| name | cmax |
| description | The default FAT umbrella. Auto-runs deepresearch + multispec decomposition + parallel /goal + verify. Opus for planning/judgment/verification; Sonnet for sub-Spec execution (escalates per router). Multispec engine selects Mode A (SDK subagents) or Mode B (Claude Code Agent Teams) automatically based on work size. Your default daily-driver. |
/cmax — default daily-driver
The umbrella. You describe what you want; the pipeline produces it — 0 → production-ready, autonomously, on a new or an existing project (it runs in the current directory).
One simple entry: cmax ask "<your goal>" (= /cmax). Hardened by default — production-ready bar (PRC) + spec self-correction (SSC) + adversarial, decomposed, evidence-gated verify. No flags to remember. Lighten only if you want: --mvp (ship an MVP), --no-ssc, --no-adversarial. Granular control lives on cmax run.
Ask. Achieve. No 10-question wizards, no model picker, no decomposition-by-hand. The SOTA-2026 spec-driven + deepresearch-backed pipeline runs automatically.
Pipeline (default behavior, no flags)
- Taste check — read
taste.md + taste.vision if present. If absent and the repo has signal, suggest cmax taste init. Don't violate taste.
- /deepresearch — web-current research with source ledger. Persisted to
memory.research_sources.
2-bis. verify-research (default ON, added 2026-06-09) — decomposed per-claim verification of the brief: parallel Haiku citation-support checks on the brief's own cached excerpts (no network, 60s timeouts), Sonnet escalation (one spot-fetch + one fresh search, 120s) only for findings the excerpts can't support. Contradicted findings are stripped into openQuestions; unverified ones flagged [unverified]; >30% escalation demand fails the brief and re-researches once. Timeouts count as unverified, never as pass.
- multispec decompose — Fable 5 (while included →2026-06-22; Opus after) authors a MultiSpec: N sub-Specs (each with measurable completion conditions + verifyHints) + DAG + rollup completion conditions + per-sub-Spec writeSet.
- /specqa (parallel per sub-Spec) — gate spec quality.
- /introspect (parallel per sub-Spec) — confidence/assumption check. Block at confidence < 6.
- Mode selection — engine picks Mode A (SDK subagents in one query()) or Mode B (Claude Code Agent Teams) based on sub-Spec count, est. duration, cross-spec coordination, write-set overlap. Override with
--mode {auto|solo|teams}.
- Parallel /goal per DAG leaf — every leaf at the current frontier runs in parallel up to the hardware + credit-aware cap. Sonnet executes sub-Specs by default; Opus on router escalations.
- per-sub-Spec /verify (parallel) — blind Opus re-reads repo, re-checks each completion condition.
- rollup /verify — blind Opus checks the rollupCompletionConditions across all sub-Spec outputs.
- Memory record — run snapshot to
memory.runs + state to .claudemax/state/.
Variants
| Skill | Routing |
|---|
/cmax | Default — Fable 5 plan/decompose (while included on Max, →2026-06-22; Opus after) + Sonnet/Opus exec per era + Opus verify |
/opusolo | Opus exec for sub-Specs too (max effectiveness, ~3× cost) |
The /workflow + /opussonnet aliases were removed 2026-05-28; the opussonnet/opusolo exec-tier variants remain via cmax run --variant. Fable 5 defaults auto-demote to Opus on 2026-06-23 when Fable bills usage credits (CMAX_FABLE_ACCESS=included|credits flips the configurations); cmax loop run --fable / --tier fable reach Fable explicitly in both modes — see docs/MODEL_ROUTING.md.
Hard rules
- Never declare success on the strength of the executor's own claim. The blind /verify is the source of truth.
- Verify and spec always run on Opus, never demoted regardless of credit % or
--cheap.
- If /verify returns partial or failed, iterate /goal against the failing conditions; re-verify. If two iterations don't move it, the SPEC is probably wrong — re-spec, don't re-grind.
- Cost-guard kicks in at 70% of monthly Agent SDK credit (demote non-essential Opus → Sonnet); at 90% aggressive demote; at 95%
--force required.
When NOT to invoke /cmax
- Trivial one-line changes — just edit.
- A single SPEC that fits in one /goal loop — invoke
/goal directly on the existing SPEC.md.
- Pure research with no implementation — invoke
/deepresearch directly.