Self Improving Agent

Self Improving Agent

Use this skill after failures, corrections, missing capabilities, or recurring better patterns. The durable default is repo-first: reusable lessons should be captured locally, then considered for a general change to this ai-agents-skills checkout, normally ~/ai-agents-skills.

When to use

• a command or operation fails unexpectedly
• the user corrects you
• the user asks for a capability that does not exist yet
• a recurring better approach is discovered
• a workflow or routing rule should become durable across agents, settings, or operating systems
• a learning affects one or more installed skills, optional artifacts, manifests, runtime helpers, docs, tests, or installer behavior

Targets

Local task notes:

• .learnings/LEARNINGS.md
• .learnings/ERRORS.md
• .learnings/FEATURE_REQUESTS.md

Canonical integration targets:

• canonical/skills/<skill>/SKILL.md for reusable skill behavior
• canonical/runtime/ for portable helper behavior
• manifest/ for installability, profiles, dependencies, artifacts, and runtime smoke contracts
• installer/ai_agents_skills/docs.py plus generated README.md and docs/ for generated docs
• docs/source/index.md and docs/source/overview.md for manual docs-site landing pages
• tests/ for focused regression coverage

Codex, Claude, DeepSeek, Copilot, Grok, and other agent homes are runtime install targets, not the primary place for reusable fixes. Update them through the installer after the canonical repo change is planned, reviewed, and verified.

~/.openclaw/workspace/skills/* (including dual-route adapters such as aas-remote-bridge) is also an install product. Edit canonical/skills/ and canonical/runtime/ first, then publish or install. Never land durable skill logic only under OpenClaw.

Install-target model

• Codex, Claude, DeepSeek, Copilot, and OpenCode are normal skill targets declared in manifest/skills.yaml.
• Copilot can receive skill adapters and supported personas, but not instruction blocks, templates, or command artifacts unless the installer gains that support.
• OpenClaw is explicit-only and fake-root-only before native target evidence. Real .openclaw writes, instruction blocks, symlink/reference modes, unclassified support files, and runtime-backed skills remain blocked until the OpenClaw install-target gates are satisfied.

When a learning mentions "all targets", state this matrix explicitly and do not claim Copilot or OpenClaw support beyond the current installer contract.

OS and substrate model

Treat OS coverage as a matrix, not a slogan:

• Linux, macOS, and WSL use POSIX runtime runners.
• Native Windows uses PowerShell runtime runners; CMD runtime entrypoints are not published.
• Git Bash/MSYS may run POSIX helpers, but the installer target shape is still Windows unless the check is deliberately scoped otherwise.
• A mounted Windows profile inspected from Linux or WSL is not proof that a native Windows agent can execute the same paths.

Reusable learnings should say which OS/substrate was inspected, which was inferred, and which remains unverified.

Workflow

1. Classify the event:
   • failure -> ERRORS.md
   • correction / better pattern -> LEARNINGS.md
   • missing capability -> FEATURE_REQUESTS.md
2. If the .learnings/ files do not exist yet, create them from the templates in assets/.
3. Append a concise structured entry.
4. If the lesson is reusable, add a Canonical Integration Plan section. Include related skills/settings, affected install targets, affected OS/substrates, proposed repo files, docs generation, tests, and blocked limits.
5. Suggest the integration plan before editing canonical files. Do not mutate the repo, installed agent homes, or runtime files without an explicit user request or an already-approved implementation task.
6. When approved, implement the smallest canonical repo change that solves the reusable problem, regenerate docs, and run focused verification.

Reminder loop

The reminder loop is manual unless a caller wraps it in their own automation:

1. after a command failure, user correction, or unexpected workaround, pause before the final reply
2. ask whether the event created a reusable lesson, error pattern, or missing capability note
3. if yes, log it immediately instead of trusting memory to preserve it later
4. if the lesson affects global workflow, routing, skills, settings, install targets, or OS behavior, propose a canonical repo integration plan

Trigger checklist

Use this skill especially when any of these happened:

• a command failed in a way you did not predict
• the user corrected a factual or workflow mistake
• you discovered a non-obvious workaround or gotcha
• you almost said a capability was unavailable before checking local evidence
• the same pain point has appeared multiple times and should be made durable
• a helper command, runtime wrapper, or doc example works for one OS or target but not another

Portable runtime helpers

The target-neutral helpers are installed through the shared runtime when this skill is selected with runtime profile auto or full. Prefer the managed runtime runner over paths inside an agent skill directory, because reference install mode intentionally does not copy support files.

POSIX example:

runtime="${AAS_RUNTIME_ROOT:?Set AAS_RUNTIME_ROOT to the installed runtime root}"
bash "$runtime/run_skill.sh" \
  skills/self-improving-agent/run_self_improving_agent.sh review-pending

Windows PowerShell example:

$runtime = if ($env:AAS_RUNTIME_ROOT) { $env:AAS_RUNTIME_ROOT } else { "$env:LOCALAPPDATA\ai-agents-skills\runtime" }
& "$runtime\run_skill.ps1" "skills/self-improving-agent/run_self_improving_agent.ps1" review-pending

CMD runtime launchers are not published because CMD cannot preserve arbitrary argument vectors safely. Use the PowerShell command above on Windows.

Common helper commands:

bash "$runtime/run_skill.sh" \
  skills/self-improving-agent/run_self_improving_agent.sh review-pending --high-only

bash "$runtime/run_skill.sh" \
  skills/self-improving-agent/run_self_improving_agent.sh check-command-safety \
  "git push --force origin main"

some_command 2>&1 | bash "$runtime/run_skill.sh" \
  skills/self-improving-agent/run_self_improving_agent.sh detect-common-errors

bash "$runtime/run_skill.sh" \
  skills/self-improving-agent/run_self_improving_agent.sh integration-plan \
  --summary "Make paper-review routing durable across agents" \
  --skill paper-review \
  --target codex --target claude --target deepseek \
  --os linux --os windows

Suggested verification

For a learning-only entry, inspect the entry and mark unresolved target/OS claims as unverified. For behavior-affecting repo changes, choose the narrowest relevant checks, commonly:

make docs
make test
make runtime-smoke ARGS="--skills self-improving-agent"
make fake-root-lifecycle ARGS="--skill self-improving-agent --platform-shape all"

Run native Windows checks with ./make.ps1 on Windows before claiming native PowerShell/CMD execution. Linux-hosted --platform-shape windows checks are useful install-shape evidence, not native execution evidence.

Read only when needed

• assets/LEARNINGS.md
• assets/ERRORS.md
• assets/FEATURE_REQUESTS.md
• references/examples.md for compact sample entries

Rules

• Keep entries short, concrete, and actionable.
• Do not write durable learning entries during a report-only, diagnosis-only, review-only, or investigation-only request unless the user explicitly asks for persistence. Report the lesson or proposed integration plan first, then wait for permission to modify files.
• Prefer canonical repo changes over one-off edits in agent homes.
• Separate inspected, inferred, and blocked target/OS coverage.
• Do not depend on OpenClaw hooks, session-spawn tools, or local runtime state unless the specific task explicitly requires them.
• Treat helper commands as convenience tools; the durable skill behavior is the structured learning plus canonical integration plan.