amicus 4.5.4 → 4.6.1

package/skills/second-opinion/MODEL-NOTES.md

@@ -4,8 +4,10 @@ This file is the `second-opinion` skill's evolving memory of **how to actually d
 well**. Read it before Stage 0 (council selection and launch); update it, with the user's
 approval, at the end of each run (Stage 6). Keep it tight — merge and prune rather than append.
 
-_Last updated: 2026-07-26 (v4.4.0 fold-back: alias-resolution hygiene, `council run --run-id`,
-debate defense/re-vote waves now exercised, haiku and glm notes; see changelog)._
+_Last updated: 2026-08-03 (per-section fold-back, both directions — the haiku "hard-404"
+re-diagnosed as a `/v1`-less `ANTHROPIC_BASE_URL`, pre-degrade-era claims re-grounded in the
+announcement contract (`degrades[]`/`seatLoss`/exit 2/one Stage-1 retry), three model sections and
+the peer-consensus≠evidence rule upstreamed from the field ledger; see changelog)._
 
 ## Global operating rules (all models)
 - **Fast path:** `council run` applies `--agent Plan` / `--no-context` / `--summary-length
@@ -22,14 +24,33 @@ debate defense/re-vote waves now exercised, haiku and glm notes; see changelog).
   defined in SKILL.md Stage 0) — no size cap, no shell-quoting hazards. Never inline a briefing as
   a CLI argument.
 - **Run in the background (`run_in_background: true`); you're notified on completion. Don't poll.**
+- **A stale MCP server kills councils with a distinctive signature — zero output on every leg.**
+  If any amicus MCP tool result carries the "Amicus was upgraded on disk (running vX, on-disk vY).
+  Restart your MCP client" notice, do NOT launch councils through the MCP tools — relaunch via the
+  CLI (`amicus council run`, a fresh process) or restart the client first. A version-skewed server
+  has timed out an entire Stage-1 wave with literally zero output bytes on every leg; that
+  all-legs-zero-output shape is what distinguishes it from model/provider failures, which produce
+  partial output or errors.
 - **Read results from the JSON documents** (`--json`): a wave's `legs[].summary` / a run's
   `summary` is the model's output; `status`/`error`/`counts` are ground truth for failures. Never
   scrape stderr logs to judge success.
 - **Transient provider errors** (502s, connection drops): re-run the affected leg (solo
   `amicus start --json`, same briefing file) or the wave — see per-model notes for
   model-specific signals. Never present a half-finished run as an answer.
+- **Seat losses are announced, retried once, and exit-coded — read the surfaces, don't diff
+  finding counts.** Every lost seat/leg is announced in one voice on stderr and recorded in
+  `run.json.degrades[]`, `verdict.json.degrades[]`, the report's **"What was lost"** section, and
+  `verdict.seatLoss`, and the run exits degraded (2). A Stage-1 wave or leg that dies is relaunched exactly once (serially, after the
+  surviving launches settle; skipped when the run is already over `--max-cost`): a heal announces
+  as a `Recovered:` line and the run stays exit 0; a seat still dead after its retry is recorded
+  with both attempts named in the why. A dead route still doesn't stop a council — it shrinks it —
+  but the shrinkage is no longer silent: confirm the bench you paid for from
+  `seatLoss`/`degrades[]` after every run, because a shrunken bench weakens the tiers (fewer
+  corroborators per finding).
 - **Credentials:** keys live in `~/.config/amicus/.env`. The legacy `~/.config/sidecar/.env`
-  fallback was removed in v2.0.0 (see `docs/SHIMS.md`). Configure with `amicus setup`.
+  fallback was removed in v2.0.0 (see `docs/SHIMS.md`). Configure with `amicus setup`. If
+  `ANTHROPIC_BASE_URL` is set anywhere in the environment, it must carry its `/v1` suffix for
+  direct-Anthropic legs — see the haiku section for the full diagnosis.
 - **PowerShell `--models` quoting (Windows):** always quote comma-separated model lists —
   `--models "gemini,gpt,deepseek"`. Unquoted, PowerShell splits on commas and amicus receives one
   mangled alias → instant arg-parse failure. (Now baked into every SKILL.md example.)
@@ -40,18 +61,50 @@ debate defense/re-vote waves now exercised, haiku and glm notes; see changelog).
   gpt/deepseek/grok have handled 82k-word agentic reads; gemini(-flash) and kimi stalled (narrate-
   stall / 25-min timeout / poller "Incomplete"). Pre-select proven long-read models or inline the
   text for large-context models.
+- **Long agentic reads can "narrate-then-never-deliver" even on capable models that DO finish
+  reading — bake the anti-narration hardening into the FIRST attempt, not the retry.** Models have
+  done a genuine full chunked read (confirmed by realistic input-token counts) and then ended the
