RES — TFW-50: Minimal Agent Commit Attribution

Date: 2026-08-05 Author: Codex / Researcher Status: 🔬 RES — Complete Parent HL: HL-TFW-50 Mode: Pipeline / Focused

---

Research Context

TFW-50 was reopened because the first consumer model treated Executor/handoff as representative, while real local history contains commits authored under Coordinator, Researcher, Executor, and Reviewer roles. This bounded iteration inventoried every current workflow, adapter, installed copy, and skill with commit relevance; separated the semantic owner, conflicting point-of-use text, always-loaded references, and derived copies; challenged three placement strategies plus a universal-rule alternative; and used only the official Git commit documentation needed to define the metadata boundary. It did not implement changes or prescribe new commit timing.

Briefing

1_briefing.md fixed three questions: which consumers really create commits, which terms make the subject precise without implying identity proof, and which configuration is the smallest complete one. 2_gather.md records the grouped rg inventory and current history; 3_extract.md exposes the candidate configuration space; 4_challenge.md rejects the Extract cadence expansion and selects C7.

Decisions

#	Decision	Rationale
D1	Select C7: one universal conventions owner, one concise glossary entry, and reconciliation only where current text conflicts or overrides the universal rule	Every supported role path loads conventions. Applicability therefore covers Coordinator, Researcher, Executor, and Reviewer without duplicated workflow cues
D2	TFW-50 formats commits that occur; it does not create a commit-cadence policy	A subject rule and a requirement to commit at a particular WAIT, STOP, stage, or artifact boundary are independent. The latter is not authorized by TFW-50
D3	Keep the existing six-path implementation allowlist	Four semantic placements are required: conventions, glossary, the handoff conflict, and active release guidance. The two installed handoff files are mechanically required runtime copies of the same conflict
D4	Refine terminology only in .tfw/conventions.md and .tfw/glossary.md; preserve and verify the other four paths	Current conventions leave scope and role derivation too implicit; the glossary should expressly separate subject attribution from Git author/committer metadata and authentication. Handoff and release conflicts are already correctly reconciled
D5	Treat other workflows, adapters, installed copies, and skills as verification-only consumers	They either inherit the always-loaded owner, set grouping/inclusion without conflicting subject text, or derive behavior from an unchanged canonical source
D6	Derive role from canonical TFW workflow ownership and Role Lock, never from hybrid or maintainer labels	Conventions §15 recognizes only coordinator, researcher, executor, and reviewer; docs/release label drift must not invent a fifth value
D7	Treat scope as normalized explicit work-slice text, not a closed registry	Real history uses task, phase, iteration, docs, knowledge, and other explicit slices; an open derivation is both precise and extensible without runtime/config enforcement
D8	Treat current all-role history as applicability evidence, not causal proof	Explicit delegation prompts and rejected TFW-49 mechanisms confound why the commits complied. The subjects remain declared trace context, not authenticated identity

Exact Terminology Contract

The official git-commit documentation distinguishes commit message text from author and committer information and warns that recorded names do not authenticate a person. TFW Commit Attribution operates only in the first-line subject text.

Term	Exact definition for HL/TS
Commit Attribution	A declared structured prefix in the first-line subject of an AI-authored commit. It supplies searchable trace context; it neither changes nor authenticates Git author/committer metadata and is not actor authentication
agent	The normalized lowercase AI product name from explicit session context, such as codex or claude; not a person, account, model version, Git author, Git committer, or hosting actor
task	The canonical TFW task ID; use project only when the current work genuinely has no task
scope / work slice	The established lowercase explicit work-slice slug, or otherwise a lowercase hyphenated form of the explicit work-slice label. It is open normalized text, not a closed registry
role	The lowercase canonical TFW workflow owner from conventions §15, confirmed by an inline Role Lock where present: coordinator, researcher, executor, or reviewer
summary	The short imperative remainder of the one-line subject after one space; no numeric length protocol, required trailer, or body schema

Recommended sole normative wording:

Every AI-authored commit MUST use [agent/task/scope/role] summary: set agent to the normalized lowercase AI product name from explicit context, task to the canonical TFW task ID (project only when none exists), scope to the established lowercase explicit work-slice slug or otherwise a lowercase hyphenated form of its explicit label, and role to the lowercase canonical TFW workflow owner from §15/Role Lock; keep summary short and imperative, commit locally, and push only after explicit user approval.

