[augustl]

Problem

Delegating to an available model can fail with model_unavailable when T3's cached provider catalog is stale. This occurred with muse-wonder-1.0: the Muse CLI and a direct thread could use it, while delegation saw only the older Spark catalog and refused the request before launching Muse.

resolveTargetRechecking already refreshes a known provider instance once on provider_unavailable, but does not do so for model_unavailable.

Change

Use that same bounded refresh-and-retry for model_unavailable. Re-resolve against the refreshed catalog, preserving model validation: a still-unavailable model is refused, and an already-advertised model causes no refresh.

Scope and approval

This is a small, focused fix for an obvious stale-cache failure under CONTRIBUTING.md's exception. It extends an existing retry condition without adding settings, changing model availability, or bypassing provider/model validation. The separate MCP authentication failure observed after switching providers is outside this PR.

Verification

The regression runs through OrchestratorMcpService.delegateTask with a Codex parent and a Muse target. Before the fix, the newly-advertised model failed with Model muse-wonder-1.0 is not advertised by provider muse. and the missing-model control showed that no refresh occurred. After the fix, all three cases pass:

• Already advertised: delegates without refreshing.
• Newly advertised: refreshes once and dispatches the requested Muse model.
• Still unavailable: refreshes once, returns model_unavailable, and dispatches nothing.

Commands and results:

cd apps/server
vp test run src/mcp/OrchestratorMcpService.test.ts -t 'rechecks a delegated model' --maxWorkers 1
Before: 2 failed, 1 passed

vp test run src/mcp/OrchestratorMcpService.test.ts --maxWorkers 1
After: 62 passed

cd ../..
vp run check:changed --concurrency 1
Passed: format, lint, both Knip passes, GraphQL generation check,
typechecks for 3 affected packages, and 70 tests across 3 files.

The regression uses deterministic provider-registry snapshots and the real delegation service; no live or paid Muse turn was started. A live retry could not be checked because the current MCP session separately returns Authentication required.

Implemented with GPT-6.1 Sol (high) via the Codex harness in T3 Code.

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-09T15:52:22.426009Z	727be86	PR opened

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[macroscopeapp]

Approvability

Verdict: Approved at 727be86

Macroscope's review found this PR approvable — Delegation now refreshes a known provider’s catalog once before rejecting an unavailable model. The retry retains existing target validation and project enablement safeguards, with no change to product defaults.

^{You can add or adjust custom eligibility rules. Learn more.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 727be86cfe

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]

Retry stale models for driver-selected targets

When a supported target specifies driverKind (with or without model) rather than providerInstanceId, the instanceId computed above is undefined, so the early return installs neither this new model_unavailable catch nor the refresh. A stale catalog for the provider selected by resolveTarget therefore still rejects requests such as { driverKind: "muse", model: "muse-wonder-1.0" } without calling refreshInstance, even though this selector combination is accepted by OrchestratorMcpTarget; carry the selected instance into the retry path as well.

Useful? React with 👍 / 👎.

[coderabbitai]

No actionable comments were generated in the recent review.

Walkthrough

When target resolution returns model_unavailable, resolveTargetRechecking now refreshes the provider instance and retries. Tests cover models advertised initially, advertised after refresh, and still unavailable after refresh.

Changes: Model resolution

Layer / File(s)	Summary
Refresh and retry unavailable models apps/server/src/mcp/OrchestratorMcpService.ts, apps/server/src/mcp/OrchestratorMcpService.test.ts	resolveTargetRechecking now retries after refreshing the provider when resolution returns model_unavailable. Tests cover initial availability, availability after refresh, and continued unavailability without dispatch.

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix

Possibly related issues

• [Bug]: delegate_task model_unavailable doesn't list the models the provider advertises #31890: The change affects the model_unavailable refusal path, but it does not change the refusal message requested by the issue.

Possibly related PRs

• pingdotgg/t3code#26625: That PR adds advertised-model validation to resolveTarget, which can produce the model_unavailable result that triggers this PR’s refresh-and-retry behavior.

