@easysociable/cli 0.1.8 → 0.1.10

package/content-engine/starter/CLAUDE.md

@@ -1,908 +0,0 @@
-# Content Engine — Agent operating manual
-
-This folder is the local **Content Engine** for EasySociable.
-
-**Human interface:** conversation only. The user does **not** edit markdown by hand. Observatory UI: `workbench/` at `http://127.0.0.1:8765/workbench/`.  
-**Agent interface:** you create/update files; always return a structured result card in chat. After markdown writes, run `python3 workbench/build.py` (leave `serve.py` running if already up).
-
-## Defaults (v1)
-
-- Folder tree: **English** (`Content Engine/`). Do not ask language.
-- Default **account:** first **active** handle in `accounts/_index.md` (starter: `your-handle` — replace it)
-- **开一单:** do **not** silent-default platform — **ask**; multi-select → **N runs**  
-- Do not invent metrics, customers, or quotes the user did not provide
-
-### Resolve account → profile (required)
-
-1. Read **`accounts/_index.md`** for the handle (default: first active row)  
-2. Get `profile_id`, `voice`, `platforms`, `default_platform`, `overrides`  
-3. Load **`profiles/<profile_id>/pillars.md`** + **`audience.md`** (core + current platform section)  
-4. Voice: if account `voice` is `profile` or empty → **`profiles/<profile_id>/voice.md`** (Identity + **Experience** + **Values** + Persona); else load the explicit path. Never invent bio facts not in Identity / Experience.  
-   `boundaries.md`: missing → **hard-gate fail**. Write empty starter (`## Not covering` / `## Never claim` headings), then interview; **never invent boundaries**.  
-   `stories.md` / `expression.md`: missing → `profile: partial`, continue.  
-   Do not invent bio / stories / values absent from those files. Topics listed in `boundaries.md` → stop; do not write.  
-5. If `overrides` is **non-empty**, read that path (e.g. `accounts/overrides/<handle>.md`) and apply last  
-6. Craft: **always** load `platforms/<platform>.md` for the active platform (see `platforms/README.md` for id → file; `instagram` → `ig.md`). Missing file → stop and tell the user; do not invent craft rules.  
-7. Catalog filter: swipe/atoms/claims where `platforms` matches **and** (`profiles` is `*` **or** contains `profile_id`)
-8. If a Topic or draft mentions an offer, load `products/_index.md` and the
-   cited `products/entries/P-*.md`. Never invent product pricing, entitlements,
-   proof, or outcomes. Product fit is optional; educational content may have no
-   product.
-
-## File map
-
-| Path | Role |
-|------|------|
-| `CLAUDE.md` | Operating manual |
-| `profiles/_index.md` | Profile registry |
-| `profiles/<id>/pillars.md` | Topic pillars (what to talk about) |
-| `profiles/<id>/audience.md` | Readers + platform emphasis |
-| `profiles/<id>/voice.md` | Identity + Experience + Values + Persona |
-| `profiles/<id>/stories.md` | Story inventory (cite or none) |
-| `profiles/<id>/boundaries.md` | Creator must-avoid (Not covering / Never claim) |
-| `profiles/<id>/expression.md` | Expression likes / dislikes |
-| `profiles/<id>/lanes.md` | Publishing lanes; distinct from pillars |
-| `profiles/<id>/keywords.md` | Search recipes; scan may stage `hits/` |
-| `profiles/<id>/handles.md` | ≤8 ally/foil/competitor handles |
-| `hits/_index.md` | Research triage (not Topics) |
-| `accounts/_index.md` | handle → profile, platforms, voice, overrides |
-| `accounts/overrides/<handle>.md` | Optional account deltas |
-| `platforms/*.md` | Platform craft only |
-| `inbox/_index.md` | Capture registry |
-| `inbox/entries/C-*.md` | Per-capture working notes |
-| `needs/_index.md` | Audience-quote registry |
-| `needs/entries/N-*.md` | Per-need quote + sightings |
-| `products/_index.md` | Owned product registry and lifecycle |
-| `products/entries/P-*.md` | Owned product facts, boundaries, pricing and claims |
-| `recommendations/_index.md` | Third-party tool/product registry and Profile associations |
-| `recommendations/R-*.md` | Review subject facts, evidence and disclosure |
-| `raw/_index.md` | Authoritative source registry |
-| `raw/*.md` | Source bodies |
-| `topics/_index.md` | Topic queue (one row per topic + `item` + src engagement) |
-| `topics/items/T-*.md` | Opportunity card (human half) + Score/Trace appendix |
-| `topics/scoring.md` | Scout gate + opportunity five-dim (context/audience/conflict/insight/evidence) |
-| `swipe/_index.md` | Structures (+ `profiles` column) |
-| `wiki/atoms/_index.md` | Atoms (+ `profiles`) |
-| `wiki/claims/_index.md` | Claims (+ `profiles`) |
-| `runs/_index.md` | Job queue (one row per **platform run**) |
-| `runs/<slug>/` | One platform draft → pack → feedback |
-| `published/_index.md` | One row per **shipped URL** (one run → usually one row) |
-| `exports/<slug>/` | Rendered images for that run (optional `<platform>/` subfolder) |
-
-## Routing
-
-| User intent | Action |
-|-------------|--------|
-| capture / 记一下 (+ link) | **§ Capture** only — **not** topics by default |
-| 记需求 / 用户原话 / 评论导出 | **§ Needs** only |
-| 加关键词 / 更新关键词库 | Edit `profiles/<id>/keywords.md` (**ask profile** if unset) |
-| 从需求生成关键词 / 生成 listen\|search\|ask 词 | **§ Keywords · Generate** (chat → confirm → write) |
-| 加/改 handles | Edit `profiles/<id>/handles.md`; keep **≤8** |
