Skip to content

REVIEW — TFW-60 / Phase AB: Honest Migration

Date: 2026-08-29 Author: Claude Code (Reviewer), on_behalf_of: saubakirov, via: claude-code Verdict: 🔄 REVISE — two items, one line each, both in the phase's own subject; architecture sound, every other check passes RF: RF Phase AB — commit ecfceee TS: TS Phase AB, revision 2 Contract baseline: master HL at 810b1b8 — recovered per conventions.md §3 rule 15 Stage files: review/map.md, review/verify.md, review/judge.md This file is a synthesis of stage findings. Reference stage files for raw evidence.

1. Map

The executor removed the mechanism behind the third report's defect — an unanchored re.search() in parse_board() that extracted HD-30 from HD-30b before the anchored grammar was consulted — and replaced it with one dispatcher in gen_index.parse_identifier() that classifies a whole candidate as current, clock, legacy or malformed. Duplicate rows and duplicate directories now raise before any output path is opened; the manifest computes three partition equalities, prints each with its arithmetic under ## Guarantees checked, and lists what it did not check; _plain() strips _ only at word boundaries so normalize_text() survives migration. The current grammar {PREFIX}_{YYYYMMDD-HHMMSS}_{ABBR} is issued by plan.md and init.md with the owner approving the abbreviation before a directory exists, carried by canon, glossary, both configs, three templates and the compiler; both historical forms stay readable. update.md fell from 1380 to 852 words while gaining source-pin, tag-check, provenance-drift and allowlist rules. via is free-form and validated non-empty. Release acts were correctly left to /tfw-release per §15 — a correction of Phase AA's ruling, acknowledged by the coordinator in the ONB. 23 implementation paths, 17 counted, 0 new, 1029 lines: inside every budget with no ruling needed.

2. Verify

# What was checked Result Evidence
1 All 23 implementation paths diffed 57a5147..ecfceee; all 8 work artifacts read — 100% after escalation 21 ✅ · 2 ⚠️ verify.md V1–V25
2 Test suite re-run: -k "not repository" / -k repository 281 passed, 3 deselected · 2 passed, 1 skipped — matches RF verify.md commands 1–2
3 gen_index.py --check tasks / --check project 53 valid · consistent with 2.0.0-dirty.3 commands 3–4
4 HELPDESK_SHAPE parsed and _plain() run inline HD-30 legacy · HD-30b malformed · TFW-01_single_underscore malformed · dirty clock · current — reproduces the evidence file command 8
5 update.md word count · adapter triplets · 11 Codex skill pairs 852 · byte-identical · byte-identical commands 5–6
6 Scope census 17 counted files, +712 −317 = 1029 command 7
7 HD-30b mechanism claim confirmed at 57a5147:migrate_board.py:130 C1
8 Seven third-report §7 items each traced to a fix or a filed observation C6
9 32 knowledge citations (HL §7.2 ×29, ONB N1–N3); PV 0–4 scanned in full 32 resolve, 32 semantically verified, 0 irrelevant, 0 hallucinated; priorities 0 and 1 checked as distinct clauses verify.md Knowledge Citations
10 Nine evidence items 8 verified · 1 partial (E3: the collision clause is a transcription of the rule, not a run) verify.md Evidence Verification
11 D1 — migrate_board.py:750 **Unaccounted: 0.** is a string constant; test_migrate_board.py:237, :508 assert the literal ❌ verify.md D1
12 D2 — update.md Step 0 and Step 9: "Both temporary directories are gitignored"; .gitignore holds only .tfw/.upstream/ ❌ verify.md D2, C7

Not verified in this review: the three external corpora (97dd429, 58329e7, aec5f2d) live outside this repository and were accepted on the stated method — git ls-tree at the pin, pre-change regexes reimplemented inline, compared identifier by identifier — with this corpus's 53 confirmed independently.

