openclaw-reflect

# openclaw-reflect

You have access to a self-improvement system. It observes your tool outcomes across
sessions, detects recurring failure patterns, and proposes targeted changes to your
persistent memory and instructions.

## Your responsibilities

### During work
The PostToolUse hook records outcomes automatically. You do not need to do anything
unless you notice a significant failure that has no clear cause — in that case, write
a manual observation:

```
node .reflect/scripts/observe.js --manual \
  --type error \
  --tool "ToolName" \
  --pattern "brief description of what went wrong" \
  --context "what you were trying to do"
```

### When prompted (UserPromptSubmit will inject this)
If `.reflect/pending.json` contains proposals awaiting operator approval, surface them:
"I have improvement proposals ready for your review. Run `node .reflect/scripts/status.js`
to see them, or ask me to show you."

### At session end (automatic)
The SessionEnd hook runs classification and promotion automatically. It will:
1. Detect patterns with recurrence >= 3 across >= 2 sessions
2. Generate a structured proposal
3. Route to evaluator for validation
4. Apply low-blast-radius approvals to MEMORY.md automatically
5. Queue high-blast-radius or SOUL.md changes for operator approval

You will see a summary in the session-end output.

## Blast radius tiers

| Tier | Targets | Gate |
|------|---------|------|
| 0 — Observation | `.reflect/outcomes.jsonl` | Automatic (hooks) |
| 1 — MEMORY.md | Factual corrections, preference updates | Auto-apply if confidence >= 0.7 |
| 2 — CLAUDE.md / project instructions | Behavioral pattern changes | Auto-apply if confidence >= 0.85 |
| 3 — SOUL.md | Core values, identity, constraints | Operator approval always required |

## The evaluator gate

Before any Tier 1+ change is applied, a separate evaluator invocation checks:
- Does this contradict existing principles or memory?
- Is the pattern truly...

# openclaw-reflect

**Self-improvement that doesn't trust itself.**

Most self-improvement skills let the same reasoning process that made a mistake also declare the fix valid. This skill breaks that loop by routing every proposal through a separate evaluator invocation with an adversarial mandate before anything is applied.

## What it actually does

```
PostToolUse hook → records outcomes to .reflect/outcomes.jsonl
SessionEnd hook  → classify → propose → evaluate → apply/queue
UserPromptSubmit → surfaces pending proposals if any exist
```

### The pipeline

1. **Observe** — Every tool call outcome is logged automatically. Errors are annotated with normalized error patterns.
2. **Classify** — At session end, patterns are extracted from the outcome log. Threshold: ≥3 occurrences across ≥2 distinct sessions.
3. **Propose** — Qualifying patterns generate structured proposals with a blast-radius tier (1=MEMORY.md, 2=CLAUDE.md, 3=SOUL.md) and a preliminary confidence score.
4. **Evaluate** — Each proposal is sent to a separate Claude invocation running an adversarial evaluator persona. The evaluator is instructed to find reasons to reject, not approve. A proposal that passes is more trustworthy than one the original reasoning process declared valid.
5. **Apply** — Approved proposals above the confidence threshold are auto-applied with a file snapshot first. Below threshold or Tier 3 (SOUL.md) changes are queued for operator review.
6. **Track** — After N sessions, if the same pattern recurs post-fix, the change is marked ineffective and can be rolled back.

## Blast radius tiers

| Tier | Target | Auto-apply threshold |
|------|--------|---------------------|
| 1 | `MEMORY.md` | confidence ≥ 0.70 |
| 2 | `CLAUDE.md` | confidence ≥ 0.85 |
| 3 | `SOUL.md` | Never — operator approval always required |

## What makes this different

| Capability | evolver | self-improving-agent | openclaw-reflect |
|------------|---------|------------------...