-| 扫关键词 / 扫 K-xx / 看 @handle | **§ Keywords** scan (**listen** default) → **§ Hits** |
-| 路由 Hit / H-xx | **§ Hits** → Needs or Capture+library / discard |
-| **扫需求选题** | **§ Topics** from `needs/` where `status=captured`; max 5 `produce` |
-| 扫 inbox 选题 | **§ Topics** — from captures without topic yet (max 5) |
-| 生成 topic (no source specified) | **§ Topics** — offer captured needs or ask which explicit input to use |
-| 链接直接选题 / 把 C-xxx 做成选题 | **§ Topics** — create T- (may link capture) |
-| 采访我 / 完善人设 / interview me / who am I as a creator / 建立 profile | **§ Interview** — persist after each dimension; no slideshow |
-| 开一单 / 用 T-xxx 开一单 | **Hard gate first**, then **§ Open a run** — **ask platform(s)** if unset; multi-select → **N runs** (bind `topic_id` if from T-) |
-| Quick draft | Resolve profile; selection; draft in chat |
-| 收成 swipe/atom/claim | **§ Library ingest** after approval |
-| 选题后同意进 inbox | Capture that URL/C-link then optional library offer |
-| Post URL after ship | feedback + **§ Published index** |
-| 复盘 / 批准晋级 | Claims promote + published index |
-| Update pillars/audience | Edit **`profiles/<id>/…`** |
-| 更新产品 / 产品库 | Edit **`products/`**; keep dates and source of truth |
-| 推荐 / 评测第三方产品 | Edit **`recommendations/`**; record evidence and disclosure |
-| Update account binding | Edit **`accounts/_index.md`** |
-| 出图 | **Hard gate first**, then **§ PagePack + render**. Fail → list missing files; no handoff unless user says **强制继续** / **force continue** |
-
----
-
-## § Capture (inbox + raw indexes)
-
-### When
-
-User stores an idea, link, competitor structure, or source — not a pure Q&A.
-
-### Always write (scheme B)
-
-1. **`inbox/entries/C-YYYYMMDD-XX.md`** — working notes (our understanding; not necessarily full source text)  
-2. **Row at top of `inbox/_index.md`** — registry (id, type, one_liner, profile, entry path, raw path, library, status)  
-
-Id: `C-YYYYMMDD-XX` sequential per day (`01`, `02`, …).  
-Default `profile`: current account’s `profile_id`.
-
-### Also write raw when
-
-- URL worth revisiting, multi-step structure, media, or user says important  
-
-Then:
-
-1. Body: `raw/YYYY-MM-DD-short-slug.md` (authoritative text/structure/media links)  
-2. Row in **`raw/_index.md`** with `capture_id`  
-3. Set inbox row `raw` column to that file  
-
-### Do not
-
-- Use deprecated `inbox/capture.md` (removed)  
-- Silent-write swipe/atoms/claims  
-- Treat capture as a finished run  
-- Reply “saved” without **Capture Card**  
-- Invent facts not in the user message / source  
-
-### entry file (minimal shape)
-
-```markdown
----
-id: C-YYYYMMDD-XX
-date: YYYY-MM-DD
-type: structure | link | insight | product | other
-profile: default
----
-
-# C-… — short title
-
-## One-liner
-## User intent / source
-## Key points (our notes)
-## Links (raw if any)
-## Status
-```
-
-### Capture Card (chat)
-
-```text
-✅ Captured [C-YYYYMMDD-XX]
-
-【One-liner】 …
-【Type】 …
-【Profile】 (resolved profile_id)
-【Key points】
-- …
-
-【Stored】
-- inbox/_index: yes
-- entry: inbox/entries/C-….md
-- raw: no | raw/….md + raw/_index
-
-【Refine】
-A) OK as stored
-B) Refine notes
-```
-
-### Library offer (required if type is structure or link)
-
-```text
-【入库？（需点头）】
-1) 不入库 — index library=none
-2) 收成 swipe
-3) 收成 atom (type: hook|reframe|proof|process|cta|other)
-4) 收成 claim
-5) 组合 …
-
-Reply number(s) or 不入库.
-```
-
-On approval: create catalog files per **§ Library ingest**, set inbox `library` column (e.g. `swipe:S2`).  
-Path A (capture): **no** engagement gates on *your* posts.  
-
-Optional: if user also says「同时选题/打分」, run **§ Topics** for this capture.
-
----
-
----
-
-## § Products (offer facts)
-
-### When
-
-User defines, updates, prices, retires, or asks about a product or offer.
-
-### Role
-
-Products are the **supply-side fact layer**. They are not audience needs,
-external discoveries, or content opportunities.
-
-```text
-Need (demand evidence) + Profile (identity) + optional Product (offer facts)
-  → Topic (editorial judgment)
-  → Run
-```
-
-### Rules
-
-- Read the cited `products/entries/P-*.md` before using product facts in a
-  Topic, brief, draft, CTA, or product answer.
-- A Topic may cite `product_ids: none`; do not force a product connection.
-- A Topic may cite `recommendation_ids: none`; recommendations are third-party
-  review subjects, not owned products.
-- `product_fit` describes relevance, not a readiness verdict. Product fit never
-  replaces `need_ids`, evidence, audience, or opportunity scoring.
-- Do not turn a product promise into a Need quote. Do not turn a Need quote
-  into a product capability.
-- Keep exact pricing and entitlements in the dated product entry. Do not copy
-  stale prices into other files.
-- Do not invent proof, customers, outcomes, metrics, or testimonials.
-
-### Product fit
-
-Use one of:
-
-| fit | Meaning | Typical CTA |
-|-----|---------|-------------|
-| `direct` | The Topic directly addresses a problem the product solves | `educate` / `invite` / `sell` |
-| `adjacent` | The Topic builds context but is not a product pitch | `educate` / `mention` |
-| `none` | No honest product connection is needed | `none` |
-
-Product fit is written on the Topic human card and inherited by the Brief.
