fix(agents): safe codex tool-terminal continuation, auth-true simple-completion routes, doctor opt-out by steipete · Pull Request #108966 · openclaw/openclaw

steipete · 2026-07-16T11:46:36Z

What Problem This Solves

Three verified codex-cluster defects on current main:

[Bug]: codex app-server: tool-call-terminal turn (stopReason=toolUse, no final assistant) dies with no retry — needs safe continuation, not replay #108517 — a codex app-server turn ending stopReason=toolUse with no final assistant message dies with "Agent stopped before producing a final response… some tool actions may have already been executed" and no recovery, even though codex-rs treats that turn as a clean completion (handle_turn_complete emits turn/completed status=completed without requiring a final agent message).
OpenAI simple-completion (utility model) pairs api_key auth with the ChatGPT-backend Codex transport — narration/titles 401 #104779 — simple-completion (narration, chat titles, labels) resolves the model route before auth, so a dual-route OpenAI model materialized on the ChatGPT-backend codex transport can be paired with an openai:api-key profile → silent 401s. Upstream contract: codex-rs to_api_provider(auth_mode) derives the endpoint from the auth mode.
doctor --fix silently adds and enables the uninstalled codex plugin (routing-altering) when openai is enabled #97180 (residual) — openclaw doctor --fix flipped an explicit plugins.entries.codex.enabled: false to true. Deny-list and global disable were respected; an explicit per-entry opt-out was not.

Why This Change Was Made

Safe continuation, not replay: one bounded continuation attempt fires only when the terminal turn is provably settled — zero visible payloads, no error/abort/timeout, and every toolCall id in the terminal assistant has a matching non-error toolResult in the message snapshot. Lifecycle counters are attempt-cumulative and alias across batches, so completion is proven per tool-call id; undispatched or partially dispatched batches keep the existing fail-closed error. The continuation appends a fresh native-thread turn (codex turn/start appends a new user turn; completed tools are never re-sent), so replaySafe deliberately does not gate it — invariant documented inline.
Auth before route: for official dual-route OpenAI models, simple-completion now resolves auth first, selects the auth-compatible route via the existing selectOpenAIModelRouteAuth + buildAgentRuntimeAuthPlan + materializePreparedRuntimeModel seam (no parallel resolver, no URL swaps), then re-validates auth against the materialized route. isAuthModeAllowedForModel is now symmetric: the codex transport requires subscription OAuth, so api-key profiles fail closed with a clear error instead of a 401. Custom base URLs and explicit refs are untouched.
Explicit opt-out wins: doctor treats plugins.entries.codex.enabled: false like the deny case — warn with a fix hint, never mutate. Missing-entry and restrictive-allowlist repairs stay automatic (the managed-harness contract from Fix doctor repair for disabled Codex runtime plugin #82502 is preserved).

User Impact

Codex turns that finish on tool calls now produce the final answer instead of an error, without any risk of double-executing side effects.
Narration/titles/labels work on Codex-OAuth-primary setups with an api-key profile present; misconfigured pairings fail loudly at auth resolution instead of silently at request time.
Doctor respects an operator's explicit codex plugin disable and says what manual action would be needed.

Evidence

Focused suites green after rebase onto current main: run.incomplete-turn.test.ts (136, incl. new continuation success / exhaustion / never-started / partial-dispatch cases), simple-completion-runtime.test.ts (26: api-key→platform route, oauth→codex route kept, custom baseUrl untouched, explicit ref honored), model-auth.profiles.test.ts (22), doctor route-warnings + core-checks (234+126).
codex-rs contract personally verified at 800715d201: bespoke_event_handling.rs (clean completion without final message), turn_processor.rs + core/src/session/mod.rs (turn/start appends, never replays), model-provider-info/src/lib.rs (endpoint = f(auth mode)).
Autoreview (gpt-5.6-sol, xhigh): iterated three times — it caught idle≠completed, cumulative-counter aliasing, and a wire-shape discriminator bug in the continuation guard; final pass clean. Remote Testbox pnpm tsgo:core green on the pre-hardening state; hosted CI validates the final head.

