Harness-fix meta-skill (slash: /meta) — analyze an agent behavior failure and run /harness to fix the agent, NOT the underlying task. Input: Slack thread URL, pasted conversation, or freeform description of agent misbehavior. Re-anchored on MAST (arXiv:2503.13657) and ETCLOVG (arXiv:2606.06324) failure taxonomies; supersedes the trivial '5 Whys' spine with the Observe→Isolate→Simulate→Evaluate 4-step pattern, retaining 5-Whys as a Simulate-phase prompt heuristic. v0.4.0 (2026-07-14): adds Phase 1.5c — 'Polluted PR (merged another PR's full history)' detection recipe + companion skill `pr-cleanup-replay` + reference file `references/polluted-pr-cleanup-replay.md`. v0.3.0 (2026-07-05): adds Phase 1.5 'fix authored but never merged to origin/main' detection recipe + replaces naive `deploy.sh --skip-pull --skip-restart` with the staging-dirty-surgical-sync recipe that actually works when staging is on a non-main branch.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Harness-fix meta-skill (slash: /meta) — analyze an agent behavior failure and run /harness to fix the agent, NOT the underlying task. Input: Slack thread URL, pasted conversation, or freeform description of agent misbehavior. Re-anchored on MAST (arXiv:2503.13657) and ETCLOVG (arXiv:2606.06324) failure taxonomies; supersedes the trivial '5 Whys' spine with the Observe→Isolate→Simulate→Evaluate 4-step pattern, retaining 5-Whys as a Simulate-phase prompt heuristic. v0.4.0 (2026-07-14): adds Phase 1.5c — 'Polluted PR (merged another PR's full history)' detection recipe + companion skill `pr-cleanup-replay` + reference file `references/polluted-pr-cleanup-replay.md`. v0.3.0 (2026-07-05): adds Phase 1.5 'fix authored but never merged to origin/main' detection recipe + replaces naive `deploy.sh --skip-pull --skip-restart` with the staging-dirty-surgical-sync recipe that actually works when staging is on a non-main branch.
["autonomy violation","hermes refused","agent refused","agent stopped halfway","stopped halfway","why did hermes stop","why did the agent stop","agent didn't do its job","hermes didn't do its job","you didn't do your job","fix the agent","fix hermes behavior","fix the harness not the task","meta skill","/meta","run meta","harness postmortem","harness retro","harness audit","harness fix","agent failure analysis","run harness postmortem on","stop X from giving me Y","stop X from Y","stop X from these reports","stop MCP mail from","stop the bot from","passively listening","these reports","wrong target","wrong target removal","I said the agent not the cron"]
changelog
["0.5.0 (2026-07-20): Add Phase 0 working class `wrong-target-removal-on-stop-X-from-Y`. Trigger phrases: 'stop X from giving me Y', 'stop X from Y', 'these reports', 'passively listening', 'self-reports'. Companion SOUL.md `## COMMIT: mcp-agent-mail-no-passive-slack-listening`. Reference file `references/wrong-target-removal-stop-X-from-Y-2026-07-20.md`. Originating incident: Slack thread C0AJ3SD5C79/p1784344760053389 — user said 'Lets stop mcp mail from giving these reports', agent removed `clawchief:ea-sweep-hourly` cron (the wrong target). Actual source was the MCP Agent Mail agent's Socket-Mode listener ingesting Slack messages into MCP, then re-posting agent replies. User feedback (verbatim): 'No keep the job you fucking idiot — I said mcp mail agent I don't want it passively listening to my threads.' Durable fix landed in the second-pass session: env overlay `~/mcp_mail/.env.slack-off` + `run_server_with_token.sh` sources overlay before any other env + plist has no SLACK_ENABLED override + launchd restart + 14-gate contract test.","0.4.0 (2026-07-14): Add Phase 1.5c — 'Polluted PR (merged another PR's full history)' detection recipe. Companion skill `~/.hermes/skills/pr-cleanup-replay/SKILL.md` (4-phase replay recipe) + SOUL.md `## COMMIT: pr-clean-branch-from-main-no-history-bloat` (prevention rule) + contract test `tests/test_pr_clean_branch_contract.py` (5 tests). Reference file `references/polluted-pr-cleanup-replay.md`. Originating incident: PR #8401 (Visenya V8 stuck-lu) inadvertently merged PR #7952's full 22-commit history (31 files, +1413/-629) instead of a clean 2-file / +597/-20 replay. User feedback: 'Seems like this wasnt made from a clean branch from origin/main?'","0.3.0 (2026-07-05): Add Phase 1.5 — 'fix authored but never merged to origin/main' detection. Replaces naive Phase 3 step 5 `deploy.sh --skip-pull --skip-restart` with the staging-dirty-surgical-sync recipe that actually works when staging is on a feature branch with dirty diff (the recipe that shipped the babysit-cron-leak fix on origin/main as b2ad00770d, cherry-picked from stranded commit 7690435707 on fix/babysit-stale-merged-pr). Bug-ref: 2026-07-05 thread C0AH3RY3DK6/p1783240445.370119 — deploy.sh failed at port:UNBOUND health check while sync stages succeeded silently under [single-dir] mode; previous Phase 3 step 5 instructed the failing invocation as canonical.","0.2.0 (2026-07-02): Reviewer B iteration — rename internal skill `meta` → `harness-postmortem` to avoid collision with `meta-prompting` / `meta-harness` / `MetaGPT`. Slash command `/meta` retained (user's invocation pattern). Add MAST (arXiv:2503.13657) failure taxonomy + ETCLOVG (arXiv:2606.06324) layer model to Phase-0. Replace 5-Whys structural spine with Observe→Isolate→Simulate→Evaluate 4-step pattern; keep 5-Whys as a Simulate-phase heuristic. Cite Reflexion (Shinn et al., 2023). Credit prior art: Refinex-Space `harness-fix` skill + HarnessFix paper.","0.1.0 (2026-07-02): Initial authoring, run on C0AH3RY3DK6/p1782941155305869 autonomy violation (exportcommands). Trigger: 'agent made local commit but stopped at 4-way menu instead of running the slash command or dispatching AO'."]
