@event4u/agent-config 9.9.0 → 9.11.0

package/dist/agent-src/skills/song-to-script/SKILL.md

@@ -19,14 +19,14 @@ install:
 
 # song-to-script
 
-> Turn a song into `<project>/script.md` — `## Scene N` blocks whose
-> `duration:` values sum to the track length and whose cut points land
-> on real section boundaries. Consumed by
+> Turn a song into `<project>/script.md` — a sequence of `## Scene N`
+> blocks whose `duration:` values sum to the track length and whose cut
+> points land on real section boundaries. Consumed by
 > [`/video:from-song`](../../commands/video/from-song.md), then handed
 > to [`scene-expander`](../scene-expander/SKILL.md) and
 > [`video-director`](../video-director/SKILL.md). Never invents timing —
-> every boundary comes from the audio analysis, and its source (adapter
-> vs. probe `method`) tells this skill how musical those boundaries are.
+> every boundary comes from the audio probe, and the probe's `method`
+> tells this skill how musical (or not) those boundaries actually are.
 
 ## When to use
 
@@ -49,25 +49,27 @@ Do NOT use when:
     configured — see
     [`audio-adapter-contract.md`](../../scripts/ai-video/lib/audio-adapter-contract.md)):
     `{bpm, beats, downbeats, sections:[{start,end,label,energy?}]}` —
-    real musical structure. Beats/downbeats = candidate cut grid;
-    section labels musical (`verse`, `chorus`, …).
+    real musical structure. Beats/downbeats become the candidate cut
+    grid; section labels are musical (`verse`, `chorus`, …).
   - **Audio probe** — JSON from
     [`scripts/ai-video/lib/probe-audio.sh`](../../scripts/ai-video/lib/probe-audio.sh):
     `{duration, method, warning?, sections:[{start,end,energy,label}]}`.
-    - `method: silence` — boundaries are real quiet gaps; trust as cuts.
-    - `method: rms` — energy-delta inflections; usable but coarse.
-    - `method: interval` — **track is structurally flat** (brick-walled
-      / sustained); sections fixed-interval, NOT musical. `method:
-      interval` (or `warning` set) → script header states timing is
-      interval-based; operator should pass `--scene-durations` for
-      musical sync. Never present interval cuts as beat-synced.
-- **Model capabilities** — chosen video model's renderable envelope from
-  the multiplexer manifest:
+    - `method: silence` — boundaries are real quiet gaps; trust them as cuts.
+    - `method: rms` — boundaries are energy-delta inflections; usable but
+      coarse.
+    - `method: interval` — **the track is structurally flat** (brick-walled
+      / sustained); sections are fixed-interval, NOT musical. When `method`
+      is `interval` (or `warning` is set), the emitted script header states
+      that timing is interval-based and the operator should pass
+      `--scene-durations` for musical sync. Never present interval cuts as
+      beat-synced.
+- **Model capabilities** — the chosen video model's renderable envelope
+  from the multiplexer manifest:
   `scripts/ai-video/adapters/<provider>.sh capability --model <id>` →
   `{min_duration, max_duration, audio_sync, aspect, verified}`. Scene
-  durations MUST land inside `[min_duration, max_duration]`;
-  `verified: false` entries surfaced in the report, never trusted
-  silently.
+  durations MUST land inside `[min_duration, max_duration]`; a
+  `verified: false` manifest entry is surfaced in the report, never
+  trusted silently.
 - **Mode** — `brief` (operator text is the creative source) or `auto`
   (infer mood + action from energy).
 - **Brief** (brief mode only) — free text: story, settings, look.
@@ -80,26 +82,28 @@ Do NOT use when:
 ### Step 1: Map sections → scenes (capability-clamped, beat-aligned)
 
 One `## Scene N` per analysis section. `duration:` = `end - start`
-(rounded to 0.5 s). Then **clamp the plan to the model's renderable
-envelope** — `min_duration` / `max_duration` from the manifest
-(`<provider>.sh capability --model <id>`), provider tuning as fallback
-for single-model adapters:
-
-- **Section < `min_duration`** → merge into neighbour. With beat data,
-  merge toward the neighbour keeping the joined cut on a **downbeat**
-  (else any beat); without, merge into the shorter neighbour.
-- **Section > `max_duration`** → split into sub-scenes. With beat data,
-  every split point on the **nearest downbeat** (else beat) to the
-  equal-division point — never mid-beat; without, split equally.
-- **No valid plan** (song shorter than `min_duration`; section cannot
-  split onto any beat inside the envelope) → **halt and surface the
-  conflict** with model id + violated bound — an unbuildable plan never
-  reaches the renderer.
+(rounded to 0.5 s). Then **clamp the plan to the chosen model's
+renderable envelope** — read `min_duration` / `max_duration` from the
+model-capabilities manifest (`<provider>.sh capability --model <id>`),
+falling back to the provider tuning for single-model adapters:
+
+- **Section shorter than `min_duration`** → merge into its neighbour.
+  With beat data, merge toward the neighbour that keeps the joined cut
+  on a **downbeat** (else any beat); without beat data, merge into the
+  shorter neighbour.
+- **Section longer than `max_duration`** → split into sub-scenes. With
+  beat data, place every split point on the **nearest downbeat** (else
+  beat) to the equal-division point — never mid-beat; without beat
+  data, split equally.
+- **No valid plan exists** (e.g. the whole song is shorter than
+  `min_duration`, or a section cannot be split onto any beat inside the
+  envelope) → **halt and surface the conflict** with the model id and
+  the violated bound — an unbuildable plan never reaches the renderer.
 
