[lydakis]

Requested by George · project thread

Six fixes from a day of running bots, one commit each.

What changes

1. Every Claude bot Sonny runs keeps its prompt cache for an hour. Before this, tasks from start_task were created without --cache-ttl, and leads made before the hour-long default were still on 5 minutes. So any gap over 300 s re-read the whole context uncached. Now tasks start with 1h, and when Sonny opens an existing Claude bot whose TTL differs from the policy, it moves the bot to 1h with the new configure field cache_ttl. The protocol version is unchanged.
2. Triggers keep working when the store file is replaced. A trigger pinned the store's identity, which hashes the file's inode. A copied or restored store made every fire fail with store_mismatch, and trigger ls and the background view filtered those triggers out. Triggers now pin the store's lineage, which survives a copy. Sonny also restamps any trigger plists still carrying the old identity when it starts. A fire that is skipped or fails attaches one line to its reply-to bot (or else its target), so the agent finds out without starting a turn.
3. An empty reply ends a turn. Before, an empty completion was accepted only after a tool call. A lead that read a wake, called no tool and decided there was nothing to say failed with empty_completion. Now any empty completion completes with empty text.
4. The approver can read a write whose preview was cut before its path. A model may write {"content": …, "path": …}. Past the 2 KiB preview the path was gone, and an ungated write has no planning node to read back, so every later gated call that turn was escalated as "the call is too long to show". tool_started now names the file of a cut write or edit, as it already did for apply_patch. The prompts read keeps those paths ahead of a call it cuts at its byte budget. When a write still names nothing, the escalation says that, instead of blaming the call under review.
5. A task that resolves itself still reports. The resolve_task description now says the reply after the call is what the lead reads. Before, a task could call it and then end with nothing, so the lead's wait got empty text.
6. history reads a turn by its handle. Every tool hands out turn:BOT/N with store-wide numbers, but history's turn counts one bot's turns from 1. The tool now takes handle, and turn_not_in_history says how each one is numbered.

Checked

• cargo fmt --check, cargo clippy --workspace --all-targets -D warnings and cargo test --workspace pass.
• New runtime tests, each confirmed to fail without its fix where that applies:
  • test_configure_moves_a_bot_to_the_hour_long_cache
  • test_a_fire_finds_its_store_after_the_file_was_replaced
  • test_an_empty_reply_ends_a_turn_after_a_tool_call_and_after_a_prompt
  • test_an_ungated_write_cut_before_its_path_still_names_its_file
  • the history handle cases in test_siblings
  • the attached missed-fire line in test_trigger
• New unit tests: Sonny cache policy for tasks and existing bots, trigger restamping, and paths from cut writes and edits.
• Full runtime suite: see CI.
• Hot path: tool_started now parses the arguments of a cut write or edit once. That is one serde_json read of a call that already arrived whole, the same work a cut patch already did. The prompts read parses only the one call it cuts at its budget.

No screenshots: nothing here changes what Sonny draws.

---

Generated by Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-29T04:27:58.997923Z	be223da	New commits
🔒 Security Review	✅ Completed 2026-10-29T04:29:06.174239Z	be223da	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1bb9b60975

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".