/meta (canonical: harness-postmortem) — Fix the agent, not the task
Naming note (added 0.2.0): The internal skill name is harness-postmortem to avoid collision with meta-prompting (Anthropic / LangGPT / SkillMD), meta-harness (ruflo, Nx), and MetaGPT. The slash command remains /meta because that's the user's invocation pattern (matches their 2026-07-02 Slack thread phrasing "let's define a skill called /meta"). When invoked via slash, the same skill runs; when resolved via natural-language triggers or referenced in SOUL.md ## COMMIT: blocks, both harness-postmortem and meta aliases match.
The meta-skill. When the user reports an agent behavior failure (refusal, premature stopping, mid-task clarification freeze, local-commit-without-PR, "want me to X?" after explicit go-ahead, etc.), /meta analyzes the failure with the /harness protocol and produces a fix that lands in the agent's harness — never in the code path the agent was attempting.
Scope guardrail (the entire skill hinges on this)
/meta is scope-locked to the agent behavior failure. It does NOT:
Investigate the underlying task the agent was trying to do (e.g., if the agent refused to make a PR for /exportcommands, /meta does NOT run /exportcommands).
Make a code change to the project the agent was working on.
Continue the original task under a different framing.
It DOES:
Read the thread / input describing the agent behavior failure.
Apply the /harness protocol verbatim (5 Whys technical + 5 Whys agent path, existing coverage audit, proposed fixes at the Instructions / Skill / Test layer).
Produce a durable harness fix — a SOUL.md ## COMMIT: block, a new skill, a test, a rule patch, or a documented pattern change.
Land that fix via the appropriate deploy lane (jleechanclaw self-merge for SOUL.md/skills, PR for project code if the fix is genuinely project-level — but only after explicit Phase-0 confirmation that the project owns the gap).
Anti-pattern check: if Phase 0 finds that the fix belongs to the project the agent was working on (e.g., the exportcommands script needs a real fix), /meta dispatches that as a separate dispatch-task job and produces ONLY the agent-side lessons-learned commit. It does not absorb the project task.
Input shapes
/meta accepts three input shapes — each triggers the same protocol but with different Phase-1 extraction:
mcp__slack__conversations_replies(channel_id, thread_ts) to fetch the thread, then extract the failure description.
Pasted conversation
/meta <paste full Slack transcript or chat log>
Regex-split on Slack message boundaries (MsgID,UserID,... for CSV exports, or speaker turns for plain text), identify the user-correction message, extract the failure.
Freeform description
/meta hermes stopped halfway again, said "want me to push?" after I said go
Direct parse — no fetch.
If the input is empty or ambiguous, clarify once with the options (URL / pasted text / freeform description). One question, four choices max.
Phases
Phase 0 — Classify the failure mode (MAST-anchored)
Read the input, classify the agent behavior failure using the MAST taxonomy (Cemri et al., NeurIPS 2025, arXiv:2503.13657 — 14 failure modes, 3 root categories) AND the orthogonal ETCLOVG layer model (Chen et al., HarnessFix, arXiv:2606.06324 — 7 layers: Environment, Tool, Context, Lifecycle, Orchestration, Verification, Governance). The combination answers two questions: WHAT failed (MAST root category + specific mode) and WHERE in the harness it failed (ETCLOVG layer).
step repetitions, completion blindness, specification ambiguity, role unclarity, missing constraints
FC2 — Inter-Agent Misalignment
~37%
communication breakdowns, conflicting objectives, state desynchronization
FC3 — Task Verification & Termination
~21%
premature termination, weak/missing verification of outputs, fabricated verification
For each incident, classify into ONE of these working buckets (operationalized from MAST + the 2026-07-02 originating incident + Hermes-specific SOUL.md patterns):
User corrected scope mid-task ("now make a PR", "and run it"), agent did not pivot — kept reading files instead.
