Codex hook lifecycle events flood the thread as “Unhandled Codex event”

[skyblue]

Description

With Codex hooks enabled, bb renders every Codex hook/started and hook/completed notification as a separate Unhandled Codex event row. A single thread can accumulate hundreds of these rows, making actual assistant output and errors difficult to find.

Environment

• bb-app: 0.42.1
• Codex CLI: 0.154.0
• OS: macOS
• Provider: Codex
• Codex hooks are configured in ~/.codex/hooks.json

Reproduction

1. Enable one or more Codex hooks.
2. Run a thread in bb with the Codex provider.
3. Observe the thread timeline while hooks run.

In the affected thread, the persisted event log contained 214 hook/started and 214 hook/completed events. They were all stored as provider/unhandled, even though the expanded payloads show normal hook lifecycle notifications and completed hook runs, not failures.

The attached screenshot also shows a warning event with the message:

Skill descriptions were shortened to fit the skills context budget.

Actual behavior

• Each hook lifecycle event appears as Unhandled Codex event.
• The thread becomes very noisy and meaningful events are hard to find.
• The label suggests an error even when the hook completed successfully.

Expected behavior

• Recognize and suppress routine Codex hook lifecycle events, or render them in a compact, non-error diagnostic area.
• Render the skill-budget warning once as a warning, rather than as an unclassified event row.
• Preserve genuinely unknown provider events as unhandled diagnostics.

Related work

PR #1310, which is closed and unmerged, mentions preventing Codex hook lifecycle events from appearing as misleading unhandled rows. This behavior is still reproducible in bb-app 0.42.1.

Screenshot

[bb-slop-cop]

🚨 SLOP COP 🚨 · new-issue-autopilot

Bug · Priority: Medium · Effort: Low
Labels: providers, provider-codex, confirmed-repro. All classifications were previously missing and are now verified.

REPRODUCED · High confidence. Hook lifecycle notifications have unknown coverage, and generic warnings lack a translation handler, so both become unhandled diagnostics. Reproduction and root-cause report.

The regression test reproduced four failures in two clean trusted checkouts. PR #3478 suppresses routine hooks, translates warnings, and preserves unsuccessful hooks and unknown diagnostics. All 281 provider tests and typecheck pass; 132 changed text lines across four files. Verification covers translation and event assembly, not live hooks or browser rendering.

AGENT GENERATED