kairn-cli 2.7.2 → 2.10.0

package/dist/cli.js

@@ -218,6 +218,7 @@ async function callLLM(config, userMessage, options) {
   const maxTokens = options.maxTokens ?? 8192;
   const { systemPrompt } = options;
   const jsonMode = options.jsonMode ?? false;
+  const cacheControl = options.cacheControl ?? false;
   const providerName = getProviderName(config.provider);
   let apiKey = config.api_key;
   if (config.auth_type === "claude-code-oauth") {
@@ -238,7 +239,7 @@ async function callLLM(config, userMessage, options) {
       const response = await client2.messages.create({
         model: config.model,
         max_tokens: maxTokens,
-        system: systemPrompt,
+        system: cacheControl ? [{ type: "text", text: systemPrompt, cache_control: { type: "ephemeral" } }] : systemPrompt,
         messages
       });
       const textBlock = response.content.find((block) => block.type === "text");
@@ -461,6 +462,96 @@ var init_exec = __esm({
 });
 
 // src/evolve/scorers.ts
+function scoreCriterionDeterministic(criterionText, stdout, stderr) {
+  const combined = `${stdout}
+${stderr}`.toLowerCase();
+  const criterionLower = criterionText.toLowerCase().trim();
+  if (/^ran\b/i.test(criterionText.trim())) {
+    for (const entry of RAN_COMMAND_EVIDENCE) {
+      const matchesKeyword = entry.keywords.some(
+        (kw) => criterionLower.includes(kw.toLowerCase())
+      );
+      if (matchesKeyword) {
+        const found = entry.evidence.some((ev) => combined.includes(ev.toLowerCase()));
+        if (found) {
+          const matchedEvidence = entry.evidence.find(
+            (ev) => combined.includes(ev.toLowerCase())
+          );
+          return {
+            score: 1,
+            reasoning: `Deterministic: found evidence of '${matchedEvidence}' in output`
+          };
+        }
+        return {
+          score: 0,
+          reasoning: `Deterministic: no evidence of '${entry.keywords[0]}' found`
+        };
+      }
+    }
+    return null;
+  }
+  if (/^(zero|no)\b/i.test(criterionText.trim())) {
+    for (const entry of ABSENCE_PATTERNS) {
+      const matchesKeyword = entry.keywords.some(
+        (kw) => criterionLower.includes(kw.toLowerCase())
+      );
+      if (matchesKeyword) {
+        const found = entry.search.some((pat) => combined.includes(pat.toLowerCase()));
+        if (found) {
+          const matchedPattern = entry.search.find(
+            (pat) => combined.includes(pat.toLowerCase())
+          );
+          return {
+            score: 0,
+            reasoning: `Deterministic: found '${matchedPattern}' which should be absent`
+          };
+        }
+        return {
+          score: 1,
+          reasoning: `Deterministic: no prohibited pattern found in output`
+        };
+      }
+    }
+    return null;
+  }
+  if (/^uses?\b/i.test(criterionText.trim())) {
+    for (const entry of PRESENCE_PATTERNS) {
+      if (criterionLower.includes(entry.keyword.toLowerCase())) {
+        const found = entry.search.some((s) => combined.includes(s.toLowerCase()));
+        if (found) {
+          return {
+            score: 1,
+            reasoning: `Deterministic: found '${entry.keyword}' in output`
+          };
+        }
+        return {
+          score: 0,
+          reasoning: `Deterministic: '${entry.keyword}' not found in output`
+        };
+      }
+    }
+    return null;
+  }
+  if (/^calls?\b/i.test(criterionText.trim())) {
+    for (const pattern of CALL_PATTERNS) {
+      if (criterionLower.includes(pattern.toLowerCase())) {
+        const found = combined.includes(pattern.toLowerCase());
+        if (found) {
+          return {
+            score: 1,
+            reasoning: `Deterministic: found '${pattern}' in output`
+          };
+        }
+        return {
+          score: 0,
+          reasoning: `Deterministic: '${pattern}' not found in output`
+        };
+      }
+    }
+    return null;
+  }
+  return null;
+}
 async function passFailScorer(task, workspacePath, stdout, stderr) {
   const outcomes = Array.isArray(task.expected_outcome) ? task.expected_outcome : task.expected_outcome.split("\n");
   const commands = outcomes.map((line) => line.replace(/^-\s*/, "").trim()).filter((line) => COMMAND_PATTERN.test(line));
@@ -513,7 +604,8 @@ async function llmJudgeScorer(task, workspacePath, stdout, stderr, config) {
   try {
     const response = await callLLM(config, userMessage, {
       systemPrompt: JUDGE_SYSTEM_PROMPT,
-      maxTokens: 1024
+      maxTokens: 1024,
+      cacheControl: true
     });
     let cleaned = response.trim();
     if (cleaned.startsWith("```")) {
@@ -544,6 +636,20 @@ async function rubricScorer(task, workspacePath, stdout, stderr, config) {
   const breakdown = [];
   let weightedSum = 0;
   for (const criterion of task.rubric) {
+    const deterministicResult = scoreCriterionDeterministic(
+      criterion.criterion,
+      stdout,
+      stderr
+    );
+    if (deterministicResult !== null) {
+      breakdown.push({
+        criterion: criterion.criterion,
+        score: deterministicResult.score,
+        weight: criterion.weight
+      });
+      weightedSum += deterministicResult.score * criterion.weight;
+      continue;
+    }
     const userMessage = [
       "## Task",
       task.description,
@@ -560,7 +666,8 @@ async function rubricScorer(task, workspacePath, stdout, stderr, config) {
     try {
       const response = await callLLM(config, userMessage, {
         systemPrompt: RUBRIC_SYSTEM_PROMPT,
-        maxTokens: 512
+        maxTokens: 512,
+        cacheControl: true
       });
       let cleaned = response.trim();
       if (cleaned.startsWith("```")) {
@@ -638,7 +745,7 @@ async function scoreTask(task, workspacePath, stdout, stderr, config) {
   }
   return score;
 }
-var COMMAND_PATTERN, SHELL_METACHAR_PATTERN, JUDGE_SYSTEM_PROMPT, RUBRIC_SYSTEM_PROMPT;
+var COMMAND_PATTERN, SHELL_METACHAR_PATTERN, JUDGE_SYSTEM_PROMPT, RUBRIC_SYSTEM_PROMPT, RAN_COMMAND_EVIDENCE, ABSENCE_PATTERNS, PRESENCE_PATTERNS, CALL_PATTERNS;
 var init_scorers = __esm({
   "src/evolve/scorers.ts"() {
     "use strict";
@@ -661,6 +768,31 @@ Return ONLY valid JSON:
   "score": 0.0-1.0,
   "reasoning": "Brief explanation"
 }`;
+    RAN_COMMAND_EVIDENCE = [
+      { keywords: ["npm run build", "build", "tsup"], evidence: ["build success", "tsup", "built in", "build completed"] },
+      { keywords: ["tsc", "typecheck"], evidence: ["tsc", "typecheck"] },
+      { keywords: ["npm run lint", "eslint", "lint"], evidence: ["lint", "eslint"] },
+      { keywords: ["npm test", "vitest", "test"], evidence: ["vitest", "test files", "tests passed", "passed (", "tests "] }
+    ];
+    ABSENCE_PATTERNS = [
+      { keywords: [".then()", ".catch()"], search: [".then(", ".catch("] },
+      { keywords: ["readfilesync", "writefilesync"], search: ["readfilesync", "writefilesync"] },
+      { keywords: ["sync"], search: ["sync"] }
+    ];
+    PRESENCE_PATTERNS = [
+      { keyword: "chalk.green", search: ["chalk.green"] },
+      { keyword: "chalk.yellow", search: ["chalk.yellow"] },
+      { keyword: "chalk.red", search: ["chalk.red"] },
+      { keyword: "chalk.cyan", search: ["chalk.cyan"] },
+      { keyword: "fs.promises", search: ["fs.promises", "fs/promises"] },
+      { keyword: "fs/promises", search: ["fs.promises", "fs/promises"] },
+      { keyword: "async/await", search: ["async ", "await "] },
+      { keyword: "@inquirer/prompts", search: ["@inquirer/prompts"] }
+    ];
+    CALL_PATTERNS = [
+      "process.exit(1)",
+      "process.exit"
+    ];
   }
 });
 
