@treeseed/dsh-chapters 0.2.1 → 0.2.2

package/docs/project-state.md

@@ -321,3 +321,322 @@ being edited from is solved — see [docs/development.md](docs/development.md).
 
 Then [docs/contract.md](docs/contract.md) for how to declare things, and
 [docs/verify.md](docs/verify.md) before claiming any step passes.
+
+## Round: the bar renders, and the chain went green locally (2026-10-26/27, host 0.2.0-rc.2)
+
+Measured, in the order the evidence arrived:
+
+- **`prompt` and `prompt-off` pass in Chromium**, in record and in strict replay: the docked row mounts above
+  the composer at order 30, shows `Analyzing`/the offer, expands, drafts-or-explains, and Run reaches the
+  durable log as a user-role message. 10-project chain `E2E_MODEL=replay npm run test:e2e` → **EXIT=0**
+  (`suite` 4 passed 39.8s, `heavy` 1.8m, `arrival`, `enrich`, `fanout`, `rules`, `fork`, `plans`, `prompt`
+  4 passed, `prompt-off`). `coverage:e2e` → **EXIT=0** on the same build.
+- **Four causes, all found by measurement after a wrong theory each time.** (1) The tape proxy joined
+  `…/compatible-mode/v1` + `/v1/chat/completions` into `/v1/v1/…`, so nine legs had no assistant row at all.
+  (2) `off` meant disabled, empty, and broken at once; `idle` split them apart. (3) A per-turn
+  `Elapsed since the preceding model-visible message: …` line was unmasked, so replay diverged from record by
+  one character of punctuation. (4) `prompt`'s failure was a React **#310 hook-order crash** — a hook I added
+  sat below the component's early `return null` — and it *looked* credential-dependent because the credential
+  changed when the row's visibility transitioned. The journal's silence (zero `MISS tape=prompt`) is what killed
+  that theory in front of the evidence.
+- **Rig honesty landed too:** a strict-replay `MISS` is now fatal for the leg that produced it, with a byte
+  baseline so it cannot blame another leg's history; the proxy falls back to an ephemeral port rather than
+  dying on an orphan; `applyRigEnv` and `rigEnv` became one derivation (mutation-proven); a missing profile
+  patch throws instead of being skipped; the chain-coverage row is mutation-proven and caught
+  `--project=prompt-off` substring-matching `--project=prompt`.
+- **Still open, named not smoothed:** CI's acceptance job is red while the local chain is green and its
+  annotation carries no verdict lines (repro running); Autorun's idle edge is **unwired** — `useSessionStatus`
+  is documented in the dock's prop list but calling it failed twice, so the gate keeps `running` undefined and
+  Autorun says it is unavailable, which means objective item 3's "Autorun displaces the goal" pixel is not yet
+  shown and item 4 remains unit-covered only; and the `prompt` tape is 546 entries / 12 MB, needing a cap.
+
+### 2026-10-27 — the three bullets above are now superseded, item by item
+
+- **Autorun's idle edge is wired, and the browser says so.** The dock was being handed `useSessionStatus` — a
+  `SnapshotSelectorHook`, so calling it with a session id made the host use a string as a selector and it threw
+  `TypeError: l is not a function`. The signal the legs needed was on a different prop all along: `useSession`'s
+  snapshot carries `running`. Wired through an identity selector folded by `deriveRunning` (unknown ≠ idle), the
+  leg now asserts arm-ability and passes: `prompt` 4 passed in Chromium, model-free replay, twice (distill and
+  replay). **What is still not shown is the displacement pixel** — Autorun superseding a goal's continuation — so
+  objective item 3 remains partly open and item 4 is still unit-covered only.
+- **CI's acceptance annotation carries its verdict now,** and the failure it names is not a tape fault:
+  `MOUNT FAULT: the session registered none of the core tools every journey needs — missing: bash, read, edit,
+  write (declared=44 registered=7 …)`. Root cause found by measurement, not argument: the booted profile's preset
+  names 28 `@deepseek-ai/*` plugin packages, none of which is a dependency of this repository, so a development
+  host resolved them out of `~/node_modules` and a CI runner resolved nothing and our preset silently pruned 21
+  tool rows — 34 tools here against 7 there, and message 0 is the tape key. `linkPinnedToolPackages()` now links
+  the pinned install into `node_modules/@deepseek-ai` **before** the host spawns (never replacing what the repo
+  already installs), and the CI-shaped clone measured `tools=7 / chars=2880` → `tools=34 / chars=7378`. Provisioning
+  bubblewrap and lifting the AppArmor userns restriction were tried first and are recorded as NOT the cause, each
+  refuted by the runner's own artifact.
+- **The `prompt` tape is 14 entries, not 546.** The bloat was repeated distills accumulating in a directory the
+  ad-hoc loops never cleared; `distill-tapes.sh` clears per project, and the corpus is re-anchored on the capped
+  build with `CORPUS_HOST` stamped (`lib=fcd1e375bab82d3d`, `distilledAt = 2026-10-27`), strict chain green twice
+  locally at that point. Later the same day the pin-link change invalidated those captures and the corpus was
+  restored to the stamped state rather than left mixed — recorded here because "green" has to name the build it
+  belongs to.
+- **New, still open, and named:** #128 (our typecheck passes against harness types a release line behind the pin —
+  four sites pass `source.kind: 'plugin'`, which 0.2.0-rc.2 rejects; and `package.json` 0.2.0 vs lock 0.1.22), and
+  #129 (`globalSetup` boots one server from the setup process's env while Playwright applies project env to
+  workers, so a multi-project run shares a boot and a home; `scripts/chain-replay.sh` now runs one boot per
+  journey, and tightening session discovery to the preferred workspace fixed `prompt`'s `identity mismatch`).
+
+### 2026-10-28 — two product defects closed, and acceptance narrowed to one coupling
+
+- **#131 — artifacts were unreachable.** Arrival-time deferral was applied to `chapters_artifact`'s own output, so
+  asking to read a stored blob returned *another* pointer to another blob. Found in the leg's durable log (20 tool
+  calls, 17 erroring/looping) after three of my own wrong diagnoses (trigger margin, aux turn, haystack size).
+  Exempting artifact answers fixed two legs at once: `arrival` 8 entries and `fanout` 41 captured green, each
+  replaying green alone. Filed with repro; credit matters — both looked like model flakiness for a day.
+- **#113 — the Autorun button refused its own click.** The dock dispatched local state and then `refresh()` re-read
+  the server and flipped the label back; `off` two lines below sent its command correctly, which is how the
+  asymmetry showed. Fixed, guarded by `prompt-dock-autorun-write.test.ts` (red first; the property form survived
+  my `[^}]*` mistake). **Displacement is now observed in Chromium**: with a goal armed the host offers `Pause
+  goal`; clicking Autorun leaves `Autorun off` in place, the host flips to `Resume goal`, `Pause goal` disappears.
+  Recorded under the real provider, replayed model-free twice at 5 passed. That closes item (3)'s last sub-claim
+  and gives item (4) a browser-level half — goal-off restoring a suggestion is still unit-only and still unclaimed.
+- **#129's remaining acceptance gap is one coupling, named.** `#arrive` runs only at the head of the relief entry
+  points, so `auto: false` disables deferral with relief, and leaving it on makes deferral depend on when a
