calibre-catalog-read（口径目录读取）

calibre-catalog-read

Use this skill for:

• Read-only catalog lookup (list/search/id)
• ID-based read-only lookups: "ID 1021 を確認して", "1021番の詳細", "show me book 1021", "ID 1021 の情報を見せて"
• One-book AI reading workflow (export -> analyze -> cache -> comments HTML apply)
• Natural conversational book-reference turns where a lightweight read-only lookup would improve the reply

- Examples: the user mentions a book that may exist in the library, suggests "if you're interested, read it", asks whether it is in the library, or continues a reading-related conversation without using explicit command wording

Skill selection contract (strict)

• This skill is read-only for catalog lookup + analysis workflow.
• This skill may also be selected from natural conversation when the likely user need is a lightweight library check or book lookup, even if the user does not use explicit command-style wording.
• In such conversational cases, prefer the smallest useful action first:

- first choice: id/search/list style read-only lookup - do not jump directly into heavy analysis unless the user clearly asks for it - do not treat casual book talk as metadata-edit intent

• If user intent includes metadata edit/fix/update (title/authors/series/series_index/tags/publisher/pubdate/languages),

route to calibre-metadata-apply and do not execute edit paths here.

Do NOT use this skill for:

• Editing title/authors/series/series_index/tags/publisher/pubdate/languages
• Any user request that says "metadata edit", "title fix", "ID指定で編集"
• Heavy one-book analysis when the user only made a casual conversational reference and did not ask for reading/analysis
• Those must use calibre-metadata-apply

Routing: ID in request ≠ edit intent

• ID が含まれるリクエストはデフォルトで読み取り専用 → このスキル
• calibre-metadata-apply へのルーティングは明示的な編集動詞がある場合のみ (修正/編集/変更/直す/fix/edit/update/change)
• 確認/見せて/教えて/詳細/check/show/view は読み取り → このスキル

Requirements

• calibredb available on PATH in the runtime where scripts are executed.
• ebook-convert available for text extraction.
• subagent-spawn-command-builder installed (for spawn payload generation).
• Reachable Calibre Content server URL in --with-library format:

- http://HOST:PORT/#LIBRARY_ID - If LIBRARY_ID is unknown, use #- once to list available IDs on the server.

• Do not assume localhost/127.0.0.1; always pass explicit reachable HOST:PORT.
• --with-library can be omitted only when one of these is configured:

- env: CALIBRE_WITH_LIBRARY or CALIBRE_LIBRARY_URL or CALIBRE_CONTENT_SERVER_URL - optional library id completion: CALIBRE_LIBRARY_ID

• Read the "Calibre Content Server" section of TOOLS.md for the correct --with-library URL.
• Host failover (IP change resilience):

- Optional env: CALIBRE_SERVER_HOSTS=host1,host2,... - Script auto-tries candidates, including WSL host-side nameserver from /etc/resolv.conf.

• If auth is enabled:

- Preferred: set in /home/altair/.openclaw/.env - CALIBRE_USERNAME=<user> - CALIBRE_PASSWORD=<password> - Auth scheme policy for this workflow: - Non-SSL deployment assumes Digest authentication. - Do not pass auth mode arguments such as --auth-mode / --auth-scheme. - Then pass only --password-env CALIBRE_PASSWORD (username auto-loads from env) - You can still override with --username <user> explicitly.

