ultra-builder-pro-cli 0.5.2 → 0.6.0

package/orchestrator/dispatch-rules.cjs

@@ -13,9 +13,11 @@
 // parallel orchestrator (8B.2) gets a richer decision surface: wave state,
 // deps readiness, and custom rule injection.
 
+const { isSupportedRuntime } = require('../adapters/_shared/runtime-assets.cjs');
+
 const ROUTE_PREFERENCES = Object.freeze({
-  haiku:  ['opencode', 'gemini', 'claude', 'codex'],
-  sonnet: ['claude', 'codex', 'opencode', 'gemini'],
+  haiku:  ['opencode', 'claude', 'codex'],
+  sonnet: ['claude', 'codex', 'opencode'],
   opus:   ['claude', 'codex'],
 });
 
@@ -73,12 +75,21 @@ const DEFAULT_RULES = Object.freeze([
 ]);
 
 function evaluate(ctx, rules = DEFAULT_RULES) {
+  const normalizedCtx = {
+    ...ctx,
+    available_runtimes: Array.isArray(ctx && ctx.available_runtimes)
+      ? [...new Set(ctx.available_runtimes.filter(isSupportedRuntime))]
+      : [],
+  };
   const sorted = [...rules].sort((a, b) => (b.priority || 0) - (a.priority || 0));
   for (const rule of sorted) {
-    if (rule.when(ctx)) {
+    if (rule.when(normalizedCtx)) {
       let runtime = null;
       if (rule.action === 'spawn_agent') {
-        runtime = rule.resolve ? rule.resolve(ctx) : (rule.runtime || null);
+        runtime = rule.resolve ? rule.resolve(normalizedCtx) : (rule.runtime || null);
+        if (!isSupportedRuntime(runtime)) {
+          return { rule_id: rule.id, action: 'block', runtime: null };
+        }
       }
       return { rule_id: rule.id, action: rule.action, runtime };
     }

package/orchestrator/tests/daemon.test.cjs

@@ -15,6 +15,7 @@ const { execFileSync } = require('node:child_process');
 const { initStateDb, closeStateDb } = require('../../mcp-server/lib/state-db.cjs');
 const ops = require('../../mcp-server/lib/state-ops.cjs');
 const daemon = require('../daemon.cjs');
+const RETIRED_RUNTIME = ['gem', 'ini'].join('');
 
 const LONG_SLEEP_CMD = process.execPath;
 const LONG_SLEEP_ARGS = ['-e', 'setInterval(() => {}, 60000);'];
@@ -55,15 +56,15 @@ function cleanup(repoRoot, db, handle) {
 test('routeTask: haiku hint prefers cheap runtimes', () => {
   const runtime = daemon.routeTask(
     { id: 't', complexity_hint: 'haiku' },
-    ['claude', 'opencode', 'codex', 'gemini'],
+    ['claude', 'opencode', 'codex'],
   );
-  assert.ok(['opencode', 'gemini'].includes(runtime), `got ${runtime}`);
+  assert.equal(runtime, 'opencode');
 });
 
 test('routeTask: opus hint prefers strong runtimes', () => {
   const runtime = daemon.routeTask(
     { id: 't', complexity_hint: 'opus' },
-    ['claude', 'opencode', 'codex', 'gemini'],
+    ['claude', 'opencode', 'codex'],
   );
   assert.ok(['claude', 'codex'].includes(runtime), `got ${runtime}`);
 });
@@ -71,23 +72,23 @@ test('routeTask: opus hint prefers strong runtimes', () => {
 test('routeTask: sonnet hint middle-ground', () => {
   const runtime = daemon.routeTask(
     { id: 't', complexity_hint: 'sonnet' },
-    ['claude', 'opencode', 'codex', 'gemini'],
+    ['claude', 'opencode', 'codex'],
   );
-  assert.ok(['claude', 'codex', 'opencode', 'gemini'].includes(runtime));
+  assert.ok(['claude', 'codex', 'opencode'].includes(runtime));
 });
 
 test('routeTask: respects availability constraint', () => {
   const runtime = daemon.routeTask(
     { id: 't', complexity_hint: 'opus' },
-    ['opencode', 'gemini'],
+    ['opencode'],
   );
   // Prefers claude/codex but those aren't available → fall back to any.
-  assert.ok(['opencode', 'gemini'].includes(runtime));
+  assert.equal(runtime, 'opencode');
 });
 
