Validate Periscope config and required environment variables.

Fetch and deduplicate content into the raw stage.

Score a stage into the scored stage.

Filter scored items into the filtered stage.

Enrich filtered items into the enriched stage.

Generate a markdown summary from a stage.

Run fetch -> score -> filter -> enrich -> summarize in one call.

List recent runs and stage states.

Read run metadata.

Read items from a run stage.

Read a generated run summary.

Read in-memory server metrics.

Evidence-corpus overview: stored items by source, clusters, runs, plus claim and research-session counters. The 'what do we already know' probe.

Full-text search over every item ever collected (BM25, CJK-aware).

Most recently published items in the corpus, optionally one source.

Import a user export ({"items": [...]}) into the evidence corpus.

Each item declares its own authorship tiers through sections, so crowd text from an export cannot become evidence. Use this for sources Periscope must not scrape; see src/mcp/README.md for the payload shape.

dry_run=true reports what would be accepted and which items are rejected, without storing anything.

Claims by pipeline status (extracted | linked | graded) with verdicts, confidence, trust, ungraded_reason and independent-source counts.

One claim in full: verdict, evidence rows, source independence.

Open a long-session research task: decompose the question, investigate each sub-question against the corpus, return a cited markdown report. State persists — follow up later, even after restarts.

Push back on an existing session: the message can narrow, expand or challenge; the sub-question tree is revised and the report re-rendered.

Advance a research session by ONE round and get back the delta.

The loop chooses a verb each round: deepen a branch, fold the message into the question tree, finalise — or ask you for something only you can supply (then pending_request is non-null and the session parks). Returns the changed section ids and the new draft revision, not a whole re-rendered tree, so a round costs only what it touched.

Read the research document as a versioned draft: revision, origin (render | user_edit | merge), and per-section state — locked, stale, and the evidence item ids behind each section. Omit revision for the latest.

Rewrite one section as the author. Stored as its own revision and locked, so a later recompute marks it stale instead of overwriting your prose.

Reply to a question the session asked (skip: true to decline), then the loop continues. A parked session is not stuck: this is how you unpark it.

Session state: sub-question tree with statuses/answers/evidence ids, turn history, and the current rendered report.

Recent research sessions with their questions and statuses.

Send a webhook notification with the given variables.

Uses the webhook URL (from environment variable), request_body template, and headers from the Periscope config. Template variables #{date}, #{language}, #{important_items}, #{all_items}, #{result}, #{timestamp}, #{summary} are replaced in the URL and request_body before sending.