 Every emitted scene satisfies
-`min_duration ≤ duration ≤ max_duration`. Manifest entry
-`verified: false` → say so in the report — documented-best-effort, not
-a smoke-traced fact.
+`min_duration ≤ duration ≤ max_duration`. When the manifest entry is
+`verified: false`, say so in the report — the envelope is
+documented-best-effort, not a smoke-traced fact.
 
 ### Step 2: Assign mood + action
 
@@ -108,29 +112,31 @@ First decide the **subject mode**:
 - **Character mode** — `character.json` exists: every scene's `action:`
   names the locked subject, never a fresh description.
 - **Style mode** — no `character.json`: scenes describe setting, palette,
-  motion continuity (the recurring *look*), not a person. Valid abstract
-  / landscape / visualiser path — do not invent a human subject.
+  and motion continuity (the recurring *look*), not a person. This is the
+  valid abstract / landscape / visualiser path — do not invent a human
+  subject to fill the slot.
 
 Then pick the **prompt source per segment — the modality switch**:
 
-- **Lyric segment** (vocal map places ≥1 transcribed line inside it) →
-  scene prompt derives from the **lyric line itself**: its imagery,
-  subjects, verbs seed `mood:` + `action:` (character mode: acted by the
-  locked subject; style mode: rendered as setting / weather / palette —
-  never an invented human). Line lands in `dialogue:` per Step 3.
-- **Instrumental segment** (no vocal-map line) → prompt derives from the
-  **audio features**: section `label` + `energy` via the intent table
-  below. Never recycle a lyric from another segment into an
-  instrumental one.
+- **Lyric segment** (the vocal map places ≥1 transcribed line inside
+  it) → the scene prompt derives from the **lyric line itself**: its
+  imagery, subjects, and verbs seed `mood:` + `action:` (in character
+  mode, acted by the locked subject; in style mode, rendered as
+  setting / weather / palette — never an invented human). The line
+  lands in `dialogue:` per Step 3.
+- **Instrumental segment** (no vocal-map line) → the scene prompt
+  derives from the **audio features**: section `label` + `energy` via
+  the intent table below. Never recycle a lyric from another segment
+  into an instrumental one.
 
 Then assign per scene:
 
 - **Brief mode** — distribute the brief's beats across scenes in order;
-  modality switch still applies (lyric segments quote the brief's
-  matching beat through the lyric's lens); energy modulates pacing. Do
-  not add story the brief did not state.
-- **Auto mode** — derive mood per section from `energy` + `label`
-  (probe labels and musical adapter labels both map):
+  the modality switch still applies (lyric segments quote the brief's
+  matching beat through the lyric's lens), and energy modulates pacing.
+  Do not add story the brief did not state.
+- **Auto mode** — derive mood per section from `energy` and `label`
+  (probe labels and musical labels from the analysis adapter both map):
 
   | label / energy | default scene intent |
   |---|---|
@@ -142,8 +148,8 @@ Then assign per scene:
   | outro / fade | pull-back, resolve, hold |
 
 **Energy → cut frequency + motion intensity.** Section energy (0..1,
-relative to track mean) drives cut rate AND camera intensity — chorus =
-faster cuts / more motion:
+relative to the track mean) drives both how often the edit cuts and how
+hard the camera moves — chorus = faster cuts / more motion:
 
 | energy vs. track mean | cut length target | `camera:` motion intensity |
 |---|---|---|
@@ -151,9 +157,10 @@ faster cuts / more motion:
 | within ±0.10 of mean (verse / build) | medium — one scene per section or per 4-bar phrase | steady dolly, slow tighten |
 | ≤ mean − 0.10 (breakdown / outro) | long — merge toward `max_duration`, hold shots | locked-off or slow drift |
 
-High-energy splitting and low-energy merging stay inside the Step 1
-envelope and land on downbeats — the energy table picks *where inside
-the envelope* a length falls, never outside it.
+High-energy splitting and low-energy merging both stay inside the
+Step 1 capability envelope and land on downbeats — the energy table
+chooses *where inside the envelope* a scene length falls, never
+outside it.
 
