[Lkx-JY]

Which issue or RFC does this PR close?

Closes #1840.

Rationale for this change

Integrate MiniMax Code with PowerContext's existing setup, configuration, and diagnostic workflows. Users should be able to install the complete native plugin, connect it to a local or protected Server, and maintain that installation without manually copying files or losing existing configuration.

The integration distinguishes package installation, native plugin discovery, CLI Server access, and actual host MCP invocation. Copying files alone must not be reported as successful native activation.

What changes are included in this PR?

CLI integration and target selection

• Add MiniMax to setup, interactive and multi-agent selection, config agent, and doctor.
• Support an explicit MiniMax data directory and executable launcher.
• Resolve installation targets through explicit options, MINIMAX_DATA_DIR, MAVIS_DATA_DIR, saved targets, and the selected CLI's native marketplace directory.
• Reuse the saved installation target and launcher for subsequent configuration and diagnostics.
• Report missing or invalid launchers explicitly rather than treating package copying as complete verification.

Complete package installation and upgrades

• Install the complete native plugin, including hidden metadata, the Skill and its references, MCP configuration, icon, and license.
• Support local sources and repository refs through the existing installation conventions.
• Track managed files, source, and ref in the installation record.
• Preserve user-created files, unrelated plugins, and existing connection settings.
• Reject modified managed files and file/directory collisions before replacement.
• Validate staged packages, remove only unchanged obsolete managed files, and restore the previous installation on publication failure.

Connection configuration and preservation

• Configure local, remote, and protected Servers through the existing setup/config workflow.
• Synchronize enabled private MCP settings with the installed plugin connection so both connections use the selected endpoint and credentials.
• Preserve unrelated MCP servers, unknown fields, headers, and disabled preferences.
• Preserve disabled private endpoints unless an endpoint change is explicitly requested. Leave unchanged private configuration byte-for-byte intact.
• Allow credential rotation without implicitly enabling a disabled private entry.
• Require replacement credentials before moving an authenticated connection to another endpoint.
• Detect divergent enabled private/plugin connections and provide reconciliation guidance.

Diagnostics and native acceptance

• Verify native plugin discovery in a disposable profile without modifying the real installation.
• Preserve SQLite-backed plugin disablement during diagnostic verification.
• Report missing, disabled, omitted, or incompatible plugin records and insufficient resource counts.
• Report package verification, native discovery, Server liveness/readiness, and authenticated CLI Server access separately.
• Isolate private and plugin MCP acceptance routes through configuration and native enablement, without relying on _2 aliases.
• Require an actual MCP tool result containing the seeded Scope ID and title, associated with the emitted tool-call ID.
• Cover cold startup and separate-process Server preflight for the isolated routes and the default installation with both connections enabled.
• Capture discovery timing and model-visible tool inventories to diagnose connected-but-unavailable tools.

Documentation and CI

• Add English and Chinese installation, authentication, upgrade, recovery, and diagnostic documentation.
• Document version alignment, credential storage, explicit plugin enablement, and qualification limits.
• Document the pinned upstream plugin-readiness source fix, its launcher setup, affected-environment verification, and official-release boundary.
• Add installation/configuration regressions, model-gateway tests, and opt-in real-host acceptance.
• Include MiniMax CLI tests in Windows portability coverage, with regressions for home-directory isolation and JSON newline differences.
• Preserve both current and historical Desktop qualification records when synchronizing master, and improve macOS packaging diagnostic collection.
• Align Code publication test database selection and the persistence compatibility declaration with the merged master behavior; keep the installed-release upgrade gate required.

Are there any user-facing changes?

Users can run:

powercontext setup minimax
powercontext config agent minimax
powercontext doctor minimax --json

Setup preserves existing plugin disablement. Reinstalling the package does not automatically re-enable it. Completed file installation is reported separately from failed native verification, with a nonzero overall exit and enablement guidance.

Doctor does not enable plugins, rewrite connection settings, or initialize the real MiniMax profile. Native discovery runs in a disposable snapshot.

