| name | session-recovery |
| description | Diagnose and recover a failed or missing cron trading session (no Telegram received, missing daily file, stale data, failed commit/push, validation failure). Use when a scheduled session (09:00/12:00/15:30/18:00/21:30) didn't complete, when data looks stale, or when Jose says "yesterday's close is missing" or "no telegram arrived". |
Session Recovery
Recover a broken cron session safely — the #1 risk here is double-applying trades or fabricating a ledger entry. Diagnose first, mutate last.
Phase 1 — Diagnose (read-only, always do all of it)
tail -50 /tmp/makemerich-session.log
tail -50 /tmp/makemerich-daily.log
journalctl --user -u 'makemerich*' --since "today" --no-pager | tail -30
git -C /home/hustle/projects/makemerich status --short
git -C /home/hustle/projects/makemerich log --oneline -5
jq -r '.fetchedAt' data/.prices-latest.json
ls data/$(date +%Y-%m-%d).json 2>/dev/null || echo "MISSING daily file"
head -30 LEDGER.md
jq --arg d "$(date +%Y-%m-%d)" '[.[] | select(.date==$d)]' data/trades/$(date +%Y-%m).json 2>/dev/null
Classify the failure before touching anything:
| Symptom in log | Class | Severity |
|---|
error fetching prices / fetch-prices non-zero | A: fetch failure | Recoverable |
Claude CLI failed or timed out | B: agent timeout | Benign — fallback text used, pipeline continued |
apply-trades failed | C: trade application failure | Dangerous — verify trade log vs portfolio |
Data validation FAILED — skipping commit/push | D: integrity gate | Commit intentionally blocked — fix data first |
git push failed | E: push failure | Trivial |
| Nothing in log at all for the slot | F: session never ran | Check cron/worker upstream |
Phase 2 — Recover, per class
A — fetch failure. Usually transient (Yahoo/Coinbase). Re-run pipeline read steps only:
node scripts/fetch-prices.js && node scripts/fetch-history.js && node scripts/update-portfolio.js
node scripts/generate-signals.js && node scripts/generate-quant-signals.js && node scripts/execute-signals.js
If it was the 21:30 close, continue to the "missed close" recipe below. If intraday, nothing else needed — the next cron slot recovers on its own.
B — agent timeout. No action. The LEDGER got fallback text ("Automated close — agent unavailable"). Optionally improve the analysis line by appending nothing — never rewrite the entry. This class is working-as-designed.
C — apply-trades failure. The danger zone. Determine which side of the write it died on:
- Orders that should have applied:
jq . data/.trade-orders.json
- Trades actually logged today (Phase 1 step 4) vs holdings deltas in
portfolio.json
- Three states:
- Logged AND holdings changed → it actually succeeded; failure was downstream. Move on.
- Neither → re-run
node scripts/apply-trades.js once, then node scripts/update-portfolio.js.
- One but not the other → stop, do not re-run (re-running duplicates). Reconcile by hand: make trade log and holdings agree per CLAUDE.md data-repair quality bar, then
validate-data.js.
D — validation gate. Run node scripts/validate-data.js and read which of the 5 checks failed. Fix per CLAUDE.md invariants (source-of-truth reconstruction; flag what can't be derived). Never bypass the gate by committing manually while it fails.
E — push failure. git pull --rebase && git push. If diverged, the remote wins for history — never force-push this repo.
F — never ran. The trigger lives outside this repo (hustle BullMQ workers / systemd). Check systemctl --user status, report upstream, then run the missed session manually (below).
Missed 21:30 close — full manual replay
Only if today's LEDGER entry is missing AND market data for today is obtainable:
bash scripts/daily-update.sh
That is the whole recipe — the script is the source of truth for order-of-operations and already contains the validation gate, fallback analysis, commit convention (log: Day N — ACTION, ...), and Telegram send. Do NOT re-implement its steps by hand unless it dies mid-way; then resume from the failing step per the classes above.
If the missed day is in the past (e.g. noticed next morning): prices for "yesterday close" are gone. Create the daily file from the last known prices, mark the analysis line "Reconstructed — session missed", and note it in FINDINGS.md. Never invent close prices.
Phase 3 — Verify (always)