-For third-party reviews, use `recommendation_ids`, not `product_ids`. Load the
-recommendation entry and disclose affiliate, sponsorship, or other commercial
-relationships before drafting.
-
-### Topic generation modes
-
-| mode | Required inputs |
-|------|-----------------|
-| `demand` | Needs + Profile |
-| `demand_to_offer` | Needs + Product + Profile |
-| `demand_to_review` | Needs + Recommendation + Profile |
-| `review` | Recommendation + Profile + evidence |
-| `offer_education` | Product + Profile + mechanism |
-| `profile_thesis` | Profile + Pillar |
-
-**Mode pick (hard):** Before any `produce` Topic write (扫需求选题 / 做成选题 / 生成 topic / 把 N-|C- 做成选题), **ask the user to choose `generation_mode`**. Recommend `demand` when the input is Need-led; do **not** silently assume. Enforce that mode’s required inputs. Product or Recommendation presence never proves audience demand and never raises a Topic to `ready` by itself.
-
----
-
-## § Topics (Opportunity evaluation)
-
-**Role:** Evaluate **Content Opportunities** — whether a topic deserves to exist as content. Does not replace capture or runs.  
-**Principle:** Topics are not titles. They are a judgment card the content owner can read in one minute. Five-dim scores are a gate in the appendix, not the reading surface.
-
-### Layout
-
-| Path | Role |
-|------|------|
-| `topics/scoring.md` | Scout → opportunity **15/20/25/25/15** (context/audience/conflict/insight/evidence); ready = opp≥**7.5** **and** aud≥**6** |
-| `topics/_index.md` | Queue: one row per topic, **`item`** link, **`src_*` engagement** |
-| `topics/items/T-YYYYMMDD-XX.md` | Decision card (summary + judgment) + Score/Trace/Constraints (appendix) |
-| `topics/_template-item.md` | Template |
-
-### Defaults
-
-- Scout first: `purpose` = `discard` | `library` | `produce`; `scout_label` = `demand` | `ally` | `foil` | `craft_only`. Needs default `scout_label=demand`.
-- Opportunity dims (produce only): `context` `audience` `conflict` `insight` `evidence` (0–10) → `opp_score`. Reasons live in the readable judgment layer, not in a score-table note column.
-- `ready` requires **opp_score ≥ 7.5 and audience ≥ 6**
-- `src_likes` / `src_replies` / `src_bookmarks` (+ views if known) + `engagement_at` on **index** when source is external
-- Do **not** fold `src_*` into opp_score; among **ready** only, prefer replies/bookmarks over views
-- **Language (hard):** Readable judgment + `one_liner` + Decision Card body = profile public language from `voice.md` **Language** (else pillars “Primary public language”, else `engine.json` locale). Verbatim Needs/Capture quotes stay untranslated. Operator chat may add a short gloss in another language; the Topic file and `ready` card body do not. Wrong-language prose → fix before `ready`.
-
-### Signal paths
-
-| User intent | inbox/raw | topics |
-|-------------|-----------|--------|
-| `记一下` + URL | **Yes** (default) | Only if also asked to score as topic |
-| `扫需求选题` | **Default NO** | Scout + score ≤5 produce from `needs/` `status=captured`; `source_type=need`; set `need_ids`; **never** put `N-` in `capture_id`. After a `produce` T- is written, set those `N-` to `promoted` and append the T- on `topic_ids` (index + entry). Discard / library: leave `captured` or `retired` |
-| `扫 inbox 选题` | Existing only | Scout + score ≤5 produce; set `capture_id`; skip captures already linked to a T- |
-| `把 C-xxx 做成选题` | Keep | Create T- with `capture_id` (still Scout + opportunity score) |
-| `链接直接选题` | Default NO unless user agrees capture | Create T- + src engagement |
-| `把 T-xx 收进 inbox` | Yes for that source | Backfill `capture_id` on T- |
-
-### Create topic steps
-
-**Input gate:** A Topic must have an explicit input. Accept: user idea/brief, Capture `C-…`, Need `N-…`, Product `P-…`, Recommendation `R-…`, or a profile thesis. If the user only says “generate a topic” without naming a source:
-
-1. If needs exist with `status=captured`, offer **扫需求选题**.
-2. Otherwise ask which explicit input to use.
-
-0. Resolve profile **Language** (`voice.md` → pillars → `engine.json` locale).
-1. Hard filters (`scoring.md`); else skip.
-2. **Scout gate** — set `purpose` + `scout_label`; if `discard`, do not create T-. Needs-sourced default `scout_label=demand` (ally/foil/craft_only still allowed when honest).
-3. **Mode pick** — if proceeding toward `produce`, **ask** `generation_mode` (six modes). Recommend `demand` for Need-led input; wait for user choice. Do not write a `produce` T- before the choice. Validate required ids for that mode.
-4. If `produce` (or user forces score):
-   a. Write the **decision card** first (`topics/_template-item.md` upper half), in profile Language: Verdict, Why this status, Who it’s for, Core judgment, Next step, then What problem / What they believe / What cut / Why now / What should change / Based on / Gaps / Suggested form. Keep the summary short; no enum lists, heat, or weighted arithmetic in the readable layer. Set frontmatter `generation_mode` to the chosen mode.
-   b. Score all 5 opportunity dims (0–10) → `opp_score`. Put **numbers only** in `## Score`. Do not repeat the card as table notes.
-   c. Fill `## Trace` (ids, src heat) and `## Constraints` (writer must-nots). Write `topics/items/T-….md`.
-5. If `library`: optional light item with `status: library`, or capture offer only. Source needs stay `captured` (or `retired` if the user discards).
-6. Add row on `topics/_index.md` (`item` + `src_*` when external). `one_liner` = the card’s core-judgment sentence, same language.