Fixes #108517
Fixes #104779
Fixes #97180

…tes, doctor opt-out Three verified codex-cluster fixes: - Tool-call-terminal turns (stopReason=toolUse, no final assistant message) now get one bounded continuation instead of dying with an incomplete-turn error. Completion is proven per tool-call id (every terminal toolCall needs a non-error toolResult in the snapshot) so undispatched or partially dispatched batches keep failing closed; the continuation appends a fresh native-thread turn and never replays completed tools. (#108517) - Simple-completion (narration/titles/labels) resolves auth before finalizing official dual-route OpenAI models and re-materializes the auth-compatible physical route through the runtime-plan seam; api-key credentials are no longer legal for the ChatGPT-backend codex transport (fail closed), ending silent 401s. Custom base URLs and explicit refs stay honored. (#104779) - openclaw doctor --fix no longer flips an explicit plugins.entries.codex enabled:false; explicit user opt-out now warns with a fix hint like the deny-list case, while missing-entry and allowlist repairs stay automatic. (#97180) Fixes #108517 Fixes #104779 Fixes #97180

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d5680f408c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-07-16T11:51:19Z

+  if (openAICodexTransportRequiresOAuth(params)) {
+    return params.mode === "oauth";
+  }


Accept token profiles for Codex subscription routes

When a user has an OpenAI token profile, this check now rejects it for openai-chatgpt-responses even though OpenClaw treats oauth and token as subscription auth (src/agents/provider-model-route-auth.ts:107-109) and the Codex app-server contract explicitly accepts an OpenAI OAuth or token profile for prepared subscription routes (extensions/codex/src/app-server/run-attempt-connection.ts:185-187, covered by extensions/codex/src/app-server/auth-profile-runtime-contract.test.ts:428-456). This breaks users who authenticated with openclaw models auth paste-token --provider openai by failing getApiKeyForModel/simple-completion route materialization before the valid token profile can be handed to Codex; allow mode === "token" here while still rejecting API-key auth.

Useful? React with 👍 / 👎.

… messagesSnapshot, update sibling hints The symmetric codex-transport auth gate now accepts both subscription credential classes (oauth and ChatGPT token) instead of oauth only, so prepared-candidate fallback keeps skipping to a valid subscription profile. attempt-result threads messagesSnapshot into the incomplete-turn attempt shape, and the doctor-lint sibling test expects the manual hint now that an explicit plugins.entries.codex.enabled=false blocks auto-repair.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8269b96d7b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-07-16T12:18:58Z

+      const routeAuthDecision = selectOpenAIModelRouteAuth({
+        resolution: routeResolution,
+        sourcePlan: buildProviderModelAuthSourcePlan({
+          ownership: { reason: "provider-binding", source },
+          profiles: [],


Respect configured OpenAI auth before locking the route

With models.providers.openai.auth: "api-key" and only an ambient OPENAI_API_KEY plus an automatic ChatGPT/OAuth profile, the earlier resolveApiKeyForProvider call has no modelApi or configured-auth filter, so it can return the OAuth profile; this then wraps that first credential as a required provider-binding, and selectOpenAIModelRouteAuth can no longer apply the configured API-key route the main runtime applies in src/agents/runtime-plan/prepare-auth.ts. Simple completions can therefore keep using the subscription/Codex route despite explicit API-key config, so keep automatic route selection aligned with the runtime plan instead of binding the first unfiltered credential.

Useful? React with 👍 / 👎.

…d doctor helper

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7d1d83480d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-07-16T12:28:49Z

+    (params.attempt.messagesSnapshot ?? []).flatMap((message) => {
+      const result = message as { role?: unknown; toolCallId?: unknown; isError?: unknown };
+      return result.role === "toolResult" &&
+        result.isError !== true &&
+        typeof result.toolCallId === "string"
+        ? [result.toolCallId]


Scope tool-result completion to the terminal assistant

Here the completion proof accepts any non-error toolResult in messagesSnapshot, but that snapshot is the full session transcript (attempt-stream-settle.ts snapshots activeSession.messages) rather than just results following the terminal assistant. If a provider/runtime repeats a toolCall id used in an earlier turn, a stale successful result can satisfy completedToolCallIds even when the current terminal toolUse request never ran, so the new continuation prompts the model not to repeat the tool and can skip a side-effecting action. Scope the result lookup to the current assistant's following span, or otherwise verify the result belongs to this terminal turn.

