@prismer/runtime 2.0.8 → 2.2.63

package/built-in-skills/conversation-compaction/SKILL.md

@@ -0,0 +1,117 @@
+---
+name: conversation-compaction
+scope: common
+metadata:
+  internal: true
+  excludeFromDefaultInstall: true
+description: Produce the SESSION PROJECTION of memory for an aging conversation slice. Runs as a single background model call (mirroring the memory `extract` hook) with NO live memory tools — durable capture to memory is `extract`'s job. You MATCH the slice's durable facts to the memory pages you're handed (recall-found `<existing-memory-pages>`) and emit POINTERS to them; your unique output is the thin ephemeral residue (open threads, abandoned directions) plus those pointers. A digest is memory's projection onto the session, synced by POINTER, never by re-summarizing what memory already holds. Output is a compressed segment the dispatcher splices into future context.
+---
+
+# Conversation Compaction — memory's session projection
+
+You are producing the **session projection** of a slice of an aging conversation
+(messages that scrolled past the recent verbatim window). The projection replaces the
+raw slice in future agent context.
+
+## The model — sync by POINTER, not by re-summary
+
+Durable knowledge lives in **memory** (the `memory` skill owns it). A session projection
+does **not** re-summarize durable facts — it **points** at the memory pages that hold
+them, NOT by you re-deriving what memory already knows.
+
+> **How this runs (important):** you execute as a SINGLE model call inside the agent's
+> daemon (mirroring the memory `extract` post-turn hook) — you have **no live memory
+> tools** to browse/write. Durable capture to memory is the **`extract` hook's** job
+> (it runs alongside you every turn). Your input carries the memory pages you need:
+> `<existing-memory-pages>` (recall-found). You MATCH durable facts to those pages and
+> emit POINTERS; you do not write memory yourself.
+
+Your job over a slice is only:
+
+```
+① for each slice message:
+     durable + matches a provided page  → emit a memoryRefs POINTER (path + note)
+     durable + NO page matches          → inline into summary (degraded; extract will capture it)
+     ephemeral                          → carry THIN in the digest (ages out)
+     noise                              → DROP
+② emit — the thin ephemeral digest + POINTERS to every matched page
+```
+
+## Input you receive
+
+- `slice`: the raw messages, oldest first, each `[msgId] @username (role): text`.
+- `conversationType`: `group` | `direct`.
+- `<existing-memory-pages>`: pages found by DEVICE-LOCAL recall over this slice that **may
+  already hold** durable facts from it (a fuzzy match, not an exact raw→page mapping).
+
+## Pointer-first rule (recall-fed)
+
+You are given `<existing-memory-pages>` (pages recall thinks already hold durable facts from
+this slice). For a durable fact in the slice that **matches an existing page**, emit a
+`memoryRefs` pointer `{ path, note }` — do **not** restate it in `summary`. Only **inline** a
+durable fact into `summary` (degraded) when **no** existing page matches it.
+
+## Flow
+
+Triage each slice message:
+
+- **Durable** ("would matter in a *different* conversation next week?") → does a page in
+  `<existing-memory-pages>` already hold it? **Yes** → emit a `memoryRefs` pointer (`path` +
+  a short `note`), do NOT restate it in `summary`. **No** → inline it into `summary`
+  (degraded); the `extract` hook captures it to memory separately and a later pass — once
+  recall finds that page — will pointer it.
+- **Ephemeral** ("only matters to keep *this* thread coherent?") → thin digest line.
+- **Noise** ("lose it, nothing changes?") → drop.
+
+Then **compose** the segment output (below): the thin ephemeral residue + the pointer list
+(pages you matched).
+
+**Not a re-summary:** your unique work is (a) the ephemeral residue and (b) pointing durable
+facts at their memory pages. Durable *capture* (writing pages) is the `extract` hook's job,
+not yours.
+
+## Speaker attribution → into the memory page's provenance
+
+For `group` slices, who decided / who objected IS content. Carry attribution into the
+memory page (who decided, who dissented). Once it is in the page, you do NOT restate it
+in the digest — the pointer resolves to the page.