-test('routeTask: no hint → first available', () => {
-  const runtime = daemon.routeTask({ id: 't' }, ['gemini', 'claude']);
-  assert.ok(['gemini', 'claude'].includes(runtime));
+test('routeTask: no hint skips unsupported runtimes', () => {
+  const runtime = daemon.routeTask({ id: 't' }, [RETIRED_RUNTIME, 'claude']);
+  assert.equal(runtime, 'claude');
 });
 
 test('routeTask: no runtimes available → null', () => {
@@ -244,7 +245,7 @@ test('runDaemon applies complexity_hint route', async () => {
     });
     handle = daemon.runDaemon({
       db, repoRoot,
-      runtimes: ['claude', 'opencode', 'gemini'],
+      runtimes: ['claude', 'opencode'],
       pollMs: 50,
       command: LONG_SLEEP_CMD,
       commandArgs: LONG_SLEEP_ARGS,
@@ -252,7 +253,7 @@ test('runDaemon applies complexity_hint route', async () => {
     await new Promise((r) => setTimeout(r, 300));
     const sessions = db.prepare("SELECT * FROM sessions WHERE task_id = 'd-opus'").all();
     assert.equal(sessions.length, 1);
-    // opus should prefer claude over opencode/gemini.
+    // opus should prefer claude over opencode.
     assert.equal(sessions[0].runtime, 'claude');
   } finally {
     cleanup(repoRoot, db, handle);

package/orchestrator/tests/dispatch-rules.test.cjs

@@ -11,12 +11,13 @@ const { test } = require('node:test');
 const assert = require('node:assert/strict');
 
 const { evaluate, DEFAULT_RULES, ROUTE_PREFERENCES } = require('../dispatch-rules.cjs');
+const RETIRED_RUNTIME = ['gem', 'ini'].join('');
 
 function baseCtx(overrides = {}) {
   return {
     task: { id: 't1' },
     deps_ready: true,
-    available_runtimes: ['claude', 'opencode', 'codex', 'gemini'],
+    available_runtimes: ['claude', 'opencode', 'codex'],
     breaker_state: 'ok',
     wave: null,
     ...overrides,
@@ -42,7 +43,7 @@ test('evaluate: opus hint + only codex available → codex', () => {
 test('evaluate: opus hint + preference exhausted → first available fallback', () => {
   const d = evaluate(baseCtx({
     task: { id: 't', complexity_hint: 'opus' },
-    available_runtimes: ['opencode', 'gemini'],
+    available_runtimes: ['opencode'],
   }));
   assert.equal(d.action, 'spawn_agent');
   assert.equal(d.runtime, 'opencode');
@@ -58,13 +59,13 @@ test('evaluate: sonnet hint + all available → claude', () => {
   assert.equal(d.runtime, 'claude');
 });
 
-test('evaluate: no hint → first available fallback', () => {
+test('evaluate: retired runtimes are discarded before fallback routing', () => {
   const d = evaluate(baseCtx({
     task: { id: 't' },
-    available_runtimes: ['gemini', 'claude'],
+    available_runtimes: [RETIRED_RUNTIME, 'claude'],
   }));
   assert.equal(d.action, 'spawn_agent');
-  assert.equal(d.runtime, 'gemini');
+  assert.equal(d.runtime, 'claude');
 });
 
 // ─── block / defer paths (new in 8B.1) ────────────────────────────────────
@@ -97,20 +98,21 @@ test('evaluate: serial wave already has running task → defer', () => {
 