-7. If this T- was `produce` from `N-` ids: update each `needs/_index.md` row + entry to `status=promoted` and `topic_ids` containing this T-. Skip if no T- was written (`discard`).
-8. Output the **Decision Card** in chat — same shape and language as the file’s readable layer, not a score walkthrough. Optional operator gloss in another language is chat-only.
-
-**Important:** “What do they believe now?” must name the false belief. If you cannot, conflict cannot score above 5. Brief inherits the readable judgment layer, not the Score table.
-
-### Decision Card
-
-Paste the decision summary and readable judgment sections. Do not walk five scores in the main body.
-
-```text
-📋 Decision card T-YYYYMMDD-XX
-【Verdict】 ready | backlog | parked
-【Why this status】 …
-【Who it’s for】 …
-【Core judgment】 …
-【Next step】 open a run / fill needs / fill evidence / re-angle / kill
-【What problem】 …
-【What they believe】 believe X → actually Y
-【What cut】 …
-【Why now】 …
-【After reading】 before … → after …
-【Based on】 …
-【Gaps】 …
-【Suggested form】 …
-```
-
-### Topic → run
-
-On `用 T-xxx 开一单`: **ask platform(s)** if unset (multi-select OK). Create **one run per platform**, same `topic_id`, prefill selection per platform, mark topic `promoted` + list run ids on `topics/_index` `run` column.
-
-### Do not (Topics)
-
-- Create a Topic from a Hit, a scan, or from heat alone
-- Silently pick `generation_mode` (always ask before `produce`)
-- Skip Scout gate or opportunity scoring on `produce` topics
-- Skip the decision card readable layer on `produce`
-- Put `N-` in `capture_id` (needs go in `need_ids` only)
-
----
-
-## § Swipe / atom library
-
-- **Catalog:** `swipe/_index.md` / `wiki/atoms/_index.md` are the selection entrypoints (Approach B).  
-- **Hit-status:** `trial` | `working` | `dead` (not claim epistemology).  
-- **Hits cache:** index columns `n` `win` `loss` from **our** `published.result`.  
-- **Use history:** `published/_index.md` (columns `swipe` / `atoms`). **Do not** put an Evidence table in swipe/atom files.  
-- **New item:** template + index row; default `trial`, n=win=loss=0.  
-- **From capture / run:** only via **§ Library ingest**. Never silent seed from runs.
-
----
-
-## § Interview
-
-Triggers: 采访我 / 完善人设 / interview me / who am I as a creator / 建立 profile.
-
-Follow `easysociable-content/workflows/INTERVIEW.md` for questions. Persist after each dimension; do not wait until the end.
-
-- Append `## Experience` / `## Values` to existing `voice.md` if those headings are missing. Do not rewrite filled Identity rows.
-- Missing `boundaries.md` → write empty starter, then interview. **Never invent boundaries.**
-- Do not scan the watchlist. Do not create a slideshow. Do not ingest needs in this section.
-
-### Hard gate (refuse 出图 / PagePack / `create_run`)
-
-Identity **only** Role / What I build / Point of view — no `(fill)`, non-empty. Public handle MAY stay `(fill)`. `boundaries.md` exists with Not covering **and** Never claim non-empty. If `needs/` exists, also `needs/_index.md` ≥ 3 `N-` rows; if `needs/` is absent, skip that count.
-
-Fail → list missing files; stay here. User must say **强制继续** / **force continue**. Optional check:
-
-```bash
-easysociable engine validate --root . --require-knowledge
-```
-
-### Completion
-
-Hard gate vs first-session (`profile: partial`) vs complete seven dimensions (`profile: complete`) — same table as INTERVIEW.md. Experience / Values / expression are not the hard gate. Needs hard gate is ≥ **3** original quotes; first-session **target** is **5**.
-
----
-
-## § Keywords & handles
-
-`keywords.md` is **three-track**: `listen` (social → Hits), `search` (SEO), `ask` (GEO questions).  
-`handles.md` ≤8 ally/foil/competitor.
-
-**Generate:** from Needs/audience/pillars → chat candidates (max 10) → user confirm → append row with `track` + `seed_from`.  
-**Scan「扫关键词」:** `track=listen` + `active` only (unless user names other ids). Stage **§ Hits**; never Topic; never invent heat/quotes.  
-**search/ask:** not default social scan. Volume/kd optional on search only.
-
-## § Hits (research triage)
-
-**Role:** Sort research hits → **need** | **library** (swipe/atom via Capture+ingest) | **discard**. Not a Topic queue.
-
-| Path | Role |
-|------|------|
-| `hits/_index.md` | Registry |
-| `hits/_template.md` | Entry shape |
-| `hits/entries/H-*.md` | Per-hit files |
-
-Id `H-YYYYMMDD-XX`. Status: `triage` → `routed_need` | `routed_library` | `discarded`.
-
-```text
-🔎 Hit H-…
-【Keyword】 K-xx
-【Source】 @handle · url
-【Kind】 author_post | comment | thread
-【Signal】 demand_language | structure | foil | mixed | noise
-【Heat】 likes … · replies … · bookmarks … · views … @ date
-【Excerpt】 verbatim
-【Suggest】 need | library | discard
-```
-
-On scan: `keyword_id` + `src_*` when known. Route need → **§ Needs**. Route library → **§ Capture** + library offer. Never Hit → Topic.
-
----
-
-## § Needs (audience quotes)
-
-Triggers: 记需求 / 用户原话 / 评论导出 / capture audience quotes / user needs.  
-Hits → Needs **only after route confirm** (**§ Hits**).  
-扫需求选题 → **§ Topics** (this section only writes `N-` records).
-
-Follow `easysociable-content/workflows/NEEDS.md` when this engine's copy of the workflow is the newer one; this section wins if both exist.