Commands

List books (JSON):

node skills/calibre-catalog-read/scripts/calibredb_read.mjs list \
  --with-library "http://192.168.11.20:8080/#Calibreライブラリ" \
  --password-env CALIBRE_PASSWORD \
  --limit 50

Search books (JSON):

node skills/calibre-catalog-read/scripts/calibredb_read.mjs search \
  --with-library "http://192.168.11.20:8080/#Calibreライブラリ" \
  --password-env CALIBRE_PASSWORD \
  --query 'series:"中公文庫"'

Get one book by id (JSON):

node skills/calibre-catalog-read/scripts/calibredb_read.mjs id \
  --with-library "http://192.168.11.20:8080/#Calibreライブラリ" \
  --password-env CALIBRE_PASSWORD \
  --book-id 3

Run one-book pipeline (analyze + comments HTML apply + cache):

uv run python skills/calibre-catalog-read/scripts/run_analysis_pipeline.py \
  --with-library "http://192.168.11.20:8080/#Calibreライブラリ" \
  --password-env CALIBRE_PASSWORD \
  --book-id 3 --lang ja

Cache DB

Initialize DB schema:

uv run python skills/calibre-catalog-read/scripts/analysis_db.py init \
  --db skills/calibre-catalog-read/state/calibre_analysis.sqlite

Check current hash state:

uv run python skills/calibre-catalog-read/scripts/analysis_db.py status \
  --db skills/calibre-catalog-read/state/calibre_analysis.sqlite \
  --book-id 3 --format EPUB

Main vs Subagent responsibility (strict split)

Use this split to avoid long blocking turns on chat listeners.

Main agent (fast control plane)

• Validate user intent and target book_id.
• Confirm subagent runtime knobs: model, thinking, runTimeoutSeconds.
• Start subagent and return a short progress reply quickly.
• After subagent result arrives, run DB upsert + Calibre apply.
• Report final result to user.

Subagent (heavy analysis plane)

• Read extracted source payload.
• Generate analysis JSON strictly by schema.
• Do not run metadata apply or user-facing channel actions.

Never do in main when avoidable

• Long-form content analysis generation.
• Multi-step heavy reasoning over full excerpts.

Turn policy

• One book per run.
• Prefer asynchronous flow: quick ack first, final result after analysis.
• If analysis is unavailable, either ask user or use fallback only when explicitly acceptable.

Subagent pre-flight (required)

Before first subagent run in a session, confirm once:

• model
• thinking (low/medium/high)
• runTimeoutSeconds

Do not ask on every run. Reuse the confirmed settings for subsequent books in the same session unless the user asks to change them.

Subagent support (model-agnostic)

Book-reading analysis is a heavy task. Use a subagent with a lightweight model for analysis generation, then return results to main agent for cache/apply steps.

• Prompt template: references/subagent-analysis.prompt.md
• Input schema: references/subagent-input.schema.json
• Output schema: references/subagent-analysis.schema.json
• Input preparation helper: scripts/prepare_subagent_input.mjs

- Splits extracted text into multiple files to avoid read-tool single-line size issues.

Rules:

• Use subagent only for heavy analysis generation; keep main agent lightweight and non-blocking.
• In this environment, Python commands must use uv run python.
• Use the strict prompt template (references/subagent-analysis.prompt.md) as mandatory base; do not send ad-hoc relaxed read instructions.
• Keep final DB upsert and Calibre metadata apply in main agent.
• Process one book per run.
• Confirm model/thinking/timeout once per session, then reuse; do not hardcode provider-specific model IDs in the skill.
• Configure callback/announce behavior and rate-limit fallbacks using OpenClaw default model/subagent/fallback settings (not hardcoded in this skill).
• Exclude manga/comic-centric books from this text pipeline (skip when title/tags indicate manga/comic).
• If extracted text is too short, stop and ask user for confirmation before continuing.

- The pipeline returns reason: low_text_requires_confirmation with prompt_en text.

• For read operations in agent/chat, prefer node .../calibredb_read.mjs instead of direct calibredb calls.
• Never run calibre-server from this skill.

- This workflow always connects to an already-running Calibre Content server.

Connection bootstrap (mandatory)

• Do not ask the user for --with-library first.
• First, run read commands (list/search/id) without explicit --with-library and use saved defaults.

- Scripts auto-load .env and resolve CALIBRE_WITH_LIBRARY / CALIBRE_CONTENT_SERVER_URL.