3. Judge

# Check Status Evidence
1 DoD met? (all TS acceptance criteria) ❌ 8 of 9 gates hold; AC-2 bullet 1 leaves a prose guarantee standing beside its computed form; AC-6's rewrite introduced a false sentence. AC-8 release correctly DEFERRED
2 Purpose Check — is this what we set out to do? + design soundness ✅ Serves DoD 20 and NS1 at 810b1b8; harm removed: a shipped task silently written to TODO under a manifest asserting zero unaccounted. No excess, deferral honoured, material. Design sound on P1, P4, P5, P9, P10
3 Tech debt documented ✅ RF §6, four typed rows with reasons
4 Style & standards ✅ Commit grammar, template, identity fields, clean staging. Minor: no journal event behind the RF updated stamp
5 Observations collected ✅ O1–O4 are real; all promoted
6 RF completeness (§7-9 present) ✅ Both empty sections justified; diagram matches code
7 Evidence completeness — does it exist? ✅ EV with nine rows and five attachments; one legitimate DEFERRED
8 Evidence sufficiency — does it establish the claim? ✅ Every green signal named with what it proves; two limits stated (external corpora on method; E3 collision clause is text) — neither weakens an AC gate
9 Backward compatibility ✅ Every previously parsed identifier parses identically; compiler gains dirty-clock resolution it lacked. Gap filed: artifact naming for the current grammar (TD-201)
10 Safety ✅ Refusals precede any write; consumers read at pins, never written; fixture discarded

4. Verdict

🔄 REVISE

The phase does what amendment A5 asked and does it the right way: the cause is removed rather than the instance patched, the guarantee arithmetic runs before a file is opened, the grammar lands with the owner in the loop, and nothing historical moves. Every claim in the RF that could be re-run was re-run here and held (§2 rows 1–10). The verdict is REVISE, not APPROVE, because two sentences shipped in this commit fail the standard the commit itself sets — and this phase's Definition of Failure names both shapes explicitly: "A guarantee printed that the tool did not compute" and "A check reported as passing that never ran" has a text-side twin in a workflow that states a fact that is not true.

If REVISE — items to fix:

  1. migrate_board.py:750 — delete the constant, or compute it. **Unaccounted: 0.** Every board row is classified exactly once; … is a string literal that shows no arithmetic. It is entailed by require_guarantees() at the head of render_manifest(), but TS AC-2 bullet 1 allows exactly two states for an asserted invariant — computed under a heading naming it, or deleted, "not both left standing" — and the computed form already exists eleven lines lower. It is also the exact sentence the third report quoted over a table listing HD-30 twice. Remove the line and the two assertions that enshrine it (test_migrate_board.py:237, :508), or print len(rows) − Σ classified from computed_guarantees()[0]. Expected diff: three lines.
  2. update.md Step 0 and Step 9 — remove or make true the claim "Both temporary directories are gitignored". .tfw/.upstream-source/ is not in .gitignore, and no receiving project ignores either directory unless its owner adds them; the payload ships no .gitignore. Either add .tfw/.upstream-source/ to this repository's .gitignore and reword to an instruction ("add both to .gitignore if they are not already there"), or drop the sentence. Keep the file under 1200 words (it is at 852). Re-sync the two byte-identical copies.

On re-entry, the executor should also write the phase journal event that the RF stage lacks (row 4) — one handoff or transition event, clock-read, referencing this REVIEW — rather than back-dating anything.

Recommendation to the coordinator, separate from the verdict: TD-200 below (event template still defines via as a provider family) is one path, inside budget (24 / 18 against 30 / 15), and belongs to the same decision AC-7 closed. Admitting .tfw/templates/journal/event.md to §4 for the REVISE round costs one line of TS and closes TD-197 without a residue. If declined, TD-200 stands as filed.

5. Tech Debt Collected

Source format: Use reference patterns (compilable_contract.md §2).

