Skip to content

Native-first code-review fixes — verification ​

  • Date: 2026-11-11
  • Branch: ticket/review-fixes
  • Base: integration/native-first at 2228366
  • Issues: parent #209, review findings from the standards/spec review, follow-up #231
  • Unit gate: npm run verify (strict typecheck + full test suite)
  • Live gate: npm run test:live:native-first

Fixed findings ​

  • Per-agent durable native mirror (Agent.nativeMirror, default on), a General settings toggle row, and ingestion/reconciliation gates that skip private native prompts without stopping managed composer work.
  • Recovered native prompts map host prompt files to transcript attachment chips.
  • Reconciliation captures a binding target key and repeats against a changed target, so an adoption during a start/reconnect read recovers into the new conversation only.
  • nativeMentionRouting: false selects the legacy tool surface, restoring the explicit send_message path alongside the visible routing refusal.
  • Shadow comparison reports the planned/visible routing delta as data; the #227 driver uses a $Name shadow prompt and asserts no unexplained text, phase, or routing mismatch.
  • Native ingress no longer awaits one session's processing in handleHostEvent; different sessions ingest in parallel.
  • Delegated prompts keep the bounded last-20 group transcript with the compact roster and budget.
  • Standards: source-text wiring guards narrowed, tool-list self-assertion removed, native-direct UI derives from chatContext()?.conversation, and the glossary keeps only the domain definition.

Deferred ​

  • The ~12 parallel native service-state maps were not merged in this branch. Follow-up #231 carries the exact maps/functions and the refactor rationale.

Verification ​

CommandResult
npm run verifyPass — strict typecheck; 1809 tests passed, 0 failed, 0 skipped
npm run test:live:native-firstPass — main and restart phases, 0 gaps, shadow comparison row logged
npm run test:liveFail — real TUI check times out waiting for the workspace wizard after agent creation; follow-up #232

Released under the MIT License.