md infer from artifacts or parent directory
D4. Human view raw YAML/markers generated Markdown/CLI view bounded Markdown with machine-readable frontmatter separately maintained human note
D5. Transition write one scalar update at a stable path marker create/delete/rename whole task-directory move append transition and derive state
D6. Journal storage Markdown event table JSONL immutable event files database event table
D7. Journal breadth lifecycle only lifecycle plus handoffs/gates/claims all agent activity rolling digest, not event history
D8. Journal retention unbounded bounded window plus archive terminal compaction separate phase/run journals
D9. Ownership coordinator-only control and journal one owner per field/file lock-mediated multi-writer unrestricted merge
D10. Edition contract one shared schema shared semantics, edition serialization common kernel plus extensions separate edition models
D11. Git topology sync working files and .git sync working files; local Git metadata one landing clone plus sync-only peers shared workspace without Git
D12. Isolation separate task files in one tree branches/worktrees separate project copies service leases/transactions
D13. View freshness rebuild per read explicit refresh with revision event-triggered generation visibly stale timestamped view
D14. Cold-start entry root README scan/query command agent instruction with stable glob external dashboard

Findings

G1. Evidence boundary and name disambiguation

Evidence is labelled as follows: S = current source, test, or shipped template; D = current primary-repository documentation, not independently executed; L = local repository/history evidence. No third-party system was installed or executed.

  • GSD / Get Shit Done: Taskwarrior/GTD projects and glittercowboy/get-shit-done are different projects. gsd-build/get-shit-done is now archived (GitHub showed read-only from 2026-06-26); its v1 file implementation remains historical source evidence. gsd-build/gsd-2 says development moved; the active home observed on 2026-08-26 is open-gsd/gsd-pi. Current GSD Pi uses authoritative SQLite/WAL plus rendered Markdown, not v1's file authority.
  • BMAD: bmad-code-org/BMAD-METHOD, not wrappers.
  • Hermes: NousResearch/hermes-agent, not similarly named packages or mirrors.
  • Additional current systems: github/spec-kit, Fission-AI/OpenSpec, and eyaltoledano/claude-task-master.

G2. Current TFW read/write/ownership map

Surface Current role Writers found Readers/consumers found Ownership/concurrency observation
Root README.md Task Board task identity/title, live status, artifact links Plan sets HL_DRAFT (.tfw/workflows/plan.md:52); handoff sets ONB and later RF (handoff.md:69 and continuation); review sets verdict/KNW/DONE (review.md:116-126); init creates/transitions it (init.md:58-106,170,202); standalone research may register RES (research/base.md:37) AGENTS new-session path; all role skills; resume; release; docs generator; people Independent tasks and successive roles converge on one line-oriented file. The writer changes by stage; there is no persistent single owner.
Task folder stable task identity/group coordinator/init at creation all workflows and docs generator Already isolates task artifacts; Full does not keep live status here.
HL / TS contract and executable specification Coordinator only Researcher, Executor, Reviewer Task-local, role-owned; not an operational log.
research/iterN/*, RES.md research trace/synthesis Researcher only Coordinator/Planner Iteration-isolated; not live coordination state.
ONB executor briefing and gate questions Executor; coordinator answers under protocol Executor/Reviewer Task-local, but only one stage's dialogue.
RF/evidence execution claims/evidence Executor Reviewer/Coordinator Terminal-stage trace; routine journal duplication would distort purpose.
REVIEW independent verdict Reviewer Coordinator/knowledge/docs Terminal-stage trace; not live status.
Generated tasks page grouped task files plus status copied from README docs/scripts/gen_docs.py published docs users _generate_tasks_index() regex-parses README's first three columns (:324-341); README remains upstream authority.
Git index/commit staging and provenance every committing role/session reviewers and later archaeology Index is shared repository/worktree metadata. It is not protected by task file ownership.

Current board measurement: 60 task rows; 40 have 8 cells and 20 have 9, while the header declares 8 columns. The board exposes ID, title-like description, status, and links. It does not consistently expose distinct goal, value, owner, dependency, or last-event fields. An on-demand catalogue cannot recover those absent fields merely by scanning terminal artifacts.

