wf

Runs a task as a dynamic workflow with fixed cost defaults (agent cap, a cheap model for scan stages, a mid model for verify stages), or reports what recent workflow runs cost per phase and model so the defaults can be tuned. Claude Code only: it needs the Workflow tool.

Variables this skill expects

Rules

• Invoking /wf is the user's explicit opt-in to a workflow; no other opt-in is needed.
• Anything the user names in the task (models, size, scope) overrides the defaults below.
• Read-only unless the task asks for changes. Nothing that changes {{PROTECTED_SYSTEMS}}. Why: workflow agents run unattended and cannot ask mid-run.
• Agents in one phase share model, effort, agentType and schema, so they share the prompt cache.

Mode: cost

If the arguments start with cost, run:

python3 -I {{CORE_DIR}}/scripts/cost.py <N> {{PROJECTS_DIRS}} --scan {{SCAN_MODEL}} --verify {{VERIFY_MODEL}} --max {{MAX_AGENTS}}

N is the number after cost (default 1). Show the table as-is, then at most three lines: which phase cost most, what the ! flags say (a stage off its model, a run over the cap), and one concrete change for next time. Cache reads bill at about a tenth of fresh input, so judge cost by cache writes and output first.

Mode: run (any other arguments)

The arguments are the task.

0. Scout inline (a few Glob/Grep/ls calls, no agents) to find the real work-list. If the task fits one session (roughly under 5 items or one file), say so and do it inline; no workflow, and skip step 1.
1. Elaborate and confirm, inline (no agent: the scout already holds the context). Rewrite the task as a precise prompt (goal, scope and where to look, what counts as done, output shape, constraints, anything ambiguous). Ask the user (AskUserQuestion if available): use it, edit it, or use the original. Put the full elaborated prompt in the preview of the "use it" option and the original task in the preview of "use the original". Why: while the question dialog is open the user sees only the dialog, not text written before it, so a prompt shown only in chat is invisible at the moment they choose. Without AskUserQuestion, print the prompt in a fenced block and end the turn with the question. Wait for the answer; the chosen text is the task from here on. Skip this step if the user says "no elaborate" in the arguments.
2. Load the workflow-authoring skill.
3. Show the plan in 3-5 lines: phases, agents per phase, model per phase, rough tokens (about 40k setup per agent plus the work). Then launch; the Workflow permission prompt is the gate.
4. Write the script with these defaults:
   • At most {{MAX_AGENTS}} agents. For long lists, batch items per agent and log() anything dropped. Enforce it in code: route every call through this guard, never agent() directly: let n = 0; const run = (p, o) => { if (++n > {{MAX_AGENTS}}) throw new Error('agent cap'); return agent(p, o) }
   • Phase titles start with their stage word (Scan, Search, List, Fetch, Verify, Edit, Write, Judge, Synthesize), e.g. Scan repos. Why: Mode: cost checks each phase's model by that word and flags a phase called "Hunt" or "Recon" as ! no stage word.
   • Verify and judge in batches of 5-10 items per agent, never one agent per item. Why: one verifier per finding is what blows past the cap.
   • Per run() call: scan, search, list, fetch → model: '{{SCAN_MODEL}}', effort: '{{SCAN_EFFORT}}'; verify, judge, edit and write → model: '{{VERIFY_MODEL}}'; final synthesis → omit model (session model).
   • agentType: 'runner' only for fixed mechanical checks, if that agent type exists.
   • Parallel file edits use isolation: 'worktree'.
   • Check before launch: reread the script and confirm every run() call has the model its stage needs (synthesis has none), that no bare agent() call remains, and that the worst-case agent count, including per-item fan-out, is at most {{MAX_AGENTS}}. Fix it before calling Workflow.
5. When it completes, give the result first, then one cost line from Mode: cost with N=1 (agents, minutes, cache writes, output).