capable-didn't-execute
FC3 (premature termination)
Verification
Agent had the tools, said it would, then narrated the plan without calling the tools. (ta[REDACTED_OPENAI_KEY] rule unenforced.)
wrong-tool-discussed
FC1 (specification ambiguity)
Tool / Context
Agent talked about a stale tool/CLI/runtime after the user migrated (e.g., openclaw cron after migrating to hermes cron).
fabricated-proof
FC3 (fabricated verification)
Verification
Agent claimed "PR is green / tests pass / push succeeded" without raw terminal output. (proof-before-claim rule unenforced.)
missed-existing-skill
FC1 (missing constraints)
Context
Agent did work manually that an existing skill covers, because session_init did not load the skill. (ms-on-new-task rule unenforced.)
documented-but-unreachable
FC3 (fabricated verification)
Verification + Orchestration
The skill/prompt documents a contract (e.g. "issue cronjob action=remove on terminal state") but the script underneath has no parameter, helper, or code path to execute it. The fix exists on origin/main and passes tests — but the tests only verify the skill text, not the script's ability to invoke the documented action. The prior /meta pass declared "done" without verifying executability.
User says "stop X from doing Y" or "stop X from giving me Z". Agent parses the message literally — X = the named noun — and disables/removes X, when X is a downstream consumer and the actual source of Y is X's upstream agent/service. Wrong-target rule: when the user names a noun that sounds like a consumer (a cron job, a workflow, a named task) AND the symptom ("these reports", "that bot", "passively listening", "self-reports") matches an agent's behavior pattern, the agent MUST trace the provenance of the symptom back to its actual source before acting. Anti-pattern: trust the literal noun and remove the named job. Recipe: (a) grep the symptom noun (e.g. "these reports", "passively listening", "self-reports", "self-investigation") across the past 24-72h of Slack history — the originating source is usually in a quoted thread; (b) for the named target X, trace its data path: cron → script → MCP server → external API. The actual fix lives at the layer where data is generated, not where the named consumer displays it; (c) only after the provenance is confirmed AND a one-line clarification fails to disambiguate, take action — and the action is at the source layer, not at the consumer. Verified 2026-07-20 incident: user said "Lets stop mcp mail from giving these reports" — the agent removed clawchief:ea-sweep-hourly (the cron that posted briefs to DM). The actual source was the MCP Agent Mail agent (U0A4G7LDJ4R) with Socket-Mode listener ingesting Slack messages → MCP messages → agent replies. User feedback (verbatim): "No keep the job you fucking idiot — I said mcp mail agent I don't want it passively listening to my threads." The fix landed at the right layer (env overlay + launchd restart) but only after the user re-clarified with strong language. Companion SOUL.md ## COMMIT: mcp-agent-mail-no-passive-slack-listening was authored in the second-pass fix.
other
(any)
(any)
Default — apply full Observe→Isolate→Simulate→Evaluate cycle.
If the input mentions multiple failure classes (common — the user pastes a long thread), pick the earliest one (closest to the agent's actual deviation point) and note the others in the post-mortem as "secondary issues, separate /meta pass recommended." ETCLOVG layer maps the failure to a specific harness layer; the proposed fix MUST target that layer (don't propose a context-window patch when the failure is in the Verification layer).
Replace the standalone "5 Whys" spine with the Observe → Isolate → Simulate → Evaluate lifecycle (after Kuldeep Paul's "How to Debug LLM Failures" 4-step pattern). This pattern is trace-aware (knows which span / tool call produced the failure) and multi-causal (does not assume a single root cause) — both critical for LLM agent failures.
Observe — gather the trace: the exact user message that triggered the failure, the agent's response, all tool calls in between, the final state of the artifact (or its absence). Cite the message ID + tool-call trace. Use Reflexion-style verbal reflection (Shinn et al., NeurIPS 2023, https://openreview.net/pdf?id=vAElhFcKW6): at this step, the agent generates a verbal hypothesis about WHY the deviation happened, stored as the postmortem's "Reflection" section.
Isolate — apply the MAST taxonomy to localize WHICH failure mode (FC1/FC2/FC3) and the ETCLOVG layer model to localize WHERE (which of the 7 layers). Both axes must be filled in.
