feat(setup): probe prek and uv in the sandbox doctor

[potiuk]

Summary

• The doctor skill's probes are meant to move in lock-step with docs/setup/sandbox-troubleshooting.md, but fix(agent-isolation): let prek and uv run inside the sandbox #1359 added the prek or uv not found inside the sandbox entry without a matching probe. Probe 7 (probe-7-dev-tools.sh) now covers it:
  • ✓ when prek / uv run inside the sandbox and ~/.cache is writable;
  • ✗ when they start but cannot write their cache.
  • Inside the sandbox a read-denied ~/.local/bin looks exactly like an absent one. So when neither tool is found, the probe reads the worktree's settings.local.json (readable in the sandbox) to tell a missing grant (⚠) from a tool that is not installed (⊘).
• Follow-ups from fix(agent-isolation): let prek and uv run inside the sandbox #1359:
  • The error-hint hook and secure-agent-setup.md now count seven probes.
  • The setup doc's hint-signature table gains the command not found: prek|uv|uvx row.
  • The troubleshooting entry no longer says the new paths apply only "from the next session". Measured: the harness applies settings.local.json changes from the next sandboxed command.

Type of change

• Skill change (.claude/skills/<name>/) — eval fixtures updated below
• Tool / bridge contract (tools/<system>/*.md)
• Python package (tools/*/ with pyproject.toml)
• Groovy reference impl
• Cross-cutting (RFC, AGENTS.md, sandbox, privacy-LLM)
• Documentation (docs/, README.md, CONTRIBUTING.md)
• Project template (projects/_template/)
• CI / dev loop (prek, workflows, validators)
• Other:

Test plan

• prek run --all-files, run inside the sandbox: every hook passes except the workspace pytest hook. Two suites fail there for sandbox reasons, and both pass outside it:
  • agent-isolation's test_agent_socket_rows_follow_the_connections_held_open opens a Unix socket, which the Linux sandbox refuses.
  • pr-management-stats calls the real gh.
• Probe 7 run by hand in all four states on Linux (bubblewrap, Claude Code 2.1.280):
  • ✓: tools found, cache writable.
  • ✗: HOME pointing at an unwritable directory.
  • ⚠: tools off PATH, grant absent.
  • ⊘: tools off PATH, grant present.
• For skill changes: setup-isolated-setup-doctor eval suite, 23/23 with magpie-run-evals.sh (claude -p).
• For skill behaviour changes: interpret-probes grows a dev_tools_status field (skip when no dev-tools line is present, so existing fixtures expect skip) and the new fixture case-16-dev-tools-not-granted.
• Other:

RFC-AI-0004 compliance

• Sandbox — the probe is read-only apart from touching and removing one file in ~/.cache, which the reference config already makes writable. It adds no host access.

Linked issues

Refs #1359

Notes for reviewers (optional)

• Probe scope: the probes still don't cover every catalogue entry. The signing-overlay entries need a live touch or a terminal, and setup-isolated-setup-verify check 10 covers them. The doctor's wording now says so instead of claiming the probes cover "the six failure modes the catalog documents".
• Token and fingerprint regeneration: docs/mode-economics.md was regenerated with skill-token-count --write. The doctor's surface_hash was refreshed by the skill-surface-hash hook because its description now names prek and uv.

---

Generative AI disclosure: this change was prepared with Claude Code (Claude Opus 5) and reviewed by the author before submission.

🤖 Generated with Claude Code