Add p.bashEach and p.writeInput prompt helpers

[Copilot]

Two friction points from evaluation tasks — per-element shell iteration and dynamic write destinations — had no first-class helpers, forcing prose workarounds that are imprecise and non-declarative.

New helpers

p.bashEach(template, inputArrayField, options?) — runs a command template once per element in a caller-supplied array field, with {} as the element placeholder:

const healthProbe = agent({
  input: s.object({ endpoints: s.array(s.url) }),
  instructions: p`${p.bashEach("curl -s -o /dev/null -w '%{http_code}' {} --max-time 5", "endpoints")}`,
  output: s.object({ results: s.array(s.object({ url: s.url, status: s.string })) }),
});

p.writeInput(inputPathField, contentOutputField, options?) — like p.writeOutput but the destination path comes from an input field rather than a static string:

const renderer = agent({
  input: s.object({ outputPath: s.path }),
  instructions: p`Render the changelog. ${p.writeInput("outputPath", "rendered")}`,
  output: s.object({ rendered: s.string }),
});

Implementation

• skills/rig/rig.ts: Two new PromptIntent modes (prompt.bashEach, prompt.writeInput), new pathField property on PromptIntent, PromptHelpers type entries with JSDoc, implementations, and rendering cases in renderPromptIntentInstruction
• src/rig.test.ts: Tests for both helpers covering mode, stored fields, and options passthrough
• skills/rig/references/prompt-intents.md: Updated helper table; merged write helpers into a unified p.write / p.writeOutput / p.writeInput decision section; expanded "Dynamic shell commands" with p.bashEach
• skills/rig/SKILL.md, README.md: Updated intent selection tables and prose

• Fixes [rig-tasks] Daily rig evaluation — 2026-07-25 — 5/5 passed #106