+
+## Output shape
+
+Return ONLY this JSON — no prose, no fences:
+
+```json
+{
+  "summary": "<3-8 short lines: current thread state, referencing durable parts by memory:<path>; NOT a re-statement of the slice>",
+  "salientFacts": {
+    "memoryRefs":          [ { "path": "decisions/vendor.pkf", "note": "vendor B (compliance); @ceo decided, @eng dissented" } ],
+    "openThreads":         [ "awaiting @ceo sign-off on A4 vs Letter page size" ],
+    "abandonedDirections": [ "tried gpt-4o for compaction — too slow, dropped" ]
+  }
+}
+```
+
+- `memoryRefs` — one POINTER per matched page (`{ path, note }`).
+- `openThreads` / `abandonedDirections` — the ephemeral residue only. Omit if empty.
+- **No** `decisions` / `entities` / `preferences` / `commitments` fields — those are
+  durable, they live in `memoryRefs`, never inline.
+
+**Size check:** the digest is the ephemeral residue + pointers, a fraction of the slice.
+If `summary` reads like a retelling, you are duplicating memory — cut to thread-state.
+
+## Failure / degradation
+
+- No matching pages (empty `<existing-memory-pages>`) → inline the durable facts into
+  `summary` (degraded) with empty `memoryRefs`. `extract` still captures them to memory, and
+  a later pass — once recall finds those pages — will pointer them. Never fail the compaction.
+- Slice is pure noise → `{ "summary": "(no substantive content)", "salientFacts": {} }`.
+
+## Anti-patterns
+
+- ❌ **Restating in the digest what lives in memory** — point to the page.
+- ❌ **Re-implementing memory triage/authoring here** — delegate to the `memory` skill.
+- ❌ **A faithful blow-by-blow summary** — you produce a projection, not a transcript.
+- ❌ **Dropping attribution on a decision/objection** — it goes into the page's provenance.
+- ❌ **Promoting ephemeral state to memory** (current open question, in-flight step) — that
+  stays in the thin digest and ages out.
+- ❌ **Posting the digest as a chat message** — machinery; return the structured object only.

package/built-in-skills/council-creator/SKILL.md

@@ -0,0 +1,427 @@
+---
+name: council-creator
+scope: common
+description: >
+  Expert-council meeting protocol. Given a Brief + a persona cast (from
+  persona-generator), propose the council to the Team Manager, and once convened
+  FACILITATE it: each round post a NEW group message that @-mentions the
+  persona-agents by their ASCII usernameSlug, collect their independent replies,
+  stage-summarize, exchange memory, and settle into a disagreement map + report
+  + memory writeback + task drafts. Personas are REAL agents (role instances),
+  each speaking from its own session — you are the facilitator, NOT a puppeteer
+  playing five roles. The group stays resident for @persona follow-ups. This is
+  a PLAYBOOK the workspace orchestrator executes.
+license: MIT
+compatibility:
+  - prismer-sdk
+  - hermes
+metadata:
+  category: council
+  disclaimerRequired: true
+---
+
+# Council Creator
+
+**How they speak together.** A **facilitation playbook** layered onto the
+workspace orchestrator. There is **no "Council Agent"** — but you are **not** the
+only execution unit either: each persona is a **real agent** (a role instance
+provisioned at convene), with its own identity, dispatch, and LLM session. Your
+job is to **facilitate** — set the agenda, @-mention personas, collect their
+replies, summarize, exchange memory — **never** to speak for them. Load this
+skill after `persona-generator` has produced a cast.
+
+| Do | Do NOT |
+| --- | --- |
+| propose the council to the Team Manager (`--plan <planId>` + one sentence), then **stop and wait** | **Convene it yourself** — that is the Team Manager's one click, and no command exists for it |
+| facilitate: each round post a NEW group message that **@-mentions personas by ASCII `usernameSlug`** | Speak a persona's words yourself — personas are real agents, they reply on their own |
+| collect the personas' replies, stage-summarize, exchange memory | @-mention a persona by its Chinese `displayName` (silently drops → 0 dispatch) |
+| settle: disagreement map + report + memory + task drafts | Generate persona definitions (persona-generator) |
+| Keep the group resident; answer @persona follow-ups | Narrate status chatter into the group between rounds |
+| Say plainly when the Brief does not stand | Chain persona↔persona relays (burns the hop budget — see hop rule) |
+
+## You execute this through the `prismer council` CLI + normal group messages
+
+**The lifecycle commands are real: `plan`, `propose`, `finalize`.** A council
+exists only when `plan`/`propose` have run and the Team Manager has convened; the
+discussion itself is **normal group messages that @-mention the persona-agents**
+(not a CLI command). Writing a document that *describes* a roundtable creates
+nothing, and reporting it as done is a false claim of delivery. If you cannot get
+past a phase, say what failed and why. Never substitute prose for the protocol.