Evidence Synthesis

Commit-producing role/workflow matrix

Role	Canonical ownership and real write surface	Existing commit semantics	Lifecycle evidence	TFW-50 implication
Coordinator	/tfw-plan; /tfw-docs, /tfw-knowledge, /tfw-release, /tfw-init, /tfw-update, /tfw-config; resume delegates writes to plan	Docs contains an explicit grouped task-commit instruction; release guidance contains a commit action; other workflow writes do not establish a uniform mandatory cadence	18 marked Coordinator subjects across local/all-ref history, including TFW-50 9aaf1f9 and 056378a, plus docs/knowledge scopes	Universal applicability; retain only release conflict reconciliation. All other Coordinator workflows are verification-only
Researcher	/tfw-research writes Briefing, Gather, Extract, Challenge, and RES	Canonical research requires stage files before WAIT but does not itself require a Git commit at every stage	Four TFW-50 stage commits before RES: 8a190d5, ad16b1a, 5035ab3, 027dc73	Researcher is covered by the universal rule. The delegation-required stage commits prove expressibility, not canonical per-stage cadence
Executor	/tfw-handoff writes ONB, implementation/evidence, and RF	Conventions already say Executor makes incremental commits; handoff explicitly commits ONB	13 marked Executor subjects, including TFW-50 c204f8a, 46fe8b1, and 389168a	Preserve the corrected handoff cue and its two installed runtime copies; do not add commit-per-artifact behavior
Reviewer	/tfw-review writes Map, Verify, Judge, REVIEW, triage, and Reviewer-owned board updates	Canonical review requires stage/final traces but does not itself require a Git commit after every stage	Five marked Reviewer subjects in TFW-49, including 929c489, 0f23362, 9d7e702, 1ebb680, and 3935c53	Universal applicability covers Reviewer. Prior TFW-49 machinery is a causal confounder and is not restored

The grouped inventory covered 72 current framework/adapter/workflow/skill paths plus root entry files and RELEASE.md. Literal positive commit instructions reduce to four families: handoff, docs, active release guidance, and the Codex adapter installation note. Only handoff and active release guidance contradicted or overrode the universal subject/push rule.

Configuration comparison

Configuration	Semantic placements	Mechanical copies	Completeness under actual context contract	Verdict
C1 canonical-only from the pre-TFW-50 baseline	Conventions + glossary	None	All roles load it, but old handoff coupled commit+push and old RELEASE.md overrode the subject	Reject baseline; retaining current reconciliations turns it into C7
C4 one owner + cues at every literal commit action	Owner, glossary, handoff, docs, release, Codex install	Handoff and docs copies	Complete, but docs and Codex install only set grouping/inclusion and do not conflict	Reject as dominated by C7
C3 cue in every role workflow	Owner, glossary, plan, research, handoff, review	Role-workflow copies	Duplicates reminders despite mandatory conventions loading and still does not justify lifecycle workflow treatment	Reject as redundant/asymmetric
C7 universal applicability + conflict reconciliation	Owner, glossary, handoff, release = 4 semantic placements	Two installed handoff copies	Covers all four roles, resolves every actual contradiction, adds no cadence	Select

The 14-semantic-placement/20-copy/three-phase Extract candidate is not a fourth viable implementation: it converts possible commit checkpoints into new mandatory cadence. Mechanical symmetry does not justify its cost or risk.

Open Questions

#	Question	Status	Answer
O1	Does TFW-50 need another research iteration?	Closed	No. The bounded hypotheses, inventory, terminology, alternatives, and confounders are sufficient for HL/TS revision
O2	Must docs/release role-label drift be fixed in TFW-50?	Closed	No broad cleanup. Canonical ownership makes the AI value coordinator; verify that no subject uses reviewer, maintainer, or a hybrid. Normalize labels only in separate approved work if an error or explicit demand appears
O3	Should Researcher/Reviewer stage commits become canonical requirements?	Outside TFW-50	No decision here. Commit cadence requires separate evidence, scope, and approval

Hypotheses (from HL §10)

#	Hypothesis	HL Status	RES Status	Evidence
H1	One canonical owner plus cues only at actual commit actions is sufficient across all roles	needs research	🟢 Refined and supported	One owner is sufficient for universal applicability; point-of-use edits are required only where current text contradicts or overrides it, not at every possible commit action
H2	[agent/task/scope/role] summary uses the smallest precise term set for the user's searches	needs terminology challenge	🟢 Supported with terminology refinement	Five subject components are sufficient when scope is open normalized work-slice text and role derives from canonical ownership/Role Lock
H3	Role workflows, lifecycle workflows, and installed adapter copies form distinct consumer classes and should not be treated as one flat sync list	needs inventory	🟢 Supported	The inventory distinguishes semantic owner, conflict point, always-loaded reference, and derived runtime copy; each has a different edit obligation
H4	Prompt compliance is sufficient for readable declared context when authentication and automated enforcement are explicitly out of scope	bounded claim; needs cross-role evidence	🟡 Bounded support; non-causal	All four roles have readable marked subjects, but explicit delegation prompts and prior TFW-49 mechanisms confound causal attribution. No claim of guaranteed compliance, authentication, or actor detection is supported

