[ajianaz]

What

Part 1 of 2 of sibling change notices: the core mechanism. Wiring into tole run children follows in part 2.

Why

#380. Sibling child agents work blind to each other; when one edits a file another has read, the other should hear about it at its next safe point.

Changes

• sibling_watch.rs (new): children are separate processes, so each tracked child appends reads/writes to one shared append-only log (owner-only, one short O_APPEND line per access) and a SiblingWatcher folds it. A sibling WRITE to a path this child read or wrote yields one notice per sibling per path per poll; the path stays quiet until the child reads or writes it again. Changes made before the child looked at a file are not news. Bounded to 4096 tracked paths; a half-written tail waits for the next poll; garbage lines are skipped.
• Tracked tool decorator: records read_file / write_file / edit_file / delete_file on success, forwards the rest of the tool surface unchanged; recording failures never fail the tool.
• interrupt.rs/turn.rs: interrupts gain a source (owner | notice). A notice is delivered at the next safe point as an attributed entry (source: notice, "sibling notice" in the text) and never withdraws output or supersedes a final answer (only owner messages do).
• Log fields are untrusted: agent ids must match the agent-id charset, and the path is escaped, bounded (160 chars) and quoted in the model-facing text, which says it is an automated notice, not an instruction.
• docs/architecture.md, CHANGELOG.

Scope

Depth-1 siblings only; the parent is not tracked; no registry, locking or ownership (explicit anti-goal: no swarm engine). The log is a same-user channel between processes that already share the parent's authority; a forged self-line can only add noise, not suppress or escalate.

Testing

• 10 unit tests: normalisation, exactly-once notice and re-arming after a re-read, own writes and earlier changes are not news, sibling reads do not notify, partial lines, garbage/missing log, path bound, owner-only mode, the decorator (records on success only, forwards the tool surface), hostile id/path cannot forge model-facing text.
• 1 integration test: a notice is delivered attributed, does not withdraw a tool call and does not supersede a final answer; queue unit test for the source split.
• fmt, clippy -D warnings, cargo test --workspace pass; cora review --staged clean.
• Not done: live two-child run (needs the CLI wiring in part 2).

Refs #380

[github-actions]

🔍 Cora AI Code Review

✅ No issues found. Code looks good!

---

Review powered by cora-code · BYOK · MIT