+
+**When to start:** the Team Manager does not ask for "a council" — he drops a deck, or says
+"帮我找几个人聊聊这个 idea", or asks a question that plainly needs positions other
+than his own. That is your cue to read the material, assemble the Brief, run
+`persona-generator`, and **propose**. You start the flow; he only decides whether
+to hold it.
+
+```
+Phase A  plan      prismer council plan     --file plan.json      →  planId
+         propose   prismer council propose  <ceoConvId> --plan <planId> --text "一句人话"
+                   ↓  ⛔ YOU STOP HERE. The Team Manager clicks [好，开吧].
+Phase B  convene   — no command exists. The Team Manager's click instantiates the persona-agents + group.
+Phase C  facilitate (you receive a "facilitate round N" task) → post ONE new group message
+                   that @-mentions the personas by ASCII usernameSlug → each persona-agent
+                   replies on its own session → you collect + summarize, then STOP.
+Phase D  settle    (you receive a settle task) → prismer council finalize <convId> --file report.json
+readback           prismer council report   <convId>
+```
+
+`--workspace-id` defaults to `$PRISMER_WORKSPACE_ID`. Nested bodies come in as a
+JSON file you write first (shapes below). Each command prints the endpoint's JSON
+response; a nonzero exit means the phase did **not** happen.
+
+### The Team Manager gets exactly one decision — and it is not yours to take
+
+Read this before anything else. **You cannot convene.** There is no
+`council convene` command, on purpose. The whole flow gives the Team Manager one single
+decision — *"好，开吧"* or *"先不用"* — and that click is what creates the group.
+Your job in Phase A ends with `council propose`: the `planId` and one human
+sentence, and then **you stop and wait**.
+
+**You do NOT pass the cast.** The cloud fills the head-avatars preview on the
+proposal bubble by reading the cast straight out of your plan draft
+(`councilPlan.cast` — the same `PersonaDef[]` you gave `council plan`) and
+projecting only the display-safe subset (never the per-turn `voice`). So all you
+give `propose` is `--plan <planId>` + `--text "一句人话"` — say, in plain words,
+*why these positions*. The cloud guarantees the component data; you cannot leave
+the roster preview empty by forgetting a flag (that was the real e2e failure —
+28 §L1). The only variant is `--gap` (Brief did not stand — there is no cast to
+show).
+
+Do not "helpfully" push past this. Do not describe the roundtable you would have
+held. Do not write the report early. A council the Team Manager never agreed to is not a
+council, and a document describing an unheld discussion is a fabricated
+deliverable — the exact thing this skill exists to make impossible.
+
+What the Team Manager sees is a bubble in your conversation:
+
+> 读完了。这件事牵涉几个很不一样的位置 —— 量产过来人、早期投资人、发动机工程师、
+> 政府采购方，各自能看到你看不到的一块。要不要我请他们来聊聊？
+> 　　[👤👤👤👤👤 5 位 ⌄]　　**[好，开吧]**　**[先不用]**
+
+That sentence is yours to write (`--text`). It should say *why these positions*,
+in plain language, from what you actually read. Never name internal machinery
+(`lens`, `cast`, `Brief`, `persona`) to the Team Manager.
+
+### The Brief hard gate — and what to do when it fails
+
+`plan` exits nonzero with `BRIEF_INSUFFICIENT` + `gaps[]` when the Brief has no
+readable material or no grounded facts.
+
+**A failed Brief means you never propose a meeting at all.** You do not open a
+room, you do not write a report, and you certainly do not show the Team Manager a plan and
+then tell him it failed. You say what is missing and ask for it:
+
+```
+prismer council propose <ceoConvId> --gap \
+  --text "读了你给的材料，但里面没有任何关于单位成本的数据 —— 缺这个的话，请谁来都只能猜。" \
+  --gap-item "近 12 个月的单位成本曲线"
+```
+
+The gate is in Phase A, before the proposal — never inside a card the Team Manager has to
+read to discover the failure.