Cold-start dependence is structural:

  • AGENTS.md requires the Task Board in new-session context.
  • Installed TFW role skills put it in their load order.
  • /tfw-resume combines board state with artifact existence.
  • /tfw-release reads completed tasks from the board.
  • docs/scripts/gen_docs.py copies status from it.

Integrity counterexample: README links TFW-54 as though it has an HL, but its task folder currently contains PROPOSAL__TFW-54__agent_team_mode.md and no HL. Persistent catalogues can therefore be discoverable but wrong; generated catalogues can also misclassify nonstandard task shapes.

G3. Reconstruction of TD-81, TD-144, TD-175, TD-177, TD-178

Debt Concrete evidence Exposed dimension
TD-81 docs/scripts/gen_docs.py:309-366 groups pages, then regex-parses README to copy task name/status. It tolerates row-width drift only because it reads early columns. Human table grammar is an implicit machine API.
TD-144 Two sessions shared one working tree and Git index. Commit fbdf443, titled for TFW-53/B, captured TFW-56's staged deletion of review mode files; verified with git show --stat fbdf443. Independent working files do not protect a shared index; explicit-path staging is a separate control.
TD-175 BLOCKED had 0 uses across the then-measured 46 rows; actual stopped experiments used REJECTED. A strict enum can be syntactically valid but operationally unused or misnamed.
TD-177 Header/row widths disagree; current count is 40 eight-cell and 20 nine-cell rows. Regex survives; a strict consumer need not. Human-readable Markdown alone is not a stable machine protocol.
TD-178 The coordinator landed TFW-53/E board rows inside 8d9432b, whose subject names TFW-58 and does not mention TFW-53; verified with git show --stat 8d9432b. Coordinator ownership does not ensure task-specific provenance or commit granularity.

The debts expose four non-equivalent controls: separate task files, persistent writer ownership, explicit-path staging, and one-task landing commits.

G4. Current edition divergence

Edition Identity/status evidence Writer model Stable-path consequence
Light shared TASKS.md plus per-task TRACE.md manual common list remains a same-file surface.
Assisted timestamp/handle/slug ID; task under work/new|doing|review|done|blocked; TRACE.md repeats Статус: one writer per task/trace; explicitly no shared board/counter/trace every transition moves the whole task folder and edits TRACE (editions/02-assisted/AGENTS.md:65-77).
Full framework ID, root Task Board status, task-local role artifacts writer changes by lifecycle role task evidence is isolated, but status and Git index remain shared.

Assisted's folder-move model is shipped local instruction, not inference. Its hooks are documented as unproven for durable orchestration; the supported baseline is manual/file discipline. H4 therefore begins from real differences in identity, state location, transition operation, duplication, and artifacts—not only transport and Git.

G5. Read-only AFD evidence

Inspected without modification at C:\Users\c0rpa\.codex\worktrees\10b2\ai-first-devices:

  • Root README retains the live Task Board and cold-start current status (README.md:55,159); tasks/README.md routes status back to root rather than duplicating it.
  • docs/DOCS_CONTRACT.md:34 separates hand-reviewed DOCS_MAP authority from disposable, rebuildable DOCS_INDEX.jsonl projection.
  • Lifecycle is independent of location because one folder cannot encode independent axes (:41); IDs remain stable across path changes (:55).
  • A concrete bloat fixture warns on a Task Board cell above about 180 words or README above about 1,800 words. AFD-23's cell was measured at 742 words before compression to 60; README moved from 2,889 toward 2,253 (:93,101-108).
  • The owner made these size signals advisory, not hard gates. Evidence shows measurement detects bloat, not that advice prevents it.
  • Evidence stays at its source (:114); search is a read-only projection (:118); structural checks do not prove arbitrary prose true (:122).

AFD supplies local evidence for stable IDs + one semantic owner + disposable view + source references. It also supplies counter-evidence: an always-load persistent board accumulated a 742-word cell before cleanup.

