moldea 10.1.0

This release targets CLI 13.0.4 and Core 9.2.0, with supported minimums of ^13.0.4 and ^9.2.0. Core supplies authored YAML locations for pointer-addressed manifest and decision diagnostics, including cross-file findings and unmatched acceptances.

Updating the Skill alone does not update a project's CLI. Update the repository-local CLI development dependency through the project's package manager.

Install the pinned Skill:

npx skills add "moldea-ai/skill#v10.1.0"

Deterministic conformance, release identity and installation checks passed. This release retains the exact historical semantic and scripted-journey selections from Skill 10.0.3, with their disclosed failures, exclusion and evaluator limitations. No fresh semantic evaluation or provider journey was performed; these selections provide no exact current-artifact model assurance.

Release evidence and scope.

moldea 10.0.5

Install the immutable release:

npx skills add "moldea-ai/skill#v10.0.5"

This patch makes Fix moldea guide a large repair through rounds in the same conversation. Each round ends with a status block showing progress, the files the repair changed, the documents it rechecks first and why the round ended. The usual Checked, Fixed, Verified, Needs input and Not checked sections follow, then numbered next steps and a resume point with the full commit. Reply in the same conversation to continue; the coding agent never sends you to a new session. Committing between rounds is optional, and committing the repair's files spares later rounds from rechecking them. Repair also posts one short progress line per completed owner.

Repair and initialization instructions omit the model-token limit that the coding agent cannot measure. The command, moldea-call and output limits still bound each round, and the release-evaluation profiles are unchanged. The local-tooling reference's release sentence matches the skill version again.

The skill still targets CLI 13.0.3 with Core ^9.1.1. Runtime requirements remain Node.js >=22.17.0 and Git >=2.32.0, with supported CLI ^13.0.0, Core ^9.0.0, repository format 1 and CLI JSON schema 7. No schema migration is required.

Historical evidence

The maintainer separately approved reusing the unchanged historical selection for 10.0.5 only. It retains 142 final cases: 134 passed and eight recovered, with zero failed final cases, recorded under 10.0.3 instructions. The new repair report, continuation and limit changes have no model evidence. No semantic evaluation, provider journey or regrading was run, and the historical selection supplies no exact current-artifact model assurance.

The semantic bundle remains under evidence-10.0.3-20261026, with compressed SHA-256 fa6062bbad4512a4297d6bba711cbd435ea5c1a7638075fbfd21be2709b49c75. The scripted-journey bundle remains under evidence-journeys-20261013-da9709, with compressed SHA-256 96de16e7646e54a17958040125b1b8da225298e28fc4f54e7865c0ede3d9f8d5. See the 10.0.5 evidence reuse.

moldea 10.0.4

Install the immutable release:

npx skills add "moldea-ai/skill#v10.0.4"

This patch removes cosmetic managed-block warnings for LF/CRLF differences, an optional final line break, completely empty lines immediately inside the markers, and trailing ASCII spaces or tabs on marker lines. Equivalent AGENTS.md and README.md files keep their bytes and metadata. Instruction text, internal whitespace, malformed-marker checks and write safeguards remain exact.

This release also first distributes the repair-helper and managed AGENTS.md writer change from commit f4b52143: Git reads set diff.autoRefreshIndex=false to prevent automatic index refresh. That change was already on main before the formatting work and receives no fresh semantic evaluation under the historical evidence-reuse decision below.

The skill targets CLI 13.0.3, which requires Core ^9.1.1. Updating the skill alone does not update a project's CLI: request explicit moldea repair or update the project's CLI development dependency to 13.0.3 through its package manager.

Runtime requirements remain Node.js >=22.17.0 and Git >=2.32.0, with supported CLI ^13.0.0, Core ^9.0.0, repository format 1 and CLI JSON schema 7. No schema migration is required.

Historical evidence

The maintainer separately accepted the unchanged historical selection for 10.0.4 only. It retains 142 final cases: 134 passed and eight recovered, with zero failed final cases. The 160 retained trials include ten failed trials; repair-unproven-adoption remains the sole exclusion. Original identities, dates, criteria, observations, verdicts, recoveries and evaluator limitations remain unchanged.

This release-specific decision replaces the launch gates requiring zero failed cases, zero exclusions and no confirmed evaluator defects with that disclosed selection. It grants no future-release exception. No semantic evaluation, provider journey or regrading was run, and the historical selection supplies no exact current-artifact model assurance.