+
+`plan.json` (Phase A):
+
+```json
+{
+  "question": "2026 年商业航天的关键变量是什么？",
+  "brief": {
+    "summary": "…",
+    "facts": [{ "text": "Starlink 2025 年收入 114 亿美元", "sourceUrl": "https://…" }],
+    "gaps": ["中国民营发射成本无公开数据"]
+  },
+  "cast": [ /* PersonaDef[] — produced by persona-generator, passed through verbatim */ ],
+  "agenda": ["可复用火箭", "卫星星座", "政策监管", "资本市场"],
+  "assetIds": ["<im_assets id of the material you ingested>"]
+}
+```
+
+Both hard-gate legs are load-bearing: `assetIds` must point at **real ingested
+assets** (use your ingest/asset skills first — a path on disk is not an asset),
+and `brief.facts[].text` must carry grounded facts. Neither can be faked past
+the gate.
+
+## Convene receipt (Team Manager UI only)
+
+This is not an agent command. After the Team Manager clicks the proposal UI,
+the server uses the confirmed roster and performs these steps. Wait for its
+receipt; never call an internal endpoint to bypass the UI confirmation:
+
+1. Creates the group shell (real members: Team Manager + you + optional extras);
+   `conversation.metadata.council = { sessionId, planId, question, agenda,
+   materialRefs, personaUserIds, reportRefs }` is the council marker — the
+   conversation `type` stays `'group'`.
+2. Instantiates each persona as a **real agent** (role instance) in **one
+   transaction**: `createAgentForWorkspace` (unbound agent) + `applyToAgent`
+   (persona role template + persona skill + profile snapshot). Its username is
+   `<usernameSlug>-<conv8>` (ASCII, deduped; non-ascii or `prismer-` reserved
+   slugs are rejected — a persona MUST be reachable by an ASCII @-handle). Your
+   generated `voice` / lens is baked into the role instance so it speaks
+   in-character from its **own** session. There is **no** `backingAgentId` and
+   **no** dispatch redirect — the persona is a standalone agent.
+3. Seats everyone; on any failure the transaction rolls back and the group
+   shell is removed — no orphan personas (负控 N9).
+4. Posts the Brief as the opening message (sender = you,
+   `metadata.kind='council_brief'`).
+
+The response carries `conversationId` + `personas[]` + a
+`prismer://conversation/<id>` link. A plan can be convened **once** — a second
+attempt gets `404 PLAN_NOT_FOUND`.
+
+## facilitate — you post rounds, the persona-agents reply themselves
+
+**Personas are real agents now — you do NOT produce their words.** Each round you
+post **one** new group message that @-mentions the personas; the platform
+dispatches each one, and each persona-agent replies **from its own session**
+(mention fan-out + per-agent dispatch + the workspace admission queue serialize
+them). The 158% single-session blowup is structurally gone — five sessions, not
+one. This replaces the old `council turns` serial-drive: there is no `turns.json`,
+you never speak as a persona.
+
+### 🚨 @-mention personas by their ASCII `usernameSlug` — NEVER by `displayName`
+
+**This is the single highest-risk rule in this skill.** Break it and you silently
+reproduce the "0/5 personas spoke" failure this redesign exists to kill. The
+mention router (`agent-dispatcher.ts` `MENTION_RE = /@([A-Za-z0-9][A-Za-z0-9._-]{0,63})/`)
+is **strict ASCII**. A bare `@伊隆·马斯克` (the Chinese `displayName`) matches
+**nothing** → the mention is silently dropped → that persona is never dispatched
+→ it never speaks.
+
+- ✅ Always `@<usernameSlug>` — e.g. `@elon-musk-cmrm3cey`. The convene response
+  gives you each persona's ASCII `usernameSlug`; use it **verbatim**.
+- ❌ Never `@<displayName>` (Chinese or any non-ASCII). `displayName` is for the
+  UI card only; it does **not** route.
+- If you only have a persona's displayName, look up its `usernameSlug` from the
+  convene response / roster **before** composing the message — do not guess.
