AGENT SKILL

harness-postmortem

Harness-fix meta-skill (slash: /meta) — analyze an agent behavior failure (autonomy violation, refusal, premature stopping) and run /harness to fix the agent, NOT the underlying task. Input: Slack thread URL, pasted conversation, or freeform description.

License 许可
MIT
最近更新
2026-11-12
Stars
★ 6
Forks
0

01 前言 Front Matter

name
harness-postmortem
type
quality
scope
user
description
Harness-fix meta-skill (slash: /meta) — analyze an agent behavior failure (autonomy violation, refusal, premature stopping) and run /harness to fix the agent, NOT the underlying task. Input: Slack thread URL, pasted conversation, or freeform description.
execution_mode
deferred

03 SKILL.md 正文

harness-postmortem (slash: /meta) — Fix the agent, not the task

Canonical location: ~/.hermes/skills/harness-postmortem/SKILL.md (or its ~/.hermes_prod equivalent). This file is the Claude-side mirror pointer. Collision rule: if a workspace contains .claude/commands/meta.md, repo-local content takes precedence over this mirror.

The slash command is /meta because that's the user's invocation pattern. Both meta and harness-postmortem resolve to the same skill.

Inputs

$ARGUMENTS — Slack thread URL, pasted conversation, OR freeform description of agent misbehavior.

Scope guardrail

/meta is scope-locked to the agent behavior failure. It never investigates or continues the underlying task the agent was attempting — it only analyzes and fixes the harness (instructions, skills, tests, rules) that let the bad behavior happen.

Behavior

  1. Detect input shape (URL / paste / description).
  2. Fetch the thread via mcp__slack__conversations_replies if URL.
  3. Apply the scope guardrail (NEVER investigate the underlying task).
  4. Run the Phase 0/1 protocol: MAST + ETCLOVG classification → Observe → Isolate → Simulate → Evaluate.
  5. If user passed --fix or used a fullrun trigger (/a, /fullrun, /finish, "don't stop", "iterate until good"), implement + commit + push.
  6. Otherwise: output the analysis and wait for approval.

Implementation reference

Full protocol at ~/.hermes/skills/harness-postmortem/SKILL.md (canonical — read this for the complete MAST/ETCLOVG taxonomy, phase steps, and anti-patterns). Highlights:

  • Scope guardrail — top of SKILL.md.
  • Phase 0 — MAST taxonomy (Cemri et al., arXiv:2503.13657) + ETCLOVG layer model (arXiv:2606.06324) classification.
  • Phase 1 — Observe → Isolate → Simulate → Evaluate 4-step spine, with 5-Whys inside Simulate as a prompt heuristic.
  • Phase 2/3 apply + land — via the hermes-deploy-pipeline self-merge flow on the repository that tracks your Hermes harness.
  • Anti-patterns — end of SKILL.md.
  • Prior art — Refinex-Space harness-fix, HarnessFix paper, MAST, Reflexion, Observe→Isolate→Simulate→Evaluate pattern.

Examples

  • /meta — auto-detect from last conversation; if none, prompt for input.
  • /meta hermes stopped halfway again — freeform.
  • /meta https://<workspace>.slack.com/archives/<channel-id>/p<ts> — Slack thread.
  • /meta --fix <description> — analyze AND apply harness fix without confirmation.

Output

Per finish-the-job Phase 4 — end-state declaration + proof artifact + judgment calls + zero follow-up question.

Cross-references

  • /harness — ~/.claude/commands/harness.md (the protocol this skill wraps).
  • harness-postmortem skill (canonical) — ~/.hermes/skills/harness-postmortem/SKILL.md.
  • harness-engineering — ~/.hermes/skills/harness-engineering/SKILL.md (SOUL.md / TOOLS.md rules).
  • finish-the-job — ~/.hermes/skills/finish-the-job/SKILL.md (end-state contract).
  • hermes-deploy-pipeline — ~/.hermes/skills/hermes-deploy-pipeline/SKILL.md (landing recipe).

04 同仓库更多 Skills