 // ─── composability: custom rules override defaults ────────────────────────
 
-test('evaluate: custom high-priority rule overrides defaults', () => {
+test('evaluate: custom rules cannot select an unsupported runtime', () => {
   const customRules = [
     ...DEFAULT_RULES,
     {
-      id: 'force-gemini',
+      id: 'force-retired-runtime',
       priority: 999,
       when: () => true,
       action: 'spawn_agent',
-      resolve: () => 'gemini',
+      resolve: () => RETIRED_RUNTIME,
     },
   ];
   const d = evaluate(baseCtx({ task: { id: 't', complexity_hint: 'opus' } }), customRules);
-  assert.equal(d.runtime, 'gemini');
-  assert.equal(d.rule_id, 'force-gemini');
+  assert.equal(d.action, 'block');
+  assert.equal(d.runtime, null);
+  assert.equal(d.rule_id, 'force-retired-runtime');
 });
 
 // ─── ROUTE_PREFERENCES shape guard (regression: don't drift from Phase 5.4) ─

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "ultra-builder-pro-cli",
-  "version": "0.5.2",
-  "description": "Multi-runtime installer for the Ultra Builder Pro agent engineering system — deploys commands, agents, skills, hooks, and MCP server to Claude Code, OpenCode, Codex CLI, and Gemini CLI.",
+  "version": "0.6.0",
+  "description": "Multi-runtime installer for the Ultra Builder Pro agent engineering system — deploys commands, agents, skills, hooks, and MCP server to Claude Code, OpenCode, and Codex CLI.",
   "bin": {
     "ultra-builder-pro-cli": "bin/install.js",
     "ubp": "bin/install.js",
@@ -33,7 +33,6 @@
     "skills/testing-rules",
     "skills/cc-collab",
     "skills/codex-collab",
-    "skills/gemini-collab",
     "skills/ultra-verify",
     "hooks/*.py",
     "hooks/adapters/*.py",
@@ -60,7 +59,6 @@
     "claude-code",
     "opencode",
     "codex",
-    "gemini",
     "agent",
     "cli",
     "multi-runtime",
@@ -87,7 +85,7 @@
     "test:state": "node --test 'mcp-server/**/*.test.cjs'",
     "test:orch": "node --test 'orchestrator/**/*.test.cjs'",
     "test:spec": "node spec/scripts/test-all.cjs",
-    "test:rest": "node --test adapters/_shared/tests/*.test.cjs adapters/tests/*.test.cjs ultra-tools/*.test.cjs tests/conformance/**/*.test.cjs tests/install.test.cjs tests/package-smoke.test.cjs",
+    "test:rest": "node --test adapters/_shared/tests/*.test.cjs adapters/tests/*.test.cjs ultra-tools/*.test.cjs tests/conformance/**/*.test.cjs tests/install.test.cjs tests/package-smoke.test.cjs tests/retired-runtime.test.cjs",
     "test:all": "npm run test:state && npm run test:orch && npm run test:spec && npm run test:rest",
     "test:hooks": "pytest hooks/tests -q",
     "verify:release": "npm run test:all && npm run test:hooks && npm audit --omit=dev --audit-level=high",

package/skills/ultra-init/SKILL.md

@@ -133,7 +133,6 @@ CLI 最后一行是 `{ "ok": true, "data": { ... } }`（见 spec/cli-protocol.md
 | Claude  | `/ultra-init [name] [type] [stack] [git]` — 命令薄壳拉起此 skill |
 | OpenCode | `/ultra-init …` 同上（agent frontmatter 小写化） |
 | Codex   | `codex exec "run /ultra-init …"`（Phase 4.4 adapter 接） |
-| Gemini  | `gemini --prompt "run /ultra-init …"`（Phase 4.5 extension 接） |
 