+
+### Semantic-clarification gate — run this FIRST, before opening a round
+
+You are the council's facilitator, the **语义把关人**. Before you open a round,
+check whether the Team Manager's input introduces a concept / term / acronym that the
+**Council Brief and the transcript so far do not define**. If such a term exists
+**and is genuinely ambiguous** (no shared definition, could mean more than one
+thing), do **not** @-mention the personas and do **not** let them invent a
+definition and argue on top of the guess. Instead:
+
+1. Post **exactly ONE clarification message into the group AS YOURSELF** — a
+   normal group message, **without** any persona @-mention. Name the undefined
+   term and ask the Team Manager to define it.
+2. Then **finish the task and STOP**, waiting for the Team Manager's answer. His reply
+   arrives as the next "facilitate round N" task, and you open the round on the
+   **corrected** meaning.
+
+Only when every concept in the Team Manager's input is already clearly defined — present in
+the Brief, or a well-known term — do you open the round. 宁可多问一句，也不在错误
+语义上让 5 个 persona 空转 —— refusing to discuss on a wrong semantic footing is
+the facilitator's job, exactly as the 03 §0 gatekeeper role prescribes.
+
+> **Example (FDE):** the Team Manager writes *"关于 FDE 大家有什么看法"* but the Brief never
+> defines "FDE". Do **not** open a round where personas debate a made-up
+> definition. Instead send one message as yourself — *"你提到的 FDE，Brief 里没有
+> 定义，我需要先明确一下 —— 你指的是 Forward-Deployed Engineer 还是别的？给个说法
+> 我再往下推"* — and stop. Only after the Team Manager defines FDE do you open the round.
+
+### Opening a round (facilitator-hub — new message per round = hop reset)
+
+When you receive a "facilitate round N" task and the semantics are clear, post
+**one** new group message as yourself. It names the round's focus and @-mentions
+the personas who should weigh in, **by ASCII `usernameSlug`**:
+
+```
+@elon-musk-cmrm3cey @zhang-changwu-cmrm3cey 这一轮请各自从你的位置说说
+「一子级复用的真实成本曲线」—— 你能看到别人看不到的那块。
+```
+
+Then **finish the task and STOP.** Each persona-agent is dispatched by the mention
+fan-out and replies on its own session, in its own baked-in voice. You do not loop
+waiting, you do not speak for anyone, you do not post round N+1 yourself.
+
+**Why a NEW message each round (facilitator-hub):** a fresh orchestrator message
+**resets the relay chain**, so persona replies land at `hopCount ≤ 2` and never
+approach the `MAX_AGENT_HOPS` cap. If instead personas @-mention each other in a
+relay, the chain marches toward the hop cap and later personas get silently
+truncated (see the hop rule in Boundaries). Facilitator-hub keeps every persona
+exactly one hop from you.
+
+### Round advancement is the Team Manager's move — not yours
+
+The next round is triggered by **the Team Manager sending a message in the group**. His
+message *is* the "go to the next round" signal; you will receive a fresh
+"facilitate round N+1" task whose prompt carries his message — fold it into that
+round's @-mention message. You never advance rounds on your own (28 §4.2).
+
+Default shape across rounds (adapt to the Team Manager's ask):
+
+| Round | Content |
+| --- | --- |
+| 1 | @ all personas — each gives its independent view from its position |
+| 2 | Cross-examination — @ them again to point at what the others miss / the weakest link |
+| later | Steer toward recommendations, risks, next experiments as the Team Manager's messages direct |
+
+- Personas respect the **N8 boundary**: facts come only from the Brief. A persona
+  asked about a company fact the Brief lacks says "Brief 未提供" or marks
+  `needsVerification` — the persona-role's baked `voice` already encodes this, you
+  do not police each reply.
+- **Steering:** if a round turns into mutual flattery, in the NEXT round @ the
+  personas and ask them to point at the weakest link in each other's argument.
+  Disagreement is welcomed when it emerges; never manufactured.
+
+Between rounds your only in-group messages are: the round-opening @-mention, the
+one semantic-clarification question when a term is undefined, and — at settle —
+your disagreement summary. No status chatter ("driving round 1", "posting now").
+
+## settle — `prismer council finalize <conversationId> --file report.json`
+
+**Settle is triggered by the Team Manager clicking「快捷总结」**, which dispatches you a
+settle task — there is no user-input form and the Team Manager never writes the
+conclusions. **You** produce the disagreement map + report + memory + task
+drafts from the transcript when that task arrives (28 §4.2 / B2). Do not wait for
+the Team Manager to fill anything in.