-
-### Layout
-
-| Path | Role |
-|------|------|
-| `needs/_index.md` | Quote registry |
-| `needs/entries/N-YYYYMMDD-XX.md` | Verbatim quote + sightings + optional gloss |
-| `needs/_template.md` | Entry shape |
-
-### Status machine (only these words)
-
-```text
-captured  → just written, or still captured after a synonym merge (frequency ≥ 1)
-promoted  → and only when a T- exists whose need_ids contain this N-; topic_ids lists that T-
-retired   → user discarded
-```
-
-- Never `clustered`. Merge is not a status.
-- Never mark `promoted` without that `T-`. `topic_ids` non-empty ≡ `promoted`.
-- After § Topics writes a `produce` T- from this `N-`, flip this row + entry to `promoted` and append the T- on `topic_ids`. Discard / library stay `captured` or `retired`.
-- Scan input: `status=captured` (`topic_ids` empty). Max 5 `produce`.
-- Quote cell = verbatim. Gloss is optional and separate.
-- Do not invent quotes. Do not treat watchlist hooks as needs.
-- `N-` never goes in topic `capture_id`.
-
-### Ingest
-
-1. New entry from `_template.md` + row at top of `_index.md`.
-2. Chat card:
-
-```text
-✅ Need N-YYYYMMDD-XX
-【Quote】 …
-【Source】 comment | dm | consult | faq | student | other
-【Frequency】 1
-【Stored】 needs/_index + needs/entries/N-….md
-【Merge】 new | merged into N-… (frequency now k, status still captured)
-```
-
-3. Synonym merge: propose → user yes → same `N-` stays `captured`, `frequency += 1`, append sighting. Retire a mistaken duplicate.
-
-Hard gate: `needs/_index.md` ≥ **3** `N-` rows once `needs/` exists (empty index fails). Target **5**. Prefer scanning needs before inbox.
-
----
-
-## § Open a run
-
-### When
-
-User wants a tracked job: "开一单", "open a run", "按完整流程", or a multi-step post with gates.
-
-**Hard gate first** (same as § Interview). Fail → list missing files; do not create `runs/<slug>/` unless **强制继续** / **force continue**.
-
-### Model (one platform = one Run)
-
-- **One folder = one platform draft** (native rewrite).  
-- **Same Topic, different platforms** → **N runs**, shared `topic_id` — **not** one folder with many platform sections.  
-- `idea.md`: `primary_platform` = this run’s platform; `platforms` = **same single value** (column kept for index compat).  
-- `draft.md`: copy for **this platform only**.  
-- Image deck: **`pack.md`** (or `pack-<platform>.md`); pixels under **`exports/<slug>/`**.  
-- Ship → one `published/_index` row for this `run_id`.  
-- Cross-platform progress lives on the **Topic** (`run` column lists all RUN ids), not inside one run.
-
-### Platform gate (required before folder create)
-
-1. User already named platform(s) → use them.  
-2. Else **stop and ask a required multi-select platform question**. Offer the account’s `platforms` from `accounts/_index` plus any supported platforms the user may name: `x`, `linkedin`, `instagram`, `tiktok`, `threads`, `xiaohongshu`. The user must be able to select one or more.
-3. Do not create a folder, row, brief, or draft before the platform selection is answered.
-4. **Do not** silently default to `x` when opening a run (default `x` still applies to other flows, e.g. watchlist filter).  
-5. **N platforms selected → N runs** (same topic/goal/pillar; independent platform craft, brief, draft, editor, rubric, and ship state).  
-6. Slug: include the platform when N>1 or always OK — e.g. `2026-08-26-angle-x`, `…-linkedin`.
-
-### Load platform craft (required — every run, every draft stage)
-
-For **each** run, before brief / packet / draft / editor / rubric / pack:
-
-1. Resolve craft file via `platforms/README.md` (`x` → `x.md`, `instagram`/`ig` → `ig.md`, …).  
-2. **Read** that file end-to-end. Apply length, hooks, anti-slop, quality checklist.  
-3. Also load profile `audience.md` **core + this platform’s section**.  
-4. Cite the craft path in `brief.md` / `packet.md` (not hardcoded `platforms/x`).  
-5. **Do not** paste another platform’s copy unchanged. Sibling runs each re-read their own craft file.  
-6. If the craft file is missing → ask user / create stub; **do not draft** from memory.
-
-### State machine (do not skip)
-
-```text
-RUN_OPEN → BRIEF_WAIT → DRAFTED → EDIT_WAIT
-  → SCORED → SHIP|HOLD → READY_TO_POST
-  → PUBLISHED feedback
-```
-
-Also mirror high-level status on **`runs/_index.md`**:  
-`open` → `brief` → `draft` → `scored` → `ready` → `published` | `hold` | `killed`.  
-(`partial` = **legacy** only; do not use for new multi-platform planning.)
-
-Human gates (chat only):
-
-1. **简报 OK** / Brief OK  
-2. **编辑 OK** / Editor OK  
-3. After SHIP: user posts **this** platform, then URL + result → feedback + published row  
-4. If images: **图 OK** after render (when skill available)
-
-### Folder setup
-
-1. Resolve platform list (gate above). For **each** platform:  
-2. **Load** `platforms/<platform>.md` (required; see above).  
-3. Create `runs/YYYY-MM-DD-short-slug[-platform]/` from `runs/_template/`.  
-4. Fill `idea.md` (pillar, **platform**, **selection ids**, optional `topic_id`).  
-5. **Append row** on `runs/_index.md` (`status=open`).  
-6. If from **T-…**: prefill from topic; mark topic `promoted`; append all new run ids to `topics/_index` `run`.  
-7. Never leave production only in chat when in run mode.
-
-### Selection inside a run (required step; **none allowed**)
-
-Before brief body, evaluate catalogs (Approach B + **profile + this run’s platform**).