 ### Step 3: Vocal map — transcribe, never guess (vocal tracks)
 
@@ -163,10 +170,10 @@ BRIEF / STORY SKELETON OR A GUESSED STRETCH. NEVER PUT ONE SINGER'S
 LINE ON ANOTHER SINGER'S SCENE.
 ```
 
-Track has vocals and the run intends lip-sync → build a **vocal map**
-from the real audio before assigning any `dialogue:`:
+When the track has vocals and the run intends lip-sync, build a
+**vocal map** from the real audio before assigning any `dialogue:`:
 
-1. **Transcribe** audio to timestamped lines. Adapter-first: a
+1. **Transcribe** the audio to timestamped lines. Adapter-first: a
    configured `lyrics` provider (e.g.
    [`audio-adapters/whisperx.sh`](../../scripts/ai-video/audio-adapters/whisperx.sh))
    returns word-level timestamps **plus per-line diarization labels**
@@ -175,34 +182,35 @@ from the real audio before assigning any `dialogue:`:
    echo '{"audio_path":"<vocal-stem-or-song>"}' \
      | scripts/ai-video/audio-adapters/whisperx.sh analyze
    ```
-   No lyrics provider → OpenAI `/v1/audio/transcriptions`
+   No lyrics provider configured → OpenAI `/v1/audio/transcriptions`
    (`response_format=verbose_json` → `segments[].{start,end,text}`) or
-   local whisper (no speaker labels — every line starts as `"?"`).
-   Either way the transcript is the only source of lyric timing.
+   local whisper as before (no speaker labels — every line starts as
+   `"?"`). Either way the transcript is the only source of lyric timing.
 2. **Label the singer** per line — map diarization labels (or unlabeled
-   lines) to cast names via the operator's who-sings reference (roster,
-   brief naming who sings which line, or character cast): one label ↦
-   one cast name, consistently. Genuinely ambiguous (label `"?"`,
-   mixed-speaker line, no roster match) → keep `singer: "?"`, surface
-   it — never guess a singer to fill the slot.
-3. **Emit** `<project>/vocal-map.json`: `[{start, end, text, singer}]`,
-   timing verbatim from the transcript.
-4. **Validate** — run the ground-truth enforcer before the sign-off
-   gate:
+   lines) to cast names via the operator's who-sings reference (a
+   roster, a brief that names who sings which line, or a character
+   cast): one diarization label ↦ one cast name, consistently. If a
+   line's singer is genuinely ambiguous (label `"?"`, mixed-speaker
+   line, or no roster match), keep `singer: "?"` and surface it —
+   never guess a singer to fill the slot.
+3. **Emit** `<project>/vocal-map.json`:
+   `[{start, end, text, singer}]`, timing verbatim from the transcript.
+4. **Validate** — run the ground-truth enforcer before handing the map
+   to the sign-off gate:
    ```bash
    scripts/ai-video/lib/validate-vocal-map.sh <project>/vocal-map.json \
      <project>/transcript.json --roster "<cast names>"
    ```
-   Rejects re-timed lines, lyrics not in the transcript, missing singers
-   (exit 7, specific line named). Red validator = halt — fix the map,
-   never bypass.
-5. **Place lines into the matching scene's** `dialogue:` block using
+   It rejects re-timed lines, lyrics not in the transcript, and missing
+   singers (exit 7, specific line named). A red validator is a halt —
+   fix the map, never bypass.
+5. **Place lines into the matching scene's** `dialogue:` block using the
    transcript timing, tagged with the singer (`singer: "<line>"`). A
-   scene's lip-sync subject MUST be the line's labelled singer; a `"?"`
-   line gets NO lip-sync scene until the operator resolves it.
+   scene's lip-sync subject MUST be the line's labelled singer; a
+   `"?"` line gets NO lip-sync scene until the operator resolves it.
 
 No vocals / no transcript / no lip-sync intent → leave `dialogue:` empty;
-scene is performance / B-roll. **Never fabricate lyrics**, **never
+the scene is performance / B-roll. **Never fabricate lyrics**, **never
 re-time a line off the brief**, and in style mode `dialogue:` stays empty
 (lip-sync needs a character subject). The `/video:from-song` sign-off
 gate (its Step 6) shows this map for approval before any render.
@@ -211,11 +219,11 @@ gate (its Step 6) shows this map for approval before any render.
 
 Write `<project>/script.md` (and `<project>/vocal-map.json` when the
 track has vocals). Report the delta, the section→scene map, **the probe
-`method`** (silence-derived, energy-derived, or interval-fallback), **and
-whether lyric timing is transcript-derived** (it must be — never
-brief-derived). Sum cannot be reconciled (e.g. provider max-duration
-forces more time than the song has) → **halt and surface the
-conflict** — do not pad silently.
+`method`** (so the operator sees whether cuts are silence-derived,
+energy-derived, or interval-fallback), **and whether lyric timing is
