AltimateAI / AltimateAI/altimate-code
feat: deterministic trivial-plan fast-path for the explore-first mandate
- 主要语言
- TypeScript
- 星标
- 811
- 派生
- 134
- 平均合并
- 3 天 2 小时
- 30 天内合并 PR
- 50
描述
Found during v0.8.3 release review (End User / Data Engineer persona). **Deferred** because it needs design + tests for a config flag or heuristic (>30 min).
## Context
v0.8.3 added a REQUIRED "Investigate before drafting" section to `packages/opencode/src/session/prompt/plan.txt` mandating at least one read-only tool call before the plan agent drafts. There's a prose escape hatch for trivial tasks (broadened in v0.8.3 to cover single well-specified edits to already-read files), but it relies on the model's judgment of the prose — a literal-minded compliant model (the exact class this targets) may still grep on a one-line rename, adding a 3-5s round trip and chattier output every time.
## Suggested fix
A deterministic trivial-plan fast-path — a config flag or a heuristic — so single-edit / fully-specified plans skip the exploration mandate without relying on the model interpreting the prose escape hatch. Pair with telemetry on how often the explore-first mandate adds a tool call to previously-trivial sessions (latency/token cost on the easy case).
## Scope
`packages/opencode/src/session/prompt/plan.txt`, plan-agent loop in `packages/opencode/src/session/prompt.ts`, plus a config surface.
贡献指南
评估
这个 Issue 还没有评估数据。