+  turn with only process narration ("Continuing…"), never producing the review + JSON: the read
+  completed, but the turn budget ran out before the switch from reading mode to writing mode.
+  Whenever the task requires many sequential file-read tool calls before writing, the briefing
+  needs an explicit preamble: *"this is headless with no follow-up turn; you must produce the
+  complete deliverable by the end of THIS response; if running low on room, stop reading early and
+  write from what you have rather than deliver nothing."*
+- **Multi-FILE agentic reads stub out like long single reads — even WITH the anti-narration
+  preamble baked in.** Weaker readers have returned narration stubs of roughly 60–100 chars on a
+  multi-file packet despite the standard preamble. Treat a many-file briefing as a long-read task:
+  expect stubs, retry solo once with the identical briefing, then substitute the model — or
+  pre-concatenate the packet into one file for weaker readers.
+- **`--agent Plan` can trigger a literal, hard "planning mode — cannot execute" refusal on some
+  models** (terse, near-zero-token responses), not just the intended read-only tool restriction —
+  and prompt wording alone has not fixed it; switching that model to `--agent Build` has, immediately.
+  On the manual path, Stages 2-3 never touch the source file (the briefing is self-contained plus a
+  no-tools preamble), so `--agent Build` is safe there; reserve `--agent Plan` for Stage-1 legs
+  that read the source, and fall back to Build per-model on a hard refusal.
 - **Debate mode rarely fires on correctness questions — design for it or expect a no-op.**
   `--debate` only engages findings the tally marked Contested or Disputed, and judges agree far
   more than expected: **2 disputes in 123 adjudications across four councils** (gemini/gpt/qwen).
   Three of those four runs produced zero debatable findings, so the round never ran. Bug hunts
   converge — a race condition either exists or it does not. If you want a rebuttal round, brief a
   genuine judgement call (architecture, tradeoffs, "is this over-engineered"), not a defect hunt.
+- **Peer consensus is not evidence on published, checkable numbers — verify against a primary
+  source before applying such a finding, especially from the Confirmed tier and findings that
+  survived debate.** In one paid council, 2 of 23 findings were factually wrong about published
+  specs and BOTH sat in the tiers that are supposed to signal reliability: one reached Confirmed
+  on two peer agreements, the other hardened after a debate round in which a correct dispute was
+  talked out of its objection. A rebuttal is a rhetorical performance — judges re-voting on a
+  persuasive defense have no more access to the spec sheet than they did the first time. Any
+  finding that asserts a published figure (spec, capacity, rating, dimension, date) gets a
+  verification pass before you act on it.
 - **Read spend from `run.json`, never from a provider credit balance.** `usage.cost.amount` is the
   run total; per-leg costs are in `runStats[].usage.cost`. Only OpenRouter-routed legs move the
   OpenRouter balance — gemini/gpt/anthropic bill directly against their own keys, so inferring cost
   from that balance under-reports it badly (observed: ~6x low). A 3-model bench + chair + debate is
   roughly **$0.60-0.80 per run**, not cents; budget `--max-cost` accordingly or the chair gets
-  skipped mid-run (exit 2, degraded) when the debate legs push the total past the ceiling.
+  skipped mid-run (exit 2, degraded) when the debate legs push the total past the ceiling. As of
+  v4.6, `runStats` also carries Stage-2 judge rows (judge-tagged), so totals read higher than
+  pre-4.6 runs for the same bench; anything keying `runStats` by model should exclude
+  `role: 'judge'`.
 - **Expect agreement inflation in Stage-2 adjudication.** The judge contract defines `agree` by
   worked example ("an 'I missed this — it's valid' counts as agree") but gives no example for
   `dispute` and no positive definition of `neutral`, while requiring a verdict on EVERY finding —
@@ -105,7 +158,9 @@ debate defense/re-vote waves now exercised, haiku and glm notes; see changelog).
   **rescope** mechanism rather than a defense — raisers overwhelmingly AMEND (downgrade an
   overstated severity, narrow a scope) rather than DEFEND, and amendments are re-confirmed on the
   re-vote. Expect high-amend / low-defend; the idea usually survives, the severity claim often does
-  not.
+  not. In ideation councils, **cross-lens convergence** (independent lenses proposing the same
+  idea) is the strongest priority signal — compute it yourself when clustering; it is
+  complementary to, not the same as, the tally's agree/dispute tiers.
 
