Repository navigation
Conversation
After a steer, the Kiro adapter holds back a chunk end that could start Kiro's "[STEERING steer-…]" note. Nothing released it: a reply ending in "[" lost the bracket at turn end, and a "[" before a tool call moved to the start of the next message. The steering filter now releases held text as its own chunk when a non-text update (such as a tool call) or the steer's start or end arrives. The shared ACP normalizeSessionUpdate hook returns a list so a flavor can emit that extra chunk. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ApprovabilityVerdict: Approved at Macroscope's review found this PR approvable — Preserves buffered trailing text in Kiro steered replies before tool updates or steering boundaries. Other ACP providers retain their existing notification behavior, keeping the change narrowly scoped. You can add or adjust custom eligibility rules. Learn more. |
|
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
Source: pingdotgg#32759 at dc52fc4. Run compaction receipt invalidation on every normalized update.
|
Note 🤖 Opus 5.5 replying on behalf of Victor. Closing this: I could not run it with a real Kiro sign-in, which I do not have, so the steered-reply fix is shown only by a replay harness and tests. The branch stays on my fork if it is useful. |
Problem
After a steer, Kiro's model ends its reply with a
[STEERING steer-…]note meant for Kiro's own UI, and the Kiro adapter strips it from the streamed text. Kiro splits that note right after a lone[. So when a chunk ends in something that could start the note ([,[S, …[STEERING steer), the filter holds that text back until the next chunk shows whether it is the note.Nothing released the held text when no next chunk came:
steering_cleareddropped the filter's state together with the held text. A steered reply ending inx = [showedx =.Checking the list [, then a tool call, thenAll good.showedChecking the listand[All good..This only happens in a Kiro turn that received a steer, when a message ends on a prefix of the note. It is rare, but the text is lost or moved silently.
Change
makeKiroSteeringNoteFilter(packages/provider-kiro) now releases held text as its own text chunk just before the next non-text update, such astool_call, or the steer'ssteering_clearedorsteering_injected. Othersession_info_updatekinds (context_usage,turn_end, …) never release it, because they do not end the message in the shared adapter. In the Kiro recordings, the only other updates sent during a reply are tool calls, which end the message anyway, so a note that is still streaming is not leaked. Text inside a note is still dropped as before.normalizeSessionUpdate(packages/provider-acp) now returns a list, and the adapter handles each item in order. The two other users of the hook (Devin and Copilot inprovider-acp-registry) are wrapped where they are configured. Their functions and behavior are unchanged.Scope and approval
This is a small, focused fix for an obvious bug in the steering-note filter added with Kiro support in #14693. The hook change is the smallest way to emit the extra chunk, and only Kiro returns more than one item. No client, contract or settings change.
Left out on purpose:
\n\nthat can remain before a removed note.steering_cleared. No recording shows Kiro doing that.Verification
Kiro is not set up on my test machine, so there is no live Kiro run. The proof is the replay harness
KiroAdapterV2.test.tsalready uses. The new test's frames have the same shape as the live Kiro 2.33.1 recording intestkit/fixtures/kiro_message_steering, with fake text. Replay transcript (frames in → assistant messages out):vp test run src/orchestration-v2/Adapters/KiroAdapterV2.test.ts -t "steering note" --reporter=verbosefromapps/server, with the new test on main's source (a9e709f):Tests 2 failed | 21 skipped (23), for the reasons above.The same command on this branch:
Tests 2 passed | 21 skipped (23).Also on this branch:
KiroAdapterV2.test.ts: 23/23 passed.AcpAdapterV2.test.ts: 100/100 passed.KiroReplayFixtures.integration.test.ts: 17/17 passed. It includes the recordedkiro_message_steeringturn, whose note is still removed.AcpRegistryReplayFixtures.integration.test.ts: 16/16 passed (Devin and Copilot).vp run check:changedpassed: format, lint, knip, typecheck and tests.Not checked: a live Kiro turn, and a client UI pass. The change is server-only, and every client shows the same projected text.
Prior PRs
Generated with Claude Opus 5.5 in T3 Code Nightly