Diagnose "clean sweep" reports from scripted multi-agent review crons (Hermes `bug-hunt-daily.sh`, future siblings in the family — `wa-*-review-*`, `*-nightly-review-*`) when the report's PR/bug/agent-failure counts are all zero but the scan didn't actually happen. Triggers when a daily bug-hunt / nightly-review cron posts a Slack report with all-zero numbers, a sibling report from the same minute surfaces the real failure, a known review CLI was rate-limited / preflight-failed and the cron ran anyway, or the user pastes such a report and asks "is this right?". Classifies into FOUR buckets (PR-discovery suppressed, agent-preflight suppressed, agent-execution suppressed, fail-closed-not-armed), identifies whether the script's FAILURE_WARNING gate masked the failure, and produces the recipe for the next sibling cron. v1.1.0 (2026-07-23) — adds Bucket E for concurrent codex-review processes racing for the same model-list cache; verified on bug-hunt run 20260723_090145.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Diagnose "clean sweep" reports from scripted multi-agent review crons (Hermes `bug-hunt-daily.sh`, future siblings in the family — `wa-*-review-*`, `*-nightly-review-*`) when the report's PR/bug/agent-failure counts are all zero but the scan didn't actually happen. Triggers when a daily bug-hunt / nightly-review cron posts a Slack report with all-zero numbers, a sibling report from the same minute surfaces the real failure, a known review CLI was rate-limited / preflight-failed and the cron ran anyway, or the user pastes such a report and asks "is this right?". Classifies into FOUR buckets (PR-discovery suppressed, agent-preflight suppressed, agent-execution suppressed, fail-closed-not-armed), identifies whether the script's FAILURE_WARNING gate masked the failure, and produces the recipe for the next sibling cron. v1.1.0 (2026-07-23) — adds Bucket E for concurrent codex-review processes racing for the same model-list cache; verified on bug-hunt run 20260723_090145.
When a Hermes scripted multi-agent review cron (currently bug-hunt-daily.sh, future siblings in the same family) posts a Slack report that looks like a clean sweep (PRs reviewed=0, Bugs found=0, Agent failures=0/0), the dispatcher is staring at the script's failure-shape: every layer of the pipeline (PR discovery, agent preflight, agent execution) silently coerces its failure to "zero work done", and the FAILURE_WARNING block is gated only on agent-execution failures — not on the upstream gates. The 4-bucket classification below is the diagnostic; the recipe at the end is the durable fix.
This is distinct fromwa-daily-cron-failure-diagnosis (which covers GCP Cloud Run cron jobs for your-project.com with email alerts). This skill covers Hermes-side launchd crons with Slack alerts — different channel, different agent fleet, different failure shape. The two skills share Pitfall 5 ("trust the report's zero count") and Pitfall 8 ("re-verify against live state"); the buckets diverge.
User pastes such a report: "is this right?", "this says zero PRs — there should be some", "the daily bug hunt is broken".
Sibling-report signal: a bug-hunt-<TS>.md file from the SAME minute has a ## PR Discovery Failure section, OR a sibling has Agent failures: N/N instead of 0/0. Sibling reports at ls /tmp/hermes/bug_reports/bug-hunt-<same-minute-prefix>* are the smoking gun.
Cron ran but didn't surface what it should have: the launchd plist (launchctl print gui/$(id -u)/ai.hermes.schedule.<job>) shows last exit 0, the log is fresh, but Slack got a zero-result report.
The five failure buckets — classify BEFORE doing anything
Always classify into ONE of these five buckets before recommending a fix. The bucket determines the recipe.
Sibling report file contains ## PR Discovery Failure block with a GraphQL: API rate limit already exceeded for user ID <N> line, OR an HTTP 5xx, OR a could not resolve host error.
The "clean sweep" report's own JSON output files for all agents are ( size = 0, size = 0). Agents were never started because the script took the path on .
0 bytes
bug-hunt-<agent>-<TS>.json
bug-hunt-<agent>-<TS>.err
continue
if [ "$PR_COUNT" -eq 0 ]
gh api user confirms the rate-limited user ID matches the active token (gh api user --jq .id should equal the user ID in the rate-limit message).
Root cause:get_merged_prs "$REPO" 2>/dev/null || echo "[]" silently coerces a non-zero exit (rate limit, auth failure, network error) to []. The script has no way to distinguish "no PRs merged in the window" from "couldn't even ask".
Recipe:
Identify which token hit the rate limit (gh api user --jq .id vs the user ID in the rate-limit message — should match).
Wait for the rate-limit window to pass (GitHub GraphQL resets hourly; REST has its own counter).