 ## 输出锚点
 

package/skills/ultra-verify/SKILL.md

@@ -1,13 +1,13 @@
 ---
 name: ultra-verify
-description: Run three-way verification with the current host as primary and read-only Gemini and Codex advisors, then verify and synthesize the evidence.
+description: Run independent two-model verification with the current host as primary and a read-only Codex advisor, then verify and synthesize the evidence.
 ---
 
-# Ultra Verify — Three-Way Verification
+# Ultra Verify — Independent Two-Model Verification
 
 Claude Code remains primary. It writes the first independent analysis, verifies consequential
-claims, and owns the final answer. Gemini and Codex are untrusted read-only advisors. Use this
-workflow only when the user explicitly asks for cross-model or three-way verification.
+claims, and owns the final answer. Codex is an untrusted read-only advisor. Use this workflow only
+when the user explicitly asks for cross-model verification.
 
 ## Modes
 
@@ -18,11 +18,11 @@ workflow only when the user explicitly asks for cross-model or three-way verific
 
 ## Preconditions
 
-1. Confirm `gemini --version` and `codex --version` plus authentication.
+1. Confirm `codex --version` plus authentication.
 2. Define the exact workspace, question, evidence standard, and response shape.
 3. Do not send secrets, unrelated files, or an unbounded home directory.
-4. Track the four workflow steps in the current host plan: primary analysis, advisor launch,
-   completion wait, and verified synthesis.
+4. Track four steps in the current host plan: primary analysis, advisor launch, completion wait,
+   and verified synthesis.
 
 ## 1. Create the session and write the primary view
 
@@ -33,25 +33,11 @@ mkdir -p "${SESSION_PATH}"
 ```
 
 Write Claude Code's evidence-backed analysis to `${SESSION_PATH}/claude-analysis.md` before
-reading either advisor. Record the mode, checkout, scope, and evidence boundary.
+reading the advisor output. Record the mode, checkout, scope, and evidence boundary.
 
-## 2. Launch both advisors read-only
+## 2. Launch the advisor read-only
 
-Give both advisors the same bounded raw question and evidence, without the primary conclusion.
-Launch them concurrently when the host supports independent shell sessions; otherwise start both
-before collecting either result.
-
-Gemini:
-
-```bash
-gemini --approval-mode plan \
-  --output-format text \
-  -p "<bounded prompt>" \
-  > "${SESSION_PATH}/gemini-output.md" \
-  2> "${SESSION_PATH}/gemini-error.log"
-```
-
-Codex:
+Give the advisor the bounded raw question and evidence without the primary conclusion.
 
 ```bash
 codex exec -s read-only \
@@ -60,10 +46,9 @@ codex exec -s read-only \
   2> "${SESSION_PATH}/codex-error.log"
 ```
 
-Never enable auto-edit, write-capable automation, danger-full-access, or permission bypass for
-either advisor.
+Never enable write-capable automation, danger-full-access, or permission bypass for the advisor.
 
-## 3. Wait for completed outputs
+## 3. Wait for completed output
 
 Run the bundled waiter in an asynchronous or yielded shell session and poll it at bounded
 intervals; never hold one blocking tool call longer than 60 seconds.
@@ -73,28 +58,26 @@ python3 "$CLAUDE_PLUGIN_ROOT/skills/ultra-verify/scripts/verify_wait.py" \
   "${SESSION_PATH}" --timeout 1200
 ```
 
-The waiter reports each advisor as `complete`, `failed`, `empty`, or `pending`. Read an advisor
-output only after the corresponding status is `complete` and the file is non-empty. Preserve its
-error log when it fails.
+The waiter reports the advisor as `complete`, `failed`, `empty`, or `pending`. Read the output only
+after its status is `complete` and the file is non-empty. Preserve the error log when it fails.
 