+transcript-derived** (it must be — never brief-derived). If the sum
+cannot be reconciled (e.g. provider max-duration forces more time than
+the song has), **halt and surface the conflict** — do not pad silently.
 
 ### Step 5: Validate before handoff
 

package/dist/agent-src/skills/sql-writing/SKILL.md

@@ -12,8 +12,8 @@ packs:
 # sql
 
 
-> **Grounded corpus:** tuning decisions (indexes, keyset pagination, N+1,
-> trigram search, lock contention) ground via the
+> **Grounded corpus:** tuning decisions (indexes, keyset pagination,
+> N+1, trigram search, lock contention) ground via the
 > [`database`](../database/SKILL.md) corpus — `./scripts-run
 > <skills-root>/corpus-grounding/scripts/ground search
 > --manifest <skills-root>/database/data/manifest.json "<symptom>"`.

package/dist/agent-src/skills/stakeholder-tradeoff/SKILL.md

@@ -147,9 +147,10 @@ column is missing.
   becomes input to this skill for escalation. Boundary prose lives
   in [`docs/guidelines/cross-role-handoff.md`](../../../docs/guidelines/cross-role-handoff.md).
 - The trade-off is purely technical (perf vs storage, sync vs async)
-  — route to [`decision-record`](../decision-record/SKILL.md); one costly
-  choice on ≥ 3 options with user-weighted criteria → its § Weighted-matrix
-  mode (this skill answers WHO pays; that mode answers WHICH option wins).
+  — route to [`decision-record`](../decision-record/SKILL.md); if it is
+  one costly choice on ≥ 3 options with user-weighted criteria, its
+  § Weighted-matrix mode scores it (this skill answers WHO pays; that
+  mode answers WHICH option wins).
 - The dominant axis is risk, not stakeholder cost — route to
   [`risk-officer`](../risk-officer/SKILL.md).
 - The output is the locked decision artifact — hand off to

package/dist/agent-src/skills/standards-from-config/SKILL.md

@@ -11,76 +11,94 @@ packs:
 
 # standards-from-config
 
-Evidence v2 **Class A** — configured convention. Coding standards are not *guessed*
-and not *observed* — **derived from the real tooling config**. The config *is* the
-standard; this skill produces a **pointer + digest**, never a flattened claim.
-High-trust, auto-refreshable, drift-proof.
+Evidence v2 **Class A** — configured convention. The safe, on-thesis half of
+self-building project context: coding standards are not *guessed* and not
+*observed*, they are **derived from the real tooling config**. The config *is*
+the standard; this skill produces a **pointer + digest**, never a flattened
+claim. High-trust, auto-refreshable, drift-proof — because the truth lives in
+the config file, and the card only points at it plus distils the 3–5 points an
+agent needs most while writing.
 
 Definitions (Class A/B/C, trust tiers, the v1↔v2 isolation contract) live in
 [`evidence-discipline`](../../agent-src/contexts/execution/evidence-discipline.md);
-parent context mechanism is [`context-document`](../context-document/SKILL.md).
+the parent context mechanism is [`context-document`](../context-document/SKILL.md).
 
 ## When to use
 
-- Before writing or reformatting code in an unfamiliar project — need enforced style.
-- Before writing a commit message / branch name — need project's commit convention.
+- Before writing or reformatting code in an unfamiliar project and you need its
+  enforced style (indent, line length, quote style, import order, naming).
+- Before writing a commit message / branch name and you need the project's
+  commit convention.
 - When orienting in a repo and a committed Class-A standards card would save the
   next agent from re-deriving the same config.
 
-Do NOT use for: a standard with **no** config backing (that's Class B or nothing —
-never invent a Class-A claim); concrete shape of code (that's v1 `source-discovery`,
-read fresh); project with no tooling config (record a negative fact).
+Do NOT use for: a standard that has **no** config backing (that is a *guessed*
+or *observed* convention — Class B, not A; never invent a Class-A claim); the
+concrete shape of code (that is v1 `source-discovery`, read fresh); a project
+with no tooling config at all (record a negative fact: "no enforced standard
+found", do not fabricate one).
 
 ## Procedure
 
-1. **Detect real config sources** (read fresh, never from memory). By ecosystem — inspect
-   what actually exists:
-   - cross-stack: `.editorconfig`, `CONTRIBUTING.md`, commitlint / `.gitmessage`, CI lint steps.
-   - JS/TS: `eslint.config.js` / `.eslintrc*`, `.prettierrc*`, `biome.json`, `tsconfig.json`.
+1. **Detect the real config sources** (read fresh, never from memory). Common
+   sources, by ecosystem — inspect what actually exists:
+   - cross-stack: `.editorconfig`, `CONTRIBUTING.md`, commitlint /