# Source Severity File Description Action
TD-200 RF TFW-60/AB §6 O4 Med .tfw/templates/journal/event.md:49, :70 The event template — the file every event author copies — still defines via as "provider family: claude, codex, gemini" while conventions.md §4 and validate_event() now declare and check free-form non-empty text. Same drift class as TD-197, one carrier over; outside the approved 25-path surface, so filed rather than fixed → Phase AB REVISE round if the coordinator admits the path; otherwise next phase touching the payload
TD-201 REVIEW TFW-60/AB §2 (reviewer finding) Med .tfw/conventions.md §4 Artifact file naming The table's examples and its no-title rule address clock and legacy tasks only; the current grammar PREFIX_stamp_ABBR has no example and no stated rule. The phase's own AC-3 fixture produced HL-TFW_20260829-172110_ABT__approved_fixture.md — a title appended — while templates/status.md in the same commit teaches authority: HL-PREFIX_YYYYMMDD-HHMMSS_ABBR.md. A framework undefined about its own artifact names is the class Phase AB exists to end → Phase AB REVISE round (one table row and one sentence) or the 2.0.0 release pass
TD-202 RF TFW-60/AB §6 O1 Med KNOWLEDGE.md:102 (D68), :162 (§3 Legacy) Both still declare YYYYMMDD-HHMMSS__slug the current identifier and say a current event name includes its actor. D37 reserves these sections for /tfw-docs; correctly not touched by the executor → /tfw-docs after APPROVE, together with TD-186
TD-203 RF TFW-60/AB §6 O2 · third field report §7 item 2 Med .tfw/templates/team/profile.md:12–14, 27–33 Says profiles cover "humans and agents alike", then that team/ holds people and agent profiles are unusable until TFW-54. Shipped to consumers; the third report's operator read both. Outside the approved surface → scoped follow-up (payload template pass)
TD-204 RF TFW-60/AB §6 O3 · third field report §7 item 5 Med .tfw/adapters/antigravity/tfw-rules.md.template:5–7, .agent/rules/tfw.md:5–8 Adapter source requires a {version} substitution on every update while the rendered rule reads .tfw/VERSION; the substitution was missed twice in a consumer, which then announced "TFW 0.8.5" for two releases. Align source with rendering → scoped adapter update
TD-205 REVIEW TFW-60/AB §3 row 4 Low tasks/TFW-60__conflict_resistant_shared_workspace/phase-ab/journal/ The phase journal holds one event (ONB → execution, 16:55:40); the RF completion that set updated: 20260829-173155 has no event. The state file says where the phase is; the journal does not say how it got to RF → REVISE round: one clock-read event on re-entry, stating the gap; nothing back-dated

6. Traces Updated

  • [x] the phase's status.md — lifecycle: RF unchanged (REVISE returns to execution, same task, as in 20260829-062100__transition__b2dc.md); updated set from the clock; transition event written to phase-ab/journal/ as {stamp}__transition__{token}.md with from: RF, to: RF
  • [ ] HL status — unchanged; the phase is not complete
  • [x] the phase's status.md — updated reflects this review. No counter is incremented
  • [x] Other project files — TECH_DEBT.md appended TD-200–TD-205; TD-197 left ⬜ Open with a note that the canon and validator half closed at ecfceee and the template half is TD-200
  • [ ] tfw-docs: Pending APPROVE — first items TD-202 (D68, §3 Legacy) and TD-186
  • [ ] tfw-knowledge: N/A for this review — no fact candidates below; pending candidates from Phase AA still stand

7. Fact Candidates

fact-candidates: processed 2026-08-30

No fact candidates from this review. The human rulings that shaped Phase AB — the grammar lands in this phase rather than after 2.0.0, no fourth external run gates the release, HD-30b is a closed sub-item and stays untouched, ABT is approved for a disposable fixture — are already recorded in HL §12 A5, TS §2 and ONB answers 1–3, and nothing human-only was added during this review session. The reviewer's own observations (a test that asserts a literal guarantee string enshrines the constant it should be checking; a rewrite that shortens a workflow can widen a local fact into a false one) are analysis, and belong in §4 where they are.

REVIEW — TFW-60 / Phase AB: Honest Migration | 2026-08-29