Context links — operating¶

Pipeline panel¶

Settings > Pipeline shows the context toggle plus live stats (eligible / waiting / done / stories with links / links added / judge accept rate / rejections by reason) and a "Find context now" button. POST /api/admin/context[?story_id=N] triggers a run for all eligible stories or a single one (admin key).

Diagnose and evaluate¶

python -m app.context.diagnose --story-id 42            # one story, writes nothing
python -m app.context.diagnose --story-id 42 --dry-run  # queries + decisions, no writes
python -m app.context.evaluate --stories 30 --markdown  # dry run over 30 recent stories
# planner mode: gold-set recall + rejected-candidate audit + per-story traces
python -m app.context.evaluate --gold config/context_gold.example.yaml
python -m app.context.evaluate --audit-rejects 10 --labels labels.yaml
# inside Docker:
docker exec discover-feed python -m app.context.evaluate --stories 30 --markdown

Planner runs also write a context_traces row per story — the planner's raw output, every task, every query with its window and filters, each candidate's decision with scores and the rule that fired, plus stage timings (planner / search / fetch_extract / compose / total) and wait entries (llm, cooldown, budget, priority_gate). diagnose prints the latest trace as JSON; median stage timings and waiting reasons show in Settings > Pipeline.

Windows

.venv\Scripts\python -m app.context.evaluate --stories 30 --markdown

Reasons table¶

context_reason on a story that kept nothing:

Reason	Meaning
`no_angles`	No search angles could be built (LLM absent and no entities)
`skipped_cooldown`	Providers were cooling down; pass paused, story stays `pending`
`provider_errors:<kind>`	Providers failed; story is `retry` with backoff
`no_hits`	Searches ran, nothing came back
`all_rejected`	Candidates found, all rejected

Candidate rejection reasons (in context_candidates): same_domain, social, aggregator, denied, bad_url, duplicate_url, already_ingested, language, off_topic, same_event (with same_event_signal), near_duplicate, min_words, flagged, domain_cap, opinion_cap, not_article, judge_rejected, judge_failed, judge_same_event, robots, fetch/paywall failures.