HL Update Recommendations

#	What to update	Source
HL1	Replace §3's provisional consumer-scope challenge with D1/D3: the existing six paths are the final sufficient allowlist under C7	D1, D3; Challenge C2, C4, C6, C7
HL2	Refine the §3 terminology table and sole rule using the exact definitions above, especially normalized explicit work-slice scope, canonical-owner/Role-Lock role, and the Git metadata/authentication boundary	D4, D6, D7; Extract E2; Challenge C5
HL3	State expressly that TFW-50 governs commit subject format only and creates no per-stage, per-STOP, per-workflow, or per-artifact commit requirement	D2; Challenge C1
HL4	Keep the one bounded phase and six-file DoF; remove any implication that 20 installed-copy edits or three phases are needed	D3, D5; Challenge C6
HL5	Update H1-H4 and risks: H4 is bounded/non-causal, current all-role history is compatibility evidence, and explicit prompts/prior mechanisms are confounders	D8; Challenge C3, C8
HL6	Record docs/release role-label drift as a verification precision point, not as a broad role-cleanup branch	D6; Challenge C5

TS Recommendation

Use one bounded, reviewable change with the following exact inventory. No new implementation file or phase is needed.

Minimum normative core

Path	Disposition	Required TS treatment
.tfw/conventions.md	MODIFY	Refine the existing sole sentence so all five terms have the exact derivation above; keep one example and the local-commit/explicit-push boundary
.tfw/glossary.md	MODIFY minimally	Keep one concise definition and link to conventions; add the express separation from Git author/committer metadata and authentication without duplicating the grammar

Minimum conflict-reconciliation surface

Path	Disposition	Required TS treatment
.tfw/workflows/handoff.md	PRESERVE + VERIFY	Keep the corrected ONB attribution cue and explicit push approval; add no checkpoint or cadence instruction
.agent/workflows/tfw-handoff.md	PRESERVE + VERIFY derived copy	Keep the same Step 4 correction and preserve all unrelated Evidence drift
.claude/commands/tfw-handoff.md	PRESERVE + VERIFY derived copy	Keep the same Step 4 correction and preserve all unrelated Evidence drift
RELEASE.md	PRESERVE + VERIFY	Keep the compliant release example and explicit push approval

Add/exclude boundary

Group	Disposition	Reason
All other canonical workflows, including plan, research, review, resume, docs, knowledge, release, init, update, and config	EXCLUDE from edits; VERIFY	They inherit conventions and contain no independent conflicting subject/push rule. Docs grouping and canonical release workflow text do not establish an override
.agent/workflows/* and .claude/commands/* other than handoff	EXCLUDE from edits; VERIFY	No changed canonical runtime text or preserved contradiction; do not sync for symmetry
.tfw/adapters/codex/README.md	EXCLUDE from edits; VERIFY	“Commit skills with the project” governs inclusion, not subject syntax or push authority
.tfw/adapters/codex/skills/tfw-*/SKILL.md and .agents/skills/tfw-*/SKILL.md	EXCLUDE from edits; VERIFY	Thin routers already load conventions and canonical workflows; source/installed pairs are equal
Root entry prompts/rules and other adapter entry templates	EXCLUDE from edits; VERIFY	They make conventions always loaded and own no commit action
Docs/release role labels and their installed copies	EXCLUDE from TFW-50 edits; VERIFY value	Canonical §15 precedence makes the AI role coordinator; do not invent maintainer or a hybrid role
tfw-task, templates, project/state config, registries, manifests, schemas, hooks, Git config, Python/scripts, validators, runtime tests, version files, and historical task artifacts	EXCLUDE	Stale/delegating or non-consumer surfaces, or prohibited runtime/config/history expansion
New files	ADD none	The six-path allowlist remains sufficient

TS acceptance should verify the exact terminology, byte-stable preservation of the four conflict-reconciliation files unless a contradiction is found, absence of conflicting subject/automatic-push text across the verification-only corpus, representative marked subjects for all four roles, preservation of unrelated handoff drift, no new cadence, and no remote action. Representative history validates format coverage only; it must not be presented as proof that the conventions sentence caused compliance.