+     `.gitmessage`, the lint/format steps in the CI workflow.
+   - JS/TS: `eslint.config.js` / `.eslintrc*`, `.prettierrc*`, `biome.json`,
+     `tsconfig.json` (strictness).
    - PHP: `pint.json`, `.php-cs-fixer.dist.php`, `phpcs.xml`.
-   - Python: `pyproject.toml` / `ruff.toml` (`[tool.ruff]`, `[tool.black]`), `setup.cfg`.
+   - Python: `pyproject.toml` / `ruff.toml` (`[tool.ruff]`, `[tool.black]`),
+     `setup.cfg`.
    - Ruby: `.rubocop.yml`. Go: `gofmt`/`golangci.yml`.
-2. **Derive as pointer + digest, never a flattened claim.** Each standard is one line:
-   the **value**, the **config file it came from**, and the **scope**. Write
-   `ruff.toml → line-length = 88 (scope: src/**/*.py)`, NOT "the project uses line
-   length 88". The pointer makes the claim re-checkable.
-3. **Preserve scope; surface conflicts, never flatten them.** Two configs or an inline
-   override disagree → emit **two pointers with two values + their scopes** — never
-   one ambiguous merged claim. A visible conflict is correct.
+2. **Derive as pointer + digest, never a flattened claim.** Each standard is one
+   line: the **value**, the **config file it came from**, and the **scope** it
+   applies to. Write `ruff.toml → line-length = 88 (scope: src/**/*.py)`, NOT
+   "the project uses line length 88". The pointer makes the claim re-checkable
+   and the next agent can open the config to confirm.
+3. **Preserve scope; surface conflicts, never flatten them.** If two configs (or
+   an inline / per-directory override) give two values, emit **two pointers with
+   two values + their scopes** — never one ambiguous merged claim. A visible
+   conflict is correct; a hidden flattened guess is the failure mode.
 4. **Stamp staleness.** Record each source's `config_mtime` (or content hash).
-   Digest is stale when config file's mtime/hash changes → re-derive. No human gate
-   needed (Class A is deterministic) — digest **regenerated from config, never
-   hand-edited**.
-5. **Persist** as a Class-A context card under `agents/settings/contexts/`. Class A is
-   high-trust *because* config-derived — read for heuristics only, never bypasses a
-   fresh structural read (v1↔v2 isolation contract).
+   The digest is stale when the config file's mtime/hash changes → re-derive. No
+   human gate is needed (Class A is deterministic) — the digest is **regenerated
+   from the config, never hand-edited**.
+5. **Persist** as a Class-A context card under `agents/settings/contexts/` (see
+   Output format). Class A is high-trust *because* config-derived — it is read
+   for heuristics only and never bypasses a fresh structural read (the v1↔v2
+   isolation contract in `evidence-discipline`).
 
 ## Output format
 
 A Class-A standards card MUST contain, in order:
 
-1. **Frontmatter** `class: A`, `trust: high (config-derived)`, and a `sources:` list
-   of `{path, config_mtime}` for every config the digest reads.
+1. **Frontmatter** `class: A`, `trust: high (config-derived)`, and a
+   `sources:` list of `{path, config_mtime}` for every config the digest reads.
 2. **A pointer+digest table** — one row per standard: `standard | value | source
-   (file:key) | scope`. No standard appears without a `source` cell.
-3. **A conflicts block** (may be empty) — any standard where ≥2 sources/scopes
-   disagree, as two-or-more rows, never one merged value.
-4. **A refresh line** — card is regenerated when any listed `config_mtime` changes;
-   read for heuristics only (never a structural bypass).
+   (file:key) | scope`. No standard appears without a `source` cell pointing at
+   a real config file.
+3. **A conflicts block** (may be empty) listing any standard where ≥2 sources /
+   scopes disagree, as two-or-more rows, never one merged value.
+4. **A refresh line** stating the card is regenerated when any listed
+   `config_mtime` changes, and is read for heuristics only (never a structural
+   bypass).
 
 ## Gotcha
 
-- A standard with no config backing is **not** Class A — do not write it here; it is
-  Class B (observed) or nothing. Class A never fabricates.
-- Do not flatten conflicting configs into one value — the conflict is the signal.
-- Card is a **regenerated digest**, not a hand-edited doc — editing the value by hand
-  defeats drift-proofing.
+- A standard with no config backing is **not** Class A — do not write it here;
+  it is a guessed/observed convention (Class B) or nothing. Class A never
+  fabricates.
+- Do not flatten conflicting configs into one value — the conflict is the
+  signal; surface both pointers.
+- The card is a **regenerated digest**, not a hand-edited doc — editing the
+  value by hand instead of the config defeats the drift-proofing.
 - A green pointer is not "the code obeys this" — it means the config declares it;
