daily-shitrat

name	daily-shitrat
displayName	Daily SHITRAT
description	Operate, inspect, and optimize the Daily SHITRAT loop. Use when running Daily SHITRAT, reviewing a Daily SHITRAT attempt, doing a Daily SHITRAT Dream/optimizer pass, checking brain daily artifacts, fixing report/Discord metadata, or preparing ShitRat-authored PR plans.
version	0.1.0
author	joel
tags	["joelclaw","brain","daily","shitrat","automation","optimizer"]
disable-model-invocation	true

Daily SHITRAT

Operate the Daily SHITRAT loop and its optimizer pass without losing the trail.

This skill has two modes:

• Daily run: inspect or execute a real brain daily attempt.
• Dream/optimizer pass: review the latest run and write separate optimizer artifacts. The Dream pass does not create a new daily attempt.

When to Use

Use this skill when the user says any of:

• "Daily SHITRAT"
• "SHITRAT Dream"
• "daily brain run"
• "brain daily"
• "latest sealed attempt"
• "daily run optimizer"
• "daily-shitrat-run-optimizer"
• "fix the Daily SHITRAT loop"
• "why did the daily report/Discord publish break?"

Source Paths

Purpose	Path
Daily automation memory	/Users/joel/.codex/automations/daily-shitrat-agent-turn/memory.md
Optimizer memory	/Users/joel/.codex/automations/daily-shitrat-run-optimizer/memory.md
Daily data root	/Users/Shared/joelclaw/data/brain-daily
Optimizer outputs	/Users/Shared/joelclaw/data/brain-daily/optimizer/YYYY-MM-DD/
Brain CLI repo	/Users/joel/Code/joelhooks/joelclaw-brain
Discord bridge repo	/Users/joel/Code/joelhooks/pi-discord-threads
Codex sessions	/Users/joel/.codex/sessions/**/rollout-*.jsonl

Resolve Codex memory with a default. An unset CODEX_HOME is not evidence that memory is missing:

CODEX_HOME="${CODEX_HOME:-/Users/joel/.codex}"
DAILY_MEMORY="$CODEX_HOME/automations/daily-shitrat-agent-turn/memory.md"

Hard Rules

• Do not run brain daily start unless Joel explicitly asks for a new Daily SHITRAT attempt.
• For Dream/optimizer work, treat the daily memory as read-only input. Write separate output artifacts under brain-daily/optimizer/YYYY-MM-DD/.
• Do not call Managed Agents, Anthropic Dreams, or any connected memory-store API unless Joel explicitly configures that later.
• Do not mutate product/customer-facing repos during optimizer work.
• Do not merge, deploy, or make GitHub writes as Joel.
• For GitHub-visible work, prepare ShitRat-authored PR plans and use ShitRat dry-run first.
• If an attempt is already running or sealed, reuse it. Do not create duplicate attempts or duplicate Discord starts.
• Do not open duplicate VISION PRs. Reference existing PRs when they already cover the work.

Daily Run Workflow

1. Inspect current state before doing anything:

brain daily latest --base-dir /Users/Shared/joelclaw/data/brain-daily
find /Users/Shared/joelclaw/data/brain-daily -maxdepth 3 -name latest.json -print

2. If a run exists for the day, reuse it. If it is sealed, inspect, repair metadata, or publish reports; do not start another attempt unless Joel asked for that exact action.

3. Read input memory before broad scans:

CODEX_HOME="${CODEX_HOME:-/Users/joel/.codex}"
sed -n '1,220p' "$CODEX_HOME/automations/daily-shitrat-agent-turn/memory.md"

4. Append an early useful receipt within 120 seconds. During broad repo scans, restores, transcript archaeology, or report repair, append progress receipts at least every 180 seconds:

brain daily progress \
  --phase scan \
  --status active \
  --summary "Read input memory and latest manifest; inspecting sealed attempt metadata"

Progress receipts must say what was read, decided, blocked, repaired, or verified. scan-started alone is too weak.

5. Keep ledger sequence contiguous. Before sealing, verify brain-dump.jsonl, checksum, manifest, report metadata, and latest pointer.

6. Prefer brain daily finish --publish-report for a normal completion. It should be robust if another actor sealed first: inspect the sealed attempt, merge final Discord metadata, and re-render/re-publish report metadata instead of failing the loop.

Dream / Optimizer Workflow

Dream work is a local analog of the "read input memory, write output memory" pattern. It is not an Anthropic Dreams API call.

