claudemax

Ask. Achieve. The canonical user-facing entry to claudemax — describe your goal, the SOTA-2026 pipeline (deepresearch → multispec → parallel /goal → verify) handles the rest. Same engine as /cmax with a friendlier verb. CLI equivalent is `cmax ask "<goal>"`.

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.

Max-effectiveness umbrella. Opus for EVERYTHING — planning, decomposition, sub-Spec /goal execution, verification. Use for novel domains, security/auth/payments, or when you want the highest ceiling regardless of cost.

Show the model-routing decision for a task (Fable/Opus/Sonnet/Haiku) with reasoning. Use before launching a packet to sanity-check the tier and to override when you have stronger judgment than the heuristic.

Reviews a proposed diff for correctness, security, style, taste alignment. Distinct from /audit (which reviews existing code, not diffs) and /verify (which checks SPEC completion).

Spec quality gate. Every completion condition must have a mechanically-checkable verifyHint (file path, command, test name, behavior). Blocks /goal handoff if the spec wouldn't be verifiable.

Independent Opus supervisor pass that re-reads the repo blind and re-checks every SPEC completion condition. The source of truth for whether a run actually succeeded. Never skip.

Iterative web-current research with source ledger. Sonnet collects, Opus synthesizes. Parallel WebSearch + WebFetch across sub-queries. Output is a structured ResearchBrief with citations, persisted to memory.research_sources. The default first step of every umbrella when the goal is novel or time-sensitive.

SOTA-2026 deepresearch-backed audit of new Claude Code CLI releases. Triggered on every CC update (or when a third-party signal — tweet, blog, changelog mirror — claims a new feature). Fetches the official changelog + Anthropic docs + GitHub release notes, verifies claims against primary sources, decides for each change IGNORE / WRAP / INTEGRATE / DEFER. Records the verdict so the harness never ships a conflicting or redundant tool.

Durable cross-session memory for claudemax. SQLite+FTS5 hybrid-retrieval store with 5-tier CoALA taxonomy, multi-scope identity, and staleness handling. Use `recall` before claiming prior-session facts, `add` to record durable lessons, `verify` to refresh a recalled row.

Session-limit-aware mega-build orchestrator. One command takes N goals, auto-sizes parallel lanes from your hardware (os.availableParallelism + free RAM) and plan tier, runs each as an independent cmax-run subprocess with checkpointing, pauses cleanly on rate-limit saturation, and can be resumed by `cmax resume` (cron-friendly). Distinct from `/orchestrate` (no auto-sizing, no resume), `/parallel` (one goal split into packets), and `/hive` (same problem to N agents). Use when you have many independent goals to ship and want the harness to manage the envelope for you.

Resume a paused `cmax mega` run. Picks up only the lanes still pending or paused; skips finished. cron-friendly — exits 0 with no-op if nothing to do. Use after a saturation pause, or wire into systemd timer for fully-unattended resume across rate-limit windows.

Multi-goal harness — runs N `/cmax` pipelines in parallel for DIFFERENT goals at the same time. Each goal gets its own deepresearch → multispec → /goal → blind /verify, independent state, independent verdict. Rollup status at end. Effectiveness-driven defaults (--tdd --confidence 0.85). Use when you have several distinct things to ship at once and don't want to babysit them sequentially. Distinct from /parallel (which fans out one goal across N sub-Specs) and /hive (which gives N agents the SAME problem).

Periodic review of claudemax's own scaffolding against the current Opus capability. Flags load-bearing vs vestigial pieces so the harness stays sharp as models improve. Run after every major model release or quarterly, whichever comes first.

Test-driven cycle enforced as a hard sequence — write the failing test FIRST, implement only enough to make it pass, then prove the test passes. Opt-in via `cmax ask "<goal>" --tdd` or invoke `/tdd <goal>` directly. Designed for completion conditions whose verifyHint names a test command; for behavior-only verifyHints it explicitly says so instead of faking.

Governed AgentDefinition creation. Writes a typed agent spec to `agents/<name>.json` and registers it in a Hermes-style index at `agents/registry.json` for reuse across runs.

Mode B parallelism. Manually invoke Claude Code Agent Teams (multiple full Claude Code instances coordinating via shared task list + worktree isolation + Agent View). The multispec engine auto-selects this mode when the work demands it; this skill is for forcing it.

One-shot semantic decision recorder. Writes topic + decision + rationale to memory.decisions for future sessions to recall. Tiny skill; load-bearing for cross-session continuity.

Adversarial code-quality scan of existing code — look for holes, smells, edge cases, missing tests. Read-only. Distinct from /review (which scans proposed diffs) and /investigate (which root-causes a specific bug).

Multi-pattern search with relevance ranking. Sharper than raw Grep/Glob for "find me places that look like X". Haiku tier. Returns ranked hits with context windows.

3-Opus adversarial debate (proposer / critic / judge) for architectural decisions. Use when a decision is irreversible and benefits from explicit adversarial framing.

Drift detection. Re-runs /taste against current code and reports what changed between the recorded taste.vision and the repo's current state. Use periodically (monthly?) or after large refactors.

Autonomous Opus loop that pursues a SPEC across multiple turns until every completion condition is satisfied or a real blocker surfaces. Use only when the work is non-decomposable and benefits from sustained reasoning. Always preceded by /spec.

Same problem given to N agents in parallel; merge proposals. Use when you want N independent attempts at the same goal and a synthesized merge. Distinct from /parallel (N different things) and /council (adversarial debate with roles).

Confidence/assumption hard-gate before /goal handoff. Outputs confidence 0-10, assumptions, unknowns. **Blocks /goal if confidence < 6.** Cheap (Opus, ~5 turns); catches false-confidence wrong turns.

Multi-source root-cause analysis for a specific bug or failure. Depth-first on one symptom. Use when something is broken and you need the actual cause, not a workaround.

Long-running mode with file checkpointing + session resumption. `cmax overnight <spec> --budget-credits N` runs until the spec is done or the budget is exhausted, snapshotting on every turn so a SIGTERM or crash doesn't lose work.

Distinct-packet fan-out. Each packet does a different thing; the engine runs them concurrently up to the hardware + credit-aware cap. Use when work decomposes into independent tasks. Distinct from /hive (same problem N times) and /council (3-Opus debate).

Final go/no-go gate combining /verify (against SPEC) + /review (of the diff). Does NOT commit unless user explicitly authorizes. The last check before merge.

Write SPEC.md with measurable completion conditions before any code runs. Use Opus. Triggered automatically by /cmax. Run it standalone when you want a contract for a goal before committing to execution.

Auto-bootstrap project taste.md + taste.vision via /deepresearch. NO 10-question wizard. Reads repo signals + researches SOTA at current time → synthesizes taste docs with zero questions (one fallback question only if repo has no signal).