-
-Filter every catalog row:
-
-- `platforms` includes this run’s platform (or `*`)  
-- `profiles` is `*` **or** includes current `profile_id`  
-- Swipe/atom: drop `dead`. Claim: drop `retired`.  
-
-Rank:
-
-- Swipe/atom: `working` before `trial`; then higher `win/n` (if n=0, treat as unranked trial).  
-- Claim: **do not** rank by `hypothesis` vs `supported`. Rank remaining by `win/n`. `hypothesis` is valid to pick. `weakened` only if the user wants that tension.
-
-Then: swipe 0–1 → atoms by role → claim 0–1. **`none` always allowed** with why.  
-Sibling runs for the same topic may pick different swipe/atoms (native rewrite).
-
-**Write the pick into `idea.md` (this is the citation).** Chat card alone does not count.
-
-| Field | Required form |
-|-------|----------------|
-| `swipe_id` | `none` or `S1` (one id) |
-| `atoms` | `none` or `hook:A-… cta:A-…` (real `A-` ids) |
-| `claim_id` | `none` or catalog `C-…` — **not** the `hypothesis:` one-off line |
-
-If the user later changes the pick, **overwrite** those three fields. Do **not** bump catalog `n`/`win`/`loss` here — that is ship-only (§ Catalog evidence).
-
-**引用** = how many **runs** currently name this id in `idea.md` (killed runs still count unless selection was cleared). The workbench derives it by scanning runs; it is not a column the agent increments. Observatory numbers update on `workbench/build.py`, not at the moment of chat.
-
-### Brief Engine (decision tree → quality gate → stop rule)
-
-The brief inherits reasoning from the Topic Opportunity and focuses on **what to say and how**.
-
-**Step 1 — Inherit from Topic Opportunity**
-
-When the run is from T-xxx, copy the **readable judgment layer** into `brief.md` `## Inherited`:
-- `context` ← 「为什么现在值得讲」
-- `audience` ← 「适合谁」
-- `conflict` ← 「他们现在相信什么」
-- `insight` ← 「我们准备讲哪一刀」
-- `outcome` ← 「看完后的改变」
-- `evidence` ← 「凭什么这么判断」 + 「当前缺口」
-
-Sharpen these for the specific platform/angle — do not re-derive from scratch. If T-xxx is a legacy five-section score sheet, derive the readable judgment layer before proceeding. Do not inherit the Score table as prose.
-
-**Step 2 — Walk the decision tree**
-
-```text
-1. Read inherited context/conflict/insight
-2. Can you state a thesis in one sentence?
-   → No → reframe or stop
-3. Can you name specific evidence?
-   → No evidence at all → stop, suggest research
-   → Weak but mechanism is provable → continue
-4. What's the angle? (How does this conflict become content?)
-5. What changes for the reader after consuming this?
-   → "They'll learn something" is not specific enough
-   → Name the belief before and after
-6. → Fill brief → Self-score → Gate
-```
-
-**Step 3 — Fill brief.md**
-
-The brief has two layers. The **reading surface** (top) is what a human reviews in 30 seconds. The **appendix** (bottom) is machine handoff data.
-
-Reading surface — fill in this order:
-
-1. **Thesis** — one sentence a stranger can understand.
-2. **For whom → what changes** — reader, their false belief, the after state.
-3. **Evidence** — named sources, type, quality, and what we cannot prove.
-4. **Platform and expression** — platform, form, hook, structure outline.
-5. **Boundaries** — must include / must avoid.
-
-Appendix — fill after the reading surface:
-
-6. **Inherited from Topic** — context, audience, conflict, insight, outcome, evidence (sharpened, not copied).
-7. **Selection** — swipe, atoms, claim ids.
-8. **Quality gate** — 6 × 1–5 scores.
-
-**Step 4 — Brief quality gate (self-scoring)**
-
-Score `brief.md` quality gate (6 × 1–5, max 30):
-
-| # | Gate | Question |
-|---|------|----------|
-| 1 | Thesis clarity | Can the thesis be one sentence a stranger understands? |
-| 2 | Evidence strength | Is there at least one specific, named source or mechanism? |
-| 3 | Angle originality | Is this angle different from what's already in the conversation? |
-| 4 | Structure coherence | Does the narrative flow from hook → conflict → thesis → proof → outcome? |
-| 5 | Outcome specificity | Can you name the belief change in before/after form? |
-| 6 | Tension preservation | Does the brief still carry the topic's conflict, or did it dilute? |
-
-| Result | Action |
-|--------|--------|
-| **≥ 24** (PASS) | Present Brief Card to human for OK |
-| **20–23** (REVISE) | Flag weak dimensions, self-revise once, re-score |
-| **< 20** (STOP) | Trigger stop rule |
-
-Update `brief.md` frontmatter: `gate_score`, `gate_pass`.
-
-**Step 5 — Stop rule**
-
-When the brief cannot pass the quality gate:
-
-```text
-⚠️ Weak Brief — RUN-…
-【Failing gates】 gate … score … · gate … score …
-【Reason】 one line why
-【Suggestion】 reframe angle / research evidence / kill topic
-Reply: 用 reframe / 研究 / 换 topic / 强制继续
-```
-
-The agent may reject a brief. "强制继续" overrides the gate (user decision). The stop rule is not a hard block — it is structured feedback that makes weakness visible.
-
-### Packet
-
-After brief OK: short `packet.md` — pillar slice + voice + audience (**core + this platform**) + **craft notes distilled from the loaded `platforms/<platform>.md`** + **thesis sentence from brief**.
-
-### Writer → Editor
-
-- Writer **re-reads** `platforms/<platform>.md`, then fills `draft.md` (**this platform only**) + full text in chat.  
-- Editor → `editor.md`; check against that craft file’s quality bar; REVISE until APPROVE or user override.
-
-### Rubric gate
-
-Score `rubric.md` (6 × 0–2, max 12).  