 ## Per-model notes
 
@@ -118,61 +173,115 @@ debate defense/re-vote waves now exercised, haiku and glm notes; see changelog).
 - **Red-team:** takes an adversarial brief well — high variance by design; use when consensus risk is high.
 - **Blind self-votes are inconsistent** (self-#1 in some runs, self-last in others) → discount self-votes either way.
 - Alias has resolved to **flash** tiers: fast, shallowest coverage, yet a recurring sharp fact/consistency checker (it alone refuted a bench-wide date error). Cheap cross-check value.
+- **Asserts quantitative "corrections" (timings, capacities) with unearned confidence** — one such
+  "severely wrong" correction was disputed 3-0, including its own blind self-dispute. Cross-check
+  its numeric claims before weighting; its honest blind self-disputes are a useful calibration tell.
 - **Unreliable on long agentic reads** (see global rule) — inline the text or swap models for book-length material.
 
 ### DeepSeek  (`--model deepseek` → via OpenRouter)
 - **Strengths:** resilient; produces strong, well-structured, well-cited critical analysis. A good
   default reviewer and a proven chair.
 - **Quirk:** occasional transient 502 mid-run → re-run the leg.
-- Proven chair (5 clean chairings) — decisive, well-structured synthesis.
+- Proven chair (many clean chairings) — decisive, well-structured synthesis.
 - As a Stage-1 reviewer of human-facing documents it **over-escalates severity** (typos/tenure → "blocker"); discount its blocker labels against peers. Prune its self-retractions when tallying.
 - **"Agree-with-the-adversary" lean:** it has been the lone endorser of a red-team's harshest claims, turning them Contested — cross-check before treating its lone agreements as consensus.
+- **Stub-on-fanout / clean-on-solo-retry:** its Stage-1 fanout leg has returned a sub-100-char
+  narration stub despite the anti-narration preamble, then produced a full review on a plain solo
+  retry with the identical briefing — retry solo before substituting the model.
 
 ### GPT  (`--model gpt` → via OpenRouter)
 - **Strengths:** reachable via the OpenRouter key; resilient; very thorough structured critique
   (25 findings on a 1-page framework). Cleanly separates the review criteria.
 - **Quirks:** verbose — peers dinged it for volume-over-judgment (good coverage, lower
   discrimination); **self-ranked its own review #1** in cross-review → discount self-votes.
-- Ranked genuine #1 by all judges (incl. non-self) in one run — thoroughness is real, not a self-vote artifact. Handled an 82k-word agentic read cleanly.
+- Ranked genuine #1 by all judges (incl. non-self) in one run — thoroughness is real, not a self-vote artifact. Handled an 82k-word agentic read cleanly — but a later similar-size read delivered only narration until retried with the anti-narration preamble; apply the preamble to gpt by default, not just to weaker models.
 - **Asserts context-dependent facts (dates, "is this future?") without verifying** — and self-confirms them in adjudication. Cross-check any time-dependent claim it raises.
-- A good calibration anchor in cross-review: confirms observational findings, disputes interpretive overreach.
+- A good calibration anchor in cross-review: confirms observational findings, disputes interpretive overreach. A repeatable role: the seat most likely to catch **confidence inflation** in an otherwise-correct argument (theoretical risk asserted as demonstrated mechanism, proxies used as pseudo-diagnostics).
+- **When gpt disputes a specific numeric claim, weight that dispute heavily before a debate round erodes it** — it has been right and then talked out of the objection by a persuasive rebuttal (see the peer-consensus rule).
 
 ### Grok  (`--model grok` → via OpenRouter)
 - Very fast legs; credible judge and chair (rejected its own weak findings as chair; honest blind self-rank).
 - Strong red-team fit; handled an 82k-word agentic read. Weight its **observational** catches heavily and its **interpretive** verdicts cautiously (bench pattern: the former confirmed, the latter disputed).
 - Stage-1 non-red-team reviews skew to scope-inflated "missing content" majors.
+- Has **hard-refused under `--agent Plan`** ("Plan mode active… cannot execute"; near-zero tokens)
+  where `--agent Build` worked immediately — prefer Build for grok legs (see the global Plan-refusal rule).
+- **Fabrication risk on long-document tasks:** one fast, suspiciously low-fresh-token response
+  contained an invented "major"-severity finding plus unsupported flavor details, verified false
+  against the source. A later run under Build (genuine read, no fabrications on spot-check)
+  partially rebuilt confidence — usable for red-team slots under Build **with the spot-check
+  discipline**: verify its most specific/surprising claims against the source before letting them
+  into a council.
 
 ### Kimi  (`--model kimi` → via OpenRouter)
 - The bench's sharpest adjudicator (caught strawmen and misreads other judges waved through).
-- **Very slow legs (5-7 min)** — it gates wave wall-clock; budget timeouts around it.
-- Stalls on long agentic reads (poller "Incomplete" with only a preamble). Reserve for short-artifact adjudication.
+- Repeatedly the **top-ranked reviewer on consumer-practical, safety-heavy artifacts** (food
+  safety, mechanical advice) — top-ranked in two such runs (unanimous in one), with the most
+  granular findings on the bench and genuine structural catches.
+- Its specificity is also its risk: it reaches for **exhaustive quantitative claims ("every",
+  "all", "none") that are directionally right and literally wrong** — and it has successfully
+  DEFENDED one in a debate round against a correct dispute. Verify its universal quantifiers
+  against a primary source (see the peer-consensus rule).
+- **Very slow legs (4-8× its peers)** — it gates wave wall-clock; budget timeouts around it.
+- Stalls on long agentic reads (poller "Incomplete" with only a preamble). Reserve for short-artifact work.
 
 ### Mistral  (`--model mistral` → via OpenRouter)
 - Fast, broad coverage, catches real issues.
 - **Hallucination risk is real:** has invented non-existent product models/specs, disputed independently by two judges. Cross-check every specific model number or product claim it introduces.
 