Manually re-run discovery with gh pr list --state merged --search "merged:>=<since>" against each scanned repo.
The "fix now" recipe: replace the silent-coercion line with a typed status return (printf 'ERROR\t%s\t%s\n' "$repo" "$out" >&2; return 1) and stop coercing to []. The FAILURE_WARNING block must surface discovery failures too.
bug-hunt-review-preflight-<TS>.err contains ERROR: installed codex CLI lacks supported review mode or ERROR: installed codex CLI lacks explicit model selection or ERROR: codex CLI not found.
The "clean sweep" report has Agent failures: 0/0 — agents never started because REVIEW_CLI_AVAILABLE=0 took the continue path.
codex review --help 2>&1 | grep -q -- '--base' returns non-zero on this machine.
Root cause:configure_review_cli() returns 1 when the local codex CLI lacks the --base/--model flags the script requires. The script then skips the agent loop entirely, leaving AGENT_PIDS=() empty, so AGENT_FAILURES stays 0 (no agents to fail). The FAILURE_WARNING block only fires on ALL_AGENTS_FAILED (every spawned agent failed), not on "no agents spawned because preflight short-circuited".
Recipe:
Install or update codex: brew upgrade codex (or npm i -g @openai/codex).
Verify the model tier per SOUL.md ## COMMIT: subagent model routing (mandatory, 2026-07-14) — cheapest correct tier for review lanes (mini/haiku-class for pollers/monitors, mid-tier Sonnet/Codex Spark for standard review/evidence lanes).
The "fix now" recipe: change configure_review_cli to return 0 on missing CLI but still mark a typed PREFLIGHT_FAILED status; the loop should set AGENT_FAILURES=${#AGENTS[@]} when preflight fails so the warning fires.
Bucket C: Agent-execution suppressed (spawned but produced 0 bytes / errors)
Symptom markers:
JSON output files exist but are 0 bytes: bug-hunt-<agent>-<TS>.json size=0.
Error files contain codex: command not found, Error: 401, Error: 429, or timeout messages.
The "clean sweep" report has Agent failures: N/N — the warning fires correctly here, but the bug count is still 0 because all agents crashed before producing findings.
Root cause:codex review was spawned but the underlying model invocation failed (auth, quota, network). The script's validate_finding_evidence() correctly rejects empty/malformed output and counts as AGENT_FAILURES. This is the bucket the script handles correctly. The bug is upstream — quota exhaustion on the chosen tier, transient OAuth expiry, or codex binary segfaulting.
Recipe:
Read the .err files: cat /tmp/hermes/bug_reports/bug-hunt-<agent>-<TS>.err.
Identify the failure class (auth → re-login; quota → switch tier per SOUL.md subagent model routing; binary → reinstall).
Retry with a different model tier (BUG_HUNT_REVIEW_MODEL=...) or different agent fleet.
If the underlying cause is quota exhaustion on the cheapest tier, escalate per ao-spawn-minimax-worker (mid-tier fallback).
Bucket D: Fail-closed-not-armed (cron disabled / plist not loaded)
Symptom markers:
launchctl print gui/$(id -u)/ai.hermes.schedule.bug-hunt-daily shows state = not running and last exit code = 0 BUT the log mtime is hours/days old.
bug-hunt-<TS>.md files in /tmp/hermes/bug_reports/ are from older runs; no fresh timestamps.
~/.hermes/launchd/ai.hermes.schedule.bug-hunt-daily.plist exists but Disabled=true (canonical example: ai.hermes.schedule.<job>-watcher.plist pattern).
Root cause: Plist was disabled, never re-enabled after a previous test run, or KeepAlive{SuccessfulExit=false} caused launchd to remove it from the run queue. This is NOT a script bug — it's an infra bug. See dropped-thread-watcher-of-watchers for the canonical silent-cron-death pattern.
Verify next scheduled run lands and produces a real report.
Bucket E: Parallel-CLI-process race (concurrent agents hit shared resource lock)
Symptom markers:
PRs were discovered correctly (Bucket A excluded — TOTAL_PRS > 0 in the log, e.g. Found 9 merged PRs).
Preflight passed (Bucket B excluded — bug-hunt-review-preflight-<TS>.err is empty).
Per-agent output files exist and are 0 bytes: bug-hunt-<agent>-<TS>.json size=0.
bug-hunt-<agent>-<TS>.err is 0 bytes for SOME lanes but NON-empty for OTHERS — the discriminating signal. The non-empty err shows the underlying CLI's banner (OpenAI Codex v0.144.5, model: gpt-5.3-codex-spark, workdir: ...) and then a fatal line like failed to refresh available models: timeout waiting for child process to exit or codex_models_manager::manager: failed to refresh available models: ....
