feat(integrations): add MiniMax Code (mcode) agent integration (#4644)

[philo-x]

Description

Adds MiniMax Code (an AI coding assistant CLI, binary mcode) as a supported Spec Kit integration under the canonical key mcode.

MiniMax Code discovers project skills natively at .minimax/skills/<skill-name>/SKILL.md (runtime source workspace-minimax, priority 65) and invokes them via its slash shortcut (/speckit-<command>). This wires it up as a SkillsIntegration with its own isolated .minimax/skills layout (multi_install_safe=True).

Headless dispatch uses mcode exec "<prompt>" --permission full so non-interactive Spec Kit workflows can run scripts and edit files. Operators can select smart through SPECKIT_INTEGRATION_MCODE_EXTRA_ARGS="--permission smart". Optional --model and --output-format json flags are also supported.

Closes #4644

Test selection reasoning

Changed file	Affects	Test	Why
src/specify_cli/integrations/mcode/__init__.py (+ registry wiring)	new mcode integration, specify init scaffolding	T1	New integration install path; CLI scaffolding change requires at minimum /speckit.specify
src/specify_cli/_invocation_style.py	slash-invocation rendering in init next-steps and hook invocations	T1	Invocation display feeds every command's guidance
src/specify_cli/command_init.py	specify init next-steps output	T1	Init scaffolding change
integrations/catalog.json	specify integration search/info discovery	T2, T3	Catalog discovery and metadata
docs/reference/integrations.md, issue templates, agent-context-defaults.json	docs / issue forms / agent-context extension	—	No slash-command behavior

Required tests

• T1: /speckit.specify (invoked as /speckit-specify) — new mcode integration install path + init scaffolding + live agent skill execution
• T2: specify integration search mcode — catalog discovery
• T3: specify integration info mcode — catalog metadata

Manual test results

Current E2E agent: MiniMax Code mcode 0.5.4 from @minimax-ai/code@0.5.4 | OS/Shell: macOS / zsh

The earlier manual checks in the table below used mcode 0.4.12; @latest was the version selector at that time. The current 0.5.4 dispatch check is recorded in the first row.

Command tested	Notes
Current dispatch E2E: McodeIntegration.dispatch_command("specify", "Create a health check endpoint specification with status and uptime", ...) using mcode 0.5.4	PASS — the production dispatch_command() path invoked MiniMax Code headlessly with --permission full, exited 0, and generated nonempty specs/001-health-check-endpoint/spec.md and checklists/requirements.md in a fresh Spec Kit project.
Setup: specify init ... --integration mcode --non-interactive --script sh	PASS — mcode executable detected on PATH (no --ignore-agent-tools needed); scaffolded 10 skills under .minimax/skills/speckit-*/SKILL.md; next steps correctly rendered dash-notation slash commands /speckit-constitution, /speckit-specify, /speckit-plan, etc.
Runtime skill discovery: mcode exec "What skills are available in this project?"	PASS — MiniMax Code actively scanned .minimax/skills/ and listed all local Spec Kit workflow skills (speckit-specify, speckit-clarify, speckit-plan, speckit-tasks, etc.) as project-local skills.
T1: mcode exec --permission full "/speckit-specify Add health check endpoint with uptime and memory metrics"	PASS — Full agent workflow execution verified: MiniMax Code discovered /speckit-specify, ran the feature allocation script, generated specs/001-health-check-endpoint/spec.md, completed checklists/requirements.md, and persisted feature state to .specify/feature.json.
T2: specify integration search mcode	PASS — found MiniMax Code (mcode) v1.0.0 with tags cli, skills.
T3: specify integration info mcode	PASS — reported MiniMax Code (mcode) v1.0.0, author spec-kit-core, Built-in integration.

Earlier 0.4.12 log excerpt: MiniMax Code execution of /speckit-specify (T1)

Specification complete and validated. Here's the summary:

## Completion Report

**Feature**: Health Check Endpoint (feature #001)

- **SPECIFY_FEATURE_DIRECTORY**: specs/001-health-check-endpoint
- **SPEC_FILE**: specs/001-health-check-endpoint/spec.md
- **Checklist**: specs/001-health-check-endpoint/checklists/requirements.md
- Feature context persisted to .specify/feature.json for downstream commands

**What the spec covers**
- 3 prioritized user stories: P1 manual status view (status + uptime + memory), P2 automated monitoring integration (stable machine-readable responses), P3 early warning when memory usage is high
- 8 functional requirements (FR-001–FR-008), each testable
- 5 measurable, technology-agnostic success criteria
- Key entities: Health Status, Uptime, Memory Usage
- Assumptions documenting defaults and out-of-scope items

**Validation results**: All 16 quality checklist items passed on iteration 1. No [NEEDS CLARIFICATION] markers remain.

Automated test results

Ran with .venv/bin/python -m pytest per CONTRIBUTING guidance.

Integration & consistency suites

.venv/bin/python -m pytest \
  tests/integrations/test_integration_mcode.py \
  tests/integrations/test_registry.py \
  tests/test_agent_config_consistency.py -q

Earlier result: 896 passed in 48.11s.

Full test suite

.venv/bin/python -m pytest tests -q

On current commit 89ce4104, rebased onto upstream/main at c00dc055 (with network access for existing online catalog tests): 8,499 passed, 213 skipped. Focused mcode integration tests on the previous equivalent tree: 41 passed. Ruff 0.15.0, the bundled extension version guard against upstream/main, and git diff --check upstream/main HEAD all passed on 89ce4104. The dispatch regression test exercises the actual dispatch_command() path and checks the spawned argv. Both catalog top-level updated_at values are refreshed.

AI Disclosure

• I did not use AI assistance for this contribution
• I did use AI assistance (fill in the disclosure below)

AI disclosure: Prepared with AI assistance. ChatGPT (GPT-5.6 Sol) was used to review the integration design, validate the Spec Kit and MiniMax Code architecture, and identify the relevant integration surfaces. Antigravity (Gemini 3.8 Flash, High, autonomous) was used on behalf of @philo-x for implementation, test authoring, end-to-end mcode workflow verification, and documentation alignment. Codex (GPT-6 and GPT-6 Sol, autonomous) was used on behalf of @philo-x for review fixes, rebasing and scope cleanup, validation, and the final commit.

[Copilot]

Copilot review overview

🟡 Changes recommended

The unconditional hook invocation behavior lacks coverage for the ai_skills=False path.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 1

Open (1)

• Add mcode regression coverage for both ai_skills states · New

What changed in this PR

Adds MiniMax Code (mcode) as a skills-based CLI integration with catalog, documentation, invocation, and context support.

Changes:

• Registers and configures .minimax/skills scaffolding and headless dispatch.
• Adds slash-style invocation and init guidance.
• Updates discovery metadata, issue templates, documentation, and tests.

File	Description
src/​specify_cli/​integrations/​mcode/​__init__.py	Implements the integration.
src/​specify_cli/​integrations/​__init__.py	Registers mcode.
src/​specify_cli/​commands/​init.py	Adds MiniMax next-step guidance.
src/​specify_cli/​_invocation_style.py	Adds slash invocation behavior.
integrations/​catalog.json	Adds catalog metadata.
extensions/​agent-context/​agent-context-defaults.json	Maps mcode to AGENTS.md.
docs/​reference/​integrations.md	Documents MiniMax Code.
.github/​ISSUE_TEMPLATE/​feature_request.yml	Adds agent option.
.github/​ISSUE_TEMPLATE/​bug_report.yml	Adds agent option.
.github/​ISSUE_TEMPLATE/​agent_request.yml	Updates supported-agent list.
tests/​integrations/​test_integration_mcode.py	Tests integration behavior.
tests/​integrations/​test_registry.py	Verifies registration.
tests/​test_agent_config_consistency.py	Verifies configuration consistency.

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[philo-x]

The existing tests/integrations/test_integration_zed.py::TestSlashSkillsSets matrix includes mcode with both ai_skills=True and ai_skills=False, each expecting /speckit-plan; both cases pass. Posted on behalf of @philo-x by Codex (model: GPT-6 Sol, autonomous); comment fully AI-drafted.

[philo-x]

Follow-up commit 45afe5ac adds mcode to the existing hook-rendering matrix for both ai_skills=True and False, bumps agent-context to 1.0.1 in both its manifest and catalog, and removes the trailing blank line.

Validation: 140 passed, 1 skipped across the slash matrix/Zed, mcode, agent config consistency, and extension version guard contract tests. The extension version guard and git diff --check against the PR base also pass. Ruff was unavailable locally.

Please re-run Copilot review on 45afe5ac. The pending CI workflows are currently awaiting approval.

Posted on behalf of @philo-x by Codex (model: GPT-6, autonomous). This follow-up implementation, validation, and comment were AI-generated; comment fully AI-drafted.

[Copilot]

Copilot review overview

🔵 Needs a closer look

The model test uses an invalid CLI model reference, and the new context self-seeding mapping lacks regression coverage.

Review effort: Balanced
Findings: 1

Open (1)

• Add mcode regression coverage for both ai_skills states

Previously missed (2)

In code that hasn't changed since last review

Cover mcode self-seeding across supported agent-context implementations

extensions/​agent-context/​agent-context-defaults.json:28

The new mcode self-seeding path is not exercised by the agent-context tests: the existing defaults-map tests use only codex/claude. Add mcode to the self-seed parity coverage and assert that the Bash, Python, and PowerShell implementations select AGENTS.md; otherwise a typo in this entry can pass the reported suite while leaving the extension with no target file.

Use a valid provider/model reference for MiniMax Code

tests/​integrations/​test_integration_mcode.py:54

Use a valid MiniMax Code model reference here. mcode exec --model requires provider/model (or provider/model#variant), so MiniMax-Text-01 would be rejected by the CLI; this test currently proves only string passthrough rather than usable model dispatch.

[mnriem]

Please address Copilot feedback and fix test & lint errors

[philo-x]

Review round summary for commits ef465add, 7a4b8ab1, and 4d6805ea (current head).

Aligned the agent-context extension, catalog, and example bundle pins at 1.0.2 after merging current main. Kept main's consolidated bundle tests with dynamic version lookup. Added mcode self-seeding parity coverage for Bash, Python, and PowerShell, and changed the model test to the locally exercised custom_provider:ark/deepseek-v4-flash reference. The earlier ai_skills=True/False hook matrix coverage remains in 45afe5ac. Made two online step-reference tests use a controlled empty catalog so network reachability no longer changes their assertions.

Validation on head 4d6805ea: .venv/bin/python -m pytest tests -q — 8,605 passed, 17 skipped; Ruff 0.15.0, extension version guard against current main, and git diff --check — passed. The mcode provider/model reference completed a local CLI smoke call. Please review the updated PR.

Posted on behalf of @philo-x by Codex (model: GPT-6, autonomous). The changes, validation, and this comment were AI-generated; comment fully AI-drafted.

[Copilot]

Copilot review overview

🟢 Approval recommended

The integration is consistently wired across runtime, catalogs, documentation, context handling, and comprehensive tests.

Review effort: Balanced
Findings: 1

Open (1)

• Add mcode regression coverage for both ai_skills states

[mnriem]

Please rebase on the latest upstream/main so we only have your changes

[philo-x]

Rebased this PR onto upstream/main 5cc1f2a8 as commit 7201d528. The branch now has one commit and 21 changed files. Removed the unrelated remote-catalog test change while retaining the mcode integration, agent-context 1.0.2 release alignment, example bundle pins, and mcode regression coverage.

Validation on 7201d528: .venv/bin/python -m pytest tests -q — 8,437 passed, 213 skipped (network access enabled for existing online catalog tests). Ruff 0.15.0, the bundled extension version guard, and git diff --check passed. The PR description now reflects this validation.

Posted on behalf of @philo-x by Codex (model: GPT-6, autonomous). The rebase, validation, and this comment were AI-generated; comment fully AI-drafted.

[Copilot]

Copilot review overview

🟡 Changes recommended

The integrations reference omits mcode from its authoritative multi-install-safe table.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 2

Open (2)

• Add mcode to the multi-install safe integrations table · New
• Add mcode regression coverage for both ai_skills states

[philo-x]

Review-round update for 10919886: added mcode (.minimax/skills) to the multi-install-safe integrations table, addressing the latest review feedback. Rebased the single PR commit onto upstream main at 03d833f7; no additional scope changes.

Validation on 10919886: 1,029 passed, 6 skipped across the mcode integration, registry, agent configuration, Zed, and agent-context parity tests. The bundled extension version guard and git diff --check also passed. The earlier full-suite result remains documented separately in the PR description.

Posted on behalf of @philo-x by Codex (GPT-6, autonomous). The change, validation, and this comment were AI-generated.

[Copilot]

Copilot review overview

🟡 Changes recommended

The integration catalog’s top-level update timestamp remains stale.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 2

Open (2)

• Update catalog top-level updated_at when adding the entry · New
• Add mcode regression coverage for both ai_skills states

Resolved since last review (1)

• Add mcode to the multi-install safe integrations table

[philo-x]

Review-round update for d1ce5a9: refreshed the catalog's top-level updated_at for the mcode entry. The other finding is already covered in tests/integrations/test_integration_zed.py::TestSlashSkillsSets: the existing hook-rendering matrix contains mcode cases for both ai_skills=True and ai_skills=False, each expecting /speckit-plan. Validation: all 59 matrix and mcode integration tests passed; 894 catalog/registry tests passed; git diff --check passed. Please review the updated PR. Posted on behalf of @philo-x by Codex (model: GPT-6 Sol, autonomous). The timestamp change, validation, and this comment were AI-generated; comment fully AI-drafted.

[philo-x]

Review-round summary for 216a8a55: rebased onto upstream/main adbd62af and consolidated the PR into one commit. Refreshed both catalog timestamps, including extensions/catalog.json alongside the agent-context 1.0.2 update. The existing TestSlashSkillsSets matrix covers mcode with both ai_skills=True and ai_skills=False.

Validation on the final tree:

• Full test suite: 8,446 passed, 213 skipped
• Focused mcode, invocation-style, agent-context parity, and bundled-extension tests: 134 passed, 6 skipped
• Ruff 0.15.0: passed
• Extension version guard: passed
• git diff --check: passed

The PR description has also been updated to reflect the final HEAD and validation results. Please review the updated PR.

Posted on behalf of @philo-x by Codex (model: GPT-6 Sol, autonomous). The rebase/commit cleanup, validation, and this comment were AI-assisted; comment fully AI-drafted.

[Copilot]

Copilot review overview

🟡 Changes recommended

Headless dispatch does not apply the permission mode used by the successful end-to-end test and can fail on unresolved permission requests.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Headless dispatch omits explicit permissions, causing workflow failures · New
• Add mcode regression coverage for both ai_skills states

Resolved since last review (1)

• Update catalog top-level updated_at when adding the entry

[philo-x]

The production build_exec_args() now emits --permission full. The regression test calls dispatch_command() and checks the argv passed to subprocess.run(). I also ran that path with MiniMax Code 0.5.4 in a temporary Spec Kit project; /speckit-specify completed and generated the spec and requirements checklist. This is included in commit 5967398f. Posted on behalf of @philo-x by Codex (model: GPT-6 Sol, autonomous); comment fully AI-drafted.

[philo-x]

Review-round summary for 20f4bf76: MiniMax Code headless dispatch now passes --permission full, matching the successful manual workflow and allowing script and file operations without an unresolved interactive permission request. The integration reference documents the mode and the SPECKIT_INTEGRATION_MCODE_EXTRA_ARGS="--permission smart" override. Added a regression test through dispatch_command() that checks the spawned argv.

The earlier slash-invocation feedback is also covered by the existing TestSlashSkillsSets matrix, which includes mcode with both ai_skills=True and ai_skills=False, each expecting /speckit-plan.

Validation on this commit: 94 passed, 1 skipped across the mcode and slash-invocation suites; git diff --check passed. The prior full-suite result is retained in the PR description and labeled with its earlier commit. The full suite and Ruff were not rerun here.

Please review the updated PR.

Posted on behalf of @philo-x by Codex (model: GPT-6 Sol, autonomous). The change, validation, PR description update, and this comment were AI-generated.

[philo-x]

Review-round summary for 89ce4104 (rebased onto upstream/main at c00dc055): corrected the mcode extra-args test to use valid --timeout 60s --max-steps 20 options and clarified the headless permission wording in the integration reference. The existing TestSlashSkillsSets matrix covers mcode with both ai_skills=True and False; the production dispatch_command() path uses --permission full and has a regression test. I replied to the two corresponding Copilot threads where clarification was useful.

Validation on 89ce4104: full suite 8,499 passed, 213 skipped; focused mcode tests on the preceding equivalent tree 41 passed; Ruff 0.15.0, bundled extension version guard, and git diff --check passed. A real dispatch_command("specify", ...) run with MiniMax Code 0.5.4 exited 0 and generated a nonempty specification and requirements checklist. The PR description now identifies the 0.5.4 E2E and distinguishes earlier 0.4.12 checks. Five remote workflows still await maintainer approval. Please review the updated PR.

Posted on behalf of @philo-x by Codex (model: GPT-6 Sol, autonomous). The code/documentation changes, validation, and this comment were AI-generated; comment fully AI-drafted.