1. Inspect inputs first:

CODEX_HOME="${CODEX_HOME:-/Users/joel/.codex}"
sed -n '1,260p' "$CODEX_HOME/automations/daily-shitrat-agent-turn/memory.md"
test -f "$CODEX_HOME/automations/daily-shitrat-run-optimizer/memory.md" && sed -n '1,260p' "$CODEX_HOME/automations/daily-shitrat-run-optimizer/memory.md"
find /Users/Shared/joelclaw/data/brain-daily -maxdepth 3 -name latest.json -print

2. Identify the latest sealed attempt and inspect:

• manifest.json
• brain-dump.jsonl
• brain-dump.sha256
• wzrrd-report/index.html
• receipt JSON/Markdown files
• Discord request/response files
• matching Codex rollout transcript

3. Use transcript summaries before raw transcript reads:

cd /Users/joel/Code/joelhooks/joelclaw-brain
bun run src/main.ts daily transcript-summary \
  --session /Users/joel/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl \
  --max-outputs 8 \
  --suspicious-output-token-threshold 1000

Expand raw rollout ranges only when the summary points to a suspicious command, final message, brain daily command, receipt gap, or oversized output. Do not dump huge transcript chunks by default.

4. Write optimizer outputs under:

/Users/Shared/joelclaw/data/brain-daily/optimizer/YYYY-MM-DD/

Required files:

• dream-memory.md: proposed reorganized memory for the next Daily SHITRAT run.
• run-quality.json: facts, metrics, source paths, transcript ids, and verification commands.
• recommendations.md: concise findings and prioritized changes.
• patch-plan.md: exact proposed repo/code/prompt changes.

5. Update optimizer memory with pointers, top regressions, implemented fixes, verification commands, and carryover:

/Users/joel/.codex/automations/daily-shitrat-run-optimizer/memory.md

Do not overwrite the daily automation memory.

Loop Quality Checks

Analyze the loop, not just system health. Look for:

• duplicate automation starts
• abandoned attempts
• concurrent writers
• sequence jumps
• seal races
• stale latest pointers
• missing report publish metadata
• missing Discord URL metadata
• incomplete final summaries
• oversized context reads
• missing early progress receipts
• gaps over 180 seconds during broad scans
• weak event names
• duplicate/conflicting findings
• non-actionable PR plans

Good facts include counts, paths, timestamps, sequence numbers, command names, transcript ids, checksum status, report URLs, and Discord message ids/URLs.

Report and Discord Contract

Manifest and latest state must persist publish metadata:

• report.publish.url
• report.publish.expiresAt
• discord.threadUrl
• discord.startedMessageUrl
• discord.finalMessageUrl

If the Discord bridge returns ids but omits URLs, Brain should derive:

• thread URL from guildId and threadId
• message URL from guildId, threadId, and messageId

brain daily report --publish after seal may update report metadata and repair derived Discord URLs. It must not mutate brain-dump.jsonl.

If final Discord metadata lands after the first report publish, re-render and re-publish so the local and published report HTML include finalMessageUrl.

Ledger Event Shape

Use strong event names and canonical records:

• agent.receipt for source reads, progress, decisions, and verifications.
• agent.finding for one canonical finding per problem or confirmed healthy state.
• agent.blocker for actual blockers, not generic caution.
• agent.pr_plan for ShitRat-authored GitHub-visible work.
• agent.sleep for explicit "do not do this" carryover.

Collapse duplicate findings, blockers, PR plans, and carryover before final summary. The report renderer should also canonicalize duplicates and flag vague PR plans.

PR Plan Contract

Every agent.pr_plan must be action-shaped:

• one repo
• one branch
• title
• base branch
• actor
• scope
• dry-run proof command
• GitHub write status
• validation commands
• body file or body text
• approval gate
• exact next command
• existing PR URL when present

Do not write a PR plan like "open a PR for the fixes." That is not actionable.

Current Known Failure Modes

Verify these whenever touching Daily SHITRAT code or prompts:

• Starting should reuse an already-running attempt.
• Duplicate Discord starts should be avoided.
• report.publish.url and expiresAt should persist into manifest and latest.
• brain daily report --publish after seal should update report metadata.
• brain daily finish --publish-report should tolerate another actor sealing first.
• Final Discord URL metadata should survive in manifest, latest, local HTML, and published HTML.
• Transcript inspection should start with brain daily transcript-summary.
• Progress receipts should appear early and during long scans.

Useful Verification Commands

From /Users/joel/Code/joelhooks/joelclaw-brain:

bun test --timeout 15000
bun run typecheck
bunx oxfmt --check src/daily.ts src/cli.ts test/daily.test.ts
bun run src/main.ts daily report --date YYYY-MM-DD --base-dir /Users/Shared/joelclaw/data/brain-daily
shasum -a 256 /Users/Shared/joelclaw/data/brain-daily/YYYY-MM-DD/attempt-001/brain-dump.jsonl

For transcript review:

bun run src/main.ts daily transcript-summary --session /absolute/path/to/rollout.jsonl

For progress during a live run:

brain daily progress --phase scan --status active --summary "Read latest manifest and verified report publish metadata"

Closeout

A good Daily SHITRAT closeout names:

• dream output paths or sealed attempt path
• code/prompt changes made
• tests and verification commands run
• report URL and Discord URLs if relevant
• remaining loop risks

Keep it short. Joel needs the shape, the evidence, and the next useful constraint.