PROPOSAL — TFW-45 addendum: Review as Consolidator + Delegated Auditors¶
Date: 2026-08-13 Author: Coordinator (Claude Code) Status: ❄️ Addendum to a FROZEN task. Not a plan. Entry point when TFW-45 thaws:
/tfw-plan. Target: TFW-45 Phase B (Review Swarm) — the HL is not edited; this file carries the newer design Sequencing constraint: TFW-53 Phase C — see § Collision with a frozen contract Sibling: TFW-56 removes thecode/docs/specmode axis and frees the term "review mode"
Why this file exists¶
Owner session 2026-08-13 opened three questions about review at once: are the code/docs/spec
modes worth their gate, is the four-stage sequence worth its cost when one agent runs all of it,
and should the main agent become a review consolidator holding goal and value defence while the
rest runs as subagents with their own system prompts and mentality.
Owner decision the same session: the mode question ships alone as TFW-56. The subagent question belongs to TFW-45, which already owns it — so the analysis lands here rather than evaporating.
«хорошо, тогда докидай пропоузал в 45 и свои мысли. здесь оставим только удаление режимов»
TFW-45's HL was written 2026-06-15 and its review half (Phase B) is three lines long:
"Review swarm mode integration (likely review/swarm.md), review.md Step 1 update, system prompt
composition rules for Map/Verify/Judge stage agents." Everything below is newer than that, and two
items contradict it. Nothing here is approved.
The owner's design¶
Main agent = review consolidator, carrying goal defence, value defence "and something else". The remaining work runs as subagents — sequentially or in parallel, where available — each with a concrete system prompt and its own mentality.
This is the same instinct as TFW-45's original review swarm, with one structural addition the original does not have: the coordinator is no longer just a synthesizer of stage files. It holds a check of its own that is not delegable. That changes the decomposition, not just the execution model.
Thought 1 — the current stage set is the wrong thing to delegate¶
TFW-45 §3.1 sketches map_reviewer → verify_reviewer → judge_reviewer + coordinator writes
REVIEW. That is a 1:1 lift of today's stages into subagents. Delegability is not uniform across them:
| Stage | Delegable? | Why |
|---|---|---|
| Verify | Best candidate | Mechanical, by far the most token-expensive (see measurement below), output is a table the consolidator can re-check against the files. Fresh context is an advantage: a verifier that never read the RF's framing is harder to persuade |
| Map | Cheap, yes | Comprehension; output is consumed structurally by later stages |
| Judge | Worst candidate | To rule, it needs the whole reference set — and the reference set is precisely the consolidator's asset |
| Purpose / goal defence | No | See Thought 2 |
Measurement, TFW-53 Phase A review (the most recent complete review in this repo):
| Artifact | Lines | Share |
|---|---|---|
review/map.md |
88 | 26% |
review/verify.md |
179 | 52% |
review/judge.md |
76 | 22% |
→ REVIEW__phase-a__*.md (synthesis of the above) |
198 | — |
So ~540 lines are written to deliver one verdict, and the majority of genuine work sits in one stage. If delegation is about offloading token-heavy mechanical work onto a fresh context, the honest design is not three symmetrical stage agents. It is a consolidator plus one or more auditors, with Map folded in or delegated as a cheap prelude.
This contradicts TFW-45 Phase B as written. It should be treated as a re-decomposition, not as "add a swarm mode file to the existing three stages".
Thought 2 — goal defence cannot be delegated, and the reason is now formal¶
TFW-53 Phase C fixes the Purpose Check's reference set: the master HL at its frozen baseline plus
the project north star, and it declares the TS and any Phase HL invalid references (DoD-20). A
subagent handed TS + RF therefore cannot perform the check — not because it is less capable,
but because it was not given the only admissible reference. Two further properties point the same
way: the check's output is a verdict, not a document, and its scope is the whole result rather
than one dimension of it.
Owner's instinct — consolidator keeps goal defence — is therefore correct, and it now has a structural argument rather than a stylistic one.
One honest counter-consideration, worth recording because it cuts the other way: the consolidator is also the most context-contaminated agent in the run — it read everything, including the executor's own framing. For a critic role that contamination is disqualifying. For the purpose check it is less harmful (the check is anchored to an external, committed baseline, not to independent judgement) — but "less harmful" is an assumption, not a measurement.