 ## 4. Verify and synthesize
 
-Compare `claude-analysis.md` with each completed advisor output:
+Compare `claude-analysis.md` with `codex-output.md`:
 
 1. Verify claims against the current checkout, tests, runtime, or primary documentation.
-2. Separate consensus, majority views, useful dissent, and unsupported assertions.
-3. Explain scope, version, and assumption differences before scoring agreement.
+2. Separate verified agreement, useful dissent, and unsupported assertions.
+3. Explain scope, version, and assumption differences before judging the result.
 4. Write `synthesis.md` and `metadata.json`, then return one Claude Code-owned conclusion.
 
-Use `references/confidence-system.md` for scoring and
-`references/cross-verify-modes.md` for mode-specific evidence expectations. Model agreement never
-overrides failing tests or authoritative runtime evidence.
+Use `references/confidence-system.md` for evidence-based confidence and
+`references/cross-verify-modes.md` for mode-specific expectations. Model agreement never overrides
+failing tests or authoritative runtime evidence.
 
 ## Degraded operation
 
-- One advisor fails: continue with the primary view plus the available advisor and name the gap.
-- Both advisors fail: return primary-only analysis with an explicit single-source warning.
-- Never block the underlying user task solely because an advisor is absent, unauthenticated, or
+- Advisor failure: return the primary analysis with an explicit single-source warning.
+- Never block the underlying user task solely because the advisor is absent, unauthenticated, or
   slow.
 