Useful? React with 👍 / 👎.

steipete · 2026-07-16T12:36:01Z

Merged via squash.

Prepared head SHA: 7d1d83480d8cd90e2e7877189cae5d8cc92714a0
Landed commit: 2848acbbaa18f120d257557b1cfb6939f383dcc6

…completion routes, doctor opt-out (openclaw#108966) * fix(agents): safe codex continuation, auth-true simple-completion routes, doctor opt-out Three verified codex-cluster fixes: - Tool-call-terminal turns (stopReason=toolUse, no final assistant message) now get one bounded continuation instead of dying with an incomplete-turn error. Completion is proven per tool-call id (every terminal toolCall needs a non-error toolResult in the snapshot) so undispatched or partially dispatched batches keep failing closed; the continuation appends a fresh native-thread turn and never replays completed tools. (openclaw#108517) - Simple-completion (narration/titles/labels) resolves auth before finalizing official dual-route OpenAI models and re-materializes the auth-compatible physical route through the runtime-plan seam; api-key credentials are no longer legal for the ChatGPT-backend codex transport (fail closed), ending silent 401s. Custom base URLs and explicit refs stay honored. (openclaw#104779) - openclaw doctor --fix no longer flips an explicit plugins.entries.codex enabled:false; explicit user opt-out now warns with a fix hint like the deny-list case, while missing-entry and allowlist repairs stay automatic. (openclaw#97180) Fixes openclaw#108517 Fixes openclaw#104779 Fixes openclaw#97180 * fix(agents): accept ChatGPT token profiles on codex transport, thread messagesSnapshot, update sibling hints The symmetric codex-transport auth gate now accepts both subscription credential classes (oauth and ChatGPT token) instead of oauth only, so prepared-candidate fallback keeps skipping to a valid subscription profile. attempt-result threads messagesSnapshot into the incomplete-turn attempt shape, and the doctor-lint sibling test expects the manual hint now that an explicit plugins.entries.codex.enabled=false blocks auto-repair. * chore(agents): route test mock casts through unknown, un-export unused doctor helper

steipete requested a review from a team as a code owner July 16, 2026 11:46

openclaw-barnacle Bot added commands Command implementations agents Agent runtime and tooling size: L maintainer Maintainer-authored PR labels Jul 16, 2026

chatgpt-codex-connector Bot reviewed Jul 16, 2026

View reviewed changes

chore(agents): route test mock casts through unknown, un-export unuse…

7d1d834

…d doctor helper

chatgpt-codex-connector Bot reviewed Jul 16, 2026

View reviewed changes

steipete merged commit 2848acb into main Jul 16, 2026
118 checks passed

steipete deleted the claude/codex-round-d1 branch July 16, 2026 12:36

github-actions Bot mentioned this pull request Jul 16, 2026

📡 Upstream Digest — 2026-07-16 13:24 UTC curtismercier/openclaw-mods#1312

Open

kagura-agent mentioned this pull request Jul 16, 2026

fix(agents): add tool-use terminal continuation for codex app-server toolUse stops #108724

Closed

steipete mentioned this pull request Jul 16, 2026

fix(agents): scope tool-terminal completion proof to post-assistant results #109259

Merged

belfortfilho mentioned this pull request Jul 24, 2026

[Bug]: lane task timeout races legitimate turn completion, discarding the final assistant text (~26s window) #113182

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

fix(agents): safe codex tool-terminal continuation, auth-true simple-completion routes, doctor opt-out#108966

fix(agents): safe codex tool-terminal continuation, auth-true simple-completion routes, doctor opt-out#108966
steipete merged 3 commits into
mainfrom
claude/codex-round-d1

steipete commented Jul 16, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Jul 16, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Jul 16, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Jul 16, 2026

Uh oh!

Uh oh!

steipete commented Jul 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Uh oh!

Conversation

steipete commented Jul 16, 2026

What Problem This Solves

Why This Change Was Made

User Impact

Evidence

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Jul 16, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Jul 16, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Jul 16, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

steipete commented Jul 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant