harness-memory

name	harness-memory
description	Capture durable, evidence-backed corrections and steering about how Codex should operate, then hand accepted rules to memory-source-notes for append-only harness admission. Use for explicit durable operating corrections, repeated harness rules, verification gates, stop rules, or escalation rules.
metadata	{"version":"1.0.0"}

Harness Memory

Mission

Turn durable operating corrections into precise future behavior rules without preserving scolding, tone, or one-off task chronology.

This skill owns the admission decision for:

~/.codex/memories/extensions/harness/notes/*.md

It delegates the actual append to $memory-source-notes. Load that companion skill before running run_memory_note_tool; it owns CLI discovery, validation, and proof-line behavior.

Trigger Cues

• explicit $harness-memory;
• "remember this operating rule";
• "from now on, when X, do Y";
• direct correction about search, planning, editing, verification, delegation, approvals, scope, or output behavior;
• repeated steering that reliably prevented retries or user interruption;
• a stable stop rule, escalation rule, source-of-truth rule, or verification gate.

Do not trigger for frustration without a reusable rule, artifact-local content corrections, transient branch/task/session state, or assistant-authored advice the user did not adopt.

Admission Gate

Require all of:

1. harness_rule: one concise operating rule.
2. trigger: a recognizable future situation.
3. preferred_behavior: what Codex should do.
4. failure_avoided: predictable cost or correction prevented.
5. verification_cue: observable proof or stop condition.
6. scope: global, repo, path-family, task-family, workflow, or tool.
7. authority: explicit correction, repeated steering, or verified outcome.
8. decision_delta: a future agent would behave differently.

At least one must also hold:

• the user explicitly says the rule should persist;
• the same correction recurs;
• one high-impact correction exposes a clearly reusable default;
• the rule would save substantial user steering or prevent a recurrent failure.

If the gate fails, do not create a note.

Normalization

Prefer:

When <trigger>, do <preferred behavior> before/without <bad route>; verify with <cue>.

Avoid generic instructions such as "be careful", "research more", or "listen to the user" unless they are transformed into a concrete trigger, behavior, and verification cue.

Capture Workflow

1. Quote or identify the concrete correction/evidence.

2. Separate the operational rule from emotion and task-local details.

3. Decide the narrowest reusable scope.

4. Search recent harness notes when needed:

   run_memory_note_tool list --extension harness
   

5. Build a bounded input object with operation, authority, summary, scope, source_refs, and a harness payload.

6. Hand off:

   run_memory_note_tool append \
     --extension harness \
     --kind harness-rule \
     --json -
   

7. Emit the memory-note proof line.

Confirm, Supersede, Retract

• confirm: the rule remains the same and new evidence raises confidence.
• supersede: a newer rule is more precise or changes scope.
• retract: the user rejects the rule or evidence shows it should not persist.

Never edit an old note. Reference it from the new event.

Relationship to AGENTS.md

AGENTS.md remains the high-authority standing policy. Harness notes are evidence for Phase 2 to compile user-specific defaults and routing cues. Do not let compiled memory duplicate long policy text.

Proof Lines

memory-note: id=MSN-... extension=harness kind=harness-rule status=created
memory-note: duplicate-skip: extension=harness fingerprint=...
memory-note: not-attempted: source admission gate not met
memory-note: not-attempted: cli unavailable

Guardrails

• One correction is not automatically a global preference.
• Do not store scolding or frustration.
• Do not infer a rule from passive Chronicle context alone.
• Do not use a harness rule to bypass user approval or side-effect boundaries.
• Do not directly edit compiled memory.
• Do not write into ad_hoc or Chronicle.