Bug: daemon caches resolved binary path across Claude upgrades (stale symlink)

[c-h-]

What happened

After upgrading Claude Code from 2.1.52 to 2.1.70, the agentctl daemon kept trying to spawn /Users/ms/.local/share/claude/versions/2.1.52 — the path it resolved at startup. New session launches failed with:

[claude-code] spawn error: spawn /Users/ms/.local/share/claude/versions/2.1.52 ENOENT

The symlink at ~/.local/bin/claude correctly pointed to 2.1.70, but the daemon's in-memory cache (resolvedCache in resolve-binary.ts) held the old resolved realpath from when it first started.

What was expected

The daemon should either:

1. Clear the binary cache on each launch (not just daemon restart), or
2. Re-resolve symlinks periodically (e.g., on spawn failure), or
3. Not resolve through symlinks at all — spawn the symlink path directly and let the OS resolve it

How to reproduce

1. Start agentctl daemon with Claude Code 2.1.52 installed
2. Upgrade Claude Code to 2.1.70 (symlink updates, old binary may or may not remain)
3. Try agentctl launch claude-code — fails with ENOENT for old version path

Workaround

agentctl daemon restart clears the in-memory cache and re-resolves to the current symlink target.

Root cause

src/utils/resolve-binary.ts calls fs.realpath() to resolve symlinks, then caches the result in a Map for the daemon's lifetime. The cache is never invalidated except on process restart. clearBinaryCache() exists but is never called.

Proposed fix

Option A (simplest): On ENOENT spawn error, call clearBinaryCache() and retry once.
Option B: Don't realpath() — just verify the symlink is executable and spawn it directly.
Option C: Call clearBinaryCache() before each launch.

Environment

• agentctl: latest (linked from ~/personal/agentctl)
• macOS Darwin 25.2.0 arm64
• Claude Code upgraded 2.1.52 → 2.1.70

[c-h-]

Update: Intermittent spawn ENOENT even after daemon restart

The stale cache was the original issue, but after restarting the daemon (which re-resolves to the correct 2.1.70 path), the spawn still fails intermittently:

[claude-code] spawn error: spawn /Users/ms/.local/share/claude/versions/2.1.70 ENOENT

Verified:

• Binary exists: -rwxr-xr-x 197364336 bytes, Mach-O 64-bit executable arm64
• claude --version → 2.1.70 (Claude Code) works fine from shell
• ~/.local/bin/claude symlink → 2.1.70 is correct
• Daemon is running (fresh restart, PID confirmed)
• PATH in daemon-env.json includes ~/.local/bin

Behavior:

• agentctl launch returns successfully (Launched session pending-XXXX)
• Session immediately goes to stopped status
• Daemon stderr shows spawn ENOENT
• Sometimes a second launch works, sometimes not — it's intermittent
• The daemon itself crash-loops on the ENOENT when it kills the supervisor

Likely root cause:

Node.js child_process.spawn() ENOENT can mean:

1. The binary doesn't exist (not our case)
2. The binary's interpreter/loader can't be found (unlikely for native Mach-O)
3. A race condition in the daemon's process management — spawning before the previous cleanup completes?
4. The daemon's cwd for the spawn doesn't exist (the --cwd dir is valid though)

Impact:

Claude Code sessions are unreliable — multiple launch attempts needed, QA work blocked. This is not just the cache issue from the original report; there's a deeper spawn reliability problem.

Environment:

• agentctl 1.6.0 (linked from ~/personal/agentctl)
• Claude Code 2.1.70
• macOS Darwin 25.2.0 arm64
• Node v24.13.0