feat(acp): add MiniMax Code as a builtin Registry npx agent

[kaizhou-lab]

Summary

Adds MiniMax Code as a builtin ACP agent. It was listed on the ACP Registry in snapshot v2026.09.12-e6ee445 (the Registry's 41st id) with an npx distribution, which makes it the first new agent since the Registry sync started that qualifies for automatic integration under the standing rules — the two earlier arrivals (antigravity-acp, kimchi) were binary-only and remain deferred.

This PR is intentionally separate from today's lock-only bump (#989) so that bump kept its routine self-merge shape. This PR is for human review: it adds a migration, a logo, and a lock entry.

item	value	evidence
Public Registry card	MiniMax Code → https://agent.minimax.io, version 0.2.7	registry.md card; CDN JSON id: minimax-code, name: MiniMax Code, license: MIT
Distribution	npx @minimax-ai/code@0.2.7, args ["acp"]	CDN JSON; no binary or uvx distribution
Product CLI (binary_name)	mcode	official README: npm install -g @minimax-ai/code → mcode; mcode --version documented, so the default PATH probe applies (no skip_version_probe)
ACP entrypoint	mcode acp	README "ACP clients" section; Registry declares the same acp arg
Chosen backend	minimax-code	product name + the agent's own agentInfo.name from the live handshake; follows the mimo-code precedent (029) for " Code" products. Not a copy of the JSON lookup id, which merely coincides
npm package identity	@minimax-ai/code, bin mcode at 0.2.7; engines >=22.19 <23 || >=24 <27	npm view; AionCore's managed Node is 24.11.0 (MANAGED_NODE_VERSION), inside the range
Probe — initialize	ok, protocolVersion 1, agentInfo {name: minimax-code, title: MiniMax Code, version: 0.2.7}	acp_probe.mjs against npx -y @minimax-ai/code@0.2.7 acp, clean temporary HOME, no inherited credentials
Probe — session/new	-32000 "Authentication required: Run mcode login and try again."	clearly classified auth requirement (rubric level: Starts; session catalog not observable unauthenticated)
agentCapabilities	loadSession: true; mcpCapabilities {http: true, sse: true}; promptCapabilities all false; sessionCapabilities {list, fork, resume, close}	seeded snake_case in the migration per the migration 003 contract
authMethods	none advertised by initialize	auth_methods left NULL — nothing is synthesized. The session error itself tells the user to run mcode login

What the migration seeds and why each nullable field is what it is

044_add_minimax_code_builtin_agent.sql (template: 031, policy shape: 034):

• id/agent_id = ec619063 (mirrored per the post-030 NOT NULL + UNIQUE rule), user_id NULL.
• command: npx, args: ["-y","@minimax-ai/code","acp"], env: []. The exact version is not stored here; pin_registry_npx_args rewrites the package token to @minimax-ai/code@0.2.7 from the lock at spawn.
• agent_source_info: {"binary_name":"mcode","bridge_binary":"npx"} — no registry_json_id / package_name in metadata.
• agent_capabilities: seeded from the probe (snake_case). Not in the ON CONFLICT DO UPDATE list, nor is auth_methods, so a re-seed can never reset what a live handshake later learns.
• yolo_id: NULL. Source-verified against the 0.2.7 ACP server chunk (run-acp-command-*.js): availableModes is exactly default and plan. The README's Ask / Auto / Full access levels are exposed as a _permission-category config option (permissionMode = default | auto | bypassPermissions), not as a session mode, and AionCore's yolo channel resolves through session/set_mode — the same chunk rejects unknown mode ids with "Unsupported Session mode" — so storing bypassPermissions would break unattended runs rather than enable them. See open question 1.
• native_skills_dirs: NULL. The 0.2.7 source scans only $MINIMAX_DATA_DIR/skills (user data dir) and the bundled assets/skills; no project-relative skills directory exists, and the README documents none.
• behavior_policy: {"supports_side_question":false} — no supports_team, no team_capable_override (retired by 033). Team capability derives from the seeded mcp_capabilities.
• The four session-derived columns are not seeded (skill step 11); the availability probe fills them on first check.

