Listen

Turns work sessions into content. Data dir: ~/.config/opencode/listen/ (state.json, ideas/*.md, transient dump.md). Source: ~/.local/share/opencode/opencode.db (opencode only, always read-only).

Mode 1: extract (default)

1. Run:

   ~/.config/opencode/skills/listen/scripts/extract.sh

   Flags: --since <ms> (override state), --all (ignore state), --limit N (max 15). First-ever run defaults to a 3-day lookback.
2. Read the summary. If sessions_dumped: 0 → say "nothing new since <date>", stop. Never touch the backlog.
3. Read dump.md (in chunks if large).
4. Extract AT MOST 5 ideas per run, matching the gold criteria:
   • lesson — clever fixes, gotchas, tooling opinions, hard-won lessons
   • take — opinions, contrarian takes, tooling/AI philosophy
   • story — surprising, funny, or story-like moments
5. Dedup FIRST: ls ~/.config/opencode/listen/ideas/, read near-matching files. Same insight → append **resurfaced:** YYYY-MM-DD to the existing file; never fork a duplicate.
6. Write each idea to ideas/YYYY-MM-DD-<slug>.md (template below).
7. Only after idea files are written, update state (use max_time_updated_ms and session_ids from the script summary):

   cd ~/.config/opencode/listen && jq --argjson ms <MAX_MS> \
     --argjson ids '["ses_a","ses_b"]' \
     '.last_scan_ms = $ms | .scanned_sessions += $ids' \
     state.json > state.tmp && mv state.tmp state.json

   If state.json is missing, create {"last_scan_ms": 0, "scanned_sessions": []} first.
8. Report: N sessions scanned, M ideas written, K dupes linked — then one line per new idea: title (type) + hook.

Mode 2: brainstorm (/listen brainstorm)

1. List backlog: idea files with status: idea, grouped by type.
2. Empty backlog → offer to run extraction first.
3. Walk ONE idea at a time: show title, insight, hook. User picks: draft now / must-write / skip / archive.
   • must-write → draft immediately, no further questions.
4. Voice — terse lowercase:
   • all lowercase, short sentences, no emojis, no hashtags
   • line 1 = hook, then payoff, optional 1 line of context
   • threads: numbered posts, each ≤280 chars, each stands alone
   • one idea per post; concrete beats abstract
5. Accepted draft → append a ## draft section to the idea file and set status: drafted.
6. Dead idea → status: archived. Never delete files.
7. Close with counts: drafted / archived / still ideas.

Idea file template

---
status: idea
type: lesson
date: YYYY-MM-DD
source_session: ses_...
title: short title
---
the insight, paraphrased, 1-3 sentences.

**why post-worthy:** ...
**hook:** rough first line

Security (hard rules)

• Idea files NEVER contain: secrets, API keys, tokens, real usernames/emails, wallet addresses, private paths, or code quoted verbatim. Paraphrase.
• The DB is opened read-only (mode=ro). Never write to it.
• dump.md stays local; never paste large raw chunks into chat.
• No auto-posting anywhere, ever.

Edge cases

• state.json missing → extractor defaults to 3-day lookback (not --all)
• dump hit byte cap → summary reports skipped sessions; they return next run
• DB locked/missing → script exits cleanly with an error; tell the user to retry
• session with no text parts → skipped silently by the script
• duplicate insight on rescan → resurfaced line, no fork