Thought 3 — the value of staging splits in two, and only one half is proven¶
| Value | Status | Evidence |
|---|---|---|
| Evidence trail — stage files are the record that verification actually happened | Proven, and independent of who writes them | D41: single-pass review is trust-based with no trail. verify.md carries 52% of stage volume: opened files, executed commands, discrepancies. The artifact survives regardless of agent identity |
| Genuine cognitive-mode separation | Unproven with a single agent | TFW-45 §2 and S1 — the owner's own argument: one agent carrying the whole history "knows" what the previous stage found and builds on it. Performance, not a mode change |
And the sharpest datum, from TFW-53 research iter2: in AFD, judge.md returned ✅ on the very AC
that contained the violation (P8 → AC-B4 → ✅), and the reviewer later retracted his own APPROVE.
Four stages did not catch it. What would have caught it is the reference set. TFW-53 HL §4 Phase C
ranks the levers explicitly: reference-set rule > forcing function > judge.md row > identity
text. Stage count is not on that list.
Implication for TFW-45: the swarm's justification must be stated as evidence-trail quality or independence of the auditor, and never as "more stages = better review". The latter is measurably false in this repository.
Thought 4 — review can parallelize where research cannot¶
TFW-45 correctly refused to parallelize research: Extract without Gather Dimensions is garbage (S3). Review is not the same shape. TFW-45's own table says "Verify benefits from Map's understanding" — benefits, not requires. Verify's inputs (RF claims, TS acceptance criteria, the files themselves) all exist at t0.
So Map ∥ Verify is genuinely available. But TFW-45 P1 is quality over speed, so parallelism must
be justified by something other than wall-clock — e.g. an auditor that is deliberately blind to
Map's narrative. That is a quality argument, and it is testable.
Thought 5 — H1 was never tested, and the cheap harness now exists¶
H1 (a fresh agent with Mindset as system prompt produces genuinely different output than one agent switching modes mid-conversation) is the whole justification for TFW-45, and it has never been run. TFW-45's §10 filter says it plainly: "H1 false → swarm mode has no quality advantage → entire task loses primary justification."
The negative controls did not exist in June 2026 and do now:
| Control | What it is | Where |
|---|---|---|
| TFW-48/49 REVIEWs | Seven verdicts, six ✅ APPROVE, on work the owner later rejected wholesale | git show 721ca15:<path> |
| The overwritten REVISE | One genuine 🔄 REVISE (7 of 10 Judge checks FAIL) overwritten three commits later | 1ebb680 |
| AFD-38/phase-b | ✅ APPROVE that the reviewer himself retracted | AFD repo |
| AFD-48/phase-b | The false-positive precedent — a goal-based REVISE demoted after owner challenge | AFD repo |
And TFW-53 DoD-29 already mandates a replay validation of the Purpose Check against exactly the first two. One harness serves both questions. Running TFW-45's H1 on the same corpus costs the marginal run, not a new experiment.
Recommendation: do not ship a swarm mode on design reasoning alone. It is the one hypothesis in this repository that is both load-bearing and cheap to test, and TFW-53 is already paying for the stand.
Collision with a frozen contract — the reason sequencing matters¶
TFW-53's DoD are frozen, and they nail the Purpose Check to one file:
| Frozen item | What it fixes |
|---|---|
| DoD-20 / 21 / 22 / 23 | The substantive Purpose Check, its citation-and-harm field, its excess test and its third outcome live in templates/review/judge.md |
| DoD-28 | Justifies that site explicitly: the rules land in judge.md "which is a template and not a workflow", keeping review.md inside the F2 word budget |
| HL §4 Phase C | "Enforcement site — templates/review/judge.md, and it is the only one" |
If Judge becomes a delegated subagent while goal defence stays with the consolidator, the enforcement site is wrong: the rule would sit in the prompt of the one agent that must not own it. Three ways out, in preference order:
| # | Path | Cost |
|---|---|---|
| ii | Ship TFW-53 Phase C as frozen, then amend §4/§5 through §12 Amendment Log when this work is planned | One logged, evidenced, ruled amendment — and it dogfoods the mechanism TFW-53 exists to build |
| i | Ship C, then rework judge.md silently in this task | Rework in the same file, and no visible record of why the site moved |
| iii | Absorb this work into Phase C | C is already 🔴 and carries 12 DoD items; this is the scope inflation TFW-53 was written to prevent |
Either way: TFW-53 Phase C lands first. This work is downstream of it, not parallel to it.
