Repository navigation
feat(manual-mode): make edits and writes also wait for users permissions - #122
Conversation
…ject AGENTS.md (#104) Fixed - #127 a multi-line `edit` no longer deadlocks the run. pi's JSON-string recovery is a bare JSON.parse, which throws on the raw newlines a small model writes inside oldText/newText; with `write` refused for an existing file, a model that had diagnosed the bug had no way to deliver the patch. Validation runs before the tool_call hook, so this is a patch-pi.mjs patch. - #121 behind a router each model keeps its own context window. The probe asked about models[0] rather than the declared default, and then stamped one value onto every model. The window drives read-guard and the context budget, and nothing was logged because the probe "succeeded". - #96 a tool call in prose is no longer "recovered". Names are checked against ctx.getAllTools(): a tool that is not registered cannot be re-issued natively by anyone, so the nudge was an instruction nobody could follow. - #128 compaction no longer arrives late on Ollama, which reports its prompt size after truncating it. For that provider only, the watchdog measures what is in context and compacts on the larger of the two figures. - skill cards actually load in the pi package: skill-inject and knowledge-inject counted directories up to find skills/, which is wrong in the built package layout and failed silently. Added - #104 project-context reads the nearest AGENTS.md (then CLAUDE.md), capped at 4000 chars, injected once as a hidden tail message so the cached prefix survives (#73). Never loads little-coder's own AGENTS.md. - #95 `little-coder -p --plan-mode` works. The clarifying questions are answered from the research and the assumptions stated in the plan, which is written to .pi/approved-plan.md for a later /implement. - #93 LITTLE_CODER_SUBCODER_ACCESS=write gives sub-coders edit/write. Off by default; dispatch is withheld at both levels. - #62 pi-little-coder, the extension and skill layer as a pi package, by @dbmrq. Published alongside each release. Changed - #122 manual mode asks about edit and write too, by @marek2901. - `true` and `false` are whitelisted, so `2>/dev/null || true` is not refused. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ptq8zPmvafKeTBJevPsQdr
|
Merged, shipped in v1.20.0. Second one of these from you, and it closes the gap the first one left. You are right that this was the more serious half. The Headless denying by default is the right call and matches the precedent #90 set: a mode built around a human in the loop should refuse when there is no human, not silently proceed. Sub-coders are unaffected since they are pinned to Thank you, and the tests were a pleasure to review. The per-instance latch isolation test is the one that would have caught a real bug if the confirmer had been module-level state. |
2 participants
Hello,
another PR for manual mode 👋
Manual mode was only gating shell commands
edit/writeslipped through without confirmation. Fixed: they now prompt Apply/Deny like shell does (native diff still shows, we just wire the gate).Added
Apply all (this session)for convenience if you don't want to review every edit, one click opts out for the rest of the session. Headless still denies by default.🙇