Credentials may be stored in both the private MCP configuration and the installed plugin configuration. Updated connection files receive owner-only permissions on POSIX systems; Windows users must protect them with appropriate ACLs. The distributed package and installation record contain no plaintext credentials.

The CLI, Server, and plugin should come from the same repository revision or matching release tag. Plugin and Python package version numbers are independent; runtime compatibility is not automatically established by matching their version strings.

The integration uses existing Server interfaces and adds no HTTP/MCP endpoints or database tables.

Known qualification limits

The integration is documented as community/experimental.

Official MiniMax CLI 0.6.9 and 0.7.2 can report an enabled, connected plugin without making its MCP tools available to the model. Both versions reproduce the plugin-only failure in the affected Linux environment, including through deferred tool discovery in 0.7.2. Reliable plugin-only invocation remains unqualified for those unmodified official npm packages.

The default installation, retaining the enabled private MCP entry together with the complete plugin, passes cold and preflight execution on official 0.7.2: the model receives the Skill and invokes a tool that returns the seeded Scope ID/title. This qualifies that installed configuration, not the isolated plugin route. Setup continues to preserve existing disabled preferences.

The missing discovery/publication barrier before turn capability capture is fixed in MiniMax main source and the public 0.7.2 synchronization source. Turn preparation waits for current MCP discovery and publication, and final capture rechecks readiness. These community source builds pass in the affected environment; the integration guide documents the pinned 0.7.2 build and its launcher through the existing --minimax-command option. The upstream source changes live outside this PowerContext PR.

Upstream issue #570 tracks adoption. Official npm 0.6.9 and 0.7.2 do not contain the fix. The original native thread remains open to track official-release qualification; the source fix and its affected-environment validation are complete.

Native macOS, Windows, MiniMax Desktop, and external model-provider workflows require separate qualification.

How was this change tested?

Focused CLI and model-gateway regressions:

.venv/bin/python -m pytest \
  tests/test_cli_minimax.py \
  tests/test_minimax_model_gateway.py \
  --doctest-modules -q
# 105 passed, 1 skipped

Coverage includes target selection, complete package installation, upgrades and rollback, user-file conflicts, disabled endpoints, credential preservation and rotation, connection synchronization, native discovery failures, and read-only diagnostics.

Opt-in real-host acceptance:

POWERCONTEXT_MINIMAX_ACCEPTANCE_COMMAND=/absolute/path/to/mcode \
  .venv/bin/python -m pytest tests/e2e/test_minimax_host.py -q \
  --basetemp=/data/codex-tmp/powercontext-minimax-acceptance

The 2026-10-29 reference run at f4345d04 used official @minimax-ai/code@0.6.9 on Linux x86_64 with Node.js 22.23.3 and passed twenty isolated-route executions. Independent affected-environment checks after synchronizing master reproduce plugin-only failures with both official 0.6.9 and 0.7.2; their private-only cases pass. The added default-installation cases pass on 0.7.2 under cold and preflight startup. These bounded results are recorded separately in integrations/minimax/acceptance.md; the known isolated plugin failures remain strict failures.

The MiniMax source fix was validated on 2026-10-31 in the original failing Linux environment with Node.js 22.23.3. The main-based build passed all six controlled invocation cases; the 0.7.2 source build passed the complete thirteen-case real-host suite. Cold plugin-only execution also passed with 6-second and 11-second MCP inventory delays, with the private entry absent and the original Scope/tool-call-ID assertions intact. Eight upstream regression tests pass with the fix; seven reproduce failures on the unchanged main runtime.

The main-based source commit passed the full applicable Linux pnpm verify profile on clean committed source. The 0.7.2 build and focused turn-capability file passed; the full upstream verifier was not repeated on that synchronization branch. These source-build results do not qualify an unmodified npm release, other operating systems, Desktop or external model providers.

