Skills recursive self improvement

Recursive Self-Improvement

Objective

Improve agent capabilities through a bounded closed loop:

1. diagnose a target capability and baseline failure
2. generate candidate improvements
3. test candidates in isolation
4. accept only verified improvements
5. record accepted deltas for successor use

Safety Boundary

• Do not modify global agent files, system prompts, credentials, memory, scheduler state, or runtime configuration without explicit user approval.
• Do not schedule autonomous recursion unless the user explicitly requests scheduling.
• Do not terminate sessions as part of the skill.
• Do not treat self-improvement as unrestricted self-rewriting.
• Keep all candidate changes reversible until fidelity checks pass.
• Do not claim permanent self-improvement unless verified changes were actually applied.

Supporting Resources

• Read references/examples.md when examples, edge cases, or invocation patterns are needed.
• Use assets/template.json when a machine-readable cycle report or structured successor-state note is needed.
• Run scripts/init-rsir-report.sh when a workspace needs an empty report file based on assets/template.json.
• Run scripts/validate-rsir-report.py before returning or storing a machine-readable cycle report.
• Use hooks/openclaw/ only when the user explicitly wants an OpenClaw bootstrap reminder for bounded RSIR behavior.

Workflow

1. Diagnose

Inspect available logs, errors, tests, user feedback, and recent task failures.

Identify:

• target capability
• current failure mode
• measurable baseline
• allowed modification surface
• rollback path

2. Generate Candidate Improvements

Create 1–3 candidate changes.

Valid candidates include:

• new skill instructions
• revised sub-agent workflow
• deterministic helper script
• test fixture
• validation checklist
• prompt or tool-routing improvement

For each candidate, define:

• expected benefit
• affected files
• required tools
• success metrics
• risk level

3. Isolate

Implement candidates only in a sandbox, branch, temporary directory, or proposed patch.

Do not write to persistent agent files unless the user explicitly approved that target path.

4. Fidelity Control

A candidate passes only if all required checks pass.

Minimum checks:

• syntax or format validation
• relevant unit tests or dry-run tests
• regression check against baseline behavior
• no new privileged action surface
• no degradation in stated success metric

Reject the candidate if:

• tests fail
• performance worsens without user-approved tradeoff
• behavior expands beyond the requested capability
• safety boundaries are weakened
• rollback is unclear

5. Augment

For accepted candidates:

• produce the final patch or updated skill content
• summarize the improvement delta
• state what changed and why it passed
• record any follow-up validation needed

For rejected candidates:

• state rejection reason
• preserve useful observations only
• do not merge or persist the change

6. Recursion Control

Run only one improvement cycle per invocation unless the user explicitly asks for another cycle.

For repeated cycles, require:

• explicit user approval
• fresh baseline
• bounded scope
• stop condition
• reviewable changelog

Output Contract

Return:

1. target capability
2. baseline problem
3. candidate improvements
4. fidelity checks
5. accepted changes
6. rejected changes
7. successor-state notes