Suggested reviewers: macroscopeapp, juliusmarminge

Merge Risk: Minimal · up to 727be

The stale-catalog retry appears ready to merge after normal checks; no actionable merge-blocking risk was identified.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title "fix(server): refresh stale model catalogs before refusing delegation" matches the main change. resolveTargetRechecking now refreshes the provider instance and retries when resolveTarget fai…
Description check		The description follows the repository template. It has Problem, Change, Scope and approval, and Verification sections. The Problem section states the failure with the muse-wonder-1.0 example and the …
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.

Full details: Title check

Explanation

The title "fix(server): refresh stale model catalogs before refusing delegation" matches the main change. resolveTargetRechecking now refreshes the provider instance and retries when resolveTarget fails with model_unavailable, in addition to provider_unavailable. The title is specific, uses a conventional commit prefix, and names the actual behavior change.

Full details: Description check

Explanation

The description follows the repository template. It has Problem, Change, Scope and approval, and Verification sections. The Problem section states the failure with the muse-wonder-1.0 example and the existing provider_unavailable behavior. The Change section describes the bounded refresh-and-retry for model_unavailable. The Scope section argues the small-obvious-fix exception and names the out-of-scope MCP authentication issue. The Verification section lists the focused test runs, before and after results, the check:changed run, and the untested live retry. The template's leading comment asks for a conventional commit title, and the title meets that. The Scope section does not link a triaged issue or maintainer approval, and it cites an exception without quoting the approval comment. The description still covers the required sections, so the check passes.

Finishing Touches

Generate unit tests (beta)

• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: main
• File: .coderabbit.config.ts
• Commit: b8c2cfa
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Advanced
• Run ID: 291580ad-f3f0-42a1-8af4-3980e4abce72

Commits

Reviewing files that changed from the base of the PR and between b8c2cfa and 727be86.

Files selected for processing: 2

• apps/server/src/mcp/OrchestratorMcpService.test.ts
• apps/server/src/mcp/OrchestratorMcpService.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.

[maria-rcks]

Note

Written by gpt-6.1-sol on behalf of Maria

Verified in a real dev environment

Verified the exact head 727be86cfe4944bcd74840bcd4bc087609701005 on Linux with Playwright and Google Chrome, using real Codex gpt-6.1-sol turns through the local proxy. Projects, threads and provider instances were created through the web app against isolated development state. No mocks, fixtures or demo data.

Reproduced on github/main at f12bb6229c first. A second Codex instance initially used the real 0.156.0 CLI, whose native catalog lacks Sol. After replacing its executable with 0.164.0 at the same path, without changing T3 settings, delegation still returned model_unavailable.

Repeated the same external CLI upgrade on the PR. The parent capabilities result still showed the old target catalog before delegation. The actual MCP result then confirmed providerInstanceId: codex_upgrade, model: gpt-6.1-sol, status: completed, summary: REAL_CHILD_OK. The refreshed provider cache reported 0.164.0 and included Sol.

Neighboring behavior: a second delegation to the already advertised Sol completed with ADVERTISED_OK; requesting definitely-unavailable-pr33426 returned model_unavailable and created no child task.

Short runtime transcript:

main: delegate_task(codex_upgrade, gpt-6.1-sol, wait)
  model_unavailable: Model gpt-6.1-sol is not advertised by provider codex_upgrade.
PR:   delegate_task(codex_upgrade, gpt-6.1-sol, wait)
  completed: REAL_CHILD_OK
PR:   delegate_task(codex_upgrade, gpt-6.1-sol, wait)
  completed: ADVERTISED_OK
PR:   delegate_task(codex_upgrade, definitely-unavailable-pr33426, wait)
  model_unavailable
Blacksmith: cd apps/server && vp test run src/mcp/OrchestratorMcpService.test.ts --maxWorkers 1
  62 passed

No code changes or pushes were needed. Existing CI checks passed. The recording shows the interaction and child completion; it is a sampled screen capture, not a performance measurement. Uploaded playback and seeking were checked.

Not verified

Native desktop, iOS and Android clients, remote/relay connections, and other provider adapters including Muse were not exercised. This live pass used explicit provider-instance selection; driver-only selectors were not exercised. The existing review about driver-only selectors remains outside this verification result.