task

/audit:task — add a task to the manifest, promote one, close one, re-open one, move one between phases, block one, note on one, or cancel one

Every verb is one call to the writer, which takes the index lock, revalidates, rolls back on a finding and journals the write itself - so nothing here is a hand edit and no lock is held by hand. T below is python3 "${CLAUDE_PLUGIN_ROOT}/scripts/manifest/audit-task.py". An unknown or empty verb: print the hint and stop.

The flags are the interface; the dialogue is the fallback. Pass what the caller already decided as flags, and ask only for what is missing. A flag another verb reads exits 2 naming that verb (scope --outcome is refused); relay it rather than retrying.

The operator's words go in VERBATIM: a --reason is theirs, unparaphrased, and reaches the hash-chained journal, as do --text and --description; - reads the value off stdin. Print the writer's output as it came - a line naming what was written and the file, then any gate basis, ready now command and warning about the id it wrote; --verbose prints the rest.

Exit codes, every verb: 0 written. 1 the write would leave the plan invalid - rolled back, the findings printed; fix the inputs. 2 usage - the message names the choices; ask, adjust, re-run. 3 the index lock is held by a live run - stop, do not take it over. 4 the lock looks abandoned - ask the human (AskUserQuestion), then re-run with --takeover.

Subcommand: add "<title>" [--phase <id>]

1. Phase. Pass --phase when given. Without it the writer takes the one running phase, or exits 2 naming the choices: ask which (AskUserQuestion), or offer /audit:phase add for a new one - on a phase branch, /audit:phase add ... --park. A parked proposal already covering the work → offer /audit:propose materialize <PROP-id> first.
2. Gather the answers (ask only for what's missing): --description (the brief - on stdin with --description - and a quoted heredoc whenever it holds backticks or code punctuation, since a shell eats a backticked span and the writer refuses the hole it leaves); --files; --outputs for files the task produces; --tests-mode tdd|regression|gate-only; --tests-add "<path>: <what it asserts>" - the leading path is what reaches files and fileIndex; --gate only to override the derived gate, --gate-clear for none; --failing-from <runId> for a fix task after a red sign-off run; --risk; --model (sonnet is the floor for fix work, never haiku); --blocked-by / --depends-on. Skills come from one source:

   python3 "${CLAUDE_PLUGIN_ROOT}/scripts/status/audit-status.py" --json --discovery --section discovery
   

   Offer the area defaults, names that output carries (never invented ones), and null (none applies; it stops the area fallback); then --skills a,b, --skills null, or omit the flag.
3. Run it, the brief on stdin:

   T add "<title>" --phase <id> --description - --files a,b --tests-mode regression <<'BRIEF'
   <why & how>
   BRIEF
   

   --dry-run validates and writes nothing. --fixes <findingId> adds a fix task for review findings - the drive's triage does that itself. --bug <bugId> makes it the fix task for a bug and writes both halves of the link in the one write - the task's bugId in its shard, the bug's taskId and status: in_progress in the index; a bug another live task is already fixing is refused by name (/audit:bug fix is the usual door).