 ### Claude  (in-council, when toggle on)
-- Consistently the most *calibrated* reviewer (no severity inflation; findings overwhelmingly Confirmed; bench-best street-cred in recent runs) but sometimes the least *original* — it can miss the boldest single catch. Treat as a reliability floor, not a discovery engine.
-
-### haiku  (`--model haiku`) — **verify before using; it has been hard-404ing**
-- The direct-Anthropic route (`anthropic/claude-haiku-4-5-20251001`) returned a hard `Not Found`
-  on **every** invocation of a recent paid corpus — 3 of 3 legs across two separate runs, both as
-  **chair** (twice, incl. the fallback retry) and as a **bench seat** — in ~2 s with zero tokens.
-- Both runs degraded *around* it silently rather than failing: one fell back to another chair, the
-  other collapsed its bench from 3 seats to 2 and exited 2. **A dead alias does not stop a council;
-  it shrinks it.** In the bench-seat case every finding came out `confidence: "thin"` with a single
-  peer corroborator, purely because the bench had halved — and nothing in `verdict.json` said so.
-- Resolve the alias against the catalog (`amicus models --check`) before putting it on a paid
-  bench, and re-check the bench roster in `run.json` against what you asked for afterwards.
-
-### GLM  (`--model glm` → z-ai via OpenRouter)
-- Cheap and fast, and ranked best-by-peers on a clean debut — but a later run produced **35 KB of
-  prose with `conformance: unstructured` and 0 parsed findings**, twice. Its structured-output
-  reliability is **not** established; treat the debut as low-N.
-- Useful behavioural note: when its structured output failed it **refused to fabricate** on the
-  repair attempts. Honest — but an honest refusal still costs you the seat, so a bench that leans
-  on `glm` for quorum can silently adjudicate a seat short while still paying for its tokens.
-  Watch `conformance` per seat, not just the finding count.
+- Consistently the most *calibrated* reviewer (no severity inflation; findings overwhelmingly Confirmed; bench-best street-cred in recent runs) but sometimes the least *original* — it can miss the boldest single catch. Treat as a reliability floor, not a discovery engine — though on some benches its findings have anchored the entire Confirmed tier.
+- Its own checkable, arithmetic-style claims deserve the same **mechanical verification** as everyone else's: one calendar-consistency claim in a Claude first-pass review was a genuine reasoning error, caught only by computing the dates programmatically before submission (see General).
+
+### haiku  (`--model haiku`) — **the "hard-404" was the environment, not the model**
+- **The model was never the problem.** The direct-Anthropic route returned an instant `Not Found`
+  (~2 s, zero tokens) on every invocation of a paid corpus — 3 of 3 legs across two runs, as chair
+  and as bench seat — and the standing diagnosis was a rotten alias. It is not: the cause is an
+  **`ANTHROPIC_BASE_URL` set in host form, without its `/v1` suffix**. Anthropic SDKs (including
+  Claude Code itself) treat the var as a HOST and append `/v1` themselves; OpenCode's provider
+  layer treats it as the full prefix — so a value that is correct for the host app kills every
+  direct-Anthropic leg amicus launches. Proven by control pair: an identical `fanout --models
+  opus` call fails "Not Found" on the host form and completes with `/v1` appended.
+- **Symptom signature:** instant "Not Found", zero tokens, direct-Anthropic routes only —
+  OpenRouter-routed legs in the same run are unaffected. It hits every direct-Anthropic alias
+  equally (haiku, opus, sonnet, claude; `fable` is OpenRouter-only and unaffected), so a "dead"
+  cheap seat and a "dead" frontier chair with this signature share one cause.
+- **The check:** inspect `ANTHROPIC_BASE_URL` in the environment amicus actually runs in — the var
+  can live only in a parent process's env (e.g. the host app), absent from every shell profile and
+  settings file on disk. If it lacks `/v1`, that is the kill. A doctor check for this is filed on
+  the backlog; until it ships, check by hand before blaming a model or an alias.
+- Seat-loss mechanics for a dead route are the same as any other loss — announced in one voice,
+  exit 2 if still dead. A dead Stage-1 seat gets the one retry; a dead CHAIR walks the chair's own
+  chain instead — same-chair retry, then promotion of the best non-bench model from the ledger —
+  with the actual chair checkpointed into `run.json` (see the seat-loss bullet in Global
+  operating rules).
+
+### GLM  (`--model glm` → `glm-5.1` via OpenRouter; the recent observations below are from the explicit `openrouter/z-ai/glm-5.2` id)
+- The v4.4.0-era "structured-output reliability not established" warning is **withdrawn as
+  wrong-cause**: its `unstructured`-conformance results and repair refusals traced to two
+  since-fixed engine defects (the unanchored fence extractor that truncated any JSON quoting a
+  code fence, and repair prompts that omitted the artifact under repair — both fixed in v4.4.1),
+  not to the model. Conformance has been `clean` in subsequent paid runs.
+- Cheap and fast; **ranked best-by-peers on an ideation-bench debut** — a focused,
+  fewest-findings reviewer whose findings land. Promising budget-to-mid bench member.
+- **First recorded confident fact error:** it asserted an engine spec that the primary source (the
+  owner's manual) contradicts outright — and the finding reached Confirmed on two peer agreements,
+  denied only by a post-council verification pass. Treat glm's confident factual assertions as
+  unverified until checked (the peer-consensus rule exists because of findings like this one).
+- Honest under pressure both ways: it refused to fabricate on repair attempts, and it has
+  withdrawn a contested finding cleanly in debate. An honest refusal still costs the seat — watch
+  `conformance` per seat, not just the finding count.
+
+### Qwen  (`--model qwen` → qwen3.7-max via OpenRouter; distinct from `qwen-coder`)
+- Very large context (1M tokens per catalog). As a red-team substitute it has produced a thorough,
+  well-organized adversarial review with accurate, specific line citations and genuinely unique
+  catches that verified true against the source — weight its specific, cited claims heavily.
+- Same observational-vs-interpretive split as grok: its cited observational findings get
+  confirmed; its "blocker"-severity interpretive claims get disputed as genre-normative. Weight
+  its severity labels on broad interpretive claims cautiously.
+- **Do not assign qwen Stage-1 multi-file reads.** Its proven mode is single-file reads and
+  short-artifact work: a book-length single-file read succeeded on retry with the anti-narration
+  preamble, but multi-file packets have produced narration stubs twice in one run — with the
+  preamble present, under both Plan and Build. It narrates rather than refuses under Plan.
 
