systems-check

name: systems-check description: Use when the user wants the mission code review phase, or when /mission dispatches it. Trigger on "systems-check " or "/systems-check". Thin wrapper around systems-check-workflow.js — language-bucketed Systems Inspectors review the full branch diff, repair Astronauts fix actionable findings; on exhausted rounds asks the user whether to continue, skip, or stop. Requires a plan from /pre-launch.

Systems Check — Review and Repair

Run the systems-check workflow, looping interactively when repair rounds are exhausted.

Step 1: Locate the plan

ISSUE_NUM=…   # first non-flag argument; if omitted, infer from branch: claude/issue-<N>-*
STATE_DIR="${CLAUDE_PLUGIN_DATA}/mission-runs/issue-${ISSUE_NUM}"
[ -f "$STATE_DIR/plan.json" ] || { echo "No flight plan for issue #${ISSUE_NUM} — run /pre-launch ${ISSUE_NUM} first."; exit 1; }
PLAN=$(cat "$STATE_DIR/plan.json")
WORKTREE_PATH=$(echo "$PLAN" | jq -r '.worktree_path')
REPO=$(echo "$PLAN" | jq -r '.repo')

Step 2: Resolve models

Built-in defaults: director=fable, astronaut=sonnet, controller=sonnet, inspector=fable, capcom=sonnet, docking=sonnet, utility=haiku.

1. If .claude/mission.local.md exists at the repo root, Read it and take any entries under models: in its YAML frontmatter.
2. If the invocation included --models role=value,..., apply those entries on top.
3. Valid roles: director, astronaut, controller, inspector, capcom, docking, utility. Valid values: haiku, sonnet, opus, fable. Warn about and ignore any invalid entry — never abort over one.
4. Fable fallback. Fable is unavailable in some environments (e.g. headless or cron runs), where spawning an agent with it hard-errors. Before launching, for any role whose resolved value is fable, confirm Fable is among this session's available models. If you cannot confirm it, downgrade that role — director→opus, inspector→sonnet, every other role→sonnet — and tell the user which roles were downgraded. Never pass fable to an agent you cannot confirm supports it.

The merged result is MODELS, passed to the workflow as args.models.

Step 3: Enter the worktree

[ -d "$WORKTREE_PATH" ] || { echo "Worktree missing at $WORKTREE_PATH — re-run /pre-launch ${ISSUE_NUM}."; exit 1; }

Call EnterWorktree with path: $WORKTREE_PATH.

Step 4: Inspection loop

Initialize: SC_DEFERRED = [] (accumulates low-confidence findings), SC_MAX_ROUNDS = 3.

Loop:

1. PRIOR=$(cat "$STATE_DIR/sc.runid" 2>/dev/null || echo "")
   

   Call the Workflow tool with:

   • scriptPath: the literal string ${CLAUDE_PLUGIN_ROOT}/workflows/systems-check-workflow.js (expand the env var — do NOT use import() or cat)
   • resumeFromRunId: PRIOR if non-empty (resumes an interrupted run), otherwise omit
   • args: { issue_number: <ISSUE_NUM>, repo: "<REPO>", plan: <plan.json object>, initial_deferred: <SC_DEFERRED>, max_rounds: <SC_MAX_ROUNDS>, models: <MODELS>, plugin_root: "<value of $CLAUDE_PLUGIN_ROOT>" }

   Save the returned runId while the workflow runs, and clear it once the run completes (each loop iteration must start fresh):

   echo "<runId>" > "$STATE_DIR/sc.runid"     # before/while running
   rm -f "$STATE_DIR/sc.runid"                 # after the run returns
   

   If the workflow throws, present the error using the banner shape in references/halt-protocol.md — options: [1] fix the stated problem and re-run /systems-check , [2] /pre-launch --replan if the plan itself is wrong.

2. If result.status === 'clean': break.

3. If result.status === 'exhausted':

   • Summarise result.open_findings: [<severity>] <file>:<line> — <summary> (<confidence>% confident)
   • AskUserQuestion: Try more rounds / Skip and continue / Stop.
   • Try more rounds → ask how many (default 3), set SC_MAX_ROUNDS, append result.low_confidence_findings into SC_DEFERRED (dedup by file+summary), loop.
   • Skip and continue → break; note the open findings need manual attention.
   • Stop → report the open findings and exit without advancing.

Step 5: Report

Systems check complete for issue #<N>.
Next: /docking <N>  (or /mission <N> drives it automatically)

If the final result carried low_confidence_findings, list them:

Low-confidence findings not auto-fixed (<N>) — review manually:
  [<severity>] <file>:<line> — <summary> (<confidence>% confident)