Skip to content

PROPOSAL — TFW-58: What Happens After a REVISE

Date: 2026-08-13 Author: Coordinator (Claude Code) Status: ⬜ TODO — proposal only. No HL, no TS. Entry point: /tfw-plan. Sequenced after: TFW-53 Phase C — see Why not now below Splits with: TFW-54 — agent freshness on a revision round belongs there

Origin

Owner, 2026-08-13, describing the loop as actually practised:

«После ревью есть замечания, и по логике их надо сначала отправить координатору… потом дополнить TS следующим revision, потом исполнителю сказать иди доделай… С другой стороны, я часто просто отдавал ревайз исполнителю, и он сам закрывал — исключая координатора из цикла. Мне проще второй вариант, меньше чатов, быстрее добить. Но при этом я понимаю, что правильно через координатора.»

«В кодексе это иногда превращается в бесконечно долгий цикл ревайзов туда-сюда, и расширяется базовый HL, и они по-немногу плывут и выходят за рамки задачи.»

«Вроде бы мы это отплытие как раз и починили тут в TFW-53, но мне кажется при полной автономии ещё чего-то не хватает.»

The owner is right on both counts. TFW-53 closes the drift channel above the TS and leaves the one below it open — and the shortcut they feel guilty about is not a violation, because there is no rule to violate.

The canon gap, measured

Where What it says What it does not say
conventions.md:348 🔄 REVISE — specific issues → back to execution (same task) back to whom, through whom, with what artifact
resume.md:52 "flag it as needing re-execution" a flag, not a process
templates/REVIEW.md:48 ### If REVISE — items to fix: what the executor does with the list
handoff.md Phase 1 Onboarding, as if arriving fresh no entry point for "I am returning after a REVISE"
all of .tfw/ — the word revision appears zero times

The arrow exists on the lifecycle diagram and nothing implements it. Every project therefore invents its own loop, and two incompatible inventions are already running in this ecosystem.

Practice: two patterns, neither written down

[TFW-38](../TFW-38__quality_enforcement/HL-TFW-38__quality_enforcement.md) — a second full phase           AFD — revisions inside one artifact
──────────────────────────────         ─────────────────────────────────────
PhaseA  → TS ONB RF REVIEW             149 REVIEW files
PhaseA2 → TS ONB RF REVIEW             13 carry rev1..rev4   (8.7%)
                                       deepest observed loop: 4 rounds
= new work, clean trace,               arc recorded in one file:
  heavy ceremony                       "wrong APPROVE (rev1) → REVISE
                                        → rev2/rev3 fixes → APPROVE"

AFD also carries 18 numbered sub-phases (phase-a0..a3, b0..b7, c1..c3, d1..d3). Eight sub-phases of a single Phase B is a proliferation signal worth measuring: how many were planned splits, and how many were revision rounds wearing a phase name?

What TFW-53 closes, and what it leaves open

Drift channel Status after TFW-53
Research rewrites the goals ✅ frozen contract + amendment log
Coordinator absorbs findings silently ✅ classify, never update
Reviewer approves work that misses the point ✅ Purpose Check (Phase C)
Reviewer blocks on wording ✅ materiality bar — a block must rest on material impact
A REVISE loop grows the TS ❌ open — the TS traceability gate was ruled out of scope by the owner
A REVISE loop has no termination ❌ open — nothing counts rounds
Who is in the loop after a REVISE ❌ open — the canon is silent

TFW-49 drifted through the contract. This is the same failure one level down, through the spec.

Design sketch — the ladder already exists

The question "coordinator in the loop or not?" is malformed. It conflates who decides what gets fixed with who does the fixing. Separate them and the answer falls out of a mechanism TFW-53 already shipped and the owner already ruled on:

reviewer files a REVISE finding
      │
      ├─ satisfiable inside the approved TS?
      │     YES → executor fixes, reviewer re-checks.  COORDINATOR NOT NEEDED
      │            (the owner's shortcut — legitimate, and the majority case)
      │
      │     NO  → the TS must change
      │            → COORDINATOR, because only the coordinator may change a TS.
      │              An executor doing it is a Role Lock violation
      │
      └─ cannot be satisfied without touching a frozen section?
             → OWNER, through the amendment protocol

Same shape as refinement / amendment / verdict. Not an invention — the transfer of a ruled mechanism onto a second surface.

The same tripwire answers the owner's second question. Revision in place versus a new phase is not a matter of taste:

Test Artifact
Revision repair of what was already specified — needs no TS change rev2 of the RF and the REVIEW; one trace, no new TS
New work cannot be accepted under the existing TS a new phase or a follow-up task, with its own TS

If it needs a TS change, it is not a revision — it is new work wearing a revision's name.

The consequence the owner is looking for: with this rule, "расширяется базовый HL" stops being reachable from inside a revision loop. More work can only be ordered through the coordinator; the goals can only move through the owner.

Open — needs research, not a decision

  1. Loop termination. Nothing counts rounds; AFD reached four. Research has max_passes and loops_per_stage, review has no equivalent. What is the number, and what happens when it is hit — escalate to the owner, or force an APPROVE carrying tech debt?
  2. Re-entry in handoff.md. An executor returning on rev2 today re-runs a full onboarding or improvises. What must a returning executor read, and what may it skip?
  3. Sub-phase proliferation. Were AFD's eight Phase-B splits planned, or revision rounds renamed? The answer decides whether "new phase" is a healthy outlet or a leak.
  4. Does the loop shrink after TFW-53 Phase C? The materiality bar removes wording-only blocks — the documented AFD-48/B failure mode. The revise rate before and after is the number that sizes this entire task.

Deferred to TFW-54 — agent freshness

Owner's intuition: «существующие агенты склонны дрейфовать и переписываться друг с другом бесконечно, отдавая туда-сюда правки». Genuinely unresolved; the evidence pulls both ways.

Fresh agent on rev2 Same agent on rev2
AFD-38: the same reviewer approved, then retracted only under owner pressure A fresh reviewer has not seen rev1's reasoning and re-litigates settled points — the ping-pong gets worse
TFW-53 iter2 self-critique: "hindsight is uncontrolled" when the same agent re-judges its own work AFD-48: the second pass found a real defect, but the first failed on materiality, not on staleness

This is agent topology, it needs measurement, and TFW-54 already carries hypotheses about it. Recorded there, not here.

Why not now, and why not inside TFW-53

Three reasons, in order of weight:

  1. Phase C changes the size of the problem before it can be measured. The materiality bar removes the class of REVISE that produces the longest loops. Designing a protocol for a loop whose frequency is about to change means measuring the wrong thing.
  2. TFW-53 is five phases and thirteen amendments deep, with Phase B unfinished. Adding a sixth surface now is the drift the task exists to prevent. The owner's own recorded preference: «Разделять большие вопросы на отдельные задачи».
  3. It needs the research iteration 1 proved works — a corpus count against AFD's 13 revision arcs and 18 sub-phases, answering how many required a TS change and how many were repairs in place.

Cost of never doing it: under full autonomy a loop with no rule and no limit is TFW-49 repeated from below. There the drift ran through the contract; here it runs through the spec.

Prerequisite

Do not open before TFW-53 Phase C ships and a handful of reviews have run under the Purpose Check. The first deliverable of the research is the before/after revise rate — without it the task is designed against a number that no longer holds.

PROPOSAL — TFW-58: What Happens After a REVISE | 2026-08-13