G6. Primary-system comparison

System State/catalogue History Concurrency Evidence level and primary source
GSD v1, archived persistent .planning/STATE.md and ROADMAP.md; every workflow reads state first bounded digest of recent decisions, blockers, session continuity; template says under 100 lines, 3-5 recent decisions, active blockers STATE.md.lock with O_EXCL, stale detection and read-modify-write S: state.cjs, state.md template, roadmap.md. Historical lineage only.
GSD Pi, current .gsd/gsd.db is runtime authority; Markdown is rendered projection DB holds milestones/tasks/decisions/summaries SQLite/WAL, leases and worktrees; explicitly local disk and single-host; gsd.db* over network/cloud sync unsupported D/current primary: auto-mode, parallel orchestration. Non-transferable service counterexample.
BMAD Method central sprint-status.yaml; strict epic/story/retro vocabularies; status command recommends next story result detail remains in story/spec artifacts deterministic generator normalizes legacy states and merges/preserves progress; central YAML is still one file S: sprint_plan.py, sprint-planning skill. Source present, not run here.
Hermes Agent tasks/status/claims/runs in ~/.hermes/kanban.db; CLI/tools/REST/UI share it append-only task_events; 400-character event summary plus full run metadata WAL and BEGIN IMMEDIATE claims; local process profiles S+D: Kanban test, Kanban reference. DB semantics do not transfer to plain sync.
GitHub Spec Kit branch feature directory with spec/plan/research/contracts/tasks.md; strict task checkboxes and exact paths checklist progress; no project event journal located [P] only for non-conflicting work; same-file work sequential S: tasks command, tasks template, implement command.
OpenSpec per-change proposal/specs/design/tasks.md; CLI reports artifact state change artifacts are trace; no general journal located change folders isolate work, but archive moves the whole directory D/current primary: workflows.md; archive warns but does not block on incomplete tasks.
Claude Task Master central structured tasks.json; list/next create on-demand views task details, not a strict append-only journal set-status mutates central JSON; generated task text files require regeneration D/current primary: README-task-master.md.

Observed patterns, not selections:

  1. Persistent discoverable state is common; no inspected system reconstructs project meaning solely from terminal artifacts.
  2. Strictness exists in YAML/JSON, Markdown/frontmatter, and databases. Parser/writer contracts and ownership matter independently of extension.
  3. Central state is normally protected by a lock, database transaction, or serialized CLI—guarantees outside bare cross-device file sync.
  4. GSD v1 is direct counter-evidence that human-readable bounded Markdown cannot be strict; its concurrency safety nevertheless depended on source-level locking.
  5. Rigorous event history is narrow. Hermes types lifecycle/claim/run events and caps summaries; GSD keeps a digest and references deeper records.
  6. Whole-directory moves in OpenSpec and Assisted conflict with TFW-60's frozen stable-path outcome.

G7. File-sync semantics

Scenario Google Drive for desktop OneDrive/SharePoint Dropbox Plain-sync evidence floor
Independent files new files/folders sync across devices file-oriented sync file-oriented sync Different task files reduce same-file collisions; propagation order is not a transaction.
Same-file concurrent/offline edit incompatible changes can leave an edited copy under original parent/root or Lost & Found non-Office conflicts keep both versions and append device name; up to five creates named conflicted copy for simultaneous/offline/open-file edits; manual merge Conflict copies/last-writer behavior are normal.
Offline reconnect mirror retains local files; stream uses cache; unsynced recovery may use desktop recovery/Lost & Found local/server copies may require resolution/resync offline edit explicitly causes conflict copy No distributed lock or cross-file happens-before survives disconnect.
Move/rename moved/renamed Drive root requires reconnect; move/delete plus local edit may relocate the copy renaming established synced root can break relationship shared-folder rename can be member-local; some moves break membership; selective-sync folder conflicts exist Directory move is not an atomic state transition across clients.
Cross-file transaction none documented none documented none documented Treat related file changes as independently visible and reorderable.

Primary sources:

The final column is an architectural inference from primary vendor contracts, not a claim that providers share an algorithm. Reliable distributed locks, compare-and-swap, ordered multi-file propagation, and atomic directory transactions were not documented.

