What's New in v1.0.0 — The Disciplined Velocity Release

Navigation Help

Problem: After orientating in a project mid-session, it's unclear which phase to enter next. Developers waste time re-reading plans, checking git, and manually triangulating state.

Solution: /meow:help scans the project and tells you exactly what to do next.

/meow:help
# → "You're in Phase 1. Gate 1 is not yet approved.
#    Next step: run /meow:cook tasks/plans/260329-auth.md"

What it scans:

• tasks/plans/ — pending plans awaiting Gate 1
• tasks/reviews/ — pending verdicts awaiting Gate 2
• tests/ — failing tests blocking Phase 3
• git log — unshipped commits blocking Phase 5
• memory/lessons.md — pending retrospective items

The skill maps observed state to one of the 7 pipeline phases and prints a single, unambiguous next action. No hallucination about "what might be next" — only concrete evidence-based routing.

---

Hook-Based Enforcement

Problem: Behavioral rules only work when agents follow them. Rationalization slips through. An agent convinced the .env read is "necessary" can bypass Rule 4 silently.

Solution: Three shell hooks upgrade behavioral rules to enforcement — the action is blocked before it executes.

Hook	Event	What it prevents
privacy-block.sh	PreToolUse	Reads of .env, *.key, credential files
gate-enforcement.sh	PreToolUse	Source code writes before Gate 1 approval
project-context-loader.sh	SessionStart	Missing project-context.md at session start

Without hooks: agent reasons "this .env read is safe" → reads it anyway
With hooks:    hook fires → read blocked → agent must ask human first

privacy-block.sh and gate-enforcement.sh are preventive — they intercept the tool call before execution. project-context-loader.sh is proactive — it injects docs/project-context.md into context before any task runs.

Rules define why. Hooks enforce what.

---

Planning Depth Per Mode

Problem: All workflow modes used the same research approach regardless of task scope. A documentation update ran the same planning overhead as a system design task.

Solution: Each mode now declares a Planning Depth that determines how many researchers run before the plan is written.

Mode	Researchers	Approach
strict	2 parallel	Competing approaches — each researcher argues a different design
architect	2 parallel	Competing approaches — same as strict
default	1	Single researcher, standard depth
audit	1	Single researcher, security-focused
fast	0	Skip research — go straight to plan
cost-saver	0	Skip research — minimize token spend
document	0	Skip research — docs tasks don't need it

strict and architect modes force competing approaches specifically to surface trade-offs that single-researcher planning misses. The synthesis step resolves the competition into a single recommended path.

---

Summary

Feature	Solves	When it activates
Navigation Help	"What do I do next?"	Explicit /meow:help
Hook Enforcement	Rule bypass via agent rationalization	Every session, automatic
Planning Depth	Over/under-researching by mode	Phase 1, per active mode

Structured Debugging

Problem: Agents guess-and-patch bugs instead of investigating root causes. Fix one symptom, introduce another.

Solution: meow:debug enforces a 5-step workflow: reproduce → isolate → root cause → fix → verify. The iron rule: no fix without root cause. Includes a references/common-root-causes.md covering the patterns behind ~80% of bugs (null paths, race conditions, API mismatches, wrong conditions).

Post-Implementation Simplification

Problem: Code that passes tests can still be unnecessarily complex. Dead code, deep nesting, premature abstractions accumulate.

Solution: meow:simplify runs after Phase 3 (Build GREEN) and before Phase 4 (Review). The iron rule: behavior must not change — every simplification must pass the same tests. Removes dead code, flattens nesting, extracts only when 3+ duplicates exist.

Team Configuration

Problem: Setting up parallel agents requires creating worktrees, ownership maps, and task queues manually.

Solution: meow:team-config automates the entire parallel setup: analyzes the plan for independent subtasks, generates an ownership map with glob patterns, creates git worktrees, and initializes the task queue. Validates zero overlap before any agent starts.

---

Previously (shipped in same release)

Improvements

Seven capabilities shipped across two weeks. Each addresses a specific failure mode in AI agent workflows.

Scale-Adaptive Intelligence

Problem: Orchestrators guess task complexity. Guesses are wrong. A fintech payment task classified as STANDARD gets Sonnet instead of Opus — and misses a security edge case.

Solution: Domain-based routing at Phase 0 via meow:scale-routing. A CSV maps keyword signals to complexity levels. The CSV verdict overrides manual classification — no exceptions.

Task: "Add Stripe refund endpoint"
  → scale-routing detects: domain=fintech, level=high
  → Override: COMPLEX tier (Opus)
  → Manual classification: ignored

Key behaviors:

• High domains (fintech, healthcare, auth) force COMPLEX tier
• Low domains (docs, config, changelog) unlock one-shot bypass of Gate 1
• User-extensible — add rows to domain-complexity.csv for project-specific domains
• No rationalization — agents cannot argue down a high-level verdict

Model Routing docs →

---

Project Context System

Problem: Each agent infers project conventions independently. One agent uses snake_case, another uses camelCase. Context drift compounds over a session.

Solution: docs/project-context.md is the agent constitution. Every agent loads it at session start, before any task-specific context.

