opusworkflow

name	opusworkflow
description	Use this as the definitive workflow command for mutating work: Opus 4.7 high/xhigh plans and reviews when proven available, and MiniMax-M2.7-highspeed executes bounded packets. It must drive to a verified result, partial result, or blocked repair path; /opusminimax is the advanced engine underneath, not a competing daily command.

/opusworkflow

Run the definitive effectiveness-first, cost-aware workflow for:

$ARGUMENTS

Definitive workflow command:

Use /opusworkflow for normal build, fix, refactor, docs, and specialist work.
It means: Opus 4.7 high/xhigh thinks and reviews when proven available.
MiniMax-M2.7-highspeed is the executor for the bulk implementation work.
The run must end as a verified result, partial result, or blocked repair path.
Use /opusworkflow unless you are debugging the engine.

The two names:

/opusworkflow = the product command humans should use.
/opusminimax is the advanced engine inside /opusworkflow.
They are not competing commands.

Mode banner:

Opus 4.7 high/xhigh is planner, adversary, and final reviewer when proven available.
MiniMax-M2.7-highspeed is the executor for bulk coding and repair.
Default executor concurrency is 1 until provider evidence proves otherwise.
Closeout policy is verified, partial, or blocked-with-repair. No vibes.

Optional Claude-only sibling:

/opussonnet keeps the same workflow governance but requests Claude Code opusplan:
Opus 4.7 planning/judgment and Sonnet 4.6 execution, with no MiniMax token.

Model-profile selector:

--model-profile minimax    # default: Opus judgment + MiniMax execution
--model-profile opussonnet # Opus judgment + Sonnet execution
--model-profile sonnet     # Sonnet planning + Sonnet execution
--model-profile opus       # Opus planning + Opus execution, explicit high-cost route
--model-profile default    # Claude Code account default
--model-profile custom --planner-model MODEL --executor-model MODEL

Contract

• Treat /opusworkflow as the definitive workflow command and the one normal command for mutating work.
• Optimize for results, not only cost: keep going through research, plan, packet execution, repair loops, verification, and introspection until the run can honestly close as verified, partial, or blocked.
• If blocked, record the blocker and the next repair action. Never close with evidence-free optimism.
• Record outcome_policy=verified-partial-or-blocked-with-repair in run artifacts produced under the /opusworkflow outer route.
• Treat /opusworkflow as /opusminimax --mode workflow with stricter budget defaults. /opusminimax is the advanced packet/provider engine behind it.
• Do not present /opusminimax as a second daily workflow choice unless the operator is debugging engine, provider, packet, repair, or benchmark behavior.
• Keep MiniMax as the standard executor provider. Use /opussonnet or --executor-provider claude-sonnet only when the operator explicitly wants the optional Claude-only route.
• Allow explicit model freedom through --model-profile; treat it as a governed route request, not runtime identity proof.
• Record the specialist being executed as inner_contract=workflow|agentfactory|hiveworkflow|parallel|defineicp|deepretaste|demo|visualizeworkflow.
• Run /specqa after SPEC.md is created, updated, or reused and before implementation. Spec QA is an Opus 4.7 high/xhigh judgment gate when runtime identity is proven; otherwise record the missing proof and downgrade or block honestly.
• If the task asks for Hermes, Hive, ICP/taste mutation, approved visualization implementation, demo artifact production, or dense packet work, preserve that specialist contract under the /opusworkflow outer route.
• Do not claim Opus planned, reviewed, or verified unless auth/model evidence or the run artifact proves it.
• Do not read .env, .env.*, .claude/settings.local.json, .claude/*.local.json, secrets/**, private credentials, customer artifacts, or MiniMax key files.
• Do not let the planner inherit ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic.
• Do not use Opus for bulk file reading, formatting, repetitive edits, or retry loops.
• Do not use the full local parallel ceiling by default. MiniMax Token Plan capacity, supervisor review capacity, and verifier capacity are the real bottlenecks.
• Default MiniMax executor concurrency to 1 for Plus-Highspeed unless the provider doctor/runtime evidence proves a higher safe ceiling.
• Run /introspect before freezing the plan, after MiniMax execution, after failed verification, and before push or ship decisions.
• Run /verify against SPEC.md before closeout.

