v2.0.3 — code-consigliere

Adds /code-consigliere [target] to the main Skill Mesh catalog. Start from a target project or pass its path from another folder to equip private collection, inventory workflows, investigate behavior and rationale, and hand off agreed improvements as seeds.

The shared core preserves Init -> Discuss -> Signoff -> Implement, keeps recorded rationale separate from observation and inference, and maintains evidence-backed lessons and explicit terminology. Real project records stay outside repositories. The skill uses the separately installed consigliere helper (or a validated local checkout); without it, investigation stays read-only and reports that persistence is unavailable.

Includes Claude, Codex and GPT/Copilot adapters, generated workflow cards, and the scoped hermetic inventory-generation repair. Catalog: 58 skills, 55 portable and three Claude-native.

Validation:

• Independent cross-family final review: PASS after material findings were fixed.
• Final staged package: 696 passed, three skipped; 449 distribution files checksummed.
• Full repository run: 2,153 passed, two skipped, one failed inherited stale-report check. The six stale hash cells were regenerated and independently reviewed; the full smoke suite then passed 40 tests. Operator accepted this combined evidence plus final staging for this release, without another full-suite run. The historical full run remains failed.
• Native Claude synthetic trial persisted findings and vocabulary without target-file changes.
• Native Codex synthetic trial loaded the entry and investigated read-only without target-file changes. Its CLI sandbox could not access the external private store, so Codex persistence remains unverified in that configuration.

Restart host sessions after upgrading. The companion helper is a separate project and is not bundled in this archive. Synthetic validation is not real-project pilot acceptance.

Skill Mesh v2.0.2

Skill Mesh v2.0.2

Patch release over v2.0.1. Qualified source: f6e6c62.

What changed

• session-wrap now presents its verdict as plain language with separate durable, at-risk, deliberately untouched, open-work, and operator-action sections.
• user-wrap preserves that report verbatim, so returning operators can see unfinished plans and other unowned work before closing a window.
• repo-wrap now performs a small post-closeout cleanup after a successful owned-project wrap: it considers only stale build-step worktrees, removes one only when it is unlocked, clean, attached, and git cherry proves no unique patch versus the pushed default branch, uses ordinary git worktree remove without force, and keeps the branch. Ambiguous or active worktrees are left alone.
• The repo-wrap workflow card and public version links were refreshed.

No skill was added or removed: 57 Claude skills, 54 each for Codex and Copilot.

Installation and upgrade

Use Windows PowerShell 5.1, verify the archive against the attached .sha256 file, extract it to a dedicated directory, and run:

powershell -NoProfile -File tools/install-skill-mesh.ps1 -Provider claude -Home C:\path\to\consumer-home -DistDir .\dist

Use -Provider codex or -Provider gpt for the other hosts. Back up the discovery tree and .skill-mesh-install.json before upgrading, and start a fresh host session afterward.

Qualification

The committed source passed the complete package-integrity suite before the final simplification (697 passed). The final two-condition cleanup contract and link checks passed (46 passed), all three profiles rebuilt (57 Claude skills, 54 GPT, 54 Codex), and tools/release.ps1 -Provider all rebuilt a clean tracked-file stage and passed its final staged package-integrity gate: 692 passed, 3 stage-only skips. It checksummed 440 distribution files. A repository-root suite was started but deliberately stopped at 24% when the operator requested scoped testing; no failure had appeared. No lint or typecheck command is configured.

Archive SHA-256: 19dd9649381f1ca5fad100b7a6d0455403f61707276687f76c9e53b24c2ae73a.

Skill Mesh v2.0.1

Skill Mesh v2.0.1

Patch release over v2.0.0. Qualified source: cbd6544db71f5d73d324e1da19791ac2c8ae5861.

What changed

Operator-facing documents are saved as files, never published as hosted Artifacts. The new single owner is _shared/document-output.md: a skill that produces a document the operator reads later writes one self-contained HTML file to docs/<skill-name>/ inside the project repository and reports the repository-relative path. This applies identically to the Claude, Codex, and GitHub Copilot profiles.

Skill	Before v2.0.1	v2.0.1
plan-redline	Claude: private hosted Artifact (https://claude.ai/artifact/...) recorded as the Proposal: locator; Codex/Copilot: documentation/<plan-slug>-proposal.html	All hosts: docs/plan-redline/<plan-slug>-proposal.html, recorded as the Proposal: locator
repo-update Step 12	Claude: guided tour published with the Artifact tool; Codex: an unspecified file in the repository	All hosts: docs/repo-update/<project-slug>-<phase-or-version>.html, committed and pushed in its own docs: commit when Steps 9-10 ran; Tour: line carries the path

• A plan whose Proposal: line still records a hosted URL is migrated on the next /plan-redline run: the file is written, the line is replaced with the path, and the report states that the prior hosted copy remains published and is now stale. A recorded documentation/...-proposal.html path is kept as the republish target unless the operator asks for a move.
• --local-only on plan-redline (passed by plan-expedite) is now a compatibility no-op.
• The artifact-design skill is still loaded on Claude Code for page craft; it no longer decides location.
• Provider guides, parity deltas, workflow diagrams (plan-redline, repo-update), and the routing chart were updated to match. A new package-integrity test (tests/package-integrity/test_document_output_contract.py) pins the contract on every provider.

