Mission Control / Overview

Docs lane

Core docs uit /home/jbvandongen/.openclaw/workspace en tekst-output uit /home/jbvandongen/.openclaw/workspace/output zijn nu direct editable in de app.

Guardrails

Alleen allowlisted root docs en lokale outputfiles zijn schrijfbaar.
Geen verborgen autosave, alleen expliciete save button.
Draft, saved en reset states zijn zichtbaar per file.

Lokale file lane

Document browser + editor

Kies links een bestand, bekijk de inhoud en schrijf pas naar disk wanneer je bewust opslaat.

Veilige editor

HEARTBEAT.md

/home/jbvandongen/.openclaw/workspace/HEARTBEAT.md

Disk state in sync

# Heartbeat Instructions

(Note: Heartbeats run via `qwen2.5:7b` on Ollama to save tokens.)

## Daily Summary (21:00)
Read /home/jbvandongen/forecast-trader-bot/trades.jsonl. Calculate total P/L, wins/losses per strategy for S8-S15. Send a concise summary to #daily-digest on Discord.

## Discord Routing (Every Heartbeat)
Run `scripts/discord_router.py` to fetch new trades.
- For each new trade, use the `message` tool to post the trade details (entry, exit, P&L, confidence) to the corresponding #s1-s15 channel.
- Post a high-level summary (e.g., "S15 +$2.45") to #benjamin-main only if the trade is significant (>$5 P&L).

## Direct Messages
Always respond to direct messages from Jim. Never go silent on direct chat.

---

## Reporting
Heartbeat turns should usually end with HEARTBEAT_OK.
Only send a direct Telegram message when Jim needs attention.

If `memory/heartbeat-state.json` is corrupted, replace it with:
```json
{"lastChecks": {"telegram": null, "botStatus": null, "vpsHealth": null, "errorLog": null, "securityAudit": null, "lastDailyChecks": null}}
```
Then alert Jim.

## Every Hour
**RULE 1: Write to session_log.jsonl after EVERY significant action.** Do NOT wait for the heartbeat to remember.

**RULE 2: DAILY MEMORY FLUSH (MANDATORY, DETAILED).** Every heartbeat MUST append a detailed entry to `memory/YYYY-MM-DD.md` under `## HH:MM`. This is not optional and may not be a one-line "status OK".

Every heartbeat memory entry MUST include:
- What was checked or worked on this wake, including source/trigger
- Commands/scripts/tools run and compact outputs or result summaries
- Files created/modified, always full absolute paths
- Decisions made and why
- Errors, blockers, and what I tried before calling it a blocker
- Subagent results or spawned subagent handoffs, with role and purpose
- Trading facts from real data if trading was checked: new trades, stale data, WR flags, service state
- Next checkpoint or why no further action was needed

Minimum content per heartbeat: 1,500 bytes. Target daily memory size: 20KB+ when there is real activity. If nothing happened, still write a detailed "no material activity" entry with verified service checks, recent files, cron state, and trading-data freshness.

Use `echo >>`, append scripts, or the `edit` tool. NEVER use `write` on existing daily memory files.

**Heartbeat flush (append-only detailed writer):**
```bash
bash /home/jbvandongen/.openclaw/workspace/scripts/memory_30min_flush.sh
```

**Main session (me, Benjamin):** After each significant action, append to `~/.openclaw/workspace/memory/session_log.jsonl`:
```json
{"ts":"14:05","hour":"14","date":"2026-04-04","action":"fix","target":"route.ts","summary":"Phase1 extractie gefixt: dict {strat:score} ipv {strat:{score}}","files":["~/.openclaw/workspace/apps/mission-control/src/app/api/autoresearch-v6/route.ts"]}
```

**session_log.jsonl fields:**
- `ts`: HH:MM current time
- `hour`: HH (the hour to group under)
- `date`: YYYY-MM-DD
- `action`: fix | create | spawn | decision | error | blocker
- `target`: short name of what was worked on
- `summary`: 1-line description of what happened
- `files`: array of absolute paths created/modified (empty array OK)

**Flush logic:** Only flushes COMPLETE hours (not the current hour). Entries stay in session_log.jsonl until the next hour starts.

## Cron Failure Monitoring (MANDATORY — added MAR 17 after 01:00 crash miss)
**JE MOET cron failures ZELF checken, niet wachten tot Jim het vraagt!**

Wanneer cron loopt:
- Check binnen 30 minuten na cron start of het gelukt is
- Check `~/benjamin_autoresearch/logs/cron.log` voor errors
- Als het gecrasht is → fix het DIRECT en rapporteer naar #benjamin-main
- Update `memory/heartbeat-state.json` met cron check timestamp

Stappen bij cron failure (NACHT: fix eerst, rapporteer later):
1. Lees cron log, vind de error
2. Fix het probleem DIRECT (geen approval vragen)
3. Test de fix manueel
4. Nachtcron opnieuw draaien als de fix werkt
5. Log naar memory/YYYY-MM-DD.md
6. Naar #benjamin-main ALLEEN als het nog stuk is na je fix