Session Start
  1. Load docs/project-context.md   ← Always first
  2. Load memory/lessons.md
  3. Load task context
  4. Route to agent

What goes in project-context.md:

• Tech stack and versions
• Naming conventions
• Anti-patterns to avoid
• Testing approach and coverage targets
• Deployment process

Generate or update it:

/meow:docs-init   # new project — scan codebase, generate skeleton
/meow:docs-sync   # after shipping a feature — diff-aware update

---

Multi-Layer Adversarial Review

Problem: Single-pass code review misses bugs. One reviewer looks for what they expect to see.

Solution: meow:review now uses a 4-step workflow with 3 parallel reviewers and a triage step.

Step 1: Blind Hunter     — reviews with zero context (catches hidden bugs)
Step 2: Edge Case Hunter — stress-tests boundary conditions and error paths
Step 3: Criteria Auditor — verifies every acceptance criterion is actually met
Step 4: Triage           — synthesizes findings, eliminates duplicates, assigns severity

Results: catches 2-3x more bugs than single-pass review, particularly:

• Silent failures where code compiles but behaves wrong under load
• Missing error handling in exception paths
• Acceptance criteria that are technically "met" but miss the intent

This uses the step-file architecture — each step is JIT-loaded, keeping context lean.

---

Anti-Rationalization Hardening

Problem: Agents rationalize their way around rules. "This is just a simple task, we don't need security review." "The tests look fine, I won't write more."

Solution: Four explicit blocks with no self-override:

Blocked behavior	Why
Downgrading complexity after scale-routing assigns it	Domain signals are more reliable than in-context reasoning
Minimizing test coverage ("looks fine")	Tests that don't cover real paths are worse than no tests
Skipping security rules for "simple" tasks	Security issues don't respect task complexity
Dismissing WARN verdicts without explicit human acknowledgment	WARNs exist for a reason; silence is not acknowledgment

---

Party Mode

Problem: AI agents presented with an architectural question give one answer. It may be wrong. There's no structured way to stress-test the decision.

Solution: /meow:party spawns 2-4 deliberation agents with different analytical lenses. Each reaches an independent conclusion. A synthesis agent reconciles positions into a single recommendation.

/meow:party "PostgreSQL or MongoDB for the session store?"

Agent A (performance lens): PostgreSQL — ACID transactions, predictable latency
Agent B (ops lens):          MongoDB — horizontal scaling easier, ops simpler
Agent C (migration lens):    PostgreSQL — existing data in PG, no migration needed

Synthesis: PostgreSQL. Migration cost and existing infra outweigh scaling argument.
MongoDB revisit if read load exceeds 50k req/s (not projected for 12 months).

Party output feeds into an ADR. The decision is documented before any code is written.

When to use: Two reasonable engineers would disagree. Long-term consequences. Wrong choice costs > 2 days.

Party Mode skill → | Architecture workflow →

---

Step-File Architecture

Problem: Complex skills with 3+ distinct phases load everything into context at once. Token-expensive. Hard to audit. Cannot resume after context reset.

Solution: Skills decompose into JIT-loaded step files.

skills/meow:review/
├── SKILL.md           # Entry point — metadata only
├── workflow.md        # Step sequence
├── step-01-blind-review.md
├── step-02-edge-cases.md
├── step-03-criteria-audit.md
└── step-04-triage.md

Only the active step is in context. State persists to session-state/{skill}-progress.json — interrupted workflows resume from the last completed step, not from scratch.

Rules:

• One step at a time — never pre-load the next step
• Never skip steps — even short steps run fast and preserve sequence integrity
• Monolithic skills (< 150 lines) don't need decomposition

Currently step-file enabled: meow:review (4 steps).

Step-file architecture →

---

Parallel Execution

Problem: COMPLEX tasks with independent subtasks run sequentially. A backend API, its tests, and a frontend component have no dependency between them — but they run one at a time.

Solution: Up to 3 parallel agents, each isolated in a git worktree.

# Orchestrator decomposes COMPLEX task:
git worktree add .worktrees/api-agent    -b feat/api-agent
git worktree add .worktrees/ui-agent     -b feat/ui-agent
git worktree add .worktrees/test-agent   -b feat/test-agent

# Agents work simultaneously, zero file overlap
# After all complete: merge branches, run full integration tests

Constraints:

Rule	Value
Max parallel agents	3
File overlap allowed	Never
Gates (1 and 2)	Always sequential, always human-approved
Integration test after merge	Required
Eligible task tier	COMPLEX only

Party Mode and Gates are never parallelized — coordination overhead exceeds the benefit.

---

Summary

Feature	Solves	When it activates
Scale-Adaptive Routing	Wrong model tier for domain	Phase 0, automatic
Project Context System	Agent convention drift	Session start, automatic
Adversarial Review	Single-pass review misses bugs	Phase 4 via meow:review
Anti-Rationalization	Agents bypass rules via reasoning	Always, built into rules
Party Mode	No stress-test for architecture decisions	Explicit /meow:party
Step-File Architecture	Token waste + no resumability	Complex skills (meow:review)
Parallel Execution	Sequential bottleneck on independent work	COMPLEX tasks, explicit