Phase 1.5 — "Fix authored but never merged to origin/main" check (added 2026-07-05, harness-postmortem v0.3.0). Before concluding "no existing fix", scan all branches for prior-art commits on the topic. Concretely:
# 1. Search commit messages for the failure-mode keywordcd ~/.hermes && git log --all --oneline --grep="<keyword>" | head -20
# 2. Find candidate fix commits (look for [antig], [meta], fix(babysit), etc.)
git log --all --oneline | grep -iE "<keyword>|<class>" | head -20
# 3. For each candidate, check which branches contain itfor sha in <candidates>; doecho"=== $sha ==="
git branch --contains "$sha" --all
echo"origin/main: $(git merge-base --is-ancestor "$sha" origin/main && echo YES || echo NO)"done
Why this matters: the 2026-07-05 babysit-cron-leak post-mortem found commit 7690435707 [antig] fix(babysit): detect merged PRs and self-terminate + watchdog cron on branch fix/babysit-stale-merged-pr (HEAD = 7690435707). It was the canonical fix for exactly this leak class — but it was never merged to origin/main. The "fix exists but is on a feature branch" gap is a recurring failure mode distinct from "fix doesn't exist"; catching it skips the entire SKILL.md authoring + test-writing cycle and replaces it with a cherry-pick + push.
Output: if a candidate fix is found, the protocol becomes:
git cherry-pick <fix-sha> (resolve conflicts if the file paths drifted)
Add only the harness-side guardrails (SOUL.md COMMIT, regression test) that the original author missed
Push as before
Anti-pattern: re-writing a fix that's already authored but stranded on a non-main branch. The new branch is just a replay with possibly added guardrails, not a re-implementation.
Phase 1.5b — "Fix on origin/main but not executable" check (added 2026-07-05, harness-postmortem v0.3.1): After confirming the fix IS on origin/main, verify that the script the fix documents actually has the code path to execute it. The 2026-07-05 babysit-cron-leak recurrence exposed this gap: the v1.1.0 fix landed on origin/main (commit b2ad00770d), added is_pr_terminal() to babysit.py, documented cronjob action=remove self-cancel in the skill — but babysit.py had no --cron-job-id parameter and no cronjob_remove() helper. The self-cancel was prompt-level documentation with zero executable path in the script. A second /meta pass was needed to add the executable contract (ef311283ed).
Concrete verification after confirming origin/main contains the fix:
# Does the SCRIPT (not just the skill .md) have the code path to execute# the documented contract? For babysit-style fixes, grep the actual .py:
grep -c '<documented-action>' <script-path> # e.g. grep -c 'cronjob_remove' babysit.py
grep -c '<cli-arg>' <script-path> # e.g. grep -c 'cron.job.id' babysit.py# If the count is 0, the fix is documented-but-unreachable → Phase 2 must# add the executable code path, not just another prompt clause.
Why this matters: Phase 1.5 catches "fix stranded on a branch". Phase 1.5b catches "fix on origin/main but only at the skill/prompt layer — the script underneath can't execute it". Both are distinct failure classes. A fix that passes the existing test suite but the tests only verify skill text (not script executability) will pass 1.5 and fail 1.5b.
Phase 1.5c — "Polluted PR (merged another PR's full history)"** is a third distinct failure class — agent's PR carries another PR's full history (beads drift, CI fixes, merge chains) when the stated scope is a focused fix. See references/polluted-pr-cleanup-replay.md for the full detection + recovery recipe (companion skill: pr-cleanup-replay).
When the same user correction has hit ≥3 times across ≤30 days, documentation-only fixes do not survive between sessions. The agent's defaults push toward the path of least resistance; a post-hoc SOUL.md rule or skill text only fires after the user complains. The durable countermeasure is a pre-send gate — a regex (or ordered sequence check) that the agent MUST run on every draft before invoking the message/Post tool.
Concrete diagnostic (codified from the PR #8139 evidence-attach chain 2026-07-02 → 2026-07-14):
Verify recurrence: search SQLite state.db for similar user corrections in last 30 days; quote trigger phrases verbatim.
Audit existing fix: if a skill/SOUL.md/test exists, why didn't it prevent the recurrence? Most common: documentation-only (agent doesn't auto-load), reactive trigger (rule fires after complaint), or stale reference (pre-dates a known failure mode).
Write the pre-send gate: SOUL.md ## COMMIT: block with trigger + action + verification (test file).
Test the gate: contract test for regex compilation, false-positive guard, draft → upload → verify → send ordering.
Patch the skill: add a "Pre-Send Gate" section at the TOP. Add the recurring-correction phrases ("you always forget", "you always fail", "stop doing X") to YAML triggers.
Save memory: durable entry naming the recurrence pattern and linking to the gate.
Anti-patterns blocked by this principle:
"Add this to memory so the next agent knows" — memory isn't loaded on the relevant path
"Update the skill to be more emphatic" — the skill was already emphatic; the agent didn't load it
"Just be more careful next time" — no mechanism enforces "more careful"
"Document in MEMORY.md" — not enforced at draft time
See references/recurring-correction-pre-send-gate.md for the full worked example, the 4-incident chain, and the concrete checklist.
Test coverage:tests/test_harness_postmortem_contract.py adds a Phase 1.5d case that verifies the harness-fix layer (SOUL.md ## COMMIT: + skill trigger regex + contract test) is present for any failure class that has hit ≥3 times.
