cicada-librarian

name	cicada-librarian
description	Turns the user's own agent (Claude Code, Codex, Cursor — whatever runs this skill) into Cicada's consolidation engine, no Cicada API key required. Use at the end of a working session, whenever the user says "remember this" / "consolidate our chat", or right after watching a video with claude-video. Also covers capturing episodes and URLs mid-conversation.

Cicada librarian skill

Cicada's Sleep cycle normally does entity/claim extraction as a nightly batch job. This skill lets the user's own frontier-model subscription (Claude Code, Codex, Cursor — no separate Cicada API key) do that extraction live, inside the conversation, using the same cicada_write_claim write path the Sleep cycle uses. A keyless local Ollama nightly batch remains the fallback for whatever this skill doesn't catch.

When to consolidate

• End of a working session — before the conversation closes, sweep it.
• On request — user says "remember this", "consolidate our chat", "save that", etc. — do it immediately, don't wait for session end.
• After watching a video — once the claude-video chain (below) returns a transcript + summary, consolidate that content the same way.

Capture tools

• cicada_save_episode(content, title) — stage a raw conversation snippet. Use for anything worth remembering: decisions, facts, plans.
• cicada_save_url(url, note?) — save a link (article, repo, bookmark). Cicada fetches and indexes it.
• Video: run the claude-video skill's /watch <url> <question> (github.com/bradautomates/claude-video) to get transcript + frames + summary. Then:
  1. cicada_save_url(url, note=<your faithful summary>) — the transcript is the source of truth; the note is your summary of it, not embellishment.
  2. cicada_write_claim(...) for relational facts the video establishes — e.g. subject=<video title>, predicate=is-about, object=<topic>, or predicate=recommends, object=<thing recommended>.

Consolidate loop

Run this to drain the backlog (end of session, on request, or after a video save):

1. cicada_pending(limit?) — list unprocessed episodes (processed: false).
2. For each episode, read its content.
3. For each atomic fact in it, call cicada_write_claim(subject, predicate, object, observer=...) — one claim per fact, never bundle two facts into one call.
4. cicada_mark_processed([episode_ids]) — once you've extracted everything worth keeping from those episodes. An unmarked episode is still a safety net for the next real Sleep cycle, so only mark what you actually consolidated.

Observer tagging (load-bearing)

This is what gives the graph real perspective diversity — get it right:

observer	when
rodrigo	the user stated/asserted this themselves. Trust-protected — a later agent claim can never silently overwrite it.
agent	you inferred, deduced, or noticed this — not something the user said outright.
external:<name>	a claim attributed to a third party (someone else's stated belief, quoted or paraphrased). Name the source when known.

Default is agent if omitted — don't rely on the default when the user clearly stated the fact themselves; tag rodrigo explicitly.

Grounding

Write rich but strictly faithful content — only facts actually present in the episode/transcript, never invented or inferred-and-passed-off-as-stated. Prefer [[wikilinks]] to existing entities over prose restatement. If something is ambiguous, write it as an agent-observed claim at lower confidence rather than guessing and asserting it as fact.

Model note

This loop runs on the user's own frontier-model subscription (whatever runs this skill) — no Cicada API key needed. A keyless local Ollama nightly batch is the fallback consolidation path for episodes this skill doesn't drain.

Never hand-edit entity files

Same rule as recall: write only through cicada_write_claim / cicada_save_episode / cicada_save_url. Never edit entities/, hubs/, or _index.md directly — that bypasses provenance and dedup.