[knihit]

Description

On Kiro, --skill-name-filter filtered nothing. Every agent's skills server indexed and served all 82 installed skills instead of only the agent's declared ones. This PR makes the filter work. The stacked PR on feat/agent-skill-index builds on it to get agents to load their skills unprompted.

MCP skill server:

• install_manifest.rs: read_installed_paths now reads the v2 manifest the installer actually writes (strategies[].files[].path, unioned across strategies). It still reads the v1 flat files[].path, because an older install's v1 manifest stays on disk until the next install migrates it. Before this, the reader only knew v1. So it found no installed paths, treated every skill as hand-authored, and the name filter skipped all of them.
• Fail-open is unchanged: any malformed entry, or a strategy without files, still gives an empty set.
• A missing manifest stays silent. A manifest that exists but can't be read (permission denied, a directory at the path, other I/O errors), isn't valid JSON, or has no recognized shape logs one skill-lookup-mcp: WARN naming the manifest path and saying --skill-name-filter will not apply under that root. read_installed_paths returns a Result whose read error carries the io::ErrorKind, and only installed_paths_for_root logs.

Tests:

• Fixtures in index.rs, scanner.rs, initialize_handshake.rs and reload_skills_shakedown.rs used the v1 shape, which is why the bug went unnoticed. They now use v2.
• New unit tests cover a v2 manifest, two strategies, and a strategy without files.
• New e2e test in get_skill_shakedown.rs: the server runs with a v2 manifest and --skill-name-filter keep-me. reload_skills reports filtered_out: 1, get_skill drop-me returns not-found, and keep-me and a hand-authored skill are still served.
• With the reader change reverted, the new e2e test and three existing filter tests fail.
• Shared contract fixture mcp/lib/skill-lookup-core/testdata/install-manifest.v2.json, so the reader's fixtures can't drift from the writer again. In cli/konductor-rs, serialized_manifest_matches_shared_contract_fixture builds a manifest from the real types, runs the production serialize(), and asserts it equals the file byte for byte. In skill-lookup-core, read_installed_paths_parses_installer_contract_fixture asserts the paths read from the same file. Both reach it through CARGO_MANIFEST_DIR, with no cross-workspace dependency.
• installed_paths_for_root has a WARN test for each logged case (unrecognized shape, invalid JSON, a directory at the manifest path) and a test that a missing manifest logs nothing. They assert on the manifest message, not on the whole log, since other tests in the crate log through HOME without the HOME lock. HOME is restored by HomeGuard, now shared from test_home_lock, so a panic can't leak the override. The crate's tests passed 8 of 8 runs with --test-threads=16.

Changelog:

• One bullet under [Unreleased] / Fixed. No VERSION change.

Type of Change

• Other (describe below): bug fix in the skills MCP server (mcp/)

Testing

• npm test: 119 of 136 pass. The same 17 fail on main (claude-code-agent-runner, generate-agent-files, run-claude-subset), and this PR touches none of their code. Not rerun after the review fixes, which change no JS.

• cd cli && make test: 1780 passed, 0 failed. One test that runs synth from a deleted cwd failed once under parallel load on macOS, then passed alone 3/3 and on a full rerun.

• mcp make test: 327 passed, 0 failed (skill-lookup-core 176).

• make -C cli lint passes. make synth leaves no tracked changes. make test-skills-sh passed, make claude-plugin-check pass, make test-py 222 passed, kiro-power-check passed

• Smoke tested: make build && make synth, then install --from . --harness kiro-cli-v2 --no-telemetry into a scratch target and HOME. Running k-developer's skill server with its installed args over stdio, reload_skills gave indexed: 27, filtered_out: 55 (before: indexed: 82, filtered_out: 0). get_skill adr-generator returned "skill not found", and get_skill code-review returned its body.

• Benchmark: not run on this branch alone. It was run on the stacked feat/agent-skill-index branch, which includes this fix.

  Setup: same tasks T1 to T4 plus a no-skill control, 2 reps each, project-scope install with --no-telemetry, isolated HOME per run, scored from session logs.

  Harness	Variant	n	Loaded a skill	Right skill
  Kiro CLI v3	main	12	4	3
  Kiro CLI v3	stacked branch	12	9	9
  Claude Code	main	12	5	5
  Claude Code	stacked branch, with a Claude line it later dropped	12	6	6

  The stacked PR's prompt block is Kiro-only. The filter's effect on its own is narrower than the table: each Kiro agent's server now offers only that agent's skills. 0 of 28 Kiro runs reported the skills server unavailable, and no get_skill call failed. Caveats: 2 runs per task cell, and main varied by about 1 run in 12 between two benchmark passes.

  After the review fixes, the stacked branch was re-benched on Kiro CLI v3 with two more reps per task: pooled right skill 13 of 22 on its final head, against 15 of 23 before the review fixes (within noise) and 3 of 12 on main. None of those runs reported the skills server unavailable or had a get_skill call fail. See the stacked PR for the table. The fixes on this branch don't change which skills the server serves.

• No new code files. Existing files keep their SPDX headers. The new JSON fixture has no comment syntax, so it carries no header.

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own changes
• I have commented my code where necessary (doc comments name both manifest shapes)
• I have made corresponding changes to the documentation (CHANGELOG, and mcp/README.md: the --skill-name-filter row says only manifest-listed skills can be excluded and that a broken manifest turns the filter off with a WARN. .gitattributes pins the byte-compared manifest fixture to LF, so the writer test passes on Windows checkouts with core.autocrlf.)
• My changes generate no new warnings: make -C mcp lint clippy -D warnings is clean. cargo fmt --check only flags lines unchanged from main, which fail the same way on main with rustfmt 1.9.0.
• I have added or updated tests that prove my change works

Agents Affected

Kiro CLI only. Claude Code installs have no skills MCP server. Each agent's skills server now serves only that agent's declared skills plus hand-authored ones:

• konductor
• konductor-mux-orchestrator
• konductor-cmux-orchestrator
• k-developer
• k-architect
• k-researcher
• k-product-manager
• k-quality-assurance
• k-tpm
• k-browser
• k-media-analyzer

Notes for Reviewers

• Install is the only producer of --skill-name-filter, and no runtime caller relied on the unfiltered set. Every specialist's prompt and SOPs reference only its declared skills. The orchestrators' undeclared references are all delegated to a specialist or conditional.
• One soft fallback: k-plan.sop.md lets konductor apply legacy-to-agentic-estimate itself if k-tpm is unreachable. get_skill now returns not-found for it, and the SOP's next step already says to skip the projection, so it degrades gracefully.
• Six commits. The second restores two existing tests in get_skill_shakedown.rs that a bad save garbled in the first. Every pre-existing line in that file matches main. The rest address review: the shared manifest fixture, the WARN on an unrecognized manifest, the shared HomeGuard, and the WARN on an unreadable or invalid manifest.
• No conflict with fix(orchestrator): tighten the direct-work rules #114.