The semantic bundle remains under evidence-10.0.3-20261026, with compressed SHA-256 fa6062bbad4512a4297d6bba711cbd435ea5c1a7638075fbfd21be2709b49c75. The scripted-journey bundle remains under evidence-journeys-20261013-da9709, with compressed SHA-256 96de16e7646e54a17958040125b1b8da225298e28fc4f54e7865c0ede3d9f8d5.

See the complete 10.0.4 evidence-reuse decision. Current README calibration fixtures now exercise genuine instruction drift rather than tolerated blank-line formatting; this changes no selected historical record. Development tooling also fixes a Bubblewrap startup-cancellation race.

Verification and release identity

• Local full release qualification passed: 1,172 root unit tests, 77 website unit tests, 877 root integration tests with two existing platform skips, 73 website artifact tests, 277 production browser tests and five development browser tests. Cancellation qualification passed three consecutive retry-free runs.
• The original release PR passed its checks. The tag's initial CI run failed during Ubuntu runner sandbox setup before the affected Linux tests. The CI runner correction changes only the workflow and testing note; the corrected full matrix passed on commit 35b59b91746afc42ca1d298a39fcd3b79a363980 with identical portable artifact bytes. The immutable tag was not moved.
• Installation from the published tag matched all 36 portable artifact files byte for byte. Pages deployment passed on 0db512688989f0ca55388936bcd1dfd0bc0644d6; the live site displays 10.0.4.

Release commit: ff1465706a0aa81887a9775cd15fd6027cc78cb5. Qualified release tree: 84841666c68218c6dd9ca1e47f3f4f806a08f586. Signed annotated tag object: 79656ce0e8bb6efe389255d4a254ba30cc4ee34b.

moldea 10.0.3

Install the immutable release:

npx skills add "moldea-ai/skill#v10.0.3"

This release updates the evaluated tooling to CLI 13.0.2 and Core 9.0.1, improves bounded context assessment, healthy-project repair, final validation and actionable tooling reports, and includes evaluator correspondence and grading corrections. Website UI 4.5.3 renders unsupported code fences as text and correctly nests recorded response headings.

The maintainer accepted qualified historical evidence for this release. The selected snapshot contains 142 cases: 134 passed and eight recovered, none failed or pending. Its source is the completed 143-case recording on commit 20e28fa, whose result remains 134 passed, eight recovered and one failed. Publication performs no model calls or regrading and preserves the original failed trials, confirmations, responses and usage.

Not yet included: repair-unproven-adoption remains in the suite unchanged but is private. The model sometimes runs direct git status before the repair helper despite explicit instructions. No recorded run changed files or initialized anything. Commit b1098a2 adds the explicit pre-gate prohibition; its single focused recheck failed the same way. No behavior improvement is claimed.

Four recovered first-try failures were grader mistakes against correct behavior: maintain-dirty-project, readme-outside-managed-block, repair-readme-drift and skill-independent-followup. Their confirmations passed. The corrections in b1098a2 have deterministic regression coverage; the selected snapshot keeps those original grades unchanged.

The exact evidence asset is published under evidence-10.0.3-20261026, with compressed SHA-256 fa6062bbad4512a4297d6bba711cbd435ea5c1a7638075fbfd21be2709b49c75. It is historical presentation evidence and supplies no exact current-artifact assurance. Native and developer-supplied Claude Code observations retain their applicability qualifications. See the complete 10.0.3 evidence-reuse qualification.

The distributed moldea artifact is unchanged from b1098a2, with SHA-256 7749e5bb4c57e64ab54ec582faaec8899d9abdb9d5bec412933b14d5a134cad2. Runtime requirements remain Node.js >=22.17.0 and Git >=2.32.0, with supported CLI ^13.0.0, Core ^9.0.0, repository format 1 and CLI JSON schema 7. No schema migration is required.

Rollback remains available from the immutable v10.0.2 tag; never move a published tag.

Release commit: 38f8247. Reviewed tree: cb6e885610280c6707d64091341539738ad785f9. Signed annotated tag object: 6e49c1554a81d7a8f0f788230cefe490a59a776f. Main and development are synchronized.

