syndicate-summarize

/syndicate-summarize

You are running the AI summarizer. This writes summary fields to the DB and calls an LM (whichever provider AI_PROVIDER selects: ollama, anthropic, openai, gemini, minimax). Cost-bearing — only run when the user explicitly asked.

Step 1 — preflight

Run status and verify the provider matching env_present.AI_PROVIDER (default ollama when unset) has its credential:

• ollama → ollama_reachable is true
• anthropic → env_present.ANTHROPIC_API_KEY is true
• openai → env_present.OPENAI_API_KEY is true
• gemini → env_present.GEMINI_API_KEY is true
• minimax → env_present.MINIMAX_API_KEY is true

If the provider can't be reached or the required key is missing, stop with a clear error rather than burning items on transient failures.

Also report items_unsummarized from status — that's the upper bound on what this run could process.

Step 2 — summarize

Default --limit is 50. Parse $ARGUMENTS for --limit N:

cd "${SYNDICATE_REPO:-$(pwd)}" && uv run python -m pipeline.cli summarize --limit ${limit:-50}

Each item takes ~1–18 minutes depending on provider/model and whether the item has an image (vision items are slower). For local Ollama on a Mac M-series, plan ~2 minutes per item on average. The skill runs in a forked subagent context so progress logs are isolated.

Output envelope:

{"ok": true, "result": {"examined": N, "summarized": N, "skipped": N, "errors": []}, "log_path": "..."}

Step 3 — report

✓ summarize · examined=X done=Y skipped=Z

skipped includes items the summarizer classified as banter, empty_content, or reaction (which have a skip_reason persisted so they never re-process). Surface unexpected errors[] entries verbatim — those are transient failures and will retry on the next run.