Skip to content

V1_TO_V2_MIGRATION.md

Latest commit

 

History

History
92 lines (72 loc) · 6.18 KB

V1_TO_V2_MIGRATION.md

File metadata and controls

92 lines (72 loc) · 6.18 KB
Raw

minmaxing v1 → claudemax v2 migration

For users of waitdeadai/minmaxing coming to claudemax. Big-picture: v2 is Anthropic-only, lean (29 active skills vs v1's 43), and built around the Claude Code 2.1.139 /goal validator-loop + multispec decomposition.

What changed at the top

Concern v1 v2
Provider Anthropic + MiniMax Anthropic-only
Daily-driver /opusworkflow (Opus plan + MiniMax exec) /cmax (Opus plan + Sonnet exec) — semantically same as v1 /opussonnet
Spec Single SPEC.md MultiSpec (DAG of sub-Specs + rollup conditions) — default behavior
Autonomy Custom workflow loop Wraps Claude Code's native /goal (validator-loop)
Parallelism Hardware-aware packets Two modes auto-selected: SDK subagents (Mode A) or Claude Code Agent Teams (Mode B)
Verification /verify skill + agentcloseout-physics /verify (blind Opus) + dark-patterns evidence hooks
Taste /tastebootstrap 10-Q wizard /taste auto-bootstrap via /deepresearch (NO questions)
Skills 43 skills 29 active skills + 1 deprecated stub (overlap-audited; v0.2.1 added /tdd + /harness-audit)
Memory SQLite + FTS5 (5 tiers) SQLite + FTS5 (same + research_sources, taste_history, sub_specs)
Billing model Per-token USD estimates Plan-aware: detect Max5x/Max20x; show "$X • Y% of monthly credit"
Auth API key Subscription Agent SDK credit by default (Max 5x/20x); API key opt-out

Skill mapping (v1 → v2)

v1 skill v2 status Notes
/opusworkflow Renamed/folded Same behavior as v1 /opussonnet once MiniMax → Sonnet. v2 uses /cmax / /workflow / /opussonnet (aliases)
/opussonnet KEPT v1 muscle memory; routes Opus plan + Sonnet exec
/opusolo KEPT v1 muscle memory; all-Opus mode
/opusminimax, /sonnetminimax DROPPED MiniMax-specific
/workflow KEPT (alias for /cmax) v1 muscle memory
/sonnetonly, /hiveworkflow CUT Overlap with /opussonnet + /hive
/deepresearch KEPT + enhanced Source ledger persisted to memory.research_sources
/webresearch MERGED into /deepresearch /deepresearch uses WebSearch/WebFetch by default
/audit KEPT Adversarial code-quality scan
/introspect KEPT Confidence/assumption hard-gate before /goal handoff
/codesearch KEPT Multi-pattern search with ranking
/investigate KEPT Multi-source bug root-cause
/autoplan MERGED into multispec engine Engine auto-decomposes
/specqa KEPT Spec quality gate
/qa MERGED into /verify verifyHints can name test commands
/review KEPT Diff review
/ship KEPT Final go/no-go combining /verify + /review
/parallel KEPT User-facing skill for distinct-packet fan-out
/hive KEPT Same problem N times → merge
/council KEPT 3-Opus adversarial debate
/agentfactory KEPT AgentDefinition registry
/tastebootstrap REPLACED by /taste auto-bootstrap NO 10 questions; uses /deepresearch
/deepretaste KEPT Drift detection
/digestaste, /digestflow CUT (internal functions) Used by SessionStart/Stop hooks; not user-facing
/align KEPT Semantic decision recorder
/memory KEPT Persistent store
/overnight KEPT Long-running mode w/ checkpoint resume
/agent-view, /remote-control, /goal-mode DOCUMENT as native CC Built into Claude Code; just document
/agentteams NEW Manual invocation of Claude Code Agent Teams (Mode B)
/sprint, /defineicp, /icpweek, /claudeproduct, /metacognition, /leveragepath, /browse, /demo, /visualize, /visualizeworkflow DEFERRED Not core to daily SW work in v2

Behavioral differences

/cmax / /opussonnet is now FAT

In v1, /opussonnet was a thin model-routing variant of /workflow. In v2, every umbrella auto-runs the full pipeline: deepresearch + multispec decompose + specqa + introspect + parallel /goal per leaf + per-sub-spec verify + rollup verify. They differ only in sub-Spec exec tier.

Multispec is the default, no flag

v1 had /sprint and various decomposers. v2 makes multispec decomposition the default behavior of every umbrella — no --multi flag. The engine picks single-spec mode internally only if the goal fits in one Spec.

Two parallelism modes auto-selected

v1 had hardware-aware packets. v2 adds Mode B (Claude Code Agent Teams) for big swarm work with shared task list + worktree isolation. The multispec engine auto-selects based on sub-Spec count / est. duration / cross-spec coordination / write-set overlap.

Verifier never demotes

v1 sometimes demoted /verify under cost pressure. v2 hard rule: /verify, /spec, /architect always run on Opus, regardless of --cheap or monthly credit %.

/taste is auto-bootstrap, NOT 10 questions

v1's /tastebootstrap asked 10 kernel questions. v2's /taste: reads repo + /deepresearch SOTA at current time → writes taste.md + taste.vision with zero questions. One fallback question only if the repo has no signal.

Migration steps

  1. Install v2: git clone … && pnpm install && pnpm build && ln -s … cmax
  2. Per-project init: cmax init in your project root. Writes .claude/skills/* + .claude/hooks/*.
  3. Plan auto-detection: cmax doctor confirms Max5x/Max20x detection. Override with CMAX_PLAN=max20x if auto-detect fails.
  4. Dark-patterns plugin: claude plugin marketplace add waitdeadai/claude-plugins && claude plugin install llm-dark-patterns@waitdeadai-plugins.
  5. First run: cmax run "<small goal>" to confirm the multispec pipeline works end-to-end.
  6. Re-derive taste: cmax taste init to replace any v1 taste.md with the auto-bootstrap version.

What's NOT migrated

  • v1's MiniMax-specific configs and scripts (drop entirely).
  • v1's setup.sh flow (v2 has its own).
  • v1's agentcloseout-physics engine (replaced by /verify + dark-patterns evidence hooks).
  • v1's custom hooks (replaced by cmax-session-start.sh / cmax-stop.sh / cmax-post-tool-use.sh + dark-patterns plugin).