 ### minimax  (`--model minimax` → via OpenRouter)
 - Fast (~2 min review legs), cheap, `clean` findings-JSON conformance on debut.
@@ -189,6 +298,21 @@ debate defense/re-vote waves now exercised, haiku and glm notes; see changelog).
 - Ranked last on its debut bench — misses the offline, interruption-handling, and
   test-methodology gap classes stronger seats catch. Fine budget-bench filler; do not chair it.
 
+### GPT-5.6-Terra  (`openrouter/openai/gpt-5.6-terra`)
+- Debut (one paid ideation run): `clean` conformance, thorough — the most findings on its bench.
+  Handled an inlined ~13K-token digest cleanly under `--agent Build`.
+- The priciest leg on its bench (~5× the cheap seats) and it self-ranked #1 (discount self-votes,
+  as always) while peers ranked it lower — thoroughness did not convert to peer standing. Low-N;
+  re-confirm before leaning on it.
+
+### Gemini 3.1 Pro  (`openrouter/google/gemini-3.1-pro-preview` — the PRO tier; the bare `gemini` alias resolves to flash)
+- Debut (one paid ideation run): `clean` conformance, mid-bench peer standing.
+  No narration or stall — but the material was INLINED, not an agentic read (where gemini-flash
+  historically stalls); don't extend the result to agentic reads untested.
+- Use the shipped `gemini-pro` alias when you want a real Pro reviewer — it live-resolves to the
+  current Pro tier, falling back to `openrouter/google/gemini-3.1-pro-preview`; the bare `gemini`
+  alias gives you flash.
+
 ### (others — add as used)
 - Opus / o-series etc. are reachable via amicus **if their API keys are configured**. Add notes
   here the first time each is used.
@@ -209,7 +333,7 @@ This section keeps only per-model **qualitative quirks** and **structural-confor
 
 - **deepseek** — strong synthesis, resilient; occasional transient 502 → re-run the leg. Proven chair. Conforms cleanly.
 - **gpt** — thorough but verbose; peers have dinged it for volume-over-judgment. Self-ranked its own review #1 in the 2026-06-04 run → the peers-only street-cred rule (now enforced by `tally`) mitigates this. Conforms cleanly. Accessible via OpenRouter.
-- **gemini** — fast, very large context; tends toward absolute severity labels ("blocker" inflation vs peers). Conforms cleanly; watch for preamble narration — instruct it to emit the JSON block verbatim after the prose.
+- **gemini** — fast, very large context; the bare alias resolves to flash tiers (shallow coverage, cheap fact-checks); asserts numeric "corrections" with unearned confidence — cross-check them. Conforms cleanly; watch for preamble narration — instruct it to emit the JSON block verbatim after the prose.
 
 ## Free-tier models (OpenRouter `:free`)
 - Heavily rate-limited (shared daily pool); a 3-leg parallel wave + cross-review can 429 mid-run.
@@ -229,6 +353,12 @@ This section keeps only per-model **qualitative quirks** and **structural-confor
 ## General
 - Model citations are usually real but **verify any load-bearing reference before publishing**;
   watch for loosely-attached attributions (e.g., a real paper cited for the wrong claim).
+- **Checkable, arithmetic-style claims (calendar/date consistency, counts, named specific facts)
+  deserve a mechanical verification pass before publishing — including claims from Claude's own
+  review, not just other models'.** One calendar-consistency claim in a Claude first-pass review
+  was wrong; running the actual dates through a real calendar caught it before it reached the
+  council or the author. This class of claim is cheap to verify mechanically and expensive to get
+  wrong in a document going to a human — don't re-reason it a second time; compute it.
 - Prefer models from **different families** for genuinely independent opinions.
 
