Ingest Inbox Item

Workflow
Frontmatter schema
Worked example
Guardrails
Related skills

Related skills: Uses normalize-format, tag-by-topic, score-intuition-density, dedupe-against-corpus, update-topic-ledger. Called by the Librarian agent for each file in inbox/.

Workflow

Copy this checklist and track progress per file:

Ingest one inbox file:
- [ ] Step 0: Compute sha256 of file body; check .librarian-state.json for duplicate fingerprint
- [ ] Step 1: Invoke normalize-format → body + partial frontmatter
- [ ] Step 2: Invoke tag-by-topic → topics + pending_tags
- [ ] Step 3: Invoke score-intuition-density → score + signals
- [ ] Step 4: Invoke dedupe-against-corpus → CREATE | LINK | SKIPPED
- [ ] Step 5: If LINK, backlink matched seeds' related_seeds field
- [ ] Step 6: Write seed file with full frontmatter; status=seed; manual_edits=false
- [ ] Step 7: Append one line to corpus/seeds/.changelog.md
- [ ] Step 8: Invoke update-topic-ledger for each topic tag
- [ ] Step 9: mv inbox file → inbox/.processed/
- [ ] Step 10: Update .librarian-state.json with the fingerprint

Step 0: Hash the body (not frontmatter). If the fingerprint is in .librarian-state.json, exit with SKIPPED (already ingested: <seed-id>). Idempotency is a feature.

Step 1: normalize-format handles all supported formats (plain markdown, .jsonl Claude Code sessions, .json Claude.ai exports, Readwise exports, transcripts with speaker labels, link captures). Returns one or more {body, partial_frontmatter} pairs — multi-chunk outputs possible for long transcripts.

Step 2: tag-by-topic proposes 1–4 tags from the controlled vocabulary. If none match, it logs to topic-ledger.md#pending-tags and still assigns the closest existing tag.

Step 3: score-intuition-density computes 0–10 from 8 explicit signals. Auditable.

Step 4: dedupe-against-corpus returns one of SKIPPED (exact fingerprint match — exit), LINK (near-match — proceed but write related_seeds), or CREATE (no match).

Step 6: Write the seed file. Location: corpus/seeds/{id}.md where id = YYYY-MM-DD-slugified-title. If a seed with that id already exists, append -v2 rather than overwrite.

Step 7: Changelog format: YYYY-MM-DDThh:mm | ADDED | {seed-id} | from {original_path} | density={N} | topics={comma-list}

Frontmatter schema

---
id: 2026-04-23-dropout-as-ensemble-thinned-networks
title: "Dropout as ensemble over thinned networks"
created: 2026-04-21T09:14:00-07:00
source:
  type: inbox-note
  original_path: inbox/2026-04-21-dropout-as-ensemble.md
  ingested_at: 2026-04-23T14:32:01-07:00
  fingerprint: sha256:7c1d...
topics: [regularization, ensembling, dropout]
intuition_density:
  score: 8
  signals: [analogy_present, concrete_worked_example, counterfactual_offered, reframe_against_default, biology_to_ai]
status: seed
provenance:
  author: kushal
  confidence: owned
links:
  related_seeds: [2026-03-11-l2-as-gaussian-prior]
  parent_source: null
section_affinity: [agent-workshop]
word_count: 87
manual_edits: false
---

[body: preserved verbatim from normalize-format output]

Worked example

Input: inbox/2026-04-21-dropout-as-ensemble.md (87 words, plain markdown, user's own note).

Run:

Step 0: fingerprint new → proceed.
Step 1: normalize-format → body unchanged (plain markdown); title from first line or filename.
Step 2: tag-by-topic → [regularization, ensembling, dropout]; dropout new → logged to pending-tags.
Step 3: score → 8; signals fired: analogy_present, concrete_worked_example, counterfactual_offered, reframe_against_default, biology_to_ai.
Step 4: dedupe → near-match 2026-03-11-l2-as-gaussian-prior at Jaccard 0.32 (below 0.5 threshold) → CREATE (not LINK).
Step 6: Write corpus/seeds/2026-04-21-dropout-as-ensemble-thinned-networks.md with full frontmatter.
Step 7: Changelog: 2026-04-23T14:32 | ADDED | 2026-04-21-dropout-as-ensemble-thinned-networks | from inbox/2026-04-21-dropout-as-ensemble.md | density=8 | topics=regularization,ensembling,dropout
Step 8: Ledger: regularization seeds 3→4, temperature→hot; ensembling 1→2; dropout new row added.
Step 9: mv inbox/2026-04-21-dropout-as-ensemble.md inbox/.processed/
Step 10: Append fingerprint to state.

Output: one seed file, one changelog line, three ledger updates, one moved file.

Guardrails

If normalize-format fails, the inbox file stays put (not moved) and the changelog records ERROR | <file> | <reason>. No partial ingests.
If topic-tagger proposes a tag not in the controlled vocabulary that cannot map to one, log to pending-tags — never silently invent.
If a seed with the candidate id exists, append -v2; never overwrite.
Never ingest files under inbox/.processed/ or inbox/.trash/.
Run single-threaded. Parallel ingests corrupt the ledger and changelog.
If dedupe-against-corpus returns SKIPPED, move the inbox file to .processed/ with a renamed suffix -duplicate-of-{matched-seed-id} and exit — don't re-write the matched seed.

Quick reference

Idempotent: yes (fingerprint check in step 0).
Writes: one seed file + one changelog line + ledger updates + one moved inbox file + state update.
Failure mode: format error halts the pipeline for that one file; other files continue.

ingest-inbox-item