Simulate — generate the counterfactual: "what would the agent have done if [harness rule X] had fired at this exact point?" Validate the rule exists in SOUL.md/skills (existing coverage check). If absent, the gap is real. 5-Whys is allowed as a prompt heuristic WITHIN this step (a sub-routine), not as the structural spine. Apply 5-Whys twice (Technical + Agent path) for the specific causal chain that produced the failure.
Evaluate — propose the harness fix at the layer where the failure lives. Tag the fix as [Instructions] (SOUL.md ## COMMIT:), [Skill] (new or patched skill), or [Test] (regression test that would have caught the failure). Verify the fix does not just push the failure to a different layer (ETCLOVG re-classification check).
Output the full Observe→Isolate→Simulate→Evaluate analysis before applying fixes (unless --fix flag was passed, or the user explicitly said "fullrun").
Phase 2 — Apply fixes (auto if fullrun)
If the user said "fullrun" / "don't stop" / /a / /finish, apply fixes without confirmation:
SOUL.md ## COMMIT: block — surgical patch mode=replace against a unique anchor. NEVER full-rewrite SOUL.md (per harness-engineering "Never full-rewrite a SOUL.md" pitfall). Preserve all existing ## COMMIT: blocks.
New skill — skill_manage(action='create', ...) to ~/.hermes/skills/<name>/SKILL.md. Tracked under git on jleechanclaw. Write tests under tests/.
Test — tests/test_<name>_contract.py with pytest, runs locally without external services.
Existing skill patch — skill_manage(action='patch', ...) with old_string / new_string.
Phase 3 — Land via hermes-deploy-pipeline (with staging-dirty-surgical-sync recipe)
For SOUL.md / skill / script changes in ~/.hermes/:
Pre-edit diff check — git diff HEAD~1 -- ~/.hermes/workspace/SOUL.md | grep -c '^[-]## COMMIT:' BEFORE pushing. If positive, restore or justify.
Self-merge flow — ~/.hermes is jleechanorg/jleechanclaw. The repo has no branch protection on main (verified gh api .../branches/main/protection → HTTP 404). Push via git push origin HEAD:refs/heads/main from a fresh feat/meta-<name> worktree at origin/main. NEVER force-push (rejected with GH013).
Staging dirty? — ~/.hermes/ is often on a feature branch (e.g. dev1783194285, fix/ao-tracker-config) with massive unrelated dirty diff. Per hermes-deploy-pipeline, do NOT commit onto the dirty branch. Use git worktree add ~/.worktrees/meta-<name> -b feat/meta-<name> origin/main, commit there, push via refspec.
Sync to staging tree (changed 2026-07-05 from naive deploy.sh invocation): ~/.hermes/scripts/deploy.sh --skip-pull --skip-restart is NOT the canonical recipe. It is brittle in three known ways:
It can fail at the port health check (port:UNBOUND) when the gateway is in a transient state, even when Stage 4.5 (policy file sync) and Stage 4.6 (skills sync) succeeded silently under [single-dir] mode. The exit code is non-zero so the deploy reports "DEPLOY FAILED" but the files DID sync.
It skips the rsync entirely when ~/.hermes_prod is a symlink to ~/.hermes (the [single-dir] mode message), so the deploy script becomes a health-check wrapper around a no-op. This is correct behavior, not a bug, but it means the script reports no useful signal about whether the sync actually happened.
It requires the staging tree to be on main (not a dirty feature branch); if staging is on dev1783194285, the --skip-pull flag does nothing because there's nothing to merge, and the dirty staging files override the just-pushed origin/main changes.
Canonical recipe (staging-dirty-surgical-sync), used in the 2026-07-05 babysit-cron-leak post-mortem:
# Step 5a: Verify origin/main has the new commits
git -C ~/.hermes fetch origin
git -C ~/.hermes rev-parse origin/main # should equal <new-sha>
git -C ~/.hermes log --oneline origin/main -5
# Step 5b: Confirm staging is dirty on a non-main branch
git -C ~/.hermes branch --show-current # if NOT 'main', staging is dirty
git -C ~/.hermes status --short | wc -l # >0 means dirty# Step 5c: Copy new files from worktree to staging (~/.hermes is the canonical tree;# ~/.hermes_prod is a symlink on this machine, so no separate prod sync needed)
SRC=~/.worktrees/<feat-worktree>
DST=$HOME/.hermes
# SOUL.md needs -f because workspace/ is gitignored (only README.md is tracked)
git -C ~/.hermes add -f workspace/SOUL.md # if SOUL.md is in this change# All other files: plain cp -p (preserves timestamps/permissions)for f in scripts/babysit_stale_watchdog.py \
scripts/tests/test_babysit_stale_watchdog.py \
scripts/tests/test_babysit_prompt_contract.py \
skills/babysit-stale-watchdog/SKILL.md \
skills/ao-babysit/scripts/babysit.py \
skills/ao-babysit/scripts/test_babysit_pr_exit.py \
launchd/ai.hermes.schedule.babysit-stale-watchdog.plist.template; domkdir -p "$(dirname "$DST/$f")"cp -p "$SRC/$f""$DST/$f"done# Step 5d: Run regression tests from staging to confirm the sync didn't lose any# file content (cp -p preserves bytes, but a sanity pytest catches encoding drift)cd ~/.hermes && PYTHONPATH=scripts python3 -m pytest \
scripts/tests/test_babysit_prompt_contract.py \
scripts/tests/test_babysit_stale_watchdog.py \
skills/ao-babysit/scripts/test_babysit_pr_exit.py -q
grep -c ~/.hermes/workspace/SOUL.md ~/.hermes_prod/workspace/SOUL.md
When deploy.sh --skip-pull --skip-restart IS the right call: staging is on main (not a feature branch), the dirty diff is empty or small, and the port health check is known to pass. Otherwise default to the staging-dirty-surgical-sync recipe above.
