gate

name	gate
description	Keep the workspace green while you edit — the free autofix hooks, and delegating the check-and-fix loop to the tiered camas-fixer ladder so residuals never spend your reasoning. Read after a batch of edits and before declaring work done.

The camas gate keeps the workspace green as you edit, in two layers — both driven by what the project declares in Config.agent.

Fix (automatic, free). A PostToolBatch hook runs camas mcp fix after each edit batch, and a Stop hook runs it again as you finish your turn — the node the project registered as Config.agent.fix (its mutating, behavior-preserving auto-fixers: formatters, --fix linters), scoped to the just-changed files. Both run at zero model tokens and never ask you anything; with no fix registered they are a no-op.

Check (you delegate — to a tiered ladder). The check node (Config.agent.check, else the default task) is read-only: it runs the project's checks and classifies the result green or needs_reasoning. After a batch of edits, and before you declare work done, delegate the check-and-fix loop to the camas-fixer ladder rather than running the checks and chasing residuals in your own context. Subagents cannot nest — a haiku subagent cannot spawn a sonnet one — so you orchestrate the escalation yourself:

• For a lint/format residual (the diagnostic names formatters or --fix linters): spawn camas-lint-fixer-haiku with the changed paths as its scope. It takes one pass and hands back. If it hands back not green, escalate by spawning camas-lint-fixer-sonnet on the same scope — it also takes one pass. If that hands back not green, the residual is yours: take it from there.
• For a test/coverage residual (the diagnostic names a test runner or coverage tool): spawn camas-test-fixer directly — it fixes tests and coverage on a capable model in one delegation, ending by re-gating to confirm green. If it hands back not green, or says the fix is ambiguous (test wrong vs. behavior wrong), take it from there yourself.
• When diagnostics span both kinds, or you can't tell from the names, start with the lint ladder (cheaper) and send whatever remains to camas-test-fixer.

Run each fixer in the background and keep working; for independent changed scopes, spawn one per scope so they run in parallel. Each fixer hands back only what it could not settle: a green result means that scope is done; a result quoting remaining diagnostics is the residual that needs your reasoning.

The Stop-hook nudge. If your turn ends before you delegated (or a fixer's residual never got picked up), a background check runs after you stop and, if the workspace isn't green, wakes you with a reminder to launch the fixer ladder — so a skipped check-and-fix loop doesn't go silently unnoticed. It wakes you at most once per prompt and never for a configuration gap (no check node registered), so it cannot loop your turn. Prefer delegating proactively per the ladder above; treat the nudge as a backstop, not your primary signal.

Never mask a residual — yours or a fixer's: do not suppress, disable, or loosen a check to make the gate pass. A green gate must mean the work is actually correct.