Diary digest

diary-digest

Turn a fat append-only journal into durable automation. The journal already paid for the mistakes; this skill makes sure each one is paid for once. It is the repeatable version of a manual harvest — profile, mine in parallel, reflect, propose gates, persist.

Principle (see global rule on mechanical gates): a lesson that lives only as prose in the log will be skipped again under pressure. The harvest is not done until each recurring mistake has a proposed mechanical gate (hook / script / lint / test / checklist), not just a paragraph.

Steps

1. Locate & profile

Find the journal (default discipline/DIARY.md; accept a path arg). Profile it cheaply before reading — this sizes the fan-out and surfaces the penalty taxonomy:

F="${1:-discipline/DIARY.md}"
wc -l "$F"
grep -Ec '^## ' "$F"                 # entry count
grep -Eo '\[[a-z][a-z-]+\]' "$F" | sort | uniq -c | sort -rn | head -30   # penalty taxonomy, by frequency

The penalty histogram is the single highest-signal artifact. Any label with count ≥2 is a recurring mistake and MUST end the harvest with a proposed gate.

Then read the digest watermark so you don't re-mine what a prior digest already covered:

grep -n '<!-- DIGEST-WATERMARK' "$F" | tail -1   # active (latest) marker. Anchor on '<!--' so prose mentioning the token isn't matched; parse its next_start_line:N

Mine only from that next_start_line:N value to EOF. The value is stable under append-only (old lines never shift, so a recorded line number stays valid forever). No marker found → first digest, mine the whole file.

2. Fan out diary-miner (parallel)

Split the un-digested range (watermark next_start_line → EOF; the whole file if no watermark) into ~2600-line chunks. Spawn one diary-miner subagent per chunk in a single message (parallel). Give each the exact path + line range. If the diary-miner agent type isn't registered in this session, fall back to general-purpose with the same extraction contract (MISTAKES with root-cause+gate / SUCCESSES / SKILL-CANDIDATES / THEMES / GOLD; terse; reply-is-the-data).

If the model fan-out is unavailable (overload / offline), degrade gracefully to a grep pass: grep -n -E '\[[a-z-]+\]|meu erro|deveria ter|errei' "$F" and read the surrounding entries yourself. A grep harvest is worse than a miner harvest but far better than none — never abandon the harvest because the fan-out failed.

3. Synthesize — k0-k4 reflection

Collect miner outputs, dedup across chunks, then reflect by metacognitive level (Logos Probabilis framing, global rule #17):

• k0 raw impulse — "ship it / skip the gate." Which mistakes were pure opportunism under time pressure?
• k1 ungrounded doubt — "looks like a bug, I'll fix it" → reverting hardened code; "fold the migration to be clean." Doubt acted on without cross-checking git/tests.
• k2 grounded action — the patterns that worked: cross-check before fixing, forward-only migrations, lock-before-idempotence, the review gate.
• k3 inflated doubt — real risk dressed as catastrophe (paranoia³). Did any response inflate?
• k4 meta — the pattern ACROSS the mistakes. Almost always: self-discipline fails under pressure; mechanical gates are the only durable fix. Name the specific shape this journal shows.

Group the deduped MISTAKES into families (gate-skipped, tests-green-but-ships, git/migration-state, stale-spec, timezone, env-empty-string, user-string-overwrite, worktree-state, …). Families, not incidents, are what you automate against.

4. Mistake → gate (delegate the mapping)

For each recurring family, derive the mechanical gate. Use the postmortem-to-gate skill's mapping (mistake-class → gate-type → where it lives) so the proposals are consistent across harvests.

5. Persist to Mycorrhiza

Consolidate the harvest into one rich link (not many small ones — recall is substring-based, dense links recall better). mcp__mycorrhiza__remember with tags: referencia-canonica, diary-digest, the project name, k4, plus a tag per mistake-family. First person, fact-dense, note-to-future-self. Run mcp__mycorrhiza__chain_status first if it's been many sessions (integrity check).

6. Stamp the watermark (append-only)

Append a fresh watermark marker at EOF so the next digest resumes where this one stopped. This is the ONLY write to the journal — never edit existing entries.

TODAY=$(date +%F)
TOTAL=$(wc -l < "$F")     # everything up to here is now digested (cumulative)
cat >> "$F" <<EOF

---

## 📍 DIGEST WATERMARK — $TODAY

<!-- DIGEST-WATERMARK v1 | digested_through_line:$TOTAL | next_start_line:$((TOTAL + 1)) | digest_date:$TODAY -->
EOF

digested_through_line is cumulative (prior runs + this one); next_start_line is where the next digest begins. Each run appends its own marker; grep … | tail -1 always finds the active one.

7. Report & propose — do NOT auto-edit the constitution

Output: the k0-k4 reflection + the family table (family → frequency → proposed gate → status). Then propose deliverables:

• new skills/subagents for repeated manual work (SKILL-CANDIDATES with 2+ hits) → can create directly under ~/.claude/skills/ or project .claude/skills/
• new discipline rules → PROPOSE and confirm before editing any global ~/.claude/CLAUDE.md (the constitution is human-ratified; "propor" ≠ "editar").
• project-specific technical gold → the project's ETERNAL.md (on-demand, per rule #19), not here.

Guardrails

• Read-only on the journal's content. The only write allowed is appending a watermark marker at EOF (step 6) — never edit or "clean up" existing entries; the journal is append-only (rule #5).
• Don't fabricate gates. If a mistake has no clean mechanical gate, say so and route it to a checklist/HUMAN.md instead of inventing a brittle hook.
• Cross-project safe: hardcode nothing about a specific company/stack. Read the journal's own taxonomy.