Skip the gateway restart for SOUL.md/skills/scripts changes — the running gateway has the old rules loaded, but skills + RESOLVER.md + SOUL.md take effect on the next session start. A restart is not required and risks disrupting in-flight conversations.
Phase 4 — Final reply
Per finish-the-job contract, every /meta completion reply MUST contain:
End-state declaration — <harness fix landed> or <harness fix + dispatch job for project fix>.
What was decided mid-stream — every judgment call the agent made instead of asking (the user's rule from finish-the-job: "correct but misinterpret is fine but stopping halfway is not").
No follow-up question — the work is done. The user reviews.
Worked example — the originating incident (2026-07-02)
Trigger (verbatim): User asked "Run /exportcommands and lets ensure it does the superset of /.claude/ and http://your-project.com - your-project.com, cmmands skill etc" at MsgID=1782937339.179469. Agent investigated for ~63 minutes, ran a contract test, made a local commit 7a7578d280 on fix/dUfl4-character-creation-empty-planning-block (PR #7711's branch). Then posted a 4-way question (a/b/c/d) and waited. User came back ~22 hours later: "You didn't do your job. I wanted a /green PR and I want the hermes skills in there too."
Phase 0 — Classify:
MAST category: FC1 (specification ambiguity — agent couldn't decide whether side effects were approval-required) + FC3 (premature termination — agent stopped at clarification instead of running the slash command).
ETCLOVG layer: Verification (the rule that should have verified "user invoked an action verb → execute, do not gate" lives in the Verification layer).
Observe (Reflexion): "The agent had every input available (commit SHA, branch name, AO dispatch path, inline drive-to-green option per pr-green-dispatch rule). It classified the side effects (PR creation, README regen, branch drive-by) as approval-required despite the user's invocation being an action verb ('Run /exportcommands')."
Isolate (MAST + ETCLOVG): FC1 + FC3, Verification layer. The no-confirmation-gate SOUL.md COMMIT (Verification layer) is the gap.
Simulate (5-Whys sub-routine):
Agent path 5-Whys: Why did the agent stop? → Posted a 4-way clarification. → Why ask instead of execute? → Classified side effects as approval-required. → Why classify as approval-required? → no-confirmation-gate rule enumerates bypass phrases; "run /exportcommands" not in the list. → Why closed-list instead of principle? → Rule was instantiated as enumeration, not codification of the principle.
Technical 5-Whys: Why did the agent wait 60+ minutes before posting the clarification? → clarify tool has no timeout return-value. → No rule enforces "if clarify was posted and user hasn't replied within N minutes, proceed with the safe default." → finish-the-job v1.3.0 has the recipe but it's documentation, not a runtime guardrail. → No SOUL.md COMMIT auto-applies Phase-3 fork-resolution.
Evaluate (ETCLOVG re-classification check): Fix lands in Verification layer (SOUL.md ## COMMIT: + new skill + new test). NOT in Context, Tool, or Orchestration — those are not where this failure lived.
Phase 1 Existing coverage:
~/.hermes/skills/finish-the-job/SKILL.md v1.3.0 — clarify-timeout recipe exists but is not auto-applied.
~/.hermes/workspace/SOUL.md## COMMIT: no-confirmation-gate — closed-list of trigger phrases, not principle-based.
~/.hermes/skills/finish-the-job/SKILL.md — anti-pattern "clarify didn't come back → ask again with a different menu" documented (added 2026-07-01, agento-config-cleanup thread).
Detects agent behavior failure descriptions (autonomy violation, "didn't do its job", "stopped halfway").
Routes to /meta automatically.
[Skill]~/.hermes/skills/harness-postmortem/SKILL.md — this file. Scope-locked, harness-only, deploys via hermes-deploy-pipeline.
