Skip to content

TS — TFW-50: Minimal Agent Commit Attribution

Date: 2026-08-05 Author: Codex / Coordinator Status: ✅ TS — Revised from RES Iteration 1 and approved under delegated user authority Parent HL: HL-TFW-50 Research: RES Iteration 1

1. Objective

Finish one portable Markdown-level Commit Attribution rule for every AI-authored local commit, regardless of whether the active TFW role is Coordinator, Researcher, Executor, or Reviewer. Refine the exact terms, preserve the already-correct handoff/release conflict fixes, verify the whole consumer corpus, and add neither enforcement runtime nor commit cadence.

2. Scope

In Scope

  • Refine the sole normative sentence and example in .tfw/conventions.md.
  • Refine the concise .tfw/glossary.md definition without duplicating the grammar.
  • Preserve and verify the corrected handoff text in canonical, Antigravity, and Claude files.
  • Preserve and verify the corrected active release subject/push boundary.
  • Verify universal applicability across all four roles, current workflows/adapters, and representative local history.
  • Complete EV/RF and independent review locally.

Out of Scope

  • New commit requirements at stages, WAITs, STOPs, workflow completion, artifacts, files, or acceptance criteria.
  • Edits to plan, research, review, docs, knowledge, canonical release, init, update, config, resume, Codex skills, or other adapter copies solely to repeat the rule.
  • Hooks, scripts, schemas, registries, manifests, state/config keys, validators, trailers, Git history rewriting, actor authentication, or human/agent detection.
  • Push, fetch, remote tag, deploy, publish, or notify.

3. Principles Check

# Principle (HL §7) Enforced by Gate
P1 Outcome over mechanism AC-1, AC-6 Searchable subjects; no runtime
P2 Naming creates behavior AC-1, AC-2 Exact five-part terminology
P3 Single source of truth AC-1, AC-4 One owner; other files only reconcile conflicts
P4 Declared context, not identity proof AC-2, AC-5 Metadata/authentication boundary is explicit
P5 Human publication authority AC-4, AC-6 Local commit and push remain separate
P6 Proportional completeness AC-3, AC-5 All roles covered without role-specific duplication or cadence
P7 Transparent boundaries AC-4, AC-5, AC-6 Exact six paths and verification-only corpus

4. Affected Files

Implementation allowlist

File Action Description
.tfw/conventions.md MODIFY Refine the sole normative rule and exact term derivation
.tfw/glossary.md MODIFY Refine the concise definition and Git metadata boundary
.tfw/workflows/handoff.md PRESERVE + VERIFY Keep corrected Step 4; no cadence addition
.agent/workflows/tfw-handoff.md PRESERVE + VERIFY Keep derived Step 4; preserve unrelated Evidence drift
.claude/commands/tfw-handoff.md PRESERVE + VERIFY Keep derived Step 4; preserve unrelated Evidence drift
RELEASE.md PRESERVE + VERIFY Keep attributed example and explicit push approval

No new implementation file is allowed. The task's total implementation surface remains the six paths already changed in commit 389168a; the corrective execution should change only conventions and glossary unless verification discovers a contradiction inside the four preserved paths.

Verification-only corpus

  • All canonical workflows: plan, research, handoff, review, resume, docs, knowledge, release, init, update, config.
  • All .agent/.claude workflow copies and Codex source/installed skills.
  • Root and adapter entry files that load conventions.
  • .tfw/adapters/codex/README.md install/repair action.
  • Actual local TFW-50 subjects and representative Reviewer history.

Lifecycle traces

README, EV, RF, review stages, REVIEW, and later docs/knowledge markers follow their canonical workflow ownership; they are not extra implementation consumers.

5. Acceptance Criteria

AC-1: Sole normative owner and grammar

  • [ ] .tfw/conventions.md contains one Commit Attribution section, one normative sentence, and one example.
  • [ ] Grammar is exactly [agent/task/scope/role] summary.
  • [ ] No changed workflow, adapter, or glossary text becomes a second normative owner.

Gate: heading/exact-occurrence scan and full six-path diff review. Evidence: N/A — repository text is the observed contract.

AC-2: Exact terms and boundaries [depends: AC-1]

  • [ ] agent is the lowercase AI product from explicit context, not a person, model, account, Git author, or committer.
  • [ ] task is the canonical TFW task ID; project is allowed only when no task exists.
  • [ ] scope is open normalized text from the explicit work-slice label, not a registry or inferred path.
  • [ ] role comes from canonical TFW workflow ownership in conventions §15/Role Lock and is one of coordinator, researcher, executor, reviewer.
  • [ ] summary is short and imperative with no numeric limit, body schema, or trailer.
  • [ ] Commit Attribution is first-line subject trace context, separate from Git author/committer metadata and authentication; unmarked commits are not classified as human.

Gate: semantic assertions against conventions/glossary and prohibited-term scan. Evidence: N/A — repository text is the observed contract.

AC-3: Universal role applicability without cadence [depends: AC-1] [depends: AC-2]

  • [ ] The rule says every AI-authored commit, so it applies equally to Coordinator, Researcher, Executor, and Reviewer.
  • [ ] No role-specific grammar or duplicated full rule is added.
  • [ ] No new requirement says to commit per stage, WAIT, STOP, workflow, artifact, file, or AC.
  • [ ] Current docs/release label drift cannot create maintainer or a hybrid role because canonical §15/Role Lock owns the role value; no broad label-cleanup branch is added.

Gate: complete verification-only corpus scan plus comparison to RES D1/D2/D6. Evidence: N/A — this verifies methodology text and absence of cadence expansion.

AC-4: Existing conflict reconciliation remains correct [depends: AC-1]

  • [ ] Canonical and installed handoff Step 4 use Commit Attribution and require separate explicit approval before push.
  • [ ] RELEASE.md uses a compliant project/release/coordinator example and requires approval before pushing commit/tag.
  • [ ] The four preserve-and-verify files remain byte-stable during corrective execution unless a concrete contradiction is found.
  • [ ] Unrelated canonical-vs-installed handoff Evidence drift remains unchanged.

Gate: baseline/current hashes and line-scoped diffs for the four files. Evidence: N/A — no publication is authorized.

AC-5: Minimal complete consumer treatment [depends: AC-3] [depends: AC-4]

  • [ ] The total implementation allowlist remains exactly six existing paths; no new implementation file or phase is created.
  • [ ] Other workflows/adapters/skills are audited as verification-only and contain no conflicting subject or automatic-push instruction.
  • [ ] Docs grouping and Codex install inclusion wording are not mistaken for competing subject rules.
  • [ ] No hook, script, schema, config/state, validator, runtime, operation matrix, or broad adapter sync appears.

Gate: exact total-task allowlist diff from plan baseline, grouped rg scan, protected-path audit, and git diff --check. Evidence: N/A — repository-local source comparison.

AC-6: Cross-role readability, regression, and no publication [depends: AC-5]

  • [ ] Representative local subjects exist for Coordinator, Researcher, Executor, and — by final independent review — Reviewer using the same grammar.
  • [ ] Subjects accurately reflect explicit task, work slice, and canonical TFW role.
  • [ ] History is reported as compatibility/searchability evidence only; explicit prompts and prior TFW-49 machinery are named confounders, so no causal compliance or authentication claim is made.
  • [ ] Existing docs tests pass; generated outputs/caches from verification are removed without touching ignored user files.
  • [ ] Remote remains at bc6779e; no push/fetch/tag/deploy/publish/notify occurs.

Gate: git log --format=%H%x09%s bc6779e..HEAD, independent review's own attributed commit, docs tests, status/remote comparison. Evidence: Record actual representative hashes/subjects and local-vs-remote state in EV. Executor may record Reviewer completion as pending independent review rather than claim it early.

Evidence Artifact

File Description
evidence/EV__TFW-50__minimal_agent_commit_attribution.md Exact terminology/source checks, six-path preservation, all-role subjects available at execution time, regression results, and no-publication state

6. Technical Guidance

  • Apply RES Iteration 1 C7, not Extract's rejected 33-file cadence configuration.
  • Keep one normative sentence; a short glossary definition may clarify metadata without restating the grammar.
  • Preserve the four already-correct reconciliation files unless a concrete contradiction is reproduced.
  • Use conventions §15 as role authority. /tfw-docs and /tfw-release AI work therefore uses coordinator; do not create maintainer or hybrid roles.
  • Audit the wider corpus read-only. Universal applicability is the intended mechanism; absence of a local cue is not itself a defect.
  • Existing TFW-50 Coordinator/Executor/Researcher commits and earlier Reviewer commits show expressibility, but their prompts/mechanisms confound causality.

7. Definition of Failure

  • ❌ Any implementation path outside the six-file total-task allowlist changes without Coordinator return.
  • ❌ Any new framework/runtime/config/state file or phase is created.
  • ❌ The full rule is duplicated outside conventions.
  • ❌ A new commit cadence or checkpoint policy is introduced.
  • ❌ Existing handoff Evidence drift changes.
  • ❌ Attribution is claimed to authenticate an actor or identify unmarked commits as human.
  • ❌ Extra fields, fixed length limits, operation matrices, or trailers appear.
  • ❌ Any current handoff/release text authorizes push without explicit user approval.
  • ❌ Any remote action, history rewrite, or old-hook restoration occurs.
  • ❌ Tests/checks fail when RF is written, or generated debris is left behind.

8. Risks

Risk Mitigation
Universal rule is mistaken for Executor-only because handoff has a local cue AC-3 all-role semantics plus representative all-role history
Precision wording becomes an unreadable mini-spec One sentence, one example, one concise glossary definition; no registry
Research cadence proposal leaks into implementation Explicit AC-3 and Definition of Failure
Verification-only files are edited for symmetry Six-path allowlist and preserve-vs-modify dispositions
Prompted history is overclaimed as causal evidence AC-6 requires explicit confounder disclosure

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