G8. Git topology evidence

  • Linked worktrees share repository data but have per-worktree HEAD, index, and other state (git-worktree).
  • A linked worktree's .git file points into the main repository's administrative directory; refs may be shared while index/HEAD are per-worktree (gitrepository-layout).
  • Git's worktree lock protects administrative pruning for temporarily unavailable worktrees; it is not a cloud-sync write lock.

Observed consequences:

  1. Syncing .git or a worktree .git pointer also syncs or invalidates machine-local administrative paths/state; Git documentation does not present cloud-sync replication of .git as a collaboration protocol.
  2. Local Git metadata separates indexes but does not decide who commits shared working-file changes.
  3. TD-144 is an index-sharing failure; TD-178 is a landing/provenance failure. Separate indexes solve only the former.
  4. Worktrees provide same-host isolation, not cross-host file-sync transactions. Current GSD Pi reaches the same single-host boundary.

G9. Deep-mode evidence loops

Loop Hypothesis pressure Counter-evidence found Revised unknown carried forward
1. Internal + AFD task-local sources can support a derived view README uniquely carries live aggregate fields and all cold-start paths read it; AFD still retains a persistent board Which low-churn entry and task-local fields preserve zero-tool human and agent discovery?
2. Workflow systems strict YAML/markers and a separate journal might be inherently safer GSD ships bounded Markdown plus lock; DB systems depend on non-transferable transactions; generated secondary files drift Which carrier combines parser strictness, human clarity, stable path, and persistent ownership? How narrow must journal events be?
3. Sync + Git single-writer files plus Git provenance might be sufficient offline forks, move/edit ambiguity, no multi-file ordering, and TD-178's cross-task landing remain What exact one-file transition, owner, staging, and landing rules are jointly necessary?

Provisional evidence pressure only—final verdicts are deferred to Extract/Challenge:

Hypothesis Supporting observation Counter-observation/open gap
H1 task-local sources plus generated views isolate writers persistent cold-start sources are universal in the sample; TFW task folders lack some board fields; zero-tool usability untested
H2 BMAD proves strict YAML; current board grammar drifts GSD proves bounded Markdown can be parsed/locked; marker operations add rename/create-delete behavior
H3 Hermes proves typed, narrow event history; AFD proves source/reference separation no file-first example proves an unbounded journal; grammar, references and retention remain open
H4 editions share identity/lifecycle/owner/result/resume concepts Assisted and Full currently differ in ID, path semantics, transition operation and artifact model, not only transport

Checkpoint

Found Remaining for Extract/Challenge
Current TFW read/write/ownership map and all five debt reconstructions Minimal field-level shared contract and one owner per field.
Read-only AFD authority/projection/bloat evidence Transferable patterns without importing debt or knowledge architecture.
Seven-system comparison with current GSD lineage and evidence levels Coherent configurations and cold-start/human tests.
Provider same-file/offline/move semantics Testable plain-sync invariants without distributed-transaction claims.
Git index/worktree evidence separated from task ownership Local-.git landing model, explicit-path staging, and own-task commit rules.
Counter-evidence against simple H1-H4 forms Exact carrier, journal grammar/reference/retention, and edition migration boundary.

Evidence state: internal and AFD observations are local/source-level. GSD v1 lock, BMAD parser, Hermes test, and Spec Kit templates are primary source-level evidence. GSD Pi, OpenSpec, Task Master, and provider semantics are primary repository/vendor documentation not executed here. Debt and knowledge architecture remain excluded.

Recommendation at this gate: proceed to Extract. The decision dimensions, external coverage, counter-evidence, and remaining gaps are explicit; another Gather pass would widen product count rather than close a Phase-A variable.

Questions for the Coordinator: none. Existing constraints are sufficient. Service/DB alternatives should remain counterexamples, not candidates that widen Phase A.

Files written at this gate: research/iter1/1_briefing.md; research/iter1/2_gather.md.

Sufficiency: - [x] External source used? - [x] Briefing gap closed? - [x] Dimensions identified? - [x] At least one hypothesis tested and counter-evidence sought (deep mode)? - [x] Three Gather loops completed?

Stage complete: YES → User decision: WAIT — Coordinator must approve Extract or redirect the evidence gap.