kiro: isolate delivered final answers to exact child conversation and serving identity by luvs01 · Pull Request #403 · luvs01/opencodex

luvs01 · 2026-09-01T05:52:33Z

Motivation

The delivered-final-answer cache was previously bound using a parent-prioritized logging digest before route/auth/account resolution, which could let one conversation suppress another sibling that shared the same parent identifier.
The change isolates suppression to the exact child conversation and serving identity so cached final answers cannot cross conversation, principal, or provider boundaries.

Description

Defer binding the turn-termination scope until after adapter, provider, model, admission, and serving-account resolution and only bind a composite scope for the Kiro adapter.
Build the composite scope from the child-specific session lane (sessionSpecificLaneIdFromRequest: thread ?? session, with the parent included only as a coalescing prefix — a parent-only lane binds no scope), the admission identity, route.providerName, route.modelId, and the resolved serving account, then normalize/hash the composite before using it as the cache key via bindTurnTerminationScope.
Resolve the serving credential lazily at check/record time: for key-authenticated routes the element is credentialIdentity — the digest of the resolved wire credential rather than the configured $ENV/keychain: reference — so a rotation behind a stable reference terminates the old identity's scope; an OAuth snapshot account id wins when bound, and the Codex auth context is the last resort.
Revalidate the live credential selection (selectionIsCurrent/refreshDispatchAdapter) before evaluating localTerminal, so a record a since-replaced credential made cannot suppress work the send path would have moved onto the live one; a failed refresh skips the terminal entirely and the send path surfaces the credential error.
Remove the earlier premature bindTurnTerminationScope usage so no insufficiently scoped record is retained before routing and auth are known.
Add end-to-end regressions in tests/server-kiro-completion-e2e.test.ts (sibling isolation, key-pool failover following the serving credential, env-backed key rotation, parent-only lane binds nothing) plus dispatch-level coverage in tests/responses/adapter-dispatch-local-terminal.test.ts for the pre-terminal staleness check and the failed-refresh path.

Testing

bun test tests/server-kiro-completion-e2e.test.ts — 19/19 pass including the new regressions.
bun test tests/responses/adapter-dispatch-local-terminal.test.ts — 4/4 pass.
bun run typecheck, bun run privacy:scan, bun run structure:check — all pass.

Link to Devin session: https://app.devin.ai/sessions/d60838d35393455abb53c5ab4a54ca18
Open in Devin Desktop: https://app.devin.ai/desktop/session/d60838d35393455abb53c5ab4a54ca18?variant=devin
Requested by: @luvs01

github-actions · 2026-09-01T05:52:43Z

✅ Deterministic PR hygiene checks passed.

github-actions · 2026-09-01T05:52:54Z

✅ READY

all PR quality gates passed.

Hygiene

✅ Deterministic PR hygiene checks passed.

coderabbitai · 2026-09-01T05:52:55Z

Important

🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Repository: luvs01/opencodex/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 0a1ca093-9637-4f1f-9f84-f383024b99ba

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

…rability

Encrypted-task recovery rebound the conversation-only termination scope on the reparsed request after the early binding was removed. The kiro composite binding in transport overwrote it only by ordering, and a bind that produces no composite digest would leave the stale scope live for the adapter's delivered-answer suppression. Document the cross-request scope contract in structure/: the binding site, composite material, recording and consumption now live in providers/kiro.md, with the transport ownership row updated. Co-Authored-By: Epinephrine <luvs01@hanmail.net>

devin-ai-integration · 2026-09-21T16:37:11Z

Fixed in c680397 — removed bindTurnTerminationScope(reparsed, resolvedConversationId) from the encrypted-task recovery path in request-prepare.ts and dropped the now-unused import. The Kiro composite binding in request-transport.ts is now the only write of the termination scope, so correctness no longer depends on it overwriting a stale conversation-only bind (which it would skip entirely when no lane/conversation identity exists, since bindTurnTerminationScope early-returns on undefined).

devin-ai-integration · 2026-09-21T16:37:13Z

Fixed in c680397 — documented the new scope contract in structure/providers/kiro.md ("Delivered final-answer termination scope": binding site, composite digest material, recording at delivery, and consumption in the adapter), updated the request-transport.ts ownership row in structure/transports/responses.md, and refreshed the manifest scope + regenerated INDEX.md. bun run structure:check passes.