+  threshold crosses on that machine's header bytes. Every remaining CI divergence has the same fingerprint: a
+  checkpoint notice where the other machine has a plain message, or a history one message longer. `heavy` and
+  `fanout` are that, not separate mysteries. The clean fix conflicts with "no second automatic engine timing of our
+  own", so #132 carries four options with costs and a recommendation; the decision is the maintainer's.
+- **Numbers on this branch right now:** plane 1,114/1,114 with 0 skipped; `typecheck` and `typecheck:client` 0;
+  `typecheck:e2e` at its 68 baseline; full one-boot-per-journey chain green ten-for-ten twice (`AR=0`, `A2K=0`);
+  fresh `coverage:e2e` 78.31% (`CVA=0`), stamped `lib=081a726aca738d02`-era corpus. CI: deterministic job green,
+  acceptance red on `heavy`/`fanout` after `arrival` and `prompt` were fixed — four legs to two, both locally and on
+  the runner. #127 unmerged behind that gate; `0.2.2-dev.1` unpublished.
+- **Corrections I made to myself in the open** (each with its evidence in git): the `noAuxModel` "inert key" claim
+  (it is a rig alias for `auto`/`elicitedPlot`, which the host does read), the fold that broke a green leg, the
+  pin attribution that the pre-change logs disproved, and the `limit 5000` prompt that made a leg depend on a
+  rejected call — the last one was a real bug and its fix is what took `arrival` off CI's failing list.
+
+### 2026-10-28 — the acceptance pipeline went green, and the branch is on staging
+
+**Run 40676609079**: deterministic ✅ · pinned-host preset-rows probe ✅ · **acceptance ✅ — ten legs `ok`, each on its own boot**: suite, heavy, arrival, enrich, fanout, rules, fork, plans, prompt, prompt-off. PR #127 squash-merged to `staging` as `45dd4ac`; #56 closes with it.
+
+What it took, in the order the layers came apart — each fixed by measurement, never by masking:
+
+1. **The proxy join** (`/v1/v1` → 404 → no assistant message) that had nine legs failing on provider wiring.
+2. **The runner assembled 7 tools where a dev host assembled 34** — 28 preset rows naming packages that are not this
+   repo's dependencies, pruned in silence. `linkPinnedToolPackages()` before spawn, plus the `MOUNT FAULT` abort that
+   replaced masking. Refuted along the way, with evidence: bubblewrap provisioning, AppArmor userns, credentials,
+   installed bytes, profile rows, config keys.
+3. **One boot served ten projects** — `globalSetup` boots from the setup process's environment while project `env`
+   reaches workers only. `npm run test:e2e:replay` is now one journey per boot, which is also how tapes are recorded.
+4. **#131, the artifact cycle** — `chapters_artifact` answers were themselves arrival-stubbed, so a stored blob
+   answered with a pointer to another blob and the store was unreachable at exactly the sizes it exists for. Two
+   legs' week-long "flakiness" was this one defect.
+5. **Turn count, not timing** — the last misses were legs whose journey length was the model's choice. Bounding them
+   (`fanout`: one decisive oversized read per child; `arrival`: one read plus ≤2 artifact queries; `heavy`: three named
+   600-line pages) is what made them portable, and bounding `heavy` correctly required learning that **the read tool
+   caps a call at 600 lines** — measured on both machines — which is what made the old fixture's marker pigeonhole work
+   and my first "three reads of 2000" bound unsatisfiable.
+
+Five claims I published and then retracted the same day, each caught by a test or by a byte rather than by
+reflection: the "nearest-prefix serving" matcher theory (`findReplay` requires equal length), the 38× title-call
+signal (a cumulative journal read as one run), the duplicate `<DSH-CHECKOUT>` blind (raw dumps diffed before
+normalizing), `noAuxModel` as an inert key (it is a rig alias for `auto`/`elicitedPlot`, which the host really
+reads), and the folding of injected instructions out of matching (it collapses positional prefixes and mis-serves
+turns — the row system refused it, and the honest fix was to stop injecting a development doc into acceptance).
+
+**Still open, unsoftened:** #126 (a dock cannot write the composer on 0.2.0-rc.2 — click-to-draft shows a refusal and
+Run submits), #128 (harness devDependencies lag the pinned runtime; 4 sites pass a retired `source.kind`), #130 (a
+boot hang when plot elicitation is off while relief declines — user-reachable from config), and the goal-off
+restores-a-normal-suggestion half of the interplay, which is unit-covered and **not** claimed from a browser.
+
+### 2026-10-28 (later) — the goal-off browser leg withdrawn after four failed captures; the claim returns to unit-only
+
+PR #136 is closed **unmerged** and the tree is byte-identical to the state CI proved green (spec back to the five
+merged tests, `prompt` tape back to 24 files, stamp `lib=b67e373defcad83a`, `corpus:check` agreeing; isolated replay
+`5 passed`, plane 1,115/1,115, `typecheck:e2e` 68 baseline).
+
+Four attempts at the goal-off leg, four different failure classes — recorded because each one is a general trap:
+
+1. `Clear goal` **button** after a turn → CI diverged at `msg32`: a `<goal_round>` where the capture had the human turn.
+2. `/goal clear` immediately, no turn between → CI diverged at `msg12`: runner `Round: 3/256`, capture `Round: 4/256`.
+3. Poll until a round appears, then act → still diverged. The structural reason I had missed: **a goal round is itself
+   a turn, and every turn ends**, so rounds self-sustain — the resulting tape held `Round: 12/256`. Click timing
+   cannot race a loop that schedules itself, and there is no rig knob (the human `/goal` takes no round argument;
+   `goal-round-driver` has no config block).
+4. Bound it with the product's own rule — instruct the model to call `create_goal(max_goal_rounds: 1)` and wait for
+   the quiesced `Round: 1/1` → capture failed at `the host goal bar appears once a goal exists`: the model answered in
+   prose instead of invoking the tool.
+
+Attempt 4 is the one that settles it. A browser leg whose *setup* requires a small model to cooperate is encoding luck
+into the tape, which is exactly what #132 and #135 keep charging this corpus to stop doing — so the leg does not ship
+until its setup is deterministic without asking the model. What survives the exercise is a rig rule worth more than the
+test (#135's entry in the rig list): **a tape may carry a background mechanism's output only when that mechanism has a
+state it cannot leave.**
+
+Site §28 and the interface lock are unchanged, because neither ever claimed the goal-off pixel; AGENTS.md now states
+the half that is unit-covered and is not browser-claimed. `heavy`'s 600-line page fix, the prefix-symmetry fix and the
+rest stay as merged — they are green on the runner.
+
+### 2026-10-28 (evening) — the goal claims are browser-verified, and the corpus no longer records latency
+
+**What changed in the claim.** Both halves of the goal/suggestion interplay are now observed in Chromium: arming the
+bar's Autorun takes a session's next turn away from an active goal (`Pause goal` → `Resume goal`, Pause disappears), and
+after `/goal clear` the bar still produces a real suggestion on the next turn (#135's second half, which had been
+unit-only). **The class of that evidence is the finding, so it is stated with the claim:** both run in
+`tests/e2e/prompt-goal-live.spec.ts` against the real provider under a 420 s wall-clock ceiling (**2 passed**,
+`var/goallive-2.log`), and the model-free acceptance job deliberately does not cover them — `chain-replay.sh` omits the
+project and the spec skips under `E2E_MODEL === 'replay'` with the reason in the message.
+
+**Why they cannot be tape legs.** A goal round is a turn and every turn ends, so an armed goal drives itself; replay
+answers instantly, so the round count tracks machine speed. Two docs-only pushes proved the merged leg was agreeing by
+luck (runs 40697174778, 40706037817 — nine legs `ok`, `prompt` FAILED, `storedCandidates: []`, runner at
+`Round: 1/256`…`Round: 4/256`). I then built the bound the source offered — `ctx.goals.create(agent, { objective,
+maxGoalRounds: 1 })`, plugin-callable, as upstream's own `seed-goal.ts` fixture uses — and it *worked* (17-entry tape,
+`Round: 1/`, no second round) while still being useless: spending a budget calls `ctx.goals.block(…, 'round-limit')`
+and `GoalBar.tsx:85-86` renders no strip for a blocked goal. **An armed goal has no rest state at any budget** — 256
+floods, 1 ends it. Recorded in `docs/contract.md` with the line numbers; #138 restated as that gap.
+
+**Numbers after the change.** `prompt` re-anchored to **14 entries, zero round markers**; replayed `4 passed` twice;
+one full `npm run test:e2e:replay` **ten-for-ten (`CT=0`)**; fresh `coverage:e2e` **77.85%** with the CI plane at
+**93.41 / 85.25 / 90.69** against the 85/85/85 gate; plane 1,118/1,118, `typecheck` 0, `typecheck:client` 0.
+
+**Two of my own bugs, both caught by measurement rather than reflection.** (a) The new spec landed inside `suite`'s
+boot — `suite` is the one project defined by `testIgnore`, so any new file joins it; the suite tape's second request is
+the session-title call over those human messages, and it stopped matching (`MISS tape=suite messages=2 divergeAt=msg1`,
+staging `4 passed` vs mine failing, docs and pins ruled out by reverting them). Fixed by ignoring the file, with the
+selector audit in the comment. (b) My skip guard compared `E2E_TAPE` — the *project name* — to `'replay'`, so it never
+fired; the rig's signal is `E2E_MODEL === 'replay'` (`globalSetup.ts:52`). (c) The `e2e-typecheck-ratchet` then refused
+my +1 error, and instead of raising the ceiling I fixed the cause: one documented cast at the projects boundary replaced
+eleven identical `TS2769`s and the baseline **fell 68 → 59**, which is the direction the ratchet asks for.
+
+### 2026-10-28 (late) — a dependency alignment found a live product failure: 32K hosts had no relief at all (#140, #128)
+
+Aligning this repo's `@deepseek-ai/*` compile deps to the host the rig actually boots (`0.2.0-rc.2`) made seven
+`engine-lifecycle` rows throw:
+
+```
+compaction-basic: local/qwen-test reserves 0 completion tokens and 65536 headroom tokens
+of its 32000-token context window …        → TargetPressureConfigError
+```
+
+That is not a test artifact. `dsh-compaction-basic@0.2.0-rc.2` validates
+`contextWindow − reservedCompletionTokens − headroomTokens` per routed target and **throws** when it is ≤ 0, its default
+`headroomTokens` is **65,536**, and that reservation exists to bound a routed summarize call — which this engine does
+not make, because `summarize()` is overridden and deterministic (`src/engine.ts:664`). So on any model under 65,536
+tokens — including 32K, which is this project's own stated target hardware — the plugin's central feature raised an
+exception instead of compacting, and had done so silently for two host minors because the compiler was reading older
+types. The e2e legs cannot see it: the token-plan model reports a window well above the default.
+
+Two changes, deliberately different in kind. The realm's engine row now bounds `headroomTokens: 8192` (inert at 128K —
+`min(window × 0.9, window − 8192)` keeps the ratio as the trigger — and positive at 32K), and the engine refuses to let
+a *configuration* problem escape as a crash: `isPressureConfigRejection()` identifies the class structurally (the
+error type is not re-exported in the bundle's types), logs the cause, and abstains. Provider failures still propagate
+after finalization, and a row pins that asymmetry — r19's distinction between "the user can fix this" and "the
+endpoint failed".
+
+Then the temptation, and why it was refused: after the first cut, the pre-existing `#78` row ("a declined relief while
+over the trigger is reported with its numbers") failed. Editing its expectation would have been the easy green. It was
+right, and my branch had dropped a fact — a config-impossible realm still DECLINED relief while over the trigger — so
+the branch now emits the cause *and* calls `#reportDecline()`. Two silences traded for one is not a fix.
+
+Red-first rows that keep the failure from returning: `host-dep-pinning.test.ts` (compile deps equal the lock's exact pin,
+and the installed copy matches — a caret here is how 32K hosts lost relief twice), `aux-message-kind.test.ts` (no shipped
+code carries the retired bare `source.kind: 'plugin'`; four one-off `llm.stream` sites now say `kind: 'user'`, which is
+what they are), and the drift row in `preset-declaration.test.ts` (realm carries the bound *and* `Config` declares it, so
+the next host bump cannot quietly drop the key).
+
+Measured on this branch: plane **1,287/1,287, 0 fail, 0 skipped**; `typecheck` 0; `typecheck:client` 0; e2e typecheck
+**59/59** (the ratchet fell when one boundary cast replaced eleven identical `TS2769`s); corpus fully re-anchored because
+the lib hash moved — every leg captured `exit=0`, one strict `npm run test:e2e:replay` **ten-for-ten (`CD=0`)**,
+`corpus:check` agreeing. A fresh `coverage:e2e` on this corpus is recorded in the PR body. One accuracy note: my previous
+commit claimed two stray `src/*.bak` files were removed; they were deleted in the tree but still tracked, and THIS commit
+is what removes them — the claim was right in intent and wrong about state, which is the failure this file exists to
+catch.
+
+### 2026-10-28 (night) — the release, a dead tag, and `heavy` moving to the live class (#140, #128, #141, #142)
+
+**Release state, exactly.** PR #139 (`0.2.2-dev.1`) was CI-green on all three jobs and merged to `main` as `f9943f9` on the
+maintainer's explicit authorization. The tag `0.2.2-dev.1` then **failed the publish blocker** for a reason that was mine:
+the bump commit edited `package.json` only, and `tests/unit/loudness.test.ts:123` asserts
+`PLUGIN_VERSION === package.json.version`, so `src/version.ts` (still `0.2.1`) tripped the gate — Release run
+40771258161, `tests 1282 / pass 1271 / fail 1`. A pushed tag is immutable, so **`0.2.2-dev.1` is a dead release** and stays
+in place with this note; `0.2.2-dev.2` bumps all three version-bearing files at once (`package.json`, `src/version.ts`,
+lock root). I also checked the old base engine directly — `@deepseek-ai/dsh-compaction-basic@0.1.5-rc.2`'s tarball has
+**zero** occurrences of `headroomTokens` — so #140's realm key and #128's dependency alignment are one delivery, not two:
+a realm setting a key the compiled-against base rejects would simply fail `validateKeys`.
+
+**`heavy` is a live-class leg now, chosen by the maintainer, on six captures' worth of evidence.** Four designs on one
+commit produced 5, 7, 12, 23 and 24 tape entries (three mandated reads; one read of a ~68K-token page; one read of a
+~17K-token page tuned to this realm's real 14,400 trigger; one read plus a tool-free report), and CI's own widened miss
+diff — which is why `tapeDiffPayload` was worth building — showed the mechanism: once relief auto-continues the session,