Completed verification: PR #89 passed all 12 checks; tag conformance passed all ten jobs, including immutable-tag installation comparison on the supported runtimes; Pages passed verification, production build, deployment and sitemap submission. Local checks passed 1,134 unit tests, 858 integration tests with two existing platform skips, 219 production browser tests and five development browser tests. Website quality and 73 artifact tests passed, and the exact release production build passed. The live installation guide, historical qualification and 142-case evidence page were verified in a browser. No criteria or limits were relaxed.

moldea 10.0.2

Install and prepare manual testing

npx skills add "moldea-ai/skill#v10.0.2"

The skill installs from the immutable tag. Installation does not install the moldea CLI globally and does not refresh an existing repository instruction block. For an adopted repository, ask Fix moldea to assess the installed tooling and bridge. Apply a managed-region refresh only where instruction edits are permitted. Where they are prohibited, obtain the exact printed handoff for an authorized editor; integration remains incomplete until it is applied and validated. Preserve shared instruction symlinks and surrounding instructions.

Use two separate setups for instruction-routing tests: fresh 10.0.2 integration, and 10.0.1 upgraded and refreshed to 10.0.2. Start a new conversation after each setup, without a preloaded moldea entrypoint or prior troubleshooting. Setup conversations do not establish routing behavior.

Changes

• Repair attribute inspection supports contained directory symlinks and Windows junctions, including the reported 32-link layout. Resolution is bounded and retains refusals for escapes, cycles, dangling links, unsafe filters, and linked parents during written-file review. An absolute link may use the already-verified supplied root spelling or physical root; unknown outside aliases remain refused.
• The managed instruction template distinguishes loading another skill's instructional resources from actual project work before both routing branches. Actual project inputs, outputs, implementation, and skill behavior changes retain the appropriate routing.
• Standalone validation uses the launcher without a repair checkpoint. Repair keeps actor Git reads, including preliminary configuration lookup, inside the bounded helper.
• Pending authorized context maintenance can continue after an installation reply. Install-only requests, pauses, and topic changes constrain that continuation. Host implementation and release authorization remain host responsibilities.
• Upgrade guidance covers stale bridges, authorized refresh, protected-file print handoffs, ignored local instruction files, shared and local Claude-file precedence, warning eligibility, accurate owner counts, and local versus committed/cloud observations.
• Validation digests identify results rather than every canonical document body. Content changes use the relevant inspect/content identities and still require reassessment.
• Skip-link regression coverage verifies keyboard behavior in both themes, at 320px and desktop widths, after client navigation, and with an earlier hidden non-page link.

CLI 13.0.1, Core 9.0.0, CLI JSON schema 7, supported CLI/Core ranges, and the inherited dependency closure remain unchanged. Website UI remains 4.5.1, including the previously merged source-preview behavior. Portable runtime requirements remain Node.js >=22.17.0 and Git >=2.32.0. No schema migration is required.

Release identity

Identity	Value
Skill	10.0.2
Immutable tag	v10.0.2
Release commit	c2f7decd2301ef4c8953c0ebf13e5d066492f005
Reviewed and tested tree	b4db73975989bdc14ba4173e81249a270d40ea22
Annotated tag object	0c415245bfece2a33ef85c9a5cbe7bb70bccb0b4
Portable artifact SHA-256	17874ecad78914db17f7f110b5dc2b9b6ab2a51cefea482ec90d408216176ad6
Portable files / bytes	36 / 1,888,462
SKILL.md SHA-256	d34e4d19a542e5a4ffed0206c7c22cbf44a430e9a0c5882cd57f7c41f1e2f7bf
Managed template SHA-256	e7a0114a8047622092ea466057166646b340040cfc8d1a832df1cc899f3b581d
Managed writer SHA-256	9438af808e855a2ca7ac347124b476180770845c1280a1952a704e71bb371db2
Repair helper SHA-256	6b9ca29e8a5519c4741a64a4903a07246dce169366d2e064b1b8a5dce8fbf822

The tag's SSH signature was verified locally. An isolated project-local installation using skills@1.5.22 and the exact public tag matched every file and byte against the release commit, including the template and generated helpers. The developer's installed skills were not replaced.

