Repository navigation
refactor(responses): lower agent_message in a request middleware - #576
Merged
Merged
Conversation
Codex sends sub-agent tasks as `agent_message` input items. The two Responses translators lowered them to framed user messages themselves, so a native Responses upstream that does not understand the item could only receive it untouched. Add the `withOpenAIResponsesAgentMessageShim` interceptor, which rewrites every `agent_message` into the same framed user message when the target protocol is not OpenAI Responses, or when the new `openai-responses-agent-message-shim` flag is enabled. The flag defaults to off for every provider, so native targets keep receiving the item as before. The interceptor runs ahead of the rest of the chain and thus before the terminal dispatch translates the payload. Remove the `agent_message` handling from both translators; an item that reaches them unlowered now fails with the existing invalid-input-item error. Move `agentMessageContent` and its tests into the gateway next to the interceptor, and move the translated-target expectations to the interceptor and attempt tests.
yyyr-p
added a commit
to yyyr-p/Floway
that referenced
this pull request
Oct 4, 2026
Integrates upstream commits e286553..c7e4d78: - refactor(responses): lower agent_message in a request middleware (Menci#576) - fix(claude-code): update subscription request defaults (Menci#571) - fix(compaction): use the context compaction shim by default (Menci#572) - fix(codex): derive quota expiry from exhausted windows (Menci#570) - fix(tls): support IP-literal certificate identities (Menci#569) - fix(web-search): resolve the OpenAI search passthrough lazily (Menci#568) - fix(provider-codex): update stable client identity for newer models (Menci#567) - fix(provider-codex): preserve Responses Lite wire semantics (Menci#543) - feat(web): adopt the blue Floway logo (Menci#563) - feat(codex): show and redeem ChatGPT reset cards (Menci#528) - fix(responses): answer Codex WebSocket prewarms locally (Menci#552) - fix(responses): handle positional tool declarations in middleware (Menci#558) - fix(gateway): preserve first-token timing across tool continuations (Menci#561) - fix(responses): consume hosted-tool streams before billing metadata (Menci#560) - fix(codex): restore output omitted from terminal snapshots (Menci#551) - fix(codex): preserve model catalog context windows (Menci#559) - fix(web): disable inactive upstream disclosure (Menci#534) - fix(translate): isolate translated request payloads (Menci#557) - feat(models): support Claude Sonnet 5.5 (Menci#555) - feat(models): support GPT-6.1 Sol (Menci#553) - fix(web): allow imported credentials on upstream creation (Menci#556) - fix(gateway): preserve hosted tool selection across aliases and continuations (Menci#549) - fix(translate): restore Responses callable identities (Menci#547) - perf(translate): bound namespace tool name allocation work (Menci#548) - fix(translate): preserve namespace descriptions on child tools (Menci#545) - feat(requests): show each request's time to first token (Menci#550) - fix(translate): reject ambiguous flat callable kinds (Menci#542) - fix(gateway): preserve client tool identity in hosted dispatch (Menci#538) - fix(web): mask credential headers in request record exports (Menci#539) Conflict resolutions: - upstreams/access-control.tsx: keep the fork's parameterized title, description, and error props (title/description ?? defaults, plus defaultOpen/revealOn) alongside upstream Menci#534's disclosureDisabled={!override}. - upstreams/access-control_test.tsx: union the Control helper to accept both error and initialOverride, and keep both Menci#534's disabled-disclosure case and the fork's validation-error cases. - control-plane/routes.ts: union the schemas.ts named imports -- the fork's OAuth2 admin bodies plus upstream Menci#528's codexRateLimitResetCreditsBody and codexRateLimitResetConsumeBody; both sides' route blocks merged cleanly. - requests/ttft_test.tsx (Menci#550): pass the fork's required exchangeStreams={[]} to the RequestDetailPanel cases that construct it with no collected streams. Verified with pnpm run verify (typegen, lint, typecheck, 611 test files / 6644 tests, test:installers, check:agents-md, check:generated-assets, check:verify-parity, build:web).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follows the design suggested in discussion: move the
agent_messagelowering out of the translators and into a Responses middleware, gated by a flag.Behaviour
Codex delivers sub-agent tasks as
agent_messageinput items. Today the two Responses translators (to Anthropic Messages and to OpenAI Chat Completions) lower them to a plain user message framed as coming from another agent rather than the user. A native Responses upstream receives the item untouched.This PR adds the interceptor
withOpenAIResponsesAgentMessageShim, which rewrites everyagent_messageinto the same framed user message when either:openai-responses-agent-message-shimis enabled.The flag is off by default for every provider kind, so existing behaviour is unchanged. The new option lets an operator enable the lowering for a Responses-compatible upstream that drops or rejects the item. The two translators no longer handle
agent_message; an item that reached them unlowered would fail with the existing invalid-input-item error.The framing text is unchanged.
agentMessageContent(and its tests) moved into the gateway next to the interceptor, since it now has a single user. This is the one placement choice you may want to weigh in on; moving it back intotranslateis trivial if you prefer that.The interceptor sits early in the chain, before the terminal dispatch translates the payload, so no other interceptor and no translator sees the raw item.
Tests
agent_messageare untouched, un-lowerable content rejects withTranslatorInputError.attempt_test.ts: the Anthropic and Chat Completions targets receive the same framed user text as before; native target forwards the item untouched with the flag off and lowers it with the flag on.agent_message.pnpm run verifypasses except forcodex-reset-cards_test.tsx("keeps the redemption key after an ambiguous failure and reopening confirmation"), which is flaky and unrelated: it is untouched by this change, passes on its own, and failed in only some runs of the surrounding directory.