 ## Lessons changelog
@@ -237,7 +367,7 @@ This section keeps only per-model **qualitative quirks** and **structural-confor
   its occasional transient 502.
 - **2026-06-03** — v2 council upgrade: added cross-review (Stage-2 anonymized peer ranking +
   per-finding adjudication) and reviewer-reliability tracking.
-- **2026-06-04** — Trusst messaging-framework council (Gemini + GPT + DeepSeek + Claude-in-council;
+- **2026-06-04** — B2B messaging-framework council (Gemini + GPT + DeepSeek + Claude-in-council;
   DeepSeek chair). First GPT use → per-model note. First scored reviewer-reliability rows
   (deepseek 2.33/100%, gpt 2.67/92%, gemini 3.67/89%; 1–4 scale).
 - **2026-06-10** — v3 migration: transport moved to `fanout --json` + `--prompt-file` (F4); pruned
@@ -254,11 +384,28 @@ This section keeps only per-model **qualitative quirks** and **structural-confor
   a bench seat; `council run` prints nothing until terminal, so pin `--run-id` when backgrounding.
   Debate's defense/re-vote waves exercised for the first time — they work, and behave as a rescope
   (amend-heavy) rather than a defense. New per-model notes: **haiku** hard-404ed 3/3 legs across two
-  runs and both councils silently degraded around it; **glm** returned `unstructured` conformance
+  runs and both councils silently degraded around it *(both halves since superseded: the 404 was a
+  `/v1`-less `ANTHROPIC_BASE_URL`, not the alias — see the haiku section — and the silent-degrade
+  era ended with the v4.5.2→v4.6 announcement contract)*; **glm** returned `unstructured` conformance
   with 0 findings twice after a clean debut, and honestly refused to fabricate on repair — which
-  still costs the seat.
+  still costs the seat *(since re-diagnosed: both results were v4.4.0 engine defects, fixed in
+  v4.4.1 — see the GLM section)*.
 - **2026-07-14 (v2.2.0)** — Optional council elements shipped and verified on a planted-flaw
   ground-truth council (critic seat, debate mode nothing-to-debate path, chair verdict scale;
   expert lenses defined but not yet field-run). New lessons: claim-class dedup glosses
   rationale-level errors in Stage-2; minimax debut (strong critic seat, CJK-intrusion quirk);
   qwen-coder debut (fast budget filler). Debate mode's defense/re-vote waves still unexercised.
+- **2026-08-03 (fold-back, both directions — owner's ruling)** — Per-section reconciliation with
+  the machine-local field ledger after six release-cycle deferrals. Corrections: the haiku
+  "hard-404" re-diagnosed (a `/v1`-less `ANTHROPIC_BASE_URL` in host form — the model was never
+  the problem; control pair: identical `fanout --models opus` fails host-form, completes with
+  `/v1`); glm's `unstructured` era re-attributed to the since-fixed v4.4.1 fence-extractor and
+  repair-prompt defects; every pre-degrade-era "silent shrink" claim re-grounded in the current
+  contract (one-voice announcement on stderr + `degrades[]` + `seatLoss`, exit 2, one Stage-1
+  retry with `Recovered:` heals). Upstreamed from the field: the peer-consensus≠evidence rule
+  (2 of 23 findings factually wrong inside high-trust tiers, caught only by primary-source
+  verification), the anti-narration-preamble and multi-file-stub rules, the Plan-hard-refusal
+  fallback, the stale-MCP-server signature, three model sections (qwen, gpt-5.6-terra,
+  gemini-3.1-pro), and gemini/deepseek/gpt/grok/kimi/Claude enrichments. Standing practice: each
+  release cherry-picks generalizable lessons per-section (docs/publishing.md release checklist),
+  never a bulk copy.

package/src/cli-council-run-render.js

@@ -0,0 +1,51 @@
+// src/cli-council-run-render.js
+'use strict';
+
+/**
+ * @module cli-council-run-render
+ * Human-readable render of a finished `amicus council run` (the non-`--json`
+ * output path, v4.0 spec §4). Split out of cli-handlers-council-run.js (v4.6
+ * Plan 4 Task 2): that file sat at 298/300 lines and Plan 4 Task 3 adds the
+ * #81 Workspace notice, which needed the room. `renderRunHuman` is
+ * self-contained — no dependency on the handler's flag-parsing or engine
+ * dispatch — so it moves verbatim to its own leaf; the old home
+ * (cli-handlers-council-run.js) requires it back and re-exports it, so every
+ * existing import path keeps working unchanged.
+ */
+
+function renderRunHuman(run) {
+  const lines = [
+    `Council run ${run.runId}: ${run.status} (exit ${run.exitCode})`,
+    `  bench: ${(run.bench || []).join(', ')}  chair: ${run.chair}`,
+    `  dir:   ${run.options && run.options.outDir}`,
+  ];
+  // v4.4: a cost line that omits unpriced legs reads as the whole bill. The
+  // diagnosis measured council-wsgate02 printing $0.3720 for a run that really
+  // spent $0.9859. Say what we know, then say what we cannot know — and print
+  // the line even when NOTHING resolved (the old `typeof amount === 'number'`
+  // guard silently dropped it, so a fully unpriced run looked free).
+  const u = run.usage || null;
+  const unknownLegs = u && typeof u.unknownLegs === 'number'
+    ? u.unknownLegs
+    : (u && u.cost && u.cost.unpricedLegs) || 0;
+  // v4.4 Task 2: a fully-priced run can still be short. `council-wsgate01`
+  // printed an unqualified $0.2821 for a run that really spent $0.3036 — every
+  // leg `reported`, and 100% of the gap one unattributed `explore` child session.
+  const subtreeLegs = u && typeof u.subtreeUnknownLegs === 'number'
+    ? u.subtreeUnknownLegs
+    : (u && u.cost && u.cost.subtreeUnknownLegs) || 0;
+  if (u && u.cost && (typeof u.cost.amount === 'number' || unknownLegs > 0 || subtreeLegs > 0)) {
+    const known = typeof u.cost.amount === 'number' ? `$${u.cost.amount.toFixed(4)}` : '$0.0000';
+    const gaps = [];
+    if (unknownLegs > 0) { gaps.push(`${unknownLegs} leg(s) unknown`); }
+    if (subtreeLegs > 0) { gaps.push(`${subtreeLegs} leg(s) with unattributed subagent child-session spend`); }
+    const tail = gaps.length > 0
+      ? ` + ${gaps.join(' + ')} — real spend is at least this much`
+      : '';
+    lines.push(`  cost:  ${known} (${u.cost.source})${tail}`);
+  }
+  if (run.error) { lines.push(`  error: ${run.error.code}: ${run.error.message}`); }
+  return lines.join('\n') + '\n';
+}
+
+module.exports = { renderRunHuman };