Completed verification

• Release integration, PR #88 merged through the protected branch workflow.
• Tag conformance passed all 10 jobs: root correctness, repository quality, Linux Node.js 22.17.0/24.11.0/26.8.1, macOS portable integration, Windows portable/deep-path installation checks, minimum Git 2.32, classification, and the aggregate result.
• Pages verification, production build, and deployment passed for the release commit, including sitemap submission.
• The corrected local candidate passed 1,053 unit tests, 720 integration tests with two existing skips, 219 recorded browser tests, and five clean-development browser tests. The 59 repair integration tests passed three consecutive runs with retries disabled. The managed-writer fixtures and four new keyboard cases also passed three runs each without retries.
• Portable generation checks, typechecking, lint, formatting, path checks, documentation checks, release identity, website checks, selected-evidence preparation, and the 57-page production build passed. Production artifact inspection found no tests. Canonical specification validation passed with zero errors; the isolated specification worktree reported its expected missing protected-instruction bridge warning.
• The live home, documentation, and evidence pages returned HTTP 200 in all 12 theme/viewport cases: light/dark at 320px and 1440px. All 12 keyboard checks passed, with visible first skip-link focus, activation into main content, and subsequent focus within main. No horizontal overflow, page errors, or serious/critical axe violations were observed. Four home-page focus screenshots were reviewed.

Known limitations and evidence qualification

Manual checks below have not been claimed as completed by this release. No new model trials, semantic evaluation, or mock-project campaign were run. Deterministic tests establish tooling mechanics and tested browser behavior; they do not establish fresh-conversation instruction adherence or complete owner assessment by a coding agent.

The website retains the historical Skill 8.2.3 / CLI 11.0.6 semantic selection: 120 included cases, 115 passes, five recoveries, and two disclosed exclusions. It supplies no current-release semantic assurance. The ordinary-hit-launcher-unavailable semantic case remains unrecorded.

An earlier parallel browser run timed out in three unchanged mobile artifact/dialog accessibility tests. Each passed three isolated runs with retries disabled, and the corrected full candidate run passed. The original traces were retained for diagnosis. No assertion, timeout, scanner threshold, or workflow was weakened.

The deployed site still reports one moderate axe region finding on the skip link in each of the 12 audited cases. A hidden preceding /cdn-cgi/content link was observed, consistent with interference in axe's skip-link heuristic. The injecting feature has not been identified. Cloudflare settings and scanner policies are unchanged. Keyboard behavior passed independently; this residual scanner finding is the accepted release limitation.

Manual-test checklist

For each scenario, record the installed skill/template identity, CLI/Node.js/Git versions, host, exact request, gate calls and paths, instruction loading, interventions, and actual outcome. Disclose inherited personal skills, blocked scenarios, and failures. Do not count setup conversations as routing evidence, work around host denials, or fix an unrelated test project to manufacture a pass.

#	Scenario	Expected observation
1	Validate with an uncommitted contained skill-directory link.	Launcher validation without a repair checkpoint or writes.
2	Repair that checkout in a fresh conversation.	Supported checkpoint succeeds; all actor Git reads, including configuration lookup, stay inside the helper.
3	Retest the 32-junction Windows layout with multiple context owners.	Links, index, and ignore state remain unchanged; tooling and owners are assessed; final validation or an accurate completed count and continuation is reported. The issue report's complete owner assessment remains unverified until this check.
4	Run the menu-card skill in both fresh and upgraded/refreshed setups.	Loading instructional resources alone triggers neither moldea route; actual project paths retain routing.
5	Upgrade while retaining the old 10.0.1 bridge.	Existing block is preserved and reported stale; the agent does not claim the routing fix is active.
6	Refresh where instruction edits are permitted.	Only the managed region changes; the stale warning clears; repeating refresh is a no-op.
7	Refresh where instruction edits are prohibited.	Exact printed handoff; the file stays untouched; integration remains incomplete until the handoff is applied and validated.
8	Keep AGENTS.md local and ignored.	Accurate local versus committed/cloud explanation, existing acceptance eligibility, no tracking changes, and no invented cloud results.
9	Omit root CLAUDE.md; test nested shared and local fallback files separately.	The governing import subject and exact permitted repair/handoff are identified; absence of the root file does not dismiss integration. Shared files retain precedence over the local fallback.
10	Use another skill on a relevant project input or implementation path.	Normal relevance handling for actual project work.
11	Request code-only pricing work without eligible CLI tooling.	After a relevance hit, one actionable notice; no canonical reads or automatic installation.
12	Reply “install it” to a blocked authorized context update.	Install the identified tooling under host rules, refresh evidence, and finish the pending authorized maintenance. ...

semantic evidence sem-20261026014206661-66a3ed7e

Recorded semantic evidence for 10.0.3.

moldea 10.0.0

Skill 10.0.0 moves the bundled helpers, CLI 13.0.0 and Core 9.0.0 onto JG Packages, so Git reads, file writes, processes and argument parsing share one maintained implementation instead of moldea's own copies. It requires Git >=2.32.0, CLI ^13.0.0 and Core ^9.0.0, with repair target CLI 13.0.0, and keeps Node.js >=22.17.0 and every numeric limit.