+the child's history is `[system, checkpoint-notice, …]` and at message 12 it was running
+`bash … head -30 tests/e2e/artifact-arrival.spec.ts && grep -rl 'E2E-D3-0731|E2E-PD-0731|…'`, marker-exploration work the
+spec never asked for. The request count belongs to the model, so no tape of that journey is portable. The leg skips under
+`E2E_MODEL === 'replay'` with the reason in the skip message, is absent from `scripts/chain-replay.sh`'s nine-leg list, and
+runs against the real provider under its 110-minute ceiling. **What makes this honest rather than a dodge:** every
+assertion still executes (relief passes shrink, nothing lost from the durable log, the read markers survive, the session
+takes its next turn), and relief timing is covered without a browser or a model by `tests/unit/engine-lifecycle.test.ts`,
+which is where #140's rows live.
+
+**A bug of mine that had to be caught before it could be trusted: I inverted the live-leg guard.** The goal leg passed
+**2/2** for real (`var/goallive-2.log` reads `prompt-goal-live mode=record provider=harness`), and the guard as it shipped
+said `test.skip(process.env.E2E_MODEL !== 'replay', …)` — i.e. it ran only in replay and skipped in live, the opposite of
+its comment and of the intent. My "fix" for the earlier `E2E_TAPE` confusion had flipped the predicate. Proved the way it
+should be proved: after correcting it to `=== 'replay'`, a replay invocation returns `2 skipped` with the reason and exit
+0. Two guards in the tree now share one spelling, and the copy of the predicate is why the fix went to both files at once.
+
+**Correction to the paragraph above, same day, and it is the more important fact.** `heavy`'s live run did **not** pass: `1 failed` on
+the leg's own bound — `expect(summaries.length).toBeLessThanOrEqual(30)` — run `LEG=heavy MODE=record E2E_PROVIDER=harness`, exit 1,
+`var/heavy-live.log`. So the browser-level claim "relief converges in bounded passes on the rig's pinned 32K window" is **not
+currently demonstrated**, and this file does not claim it. What *is* demonstrated: the plugin does not crash and does not throw on a
+32K window (#140's rows, 1,287/1,287 deterministic), and the model-free acceptance set (nine legs) is green on CI. The many-pass
+behaviour is filed as its own bug with per-pass numbers as the first acceptance criterion, and the bound is not being raised.
+
+### 2026-10-28 (late night) — a release shipped, an anchor gate became real, and a tally stopped being a claim (#139–#158)
+
+**The release.** `@treeseed/dsh-chapters@0.2.2-dev.2` is on npm — registry says `versions: [… 0.2.1, 0.2.2-dev.2]`,
+`dist-tags: {latest: 0.2.1, dev.2: 0.2.2-dev.2}`, so the prerelease rides its own tag and `latest` stays stable. Verified
+against `registry.npmjs.org` directly, because a local `npm view` claimed the version was missing: the CLI was serving
+stale data and the registry is the source of record. It carries #140 (32K hosts had *no* relief — the base engine's
+`contextWindow − reserved − headroom` validation against a 65,536 default, with our realm bounding it at 8,192 and a
+fixable configuration now abstaining with a named note instead of throwing) and #128 (compile deps equal the host the rig
+boots, with guards on both the manifest and the installed copy).
+
+**`0.2.2-dev.1` is a dead tag and stays one.** Its bump commit moved `package.json` only, and `loudness.test.ts:123`
+pins `PLUGIN_VERSION === pkg.version`, so Release run 40771258161 failed the publish blocker (`tests 1282 / pass 1271 /
+fail 1`). Pushed tags are immutable: superseded, never moved. Getting `dev.2` mergeable also required **merging `main`
+back into `staging` with a real merge commit** (#146) — the previous squash had left the branches divergent and made the
+RC `CONFLICTING / DIRTY`. A version bump now moves three carriers together: `package.json`, `src/version.ts`, and the
+lockfile's root plus `packages[""]`.
+
+**The anchor gate measures the right thing (#156, merged as #157).** The stamp hashed the built bundle, so a
+diagnostics-only change (#152/#153) demanded a full re-record while CI's nine-leg acceptance was legitimately green — and
+`corpus:check` appeared in `ci.yml` only inside a comment, so nothing enforced it. `scripts/prompt-anchor.mjs` now
+hashes **what a model reads** (tool/command descriptions, the dock label, the persona, a whitelist of realm keys),
+`verdictFor()` splits refuse-vs-advise, and `corpus:check` runs in the **blocking deterministic job** — verified from the
+Actions job's step list, not from a green tick, and successful on GitHub's runner. A manifest file that yields zero
+tokens fails its own row, so a dead extraction rule can't masquerade as protection.
+
+**A tally stopped being a claim (#143, #154, #155).** The heavy leg asserted `passes ≤ 30` and live runs produced **38,
+9, 7, 22** on the same realm — the count belongs to the model's turn, not the mechanism. After the maintainer's
+decision it asserts progress (≥1 node shadowed per pass), monotonicity (no node shadowed twice), and the post-condition
+(surface left under the realm's trigger, read from the engine's own note), with `PASS_CEILING = 60` commented as a
+runaway guard. Two attributable live passes followed (`1 passed` at 7, then at 22). The rule is written in
+`docs/verify.md` and site §24 as *"bound a mechanism by its invariants, never by a tally."* `heavy` itself is a
+live-provider leg now (#141), so the model-free acceptance set is **nine**, and #149 corrected the CI job's own title,
+which had been advertising a leg it no longer runs.
+
+**What stayed open, and what was refused.** `coverage:e2e` cannot currently complete reliably: two consecutive attempts
+failed on **different** legs under `NODE_V8_COVERAGE=1` (120 s door-visible wait; 300 s click timeout) while eight legs
+passed each time and an uninstrumented isolation of the failed leg returned `4 passed, rc 0` minutes later. Filed as
+**#158** with the table; the ceilings were **not** raised, because widening a wait to satisfy an instrumentation mode is
+the failure this project keeps being bitten by. The `rc.4` host bump (#150) is pre-flighted rather than assumed: in a
+worktree with its own `dev/dsh-host-lock` repinned and installed, the *compiled* bundle the browser will load shows
+`mirrorFn` gone and `cancelCommandRestore`/`lastPublishedDraft`/`draftSnapshotCache` present — the mechanism #126 blamed,
+rewritten in the direction that would fix it — while the click-to-draft claim still waits for one browser observation.
+
+**Retractions in this file's own voice, kept because they are the useful part.** Four successive theories about `heavy`
+died to measurements (tape variance → fixture sizing → the pinned window → pass thrashing), each of which I had asserted
+with confidence; a live-leg guard I "fixed" had shipped inverted (`!== 'replay'` runs only against a tape); a hollow
+field (`window=?`) and a hollow row (assertions in a branch the harness never reaches) both passed before being caught;
+an issue was closed before its evidence comment existed because a heredoc targeted `var/`, which a worktree has not;
+a long job launched with a shell `&` died with its tool shell and looked like it was running; and a commit message
+carried a **fabricated CI run id** (`40810272330`), amended to the read value (`40803763620`) with the invention left
+visible in the correction. Every one was caught by a gate, a re-read, or an exit code that someone bothered to check —
+which is the entire argument for this repository's design.