Optional future work, explicitly outside TFW-50, may study Researcher stage-versus-iteration commits, Reviewer stage-versus-final commits, Coordinator workflow closure commits, Executor coherent-slice grouping, and docs/knowledge grouping. None is decided by Commit Attribution.

Fact Candidates

Cognitive mode: Pure reporting — these are human-supplied project statements, not agent-discoverable inventory findings.

#	Category	Candidate	Source	Confidence
FC1	process	Local commits in TFW are made by all TFW roles, including the Researcher; an Executor/handoff-only consumer model is incorrect	User correction, TFW-50 research delegation	★★★
FC2	constraint	TFW-50 is a commit-subject formatting task and must not introduce new commit frequency, per-stage commits, or commit-per-artifact behavior	User Challenge correction, TFW-50 Iteration 1	★★★

fact-candidates: processed 2026-08-05

Strategic Insights (Research)

#	Category	Insight	Source	Confidence
SS1	convention	Universal applicability can cover every role through the always-loaded owner; role coverage does not by itself justify duplicated workflow cues. This makes behavior, not repeated file presence, the completeness test	User, Challenge constraints 2-4	★★★
SS2	constraint	A formatting rule can silently become lifecycle policy if action cues are written as mandatory checkpoints. Future TS wording must keep the conditional “when a workflow commits” boundary explicit	User, Challenge constraint 1	★★★
SS3	philosophy	“Minimum complete” is not the fewest files in the abstract; after consumer classes and the actual context contract are modeled, it can legitimately be the smaller six-path surface	User, original goal and Challenge constraints 3, 6	★★★
SS4	process	Observed compliance is useful compatibility evidence but explicit prompts prevent a causal claim about prompt architecture. Verification should report that confounder instead of overstating reliability	User, Challenge constraint 3 and synthesis approval	★★★

Findings Map

flowchart TD
    A["One conventions owner: format commits that occur"] --> B["Always-loaded applicability"]
    B --> C["Coordinator"]
    B --> D["Researcher"]
    B --> E["Executor"]
    B --> F["Reviewer"]
    A --> G["Existing instruction conflict test"]
    G --> H["Handoff: subject cue plus separate push approval"]
    H --> I["Canonical workflow"]
    H --> J["Two installed runtime copies"]
    G --> K["Active RELEASE.md: compliant subject plus separate push approval"]
    G --> L["No conflict: verification only"]
    L --> M["Other workflows, adapters, skills, entry paths"]
    A -. "does not imply" .-> N["New cadence, per-stage commits, runtime enforcement"]

Iteration Status

• Iteration: 1 of 1 (min) / 1 (max)
• Hypotheses tested: H1 (refined and supported), H2 (supported with terminology refinement), H3 (supported), H4 (bounded support; non-causal)
• Hypotheses deferred: None
• Gaps discovered: Causal effectiveness cannot be inferred from the current prompted history; docs/release role-label drift remains a known verification-only precision issue; canonical commit cadence is intentionally unanswered because it is outside TFW-50
• Superseded decisions: D1/D2 supersede provisional Extract C2/E3, which incorrectly treated possible commit checkpoints as mandatory TFW-50 cadence; D3 supersedes Extract's 20-copy, three-phase inventory

Open Threads (for next iteration)

No open threads. Remaining work is Coordinator-owned HL/TS revision and verification, not another research iteration.

Recommendation

• [x] SUFFICIENT — proceed to /tfw-plan to update HL and write TS
• [ ] MORE NEEDED
• [ ] BLOCKED

⚠️ Coordinator decides whether to continue or proceed. Researcher recommends but does NOT decide.

Conclusion

Iteration 1 establishes that universal applicability, not duplicated role cues, is the correct completeness mechanism for Commit Attribution. The smallest complete future TS retains the six existing paths: refine only conventions and glossary terminology, preserve and verify the four already-correct conflict reconciliations, and treat every other workflow/adapter/skill as verification-only. The strongest research correction was negative: Extract initially expanded attribution into per-stage and per-workflow cadence plus 20 mechanical copy edits; Challenge separated those policies and removed the unsupported scope. The result is complete for all four roles while remaining explicit that local history demonstrates expressibility and searchability, not causation, authentication, or guaranteed compliance.

---

RES — TFW-50: Minimal Agent Commit Attribution | 2026-08-05