fix(pipeline): scratch files leaked into agent commits; install a logger

[ANonABento]

Both bugs found by actually running an agent from a column end to end — dogfooding #275. Neither was visible any other way.

The git exclude never applied in a worktree

execute_spawn_cli built the exclude path by joining .git/info/exclude onto the working dir, then guarded on .exists(). In a linked worktree .git is a file, not a directory, so that path never exists — the guard found nothing and silently skipped the whole exclusion.

So every worktree-based agent run has been committing .task.md. This is pre-existing, not new; .agent.md had just inherited it. First dogfood run:

 .agent.md       | 25 +++++++++++++++++++++++++
 .task.md        |  8 ++++++++
 agent-proof.txt |  1 +
 math.js         |  2 +-

Only the last two belong there.

git rev-parse --git-path info/exclude resolves the real common-dir file for both a plain repo and a linked worktree. exclude_from_git uses it, is idempotent per pattern, appends only what's missing, creates the parent dir, and no-ops outside a repo. Same pipeline re-run after the fix:

 agent-proof.txt | 1 +
 math.js         | 2 +-

Nothing was installing a logger

The log crate is a silent no-op until something registers one, and nothing did — no tauri-plugin-log, no env_logger, no set_logger anywhere in the tree. All 202 call sites were discarded, including 19 log::error!:

level	sites
error	19
warn	117
info	62
debug	4

That is exactly what made the first dogfood failure undiagnosable: a trigger that failed and a trigger that never fired look identical from outside. I spent several rounds guessing at DB state before realising the diagnostics I was looking for had never been emitted.

env_logger now initializes in run(), defaulting to kaitencode=warn, controllable via RUST_LOG. Turning it on immediately printed the full trigger path — argv, working dir, tmux wait-for channel — and everything after that was straightforward.

What the dogfood run proved about #275

Real agent, real column, real repo, isolated data dir:

• attached agent supplied the CLI and its model — argv showed --model sonnet from the agent config, not the workspace default
• .agent.md written and followed: the agent wrote the proof token the instructions asked for
• the attached skill was used — it ran the verification command and reported add(2, 3) → 5
• exit 0 → agent_complete → auto-advance Todo → Fix → Done
• worktree cleaned up on reaching the terminal column
• the planted bug was actually fixed: return a - b → return a + b

Roadmap

Critical-now is now empty — the stale /usr/local/bin/claude is gone (verified: which -a claude returns one path, 2.1.239).

Recorded a new env-hygiene note: the same duplicate-binary shape now exists for codex. /usr/local/bin/codex is 0.130.0 while fnm's shim is 0.145.0 — the version every codex behaviour in this repo was verified against. The fnm path is per-shell and lives in tmpfs, so a reboot or a desktop-launcher start hands the app the stale one. Checked: not currently breaking anything — both carry --skip-git-repo-check and --dangerously-bypass-approvals-and-sandbox, and both have resume --last. Logged as latent, not urgent.

Checks

cargo clippy -D warnings · cargo test 559 (+2) · tsc · eslint · test:ipc · test:type-scale · vitest 444.