opusminimax

name	opusminimax
description	Advanced engine behind /opusworkflow for provider split, packet, repair, and benchmark work, including the explicit sonnetminimax profile for Sonnet 4.6 judgment plus MiniMax-M2.7-highspeed execution. Normal build/fix/refactor work should use /opusworkflow.
argument-hint	["task",{"mode":"workflow|benchmark|repair"}]
disable-model-invocation	true

/opusminimax

Run the governed Opus planner + MiniMax executor mode for:

$ARGUMENTS

Plain rule:

If you are trying to build, fix, refactor, or ship a product, use /opusworkflow.
Use /opusminimax directly only for engine, provider, packet, repair, or benchmark work.
/opusworkflow is the human command; /opusminimax is the advanced engine.

Mode banner:

Claude is planner, adversary, and reviewer.
MiniMax-M2.7-highspeed is executor for bulk coding.
Worker summaries are claims until verified by diffs, logs, tests, or artifacts.

Non-Negotiable Contract

• Do not treat this as a model alias switch.
• Do not claim Opus planned, reviewed, or verified unless the planner profile, auth/model preflight, or runtime artifact proves it.
• Do not let the planner inherit ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic.
• Do not run MiniMax execution from the planner phase unless the user explicitly enters the executor phase.
• Do not read .env, .env.*, .claude/settings.local.json, .claude/*.local.json, secrets/**, private credentials, customer artifacts, or MiniMax key files.
• Use /deepresearch for architecture, provider, benchmark, harness, or high-stakes implementation decisions. Use /claudeproduct first for Claude product facts.
• Treat local capacity, Codex max_threads, and MAX_PARALLEL_AGENTS as ceilings. Treat MiniMax Token Plan capacity as the executor bottleneck until verified.
• Default MiniMax executor concurrency to 1 unless provider-tier evidence proves a higher ceiling.
• If executor_provider=claude-sonnet is explicit, treat it as the optional Claude-only /opussonnet route: no MiniMax base URL, executor model must be Sonnet, and the run artifact must not imply MiniMax executed anything.
• If model_profile=sonnetminimax is explicit, treat it as the governed hybrid route: planner/reviewer requests claude-sonnet-4-6, executor provider is minimax, executor model is exactly MiniMax-M2.7-highspeed, and the planner profile must still have no MiniMax base URL or MiniMax credential fields.
• If /sonnetminimax is explicit, treat it as the short power-user wrapper for model_profile=sonnetminimax with default effort max, intended for saving Opus quota while keeping MiniMax-M2.7-highspeed Token Plan execution.
• If /opusolo or model_profile=opus is explicit, treat it as the optional all-Opus route: no MiniMax base URL, planner and executor must both request Opus, default effort is high for /opusolo, and --effort max is an explicit highest-effort alias rather than the default.
• If model_profile=sonnet|opus|default|custom is explicit, treat it as a governed Anthropic-only route: no MiniMax base URL, no MiniMax executor model, and no runtime model identity claim without /status, sentinel, or artifact proof.
• When outer_route=opusworkflow, record plan_mode in the run artifact. Default plan_mode.policy=auto means auto_approved_when_gates_pass: the workflow may transition from plan to execution only after research, code audit, /introspect pre-plan, Agent-Native Estimate, SPEC.md, and /specqa allow execution.
• Run /introspect before plan freeze, after executor execution, after failed verification, and before push or ship decisions.
• Run /verify against SPEC.md after executor aggregation.

Phase 0: Provider And Capacity Preflight

Use static proof first:

bash scripts/opusminimax-doctor.sh --static
bash scripts/parallel-capacity.sh --json

When runtime planner execution is requested, diagnose and repair safe local profile drift before model invocation:

bash scripts/opusminimax-doctor.sh --runtime --fix-local-profiles

If auth, account access, ANTHROPIC_API_KEY, or model availability still blocks the planner, stop with repair steps. Do not silently downgrade or claim Opus.

Record:

• planner profile path and whether it is provider-neutral
• executor profile path and whether it uses MiniMax-M2.7-highspeed
• selected model_profile and whether it is default cost-optimized or an explicit user override
• requested planner model
• requested executor model
• local capacity ceiling
• MiniMax provider ceiling or unverified-default-1
• effective executor budget

Budget formula:

effective_executor_budget = min(
  local_recommended_ceiling,
  provider_executor_ceiling,
  independent_packets,
  supervisor_review_capacity,
  verifier_capacity
)

Phase 1: Research And SPEC

For file-changing work, follow the repo's normal lifecycle:

1. Taste/current-state check.
2. Metacognitive route with capacity evidence.
3. /deepresearch or local-only research brief.
4. Code audit before planning.
5. /introspect pre-plan.
6. Concrete plan.
7. Agent-Native Estimate.
8. SPEC.md active contract.
9. /specqa execution-allowed decision and plan_mode transition checkpoint before implementation packets run.

Do not skip /workflow discipline. /opusminimax wraps the workflow with a provider split and packet contract.