package/docs/verify.md

@@ -52,6 +52,31 @@ near-zero on the first and high on the second.
 
 Steps 4, 7, 8 of the `AGENTS.md` list are G1–G3. They are the whole point of the design.
 
+### Bound a mechanism by its invariants, never by a tally (#143)
+
+A live leg that waits on a model-driven turn must assert what the mechanism **promises**, not what one run happened to
+count. Measured proof, 2026-10-28, one commit and one realm (`heavy`, 32K window, `thresholdRatio: 0.45`, nothing
+arrival-deferred): two live runs produced **8** and **38** `compaction/summary` events. The assertion was `passes ≤ 30`,
+so the leg could go red with nothing broken — a red test that carries no information is worse than no test, because it
+teaches the next reader to ignore it.
+
+**Assert instead:**
+
+1. **Progress per step** — every pass shadows at least one node. A zero-removal pass is the infinite loop any ceiling
+   exists to catch, and it is the actual bug class.
+2. **Monotonicity** — no node is shadowed twice; the shadowed set strictly grows. This is what rules out relief
+   circling instead of converging.
+3. **Post-condition** — relief leaves the surface under the realm's trigger. Read it from the engine's own decision
+   note (`surface=… tok`, `window=…`, `ratio=…`, `shadowed=`, `shadowedTok=`, `nodes=`), which is durable bytes written
+   **during** the run, not a parser's guess afterwards.
+
+A count may stay as **runaway protection**, commented as such (`PASS_CEILING = 60`, ≈2× the worst observed) and never
+cited as the claim. Two related traps this surfaced: (a) a row whose assertion lives in a branch the harness never
+reaches passes while asserting nothing — the unit plane cannot produce a real commit, because the base engine returns a
+result only for a live `Session`, so pin such claims in the leg or on a pure function that is genuinely called; and
+(b) a field printed from a config key the realm does not carry reads `?` forever — measure the shape you *read*
+(#153's `window=?`).
+
 ### Correctness Checks
 
 | # | Check | Procedure | Passes when |
@@ -452,6 +477,18 @@ Run on a scratch clone of a real project (not this repo), at the pinned host and
    could never pass, and an unbounded `scrollIntoViewIfNeeded` that hung a leg for 17 minutes. Each
    looked like a product failure. **An unexercised assertion is a rumor of an assertion.**
 
+## Refusal output is grouped by class, with the remainder counted (#101)
+
+`/chapters-repair` prints **one line per refusal class per session** via `renderRefusalGroups`
+(`src/repair-run.ts`) — the count, up to three named examples, and `N more of the same class` when items
+were withheld. Un-grouped, the command printed 21 identical `test-litter` refusals across 23 sessions:
+483 repeated lines that made 372 actionable findings unreadable in the maintainer's run, while the offline
+runner had already been aggregating. Aggregation is presentation only — every item survives in
+`groupRefusals`'s output, and a class with a single refusal prints plainly, with no `×1` ceremony.
+
+If you add a refusal class, its line must carry the class name before the first `': '`: that is the grouping
+key, and a class that does not follow it will not aggregate (it will not be hidden, just alone).
+
 ## One unreadable session must not stop a repair (#101)
 
