CAF — Current-State Teardown

Scope. A diagnostic snapshot of the Catalyst Assessment Framework as it stands today. Teardown only — no restructure recommendations in this document. Based on a close read of UNDERSTANDING.md, the root CONTRACTS.md, and a representative leaf skill (1-1-input-shape-classifier); details that need the live repo to confirm are flagged [verify-against-repo].

What CAF is

A library of 106 agent "skills" (markdown prompt-specs) that, run in sequence, walk a consultant through assessing a client engagement and emit a build manifest: a plan of what agents/skills should be built for that client. CAF plans; it does not build. Every skill declares allowed-tools: Read Grep; none carry Skill, Agent, or Task tools, so no skill can invoke another skill, spawn a sub-agent, or execute anything against a real system.

Architecture: 7 phases + Meta-Council

Phase structure (v4 numbering):

1. Engagement Framing
2. Pre-Engagement Discovery & Readiness
3. Discovery / Classification / Archetype Tagging
4. Prioritization / Governance / Platform-Native Inventory
5. Skill Contracts / Composition / Validation Archetype
6. Execution & Iteration
7. Scale / Replicate / Productize

Above the phases sits a Meta-Council with cross-phase VETO/HALT authority: m-1 portfolio strategist (HALT authority), m-2 prioritization integrity, m-3 expectation steward. Some phases also carry per-phase meta-personas (2-m, 3-m).

Phase 4 additionally contains lettered sub-paths (4A, 4B, 4C, 4D, 4G, S4) whose naming scheme is not explained by the root docs — the selection logic between paths appears to live in operator judgment rather than in any router skill [verify-against-repo].

The handoff contracts (Contract 0–6)

Seven contracts govern the phase boundaries. They are structural-semantic rather than rigid schemas: each specifies the categories of information that must flow, carries confidence/provenance metadata and minimum completeness thresholds, and requires archetype tags, platform target, and domain rules to propagate unchanged downstream. Contracts exist only as YAML blobs inside the conversation; nothing is written to disk.

Four structural gaps

1. No marketplace plugin. Install is a bespoke bin/install.sh copying skills/ into ~/.claude/skills/ plus 6 other targets, with versioning via a custom .caf-skills.lock. The root SKILL.md is a catalog document, not an invokable command; there is no master /catalyst entry point.
2. No persistence / not resumable. Because contracts flow as in-conversation YAML only — no manifest file, no engagements/ directory, no run ledger — a fresh session cannot resume an engagement unless the operator pastes prior YAML back in by hand.
3. No live orchestrator. "Orchestration" today means the operator runs 106 skills by hand in phase order. The two synthesizer skills (2-10-orchestrator, 3-10-readiness-synthesizer) aggregate a phase's outputs into the handoff contract but invoke nothing.
4. Skills are inert against real systems. Pure Read/Grep plus YAML emission: the skills describe deterministic checks (SQL re-execution, lineage verification) but cannot perform them. The co-installed ds-* / agent-* / sql-* CLI toolkit is not wired into CAF reasoning.

Migration debt (v3→v4)

A codemod renumbered phases (+2) and flattened nested phaseX/ directories into a flat skills/. The migration is half-finished:

• Frontmatter is correct, bodies are not. Skill bodies still contain old phase numbers (e.g. 1-1-input-shape-classifier says "Phase 3 (3.1)" where it should read Phase 1).
• At least one corrupted threshold. A naive numeric pass turned 0.7 → 2.7 on a 0–1 confidence scale, so a Hard Constraint now reads "confidence below 2.7" — always true; that gate is effectively broken.
• Stale root docs. SKILL.md, README.md, and CONTRACTS.md are still v3 and contradict the live v4 layout (skills/, CATALOG.md, docs/PHASE_TAXONOMY.md). CONTRACTS.md is only half-renumbered: headers say Contract 0 = Phase 1→2 while frontmatter and body say Phase 6→7.

Per-skill anatomy (why the count is 106)

Each leaf skill is a multi-hundred-line specification for a single step: Overview, Key Questions, a Hard Constraint with a PROHIBITED-from list, Produces, a verbatim Role Prompt, Capabilities, a weighted Evaluation Criteria table, Instructions, and YAML I/O. This is heavy prescription of how rather than what — a prompting style tuned for older, less capable models. It is the direct cause of the 106-skill count: every judgment is decomposed into its own hand-walked micro-step.

Summary

CAF today is a thorough, well-thought-out planning rubric executed entirely by operator labor: rich phase taxonomy, real contract discipline, genuine HITL gates — but no install story beyond a copy script, no on-disk state, no orchestration, no execution capability, and unresolved v3→v4 migration debt that includes one silently broken quality gate. The intellectual content is strong; the delivery vehicle predates current agent-platform capabilities.

A companion restructure proposal (separate document) exists; this teardown intentionally stops at diagnosis.