Changes

• Git through jg-git-commands. The repair helper reads the stored attribute source, the index, the status, the checkpoint's commit and its diffs with isolateConfig, so it still sees only the repository's own config. The writer reads the index, the untracked listing and core.symlinks the same way CLI 13 does. A failed repair Git read reports jg-git-commands' failure message instead of Git's own output.
• Shared file, process and argument helpers. The helpers write files with jg-fs, run processes with jg-process and parse their arguments with parseStrictArgs.
• Narrower --save-response cancellation. A cancelled capture stops before its atomic write starts; a write already started may complete, and the invocation still fails with no stdout.
• Git 2.32 floor. Git 2.32.0 or later is required, up from 2.30.0.

Verification

• Repair Git, writer and managed-agents integration tests run on current Git on Linux, Windows and macOS and, in a new conformance job that builds Git 2.32.7 from source, on the new minimum.
• The sandbox Git policy tests run jg-git-commands' reads of the bundled helpers and of CLI 13 through the evaluator's wrapper, and the sandboxed host tests run the published CLI 13.0.0 through it.
• Recomputing the 6 journey definition digests, the 125 case definition digests and the case suite digest with the new deterministic JSON serialization reproduces the 9.0.0 values.
• Before this release, the release-candidate workflow passed all 15 package-manager profiles (npm 7 through 12, pnpm 8.3.1 through 12.3.1, Yarn 1.22.22, 4.14.1 and 4.18.0) against the CLI 13.0.0 candidate closure, byte-identical to the published packages.
• Evaluator protocol 53 admits jg-git-commands' command lines and CLI 13's in its sandbox. It has not been recorded.

Pinned semantic evidence

The semantic evidence was recorded on Skill 8.2.3 with CLI 11.0.6. None of the 10.0.0 changes was re-evaluated. Whether a coding agent follows the reworded cancellation rule is not established by that evidence, and deterministic checks do not establish behavioral reliability.

The selected bundle remains semantic-sem-20261018051032668-1465f9ee.json.gz under evidence-8.2.3-20261018: 120 cases, 115 passed and five recovered, with the same two disclosed exclusions. Its compressed SHA-256 is baa883450e1087efafbf3c685e311859fb63575dfa839761c20db48de292342d. No paid run, regrading or evidence repacking was performed for 10.0.0.

Install this exact release:

npx skills add "moldea-ai/skill#v10.0.0"

moldea 9.0.0

Skill 9.0.0 applies one rule to instruction links: validation and the bundled writer follow a link only to a regular file inside the repository, the same way for real links, Git links checked out as plain files and hosted repositories. It requires CLI ^12.0.0 and Core ^8.0.0, with repair target CLI 12.0.0, and keeps Node.js >=22.17.0, Git >=2.30.0 and every numeric limit.

Changes

• One link rule in validation. validate reads each link's stored text and follows it only inside the repository. A link that leads anywhere else reports MOLDEA_INTEGRATION_LINK_UNRESOLVED, which replaces MOLDEA_INTEGRATION_UNVERIFIED, with its reason as the qualifier: outside, missing, loop, not-file or unsupported. An accepted warning matches only that reason.
• The writer follows links exactly as validation does. managed-agents.mjs asks Git how each visited path is stored, with the configuration files the CLI reads, so files Git ignores and Git links checked out as plain files behave as they do in validation. It writes only to a regular file inside the repository, keeps the link, and refuses every other link with its reason and next step.
• A next step for every reason. Repair guidance explains each reason: replace an absolute link with a relative one, commit an ignored or uncommitted target or point the link at a committed file, fix a loop, or point the link straight at a file. It also explains why Cloud reports a missing AGENTS.md block when the file exists only locally.

