git:20261004.5baa412 to git:20261006.5699391

51 added, 7 removed. Audit A to A.

---

name: release-prep

description: Clean verified release artifacts and prepare MemPalace metadata, installed acceptance, and local qualification before separately authorized publication.

disable-model-invocation: true

---

# Release preparation

## Entry and authority

- Read `AGENTS.md`, `docs/RELEASING.md`, `CONTRIBUTING.md`, and

`docs/quality/acceptance/README.md`. Those files own gates and report format.

- Derive permitted edits, cleanup, commits, and exclusions from the active request.

- Reuse existing authorization. Inspection alone permits no mutations.

+ Reuse existing authorization, including disposable-platform tests and local

+ source delivery. Check each external effect separately. Inspection alone permits

+ no mutations; do not ask again for an already-authorized effect.

- Verify cwd, Git root/common directory, status, worktrees, processes and open files.

Use `autopilot doctor --json` when available; otherwise use the repository's

documented manual ownership check. Preserve initial edits and stop overlapping writes.

- Never invoke Codex, Claude, Gemini, another model/provider, or an authenticated

client as a preparation or qualification gate. Never inspect or require

credentials, tokens, keychains, auth files or paid-account state.

- Push, tag creation, public ref promotion, settings changes and publication each

require explicit operator authorization. Preparation grants none of them.

Never use `--force` or rewrite a public ref.

## Inspect and clean

- Read current backlog through the project-bound `backlog-utility` interface.

Do not create tasks, worksets or a second release plan merely to run this skill.

- Resolve public history with `python scripts/release_public_read.py --version-tags`.

Local tags and local development `main` are insufficient release identity.

Separate newest public tag, last published release and pending local version.

- Compare public history with the current behavior owners. Choose the smallest

justified version change; keep an already prepared unpublished version when suitable.

- Inspect untracked/ignored files, stashes and commits unique to affected worktrees

before cleanup. Remove only identified disposable artifacts with verified ownership

and no open writer; apply the repository's target-identity gate before deletion.

Preserve palaces, databases, backups, handoffs, evidence, environments and unique work.

Missing ownership blocks only cleanup of that item. Do not blanket-clean or prune.

## Prepare the existing owners

- Use `pyproject.toml`, `CHANGELOG.md`, public docs and existing generators.

Keep `## Unreleased`; consolidate user-visible changes under the pending version.

Exclude private operational history and local machine details from public notes.

- State exact edited paths and acceptance commands before writing; verify all four

Git views afterward. Do not add another release controller, gate or procedure.

+ - Audit current public entrypoints for each changed behavior: README, feature guide,

+ agent installation, LLM rules and examples; check packaged instructions only for

+ their declared capabilities. Preserve dated historical reports. Resolve claims

+ against source and link the existing feature owner instead of copying a procedure.

+ - Require RELEASING section 2's privacy audit before public delivery. Use Git's

+ inventory to include hidden LLM instructions/prompts/settings and new staged

+ backlog history. Review actual private identifiers/context, not just token patterns;

+ classify examples and public authorship with evidence. Inspect both archive contents

+ and exact outgoing ancestry. Preserve local private evidence, retain only redacted

+ findings and block delivery on confirmed disclosure or consequential unknowns.

+ - Inspect newly staged canonical records for private data. Preserve sealed local

+ records and use existing ignore rules and supported backlog operations; never

+ scrub immutable history or exempt it from the scanner. Before switching or

+ integrating refs that still track excluded files, retain validated sealed bytes

+ through the existing backup/export route and verify hashes/store validity afterward.

- Follow section 1a of `docs/RELEASING.md` for installed-candidate acceptance and

write its report at `docs/quality/acceptance/vX.Y.Z.md`. Use disposable installs,

palaces and profiles; preserve real data and user configuration.

+ - Require direct coverage of installed CLI commands, delegated subcommands, every

+ enabled MCP tool in every profile, both protocol handshakes, and supported extras.

+ Retain wheel-bound invocations, expected/actual results and relevant failure/retry

+ postconditions. Classify help, parser errors and listings as discovery/guidance;

+ classify documented retired/unauthorized-operation refusals explicitly. Missing

+ required execution is blocking UNRUN, even when source tests or CI pass.

- Reuse unchanged, identity-bound evidence. Rebuild and rerun affected acceptance

after package changes; do not count source tests as installed CLI/MCP execution.

Never inspect or import real user conversations to satisfy synthetic coverage.

