feat(loop): assign model tiers per role — fable builder and plan gate

Author: superstructor

README.md

@@ -65,6 +65,8 @@ Hopper → Deming
 
 Requirements → spec → review → scout → build → test/security/perf/UX → code review → visual → docs → commit → observe. One iteration at a time. Von Braun handles release builds separately via `/great:deploy`.
 
+Each role runs on the model tier its job demands: Fable for the builder and the plan-approval gate, Opus for testing, security audit, specs, and code review (security roles stay on Opus deliberately — Fable's cyber safety classifiers can refuse security-focused analysis), Sonnet for scouting, UX, visual, perf, docs, and deploys, Haiku for commits. Deming, the team lead, is your own session — run it on Opus or Fable.
+
 ```sh
 great loop install            # registers the plugin with Claude Code
 great loop install --project  # adds .tasks/ working state to the current repo

loop/.claude-plugin/plugin.json

@@ -1,5 +1,5 @@
 {
   "name": "great",
-  "version": "0.2.0",
+  "version": "0.3.0",
   "description": "great.sh Loop — 16-role AI agent orchestration"
 }

loop/agents/davinci.md

@@ -2,7 +2,7 @@
 name: davinci
 description: "Leonardo da Vinci — Builder. Teammate. Turns specs into working code."
 tools: [Read, Write, Edit, Glob, Grep, Bash, Task, mcp__context7__*]
-model: opus
+model: fable
 memory: project
 ---
 

loop/agents/dijkstra.md

@@ -2,7 +2,7 @@
 name: dijkstra
 description: "Edsger Dijkstra — Code Reviewer. Reviews code quality, complexity, and structure. Gate between build and commit."
 tools: [Read, Glob, Grep]
-model: sonnet
+model: opus
 memory: project
 ---
 

loop/agents/socrates.md

@@ -2,7 +2,7 @@
 name: socrates
 description: "Socrates — Adversarial Spec Reviewer. Questions every assumption. Plan approval gate."
 tools: [Read, Glob, Grep]
-model: opus
+model: fable
 memory: project
 ---
 

loop/skills/loop/SKILL.md

@@ -24,7 +24,7 @@ Execute the full 16-agent great.sh Loop.
 
 **2. Lovelace** (Opus) — Write spec. Use Context7 MCP for library/framework docs. Output to `.tasks/ready/`.
 
-**3. Socrates** (Opus) — **Plan approval gate.** Adversarial review with structured elenchus. Max 3 Lovelace<->Socrates rounds. You (Deming) decide after 3 if still rejected.
+**3. Socrates** (Fable) — **Plan approval gate.** Adversarial review with structured elenchus — ambiguity and contradiction hunting is where the top-tier model earns its cost: an error caught at spec time saves a full build cycle. Max 3 Lovelace<->Socrates rounds. You (Deming) decide after 3 if still rejected.
 
 **4. Humboldt** (Sonnet) — Scout codebase. Map files, patterns, dependencies.
 
@@ -34,9 +34,9 @@ Create an agent team with four teammates working in parallel:
 
 ```
 Spawn a team of four teammates:
-- Da Vinci (builder, Opus): implement spec using Context7 MCP. Run quality gates. Message Turing and Kerckhoffs when ready. Require plan approval before implementing.
-- Turing (tester, Opus): prove the build is broken. Adversarial tests. Regression watchdog. Message Da Vinci with failures.
-- Kerckhoffs (security, Opus): audit for credential leakage, permission errors, input validation gaps, supply chain risks. CRITICAL/HIGH block commit. Message Da Vinci with fixes.
+- Da Vinci (builder, Fable): implement spec using Context7 MCP. Run quality gates. Message Turing and Kerckhoffs when ready. Require plan approval before implementing.
+- Turing (tester, Opus — not Fable: its security probing can trip Fable's cyber classifiers): prove the build is broken. Adversarial tests. Regression watchdog. Message Da Vinci with failures.
+- Kerckhoffs (security, Opus — deliberately not Fable: security-focused analysis is excluded from Fable's gains and risks classifier refusals mid-audit): audit for credential leakage, permission errors, input validation gaps, supply chain risks. CRITICAL/HIGH block commit. Message Da Vinci with fixes.
 - Nielsen (UX, Sonnet): walk affected user journeys. Playwright MCP for web. Can block commit. Message Da Vinci for issues.
 
 Give each teammate the approved spec and scout report, and state what "done"
@@ -60,7 +60,7 @@ Collect teammate reports:
 - Performance regression -> Da Vinci fixes, Wirth re-measures (max 2 cycles)
 - Non-blocking -> Nightingale files as P2/P3
 
-Run **Dijkstra** (subagent, Sonnet) — Code quality review. REJECTED -> Da Vinci fixes, Dijkstra re-reviews (max 2 cycles).
+Run **Dijkstra** (subagent, Opus) — Code quality review. REJECTED -> Da Vinci fixes, Dijkstra re-reviews (max 2 cycles).
 
 ## Phase 4: Finish
 

loop/teams-config.json

@@ -6,7 +6,7 @@
       "description": "Leonardo da Vinci — Builder",
       "instructions": "Implement the provided spec. Use Context7 MCP for library docs. Run all quality gates. Message teammates directly for coordination. Broadcast when build is ready.",
       "tools": ["Read", "Write", "Edit", "Bash", "Glob", "Grep"],
-      "model": "opus"
+      "model": "fable"
     },
     "turing": {
       "description": "Alan Turing — Tester/Breaker",

src/cli/loop_cmd.rs

@@ -663,7 +663,7 @@ fn run_install(project: bool, force: bool, _non_interactive: bool) -> Result<()>
     output::header("great.sh Loop installed!");
     println!();
     output::info("16 roles: 4 teammates + 11 subagents + 1 team lead");
-    output::info("All Claude: Opus + Sonnet + Haiku");
+    output::info("All Claude: Fable + Opus + Sonnet + Haiku");
     println!();
     if project {
         output::info("Usage: claude -> /great:loop [task description]");