package/src/cli-handlers-council-run.js

@@ -13,6 +13,10 @@ const path = require('path');
 const { failJson, buildErrorDoc, ERROR_CODES } = require('./utils/error-doc');
 const { validateTaskId } = require('./utils/validators');
 const { GATEWAY_MODES } = require('./utils/model-descriptor');
+// v4.6 Plan 4 Task 2: renderRunHuman moved to its own leaf (size gate); this
+// file re-exports it below so every existing require() of this path still
+// resolves it unchanged.
+const { renderRunHuman } = require('./cli-council-run-render');
 
 const CHAIR_DEFAULT = 'deepseek';
 
@@ -40,9 +44,9 @@ function sanitizeCouncilName(name) {
 
 /**
  * Resolve bench models from --models XOR --council (mirrors handleFanout).
- * Also returns `presetName` (v4.3 Task 3, spec §7.1): the trimmed --council
- * name when that branch was taken, else null — threaded into runCouncil's
- * `councilName` option so council ledger rows can be attributed to a preset.
+ * Also returns `presetName` (v4.3 Task 3, spec §7.1: trimmed --council name,
+ * else null) and `droppedMembers`: a preset's own drops, or — bare --models —
+ * the parsed `--dropped-members` MCP→child passthrough (v4.6 Plan 4 Task 4b).
  */
 function resolveBench(args, useJson) {
   const hasModels = typeof args.models === 'string' && args.models.trim();
@@ -68,54 +72,40 @@ function resolveBench(args, useJson) {
     if (expanded.error) {
       return { fail: failJson(useJson, { code: ERROR_CODES.BAD_ARGS, message: `Error: ${expanded.error}` }) };
     }
-    if (expanded.dropped && expanded.dropped.length && !useJson) {
-      process.stderr.write(`Notice: dropped unavailable council member(s): ${expanded.dropped.join(', ')}\n`);
-    }
-    // v4.5 Wave 2: threaded into runCouncil's options — the ONLY prior signal
-    // was the stderr-only Notice above, which --json mode never even prints.
+    // v4.5 Wave 2 → Plan 4 Task 4: threaded into runCouncil's options — the
+    // sink now announces each dropped member, with reason, on every transport and surface.
     return { bench: expanded.models, presetName, droppedMembers: expanded.droppedMembers || [] };
   }
-  return { bench: parseList(args.models), presetName: null, droppedMembers: [] };
+  if (args['dropped-members'] === undefined) {
+    return { bench: parseList(args.models), presetName: null, droppedMembers: [] };
+  }
+  let dm; try { dm = JSON.parse(args['dropped-members']); } catch { dm = null; }
+  if (!Array.isArray(dm) || !dm.every(d => d && typeof d.member === 'string' && typeof d.reason === 'string')) {
+    return { fail: failJson(useJson, { code: ERROR_CODES.BAD_ARGS,
+      message: 'Error: --dropped-members must be a JSON array of {member, reason} entries' }) };
+  }
+  return { bench: parseList(args.models), presetName: null, droppedMembers: dm };
 }
 
