Best possible solution:

Land the narrow public hook metadata change after normal maintainer review and CI, preserving the no-raw-final-text boundary.

Do we have a high-confidence way to reproduce the issue?

Not applicable as a bug reproduction. The feature is source-verifiable by comparing current main, where the ended hook lacks final, with the PR path that emits metadata from an existing frozenResultText record.

Is this the best way to solve the issue?

Yes. Reusing the existing frozen-result capture and adding a privacy-minimal object to the existing ended-hook event is narrower than adding a new delivery path, exposing raw text, or inferring state from channel routing.

What I checked:

• Current main hook payload has no final metadata: emitSubagentEndedHookOnce currently emits target, reason, delivery, run, timing, outcome, and error fields, but no final field. (src/agents/subagent-registry-completion.ts:94, f4f98f45c72f)
• Current main public hook type has no final metadata: PluginHookSubagentEndedEvent currently ends at error?: string, so plugin authors have no typed final-result metadata field on main. (src/plugins/hook-types.ts:575, f4f98f45c72f)
• Existing frozen-result source is available before ended-hook emission: The lifecycle path freezes completion output into frozenResultText and frozenResultCapturedAt before deciding whether to emit the ended hook. (src/agents/subagent-registry-lifecycle.ts:750, f4f98f45c72f)
• PR implementation is narrow: The patch adds buildSubagentEndedFinalMetadata, hashes exact UTF-8 bytes, reports byte length and capture time, omits blank finals, and attaches the object as event.final. (src/agents/subagent-registry-completion.ts:64, 4f3665d4ac2f)
• PR extends the public hook contract: The patch adds a documented optional final object to PluginHookSubagentEndedEvent and explicitly keeps raw final text out of the payload. (src/plugins/hook-types.ts:582, 4f3665d4ac2f)
• Focused tests cover metadata behavior: The patch asserts exact UTF-8 hash and byte length, missing timestamp handling, blank omission, and no raw final text leakage in the event JSON. (src/agents/subagent-registry-completion.test.ts:131, 4f3665d4ac2f)

Likely related people:

• steipete: Recent current-main history and local blame show Peter Steinberger maintaining the subagent lifecycle/completion extraction area and adjacent plugin-hook surfaces shortly before this PR. (role: recent maintainer and adjacent owner; confidence: high; commits: 59fb9e5ca7fe, 2218ce46fe2e, 6f3b5f8666f3; files: src/agents/subagent-registry-completion.ts, src/agents/subagent-registry-lifecycle.ts, src/plugins/hook-types.ts)
• onutc: The thread-bound subagents feature introduced the subagent lifecycle hook events, the subagent_ended contract, and the channel-agnostic hook architecture this PR extends. (role: introduced behavior; confidence: high; commits: 8178ea472db1; files: src/plugins/hook-types.ts, src/plugins/hooks.ts, src/agents/subagent-registry-lifecycle.ts)
• tyler6204: Tyler Yust authored the frozen-result refresh and push-based announce delivery work that provides the stored final-result fields this PR exposes as metadata. (role: adjacent owner; confidence: medium; commits: 81b93b9ce04b, e554c59aac68; files: src/agents/subagent-registry.types.ts, src/agents/subagent-registry-lifecycle.ts, src/agents/subagent-announce.ts)
• vincentkoc: Vincent Koc recently touched plugin hook/type surfaces and subagent ended hook loading behavior, making him relevant for public hook contract review. (role: recent hook/type maintainer; confidence: medium; commits: 48042c387556, 1b25dcf57a9f; files: src/plugins/hook-types.ts, src/agents/subagent-registry-completion.ts)

Remaining risk / open question:

• Focused tests were not executed in this read-only review; the PR author's stated Vitest run was inspected from the PR body but not independently rerun.

Codex review notes: model gpt-5.5, reasoning high; reviewed against f4f98f45c72f.