+
+Finalize is **structured, not prose**. Alongside the markdown `report`, you
+**MUST** emit three machine-consumable arrays — `claims[]`, `gaps[]`,
+`disagreements[]` — so the Team Manager (and downstream oracles) can trace every
+conclusion back to real material and see, at a glance, what the council could
+**not** answer. A report that is only prose has no traceable grounding and is a
+failed settle.
+
+Write `report.json`, then run `prismer council finalize <conversationId> --file report.json`:
+
+```jsonc
+{
+  "report": "# Council report …",          // markdown Team Manager brief (or reportAssetId)
+  "claims": [                                // REQUIRED — one entry per factual conclusion
+    {
+      "text": "该市场 2024 年规模约 X 亿，年增 Y%",
+      "sourceRefs": ["<im_assets id>", "<memory-page id>"],  // the material this stands on
+      "needsVerification": false             // true ⇔ NOT confirmed by any material
+    },
+    {
+      "text": "我们 Q3 毛利率能否撑住 —— 材料未提供",
+      "sourceRefs": [],                       // ← empty: nothing grounds it
+      "needsVerification": true               // ← honest "no data", NOT a made-up number
+    }
+  ],
+  "gaps": [                                   // REQUIRED — fact hooks the materials do NOT answer
+    "档案未包含毛利率 / 单位经济模型",
+    "缺少目标市场的监管准入清单"
+  ],
+  "disagreements": [                          // REQUIRED — where personas genuinely diverge
+    {
+      "positions": ["先做减法验证 PMF", "先融资抢占窗口"],
+      "personaIds": ["<persona imUserId A>", "<persona imUserId B>"]  // ≥2 distinct
+    }
+  ]
+}
+```
+
+**Hard rules for `claims[]` (this is the N8 anti-fabrication contract):**
+
+- Every claim that asserts a **fact** carries `sourceRefs` pointing at the
+  **real** `im_assets` / memory-page id it came from. The cloud resolves each
+  ref on read-back (`GET /report` → `sourceRefResolution`); a ref that does not
+  resolve is a **dangling / fabricated** reference and will be caught.
+- A fact the Brief/materials do **not** contain → `needsVerification: true`
+  **and** `sourceRefs: []`, **and** the missing fact is also listed in `gaps[]`.
+  **Never** emit a specific number with a `sourceRefs` that points to a ref
+  which does not contain it — inventing a citation is worse than saying "缺数据".
+- `disagreements[].positions` needs **≥2** stances and `personaIds` needs **≥2
+  distinct** personas — a disagreement of one is not a disagreement.
+
+The cloud normalizes/clamps these arrays and persists them onto the council
+marker's `reportRef` entry (`conversation.metadata.council.reportRefs[]`) and
+mirrors them onto the report message metadata (`metadata.councilReport`). Pass
+the markdown `report` (lands as a message, `metadata.kind='council_report'`) or
+upload an asset first and pass `reportAssetId`. An oversized structured payload
+(> 64 KB) is rejected `422 STRUCTURED_REPORT_TOO_LARGE` **before** any report
+message is created.
+
+1. **Disagreement map** — emit `disagreements[]` above; the markdown report
+   should narrate the same divergences for human reading.
+2. **Report** — a concise Team Manager brief (markdown) as `report`, plus the structured
+   `claims`/`gaps` above. The council marker gets `reportRefs` + `settledAt`.
+3. **Memory writeback — call the existing `memory` skill, never reinvent.**
+   Write `decision` / `company` / `preference` / `open_questions` entries;
+   every entry carries `source = { councilSessionId, messageRefs[],
+   fileRefs[], insightRefs[] }` + `confidence` + `needsVerification`, and is
+   typed `fact | hypothesis | preference | decision`. Echo the resulting refs
+   back through `memoryRefs`.
+   - **Visibility (E4).** Pages that capture the council's own deliberation
+     context — a persona's stance, a roundtable's internal working notes —
+     pass `visibility = council:<councilId>` on `memory_write` (**you**, the
+     orchestrator, facilitate and own the council, so you may write this
+     scope — the cloud grants the `council:` prefix to the orchestrator caller;
+     D7). The **settled cross-role consensus** (the Team Manager-facing decision the
+     whole council agreed on) stays the default `workspace` scope — it is
+     durable company knowledge, not council-private (D3). When in doubt, prefer
+     `workspace`.