Verification

• Writer tests with real Git cover inside links, Git links checked out as plain files with their bytes unchanged, a committed link replaced by a real file, conflicting index modes, every refusal reason, absolute links, ignored targets and ignored instruction files, Git failures, a name with pathspec magic, exact-path output matching, global core.symlinks and core.excludesFile settings, the recorded Git calls and filesystem containment.
• Disposable repositories driven through the launcher with the published CLI 12.0.0 covered initialization, repair, acceptance by reason, the three Claude Code link chains, absolute and ignored-target links with the writer's next steps, an ignored CLAUDE.local.md and a Git link checked out as a plain file.
• Before this release, the package-manager harness passed locally against the CLI 12.0.0 closure, byte-identical to the published packages, on every release-candidate entry (npm 7 through 12, pnpm 8.3.1 through 12.3.1 including Vitest 3.2.4 coexistence, Yarn 1.22.22, 4.14.1 and 4.18.0) and Yarn 3.8.7.
• Evaluator protocol 52 admits the writer's two Git reads in its sandbox. It has not been recorded.

Pinned semantic evidence

The semantic evidence was recorded on Skill 8.2.3 with CLI 11.0.6. None of the 9.0.0 changes was re-evaluated. Whether a coding agent follows the per-reason guidance is not established by that evidence, and deterministic checks do not establish the new rules' behavioral reliability.

The selected bundle remains semantic-sem-20261018051032668-1465f9ee.json.gz under evidence-8.2.3-20261018: 120 cases, 115 passed and five recovered, with the same two disclosed exclusions. Its compressed SHA-256 is baa883450e1087efafbf3c685e311859fb63575dfa839761c20db48de292342d. No paid run, regrading or evidence repacking was performed for 9.0.0.

Install this exact release:

npx skills add "moldea-ai/skill#v9.0.0"

moldea 8.3.0

Skill 8.3.0 moves the Claude Code integration check from instruction text into Core 7.1.0 and the bundled writer, and supports instruction files that are symbolic links. It requires CLI ^11.1.0 and Core ^7.1.0, with repair target CLI 11.1.0, and keeps Node.js >=22.17.0, Git >=2.30.0 and every numeric limit.

Changes

• Core checks linked and Claude Code instruction files. validate checks a symbolic-link AGENTS.md on its regular-file target, inside or outside the repository. It reports MOLDEA_INTEGRATION_CLAUDE_IMPORT_MISSING for the Claude Code instruction file that does not load AGENTS.md, and MOLDEA_INTEGRATION_UNVERIFIED for a link it cannot verify, such as a broken link or a network target. CLI 11.1.0 passes Git-ignored instruction files to this check and to no other.
• The writer follows inside links. managed-agents.mjs resolves links only through the Repository FS resolver that validation uses. It updates a linked AGENTS.md through its regular-file target inside the repository, keeps the link, and refuses any other link with its reason, so the coding agent hands off instead.
• The writer adds the Claude Code import. managed-agents.mjs --claude <file> appends @AGENTS.md, or @../AGENTS.md in .claude/CLAUDE.md, to the existing file named by the warning, keeping its line ending and doing nothing when an equivalent import exists. --claude <file> --print prints only the line. The helper never creates a Claude Code file.
• Acceptance only for shared files. The coding agent offers acceptance of an integration warning only when its file is shared and tracked by Git. An accepted warning about CLAUDE.local.md or an ignored file would report MOLDEA_ACCEPTANCE_UNMATCHED on every other checkout, so only the fix or the handoff is offered there.

Verification

• Real-link writer tests run on Linux, Windows and macOS and compare each written file with the file the operating system opens through the link. They cover refusals of outside, broken, looping and network targets, .. after a directory link, Claude Code import insertion, idempotence and CRLF, the --print forms, and never creating a Claude Code file.
• Disposable repositories driven through the launcher with the published CLI 11.1.0 covered initialization, repair and warning acceptance; a CLAUDE.md link to AGENTS.md, an AGENTS.md link to CLAUDE.md, and .claude/CLAUDE.md with @../AGENTS.md; an AGENTS.md linked outside the repository; and an accepted /CLAUDE.local.md warning reporting MOLDEA_ACCEPTANCE_UNMATCHED on a second checkout.
• Before this release, the package-manager harness passed locally against the published CLI 11.1.0 closure on every release-candidate entry (npm 7 through 12, pnpm 8.3.1 through 12.3.1 including Vitest 3.2.4 coexistence, Yarn 1.22.22, 4.14.1 and 4.18.0) and Yarn 3.8.7.
• Evaluator protocol 51 recognizes the two new helper forms and admits CLI 11.1.0's instruction-file Git probe in its sandbox. setup-check-claude-instructions now expects Core's warning and the exact @../AGENTS.md line. Neither change has been recorded.

