技能 2026-10-12 11:45 UTC
caliperlatest
版本【latest】由提交人 edonadei 提交,主要更新类型为:功能特性新增与扩展。提交说明:Plant a history so a claude-code attempt starts mid-conversation (#292)。为项目引入了新的能力与模块支持。
Commit提交信息原文
Plant a history so a claude-code attempt starts mid-conversation (#292) * Plant a history so a claude-code attempt starts mid-conversation A task's history: names a file of earlier user and assistant turns, shell calls included. claude-code writes them as a session in the attempt's isolated home and resumes it with --resume, so the prompt is the next user turn. Codex, pi and hermes refuse such a spec, naming the backend. The planted turns stay out of the transcript, since the resumed stream does not replay them, and the judge reads them as labelled prior context. They are read once per run and handed to both the harness and the judge. Records the decision as ADR 0036 and documents history: in the README, the spec reference, backends, results and both skill references. Closes #278 * Let the planted-turn type hold its own shape PlantedTurn validates its own invariants, so the history loader only words pydantic's first error in the file's terms instead of re-checking every key by hand. The claude-code session writer builds each assistant turn's tool calls once, and the history tests import at the top. * Add an example that starts mid-conversation A slash-only repro-first mode can't be discovered, so the task plants the turns where the user entered it and the agent found the bug, and checks the failing test still comes before the fix. * Plant a skill by name so the history never copies it A skill: turn names a declared skill the user entered by slash command. claude-code expands it from the installed SKILL.md each attempt, so an edit to the skill reaches the next run; an ablated skill plants the bare command. The example now enters repro-first this way, and its last planted turn no longer promises the test-first fix the skill should cause. * Enter a planted skill with /name in a user turn A user turn opening with /name of an installed skill is that slash command, expanded from the installed SKILL.md; anything else stays as typed. The history reads like the conversation, with no extra turn kind, and an ablated skill's command stays as typed, as the CLI would leave it. A /name matching no declared skill is warned about, not refused, since a turn may open with a path. * Name history files *.history.yaml A convention, not a rule: history: still takes any path. The suffix reads as a history beside the *.eval.yaml it belongs to.