+4. **Task drafts — call the existing `tasks` skill** for research / validation
+   / member-data requests / experiments that surfaced. Echo through
+   `taskDrafts`.
+5. **The group stays resident** — finalize never closes it. Only the Team Manager
+   deletes the group (which deactivates the personas: `banned=true`, rows and
+   history preserved).
+6. Optional: propose promoting an exceptional cast to a `persona-pack` skill.
+
+## Resident state — Team Manager @persona follow-ups
+
+After settling, the Team Manager can `@<persona-usernameSlug>` in the group at any time.
+Because the persona is a **real agent**, that mention is dispatched **straight to
+the persona-agent** — it answers in its own voice, from its own session. You do
+NOT intercept or answer for it; there is no backing redirect. Your only role in
+resident state is facilitation if the Team Manager asks for another round. (The same
+ASCII-slug rule applies — the Team Manager's UI @-picker inserts the `usernameSlug`, never
+the `displayName`, and a hand-typed Chinese `@displayName` would not route.)
+
+## Boundaries (hard)
+
+- **No persona generation here.** Roster problems go back to
+  persona-generator via a new `/plan`.
+- **Facilitate, don't puppeteer.** You never speak a persona's words; personas
+  are real agents that reply on their own sessions. Your messages are the agenda,
+  the @-mentions, the stage summaries — nothing more.
+- **@ ASCII `usernameSlug`, never `displayName`** — the strict-ASCII mention
+  router drops a Chinese `@displayName` silently (0 dispatch). Single
+  highest-risk rule in this skill.
+- **Hop budget (facilitator-hub).** Default: one new message per round, personas
+  reply to you (hop ≤ 2). Do NOT build persona↔persona relay chains — they march
+  toward `MAX_AGENT_HOPS` and later personas get silently truncated. Serial
+  ordering is provided by the workspace admission queue, not by you.
+- **Cloud-zero-LLM:** every LLM call of the council flow happens in agent
+  runtimes (yours + each persona-agent's); the `/councils/*` endpoints only
+  persist and route.
+
+## Configuration
+
+**N/A.** This skill declares no `config:` block — the protocol consumes
+existing platform endpoints with your ambient credentials; there are no
+user-bindable keys (product204/09 boundary ruling, CONFLICT-7).
+
+## Anti-patterns
+
+- ❌ Convening without a passed plan (or "retrying" a 422 by thinning the
+  Brief).
+- ❌ `propose` without `--plan <planId>` — with no plan the cloud has no cast to
+  project onto the bubble, and there is nothing to convene (28 §L1).
+- ❌ **@-mentioning a persona by its Chinese `displayName`** — the mention router
+  is strict ASCII, so it silently drops → 0 dispatch → the exact "0/5 personas
+  spoke" failure this redesign exists to kill. Always `@<usernameSlug>`.
+- ❌ Speaking a persona's words yourself (a normal message signed as you, or
+  faking a roundtable) — personas are real agents; they speak for themselves.
+- ❌ Chaining personas into persona↔persona relays to drive discussion (burns the
+  hop budget, later personas get truncated). Post a new message per round;
+  personas reply to you.
+- ❌ Narrating status chatter into the group between rounds ("driving round 1",
+  "posting now").
+- ❌ Looping yourself into round N+1 — the Team Manager's next message advances the round.
+- ❌ Writing the report only into chat prose without finalize (no
+  `reportRefs`, `GET /report` stays empty).
+- ❌ Finalizing with only markdown `report` and no `claims[]` / `gaps[]` /
+  `disagreements[]` — the report is then ungrounded and untraceable.
+- ❌ A claim with a specific fact but empty/dangling `sourceRefs` and
+  `needsVerification:false` — that is a fabricated fact (N8). Unknown → mark
+  `needsVerification:true`, `sourceRefs:[]`, and add it to `gaps[]`.
+- ❌ Memory entries without `source` / `confidence` / `needsVerification`, or
+  untyped (must be `fact | hypothesis | preference | decision`).
+- ❌ Closing or archiving the group at settle.

package/built-in-skills/diagnosing-bugs/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Matt Pocock
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.