…kiro route The composite termination scope now binds only in transport for the kiro adapter, so the openai-chat spy could no longer observe a bound scope. Record the delivered answer through a real kiro turn and replay against the kiro route, preserving the reparse-rebind and follow-up assertions. Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

devin-ai-integration · 2026-09-22T02:10:39Z

Already addressed in c680397 — the recovery reparse no longer calls bindTurnTerminationScope; the transport binding is now the only write of the scope.

devin-ai-integration · 2026-09-22T02:10:42Z

Already addressed in c680397 — the scope contract is documented in structure/providers/kiro.md (binding site, composite material, recording, consumption), with the request-transport.ts ownership row updated in structure/transports/responses.md. bun run structure:check passes.

…d time A Kiro 429/401 can rotate the serving credential after the termination scope is bound, so an eagerly-hashed servingAccount pinned every record to the credential that failed -- and for key-authenticated routes it was always null, letting one key's delivered answer suppress a replay that belonged to a different key. Bind a resolver instead: conversation, admission, and route stay eager, while the serving credential resolves when the answer is checked or recorded, tracking route.provider through every failover site. Key routes now contribute their non-secret apiKeyAccountLogLabel so distinct keys never share a scope. Co-Authored-By: Epinephrine <luvs01@hanmail.net>

…terminal Two follow-on Devin Review findings on the termination scope: - A request carrying only x-codex-parent-thread-id bound the shared parent lane as its scope, so every sibling under one parent collapsed into a single conversation and could suppress each other. The bind now takes the child-specific lane only (sessionSpecificLaneIdFromRequest) and falls back to the Cursor conversation id; a parent-only request binds no scope rather than treating a coalescing group as one child. - The lazy credential resolver made the check read whatever selection the request was bound with, which can go stale between prepare and dispatch. prepareAdapterExchange now runs the same selection-binding revalidation dispatch uses (selectionIsCurrent/refreshDispatchAdapter) before evaluating localTerminal, then rebinds the reasoning replay scope, so a record made by a replaced credential cannot suppress work the send path would have moved onto the live one. Regressions: an e2e asserting two requests sharing only a parent header both reach upstream, and unit coverage of the stale-binding refresh ordering at the local-terminal boundary. Co-Authored-By: Epinephrine <luvs01@hanmail.net>

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

The serving-credential element hashed the configured env:/keychain: reference, so a rotation behind a stable expression kept the old identity's scope. credentialIdentity digests the resolved wire credential instead — the same primitive reasoning-metadata uses to bind learned refusals to what was actually sent. A failed dispatch refresh left the route stale but still evaluated localTerminal, letting a stale-credential record suppress a turn and return success where the credential error belonged. The terminal is now skipped after a failed refresh; the send path re-validates and surfaces the mapped error on its own terms. Adds an env-rotation e2e regression and a failed-refresh dispatch test, and keeps structure/providers/kiro.md in sync. Co-Authored-By: Epinephrine <luvs01@hanmail.net>

devin-ai-integration · 2026-09-22T06:54:35Z

End-to-end verification of kiro: bind termination identity to the resolved wire credential on 72c4e48.

Environment: Windows, real startServer proxy over HTTP against a scripted Kiro upstream (amazon eventstream, codex_kiro_final_answer toolUseEvents). Manual driver independently re-derives the feature behavior; bun test tests/server-kiro-completion-e2e.test.ts tests/responses/adapter-dispatch-local-terminal.test.ts → 23 pass, 0 fail.

