Task evidence

Task records#

A task is either inspection or execution. “Continue” and status questions keep the current task. Explicit inspection stays read-only for recognised edit operations until you ask for changes.

honey task status --json
honey task set --mode execute --operation code --cwd /actual/project --check tests --check lint
honey task verify --check tests --tool-call <actual-test-call-id>
honey task verify --check lint --tool-call <actual-lint-call-id>

Codex uses CODEX_THREAD_ID; otherwise add --agent claude --session <session-id>. Bind the project before running its checks. task set starts a new contract and can't promote an inspection task to execution without your request.

What counts as proof#

The default code check recognises direct test and lint commands. Explicit checks link a fresh, successful result — the agent still has to confirm the command proves the named requirement. Later edits, potentially mutating shell commands, failed reruns and incomplete results invalidate proof. Checks started before a mutation, or in a previous task, can't verify the new state.

Completion claims without evidence trigger one reminder per evidence cycle. Honest unresolved reports and inspection answers are allowed.

Uninstall and startup operations#

--operation uninstall requires the checks application, helpers, startup and preserved-data. Declare concrete targets with --target; --absent and --present measure and recheck the real state.

honey task verify --check application --absent /Applications/Example.app

--operation startup requires its resulting registration to be checked.

Hosts that hide tool events#

When a host doesn't deliver inner events for a wrapped tool, run the check through honey. It executes without a shell and records the exit status plus a source snapshot.

honey task set --mode execute --operation code --cwd /path/to/project
honey task run --check verification -- python3 tests/run.py

Printed success can't manufacture proof: changed source, HEAD or index makes the result stale. The snapshot covers tracked and non-ignored files (up to 20,000 files or 64 MiB), rejects source symlinks, and doesn't claim to cover ignored dependencies or the external environment.

Metrics#

honey task status shows task and session tool counts, repeated reads and checks, failures, unknown results and reported durations. /rsi coach aggregates them with denominators. Missing durations are never estimated, and these observations are not measured speedups.