fix(daemon): correct Claude Code --add-dir capability detection (fix …

[1119302165]

Description

This PR fixes issue #430 where Claude Code failed to read skill files due to incorrect directory access policy enforcement.

Root Cause

The Claude Code agent definition was probing global claude --help output for capability flags, but --add-dir and --include-partial-messages are subcommand-specific flags that only appear under claude -p --help, not in the global help output.

This caused:

1. Capability detection to incorrectly set caps.addDir = false for users with working Claude Code installations
2. The --add-dir flag was therefore never passed on agent spawn
3. Claude Code's sandbox enforcement blocked all absolute-path reads outside the project cwd
4. Skill files, skill assets, and design-system specs could not be read by the agent

Note: The "probing failed / timeout" path already worked correctly, as it fell back to an empty capabilities object {} which defaulted to passing --add-dir. The failure mode only manifested when the help probe actually succeeded.

---

Impact Analysis

Scope

Aspect	Impact
Agents affected	Claude Code ONLY. All 12+ other agents (Codex, Cursor, Devin, Gemini, etc.) are completely untouched.
Code change surface	1 line changed in helpArgs + comment. No logic changes to any other agent.
API compatibility	No external API changes. helpArgs is stripped before serialization.
Backward compatibility	Fully backward compatible. Older Claude Code versions continue to work via fallback mechanism.

Execution Paths Affected

Path	Before	After
Daemon agent detection	execFile('claude', ['--help'])	execFile('claude', ['-p', '--help'])
Agent spawn buildArgs	--add-dir never passed when probe succeeded	--add-dir correctly passed when detected
/api/agents endpoint	helpArgs internal-only field stripped	No change — still stripped

---

Changes Made

File	Change
apps/daemon/src/agents.ts	Updated Claude Code helpArgs to probe subcommand help
apps/daemon/tests/agents.test.ts	Added 3 regression tests for issue #430

---

Regression Protection

Tests added:

1. ✅ Default --add-dir propagation when directories are supplied
2. ✅ Correct filtering of empty/null directory entries
3. ✅ Explicit assertion that ['-p', '--help'] is used for probing (prevents regression)

---

Testing

1. Before this fix: Running claude --help | grep --add-dir returns no matches
2. After this fix: Running claude -p --help | grep --add-dir correctly finds the flag
3. All existing agent tests continue to pass
4. New regression tests prevent this specific failure mode from recurring

[lefarcen]

Hi @1119302165! 🎉

Thanks for the contribution — this is a precise root-cause fix with thorough testing and impact analysis.

I will run a deep review and get back to you within 24h.

Thanks for making open-design better!
— open-design team

[mrcfps]

@1119302165 I reviewed the Claude Code capability-detection change in apps/daemon/src/agents.ts and the added regression coverage in apps/daemon/tests/agents.test.ts. The fix is tightly scoped to probing the claude -p --help subcommand where --add-dir is advertised, and the tests cover the default add-dir behavior, directory filtering, and the root-cause help arguments. I did not find any actionable issues in the changed ranges; thanks for the clear, focused fix and the helpful regression protection. 🙌

_{Generated by Looper 0.5.2 · runner=reviewer · agent=opencode}

[lefarcen]

I've reviewed the changes as promised. @mrcfps's review above covers all the code correctness and test coverage points comprehensively — the fix is tightly scoped and well-protected by regression tests.

No additional actionable issues from my end. Thanks for the clear, focused fix! 🎉

[lefarcen]

🎉 Merged! Thanks @1119302165 for the precise root-cause fix and thorough regression tests.

The capability detection now correctly probes claude -p --help where --add-dir is actually advertised, and the tests guard against this exact failure mode recurring.

Appreciate the clear analysis and careful implementation! 🙏