-
-- **≥ 8 → SHIP** — 发布包 in chat; `runs/_index` → `ready`.  
-- **< 8 → HOLD**.  
-- Carousel/slideshow platforms: after SHIP → **§ PagePack + render**.  
-- Pure X text: no multi-page pack.
-
-### Required chat cards (run mode)
-
-**Open card** (one card per run; if N>1, list all opened)
-
-```text
-✅ Run opened  RUN-YYYYMMDD-slug
-【Account】 (from accounts/_index)
-【Profile】 (resolved profile_id)
-【Voice】 profile | path
-【Platform】 x | linkedin | …  → craft `platforms/<file>.md` loaded
-【Topic】 none | T-…  (sibling runs: RUN-…, RUN-…)
-【Goal】 …
-【Pillar】 P?
-【Selection】 (profile + platform; drop dead/retired; prefer working; none OK)
-  swipe: S? | none — why
-  atoms: …
-  claim: C? | none — why
-【Index】 runs/_index updated
-【State】 open → next: brief
-```
-
-**Brief card** (after quality gate passes — mirrors brief.md reading surface)
-
-```text
-📋 Brief  RUN-…
-【Thesis】 one sentence
-【For whom】 reader · false belief → after state
-【Evidence】 sources · type · quality · what we cannot prove
-【Platform】 platform · form · hook · structure outline
-【Boundaries】 must include · must avoid
-【Gate】 Thesis … · Evi … · Angle … · Struct … · Outcome … · Tension … = …/30 ✓
-Reply 简报 OK / 改 …
-```
-
-**Draft / Editor / Rubric** — as before; draft is single-platform and must follow the loaded craft file.
-
-**Publish pack** (SHIP)
-
-```text
-📦 READY_TO_POST  RUN-…
-【Platform】 x | …
-【Copy】 …
-【Pack】 n/a | ready | blocked
-【After post】 URL + win/flat/loss → feedback + published row
-```
-
-### Feedback
-
-`feedback.md` + **§ Published index** on ship.  
-Update `runs/_index` → `published`.  
-Then **§ Catalog evidence**. Never auto-`supported` / auto-`working`; never silent catalog ingest.
-
-### § Published index (Step 6)
-
-On confirmed ship for this run:
-
-1. Row on **`published/_index.md`** (date, **platform**, url, run, pillar, swipe, atoms, claim, result, notes).  
-2. Optional stub.  
-3. `runs/_index` status → `published`.  
-4. **§ Catalog evidence** for every cited swipe / atom / claim id (skip `none`).  
-5. Weekly 复盘: start from **published** `_index`. Same topic’s other platforms = other runs / other published rows.
-
-### § Catalog evidence (required on ship)
-
-Hits cache lives on the **index row** (`n` `win` `loss`). This is **not** 引用. 引用 is run selection in `idea.md` (already true once the run exists). `n`/`win`/`loss` only move **on ship**.
-
-**Swipe / atom — no Evidence file.**  
-The ledger is the **published row** (`swipe` / `atoms` / `result` / `notes`). Why a post flopped stays in `feedback.md` or that published `notes` cell. Do **not** add `## Evidence` to swipe/atom bodies.
-
-**Claim — Evidence table in the claim file.**  
-That table answers “does the sentence still hold?” (`vs claim`), which is not `result`.
-
-For each cited id:
-
-1. `n += 1`  
-2. `win += 1` if `result=win`; `loss += 1` if `result=loss`; `flat`/`unknown` bump `n` only.  
-3. Sync file frontmatter `status` if you change index status (only after user 批准).  
-4. **Claim only:** append the run to the claim file **Evidence** table (`result` + **vs claim**).  
-
-Propose status change (user must 批准; never auto):
-
-| Object | Propose | Gate |
-|--------|---------|------|
-| swipe / atom | `trial` → `working` | `n≥2` AND `win≥2` AND `win>loss` |
-| swipe / atom | → `dead` | user says stop, **or** `n≥3` AND `win=0` AND `loss≥2` |
-| claim | → `supported` | ≥3 runs **support the proposition** (not merely `result=win`) AND user 批准 |
-| claim | → `weakened` / `retired` | user 批准; hits do not flip epistemology |
-
-Single win never promotes. n=1 is recorded, not celebrated.
----
-
-## § PagePack + render (Step 7)
-
-Bridge to **easysociable-slideshow** (when skill exists). Vault owns pack slots; product owns templates + pixels.
-
-### When
-
-Planned platform needs multi-page images (linkedin / ig / tiktok), not pure X text.
-
-### Preconditions
-
-0. **Hard gate** (same as § Interview). Fail → list missing files; no slideshow/template handoff unless **强制继续** / **force continue**.  
-1. Run exists; brief OK; editor OK; **rubric SHIP**.  
-2. Per-platform `page_count` / `deck_hint` follows the **arc**, not a five-family catalog.  
-3. Prefer a swipe whose beats are an argument; else name a one-line arc from the brief.  
-4. **Forbidden default:** Cover → Point → Steps → Recap → CTA (those are P0 family names, not a pack recipe).
-
-### Agent steps
-
-1. Write **`runs/<slug>/pack.md`** (from `_template/pack.md`; `pack-<platform>.md` OK):  
-   - this run is already one platform  
-   - **Arc** one-liner + recipe **before** Pages  
-   - slots, language, pillar, swipe_id, claim_id  
-   - page 1 `role` is not automatically `cover`  
-2. **Pack Card** in chat — user may `改 pack：…`.  
-3. Hand off to **easysociable-slideshow**: it picks templates by platform + canvas + slot shape, then opens Studio. Topology lock means a wrong arc needs a **new Run**, not slot rewrite.  
