fix(splitting): remove PR-creation references from prompt template

[aaight]

Summary

Fixes MNG-1084.

The splitting agent has PM-only capabilities (no scm:pr) — it never opens
pull requests. But the prompt template's Step 6 told the agent to "post
summary comment ... once you've confirmed PR creation" and to use "real and
confirmed PR numbers and URLs - see output from CreatePR". That wording was
left over from a copy/paste with the implementation prompt and contradicted
the agent's actual capability contract, leading to confused output on
real splitting runs.

Changes

• src/agents/prompts/templates/splitting.eta — replaced the two
  action-oriented PR-creation references in Step 6 with wording grounded in
  the splitting workflow's actual side-effects (story work items returned by
  CreateWorkItem). Story sizing language ("One PR, one day" heuristic and
  the - **S**mall - One PR, one day... INVEST bullet) is intentionally
  preserved because it describes target story size, not an action the agent
  should perform.
• tests/unit/agents/prompts.test.ts — added two narrow regression
  tests:
  • One that asserts the rendered splitting prompt does not contain
    CreatePR, confirmed PR creation, or real and confirmed PR numbers,
    and does still contain CreateWorkItem, Stories Created,
    [Story Title](URL), and URLs returned by `CreateWorkItem`.
  • One that pins the preserved "One PR, One Day" sizing-heuristic wording
    so a future cleanup pass cannot strip it accidentally.

Out of scope

• No changes to src/agents/definitions/splitting.yaml — capabilities
  already correctly omit scm:pr and the existing
  agent-profiles.test.ts > splitting includes file editing but not CreatePR
  test continues to pass.
• No changes to src/gadgets/session/core/finish.ts or backend PR-validation
  code — splitting remains a no-PR workflow by capability/agent contract.
• No documentation updates — README.md and docs/architecture/04-agent-system.md
  already correctly describe splitting as creating smaller work items with
  fs, session, pm capabilities.
• Custom DB-stored splitting prompt overrides (per-project) may still
  contain stale copy; operators with overrides will need to update those
  separately.

Test plan

• npx vitest run --project unit-core tests/unit/agents/prompts.test.ts — 129 pass (added 2)
• npx vitest run tests/unit/backends/agent-profiles.test.ts — 56 pass; the existing splitting includes file editing but not CreatePR invariant still passes
• npx vitest run --project unit-core tests/unit/agents/capabilities/resolver.test.ts — 37 pass
• npx vitest run --project unit-core tests/unit/agents/ — all 802 agent unit tests pass
• npm run typecheck — clean
• npx biome check on the two touched files — clean

🕵️ claude-code · claude-opus-4-7 · run details

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[nhopeatall]

LGTM - the splitting prompt now grounds the summary step in CreateWorkItem story URLs without removing the intentional PR sizing heuristic. I verified the focused prompt regression test plus the existing capability/profile guards locally.

🕵️ codex · gpt-5.5 · run details