What TFW-56 changes for this task¶
TFW-56 deletes the
code/docs/spec axis. Two consequences here:
- Smaller surface.
review.mdloses a step and its WAIT gate;judge.mdloses its Mode-Specific section; three mode files disappear. A swarm redesign touches less. - The term is freed. Today "review mode" means
code/docs/spec. After TFW-56 nothing holds the name. If swarm ships as a mode, the word must carry exactly one meaning (D28: one name = one behaviour) — and TFW-45 H8 already doubted that "mode file" is even the right abstraction for an execution model, as opposed to a parameter set.
Open questions this addendum adds¶
| # | Question |
|---|---|
| Q1 | Is the target consolidator + N auditors (a re-decomposition) or the existing 4 stages, delegated (TFW-45 Phase B as written)? These are different tasks |
| Q2 | Does Map survive as its own artifact, or fold into Verify's preamble? 26% of stage volume, and its consumer is the next stage rather than the owner |
| Q3 | Where does the Purpose Check live once Judge is delegated — templates/REVIEW.md, a new consolidator-owned stage file, or the workflow itself? DoD-28's word-budget argument must be re-answered, not ignored |
| Q4 | Does an auditor deliberately blind to Map produce better verification than one primed by it? This is the parallelism question stated as a quality question |
| Q5 | Degradation contract: what exactly does a tool without spawn run? TFW-45 DoD-5 demands zero regression for single-agent users, and that promise now has to cover a re-decomposition, not just a mode switch |
| Q6 | Does the consolidator's context contamination hurt the purpose check? Assumed harmless in Thought 2 — unmeasured |
Hypotheses added to TFW-45's set¶
| # | Hypothesis |
|---|---|
| H10 | The delegable work is asymmetric: Verify carries the majority of mechanical volume, Judge and the purpose check are not honestly delegable, so a consolidator + auditor topology beats a 1:1 stage-to-subagent lift |
| H11 | A subagent given only TS + RF structurally cannot perform the Purpose Check, because TFW-53 DoD-20 makes both invalid references — so goal defence stays with the consolidator by construction, not by preference |
| H12 | Map ∥ Verify is safe in review (unlike Gather → Extract in research), and a Map-blind auditor produces better verification, not merely faster |
| H13 | The consolidator's full-context contamination is acceptable for the purpose check specifically, because the check is anchored to an external committed baseline rather than to independent judgement |
| H14 | The stage-file evidence trail (D41) retains its full value under delegation — the artifact matters, the author does not |
Strategic insights carried over¶
| # | Insight | Category | Source |
|---|---|---|---|
| S9 | Owner separates the two review questions deliberately: mode deletion ships alone and immediately; the subagent re-architecture waits. Small reversible cleanup is not held hostage to a large unproven redesign | process | User, 2026-08-13 |
| S10 | Owner's consolidator formulation places goal and value defence at the top of the review, not inside a stage: «основному быть агентом ревью-консолидатором с защитой целей ценностей». That is an ordering claim — the purpose check outranks the quality checks, rather than sitting beside them | philosophy | User, 2026-08-13 |
| S11 | Owner explicitly does not want subagents forced: «мы не форсируем использование субагентов». So the design must be genuinely optional, and the single-agent path stays a first-class citizen rather than a degraded fallback | constraint | User, 2026-08-13 |
| S12 | Owner's phrasing — «с конкретными системным промптом и ментальностью» — keeps the June 2026 premise intact: identity via system prompt, not instruction. The premise is unchanged and still untested (H1) | philosophy | User, 2026-08-13 |
Prerequisite¶
Do not start before TFW-53 Phase C is complete, and read TFW-56's RF first — both change the files this work would touch. Running before either means designing a swarm around a review flow that is about to change twice.
PROPOSAL — TFW-45 addendum: Review as Consolidator + Delegated Auditors | 2026-08-13