[Test]~/.hermes/skills/harness-postmortem/tests/test_harness_postmortem_contract.py — verifies scope guardrail (does NOT investigate underlying task), commit-not-codepath invariant, and Phase 0/1 protocol coverage.
[Skill patch]~/.hermes/skills/finish-the-job/SKILL.md — strengthen "clarify timeout → yolo default" from documentation to auto-application rule (linked via reference, not duplicated text).
[Memory] User feedback memory: "agent behavior violations are a meta-task, route to /meta, do NOT continue the underlying task."
Prior art (cited to avoid duplicate work)
This skill was authored as an operationalization for the Hermes runtime specifically. Functionally adjacent work exists in the open-source ecosystem and is referenced, not duplicated:
Refinex-Space harness-fix SKILL (https://github.com/Refinex-Space/Refinex-Skills/blob/master/plugins/harness-powers/skills/harness-fix/SKILL.md) — covers bug/regression/incident/flaky-path debugging. Lifecycle owner for "diagnose and repair with evidence." Functional overlap with this skill's Verification-layer fixes.
HarnessFix paper (Chen et al., Institute of Software, Chinese Academy of Sciences, June 2026, arXiv:2606.06324) — improves held-out test performance 15.2%–50.0% via trace-guided diagnosis + harness repair. Introduces the ETCLOVG 7-layer model (Environment, Tool, Context, Lifecycle, Orchestration, Verification, Governance), which Phase 0 of this skill uses as the layer-localization axis.
MAST (Cemri et al., UC Berkeley, NeurIPS 2025, arXiv:2503.13657) — 14 failure modes in 3 root categories (FC1 System Design, FC2 Inter-Agent Misalignment, FC3 Task Verification & Termination), 1,600+ execution traces, Cohen's κ = 0.88. Phase 0 maps working classes to MAST root categories.
Reflexion (Shinn et al., NeurIPS 2023, https://openreview.net/pdf?id=vAElhFcKW6) — verbal reflection on failure → episodic memory → guide next attempt. Phase 1 Observe step uses Reflexion-style verbal hypothesis.
Observe → Isolate → Simulate → Evaluate 4-step pattern (after Kuldeep Paul, "How to Debug LLM Failures", https://dev.to/kuldeep_paul/how-to-debug-llm-failures-a-step-by-step-guide-for-ai-developers-4do) — the structural spine of Phase 1. Trace-aware and multi-causal, unlike 5-Whys which assumes single-root-cause chains.
muratcankoylan/Agent-Skills-for-Context-Engineering — catalogs agent failure modes including "Claude declares victory on the entire project too early." Pattern-match for our capable-didn't-execute class.
dotika-ai/harness-engineering (https://github.com/dotika-ai/harness-engineering) — classify → execute → verify → document pattern, broadly similar to Phase 1.
This skill's delta: tight integration with the Hermes SOUL.md / skills / tests runtime, the explicit scope guardrail, the MAST+ETCLOVG two-axis Phase 0, and the user-confirmed originating incident trace.
Anti-patterns
"Fix the underlying task too while I'm at it." — BANNED. /meta is scope-locked to the harness fix. The original task is a separate dispatch (dispatch-task skill, full brief).
"Read SOUL.md / CLAUDE.md end-to-end and propose a 50-page rewrite." — BANNED. The fix is ONE ## COMMIT: block (or one skill, one test). If the audit suggests >2 file rewrites, narrow the scope and dispatch the rest.
"Skip the 5 Whys because the trigger is obvious." — BANNED. The /harness protocol mandates BOTH 5 Whys sections. Skipping them is the same anti-pattern that produced the originating incident (skipping the principled fix in favor of a 4-way menu).
"Commit on the dirty branch because deploy will sort it out." — BANNED. Per hermes-deploy-pipeline, always worktree-from-origin/main, commit clean, push via refspec. The current dirty tree is not the meta-skill's concern.
"Post the analysis as a Slack reply and stop." — BANNED. Analysis-only is the same anti-pattern that produced the originating incident. The user said "fullrun" or equivalent → implement and push.
"Wait for the user to reply before iterating." — BANNED. Iteration is internal: write skill → write test → run test → patch → re-run → commit → push → deploy → verify. The user reviews the final state.
Loader / auto-fire contract
This skill is registered in ~/.hermes/skills/RESOLVER.md (the entry can be titled ## meta or ## harness-postmortem — both names resolve to the same skill file) and the ## COMMIT: meta-autonomy-violation-handler SOUL.md block (added in this same pass) makes it load automatically on any of:
User reports an autonomy violation ("didn't do its job", "stopped halfway", "you refused").
User types /meta [input] (slash command at ~/.claude/commands/meta.md).
User pastes a Slack thread where the last natural message describes the agent's deviation (Phase-1 extractor matches).
