Self Improve

Self-improve

Audit the durable guidance layer, prove how it behaves in the active runtime, and apply only approved changes. Verify current capabilities from live evidence and official sources rather than hardcoding release or model names.

Scope and depth

An explicit or natural-language self-improve request performs the requested scope. A wrap-up caller starts with skills used, guidance/memory/config changed, and failures, workarounds, conflicts, or discovery gaps observed this session. Do not turn routine wrap-up into a release-wide audit without such evidence.

Route major Codex model/app/CLI or OACP release migrations, multi-skill capability changes, and context-budget audits to $audit-oacp-skills. Route a changelog summary without guidance-impact analysis to $codex-changelog.

Treat audit, review, diagnosis, and report requests as read-only. A numbered approval authorizes only its listed local edits. Commit, push, PR, merge, messaging, and other external writes remain separate unless explicitly listed.

1. Resolve targets and ownership

Use the active Skills catalog and exact locators first. Read every selected SKILL.md completely, then load only the references or scripts required for the selected scope.

Classify each target before proposing edits:

• Authored source: editable only after approval.
• Symlinked skill or guidance: edit the resolved source and name its repo.
• Installed artifact: plugin cache, bundled/system skill, or provider resource; inspect read-only and locate an authoring/update path.
• Cross-runtime namesake: report drift without synchronizing another runtime automatically.

Generated Codex memory under ${CODEX_HOME:-$HOME/.codex}/memories/ is generated state, not hand-maintained project memory. Do not hand-edit or apply line-budget trimming to it. Inspect configuration through an explicit key allowlist and redact tokens, credentials, environment values, and private endpoints. Never broadly search credential-adjacent config or session trees.

Load only the applicable section of references/audit-checks.md for detailed skill, curated-memory, or AGENTS/config checks.

2. Prove effective state

For every target, verify the relevant chain:

authored source -> discovery/configuration -> callers/dependents -> observed behavior

Source validity alone is insufficient. Check applicable auxiliary metadata, canonical discovery paths, symlinks, manifests, hooks, setup generators, and direct skill/AGENTS callers within explicit authored roots. Run one live or test probe when available. Mark an unavailable probe as unverified; do not silently treat authored intent as effective behavior.

If a dependent is outside the approved scope, report it as follow-up instead of editing it. This step must catch valid-but-undiscoverable skills, stale caller syntax, generated configuration that differs from its source, and scripts whose documented effect is not observable.

For current OpenAI/Codex behavior, use $openai-docs or $codex-changelog as appropriate. Capability-check unstable CLI surfaces and distinguish authored, configured, and runtime-observed states.

3. Analyze and capture acceptance evidence

Use these tags:

Check trigger behavior, workflow inputs/outputs, stopping conditions, approval boundaries, safety, current paths/commands/config, ownership, duplication, run learnings, and validation. Promote repeatable run lessons with $consolidate-learnings; never append a second procedure as a learned log.

For every proposed change, define a before/after acceptance probe and capture its current result before editing. Use context measurement only when the finding concerns a hot-path skill, progressive disclosure, a release-wide migration, or a context budget. Reuse $audit-oacp-skills measurement with the same tokenizer for the comparison; routine correctness fixes do not require a token baseline.

4. Report and request approval

Group findings by requested scope. For each finding provide the target and line when available, live evidence, impact, specific outcome, and acceptance probe. Say No issues found for an in-scope category with no findings.

End a changeable audit with a self-contained numbered menu containing:

• the recommended choice first when clear;
• every proposed change number;
• target files/repos and exact outcomes;
• allowed replies: all, none, or item numbers;
• whether commit, push, PR, messaging, or another external action is included.

Do not edit before approval unless the original request already authorized the exact change.

5. Apply and verify approved changes

1. Re-read every live target and nearest AGENTS.md; confirm repo, branch, status, ownership, and unrelated concurrent changes.
2. Preserve user and other-agent work. In a shared skills repository, edit only Codex-owned codex/ or explicitly approved shared files.
3. Keep installed/plugin/system/provider artifacts read-only. Coordinate with another runtime only when messaging is authorized.
4. Before editing OACP runtime memory, acquire the project's .memory-write.lock; defer when another valid writer holds it and remove a lock created by this run on exit.
5. Apply only approved edits with apply_patch, inspect the diff, and confirm the intended files changed.
6. Re-run the exact acceptance probe captured before editing, followed by proportional project-native validation. For a changed skill, run the skill validator, check frontmatter and direct links, exercise representative positive/negative triggers, and verify effective discovery when applicable.
7. When context measurement was warranted, compare with the same tokenizer and explain increases as well as reductions. Lower token count is not a correctness result.

For an OACP runtime-memory edit, use an explicit owner token because a shell trap cannot span separate Codex tool calls. Never remove a valid writer lock solely because it is old.

Acquire before the first edit:

LOCK_DIR="$OACP_ROOT/projects/${PROJECT}/.memory-write.lock"
LOCK_OWNER_FILE="$(mktemp)"
python3 - "$LOCK_OWNER_FILE" <<'PY'
import datetime
import json
import secrets
import socket
import sys

owner = {
    "token": secrets.token_hex(16),
    "host": socket.gethostname(),
    "created_at_utc": datetime.datetime.now(datetime.timezone.utc).isoformat(),
}
with open(sys.argv[1], "w", encoding="utf-8") as handle:
    json.dump(owner, handle)
PY
chmod 600 "$LOCK_OWNER_FILE"
if mkdir "$LOCK_DIR" 2>/dev/null; then
  chmod 700 "$LOCK_DIR"
  cp -- "$LOCK_OWNER_FILE" "$LOCK_DIR/owner"
  chmod 600 "$LOCK_DIR/owner"
  echo "Memory lock acquired"
else
  echo "Memory lock held by another writer; defer memory edits"
fi

If acquisition fails, leave the existing directory untouched and defer all runtime-memory edits. Lock recovery requires separately proving the prior writer is gone; age alone is never proof. Release after the last edit, including an edit failure, only when the stored owner is byte-identical to this run's token:

if [ -d "$LOCK_DIR" ] && cmp -s "$LOCK_OWNER_FILE" "$LOCK_DIR/owner"; then
  command rm -f -- "$LOCK_DIR/owner"
  rmdir -- "$LOCK_DIR"
else
  echo "Memory lock ownership changed or was not acquired; refusing release"
fi
command rm -f -- "$LOCK_OWNER_FILE"

6. Commit or publish only when requested

Commit, push, PR creation, merge, and messaging are separate actions. When explicitly requested, follow each owning repo's AGENTS.md, reconfirm its root, branch, and status, stage only approved files, and keep repositories independent. Runtime memory and unrelated untracked state are never included merely because guidance changed.

7. Summarize

Report scope and sources, effective-state evidence, findings applied or deferred, files changed by owning repo, exact validation and before/after probe results, external-action state, and remaining risks.

Keep changes surgical. Do not invent current facts, hardcode ephemeral model or release details, or let an audit drift into unapproved runtime namespaces.