Skip to content

fix(mcp): expose account pool members for delegation checks - #70

Merged
ScriptedAlchemy merged 2 commits into
mainfrom
t3/fix-pool-provider-mismatch
Nov 6, 2026
Merged

ScriptedAlchemy merged 2 commits into
mainfrom
t3/fix-pool-provider-mismatch

Conversation

@ScriptedAlchemy

@ScriptedAlchemy ScriptedAlchemy commented Nov 6, 2026 •

Copy link
Copy Markdown
Owner

Problem

A delegation targeting pool_claudeAgent reports the selected concrete account, such as claudeAgent. Capabilities advertised the pool but omitted its member IDs, so agents could not verify that relationship and reported an unconfirmed provider mismatch even when model and effort matched.

Change

Expose poolMemberInstanceIds for account pools in orchestrator_capabilities and clarify the target/result schemas: compare the returned account with the advertised members, then check model and options separately. Extend the existing pool delegation test and add eval cases for an expected member and an unrelated account.

Scope and approval

This is a focused fix for missing metadata in the existing delegation capability. It qualifies for the small obvious-bug exception: routing and execution behavior are unchanged; the response now exposes the relationship needed to interpret the result.

Prior implementation verification

  • The extended pool delegation test failed before the fix because poolMemberInstanceIds was absent, then passed afterward. Non-pool providers omit the field.
  • Focused service, pool-routing, and contract tests: 112 passed.
  • vp run check:changed HEAD: passed formatting, lint, knip, GraphQL validation, native-addon build, affected-package typechecks (36 packages), and adjacent tests (90 tests in four files).
  • Both new MCP eval cases passed the runner's dry run. Model runs were attempted but blocked by missing local vibeproxy configuration; no model-behavior pass is claimed.

Observed in the focused service test:

capabilities: pool_codex -> [codex, codex_work]
delegate_task(driverKind=codex): providerInstanceId=codex_work, model=gpt-5.4

The original reported task was not independently traced; this change addresses the confirmed metadata gap.

Model: GPT-6.1 Sol (xhigh). Harness: Codex in T3 Code.

Current audit verification

Updated this branch with current main and the shared Android test-environment fixture fix. The published tree is the same tree verified locally.

  • The changed-file gate passed all 90 tests in four files, formatting, lint, both Knip passes, GraphQL validation, and the native project-memory build.
  • The gate's cached typecheck task was blocked by this environment's seccomp_unotify process-tracing restriction. All 41 affected-package typecheck tasks passed when rerun serially with vp run --no-cache --parallel --concurrency-limit 1 and the same package filters.
  • Both provider-pool eval cases constructed valid requests in a fresh dry run. This is request validation; model behavior was not rerun.
  • Review found no additional defect in the capability field or the separate provider-membership, model, and option checks.

Audit and fixes: GPT-6 Astra Pro in ChatGPT.

@changeset-bot

changeset-bot Bot commented Nov 6, 2026 •

Copy link
Copy Markdown

⚠️ No Changeset found

Latest commit: aadcc03

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

@ScriptedAlchemy
ScriptedAlchemy merged commit c81b119 into main Nov 6, 2026
16 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant