feat(hooks): error event — fires when a turn exits unrecoverably

[emal-avala]

Summary

Adds HookEvent::Error, a lifecycle event the turn driver fires when an LLM call can't be recovered via retry or reactive compaction. Fills a real gap: today there's no way for a hook to catch turn failures without tailing stderr or polling session files.

Hook context

{
  "session_id": "…",
  "turn": 3,
  "stage": "llm_call_failed",
  "message": "<error text>"
}

The stage field is a coarse tag ("llm_call_failed" today) so future fire-sites — tool-dispatch failures, permission denies — can reuse the same hook surface with a different tag. Hooks can filter on stage if they care about one category.

Example use

[[hooks]]
event  = "error"
action = { type = "http", url = "https://pager.example.com/alert", method = "POST" }

Integration

• Schema: new HookEvent::Error variant with rustdoc.
• Dispatcher: run_hooks test confirms event matching.
• QueryEngine::fire_error_hooks(stage, message) helper plus one fire-site at the only return Err(...) path in run_turn_with_sink.
• /hooks events catalog, format_hook_event, and parse_hook_event all teach the new name.

Tests

• Serde roundtrip: HookEvent::Error ↔ "error"
• Dispatcher: a registered Error hook actually fires when the event is dispatched
• CLI invariant tests continue to pass (hook_event_catalog_has_unique_names, parse_hook_event_covers_every_variant_in_catalog)

Full hook + schema suites pass. Clippy clean under -D warnings.

Test plan

• cargo test -p agent-code-lib --lib hooks::
• cargo test -p agent-code-lib --lib hook_event_serde
• cargo test -p agent-code --bin agent parse_hook_event
• cargo clippy --workspace --tests --no-deps -- -D warnings
• cargo fmt --all --check

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.