The "clean sweep" report has Agent failures: N/N where N == number of agents — FAILURE_WARNING fires correctly here.
ps aux | grep -E "<cli-binary>" shows multiple instances of the same CLI binary running concurrently with overlapping start times.
Root cause: The script spawns N agents in parallel as background processes from a single bash subshell, all targeting the SAME CLI binary with the SAME model. Each CLI process tries to refresh the local model-list cache on startup; the cache has a single-writer lock and the loser times out. One process wins, the others produce 0-byte output and either swallow the failure to a 0-byte .err (race winner ate the error stream) or write the banner + timeout to .err before exiting.
This was confirmed in the 2026-07-23 09:01 PT bug-hunt run: three concurrent codex review -c "model=\"gpt-5.3-codex-spark\"" - processes from bug-hunt-daily.sh:230-238. bug-hunt-claude-20260723_090145.{json,err} and bug-hunt-codex-20260723_090145.{json,err} were both 0 bytes. bug-hunt-minimax-20260723_090145.json was 0 bytes; bug-hunt-minimax-20260723_090145.err contained the actual Codex banner + codex_models_manager::manager: failed to refresh available models: timeout waiting for child process to exit. All three agents used the SAME gpt-5.3-codex-spark model — the labels claude/codex/minimax were cosmetic, all three were the same Codex CLI invocation.
Recipe:
Detect: count err-files-with-content vs total.for f in /tmp/hermes/bug_reports/bug-hunt-*-<TS>.err; do echo "$f $(stat -f%z "$f") bytes"; done. Mix of 0-byte and non-zero = Bucket E.
Stop spawning the same CLI in parallel. Two durable fixes, pick one:
Switch to per-agent CLIs (the canonical fix in PR #792): hermes -z -m <model> per agent label where each label maps to a DIFFERENT CLI (claude → Claude Code, gemini → Gemini CLI, minimax → minimax Anthropic-API shim). Each CLI has its own model-list cache, no contention.
Serialize the agents: change the bash loop from & backgrounding to wait after each spawn. Loses parallelism but eliminates the race.
Surface the race in the report. Add a per-agent stderr-capture guard that copies non-empty .err content into the ## Results section BEFORE the FAILURE_WARNING block, so a single-line codex_models_manager::manager: failed to refresh available models: timeout waiting for child process to exit shows up in Slack instead of being swallowed.
Add a regression test that monkey-patches the agent spawn to return the timeout error string and asserts the report's ## Results section contains the literal failed to refresh available models.
Related skill for the canonical fix: the durable fix lives in jleechanorg/jleechanclaw PR #792 (fix/durable-bug-hunt-harness, MERGEABLE, CodeRabbit APPROVED, blocked only by Gate 3 stale state + Gate 5 unresolved comment as of 2026-07-23). Reference: references/2026-07-23-bug-hunt-codex-review-race.md.
The 7-step investigation protocol (always run in order)
Step 1 — Read the report file the bot posted
cat /tmp/hermes/bug_reports/bug-hunt-<TS>.md
Capture: PRs reviewed, Bugs found, Agent failures. If all three are zero, suspect suppression (Buckets A/B/C/D).
Step 2 — List sibling reports from the same minute
ls -la /tmp/hermes/bug_reports/bug-hunt-<same-minute-prefix>*
If you see ≥2 reports from the same minute, the cron ran more than once and one of them likely surfaced the real failure. Sibling reports are the smoking gun — Bucket A/B always leaves a sibling that reports correctly.
for f in /tmp/hermes/bug_reports/bug-hunt-{claude,codex,gemini,minimax}-<TS>.{json,err}; doecho"=== $f ($(stat -f%z "$f" 2>/dev/null) bytes) ==="
[ -s "$f" ] && head -c 500 "$f"echodone
Empty JSON + empty err = Bucket A or D (agents never spawned). Empty JSON + non-empty err = Bucket C (agents spawned, failed). Mixed empty-JSON-and-err across agents = Bucket E (parallel-CLI-process race) — see Bucket E's symptom markers; the lane that captured the banner-then-timeout is the one that lost the model-list cache lock. Preflight .err present = Bucket B.
(Adjust <TS> and mergedAt >= "<since>" to the actual window.) This is the live ground truth; the report is the suspect. Compare counts. Mismatch = Bucket A.
Step 5 — Classify into one of the four buckets
Apply the markers from each bucket section above. Buckets can compound (e.g. Bucket A on first repo → all subsequent repos loop returns [] → no agents → clean sweep). When in doubt, start with Bucket A — it is the most common and the most silent.