Phase 2: Executor Packetization

MiniMax only receives bounded packets. Write or request packets with:

{
  "artifact_type": "opusminimax-packet",
  "run_id": "YYYYMMDD-HHMMSS-task",
  "packet_id": "P1",
  "objective": "One bounded implementation objective",
  "context_summary": "Planner-provided context, not full hidden reasoning",
  "owned_paths": ["path/or/glob"],
  "forbidden_paths": [".env", ".claude/*.local.json", "secrets/**"],
  "commands_allowed": ["bash scripts/specific-check.sh"],
  "acceptance_checks": ["observable check"],
  "risk_notes": ["what can go wrong"],
  "rollback_plan": "How to revert this packet safely",
  "expected_outputs": ["diff", "worker-result", "logs"],
  "stop_conditions": ["ambiguous ownership", "secret requested"]
}

Hard packet rules:

• One owner per packet.
• Clear write ownership; no overlapping files unless a sync barrier exists.
• No broad ** ownership for implementation packets unless the packet is explicitly read-only.
• No credential paths in packet context.
• No benchmark gold/hidden data in executor-visible prompts.

Phase 3: MiniMax Execution

Use the executor bridge:

bash scripts/minimax-exec.sh --packet .taste/opusminimax/{run_id}/packets/P1.json --run-dir .taste/opusminimax/{run_id}

The bridge validates the packet and writes a sidecar. Runtime model calls are opt-in:

bash scripts/minimax-exec.sh --packet ... --run-dir ... --execute

Do not trust the executor closeout by default. Required evidence:

• diff or touched files
• commands run
• logs or summaries tied to commands
• acceptance checks
• unresolved failures
• parent verification status

Phase 4: Claude Review And Verification

Claude/Opus reviewer duties:

• Compare executor output to SPEC.md.
• Verify claimed commands and changed files.
• Reject out-of-ownership changes.
• Run or inspect static gates appropriate to the change.
• Record failed verification as blocked, not positive closeout.
• Run /introspect post-implementation.

Final status must be one of:

• verified: static/runtime evidence supports the claim
• blocked: execution cannot be trusted yet
• partial: some packets accepted, others rejected
• runtime-pending: static contract passes, provider runtime not exercised

Benchmark Repair Mode

Use for SWE-bench-style or Terminal-Bench-style tasks only when the user asks for benchmark/repair mode.

Pipeline:

task intake -> gold/hidden quarantine -> localization -> repro test attempt ->
MiniMax candidate patches -> validation -> Claude adversarial selection ->
prediction/export

Benchmark rules:

• The solver sees only visible prompt and pre-fix repo state.
• Gold patches, hidden tests, and judge-only data stay quarantined.
• Aggregate scores require per-task result artifacts.
• Static harness evals are not benchmark proof.
• Do not claim SWE-bench, Terminal-Bench, SWE-Bench Pro, or private benchmark strength without reproducible per-task evidence.

Required Run Artifact

When /opusminimax executes or prepares a real run, produce:

{
  "artifact_type": "opusminimax-run",
  "run_id": "YYYYMMDD-HHMMSS-task",
  "outer_route": "opusworkflow",
  "inner_contract": "workflow",
  "model_profile": "minimax",
  "executor_provider": "minimax",
  "planner_identity_status": "blocked",
  "executor_identity_status": "configured",
  "fallback_status": "none",
  "provider_profiles": {
    "planner": {"path": ".claude/settings.opusminimax-planner.example.json"},
    "executor": {"path": ".claude/settings.minimax-executor.example.json"}
  },
  "model_ids": {
    "planner_requested": "claude-opus-4-7",
    "executor_requested": "MiniMax-M2.7-highspeed"
  },
  "model_route": {
    "profile": "minimax",
    "planner": {"provider": "anthropic", "requested_model": "claude-opus-4-7", "identity_status": "blocked"},
    "executor": {"provider": "minimax", "requested_model": "MiniMax-M2.7-highspeed", "identity_status": "configured"},
    "fallback_policy": "fail-closed-unless-explicit"
  },
  "plan_mode": {
    "enabled": true,
    "policy": "auto",
    "checkpoint": "pre-implementation-plan-approval",
    "auto_approval": {
      "status": "auto_approved_when_gates_pass",
      "execution_allowed_after": [
        "research_brief_recorded",
        "code_audit_recorded",
        "pre_plan_introspection_pass",
        "agent_native_estimate_recorded",
        "spec_created_updated_or_reused",
        "specqa_execution_allowed"
      ]
    },
    "does_not_replace": ["SPEC.md", "/specqa", "/introspect", "/verify", "runtime_model_identity_proof"]
  },
  "capacity": {
    "local_ceiling": 10,
    "provider_ceiling": 1,
    "task_packet_count": 1,
    "safety_cap": 1,
    "effective_concurrency": 1
  },
  "packets": ["P1"],
  "verification": {"status": "runtime-pending"},
  "failures": [],
  "retries": 0,
  "final_confidence": "medium"
}

For the SonnetMiniMax hybrid, the same schema records:

{
  "model_profile": "sonnetminimax",
  "executor_provider": "minimax",
  "model_ids": {
    "planner_requested": "claude-sonnet-4-6",
    "executor_requested": "MiniMax-M2.7-highspeed"
  },
  "spec_qa": {
    "requested_reviewer": "claude-sonnet-4-6",
    "identity_status": "blocked"
  }
}

Validate sidecars:

bash scripts/artifact-lint.sh .taste/opusminimax/{run_id}/opusminimax-run.json
bash scripts/artifact-lint.sh .taste/opusminimax/{run_id}/packets/P1.json

Anti-Patterns

• "Opus planned this" with no model identity evidence.
• Planner and executor sharing one provider profile.
• Using the full 10-lane local ceiling because it looks powerful.
• Giving MiniMax an unbounded repo-wide task.
• Calling static fixtures benchmark proof.
• Hiding runtime failures behind "ready" or "done".
• Letting executor summaries replace parent verification.