 ## Session files
@@ -102,8 +85,6 @@ overrides failing tests or authoritative runtime evidence.
 ```text
 .ultra/collab/<SESSION_ID>/
   claude-analysis.md
-  gemini-output.md
-  gemini-error.log
   codex-output.md
   codex-error.log
   metadata.json

package/skills/ultra-verify/evals/evals.json

@@ -4,47 +4,45 @@
     {
       "id": 1,
       "prompt": "/ultra-verify decision 'Should we use PostgreSQL or MongoDB for our new microservice?'",
-      "expected_output": "Claude answers first, then launches Gemini and Codex in parallel, collects all three analyses, and presents a synthesis with consensus scoring",
+      "expected_output": "Claude writes its analysis first, launches Codex read-only, verifies both analyses, and returns one evidence-backed synthesis",
       "assertions": [
-        "Claude writes its own analysis BEFORE reading external AI output",
-        "Launches Gemini and Codex in parallel (run_in_background)",
-        "Creates session directory under .ultra/collab/ with verify-decision in name",
-        "Reads all three output files with Read tool",
-        "Presents confidence level (consensus/majority/no consensus)",
-        "Includes per-AI recommendation summary and trade-off matrix"
+        "Claude writes its own analysis before reading advisor output",
+        "Launches Codex with a read-only sandbox",
+        "Creates a session directory under .ultra/collab/",
+        "Verifies consequential claims against authoritative evidence",
+        "Separates verified agreement, useful dissent, and unsupported assertions"
       ]
     },
     {
       "id": 2,
-      "prompt": "/ultra-verify diagnose 'API returns 500 intermittently under load, logs show connection pool exhaustion'",
-      "expected_output": "Three sets of hypotheses collected and ranked by consensus count",
+      "prompt": "/ultra-verify diagnose 'API returns 500 intermittently under load'",
+      "expected_output": "Independent hypotheses are reconciled through discriminating checks and evidence strength",
       "assertions": [
-        "Each AI provides top-3 hypotheses",
-        "Hypotheses ranked by consensus count (3/3 > 2/3 > 1/3)",
-        "Each hypothesis includes evidence and verification steps",
-        "Recommended investigation order based on consensus"
+        "Both analyses provide ranked hypotheses",
+        "Duplicate hypotheses are merged by root cause",
+        "Each retained hypothesis includes a verification step",
+        "Investigation order follows evidence strength and verification cost"
       ]
     },
     {
       "id": 3,
       "prompt": "/ultra-verify audit",
-      "expected_output": "Three-way code audit with findings graded by consensus (3=critical, 2=high, 1=investigate)",
+      "expected_output": "Findings are deduplicated and verified against the current checkout without vote-based severity",
       "assertions": [
-        "Sends same audit prompt to all three AIs",
-        "Findings graded by how many AIs found them",
-        "3/3 findings marked Critical, 2/3 marked High, 1/3 marked Investigate",
-        "Action items prioritized by consensus count"
+        "Uses the same scope and severity model for both analyses",
+        "Verifies every reported location",
+        "Retains uniquely reported findings when source evidence confirms them",
+        "Severity does not depend on model agreement"
       ]
     },
     {
       "id": 4,
-      "prompt": "/ultra-verify estimate 'Implement OAuth2 login with Google and GitHub providers'",
-      "expected_output": "Three effort estimates compared for convergence",
+      "prompt": "/ultra-verify estimate 'Implement OAuth2 login with two providers'",
+      "expected_output": "Two estimates are reconciled by scope, assumptions, dependencies, and risk",
       "assertions": [
-        "Each AI provides estimate with breakdown and assumptions",
-        "Convergence analysis (within 20% = high confidence)",
-        "Outlier estimates investigated",
-        "Final recommended estimate with confidence level"
+        "Both analyses include work breakdowns and assumptions",
+        "Large differences are explained rather than averaged mechanically",
+        "Final range is supported by the reconciled work breakdown"
       ]
     }
   ]

package/skills/ultra-verify/references/confidence-system.md

@@ -1,38 +1,21 @@
-# Confidence System
+# Evidence-Based Confidence
 
-Confidence scoring based on AI consensus for cross-verify results.
+Cross-model agreement is a signal, not an authority. Assign confidence only after checking claims
+against source code, tests, runtime output, or primary documentation.
 
 ## Levels
 
-### Consensus (3/3 Agree)
-- **Confidence**: Highest
-- **Action**: Strongly recommended — proceed with high confidence
-- **Display**: `[CONSENSUS 3/3]`
+- **Verified agreement**: both analyses agree and the consequential claims have authoritative
+  supporting evidence.
+- **Resolved dissent**: the analyses differ, but current evidence supports one position and explains
+  the disagreement.
+- **Unresolved**: available evidence cannot distinguish the positions or key assumptions remain
+  untested.
+- **Single source**: the advisor failed or returned no usable output; no cross-model corroboration is
+  available.
 
-### Majority (2/3 Agree)
-- **Confidence**: High
-- **Action**: Recommended — investigate the dissenting view for edge cases or alternative perspectives
-- **Display**: `[MAJORITY 2/3]` + note which AI dissented and why
+## Reporting
 