### ZNKE Daily Meeting (Every Heartbeat — 08:00-22:00)
After the 08:00 cron runs `znke_daily_meeting.py`, the digest file is written to `~/.openclaw/workspace/logs/znke_digest_YYYY-MM-DD.txt`.
On each heartbeat:
1. Check for new digest file: `ls ~/.openclaw/workspace/logs/znke_digest_$(date +%Y-%m-%d).txt`
2. If exists AND `znke_digest_YYYY-MM-DD_sent.txt` does NOT exist → run `znke_digest_poster.py` to post to Discord #znke-digest (1486073780089524245)
3. This handles the case where the 08:00 cron ran but Discord post failed

## ZNKE Autonomous Tasks (RE-ACTIVATED — 2026-04-01)
**Jim directive 2026-04-01:** ZNKE mag weer autonoom worden opgepakt naast trading en Toerie. Niet meer skippen.

ZNKE taken zijn weer actief in heartbeats. Werk parallel waar logisch, maar houd trading health checks en Toerie control-room werk ook bij.

## Trading Autonomous Tasks (Every Heartbeat — day only, 08:00-22:00)
**VERIFICATION REMINDER:** Before posting ANY trading claim (WR%, P&L, win rate): run the actual data check first. Say "niet gecheckt" if you have not verified. Never post numbers you have not confirmed.

**AUTONOMOUS PIPELINE PLAN — READ THIS FIRST:**
1. Read `~/.openclaw/workspace/planning/AUTONOMOUS_PIPELINE_PLAN.md`
2. Find the first incomplete unblocked task in Phase 1 (critical fixes)
3. Execute it directly OR spawn a Codex subagent (for code tasks)
4. Update status in TRADING_AUTONOMOUS_TODO.md when done
5. Log to memory/YYYY-MM-DD.md
6. Post summary to #trading-digest on Discord (1486278316544491530)

**Autonomous rules:**
- Jim is offline — work autonomously without asking
- Only ask if: cost > $10, something broken, or day loss > $10
- Prioriteit trading: SWARM fix > GPU burst > S25-S30 research
- ZNKE taken weer actief waar heartbeats of todo's daarom vragen

### Service Health Check (every heartbeat):
- V6 autoresearch: `systemctl --user is-active benjamin-v6.service` → must be "active"
- Blitz sim: `systemctl --user is-active benjamin-blitz.service` → must be "active"
- Mission Control: `systemctl --user is-active mission-control.service` → must be "active"
- MiroFish: Check JSON freshness — `stat -c %Y ~/.openclaw/workspace/logs/health_watchdog.log` or check `mirofish_cache.json` age directly. Regime comes from JSON files, not HTTP. If MiroFish service is expected, check `systemctl --user is-active benjamin-mirofish.service`. Otherwise skip.
- If any crashed → restart it immediately, log to memory

### Paper Monitor (every heartbeat, day only):
- Run `python3 ~/.openclaw/workspace/scripts/paper_monitor.py`
- Check for demotion triggers (WR<45%, 5x loss, DD>8%)
- If demoted → log to memory, update strategies.json

### Daily Digest (once daily, 21:00):
- Read `~/benjamin_autoresearch/logs/paper_trades.jsonl`
- Calculate P&L per strategy, win rate, total trades
- Post summary to #daily-digest on Discord

### Trading Daily Autonomous Actions (every heartbeat, day 08:00-22:00)
Every heartbeat check and act:
1. **Count new trades per strategy** since last check
   - Read `~/benjamin_autoresearch/logs/paper_trades.jsonl`
   - Track which strategies generated new trades
2. **WR check for all strategies** (rolling window of last 50 trades)
   - For each strategy with ≥50 trades: calculate WR on last 50
   - Flag strategies where WR < 55% (below live gate)
   - Flag strategies where WR < 45% (critical review trigger)
3. **Immune response for trading wounds** — spawn the right subagent chain immediately:
   - No trades from a strategy for 24h while enabled → spawn Victor to diagnose scanner/config/threshold/data cause
   - WR < 55% on last 50 → spawn Victor for signal diagnosis, then Codex/Sam only if code/config fix is needed
   - Service down or stale paper data → spawn Codex/Sam to fix service/log/data path, then verify with real commands
   - OOS/overfitting concern → spawn Diocles/Victor to challenge the calibration before promotion
4. **auto_wr_boost.py** — run for flagged strategies
   - If WR < 55%: run `python3 ~/.openclaw/workspace/scripts/auto_wr_boost.py --strategy <name> --train-only`
   - Accept no live candidate until OOS 70/30 validation passes
   - Log action and outcome to memory/YYYY-MM-DD.md
5. **Log to TRADING_AUTONOMOUS_TODO.md** — mark daily tasks complete as executed

## Autonomous Body / Immune System Loop (Every Heartbeat)
This is the 24/7 subagent handoff model Jim asked for. Treat problems like a body treats a wound: detect, send the right cells, repair, verify, remember.