No skill was added or removed: 57 Claude skills, 54 each for Codex and Copilot. Other skills that write Markdown reports or run evidence under their own existing defaults (judge-ui, judge-motion, tier-escalate, tier-offload, skill-iterate, skill-evolve, user-brainstorm) are unchanged in this patch; they never used the Artifact tool.

Installation and upgrade

Same procedure as v2.0.0. Use Windows PowerShell 5.1, verify the archive against the attached .sha256, extract to a dedicated directory, and run the installer against the existing home:

powershell -NoProfile -File tools/install-skill-mesh.ps1 -Provider claude -Home C:\path\to\consumer-home -DistDir .\dist

Use -Provider codex or -Provider gpt for the other hosts. Retain the v2.0.0 distribution and back up the discovery tree and .skill-mesh-install.json before upgrading. Start a fresh host session after installation.

Qualification

Qualification on the qualified source: pytest over tests/package-integrity, tests/distributions, tests/release, tests/router and tests/smoke -- 1453 passed, 1 skipped (4h41m). tools/release.ps1 -Provider all then rebuilt the distributions into a clean stage and re-ran the package-integrity gate against the staged artifacts before writing CHECKSUMS.txt (440 dist files). The full repository-root suite (calibration, experiments, production-toolchain, telemetry) was not re-run for this patch; it last ran green at v2.0.0.

The v2.0.0 known limitations are unchanged.

Skill Mesh v2.0.0

Skill Mesh v2.0.0

Qualified source: ddce08eef3ac65084323d75a150595512ef133c9.

The catalog includes 57 Claude skills and 54 portable skills for Codex and GitHub Copilot. All 42 planned rebuilds and 15 retained candidates have source dispositions. Skills are generated from shared behavior contracts and host adapters; this release does not promise every skill will be shorter.

Installation

Use Windows PowerShell 5.1. Extract the release bundle to a dedicated directory. From that directory, select ONE provider and an explicit target home. The distribution root must contain the provider subdirectory:

powershell -NoProfile -File tools/install-skill-mesh.ps1 -Provider claude -Home C:\path\to\consumer-home -DistDir .\dist

For Codex use -Provider codex; resolve the effective Codex home with tools/probe-codex-skills.ps1 as described in documentation/providers/codex.md. For GitHub Copilot use -Provider gpt. Discovery roots are .claude/skills, .agents/skills, and .github/skills, respectively. Start a fresh host session after installation. Installing a profile does not authenticate its host or prove every native capability is available.

Upgrade and rollback

Before upgrading, retain the previous distribution and a separate backup of the target discovery tree and .skill-mesh-install.json. Stop sessions that might write those paths while changing installations. Run the same installer against the new distribution and existing home. It verifies ownership and current bytes; if it refuses a collision, inspect the conflict rather than adding -Force routinely.

To remove this profile's unchanged owned files:

powershell -NoProfile -File tools/install-skill-mesh.ps1 -Provider claude -Home C:\path\to\consumer-home -Uninstall

To restore the preserved previous generated profile, use this release's installer with -DistDir pointing to the preserved previous distribution after the current profile is uninstalled. Qualification exercises this installer restoring preserved pre-rebuild profile bytes; it does not establish compatibility of an older installer or every historical v1 release. Legacy or hand-maintained installations require their own preserved backup and the documented migration procedure; uninstall alone does not reconstruct legacy files. Modified files can require explicit resolution. Keep unrelated user files intact.

Known limitations

• Session advice passes the selected safety/provenance checks and emits a four-line final report, but may add a progress sentence and vary triage-line formatting. The operator accepted this presentation limitation for v2; strict formatting is deferred to v3.

• Native UAT's selected fail-stop smoke passed. Its first command's exit-code wrapper did not independently capture the inner printf exit; matching output was observed. This is a scoped result, not blanket workflow qualification.

• Judge Motion is Claude-native and needs its packaged runtime dependencies, browser runtime, independent judge capability, and native permission approval. The selected live smoke passed: six mechanical assertions, three independent Sonnet judge returns, calibration PASS, and a real missing-rubric capture that correctly escalated without judging. Protected report writes beneath .claude required native safety review; command allow rules alone did not enable those writes.

• Session-dependent actions require authoritative host session identity and the permissions named by the skill. Earlier project SET/CLEAR observations without session identity and Codex read-only holds remain scoped holds; this release does not claim those unavailable write paths succeeded.

• claude-oauth-auth, context-slim, and judge-motion are Claude-only. PowerShell 7 compatibility is not claimed.

• The citation utility dependency is published on skill-mesh-v2/overview-filters in aberson/citation-needed, not merged into that project's main branch.

Qualification

Full root suite: 2128 passed, 16 skipped in 15560.49s (4:19:20). All three final profiles passed fresh install, repeat install, upgrade from preserved pre-rebuild profiles, uninstall preserving unrelated files, and prior-profile byte restoration with the final installer. This is not proof of every historical v1 installer or restoration of user-modified files. The preserved prior profiles were not a published v1 release.

Archive SHA-256 is in the attached checksum file; generated-file hashes are in the bundle CHECKSUMS.txt. Normal installed profiles were not changed during qualification.

Workflow documentation

Browse the v2 README, complete skill reference, and 57 source-linked workflow diagrams. Documentation was published after the qualified bundle; the release tag and assets are unchanged.