Switch to per-agent CLIs (e.g. hermes -z -m <model> per label) OR serialize the spawn loop; surface the cached timeout in the report
Step 7 — Produce a named end-state
Per diagnosis-requires-followthrough-or-handoff, every diagnosis must end with one of:
"apply now" — fix is ≤10 lines, single-file, reversible (e.g. broaden FAILURE_WARNING gate). Apply inline in same turn. Reply: "Diagnosis complete. Applied the fix inline — <summary + file:line + verification>."
"dispatch now" — multi-file or needs PR (the script fix + pytest). ao spawn with a one-line brief OR bring-to-green babysit cron. Reply: "Diagnosis complete. Dispatched on branch with the recipe."
"hand off explicitly" — fix is destructive/expensive (data migration, schema, billing, prod deploy) OR user has been upset by auto-fixes. Paste exact 3-5 line shell block + offer to dispatch. Reply: "Diagnosis complete. NOT applying because . To finish: ."
Bare "Investigate" / "look at this report" (no /a, /finish, /auto) means diagnose-then-hand-off: do NOT auto-dispatch. Reply with the diagnosis + end-state ask.
Pitfalls
Pitfall 1: Trusting the "clean sweep" report's zero counts
The script's three zero counts (PRs=0, Bugs=0, Agent failures=0/0) are all three independently suppressible by upstream failures. Always run Step 2 (sibling reports) and Step 4 (re-derive ground truth) before believing any of the three. The 2026-07-22 case: zero counts on the main report, sibling at the same minute had ## PR Discovery Failure with the rate-limit message verbatim, and Step 4 found 8 merged PRs the bot should have reviewed.
Pitfall 2: Diagnosing the report's surface format instead of the script's pipeline
The user often asks "is this report formatted right?" — that's the wrong question. The format is fine; the content is suppressed. Pivot to the pipeline: PR discovery → agent preflight → agent execution → report generation. Each stage has a silent-failure path; identify which one is firing in this run.
Pitfall 3: Skipping the launchd state check (Bucket D) on first encounter
If bug-hunt-<TS>.md files are stale (no fresh timestamps in the last 24h), check the launchd plist state FIRST. No amount of script debugging helps if the cron isn't running. launchctl print gui/$(id -u)/ai.hermes.schedule.bug-hunt-daily 2>&1 | grep -E 'state|last exit' is the 5-second check.
Pitfall 4: Reporting "all three agents failed" when the actual cause was preflight
The "clean sweep" report with Agent failures: 0/0 often masks a preflight failure (Bucket B). The script's FAILURE_WARNING gate only fires on AGENT_FAILURES == ${#AGENTS[@]} (every spawned agent failed), not on REVIEW_CLI_AVAILABLE == 0 (no agents spawned because preflight short-circuited). Read the preflight .err file before declaring "agents failed".
Pitfall 5: Trusting prior-session conclusions about cron health
EA sweeps (e.g. memory/briefings/YYYY-MM-DD/<time>-ea-sweep.md) often cite "bug-hunt cron is healthy" or "no PRs to review" from brief-time telemetry. Re-derive live per Step 4. The same "session-of-record drift" bug that wa-daily-cron-failure-diagnosis Pitfall 10 catches for daily-cron telemetry applies here.
Pitfall 6: Reading the bot's Slack-token vs the user's GitHub-token
The Slack post lands under the bot's identity (XOX-B). The GitHub gh pr list rate-limit is on the user's GitHub token (XOX-P for the user, or GITHUB_TOKEN for the bot identity). When triaging rate-limit errors, the user ID in the message is the GitHub user (gh api user --jq .id), not the Slack user. They're different credentials.
Pitfall 7: Posting the diagnostic to the wrong Slack thread
Per SOUL.md slack-reply-inherit-thread-ts + slack-channel-routing-policy, the diagnostic reply belongs in the originating thread (the bug-hunt report thread) — NOT a home-channel orphan. The report itself is the durable artifact; if you reply home-channel instead of in-thread, the user will treat the home-channel post as "your reply" and re-ask in-thread. Always verify ThreadTs == <correct_ts> on the verification call (conversations.replies).
Pitfall 8: Auto-dispatching on bare "is this right?"
no-confirmation-gate: bare "is this right?" / "look at this report" / "the bug hunt is broken" are NOT dispatch commands. Only /a, /fullrun, /finish, /auto, /f, /fin bypass the confirmation gate. If the user's message is bare, diagnose and name the end-state — let the user invoke the dispatch command. Posting "Should I spawn an AO worker? Y/N" is a SOUL.md violation.