Scenario	Result
First turn delivers `phase:"final_answer"`	1 upstream send, real answer returned
Replay same conversation (client dropped `phase`)	`output:[]`, upstream stays at 1 — suppressed
Replay again	still suppressed — record survives use
Sibling `session_id`, identical body	sent upstream — lane isolation holds
Trailing NEW user message after answer	sent upstream — only a trailing answer suppresses
Rotate env-backed key (`apiKey:"$VAR"` a→b), replay	sent upstream — rotated credential = new scope
`x-codex-parent-thread-id` only (no session)	both siblings sent — parent-only binds no scope
Streamed replay	`response.completed` only, no deltas, no `empty_completion_retry_failed`
Env var deleted before replay	400 `kiro token missing` — real error, not fake empty success

Request log rows for suppressed turns: localTerminalReason:"kiro_final_answer_already_delivered", usageStatus:"reported", zero usage (not estimated).

GUI golden path: GET / serves the built dashboard (Online, provider count, token totals); Logs lists all rows — suppressed replays show 0 tokens / "Upstream sends: 0" while the delivered turn shows real usage.

Tested via Devin session: https://app.devin.ai/sessions/d60838d35393455abb53c5ab4a54ca18

luvs01 · 2026-09-24T14:24:20Z

Closing: current dev already isolates delivered final answers — src/responses/turn-termination.ts has �indTurnTerminationScope keyed by a WeakMap on the parsed request, bound from
equest-prepare.ts with the resolved conversation identity, which is the exact-child-conversation isolation this PR introduced.

devin-ai-integration · 2026-09-24T14:24:24Z

Devin is currently unreachable - the session may have died.

View session

fix(kiro): isolate delivered final answers

dd0c6ff

luvs01 added codex aardvark labels Sep 1, 2026 — with ChatGPT Codex Connector

luvs01 added codex aardvark labels Sep 1, 2026

github-actions Bot added the bug Something isn't working label Sep 1, 2026

github-actions Bot changed the title ~~kiro: isolate delivered final answers to exact child conversation and serving identity~~ [WRONG BRANCH] kiro: isolate delivered final answers to exact child conversation and serving identity Sep 1, 2026

github-actions Bot marked this pull request as draft September 1, 2026 05:52

This comment was marked as resolved.

Sign in to view

luvs01 changed the title ~~[WRONG BRANCH] kiro: isolate delivered final answers to exact child conversation and serving identity~~ kiro: isolate delivered final answers to exact child conversation and serving identity Sep 3, 2026

luvs01 changed the base branch from main to dev September 3, 2026 06:22

github-actions Bot marked this pull request as ready for review September 3, 2026 06:28

ci: retrigger checks (empty commit; dev merge conflicts)

ac13e49

This comment was marked as resolved.

Sign in to view

Merge branch 'dev' into codex/propose-fix-for-kiro-final-answer-vulne…

1536450

…rability

This comment was marked as resolved.

Sign in to view

test(layout): register adapter-dispatch-local-terminal in layout tables

fdf849c

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

luvs01 closed this Sep 24, 2026

Conversation

luvs01 commented Sep 1, 2026 • edited by devin-ai-integration Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Description

Testing

Uh oh!

github-actions Bot commented Sep 1, 2026

Uh oh!

github-actions Bot commented Sep 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✅ READY

Hygiene

Uh oh!

coderabbitai Bot commented Sep 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

devin-ai-integration Bot commented Sep 21, 2026

Uh oh!

devin-ai-integration Bot commented Sep 21, 2026

Uh oh!

devin-ai-integration Bot commented Sep 22, 2026

Uh oh!

devin-ai-integration Bot commented Sep 22, 2026

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

devin-ai-integration Bot commented Sep 22, 2026

Uh oh!

luvs01 commented Sep 24, 2026

Uh oh!

devin-ai-integration Bot commented Sep 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

luvs01 commented Sep 1, 2026 •

edited by devin-ai-integration Bot

Loading

github-actions Bot commented Sep 1, 2026 •

edited

Loading

coderabbitai Bot commented Sep 1, 2026 •

edited

Loading