Hacktoberfest 2026: le issue che i maintainer hanno segnato per ottobre, aperte e adatte ai principianti. Sfoglia le issue Hacktoberfest

[Bug]: Plugin MCP tools are missing from the first turn when discovery publication races capability capture

Aperta
  • Qualcosa non va?
#570 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
3-5 giorni
Idoneità per principianti
25/100
Tipo di issue
Bug
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
node.js, typescript
Ambito
backend-api-design, cli

Direzione di ricerca

Start at prepareTurnCapabilities(signal) and read the turn-capability tests; the issue reports that waitUntilReady() can finish before plugin MCP discovery and publication. Verify that turn capture waits for the current discovery and publication, while cancellation, failed-server behavior, and read-only inspection remain as described. Done means the first turn sees the published plugin inventory and the behavioral regressions pass.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

An enabled native plugin can complete MCP tools/list successfully while the first headless turn captures an empty plugin tool inventory. The model then cannot discover or invoke those tools. I have prepared a source fix and regression coverage in a fork; this issue follows the repository's current external-contribution process.

Product or interface

CLI - headless

Version
  • Reproduced with unmodified official npm @minimax-ai/code 0.6.9 and 0.7.2.
  • Reproduced without injected latency using public main at 50448abe3c753391cbfb24df835518d844504862 (source version 0.6.5).
  • Fix also applied to the public 0.7.2 synchronization branch at 0bd2c5a3f425a4e5fdafdf2eb954548e4d20fdfd.
  • Node.js 22.23.3; pnpm 9.12.0.
Platform

Linux — Alibaba Cloud Linux 3 (Soaring Falcon), update 13, x86_64.

Issue area

MCP / OAuth / integration

Steps to reproduce

Use an isolated MiniMax profile and an enabled local plugin declaring a streamable HTTP MCP server. The real PowerContext Server exposes list_scopes, with a seeded synthetic Scope ID/title. Remove the corresponding private entry from mcp.json to isolate the plugin connection. A separate private-only control succeeds.

A complete real-host reproduction is available in PowerContext at 319e389d3d61bb800cf3d8d7541e9e4522e54f7f, in tests/e2e/test_minimax_host.py:

POWERCONTEXT_MINIMAX_ACCEPTANCE_COMMAND=/absolute/path/to/mcode \
  uv run --frozen --extra cli --extra server python -m pytest \
  tests/e2e/test_minimax_host.py \
  -k 'controlled_inference and plugin-mcp' -q \
  --basetemp=/path/to/disposable/test-data

The test launches a real PowerContext Server, SQLite and MiniMax, with synthetic local authentication and an OpenAI-compatible local inference fixture. It uses mcode exec --permission full --max-steps 5 --timeout 40s --output-format json. The fixture selects the discovered tool, invokes it, and requires the seeded Scope ID/title associated with that tool-call ID. No external model account or provider credential is needed; native plugin enablement and Skill delivery are checked independently.

Expected and actual behavior

Expected: the first turn captures the completed, published inventory for the current plugin configuration, so the model can discover and invoke list_scopes.

Actual: both cold and separately preflighted plugin-only executions fail. Native inventory responses include 63 tools and list_scopes; the model sees no Scope tool. The 0.7.2 deferred-discovery path returns an empty tool_search result. Switching to canonical /mcp/ also fails. On unmodified public main, the natural failure reproduces with a successful inventory response in about 264 ms; no artificial delay is required.

Expected one isolated list_scopes tool, discovered []
invoked_tools = []
scope_result_matched = false
Proposed source fix

waitUntilReady() can finish before background plugin MCP discovery/publication. Turn acquisition then captures and retains that incomplete generation; later inventory arrival cannot replace the turn's immutable view. The patch waits at prepareTurnCapabilities(signal) for current discovery and publication, and rechecks readiness at final synchronous capture. The MCP readiness wait has a 20-second bound. Cancelling one waiter leaves shared discovery available to other turns. A published failed server retains the existing fail-open behavior, with healthy servers still available and no wait for retry timers. Read-only plugin inspection remains non-blocking.

The patch changes two runtime files and adds eight behavioral regressions to an existing test file. No vendor bundle, private connection fallback, fixed sleep, or relaxed isolated-route assertion is used.

Validation
  • Main-based commit: pnpm verify passed the full applicable Linux profile on the clean committed source, including source export, lint, type checking, build, artifact checks, capability tests, Code SDK tests, CLI smoke, offline BYOK and permissions policy. The capability suite reported 5,557 passed / 18 skipped. Windows and macOS sandbox gates were inapplicable; the npm release-archive installation gate was not run because this is a source build.
  • The focused turn-capability file passed on both branches: 13 tests on main and 17 on the 0.7.2 synchronization source, preserving its existing Agent roster coverage. Seven of the new eight regressions fail against the unmodified main runtime; all eight pass with the fix.
  • The main fixed build passed all six real-host controlled-inference cases: private-only, plugin-only and default installation, each under cold and preflight startup.
  • The 0.7.2 fixed build passed the complete 13-case PowerContext real-host suite, covering setup/diagnostics, authentication, disabled private settings and all six invocation cases. Its build passed; the full upstream pnpm verify was run on the main-based fix, not repeated on this synchronization branch.
  • Cold plugin-only execution also passed with 6-second and 11-second delays in the actual MCP inventory response (recorded durations 6,023 ms and 11,013 ms). The unmodified main source failed both without injected delay and with the 11-second delay. Every successful invocation returned the seeded Scope ID/title with the matching tool-call ID; the plugin Skill was present and the private MCP entry absent.

These are source-build results, not a claim that official npm 0.6.9 or 0.7.2 has been updated. macOS, Windows, Desktop and external model providers were not exercised. OpenAI Codex assisted with implementation, investigation and verification.

Before submitting
  • I have searched existing issues.
  • I have included my version and removed sensitive information.
Lingua principale
TypeScript
Stelle
2.1k
Fork
271
Merge medio
15h 10m
PR unite (30g)
34

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di MiniMax-AI/minimax-code

Tutte le issue di MiniMax-AI/minimax-code

Issue simili

Altre issue su TypeScript

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.