Other files:

• crates/aionui-runtime/resources/acp-registry-npx-lock.json — new entry minimax-code → @minimax-ai/code 0.2.7 with registry_json_id alias.
• crates/aionui-assets/assets/logos/acp-registry/minimax-code.svg — the Registry's official icon (…/registry/v1/latest/minimax-code.svg, 169 bytes, currentColor fill), keyed by backend; added to registry_agent_logos_are_embedded_as_svg.
• crates/aionui-db/tests/minimax_code_builtin_agent_migration.rs — new: launch fields, handshake columns (seeded caps, NULL auth/yolo/skills, policy shape), lock pin is an exact semver under the backend key, and a bad-path test that mcode / minimax / minimax-ai / @minimax-ai/code do not resolve to a builtin row.
• crates/aionui-db/tests/team_capability_criteria_migration.rs — minimax-code added to the mcp-capabilities table (http+sse) and to the NULL-auth-methods-by-design list.
• crates/aionui-db/src/repository/sqlite_agent_metadata.rs — seed row count 43 → 44.
• crates/aionui-runtime/src/registry_npx_lock.rs — the every_lock_entry_has_an_exact_version guard counts lock entries; 12 → 13. Its exact-semver check now also covers the new entry.
• crates/aionui-ai-agent/src/registry.rs — three more count guards: hydrate_loads_seed_rows (43 → 44), every_builtin_npx_agent_has_a_release_lock (12 → 13 pinned npx builtins), diagnostic_snapshot_pairs_rows_with_reasons (43 → 44), and list_by_agent_type_counts_seed_rows (ACP rows 39 → 40). CI surfaced these one fail-fast round at a time (753/9388, then 772/9388 tests run); the whole registry::tests module was then run locally by filter (30 passed) before the final push.

Open questions for review

1. Unattended/Team runs. With yolo_id NULL, AgentType::full_auto_mode_id falls to its _ => "yolo" default, which this agent will reject as an unsupported mode — identical to the current situation for omp, dimcode, kilo, dirac, grok and every other Registry agent seeded without a yolo_id. The vendor's actual unattended switch is the permissionMode=bypassPermissions config option; supporting that would be an AionCore launch-configuration/config-option path, not a metadata value, and is out of scope here.
2. Sign-in affordance. Because initialize advertises no authMethods, the UI cannot offer a login button until the runtime learns otherwise; the user gets the agent's own actionable error ("Run mcode login"). If a later version starts advertising methods, the live handshake will populate the column.
3. Registry pin lags upstream. npm latest is already 0.4.5 (the vendor ships several times a day); the Registry declares 0.2.7 and the lock follows the Registry, as for every other agent. The nightly sync will bump it as the Registry moves.
4. mcode-tools. Newer package versions (≥0.3.x) add a second bin; 0.2.7 ships mcode only. Nothing here depends on it.

AionUi issues

None linked. Per the workflow, issue linking requires an explicit request; no AionUi issue was searched, edited, or closed.

Validation

• just migration-check — pass
• Targeted tests (run with name filters, per the repository rule against unfiltered local runs):
  • cargo test -p aionui-db --test minimax_code_builtin_agent_migration
  • cargo test -p aionui-db --test team_capability_criteria_migration
  • cargo test -p aionui-db --lib seed_rows_populated_after_migrations
  • cargo test -p aionui-assets --lib registry_agent_logos_are_embedded_as_svg
  • cargo test -p aionui-runtime --lib registry_npx_lock
  • cargo test -p aionui-ai-agent --lib -- registry::tests (whole module, 30 passed)
• just lint-fix (cargo fix + clippy --fix --workspace -D warnings) — clean
• just fmt — clean
• Local full cargo nextest was not run: this PR was prepared during working hours, when the full workspace suite is not permitted on this host; the Test check on this PR is the authority. This is the reason the PR is left open for review rather than self-merged, in addition to it not being a lock-only change.

Logging

No logging changes: metadata-only integration. Existing startup/session error paths already identify a failing agent by backend, and the agent's own session/new error carries the actionable instruction.