[keys-i]

Summary

Closes #21.

• make the selected manual independently focusable and scrollable by keyboard and mouse
• simplify the catalog, debounce directory filtering, and keep the existing pinned Textual stack
• repair default marketplace previews/applies with the supported all-agent target and pin Skills CLI 1.5.22
• add safe developer-clutter discovery for manifest-backed outputs and language/package caches
• show Linux/macOS filesystem and temporary-directory targets in every dry run

Scope

• bin/
• dev/
• lib/
• sys/
• Documentation or repository settings

Testing

Commands run:

uvx ruff@0.16.0 format --check --target-version=py311 ...
uvx ruff@0.16.0 check --target-version=py311 ...
uv run dev/gui.py --self-test
uv run --script bin/agent --self-test
uv run --script bin/cleaner --self-test
zsh -n sys/macos/clean && zsh sys/macos/clean --self-test
sh -n sys/linux/clean && bash sys/linux/clean --self-test
bash run.sh --self-test
bash -s -- --self-test < run.sh
npx --yes skills@1.5.22 --version
npx --yes skills@1.5.22 add vercel-labs/agent-skills --agent '*' --list

All listed checks passed, along with every Python command self-test, uv audit, Bandit 1.9.4, and Prettier 3.9.6. The live marketplace command listed nine skills and did not install them. Color and NO_COLOR=1 wide dashboard renders were visually reviewed. A read-only developer inventory scanned 77,286 entries / 1.8 GiB of findings in 1.46 seconds; no deletion was applied.

Comparable dashboard self-test measurements on the same macOS host used hyperfine --warmup 1 --runs 7 against the merge base and final tree:

Measurement	Merge base	Final	Delta
Mean time	4.435 ± 0.278 s	6.375 ± 0.248 s	+1.939 s / +43.73%
Range	4.041–4.759 s	6.011–6.678 s	—
dev/gui.py size	85,381 B	88,027 B	+2,646 B / +3.10%

The timing increase reflects expanded self-test coverage; it is not a performance improvement claim.

Tested systems and interpreter versions:

macOS host; local uv-selected Python 3.14
Linux shellcheck and Windows PowerShell remain CI-only

Checklist

• The change is focused and contains no unrelated files.
• The code is clean and comments explain intent or risk.
• New or changed behaviour has a repeatable test or documented manual check.
• Help text and dependencies are current.
• Destructive behaviour and elevated permissions are clearly disclosed.
• No credentials, private data, or machine-specific paths are included.

Reviewer Notes

No UI dependency was added; Textual remains pinned at 8.2.8. Cleanup remains preview-first, home-bounded, same-filesystem, and symlink-rejecting. The macOS apply path previews, confirms, rediscovers, and only then deletes. Rollback is a revert of commits 8d4fe60 and a3a2254.

The mandatory fixed-model gatekeeper was unavailable for this ChatGPT account. The contract-authorized gpt-5.6-sol ultra fallback reviewed the final diff and evidence and returned VERDICT: PASS with no blocking findings.