- Triage findings in the existing evidence. Fix a small blocking defect only when

required and within current authority; otherwise report its smallest correction.

No automatic backlog decomposition or separately mandated task per fix.

## Qualification

Use the repository virtualenv. Run the exact existing gate set in RELEASING

section 1 and CONTRIBUTING; retain raw logs outside the published tree.

Core preparation checks:

```bash

python scripts/docs_drift_guard.py

python scripts/public_safety_scan.py --tracked --staged

+ python scripts/public_safety_scan.py --committed

+ python scripts/gitleaks_scan.py full-history --artifact-dir .codex-local/release-privacy

python scripts/quality_scorecard.py --check

python scripts/gen_code_intelligence_packet.py --check

python scripts/release_preflight.py --json

```

- Build the reviewed wheel, then run the existing installer matrix and installed

+ Build both distribution formats, then run the existing artifact checker, installer matrix and installed

application owner; get all flags and platform prerequisites from RELEASING:

```bash

- python -m build --wheel --outdir dist

- python scripts/release_install_metadata_smoke.py --all-installers --install-spec . --json

+ python -m build --outdir dist

+ MEMPALACE_RELEASE_WHEEL="$(pwd)/dist/mempalace_code-X.Y.Z-py3-none-any.whl"

+ python scripts/release_artifact_gate.py --dist dist --require-wheel --require-sdist --json

+ python scripts/release_install_metadata_smoke.py --all-installers --install-spec "$MEMPALACE_RELEASE_WHEEL" --json

python scripts/release_readiness_gate.py --installed-golden-wheel "$MEMPALACE_RELEASE_WHEEL" --json

```

- Bind `MEMPALACE_RELEASE_WHEEL` to the exact built wheel before invocation.

+ Replace `X.Y.Z` with the pending version and retain the exact wheel digest.

+ Use that wheel in every installer and platform contour; a source install is not

+ evidence for the built artifact. Set `MEMPALACE_RELEASE_SYSTEMD_USER=1` only in the

+ congruent disposable Linux fixture described in RELEASING.

The required Linux systemd-user receipt cannot be replaced by macOS results.

Missing model provenance or platform prerequisites remain blocking `UNRUN`;

provision only within current authority, then retry the existing owner command.

No confirmed high/critical issue may remain; record other findings and explicit

deferrals per the acceptance-report contract. Never invent PASS or weaken a gate.

An authorized local commit follows the repository's independent source-review

policy; that review is separate from credential-free release qualification and

is not application-acceptance evidence. Do not launch provider clients from this

- workflow. Once the acceptance report is committed for the exact package ids:

+ workflow. Write the acceptance report after installed qualification, inspect every

+ public path including newly staged records, rerun public-safety on the complete

+ staged tree, and commit within existing authority. Once the acceptance report is

+ committed for the exact package ids:

```bash

python scripts/release_preflight.py --tag vX.Y.Z --require-clean --json

```

For the opt-in read-only pre-tag check at its documented stage:

```bash

python scripts/release_preflight.py --tag vX.Y.Z --require-clean --check-live-upstream

```

Clean-tree or committed-report checks remain pending until their prerequisites exist.

+ After documentation-only edits, rebuild archives and compare runtime/plugin members,

+ dependency/entrypoint metadata and bound package ids with the tested wheel. A README-only

+ metadata delta needs fresh artifact and installed metadata checks; record both wheel

+ digests and the exact delta before reusing application evidence. Other package changes

+ need affected installed checks and a refreshed report. Preserve unaffected evidence.

Candidate creation follows RELEASING and requires installed acceptance first.

## Result and recovery

- Report version/public baseline, reviewed source SHA, exact changes and cleanup,

+ Report version/public baseline, reviewed source SHA, documentation surfaces checked,

+ exact changes and cleanup,

wheel/source identities, checks run versus reused, unrun/failed gates and one

existing recovery command. Mark local readiness separately from public admission.

+ For each platform name direct installed CLI/MCP coverage, installer/systemd receipts,

+ fixture cleanup and live-state preservation; distinguish them from source tests.

+ Report privacy outcomes separately for the current public tree, LLM instructions,

+ archives, outgoing ancestry and retained local state. Detector PASS is not proof that

+ private operational data is absent. Name unresolved findings before release handoff.

Hand off to `/release` only after local qualification and committed acceptance pass.

Push, tag, release and live-palace actions remain pending unless explicitly authorized.