Pinned semantic evidence

The semantic evidence was recorded on Skill 8.2.3 with CLI 11.0.6. None of the 8.3.0 changes was re-evaluated. Whether a coding agent uses the new helper modes or follows the acceptance rule is not established by that evidence, and deterministic checks do not establish the new rules' behavioral reliability.

The selected bundle remains semantic-sem-20261018051032668-1465f9ee.json.gz under evidence-8.2.3-20261018: 120 cases, 115 passed and five recovered, with the same two disclosed exclusions. Its compressed SHA-256 is baa883450e1087efafbf3c685e311859fb63575dfa839761c20db48de292342d. No paid run, regrading or evidence repacking was performed for 8.3.0.

Install this exact release:

npx skills add "moldea-ai/skill#v8.3.0"

moldea 8.2.8

Skill 8.2.8 closes gaps an external review found in 8.2.7: repair Git evidence in partial clones and on older Git, Claude Code integration checks, and CLI setup in Yarn workspaces and when the CLI was declared as a regular dependency. It retains CLI 11.0.6, Node.js >=22.17.0, Git >=2.30.0, every numeric limit and the existing continuation behavior.

Changes

• The repair Git helper never fetches. Every helper Git read disables lazy fetching and allows no transport protocol, so an object missing from a partial clone becomes unavailable evidence and repair reassesses every completed owner instead of contacting the remote.
• The file monitor is disabled on every supported Git. The helper's override is now core.fsmonitor= (empty). Git 2.35 and earlier treated the previous false value as the name of a program to run.
• Claude Code integration is checked. When a repository has a root CLAUDE.md, .claude/CLAUDE.md or CLAUDE.local.md, Claude Code does not load AGENTS.md by default unless one of those files imports it or is a symbolic link to it. Initialization, setup checks and repair now check this. Imports resolve relative to the file that contains them, so a root file uses @AGENTS.md and .claude/CLAUDE.md uses @../AGENTS.md. A shared file must carry the integration, because CLAUDE.local.md covers only its owner. The coding agent adds the line only when host instructions permit editing that file, never writes through a link, and otherwise reports the exact line and file.
• Yarn Classic workspace roots. Yarn Classic 1.x installs and removes with -W, which a workspace root requires and other projects accept.
• Moving a CLI declared as a regular dependency. With every manager, a CLI declared outside root devDependencies is first removed with the manager's targeted removal and then added as an exact root development dependency, both with scripts disabled. Adding alone left the declaration in dependencies with Yarn Classic and was refused by Yarn 4.

Verification

• The repair Git helper's integration tests run on Linux, Windows and macOS and, in a new required conformance job, on Debian Bullseye's Git 2.30. In a blobless partial clone, continuation reports unavailable evidence and leaves Git storage unchanged, while a plain Git read of the same object does fetch. A stored file-monitor index extension never runs a false program from PATH. Debian's Git 2.30 package backports GIT_NO_LAZY_FETCH, so each no-fetch setting was also shown to block the fetch on its own.
• The package-manager harness now installs Yarn at a real workspace root and moves a CLI declared in dependencies for every manager, preserving an unrelated dependency's declaration and lock entries. Before this release it passed locally against the published CLI 11.0.6 closure on every release-candidate entry (npm 7 through 12, pnpm 8.3.1 through 12.3.1 including Vitest 3.2.4 coexistence, Yarn 1.22.22, 4.14.1 and 4.18.0) and Yarn 3.8.7, and a Yarn Classic workspace run without -W failed as expected.
• Evaluator protocol 50 adds the setup-check-claude-instructions scenario for the .claude/CLAUDE.md variant and tightens the ordinary final-validation exemption so an unreadable patch never earns it. Neither has been recorded.

Pinned semantic evidence

The semantic evidence was recorded on Skill 8.2.3 with CLI 11.0.6. None of the 8.2.8 changes was re-evaluated. Whether a coding agent checks Claude Code's files or chooses the correct import or package-manager command is not established by that evidence, and deterministic checks do not establish the new rules' behavioral reliability.

The selected bundle remains semantic-sem-20261018051032668-1465f9ee.json.gz under evidence-8.2.3-20261018: 120 cases, 115 passed and five recovered, with the same two disclosed exclusions. Its compressed SHA-256 is baa883450e1087efafbf3c685e311859fb63575dfa839761c20db48de292342d. No paid run, regrading or evidence repacking was performed for 8.2.8.

Install this exact release:

npx skills add "moldea-ai/skill#v8.2.8"