Chapter 66: Kaizen Roadmap — R4 Research Tickets (KAIZEN-0017..0044)

In the pmat project, kaizen (改善, “continuous improvement”) refers to a structured, evidence-driven backlog of enhancements derived from external research rather than internal feature planning. Round 4 of this research — tracked in GitHub issue #337 and opened on 2026-04-18 — indexes 28 new tickets (KAIZEN-0017 through KAIZEN-0044). Seventeen tickets originate from newly published arxiv papers in the 2604 series (post-dating the R1/R2/R3 rounds); eleven come from integration gaps exposed by Claude Code 3.7+ changelogs (v2.1.84 through v2.1.113). Each ticket maps a specific research finding or upstream feature to a concrete pmat capability, and this chapter catalogs them so book readers can trace proposed commands and MCP tools back to their scientific or product justification.

Ticket Index (KAIZEN-0017..0044)

Effort labels use the same convention as prior rounds: S ≈ one-sprint ticket, M ≈ multi-sprint with design doc.

Execution Plan

The 28 tickets cluster into six themes. Sequencing below reflects dependency order and the issue-level triage notes: S-effort Claude Code gaps are queued first because they unblock downstream tickets, while larger arxiv-driven capabilities land after the CC integration surface is stable.

1. Claude Code Integration & Hook Ecosystem

• KAIZEN-0034 — PreCompact hook preserving .pmat-work/ state
• KAIZEN-0042 — WorktreeCreate hook triggering pmat context build --worktree
• KAIZEN-0039 — monitors manifest entry for background quality monitor
• KAIZEN-0038 — pmat session-recap writing to .pmat/recap/<date>.md

These are the minimum-viable hook integrations needed so that pmat state survives compaction and worktree creation across a CC session.

2. MCP Conformance & Security

• KAIZEN-0023 — ClawGuard MCP middleware with allowlist schema
• KAIZEN-0024 — MCP-DPT audit across pmat’s 6 defense layers
• KAIZEN-0022 — ClawVM typed-page state abstraction (pmat-state MCP tool)
• KAIZEN-0035 — Strict Bash(pmat *) denyRules template
• KAIZEN-0036 — sandbox.network.deniedDomains recommendations

Security work must land before external-facing cost optimizations (theme 4) because ClawGuard validates adversarial tool returns that the cascading router would otherwise trust.

3. Settings, Skills, and SDK Updates

• KAIZEN-0044 — paths: globs on shipped skills for autoload scoping
• KAIZEN-0040 — Documented skill chains (coverage-gaps → tests → /compact)
• KAIZEN-0041 — pmat settings sync --remote for managed team settings
• KAIZEN-0037 — /effort xhigh advisor mapping pmat tasks to effort levels

4. Cost-Aware Cascading & MOO Delegation

• KAIZEN-0017 — pmat agent-opt UCB-E model selection
• KAIZEN-0018 — pmat agent-split --tactics t1,t2,t3 local-route + compress
• KAIZEN-0019 — pmat skills-optimize --goals pass,cost,latency NSGA-II
• KAIZEN-0026 — pmat ask --cascade haiku→sonnet→opus by confidence

Quantified gains from the source papers: 13-32× cost gap (AgentOpt), 45-79% token savings (Local-Splitter), +131% pass rate (SkillMOO), +26.75% Pareto improvement (CascadeDebate).

5. Memory Fidelity & Context Curation

• KAIZEN-0020 — pmat memory-distill --session <id> TSUBASA-style distillation
• KAIZEN-0021 — pmat context curate --budget 50000 RL entropy reduction
• KAIZEN-0043 — pmat observe agents subagent + MCP-latency tail

6. Trajectory Integrity, Scaffolds & Debugging

• KAIZEN-0027 — pmat agent scaffold --pattern react|plan-exec|tree
• KAIZEN-0030 — pmat trace verify --gates-per-step contract validation
• KAIZEN-0028 — pmat work classify-bug --taxonomy agentic auto-label + Five Whys
• KAIZEN-0031 — pmat skills-mine --from-history reusable-sequence mining
• KAIZEN-0032 — pmat poc-gen --from-bug-report regression-grade PoC
• KAIZEN-0029 — pmat security-scan --agents argus Argus SAST orchestrator
• KAIZEN-0025 — pmat optimize --cooperate rustc,clippy,miri
• KAIZEN-0033 — pmat deps autopatch --gate auto-validate

This theme is the R4 “shift from endpoint accuracy to trajectory integrity” — every tool call is a verifiable step, mirroring pmat’s existing Five Whys + contract-falsification posture.

Cross-References

• Round 1–3 tickets (KAIZEN-0001..0016) were indexed in the previous chapter (see Chapter 55 — Autonomous Continuous Improvement for the kaizen execution engine itself).
• Ticket status updates land on the GitHub issue first, then propagate here on each release of the book.
• For ClawVM (KAIZEN-0022), check whether the existing trueno-graph O(1) context layer already satisfies the typed-page requirement before reimplementing.
• For the agentic bug taxonomy (KAIZEN-0028), reconcile the 5-layer taxonomy against the categories already encoded in pmat five-whys.