Pitfall 9: Skipping the "named end-state" — finishing with "want me to fix?"
Every diagnostic reply must end with ONE of: "apply now" / "dispatch now" / "hand off" named explicitly. A reply ending with "Want me to ...?" or "Should I ...?" is a SOUL.md no-pick-one-menus violation. The user explicitly opted into autonomous execution by building this skill family, but a structured diagnosis with a named handoff lets them pick precisely.
Pitfall 11: Treating a mixed empty/non-empty .err pattern as Bucket C only
The Bucket C markers say "empty JSON + non-empty err = Bucket C". When ALL lanes show that pattern, it's Bucket C. But when the pattern is mixed (some lanes 0-byte err, some non-empty err, ALL lanes 0-byte JSON), it's Bucket E — a parallel-CLI-process race where the cache-lock winner ate its error stream and the loser(s) got the timeout. The single signal that disambiguates Bucket C from Bucket E is which CLI is in the .err file. If every non-empty .err file shows the SAME CLI banner and the SAME model — and the script's TASK_PROMPT was sent to N distinct labels that all resolved to the same CLI/model — it's Bucket E, not C. The "fix CLI" recipe in Bucket C won't help; the fix is "stop running this CLI N times in parallel" (Bucket E recipe).
A subtler pitfall inside this one: the cosmetic agent labels (claude/codex/minimax) often DO NOT correspond to distinct CLI binaries. When the script dispatches via a single codex review -c "model=\"…\"" for all three labels, the labels are cosmetic and all three agents are the same Codex CLI. The bug-hunt run 20260723_090145 had labels claude/codex/minimax and a single banner OpenAI Codex v0.144.5 / model: gpt-5.3-codex-spark in the surviving .err. Diagnosing this as "claude and codex failed but minimax worked" is wrong — minimax didn't work either; it just produced a non-empty .err before the timeout.
Pitfall 10: Including a recipe without proof
Per SOUL.md proof-before-claim, the diagnostic reply must include raw terminal output proving the classification — sibling report contents, .err file lines, gh api output for the re-derivation. A "Bucket A" classification with no cat output of the sibling report's ## PR Discovery Failure section is unprovable and is a SOUL.md fabricated proof violation.
Reference and cross-links
references/2026-07-22-bug-hunt-rate-limit-suppression.md — session-specific evidence: the three sibling reports from 16:40:07, 16:40:09, 16:40:11 PT; the GraphQL rate-limit message verbatim; the script's lines 112-114, 137, 405-416 trace; the 8 PRs the bot failed to review; the durable-fix recipe for the script + pytest.
references/2026-07-23-bug-hunt-codex-review-race.md — session-specific evidence for Bucket E: three parallel codex review processes racing for the same model-list cache; one lane captured the banner + codex_models_manager::manager: failed to refresh available models: timeout waiting for child process to exit, the other two produced 0-byte .err; the cosmetic agent labels all resolved to the same CLI; in-flight durable fix is jleechanorg/jleechanclaw#792 (fix/durable-bug-hunt-harness).
~/.hermes/skills/worldarchitect/wa-daily-cron-failure-diagnosis/SKILL.md — sibling skill for GCP-cron failures (4-bucket classification pattern shared, but different cron family).
~/.hermes/skills/devops/gh-rate-limit-and-transient-failures/SKILL.md — the GraphQL rate-limit-and-recovery playbook that Bucket A relies on.
~/.hermes/skills/dropped-messages/SKILL.md — Bucket D detection (silent launchd death).
~/.hermes/skills/hermes-imports/dispatch-task/SKILL.md — AO worker dispatch mechanics for the "dispatch now" end-state.
~/.claude/skills/drive-pr-to-green/SKILL.md — for driving the resulting PR through to merge.
~/.claude/skills/advice/SKILL.md — for /advice second-opinion on the script fix.
~/.cursor/rules/pr-hyperlink.mdc — for the PR-hyperlink rule when reporting the missed PR list to the user.
~/.hermes/scripts/bug-hunt-daily.sh — the script under diagnosis; lines 112-114 (silent coercion), 137 (|| echo "[]"), 405-416 (FAILURE_WARNING gate).
One-line summary
Treat every "clean sweep" report as suspect until proven otherwise — sibling reports in the same minute and live gh api re-derivation are the two smoking guns. Classify into A/B/C/D/E (E = parallel-CLI-process race: mixed empty/non-empty .err across lanes, all targeting the same CLI/model), fix the right surface, name the end-state explicitly.