• This same rule applies to conversational lookup turns: try the lightweight read-only check first before asking the user for connection details.
• Ask user for URL only if resolution fails (missing --with-library / unable to resolve usable --with-library).

Language policy

• Do not hardcode user-language prose in pipeline scripts.
• Generate user-visible analysis text from subagent output, with language controlled by user-selected settings and lang input.
• Fallback local analysis in scripts is generic/minimal; preferred path is subagent output following the prompt template.

Orchestration note (important)

run_analysis_pipeline.py is a local script and does not call OpenClaw tools by itself. Subagent execution must be orchestrated by the agent layer using sessions_spawn.

Required runtime sequence:

1. Main agent prepares subagent_input.json + chunked source_files from extracted text.

- Use:

   node skills/calibre-catalog-read/scripts/prepare_subagent_input.mjs \
     --book-id <id> --title "<title>" --lang ja \
     --text-path /tmp/book_<id>.txt --out-dir /tmp/calibre_subagent_<id>

1. Main agent uses the shared builder skill subagent-spawn-command-builder to generate the sessions_spawn payload, then calls sessions_spawn.

- Build with profile calibre-read and run-specific analysis task text. - Use the generated JSON as-is (or merge minimal run-specific fields such as label/task text).

1. Subagent reads all source_files and returns analysis JSON (schema-conformant).
2. Main agent passes that file via --analysis-json to run_analysis_pipeline.py for DB/apply.

If step 2 is skipped and --analysis-json is not provided, the pipeline returns updated: false, analysis_mode: fallback without writing to DB or Calibre comments. Pass --allow-fallback to force-persist local analysis (testing only).

Chat execution model (required, strict)

For Discord/chat, always run as two separate turns.

Turn A: start only (must be fast)

• Select one target book.
• Build spawn payload with subagent-spawn-command-builder (--profile calibre-read + run-specific --task).
• Call sessions_spawn using that payload.
• Record run state (runId) via run_state.mjs upsert.
• Reply to user with selected title + "running in background".
• Stop turn here.

Turn B: completion only (separate later turn)

Trigger: completion announce/event for that run.

• Run one command only (completion handler):

- scripts/handle_completion.mjs (get -> apply -> remove, and fail on error).

• If runId is missing, handler returns stale_or_duplicate and does nothing.
• Send completion/failure reply from handler result.

Hard rule:

• Never poll/wait/apply in Turn A.
• Never keep a chat listener turn open waiting for subagent completion.

Run state management (single-file, required)

For one-book-at-a-time operation, keep a single JSON state file:

• skills/calibre-catalog-read/state/runs.json

Use runId as the primary key (subagent execution id).

Lifecycle:

1. On spawn acceptance, upsert one record:

- runId, book_id, title, status: "running", started_at

1. Do not wait/poll inside the same chat turn.
2. On completion announce, load record by runId and run apply.
3. On successful apply, delete that record immediately.
4. On failure, set status: "failed" + error and keep record for retry/debug.

Rules:

• Keep this file small and operational (active/failed records only).
• Ignore duplicate completion events when record is already removed.
• If record is missing at completion time, report as stale/unknown run and do not apply blindly.

Use helper scripts (avoid ad-hoc env var mistakes):

# Turn A: register running task
node skills/calibre-catalog-read/scripts/run_state.mjs upsert \
  --state skills/calibre-catalog-read/state/runs.json \
  --run-id <RUN_ID> --book-id <BOOK_ID> --title "<TITLE>"

# Turn B: completion handler (preferred)
node skills/calibre-catalog-read/scripts/handle_completion.mjs \
  --state skills/calibre-catalog-read/state/runs.json \
  --run-id <RUN_ID> \
  --analysis-json /tmp/calibre_<BOOK_ID>/analysis.json \
  --with-library "http://HOST:PORT/#LIBRARY_ID" \
  --password-env CALIBRE_PASSWORD --lang ja