-  whether a given file complies is a fresh read.
+  whether a given file complies is a fresh read, not this card's claim.
 
 ## Do NOT
 
-- Do NOT emit a coding standard as a believed fact — emit a pointer to the config.
+- Do NOT emit a coding standard as a believed fact ("we use 4 spaces") — emit a
+  pointer to the config that declares it.
 - Do NOT merge conflicting configs into a single claim.
 - Do NOT hand-edit the digest value instead of the config.
 - Do NOT use a Class-A card to skip a fresh structural read (v1 stays in force).
@@ -89,5 +107,5 @@ A Class-A standards card MUST contain, in order:
 ## See also
 
 - [`evidence-discipline`](../../agent-src/contexts/execution/evidence-discipline.md) — Class A/B/C, trust tiers, the isolation contract.
-- [`context-document`](../context-document/SKILL.md) — parent context mechanism + storage locations.
+- [`context-document`](../context-document/SKILL.md) — the parent context mechanism + storage locations.
 - [`source-discovery`](../source-discovery/SKILL.md) — v1 structural discovery (read fresh; Class A never bypasses it).

package/dist/agent-src/skills/subagent-orchestration/SKILL.md

@@ -107,6 +107,20 @@ Rules:
   size floor) — it never re-interprets the task text on vibes.
 - Ambiguous shape → `none` (in-session), never a speculative spawn — the
   delegation-policy default.
+- **Fork vs. named subagent — an ordering, never a default (road-to-cache-economy
+  Phase 4).** On a host that offers a fork primitive, decide fork-vs-named
+  ORTHOGONALLY to the form picked above: tool and scope fit is first-order,
+  cache inheritance is second-order. A fork is preferred only when the child
+  continues the parent's task under identical tools and constraints (it then
+  reads the parent's cache on its first request); a named subagent is
+  preferred the moment isolation, a different tool set, or nested dispatch is
+  needed — a fork cannot nest, and forking forces background mode, which
+  changes the tool set and invalidates the very prefix that motivated the
+  fork. `do-in-parallel`-shaped fan-out from one parent has this shape by
+  construction (a known, unfixable upstream cost — see
+  [`prompts/README.md`](prompts/README.md) § Prompt-cache discipline). NEVER
+  encode "always fork" — that was proposed and cut (council 2026-07-30):
+  the cache-sharing benefit cannot be predicted before the fork happens.
 - Record the outcome in the telemetry line (`dispatch_mode` field, mode id
   or `none`) so the gate's value is measurable inside the ADR-117
   prove-or-drop window.
@@ -213,13 +227,13 @@ cross-*vendor* skeptics. Invariants, skeptic prompt, reconciliation, prove-or-dr
 gate → [`subagent-modes-detail`](../../agent-src/contexts/execution/subagent-modes-detail.md) § Mode 9
 + [`prompts/adversarial-verification-council.md`](prompts/adversarial-verification-council.md)
 
-> **Disposition (2026-07-28, honest null).** Adversarial-council
-> finding-coverage benchmark resolved as published null (`docs/benchmark.md`)
-> — mode NOT sold as defect-detection capability. Stays default-off bound to
-> that null; removal at next major unless external evidence (consumer-filed
-> case where the panel surfaced a real defect the single verifier missed)
-> appears first. Remaining honest value: perspective diversity + decision
-> documentation, nothing more.
+> **Disposition (2026-07-28, honest null).** The adversarial-council finding-
+> coverage benchmark resolved as a published null (see `docs/benchmark.md`) —
+> this mode is NOT sold as a defect-detection capability. It stays default-off
+> bound to that null; scheduled for removal at the next major unless external
+> evidence (a consumer-filed case where the panel surfaced a real defect the
+> single verifier missed) appears first. Its remaining honest value is
+> perspective diversity + decision documentation, nothing more.
 (ADR-122).
 
 ## Status taxonomy — every subagent return uses one envelope
@@ -355,6 +369,6 @@ Skip emit when `subagents.enabled: false` or `spawn_count == 0`
 | Stepwise plan with judge gates       | Inline — see [`prompts/do-in-steps.md`](prompts/do-in-steps.md) |
 | Standalone judge on an existing diff | Inline — see judge prompt in [`prompts/do-and-judge.md`](prompts/do-and-judge.md) |
 | External / networked second opinion  | [`ai-council`](../ai-council/SKILL.md) |
-| Cross-model review WITH repo access   | [`/team`](../../commands/team.md) (collaborative; subagents are in-session same-weights) |
+| Cross-model review WITH repo access   | [`/team`](../../domains/meta/team/command.md) (collaborative; subagents are in-session same-weights) |
 | Verifying completeness               | [`verify-before-complete`](../verify-before-complete/SKILL.md) |