### Generic workflow lane
1. **Detector agent** notices a signal: trend, bug, stale data, no trades, bad WR, content opportunity, tool failure.
2. **Diagnosis agent** explains the likely cause and required next action.
3. **Builder/content agent** creates the fix, post concept, data patch, automation, or script.
4. **Tool scout agent** checks which tool should execute or validate it: n8n, Postiz/content scripts, Forecast Trader, Mission Control, Discord, Toerie, ZNKE, TRIBE v2 only for multimodal neuroscience/brain-response analysis, or local scripts.
5. **Validator agent** challenges the plan, runs evidence checks, and rejects overfit/unsafe work.
6. **Benjamin** applies only verified actions, logs the result, and reports to Jim only for meaningful progress, blockers, or approval needs.

### Content/marketing lane
- Trend found → Lisa researches the signal → Vera drafts the post → Sam checks tooling through n8n/Postiz/content scripts → Diocles validates whether the idea can work → Benjamin schedules/posts only after approval if public.
- TRIBE v2 is not the team/workflow orchestrator. Use it only when the task explicitly needs multimodal brain-response or in-silico neuroscience analysis of video/audio/text stimuli.

### Trading lane
- No trades → Victor checks config/scanner/threshold/data → Codex fixes code if needed → Victor validates paper-trade generation.
- Bad trades / WR <55% → Victor diagnoses regime/edge problem → auto_wr_boost runs train-only → Diocles checks overfitting → OOS 70/30 before live candidate.
- Service crash/stale data → Sam/Codex fixes service → Benjamin verifies `systemctl`, logs, and curl/API output.

### Rule
Quiet monitoring alone is not enough when a wound exists. If the heartbeat finds a real wound, it must either spawn the right subagent, fix it directly, or log a genuine blocker with proof.

### Morning Report (every day 07:00 CET — #trading-digest 1486278316544491530)
Post to #trading-digest with:
- How many new trades generated (since yesterday)
- Per strategy: current WR on last 50 trades
- Flag strategies above 55% WR (eligible for live consideration)
- Flag strategies below 55% WR (needs attention / auto_wr_boost ran)
- Total P&L yesterday
- Any service restarts or errors overnight
Format: compact bullet list, no markdown tables, plain text

## Every Heartbeat
- **Read `~/.openclaw/workspace/daily-plan.md`** — follow today's plan tasks. Find first incomplete task (⬜), execute it, mark done (✅). Skip if no tasks or all done.
- Update `memory/heartbeat-state.json` timestamps for checks performed
- Git backup: run auto-git-sync if configured. Alert Jim only for real breakages.
- System health check (alert if critical issues found)
- Cron failure check (alert on new failures only, not repeats already reported)

## Once Daily
- Check bot status: any crashes, daily loss limit hit?
- Check VPS Chicago: responsive?
- Memory index: read recent `memory/YYYY-MM-DD.md` and update `MEMORY.md` if needed

## Weekly
- Verify gateway: `openclaw gateway probe` → must show RPC: ok
- Verify gateway service: `systemctl --user status openclaw-gateway.service` → must be active
- Review `MISTAKES.md` — anything worth escalating to Jim?

## Diocles Check (Every Heartbeat, 08:00-22:00)
After all checks above, call Diocles (devil's advocate):
1. Spawn a quick minimax/MiniMax-M2.7 subagent (timeout=60s)
2. Prompt: "Lees memory/2026-MM-DD.md (laatste entries) en heartbeat-state.json. Stel 2-3 kritische vragen over recente beslissingen. Format: [DEMANDING_QUESTIONS]...[/DEMANDING_QUESTIONS]. Als alles OK is: 'Geen kritische vragen deze ronde.'"
3. Als Diocles disputed items vindt: log naar memory/YYYY-MM-DD.md, post samenvatting naar #benjamin-main
4. Diocles rapporten: ~/.openclaw/workspace/logs/diocles_YYYY-MM-DD.md

## When to Reach Out (Proactively)
- Bot crashed or hit $30 daily loss limit
- VPS offline for >10 minutes
- Something genuinely urgent

## When to Stay Quiet (HEARTBEAT_OK)
- Late night (23:00-08:00) unless urgent
- Nothing new since last check
- You just checked <30 minutes ago
- All systems nominal

## Night Runner Check (EVERY SESSION START)
When asked about plans or schedule: ALWAYS check cron list first.
Key overnight jobs (never forget):
- 01:00 night_runner.sh (pipeline v3 — MiroFish + AutoResearch loop)
- 02:30 AutoResearch optimizer
- 03:00 CRON MONITOR (self-fix if night_runner failed)
- 05:00 Daily planner
- 05:30 Team briefing
- 06:30 Briefing → Jim

Save flow

Scope

Docs

Bestandsgrootte

12.8 KB

Laatst opgeslagen

03 mei 2026, 21:01

Write safety

Alleen expliciet toegestane lokale bestanden zijn schrijfbaar.
Geen shell writes, alleen server-side filesystem writes binnen Mission Control.
Je ziet altijd exact welk absolute pad je opslaat.

Laatste save resultaat

Nog niet opgeslagen in deze sessie.