236ed2c6
docs: specify sprint models
a73x 2026-09-06 13:44
Commit message
docs/skills/sprint-delivery/SKILL.md
| Old | New | ||
|---|---|---|---|
| @@ -64,13 +64,41 @@ commit IDs, or temporary service URLs in this skill. | |||
| 64 | test success does not imply user acceptance. Begin the next sprint only when | 64 | test success does not imply user acceptance. Begin the next sprint only when |
| 65 | authorized, preserving any acceptance requirement in the agreed plan. | 65 | authorized, preserving any acceptance requirement in the agreed plan. |
| 66 | 66 | ||
| 67 | ## Model and effort routing | ||
| 68 | |||
| 69 | Use a cheap model for conversation and routine work, reserving stronger models | ||
| 70 | for planning and review: | ||
| 71 | |||
| 72 | - `gpt-5.6-luna`: conversation and implementation. Use `none` for status, | ||
| 73 | explanation, and obvious edits; `low` for routine implementation and focused | ||
| 74 | debugging; `medium` when inspecting several files or a small interaction. | ||
| 75 | - `gpt-5.6-terra`: adversarial review. Use `medium` for normal review and | ||
| 76 | `high` for cross-component review or tricky regressions. | ||
| 77 | - `gpt-6-astra`: sprint planning and architectural decisions. Use `high` for | ||
| 78 | ordinary planning, `xhigh` for protocol or ownership changes, and `max` only | ||
| 79 | for difficult design disputes or repeated failed attempts. | ||
| 80 | |||
| 81 | Raise effort before raising the model tier when the task remains local. Escalate | ||
| 82 | to a stronger model when the task crosses component boundaries, changes a | ||
| 83 | protocol or ownership invariant, remains ambiguous, or has failed twice for the | ||
| 84 | same reason. Do not spend maximum effort on routine work. | ||
| 85 | |||
| 86 | The planning handoff must be durable in the repository (usually the sprint | ||
| 87 | plan): goal, acceptance scenarios, affected components, reusable existing code, | ||
| 88 | architectural decisions and invariants, implementation slices, review questions, | ||
| 89 | and known risks. Luna works from that artifact rather than asking Astra to | ||
| 90 | reconstruct repository context on every turn. | ||
| 91 | |||
| 92 | With collaboration tools, use `fork_turns="none"` and provide a short task | ||
| 93 | brief. A full-history fork inherits the parent model and defeats this cost | ||
| 94 | choice. Reuse agents with the intended model; an old agent with an inherited | ||
| 95 | expensive model is not a cheap substitute. | ||
| 96 | |||
| 67 | ## Economical pairing | 97 | ## Economical pairing |
| 68 | 98 | ||
| 69 | - Explicit model defaults: `gpt-5.6-luna` for implementation and | 99 | The implementation/review pairing remains Luna for implementation and Terra for |
| 70 | `gpt-5.6-terra` for adversarial review. With collaboration tools, use | 100 | adversarial review. Astra is added only for planning and the escalation cases |
| 71 | `fork_turns="none"` and provide a short task brief. A full-history fork inherits | 101 | above; it is not required for every slice. |
| 72 | the parent model and defeats this cost choice. Reuse agents with the intended | ||
| 73 | model; an old agent with an inherited expensive model is not a cheap substitute. | ||
| 74 | - Give each agent the deliverable, relevant file paths, edit ownership, | 102 | - Give each agent the deliverable, relevant file paths, edit ownership, |
| 75 | constraints, acceptance checks, and its partner's name. They should read only | 103 | constraints, acceptance checks, and its partner's name. They should read only |
| 76 | necessary code. The main session coordinates, integrates, and independently | 104 | necessary code. The main session coordinates, integrates, and independently |