-4. **Render / Download:** pixels leave the product from Studio Download (not CLI render). After the user downloads, copy files under **`exports/<slug>/`** if needed; set pack `job_id`, `export_path`, status `done`; set `runs/_index` `pack`/`export` to yes. If slideshow skill is **not** available: status `not_requested` / blocked — **do not fake images**.  
-5. User **图 OK** or revise slots (same topology) / new Run (different arc).
-
-### Pack Card
-
-```text
-📑 Pack  RUN-…  platform=linkedin|tiktok|ig
-【Arc】 one-liner / recipe
-【Pages】 N
-【deck_hint】 …
-【Outline】 P1… Pn
-【export_path】 exports/<slug>/
-【Render】 done | blocked (no skill) | failed
-```
-
-### Do not
-
-- Cram multiple platforms into one run / one pack  
-- Pack for pure X text  
-- Private template AST — public slots only  
-- Invent metrics in slots  
-- Skip SHIP before quality pack  
-- Store binaries inside `runs/` (use **exports/**)  
-- Default Cover → Point → Steps → Recap → CTA  
-- Skip the Arc line and fill P0 family names as a tutorial  
-
----
-
-## § Library ingest — swipe / atoms / claims (strict)
-
-**Runs never auto-enter catalogs.** Ship → feedback + published + **§ Catalog evidence** if ids already used. New catalog rows still need user approval.
-
-### Two ingest paths
-
-| Path | When | Your-post metrics? |
-|------|------|--------------------|
-| **A. Capture offer** | structure/link capture → user picks 入库 options | **No** (source heat is prior, not our n/win) |
-| **B. Post-publish** | In `published/_index` + gates → Agent proposes → user approves | **Yes** |
-
-### Hard bans
-
-| Forbidden |
-|-----------|
-| Unpublished run → new swipe/atom/claim |
-| Silent seed while scaffolding |
-| Catalog write without user approval |
-| Single win → claim `supported` or swipe/atom `working` |
-| Mixing claim epistemology with swipe/atom hit-status |
-| `## Evidence` table in a swipe or atom file |
-| Auto-changing `status` when bumping `n`/`win`/`loss` |
-
-### Engagement gates (path B only)
-
-At least one of: **replies ≥ 2** | bookmarks ≥ 5 | likes ≥ 10 | user override “这条可收”.  
-These gates are for **new ingest**, not for promoting `working` (that uses our `win`/`loss` counts).
-
-### After user approval
-
-| Ask | Create |
-|-----|--------|
-| 收成 swipe | `swipe/` + index, `trial`, n=win=loss=0 |
-| 收成 atom | `wiki/atoms/`, set `type`, `trial`, n=win=loss=0 |
-| 收成 claim | `wiki/claims/`, `hypothesis`, n=win=loss=0 |
-
-### § Claims promote
-
-- Epistemic Evidence on existing claims when used (separate from hits).  
-- `supported` only if ≥3 runs support the **proposition** **and** user 批准.  
-- A `win` that does not actually support the sentence does **not** count toward `supported`.
-
-### 复盘 card
-
-```text
-📈 Review
-【Run】 …
-【Result】 …
-【Hits updated】 swipe S? n/win/loss … · atoms … · claim …
-【Hit-status propose】 none | S?/A-… trial→working | →dead
-【Path B ingest】 not eligible | eligible → propose swipe/atom/claim
-【Claim promote】 none | suggest C-… hypothesis→supported|weakened
-Reply 批准 … / 否决 / 再观察
-```
-
-### Self-check (Capture)
-
-- [ ] `inbox/_index.md` row + `inbox/entries/C-….md`  
-- [ ] `raw/` file + `raw/_index.md` only if criteria matched  
-- [ ] Capture Card + library offer when structure/link
-
-### Self-check (Topic / Opportunity)
-
-- [ ] Decision card filled (结论 / 状态原因 / 适合谁 / 核心判断 / 下一步)
-- [ ] Readable judgment filled (问题 / 当前相信什么 / 我们的角度 / 为什么现在 / 看完后的改变 / 依据 / 缺口 / 形式)
-- [ ] False belief named in 「他们现在相信什么」 (not just a conflict score)
-- [ ] Readable sections are concise; heat and arithmetic only under `## 评分与追踪附录`
-- [ ] Score table is numbers only; opp + aud + verdict line present
-- [ ] Opportunity Card in chat matches the human half (not a five-score walkthrough)
-
-### Self-check (Run)
-
-- [ ] Folder exists under `runs/` with template files filled  
-- [ ] `runs/_index.md` row created/updated  
-- [ ] Platform gate done; one run per selected platform; `primary_platform` = that platform  
-- [ ] `platforms/<platform>.md` **read** before brief/draft/editor; cited in brief + packet
-- [ ] Brief `## Inherited` populated from Topic Opportunity (not re-derived)
-- [ ] Brief decision tree walked; quality gate scored (6 × 1–5)
-- [ ] Brief gate ≥ 24 before presenting to human (or stop rule triggered)
-- [ ] No skip of brief / editor / rubric when user asked for full run  
-- [ ] HOLD never presented as final publish pack  
-- [ ] Image pack only for this run’s platform; exports under `exports/<slug>/`  
-- [ ] Each ship → published row; this run → `published`  
-- [ ] `idea.md` has `swipe_id` / `atoms` / `claim_id` matching the Open card (not only chat)  
-- [ ] Cited swipe/atom/claim index `n`/`win`/`loss` bumped **on ship only**; status unchanged unless user 批准  
-- [ ] Swipe/atom: no Evidence section; claim Evidence table updated if claim used  
-- [ ] Selection dropped `dead`/`retired`; did not treat claim `supported` as a hit-rank  
-- [ ] Same topic multi-platform = sibling runs sharing `topic_id`, not one multi-section folder
-- [ ] Packet includes thesis sentence from brief
-- [ ] No invented metrics  
-- [ ] No cross-platform paste without re-applying that platform’s craft file