Native tests use real files, subprocesses, a PowerContext Server, SQLite, authenticated MCP, and a controlled local inference endpoint. They also cover repeated setup, disablement preservation, loader omission, authentication failures, and diagnostics that leave the real profile unchanged. Simulated CLI output in unit tests is not counted as native qualification.

Repository hooks, including Ruff and type checks, passed:

uv run prek run -a

The acceptance record documents the independent failure and the limits of local passing results. Required CI checks remain a separate merge requirement.

AI usage statement

OpenAI Codex assisted with source tracing, implementation, regression-test design, failure investigation, documentation, and validation. The resulting changes were reviewed locally.

[hidb4ai]

One non-blocking documentation issue in the protected-server instructions.

[hidb4ai]

please update & resolve conflicts

[hidb4ai]

LGTM

[hireview001]

• Resolve merge conflicts with the target branch.

[hireview001]

• CI failed: Upgrade release gate (failure).
• CI failed: upgrade-acceptance / Prepare upgrade artifacts and matrix (failure).
• CI failed: Main (failure). Failed steps: upgrade-acceptance / Prepare upgrade artifacts and matrix / Run upgrade gate regressions; Upgrade release gate / Require the selected parallel acceptance workflow.

[hidb4ai]

The native acceptance test does not reliably distinguish the private and plugin MCP routes. Please address the reproduced issue below before relying on the documented two-route qualification.

[hidb4ai]

[P2] Isolate the MCP route instead of assuming a _2 alias

With the documented @minimax-ai/code@0.6.9 on Linux, the native execution log shows the plugin MCP server connected, but the model receives mcp__powercontext__list_scopes, without _2. This filter then selects no tool for plugin_connection=True, and test_real_host_invokes_powercontext_mcp_with_controlled_inference[plugin-mcp] fails its seeded-scope assertion in both the full run and an isolated rerun. The private-mcp case leaves the same plugin enabled and its execution log also shows the plugin connection, so its passing result does not establish the other route. Please make each case enable only its intended connection, select the actual discovered tool, and verify the returned Scope before claiming both native paths in integrations/minimax/acceptance.md. The equivalent _2 checks at lines 278 and 443 need the same treatment.

[hidb4ai]

Follow-up at dfa6d6a5c85c07e35e75cc9afee655e1717bd7e6: the alias assumption has been removed, but the plugin-only native qualification still fails on the documented Linux x86_64 / official @minimax-ai/code@0.6.9 / Node.js 22.23.3 environment.

POWERCONTEXT_MINIMAX_ACCEPTANCE_COMMAND=/path/to/mcode python -m pytest tests/e2e/test_minimax_host.py -k test_real_host_invokes_powercontext_mcp_with_controlled_inference -q

The private-mcp parameter passes; plugin-mcp fails at line 473 with Expected one isolated list_scopes tool, discovered []. The bundle URL/Authorization and native plugin-enabled assertions pass, and the native log records an MCP connection, but the model request exposes no PowerContext MCP tools. The controlled model then returns its final response without invoking MCP. The same failure also reproduces in the full suite and an isolated plugin-case rerun with Node 22.22.3.

Please reconcile native tool-discovery readiness/configuration with this fixture and demonstrate the authenticated plugin call while retaining the Scope ID/title assertion. integrations/minimax/acceptance.md:38-40 currently claims both isolated cases return the seeded Scope; that qualification is still not reproducible here. This evidence establishes a failing native acceptance path; it does not establish a Server connection failure.

[hidb4ai]

Status recheck on 2026-10-29: the PR head is still dfa6d6a5c85c07e35e75cc9afee655e1717bd7e6, with no new commits since the reproduction above. Main workflow attempt 2 completed with failure; the PR still shows 11 failed checks and merge conflicts. The native qualification P2 remains unresolved. Its evidence is the previously executed real-host tests on this identical commit.

There are no added HTTP/MCP endpoints or database tables; MiniMax extends the existing CLI entry points. Merge readiness remains 40/100, so the approval conditions are still unmet.

[AlexStocks]

补充 head 前进后的状态证据（不另开线程，本条根因与本线程一致）。

head 已从 dfa6d6a5 前进到 f38965e1，增量只含两个作者提交：f4345d04（test/docs）与 f38965e1（integrations/minimax/acceptance.md），无产品代码改动；该 head 76 个 check run 为 75 success + 1 skipped，无 failed。

关于 _2 别名：tests/e2e/test_minimax_host.py 在当前 head 已不含任何 _2 字面量（dfa6d6a5 时也已不含），路由不再靠连接名后缀推断，而由参数化的 plugin_connection 显式区分，证据文件新增 "route": "plugin-mcp" | "private-mcp"（test_minimax_host.py:623），并配合 len(set(gateway.invoked_tools)) == 1 与 plugin 分支的 powercontext-project-context 断言。

同时 acceptance.md 把此前“未定位根因”的表述收紧为可复现的实验结论：只延迟 /mcp/ 的 tools/list 6 s 时，冷启动 plugin-only 用例失败，模型在约 3.90 s 时收到 18 个内置工具、无 PowerContext 工具，Server 于 6.04 s 才返回 56 个工具；零延迟对照暴露 74 个工具并通过。该文档只声明这是“注入延迟下的启动竞态”，未声称解释外部自然失败，也未把未提交的本地 readiness 原型当作官方 CLI 验收。这部分表述与代码/测试一致，不再构成额外的定级依据。

[hidb4ai]

Fresh diagnostic recheck at f38965e10eac66847a493064a6638ad69df65705 in the originally failing Linux environment, using the unchanged official npm @minimax-ai/code@0.6.9 and Node.js 22.23.3, with no injected latency:

POWERCONTEXT_MINIMAX_ACCEPTANCE_COMMAND=/path/to/mcode python -m pytest tests/e2e/test_minimax_host.py -k controlled_inference -q

Both private cases return the seeded Scope ID/title; both plugin cases fail with Expected one isolated list_scopes tool, discovered []. The new capture produced the following sanitized evidence. All times are milliseconds relative to each native invocation's first observed MCP HTTP request, using the same Python monotonic clock. Setup/preflight traffic was cleared before execution, and HTTP 307 redirects are omitted. Every listed HTTP 200 inventory contains 56 tools, including list_scopes.

Startup / route	tools/list starts	HTTP 200 inventory finishes	First model catalog received	Model tools: total / PowerContext	Seeded Scope matched
Cold / private	84.61	195.59	1141.79	74 / 56	Yes
Cold / plugin	136.80	214.86	1422.50	18 / 0	No
Preflight / private	143.66	389.29	1265.84	74 / 56	Yes
Preflight / plugin	115.29	195.45	1720.14	18 / 0	No

In the failing plugin cases, invoked_tools is empty and scope_result_matched is false. The inventory finished 1207.64 ms and 1524.69 ms before the corresponding model request was received. Unlike the injected-delay example, these failures do not have an inventory response still pending at model-request receipt. This capture does not establish when the host constructed its tool snapshot; native capability capture/publication still needs correlation. No credentials, headers, tool arguments, or tool-result contents are included here.

The documentation now states the qualification limit accurately, and CI/merge conflicts have cleared. The official-host native qualification issue remains reproducible, so I am keeping this P2 open; the separate source-build readiness prototype does not establish a fix for these executions.

[hidb4ai]

Rechecked the affected Linux environment after merging master 0c1dfc289, and pushed the acceptance updates in 319e389d3d61bb800cf3d8d7541e9e4522e54f7f. Both hosts are unmodified official npm packages, using Node.js 22.23.3, without injected latency.

Official CLI	Private cold/preflight	Plugin cold/preflight
0.6.9	Seeded Scope ID/title returned	No Scope tool available
0.7.2	Seeded Scope ID/title returned	Empty tool_search result

Every native inventory response included 63 tools and list_scopes. The plugin cases still fail Expected one isolated list_scopes tool, discovered []; using the canonical /mcp/ URL also fails. Inspection of the shipped 0.7.2 runtime shows that readiness can finish before background MCP discovery/publication, while the turn retains the capabilities captured at acquisition. A later inventory response does not replace that captured view.

The new, separately labelled installed-default cases retain both the enabled private entry and complete plugin. Official 0.7.2 passes cold and preflight execution, with the seeded Scope ID/title tied to the actual tool-call ID and the Skill present. This establishes a working default configuration, not plugin-only qualification. The original isolated-route assertions remain strict; neither vendor patching nor private-route substitution is used to make them pass. Keeping this P2 open until a released host fix passes the isolated cases in this environment.

[hidb4ai]

Implemented and pushed the host fix: main-based commit 50ee352 and 0.7.2 source commit 0ea69bd. Turn preparation now waits for MCP discovery and inventory publication, with cancellation and a bounded readiness wait; synchronous capture rechecks the current generation.

The original failing Linux environment now completes cold and preflight plugin-only invocation using the fixed source, returning the seeded Scope ID/title with the matching tool-call ID and delivering the Skill. Cold plugin-only runs also pass with 6-second and 11-second inventory delays. The private MCP entry stays absent, and the strict route assertions are unchanged. Full applicable Linux verification passed on the main-based fix.

PowerContext commit c88aca0 documents the pinned build, launcher setup and acceptance evidence. Upstream #570 tracks adoption under the upstream contribution policy. Official npm 0.6.9/0.7.2 still lack the patch, so this thread remains open for official-release qualification. No HTTP/MCP endpoints or database tables were added.

[hireview001]

• Resolve merge conflicts with the target branch.

[AlexStocks]

结论：无新增 Finding；head 的 11 条红与本 PR 变更面无关

head dfa6d6a5（31 文件，+2437/-32），tests/test_cli_minimax.py 84 passed、tests/test_minimax_model_gateway.py --doctest-modules 8 passed；desktop.yml 的 macOS 打包改动语义等价（该 job 有 defaults.run.working-directory: desktop，pnpm tauri build --ci -- --locked 与原 pnpm desktop:build 同义，新增仅 --verbose + tee）。两条旧 P2 已在各自线程内给出闭合证据。

11 条红 job 的归因（均不在本 PR 变更面内）

• 签名 A：quality 与 tests (3.11|3.12|3.13|3.14, unit) 都是 upgrade_release_gate.GateError: Persistence changed (file set or content; CRLF/LF normalized)，出自 test_repository_baseline_covers_all_three_databases；Upgrade release gate 是该 job 的聚合。
• 签名 B：tests (*, e2e) 都是 tests/e2e/test_owner_recovery_http.py::test_http_creation_and_restart_recover_the_original_owner → MigrationError: incompatible_schema；upgrade-acceptance / Prepare upgrade artifacts and matrix 同源。
• 本 PR 31 文件不含 persistence/、server/identity.py、server/authz、limits.py（MANAGED_INPUTS 全集）。

定量证据（managed_digest(root) 对 release/upgrade-support.json 的声明值）

树	实测	声明
master aedf7beb	99d448c2…	99d448c2… 一致
本 PR base 927bb6ee	f0a1d364…	9b67d639… 不一致
bad4c0c2 / refs/pr/2220	9b67d639…	一致

即 base 本身已不再过门禁；修复来自 master 上 de435aa3、59572c6b（#2212，注册 durable owner recovery storage 与 payload charset）。这两个签名属"上游修复后本 PR 未 rebase"的基线滞后，rebase 到当前 master 即可消除，不需要在本 PR 内改动。

[hidb4ai]

The native qualification P2 remains open: the isolated plugin MCP case still fails on the documented CLI 0.6.9 / Linux / Node 22.23.3 environment. I have added reproduction evidence to the existing thread.

Merge readiness: 40/100. There are also conflicts in .github/workflows/desktop.yml and desktop/VALIDATION.md against master aedf7beb, and 11 failed checks. The persistence-digest mismatch and Owner recovery schema failure reproduce on base 927bb6ee; those are inherited baseline failures. Please sync master, preserve the qualification records when resolving conflicts, and obtain passing required checks.

The PR adds no HTTP/MCP endpoints or database tables; it extends the existing setup/doctor/config CLI with MiniMax support. Approval conditions are not met.

[hireview001]

• Resolve merge conflicts with the target branch.
• CI failed: Upgrade release gate (failure).
• CI failed: tests (3.11, e2e) (failure).
• CI failed: tests (3.12, unit) (failure).
• CI failed: tests (3.13, unit) (failure).
• CI failed: tests (3.11, unit) (failure).
• CI failed: tests (3.14, unit) (failure).
• CI failed: tests (3.12, e2e) (failure).
• CI failed: tests (3.13, e2e) (failure).
• CI failed: tests (3.14, e2e) (failure).
• CI failed: upgrade-acceptance / Prepare upgrade artifacts and matrix (failure).
• Additional CI issues: PR checks.

[Lkx-JY]

Local investigation reproduced a plugin tool-discovery race: the host can construct the model’s tool catalog before plugin MCP discovery and inventory publication complete.
A source-level readiness fix makes the controlled 11-second discovery-delay case pass while preserving the seeded Scope ID/title assertions. The full native acceptance suite also passes locally.
This validation uses a build of the public MiniMax 0.6.5 source. It does not establish that official npm 0.6.9, or the originally failing environment, is fixed.
Could you rerun the cold plugin-only case with the diagnostic capture and share the sanitized discovery timeline and acceptance evidence? We need to correlate tools/list completion with the first model tool catalog. The original P2 should remain open pending validation in that environment.

[hidb4ai]

The existing native qualification P2 remains reproducible on official MiniMax CLI 0.6.9 / Node 22.23.3: cold and preflight private cases return the seeded Scope, while both plugin cases expose no PowerContext tools to the model. The existing thread now contains sanitized discovery/model timelines from this head. The documentation correctly records the limitation; no separate new installer defect was found.

Merge readiness: 55/100. The PR is mergeable and the previous CI failures have cleared, but native plugin-route acceptance is still unresolved, so approval conditions are not met. No HTTP/MCP endpoints or database tables are added; the CLI gains MiniMax host support.

[hidb4ai]

Target-branch recheck at unchanged PR head f38965e10eac66847a493064a6638ad69df65705: master has advanced to 90355b8e96d68e2932d7980b3e76957f35ea2b44. GitHub reports CONFLICTING, and git merge-tree --write-tree --name-only confirms a content conflict in desktop/VALIDATION.md, in the Compatibility scope paragraph. Please sync master and resolve it using the current compatibility scope while retaining the historical qualification evidence, then rerun the required checks.

The existing native qualification P2 remains open. Its evidence comes from the previously executed native tests on this identical head; the unchanged suite was not rerun for this status check. Merge readiness is now 45/100, so the existing request for changes remains appropriate. No HTTP/MCP endpoints or database tables are added; the existing CLI commands gain MiniMax support.

[hidb4ai]

Pushed the conflict resolution and repository fixes to 319e389d3d61bb800cf3d8d7541e9e4522e54f7f on this PR's existing branch, preserving the author's commits.

The merge retains both current and historical Desktop qualification records. The Code publication checks now use the intended database explicitly, and the persistence compatibility declaration matches the merged implementation without relaxing the installed-release upgrade gate. Added native coverage and bilingual guidance for the normal MiniMax installation, with both the private MCP entry and complete plugin enabled; also corrected the invalid-manifest fixture for newer host validation.

The installed configuration returns the seeded Scope ID/title and loads the Skill on official MiniMax 0.7.2 under both cold and preflight startup. The plugin-only P2 remains reproducible with official 0.6.9 and 0.7.2 and requires a supported host fix. Its strict assertions remain intact. No HTTP/MCP endpoints or database tables are added relative to current master.

Local hooks/type checks and the affected regressions passed. GitHub now reports the branch as mergeable; checks for the new commit are pending. The three previously fixed review threads have been resolved; the native P2 remains open.