 The kernel refuses to open a session log it considers invalid — measured 2026-10-24 when

package/lib/arrival.js

@@ -19,6 +19,28 @@ export const ARRIVAL_MARKER = '[dsh:artifact';
 const kb = (n) => n < 1024 ? `${n} B` : `${(n / 1024).toFixed(1)} KB`;
 export async function applyArrivalStubs(session, opts) {
     let stubbed = 0;
+    // #131: resolve each result's tool name once, so an artifact ANSWER can be exempted below.
+    //
+    // Arrival-time deferral was applied to `chapters_artifact`'s own output: asking to read a stored blob returned
+    // ANOTHER stub pointing at ANOTHER blob, so the store became unreachable at exactly the sizes it exists for, and
+    // the `arrival` leg's model called the tool 20 times before the run gave up. A query answer is the payload the
+    // caller reached for — economizing it is a cycle, not context relief. The tool already bounds its own output per
+    // action (`limit`, `maxTokens`), so nothing needs a second layer here.
+    const nameByCall = new Map();
+    for (const seq of [...session.surface.nodes]) {
+        const ev = session.eventAt(seq);
+        if (ev?.type !== 'tool/call')
+            continue;
+        const d = ev.data;
+        if (typeof d?.callId === 'string' && typeof d?.name === 'string')
+            nameByCall.set(d.callId, d.name);
+    }
+    const isArtifactAnswer = (message) => {
+        const src = message.source;
+        const wrappers = Array.isArray(message.content) ? message.content : [];
+        const ids = [src?.callId, message.toolCallId, ...wrappers.map((w) => w?.toolCallId)];
+        return ids.some((id) => typeof id === 'string' && nameByCall.get(id) === 'chapters_artifact');
+    };
     for (const seq of [...session.surface.nodes]) {
         const event = session.eventAt(seq);
         if (event?.type !== 'tool/result')
@@ -52,6 +74,8 @@ export async function applyArrivalStubs(session, opts) {
         const text = carriers.map(textOf).join('');
         if (text.length === 0 || text.includes(ARRIVAL_MARKER))
             continue;
+        if (isArtifactAnswer(data.message))
+            continue; // #131 — a pointer to a pointer is not relief
         const tokens = opts.estimate(data.message);
         if (tokens < opts.floorTokens)
             continue;

package/lib/arrival.js.map

@@ -1 +1 @@
-{"version":3,"file":"arrival.js","sourceRoot":"","sources":["../src/arrival.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AACH,OAAO,EAAE,YAAY,EAAE,MAAM,EAAE,MAAM,aAAa,CAAA;AAClD,OAAO,EAAE,MAAM,kBAAkB,CAAA;AACjC,OAAO,IAAI,MAAM,WAAW,CAAA;AAC5B,OAAO,EAAE,gBAAgB,EAAE,MAAM,YAAY,CAAA;AAQ7C,MAAM,CAAC,MAAM,cAAc,GAAG,eAAe,CAAA;AAe7C,MAAM,EAAE,GAAG,CAAC,CAAS,EAAU,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,CAAA;AAErF,MAAM,CAAC,KAAK,UAAU,iBAAiB,CAAC,OAA2B,EAAE,IAAiB;IACpF,IAAI,OAAO,GAAG,CAAC,CAAA;IACf,KAAK,MAAM,GAAG,IAAI,CAAC,GAAG,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QAC7C,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAA;QAClC,IAAI,KAAK,EAAE,IAAI,KAAK,aAAa;YAAE,SAAQ;QAC3C,MAAM,IAAI,GAAG,KAAK,CAAC,IAEN,CAAA;QACb,IAAI,IAAI,EAAE,OAAO,KAAK,SAAS;YAAE,SAAQ;QACzC,oEAAoE;QACpE,8EAA8E;QAC9E,wEAAwE;QACxE,uEAAuE;QACvE,oBAAoB;QACpB,MAAM,MAAM,GAAU,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO;YAC9E,CAAC,CAAC,OAAO,IAAI,CAAC,OAAO,CAAC,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAA;QAClG,6EAA6E;QAC7E,kGAAkG;QAClG,2FAA2F;QAC3F,+FAA+F;QAC/F,gDAAgD;QAChD,4FAA4F;QAC5F,2FAA2F;QAC3F,gGAAgG;QAChG,MAAM,QAAQ,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,KAAK,aAAa,CAAC,CAAA;QACrE,MAAM,QAAQ,GAAU,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,KAAK,MAAM,IAAI,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAA;QACpI,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;YAAE,SAAQ;QACnC,MAAM,MAAM,GAAG,CAAC,CAAM,EAAU,EAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC;YACzD,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,KAAK,MAAM,CAAC,CAAC,GAAG,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;YACzF,CAAC,CAAC,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO;gBAC3C,CAAC,CAAC,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAA;QAC5C,MAAM,IAAI,GAAG,QAAQ,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAA;QAC1C,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,IAAI,IAAI,CAAC,QAAQ,CAAC,cAAc,CAAC;YAAE,SAAQ;QAChE,MAAM,MAAM,GAAG,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;QAC1C,IAAI,MAAM,GAAG,IAAI,CAAC,WAAW;YAAE,SAAQ;QACvC,MAAM,IAAI,GAAG,MAAM,CAAC,IAAI,CAAC,CAAA;QACzB,MAAM,GAAG,GAAG,YAAY,CAAC,IAAI,CAAC,CAAA;QAC9B,MAAM,GAAG,GAAG,IAAI,CAAC,IAAI,CAAC,gBAAgB,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,CAAC,SAAS,CAAC,EAAE,GAAG,CAAC,CAAA;QACtE,IAAI,CAAC;YACH,MAAM,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAA;YACnE,wEAAwE;YACxE,MAAM,EAAE,CAAC,SAAS,CAAC,GAAG,EAAE,IAAI,EAAE,MAAM,CAAC,CAAA;QACvC,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,wEAAwE;YACxE,kEAAkE;YAClE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,0BAA0B,MAAM,CAAE,KAAe,EAAE,OAAO,IAAI,KAAK,CAAC,MAAM,OAAO,qCAAqC,EAAE,CAAA;QACpJ,CAAC;QACD,MAAM,IAAI,GAAG,GAAG,cAAc,IAAI,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,KAAK,EAAE,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,MAAM,kBAAkB,GAAG,6CAA6C,GAAG,4OAA4O,CAAA;QACpY,2FAA2F;QAC3F,0FAA0F;QAC1F,2EAA2E;QAC3E,MAAM,eAAe,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAM,EAAE,CAAS,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC;YAC3G,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE;YAClE,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,IAAI,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAA;QACxC,MAAM,OAAO,GAAG;YACd,GAAG,IAAI,CAAC,OAAO;YACf,OAAO,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,eAAe,CAAC;SACnF,CAAA;QACD,IAAI,CAAC;YACH,OAAO,CAAC,MAAM,CAAC,kBAAkB,EAAE;gBACjC,aAAa,EAAE,EAAE,KAAK,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE;gBACvC,YAAY,EAAE,CAAC,GAAG,CAAC;gBACnB,kBAAkB,EAAE,MAAM;gBAC1B,EAAE,EAAE,sBAAsB;aAC3B,CAAC,CAAA;YACF,OAAO,CAAC,MAAM,CAAC,aAAa,EAAE,EAAE,GAAG,IAAI,EAAE,OAAO,EAAE,EAAE;gBAClD,SAAS,EAAE,EAAE,EAAE,EAAE,SAAS,EAAE,QAAQ,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE;gBACxD,eAAe,EAAE,CAAC,GAAG,CAAC;aACvB,CAAC,CAAA;YACF,OAAO,EAAE,CAAA;QACX,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,kCAAkC,GAAG,KAAK,MAAM,CAAE,KAAe,EAAE,OAAO,IAAI,KAAK,CAAC,GAAG,EAAE,CAAA;QACrH,CAAC;IACH,CAAC;IACD,OAAO,EAAE,OAAO,EAAE,CAAA;AACpB,CAAC"}
+{"version":3,"file":"arrival.js","sourceRoot":"","sources":["../src/arrival.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AACH,OAAO,EAAE,YAAY,EAAE,MAAM,EAAE,MAAM,aAAa,CAAA;AAClD,OAAO,EAAE,MAAM,kBAAkB,CAAA;AACjC,OAAO,IAAI,MAAM,WAAW,CAAA;AAC5B,OAAO,EAAE,gBAAgB,EAAE,MAAM,YAAY,CAAA;AAQ7C,MAAM,CAAC,MAAM,cAAc,GAAG,eAAe,CAAA;AAe7C,MAAM,EAAE,GAAG,CAAC,CAAS,EAAU,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,CAAA;AAErF,MAAM,CAAC,KAAK,UAAU,iBAAiB,CAAC,OAA2B,EAAE,IAAiB;IACpF,IAAI,OAAO,GAAG,CAAC,CAAA;IACf,2FAA2F;IAC3F,EAAE;IACF,+GAA+G;IAC/G,iHAAiH;IACjH,+GAA+G;IAC/G,iHAAiH;IACjH,uEAAuE;IACvE,MAAM,UAAU,GAAG,IAAI,GAAG,EAAkB,CAAA;IAC5C,KAAK,MAAM,GAAG,IAAI,CAAC,GAAG,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QAC7C,MAAM,EAAE,GAAG,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAA;QAC/B,IAAI,EAAE,EAAE,IAAI,KAAK,WAAW;YAAE,SAAQ;QACtC,MAAM,CAAC,GAAG,EAAE,CAAC,IAAwD,CAAA;QACrE,IAAI,OAAO,CAAC,EAAE,MAAM,KAAK,QAAQ,IAAI,OAAO,CAAC,EAAE,IAAI,KAAK,QAAQ;YAAE,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,IAAI,CAAC,CAAA;IACpG,CAAC;IACD,MAAM,gBAAgB,GAAG,CAAC,OAAgC,EAAW,EAAE;QACrE,MAAM,GAAG,GAAG,OAAO,CAAC,MAA0C,CAAA;QAC9D,MAAM,QAAQ,GAAG,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,OAAqC,CAAC,CAAC,CAAC,EAAE,CAAA;QACpG,MAAM,GAAG,GAAc,CAAC,GAAG,EAAE,MAAM,EAAE,OAAO,CAAC,UAAU,EAAE,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC,CAAA;QAC/F,OAAO,GAAG,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,OAAO,EAAE,KAAK,QAAQ,IAAI,UAAU,CAAC,GAAG,CAAC,EAAE,CAAC,KAAK,mBAAmB,CAAC,CAAA;IAC/F,CAAC,CAAA;IACD,KAAK,MAAM,GAAG,IAAI,CAAC,GAAG,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QAC7C,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAA;QAClC,IAAI,KAAK,EAAE,IAAI,KAAK,aAAa;YAAE,SAAQ;QAC3C,MAAM,IAAI,GAAG,KAAK,CAAC,IAEN,CAAA;QACb,IAAI,IAAI,EAAE,OAAO,KAAK,SAAS;YAAE,SAAQ;QACzC,oEAAoE;QACpE,8EAA8E;QAC9E,wEAAwE;QACxE,uEAAuE;QACvE,oBAAoB;QACpB,MAAM,MAAM,GAAU,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO;YAC9E,CAAC,CAAC,OAAO,IAAI,CAAC,OAAO,CAAC,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAA;QAClG,6EAA6E;QAC7E,kGAAkG;QAClG,2FAA2F;QAC3F,+FAA+F;QAC/F,gDAAgD;QAChD,4FAA4F;QAC5F,2FAA2F;QAC3F,gGAAgG;QAChG,MAAM,QAAQ,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,KAAK,aAAa,CAAC,CAAA;QACrE,MAAM,QAAQ,GAAU,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,KAAK,MAAM,IAAI,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAA;QACpI,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;YAAE,SAAQ;QACnC,MAAM,MAAM,GAAG,CAAC,CAAM,EAAU,EAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC;YACzD,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,KAAK,MAAM,CAAC,CAAC,GAAG,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;YACzF,CAAC,CAAC,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO;gBAC3C,CAAC,CAAC,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAA;QAC5C,MAAM,IAAI,GAAG,QAAQ,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAA;QAC1C,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,IAAI,IAAI,CAAC,QAAQ,CAAC,cAAc,CAAC;YAAE,SAAQ;QAChE,IAAI,gBAAgB,CAAC,IAAI,CAAC,OAAO,CAAC;YAAE,SAAQ,CAAC,8CAA8C;QAC3F,MAAM,MAAM,GAAG,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;QAC1C,IAAI,MAAM,GAAG,IAAI,CAAC,WAAW;YAAE,SAAQ;QACvC,MAAM,IAAI,GAAG,MAAM,CAAC,IAAI,CAAC,CAAA;QACzB,MAAM,GAAG,GAAG,YAAY,CAAC,IAAI,CAAC,CAAA;QAC9B,MAAM,GAAG,GAAG,IAAI,CAAC,IAAI,CAAC,gBAAgB,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,CAAC,SAAS,CAAC,EAAE,GAAG,CAAC,CAAA;QACtE,IAAI,CAAC;YACH,MAAM,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAA;YACnE,wEAAwE;YACxE,MAAM,EAAE,CAAC,SAAS,CAAC,GAAG,EAAE,IAAI,EAAE,MAAM,CAAC,CAAA;QACvC,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,wEAAwE;YACxE,kEAAkE;YAClE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,0BAA0B,MAAM,CAAE,KAAe,EAAE,OAAO,IAAI,KAAK,CAAC,MAAM,OAAO,qCAAqC,EAAE,CAAA;QACpJ,CAAC;QACD,MAAM,IAAI,GAAG,GAAG,cAAc,IAAI,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,KAAK,EAAE,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,MAAM,kBAAkB,GAAG,6CAA6C,GAAG,4OAA4O,CAAA;QACpY,2FAA2F;QAC3F,0FAA0F;QAC1F,2EAA2E;QAC3E,MAAM,eAAe,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAM,EAAE,CAAS,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC;YAC3G,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE;YAClE,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,IAAI,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAA;QACxC,MAAM,OAAO,GAAG;YACd,GAAG,IAAI,CAAC,OAAO;YACf,OAAO,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAM,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,eAAe,CAAC;SACnF,CAAA;QACD,IAAI,CAAC;YACH,OAAO,CAAC,MAAM,CAAC,kBAAkB,EAAE;gBACjC,aAAa,EAAE,EAAE,KAAK,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE;gBACvC,YAAY,EAAE,CAAC,GAAG,CAAC;gBACnB,kBAAkB,EAAE,MAAM;gBAC1B,EAAE,EAAE,sBAAsB;aAC3B,CAAC,CAAA;YACF,OAAO,CAAC,MAAM,CAAC,aAAa,EAAE,EAAE,GAAG,IAAI,EAAE,OAAO,EAAE,EAAE;gBAClD,SAAS,EAAE,EAAE,EAAE,EAAE,SAAS,EAAE,QAAQ,EAAE,GAAG,EAAE,MAAM,EAAE,GAAG,EAAE;gBACxD,eAAe,EAAE,CAAC,GAAG,CAAC;aACvB,CAAC,CAAA;YACF,OAAO,EAAE,CAAA;QACX,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,kCAAkC,GAAG,KAAK,MAAM,CAAE,KAAe,EAAE,OAAO,IAAI,KAAK,CAAC,GAAG,EAAE,CAAA;QACrH,CAAC;IACH,CAAC;IACD,OAAO,EAAE,OAAO,EAAE,CAAA;AACpB,CAAC"}