-function renderRunHuman(run) {
-  const lines = [
-    `Council run ${run.runId}: ${run.status} (exit ${run.exitCode})`,
-    `  bench: ${(run.bench || []).join(', ')}  chair: ${run.chair}`,
-    `  dir:   ${run.options && run.options.outDir}`,
-  ];
-  // v4.4: a cost line that omits unpriced legs reads as the whole bill. The
-  // diagnosis measured council-wsgate02 printing $0.3720 for a run that really
-  // spent $0.9859. Say what we know, then say what we cannot know — and print
-  // the line even when NOTHING resolved (the old `typeof amount === 'number'`
-  // guard silently dropped it, so a fully unpriced run looked free).
-  const u = run.usage || null;
-  const unknownLegs = u && typeof u.unknownLegs === 'number'
-    ? u.unknownLegs
-    : (u && u.cost && u.cost.unpricedLegs) || 0;
-  // v4.4 Task 2: a fully-priced run can still be short. `council-wsgate01`
-  // printed an unqualified $0.2821 for a run that really spent $0.3036 — every
-  // leg `reported`, and 100% of the gap one unattributed `explore` child session.
-  const subtreeLegs = u && typeof u.subtreeUnknownLegs === 'number'
-    ? u.subtreeUnknownLegs
-    : (u && u.cost && u.cost.subtreeUnknownLegs) || 0;
-  if (u && u.cost && (typeof u.cost.amount === 'number' || unknownLegs > 0 || subtreeLegs > 0)) {
-    const known = typeof u.cost.amount === 'number' ? `$${u.cost.amount.toFixed(4)}` : '$0.0000';
-    const gaps = [];
-    if (unknownLegs > 0) { gaps.push(`${unknownLegs} leg(s) unknown`); }
-    if (subtreeLegs > 0) { gaps.push(`${subtreeLegs} leg(s) with unattributed subagent child-session spend`); }
-    const tail = gaps.length > 0
-      ? ` + ${gaps.join(' + ')} — real spend is at least this much`
-      : '';
-    lines.push(`  cost:  ${known} (${u.cost.source})${tail}`);
-  }
-  if (run.error) { lines.push(`  error: ${run.error.code}: ${run.error.message}`); }
-  return lines.join('\n') + '\n';
+/**
+ * Default real helpers; tests override via depsOverride (mirrors
+ * cli-handlers-spend.js's realDeps()/depsOverride convention).
+ */
+function realDeps() {
+  return {
+    // #81 (spec §2): same pure presence probe doctor's electron checks use (src/cli-handlers-doctor.js).
+    getElectronPath: () => require('./sidecar/interactive-process').getElectronPath(),
+  };
 }
 
-/** @param {object} args parsed CLI args @returns {Promise<number>} exit code */
-async function handleCouncilRun(args) {
+/**
+ * @param {object} args parsed CLI args
+ * @param {object} [depsOverride] test seam (getElectronPath)
+ * @returns {Promise<number>} exit code
+ */
+async function handleCouncilRun(args, depsOverride = {}) {
   const useJson = !!args.json;
+  const deps = { ...realDeps(), ...depsOverride };
 
   // v4.5 Task 12 (B7/F5): resolve --pack FIRST, above the Task-5 template
   // block, so a pack-filled args.template renders through that single
@@ -243,6 +233,17 @@ async function handleCouncilRun(args) {
   const { readCache } = require('./utils/model-catalog');
   const { runCouncil } = require('./council/run');
   const cfg = loadConfig() || {};
+
+  // #81 (spec §2): the GUI's existence was announced on NO surface from the
+  // CLI path — MCP launches auto-open, the CLI stayed silent. Auto-open
+  // parity is a product decision (deliberately not taken here); the SILENCE
+  // is the spec's to fix. Presence probe only — never launches. Placed here
+  // (runId/runDir already resolved, still before the engine await) so the
+  // notice is useful WHILE the run is live, not just after it finishes.
+  if (!useJson && deps.getElectronPath()) {
+    process.stderr.write(`Notice: the Council Workspace can render this run live — open it with: amicus watch ${runId} --ui\n`);
+  }
+
   const { exitCode, run } = await runCouncil({
     briefing: promptRes.prompt, models: bench, chair, critic, lenses,
     project, runId, runDir,

package/src/cli-handlers-council.js

@@ -8,7 +8,7 @@ const { sumWaveUsage, formatCost } = require('./utils/pricing');
 const { failJson, ERROR_CODES } = require('./utils/error-doc');
 const { buildReport } = require('./council/report');
 const { validateFindings, buildValidateDoc } = require('./council/findings');
-const { buildVerdict, readOverallVerdict, writeVerdictAtomic } = require('./council/verdict');
+const { buildVerdict, readOverallVerdict, readPriorVerdictSurfaces, writeVerdictAtomic } = require('./council/verdict');
 const {
   runSave: runCouncilSave,
   runList: runCouncilList,
@@ -169,8 +169,14 @@ function runVerdict(args, useJson) {
     // one of only two homes of the chair's synthesis (the other is
     // chair-output.md); tally.json carries no copy. Recover it from the RUN
     // folder — the tally's own directory, not `-o` — before rebuilding.
-    const overallVerdict = readOverallVerdict(path.dirname(path.resolve(tallyPath)), record.meta.runId);
-    verdict = buildVerdict(record, decisions, { overallVerdict });
+    const runDir = path.dirname(path.resolve(tallyPath));
+    const overallVerdict = readOverallVerdict(runDir, record.meta.runId);
+    // #87: tally.json carries neither seatLoss nor degrades — recover both from
+    // the run folder's verdict the same way the chair line is recovered.
+    const prior = readPriorVerdictSurfaces(runDir, record.meta.runId);
+    verdict = buildVerdict(record, decisions, { overallVerdict,
+      ...(prior.seatLoss ? { seatLoss: prior.seatLoss } : {}),
+      ...(prior.degrades ? { degrades: prior.degrades } : {}) });
   }
   catch (e) {
     return failJson(useJson, { code: ERROR_CODES.BAD_ARGS, message: `cannot build verdict: ${e.message}`,