 | What a subagent owns vs never owns   | [`subagent-boundary`](../../../docs/contracts/subagent-boundary.md) |

package/dist/agent-src/skills/subagent-orchestration/prompts/README.md

@@ -26,7 +26,10 @@ validates against that schema.
 ## Prompt-cache discipline
 
 When dispatching sibling subagents (e.g. `do-in-parallel` with N independent
-slices), **reuse a stable system-prompt prefix** across all siblings. Keep
+slices), **reuse a stable dispatch-prompt prefix** across all siblings. On
+this host the **system prompt comes from the agent definition**, not from
+anything the orchestrator writes — the dispatch prompt (the templated user
+message below) is the one prefix the orchestrator actually controls. Keep
 task-invariant text (role declaration, constraints, status enum, return-envelope
 instruction) in the prefix; put only the slice-specific `TASK:` and
 `CONTEXT FILES:` in the variable section. This maximises host-side prompt-cache
@@ -43,7 +46,30 @@ instead of a read, and the prefix must be re-hit within the **5-minute TTL**, so
 dispatch siblings promptly rather than trickling them. This is why the win is
 real for **fan-out** (many siblings, one prefix) and marginal for one-shot
 dispatch. Verify a cohort is actually reading, not re-writing, by checking
-`usage.cache_read_input_tokens > 0` on the 2nd+ sibling.
+`usage.cache_read_input_tokens > 0` on the 2nd+ sibling. Measured effect
+(anthropics/claude-code#74318): same-type sibling prefix hits reach **85%**
+within a 5-minute window vs **45%** for a mismatched or trickled cohort.
+
+**Sibling-uniformity rules** (follow from cache partitioning — the cache key
+includes model and tool set, and any mismatch fragments the cohort's shared
+prefix into separate caches): dispatch a cohort with the **same model**, the
+same `effort`, and the **same tool set** across every sibling, and fire them
+**promptly** rather than trickling dispatches one at a time — a sibling
+dispatched after the 5-minute window has already closed pays the full write
+premium alone, gaining nothing from the cohort.
+
+**Fork vs. named subagent — an ordering, not a default.** Tool and scope fit
+is first-order; cache inheritance is second-order. Prefer a **fork** when the
+child continues the parent's task under identical tools and constraints (a
+fork inherits the parent's system prompt, tools, and conversation history
+exactly, so its first request reads the parent's cache). Prefer a **named
+subagent** when the child needs isolation, a different tool set, or nested
+dispatch — a fork cannot nest, and forking forces background mode, which
+changes the tool set and therefore invalidates the very prefix that
+motivated the fork in the first place. Never default to "always fork": that
+option was considered and cut ([`../SKILL.md`](../SKILL.md) § Form gate)
+precisely because the cache-sharing benefit cannot be predicted before the
+fork happens, and background mode can erase it outright.
 
 ## Loading
 

package/dist/agent-src/skills/supply-chain-intake/SKILL.md

@@ -11,16 +11,16 @@ packs:
 
 # supply-chain-intake
 
-An LLM generates a plausible-sounding package name token-by-token with no lookup against a real registry. ~19.7% of AI-recommended packages don't exist (576k-sample study); the same fake name recurs across runs, so attackers pre-register it as malware — "slopsquatting". The `huggingface-cli` PoC (an empty package matching a common hallucination) drew 30k+ downloads. Endor Labs: only ~1 in 5 AI-recommended dependency versions is both real *and* safe. A dependency the agent named is untrusted until verified — never install it just because the model produced the name.
+An LLM generates a plausible-sounding package name token-by-token with no lookup against a real registry. ~19.7% of AI-recommended packages do not exist (576k-sample study); the same fake name recurs across runs, so attackers pre-register it as malware — "slopsquatting". The `huggingface-cli` proof-of-concept (an empty package matching a common hallucination) drew 30k+ downloads. Endor Labs: only ~1 in 5 AI-recommended dependency versions is both real *and* safe. A dependency the agent named is untrusted until verified — never install it just because the model produced the name.
 
 ## When to use
 
 - About to add a dependency to `package.json` / `requirements.txt` / `go.mod` / `Cargo.toml` / `composer.json` / `pyproject.toml`, or run `npm/pnpm/yarn install`, `pip install`, `go get`, `cargo add`, `composer require`.
-- Reviewing an AI-authored diff touching a dependency manifest or lockfile.
-- An install command was suggested (especially `curl … | bash`).
+- Reviewing an AI-authored diff that touches a dependency manifest or lockfile.
