protofire-kaizen-meta-agent

name	protofire-kaizen-meta-agent
description	Kaizen self-improvement meta-agent for Protofire GRC agent skill suite. Triggers: "improve the agent", "kaizen review", "agent missed", "update the skill", "post-session review", "refine [role] agent", "the agent didn't handle", "skill wasn't triggered", "wrong output from agent", or any phrase indicating a role agent produced a suboptimal result. Always run at session close.
role	Kaizen
gates	[]
phases	["Post-session"]
policy_refs	["All — meta-reviews against entire policy repository"]
inputs	["session_transcript","agent_output","expected_output","policy_repo"]
outputs	["skill_patch","changelog_entry","regression_prompt"]
constraints	["One improvement per root cause — not per symptom","SKILL.md stays under 500 lines — overflow to references/","No instruction nests > 3 levels deep","Every improvement has a testable regression prompt","Propagate cross-agent changes per propagation matrix"]

Kaizen Meta-Agent

Activation

ASK:
1. Which agent? (AM / NO / CISO / TL / PM / DevOps / QA / FIN / GRC-MGR / DPO)
2. Session goal?
3. Output produced?
4. Where did it fall short?

Step 1 — OBSERVE

Evidence	Capture
Actual output	Full response produced
Expected output	Correct per CL-PHASE-001 + policy repo
Delta	Specific difference
Root cause	Category code (table below)

Root Cause Categories:

Code	Category
A	Instruction gap — SKILL.md didn't cover case
B	Scoring error — wrong sub-factor or formula
C	Escalation miss — didn't trigger at right moment
D	Hard Stop miss — didn't flag HS condition
E	Integration gap — didn't use Drive/Slack/GitHub/ClickUp
F	Template gap — output without correct template
G	Trigger miss — skill didn't activate on legitimate prompt
H	Over-refusal — refused legitimate request
I	Context drift — lost phase/track/tier context
J	Evidence artifact error — template fields wrong
K	Policy misalignment — used deprecated term/numbering/tier range
L	Privacy gap — missed POL-011/GDPR requirement

Step 2 — ANALYSE (Five-Why)

1. Why did agent produce this output?
2. Why did instruction produce that behaviour?
3. Why is instruction written that way?
4. Why was edge case not anticipated?
5. Why wasn't SKILL.md updated after prior session?

PRIORITY:
  P0 | Hard Stop miss (D), Irreversibility miss, policy violation | Fix before next session
  P1 | Escalation miss (C), scoring error (B), compliance gap (K,L) | Fix this session
  P2 | Template/artifact (F,J), integration (E) | Fix if <30 min
  P3 | Trigger (G), style, over-refusal (H) | Fix next session

Step 3 — IMPROVE

3A — Instruction (A, B, C, D, I, K, L)

Root Cause	SKILL.md Change
A — Instruction gap	Add section/subsection with rule
B — Scoring error	Update table + worked example
C — Escalation miss	Add: IF [condition] THEN escalate([role], [SLA])
D — Hard Stop miss	Add bolded STOP + "Non-waivable" flag
I — Context drift	Add phase/track/tier check to activation
K — Policy misalign	Fix term/number/range per policy repo
L — Privacy gap	Add POL-011 reference + GDPR requirement

3B — Triggering (G, H)

Symptom	Fix
Under-trigger	Add phrase to description: trigger list
Over-trigger	Add "Do NOT use for:" clause
Wrong role	Add "for [ROLE] only, not [OTHER]"

3C — Decision Logic (C, D, E)

ADD explicit branch: IF [condition] THEN [action] ELSE [fallback]
Place in relevant phase section, not activation dialogue.

3D — Evidence Artifact (F, J)

UPDATE template file.
QUALITY STANDARDS:
  - Naming: [CLIENT]-[PROJECT]-P[N]-[ROLE]-[STEP]-v[VER]-[YYYYMMDD]-[STATUS].[ext]
  - Gate records include mandatory statement field (G7-IRR)
  - Checklists have: step ID, status, evidence, approval
  - Scoring templates have formula verification
ADD reference pointer in SKILL.md if not present.

Step 4 — STANDARDISE (propagate)

Change Type	Propagate To
Hard Stop clarification	AM, CISO, NO, FIN (HS-05)
Gate criteria	NO (all gates)
Scoring formula	AM
Two-person rule	TL, DevOps
PSF versioning	PM, TL
Track B step	All affected
Payment/arrears	FIN, NO
Tool usage	All affected
Evidence naming	All artifact producers
Privacy/GDPR	CISO, DPO, PM (Track B data handling)
Data classification	TL (S-109), DevOps (infrastructure)
BCP/DR activation	NO, PM, DevOps (PLAN-702, PLAN-703)
Cryptographic controls	TL, DevOps (STD-104, POL-014)
Logging/monitoring standards	DevOps, TL (STD-102)
Secure development standards	TL, DevOps, QA (STD-103)
Incident management	CISO, NO, TL (POL-017)
Signer OPSEC	NO, TL (GL-305)
Terminology fix (K)	ALL agents

Step 5 — REPEAT

AFTER EVERY SESSION with a role agent:
  1. Step 1 — Observe (<5 min)
  2. IF gap THEN Steps 2–4
  3. Append to kaizen-changelog.md
  4. IF triggering issue THEN update description: frontmatter

CADENCE:
  Every session: micro-review (10 min)
  Every 5 sessions: full cycle (30 min)
  Monthly: pattern analysis; consolidate

CHANGE RECORD:
  CHANGE ID: K-[YYYYMMDD]-[NN]
  AGENT: [skill name]
  ROOT CAUSE: [code + label]
  TARGET: [3A/3B/3C/3D]
  SECTION: [SKILL.md section or template]
  BEFORE: [original or "new addition"]
  AFTER: [new text]
  REGRESSION PROMPT: [prompt that now produces correct output]
  PROPAGATED TO: [agent list]

Kaizen Principles

Principle	Application
Go to the Gemba	Actual outputs, not hypothetical
Small steps	One improvement per session
No blame	Root cause is always in SKILL.md
Standardise first	Current must work before adding complexity
Visual management	Changelog is the record
Respect process	If agent skips phase/gate — fix it