-### No Consensus (All Differ)
-- **Confidence**: Low
-- **Action**: Decompose the problem into smaller questions, gather more data, or accept that the answer is genuinely ambiguous
-- **Display**: `[NO CONSENSUS]` + summary of each position
-
-## Application by Mode
-
-| Mode | Consensus | Majority | No Consensus |
-|------|-----------|----------|--------------|
-| decision | Strong recommendation | Recommended + dissent analysis | Present all options, suggest decomposition |
-| diagnose | Investigate first | Investigate second | Check all, may need more data |
-| audit | Critical — fix now | High — likely real | Investigate — may be false positive |
-| estimate | Use average, high confidence | Investigate outlier | Decompose task further |
-
-## Degraded Confidence
-
-When only 2 AIs respond (one failed):
-
-- **2/2 Agree**: Equivalent to Majority (not Consensus — missing perspective)
-- **2/2 Differ**: Low confidence — note missing third opinion
-
-When only 1 AI responds: Claude-only analysis — explicitly mark as single-source, no consensus scoring.
+Record which claims were verified, the evidence used, useful dissent, unsupported assertions, and
+remaining uncertainty. Model agreement never upgrades a claim contradicted by tests or runtime
+evidence, and a uniquely reported finding remains valid when independently verified.

package/skills/ultra-verify/references/cross-verify-modes.md

@@ -1,73 +1,28 @@
 # Ultra Verify Modes
 
-Four modes for three-way AI verification. All modes follow the safe orchestration contract in the parent `SKILL.md`.
+All four modes compare an independently written primary analysis with one read-only external
+advisor. Agreement is useful corroboration, not proof; evidence decides the final result.
 
 ## 1. Decision (`decision`)
 
-For architecture decisions, technology choices, or design trade-offs.
-
-**Flow:**
-1. Present the decision context, constraints, and options to all three AIs independently
-2. Each AI provides: recommendation, pros/cons, risks, trade-offs
-3. Synthesis ranks options by consensus:
-   - 3/3 recommend same → **Strong recommendation**
-   - 2/3 agree → **Recommended** with dissenting view analyzed
-   - All different → Present all three with trade-off matrix
-
-**Output structure:**
-- Per-AI recommendation summary
-- Consensus matrix (which AI recommends what)
-- Trade-off analysis
-- Final recommendation with confidence level
+Give both analyses the same decision context, constraints, and options. Compare recommendation,
+trade-offs, risks, and assumptions. When they differ, identify the assumption or evidence that
+caused the split and resolve it before recommending an option.
 
 ## 2. Diagnose (`diagnose`)
 
-For bug diagnosis, root cause analysis, or troubleshooting.
-
-**Flow:**
-1. Provide symptoms, error messages, relevant code, and recent changes to all three AIs
-2. Each AI provides its top-3 hypotheses with evidence and verification steps
-3. Synthesis:
-   - Hypotheses appearing in 3/3 lists → **Most likely** (investigate first)
-   - Hypotheses in 2/3 lists → **Probable** (investigate second)
-   - Unique hypotheses → **Worth checking** (may catch edge cases)
-
-**Output structure:**
-- Ranked hypothesis list (by consensus count)
-- Per-hypothesis: description, evidence, verification steps, suggested fix
-- Recommended investigation order
+Provide symptoms, exact errors, relevant code, and recent changes. Each analysis should rank its
+root-cause hypotheses and provide discriminating checks. Merge duplicate hypotheses, retain useful
+unique ones, and order investigation by evidence strength and verification cost.
 
 ## 3. Audit (`audit`)
 
-For code review, security audit, or quality assessment.
-
-**Flow:**
-1. Provide the code scope to all three AIs with the same audit prompt
-2. Each AI reports findings with severity and location
-3. Synthesis grades findings by consensus:
-   - Found by 3/3 → **Critical** — fix immediately
-   - Found by 2/3 → **High** — likely real, investigate
-   - Found by 1/3 → **Investigate** — may be false positive or edge case
-
-**Output structure:**
-- Findings table: description, severity, consensus count, which AIs found it
-- Grouped by consensus level (Critical → High → Investigate)
-- Action items prioritized by consensus
+Use the same code scope and severity model for both analyses. Deduplicate findings by root cause,
+verify every location against the current checkout, and retain a single-model finding when source
+evidence confirms it. Never derive severity from model agreement alone.
 
 ## 4. Estimate (`estimate`)
 
-For effort estimation, complexity assessment, or timeline planning.
-
-**Flow:**
-1. Describe the task, requirements, and constraints to all three AIs
-2. Each AI provides: estimate (time/effort), breakdown, assumptions, risks
-3. Synthesis evaluates convergence:
-   - All within 20% of each other → **High confidence** estimate (use average)
-   - One outlier → **Moderate confidence** — investigate the outlier's reasoning
-   - All significantly different → **Low confidence** — task needs decomposition
-
-**Output structure:**
-- Per-AI estimate with breakdown and assumptions
-- Convergence analysis
-- Recommended estimate with confidence level
-- Risk factors that could shift the estimate
+Require a breakdown, assumptions, dependencies, and risk range from both analyses. Explain large
+differences by locating mismatched scope or assumptions. Return a range supported by the reconciled
+work breakdown rather than averaging estimates mechanically.