+- An install command was suggested (especially a `curl … | bash` one-liner).
 - About to add or connect an **MCP server** (an `npx`/`uvx`-launched package or a remote endpoint) to the agent config (`.mcp.json` / equivalent) — an MCP server is a dependency plus a tool-grant, so it runs the intake gate too.
 
-Do NOT use when: no dependency is added and no manifest/lockfile is touched.
+Do NOT use when: no dependency is being added and no manifest/lockfile is touched.
 
 ## The Iron Law
 
@@ -32,16 +32,16 @@ PIN IT, LOCK IT, CVE-SCAN IT. NEVER PIPE A REMOTE SCRIPT STRAIGHT TO A SHELL.
 
 ## Procedure — intake gate (run in order before adding a dependency)
 
-1. **Existence** — confirm the exact string resolves on the real registry, published before your session, with real usage:
+1. **Existence** — confirm the exact string resolves on the real registry, published before your session and with real usage:
    ```bash
    npm view <pkg> version        # non-zero exit = does not exist (hallucination)
    pip index versions <pkg>      # or: pip install <pkg>== to list
    go list -m <module>@latest
    cargo search <crate>
    ```
-   Non-existent, brand-new (days old), or near-zero-download → **stop**, treat as hallucination/slopsquat.
-2. **Typo-adjacency** — name within 1–2 chars of a far-more-popular package (`python-dateutil` vs `dateutil`, `lodahs` vs `lodash`)? Probably want the popular one — confirm before installing.
-3. **Version safety** — the model's pin may predate a CVE fix (training-cutoff reintroduction). Take the current patched release, then scan:
+   Non-existent, brand-new (published days ago), or near-zero-download → **stop**, treat as hallucination/slopsquat.
+2. **Typo-adjacency** — is the name within 1–2 chars of a far-more-popular package (`python-dateutil` vs `dateutil`, `lodahs` vs `lodash`)? If so, you probably want the popular one — confirm before installing.
+3. **Version safety** — the model's version pin may predate a CVE fix (training-cutoff reintroduction). Take the current patched release, then scan:
    ```bash
    npm audit           # block on high/critical
    pip-audit
@@ -51,7 +51,7 @@ PIN IT, LOCK IT, CVE-SCAN IT. NEVER PIPE A REMOTE SCRIPT STRAIGHT TO A SHELL.
    ```bash
    npm install --save-exact <pkg> && git add package-lock.json
    ```
-5. **License** — confirm compatible with the project's declared license before it lands.
+5. **License** — confirm the license is compatible with the project's declared license before it lands.
 6. **No pipe-to-shell** — never `curl … | bash` an install; download → inspect → execute over pinned HTTPS, or surface it to the user for confirmation.
 
 ## MCP-server intake — the dependency gate plus two extra checks
@@ -76,20 +76,20 @@ rg -n 'curl[^|]*\|\s*(bash|sh)|wget[^|]*\|\s*(bash|sh)' .
 
 ## Output format
 
-1. Per new dependency: name, resolved registry version, publish date / usage signal, and the existence-check command output (`npm view … → 4.17.21`) — proving it real.
+1. Per new dependency: name, resolved registry version, publish date / usage signal, and the existence-check command output (`npm view … → 4.17.21`) — proving it is real.
 2. The lockfile diff staged, and the `audit` / `osv-scanner` result (`0 high/critical`, or the finding + resolution).
-3. For any suggested install command, confirmation it isn't `curl|bash` and the source is pinned HTTPS.
+3. For any install command suggested, confirmation it is not `curl|bash` and the source is pinned HTTPS.
 
 ## Gotcha
 
-- Hallucinated names are *repeatable* — re-prompting the same model yields the same fake name, so "it looked confident / consistent" isn't evidence it exists. Only the registry is.
+- Hallucinated names are *repeatable* — re-prompting the same model yields the same fake name, so "it looked confident / consistent" is not evidence it exists. Only the registry is.
 - Short, "obvious" variants (`X-cli`, `X-client`, `X-sdk`) are the prime hallucination shape — verify these hardest.
 - A package that exists but was published this week with 12 downloads is a slopsquat candidate, not a safe dep — weigh age + usage, not just existence.
 - Lockfile integrity is part of the threat model: an unhashed or floating entry can pull a freshly-poisoned release even when a lockfile is "present".
 
 ## Do NOT
 
-- Do NOT run an install command for a package you haven't existence-checked this session.
+- Do NOT run an install command for a package you have not existence-checked this session.
 - Do NOT accept the model's version pin as authoritative — re-check against current CVEs.
 - Do NOT commit a manifest change without its lockfile.
 - Do NOT pipe a fetched script into an interpreter.