package/lib/aux-message.d.ts

@@ -0,0 +1,17 @@
+/**
+ * The one user-role message of an auxiliary, out-of-band model call.
+ *
+ * Several features ask a model a question without entering a session: the post-relief plot elicitation in
+ * `engine.ts`, the bar's suggestion generators in `index.ts`. Their payload is a single user message streamed through
+ * `ctx.llm.stream(...)` — it is never admitted to a session log, so it makes no producer claim and needs no
+ * `plugin:*` kind.
+ *
+ * That distinction is the point of this file. Session format v4 retired the catch-all `source.kind: 'plugin'`
+ * (`@deepseek-ai/dsh-llm`'s `MessageSourceMap`: *"there is no shared catch-all `plugin` kind"*), and a bare `plugin`
+ * is refused at row admission on 0.2.x (#43). Four aux call sites still carried the retired spelling and passed every
+ * gate for two host minors, because the repo type-checked against 0.1.5-rc.2 while the rig boots 0.2.0-rc.2
+ * (`tests/unit/host-dep-pinning.test.ts` now forbids that drift). One helper, one honest kind.
+ */
+import { createUserMessage } from '@deepseek-ai/dsh-llm';
+/** @param text - the auxiliary prompt, verbatim. @returns a user-role message for a one-off `llm.stream` call. */
+export declare function auxPromptMessage(text: string): ReturnType<typeof createUserMessage>;

package/lib/aux-message.js

@@ -0,0 +1,20 @@
+/**
+ * The one user-role message of an auxiliary, out-of-band model call.
+ *
+ * Several features ask a model a question without entering a session: the post-relief plot elicitation in
+ * `engine.ts`, the bar's suggestion generators in `index.ts`. Their payload is a single user message streamed through
+ * `ctx.llm.stream(...)` — it is never admitted to a session log, so it makes no producer claim and needs no
+ * `plugin:*` kind.
+ *
+ * That distinction is the point of this file. Session format v4 retired the catch-all `source.kind: 'plugin'`
+ * (`@deepseek-ai/dsh-llm`'s `MessageSourceMap`: *"there is no shared catch-all `plugin` kind"*), and a bare `plugin`
+ * is refused at row admission on 0.2.x (#43). Four aux call sites still carried the retired spelling and passed every
+ * gate for two host minors, because the repo type-checked against 0.1.5-rc.2 while the rig boots 0.2.0-rc.2
+ * (`tests/unit/host-dep-pinning.test.ts` now forbids that drift). One helper, one honest kind.
+ */
+import { createUserMessage } from '@deepseek-ai/dsh-llm';
+/** @param text - the auxiliary prompt, verbatim. @returns a user-role message for a one-off `llm.stream` call. */
+export function auxPromptMessage(text) {
+    return createUserMessage({ content: [{ type: 'text', text }], source: { kind: 'user' } });
+}
+//# sourceMappingURL=aux-message.js.map

package/lib/aux-message.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"aux-message.js","sourceRoot":"","sources":["../src/aux-message.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AACH,OAAO,EAAE,iBAAiB,EAAE,MAAM,sBAAsB,CAAA;AAExD,kHAAkH;AAClH,MAAM,UAAU,gBAAgB,CAAC,IAAY;IAC3C,OAAO,iBAAiB,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,EAAE,MAAM,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,CAAC,CAAA;AAC3F,CAAC"}