fix(guard): note each unresolved command once

[AbysmalBiscuit]

TL;DR (human written)

---

Closes #196

---

Agent summary

The pre-tool-use guard repeated its "could not be fully resolved" notes once per loop iteration, and added the tasks-and-apps note to commands nothing in the project runs. A 4-iteration loop with five segments came back with 20 notes. Most of them named printf, jq or node, so the agent could not act on them.

Fix

• decide keeps each distinct finding once per hook call. Blocks go through the same path, so cargo build repeated in a loop returns one deny reason instead of one per iteration.
• The tasks-and-apps note now needs a program something could claim: a task run, an app launch, or a catalog dev server such as vite. Both sides are unwrapped the way the matcher does it, so a doppler run -- cargo build task still counts as cargo. When the program word itself is unresolved, the note stays whenever the project has tasks or apps.

That same loop now yields 2 notes: one for cargo build "$(date)" and one for the [harness.commands] rule node "$y" might hit.

A side effect: vite "$x" in a project with no tasks or apps now gets the note. The guard denies a resolved vite there too, so the note is accurate.

Tests

Three end-to-end tests in tests/harness_guard.rs send a hook payload through the devkit hook pre-tool-use binary:

• The loop test failed before the fix with 9 notes, 3 of them distinct.
• The printf/jq test failed before the fix because a printf note came back.
• The keep-note test (cargo build "$x", nitro "$x", vite "$x", "$tool" build) passes both before and after. It checks that the fix does not drop notes it should keep.

Not fixed here

The guard redirects a filtered test run to a task: cargo nextest run --test harness_guard to test, and cargo test --test harness_guard to test-doc, which runs only doctests. A filtered run is not the task. That belongs in its own issue.

---

🤖 Generated with Claude Opus 5.5 (claude-opus-5-5) via Claude Code