Budget Policy

Use Claude/Opus only at high-leverage judgment gates:

Phase	Default Model	Rule
Intake and taste check	Claude normal effort	Keep concise.
Repo audit and exploration	local tools or MiniMax packet	Avoid Opus-wide repo reads.
Plan and SPEC freeze	Opus 4.7 if proven available	Use high/xhigh only here when warranted.
Spec QA	Opus 4.7 if proven available	Review requirements quality, SOTA/currentness, critical risks, and improvement suggestions before execution.
Implementation	MiniMax-M2.7-highspeed	Bounded packets only.
Test failure loop	MiniMax first	Limit retries per packet.
Adversarial review	Opus 4.7 if proven available	Challenge architecture, risk, and evidence.
Final ship/no-ship	Opus 4.7 if proven available	Verify claims, do not rubber-stamp.

Practical target for the $20 Claude + $40 MiniMax setup:

80-90% mechanical work: MiniMax-M2.7-highspeed
10-20% judgment work: Opus 4.7 high/xhigh when account state proves it
executor lanes: 1 default, 2 only after explicit runtime proof

Optional opussonnet target:

planner/judgment: claude-opus-4-7 through opusplan
execution: claude-sonnet-4-6 through opusplan/Sonnet profile
MiniMax token: not required

Explicit Anthropic-only profiles:

sonnet: claude-sonnet-4-6 for planning and execution
opus: claude-opus-4-7 for planning and execution; use intentionally
default: Claude Code account default; confirm with /status before claims
custom: explicit planner/executor model IDs; static gates only prove request shape

Workflow

1. Run provider/capacity preflight:

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

2. Create or update the workflow artifact and SPEC.md using the normal /workflow lifecycle.

3. Run /specqa against the active SPEC.md. Record spec_qa_model_identity_status, requested/proven reviewer model, webresearched actual-time data source ledger, critical findings, improvement suggestions, and execution-allowed decision. Do not claim Opus 4.7 reviewed the spec unless runtime identity evidence proves it.

4. Route implementation through /opusminimax packet artifacts:

bash scripts/opusworkflow.sh --task "$ARGUMENTS"

For specialist mutation, pass the contract explicitly:

bash scripts/opusworkflow.sh --task "$ARGUMENTS" --inner-contract agentfactory
bash scripts/opusworkflow.sh --task "$ARGUMENTS" --inner-contract hiveworkflow

For the optional Claude-only executor:

bash scripts/opusworkflow.sh --task "$ARGUMENTS" --executor-provider claude-sonnet

For explicit model profiles:

bash scripts/opusworkflow.sh --task "$ARGUMENTS" --model-profile sonnet
bash scripts/opusworkflow.sh --task "$ARGUMENTS" --model-profile opus
bash scripts/opusworkflow.sh --task "$ARGUMENTS" --model-profile custom --planner-model claude-sonnet-4-6 --executor-model claude-sonnet-4-6

5. MiniMax executes only planner-approved packets with owned paths, forbidden paths, allowed commands, acceptance checks, rollback notes, and stop conditions.

6. Parent Claude verifies:

• diffs and touched files
• command evidence
• tests and logs
• packet ownership
• unresolved failures
• final confidence

Runtime Opt-In

Static preparation is the default. Runtime model calls require explicit opt-in:

bash scripts/opusworkflow.sh --task "$ARGUMENTS" --execute-planner

For a tiny Opus planner identity check, use:

claude --model claude-opus-4-7 --settings .claude/settings.opusminimax-planner.local.json -p 'Reply exactly: OPUSWORKFLOW_AUTH_OK'

Only treat the Opus planner path as runtime-proven when the response is exactly OPUSWORKFLOW_AUTH_OK or an equivalent run artifact proves the model identity.

The planner must still refuse Opus claims if the actual model identity is not proven. If Claude Code falls back to Sonnet or the subscription does not expose Opus, record that honestly and continue with downgraded confidence or ask for runtime/account action.

Spec QA uses the same proof rule. requested_reviewer=claude-opus-4-7 is not runtime proof; only /status, a sentinel, or a durable model-identity artifact can set claims_opus_review=true.

Anti-Patterns

• "One command install plugs my subscription in" without claude auth login.
• ANTHROPIC_API_KEY present while intending subscription billing.
• Planner and executor sharing the MiniMax base URL.
• Ten MiniMax executor lanes on Plus-Highspeed because the machine supports ten local agents.
• Opus doing every edit because it feels smarter.
• MiniMax summaries accepted without parent verification.
• /specqa skipped after SPEC.md and before execution.
• Pay-as-you-go fallback enabled silently.