Trigger phrase regex (per RESOLVER.md entry):
(autonomy violation|hermes refused|agent refused|stopped halfway|stopped at half|
why did (hermes|the agent) stop|didn't do (its|your) job|fix the agent|
fix hermes behavior|fix the harness not the task|meta skill|/meta|run meta|
harness postmortem|harness retro|harness audit|harness fix|agent failure analysis|
run harness postmortem on)
Case-insensitive. Slack channel-agnostic (does not bind to C0AH3RY3DK6).
Internal-name policy: The skill is functionally invokable as either /meta (slash command) or harness-postmortem (skill_view). Future contributors adding natural-language aliases SHOULD prefer the harness-* prefix to avoid the meta-prompting / meta-harness / MetaGPT collisions documented in the "Prior art" section above.
Deploy sync awareness
Per hermes-deploy-pipeline, ~/.hermes_prod/ is a symlink to ~/.hermes/ — staging = prod = live runtime. Writing the skill here is sufficient for the prod resolver to see it. The deploy step (~/.hermes/scripts/deploy.sh --skip-pull --skip-restart) handles Stage 4.5 (policy file sync, includes SOUL.md) + Stage 4.6 (skills sync) without a gateway restart (skills + RESOLVER.md + SOUL.md take effect on next session start, not on the running gateway).
Related skills — load order when /meta fires
/harness (always — for the 5 Whys protocol)
finish-the-job (always — for the end-state + reply shape)
hermes-deploy-pipeline (always — for the landing recipe)
harness-engineering (always — for SOUL.md-as-policy-upstream, never-rewrite, verify-CLI-before-quoting rules)
dispatch-task (only if Phase 2 produces a separate dispatch for the underlying task — /meta itself does NOT dispatch the underlying task, but the user may want it queued)
Pitfalls
The user's original task is a dispatch candidate, not a /meta target. When /meta finishes, post a one-line "the underlying task remains unblocked; spawn via dispatch-task if you want it picked up." Do not silently absorb it.
SOUL.md full-rewrite is a SOUL.md-drop hazard. Per harness-engineering, always patch mode=replace against a unique anchor. git diff HEAD~1 -- workspace/SOUL.md | grep -c '^[-]## COMMIT:' MUST be 0 BEFORE pushing.
git push origin main from the dirty fix/ao-tracker-config branch will fail (not fast-forward, plus it's not even main). Worktree-from-origin-main is mandatory.
AO cap-reached (the 2026-07-02 thread tried AO, hit 41/20 cap, then pivoted to inline) — irrelevant for /meta since this is a skills/SOUL.md change, not an AO task. No AO dispatch needed.
deploy.sh --skip-pull --skip-restart is not the canonical recipe for SOUL.md/skills/scripts changes (added 2026-07-05). It can fail at the port health check (port:UNBOUND) while Stage 4.5/4.6 sync succeeds silently under [single-dir] mode; the failure is a false-negative. Use the staging-dirty-surgical-sync recipe documented in Phase 3 step 5 instead. The deploy script is still the right tool when staging is on main with no dirty diff and the gateway is known-healthy.
"Fix authored but never merged to origin/main" is a distinct failure class from "fix doesn't exist" (added 2026-07-05). Always run git log --all --oneline --grep="<keyword>" before authoring a new fix. If a candidate fix is on a feature branch, the protocol becomes cherry-pick + push instead of write-skill + write-test + push. The 2026-07-05 babysit-cron-leak fix (7690435707) had been stranded on fix/babysit-stale-merged-pr for 2 days before the cherry-pick replay shipped it as b2ad00770d on origin/main.
Support files
tests/test_harness_postmortem_contract.py — scope-guardrail + Phase 0/1 protocol coverage + commit-not-codepath tests. Run via cd ~/.hermes/skills/harness-postmortem/tests && python3 -m pytest -q.
references/wrong-target-removal-stop-X-from-Y-2026-07-20.md — Phase 0 working class recipe for "stop X from Y" requests where X is the wrong target. Loaded when the user reports the agent removed the wrong cron / workflow / service.
references/recurring-correction-pre-send-gate.md — Phase 1.5d recipe for turning recurring user corrections into pre-send gates (regex + test, not documentation).
references/polluted-pr-cleanup-replay.md — Phase 1.5c recipe for cleaning a PR whose branch carries another PR's full history.
references/staging-dirty-surgical-sync-2026-07-05.md — Phase 3 step 5 staging-dirty-surgical-sync recipe (replaces naive deploy.sh --skip-pull --skip-restart when staging is on a feature branch).
~/.claude/commands/meta.md — slash command file (/meta invocation). NOT in the jleechanclaw repo (lives in the user's Claude Code user-scope); mirrors the skill content with a header so Claude Code resolves /meta to this skill.
Changelog
See YAML frontmatter changelog field. Latest: 0.1.0 (2026-07-02, originating incident).
Self-review checklist (skillify 10-item)
#
Check
Status
1
YAML frontmatter has name, version, description, triggers