@@ -1329,7 +1461,8 @@ async function propose(iteration, workspacePath, harnessPath, history, tasks, co
   const response = await callLLM(proposerConfig, userMessage, {
     systemPrompt: PROPOSER_SYSTEM_PROMPT,
     maxTokens: 8192,
-    jsonMode: true
+    jsonMode: true,
+    cacheControl: true
   });
   return parseProposerResponse(response);
 }
@@ -1831,7 +1964,22 @@ async function parseAgents(harnessPath) {
     if (Array.isArray(disallowedTools)) {
       node.disallowedTools = disallowedTools;
     }
-    const knownKeys = /* @__PURE__ */ new Set(["name", "model", "disallowedTools"]);
+    const modelRouting = frontmatter["modelRouting"];
+    if (typeof modelRouting === "object" && modelRouting !== null) {
+      const mr = modelRouting;
+      if (typeof mr["default"] === "string") {
+        node.modelRouting = {
+          default: mr["default"]
+        };
+        if (typeof mr["escalateTo"] === "string") {
+          node.modelRouting.escalateTo = mr["escalateTo"];
+        }
+        if (typeof mr["escalateWhen"] === "string") {
+          node.modelRouting.escalateWhen = mr["escalateWhen"];
+        }
+      }
+    }
+    const knownKeys = /* @__PURE__ */ new Set(["name", "model", "disallowedTools", "modelRouting"]);
     const extra = {};
     for (const [key, value] of Object.entries(frontmatter)) {
       if (!knownKeys.has(key)) {
@@ -2457,8 +2605,9 @@ function renderRuleWithFrontmatter(rule) {
 function renderAgentWithFrontmatter(agent) {
   const hasModel = agent.model !== void 0;
   const hasDisallowed = agent.disallowedTools !== void 0 && agent.disallowedTools.length > 0;
+  const hasRouting = agent.modelRouting !== void 0;
   const hasExtra = agent.extraFrontmatter !== void 0 && Object.keys(agent.extraFrontmatter).length > 0;
-  if (!hasModel && !hasDisallowed && !hasExtra) {
+  if (!hasModel && !hasDisallowed && !hasRouting && !hasExtra) {
     return agent.content;
   }
   const yamlLines = ["---"];
@@ -2471,6 +2620,16 @@ function renderAgentWithFrontmatter(agent) {
       yamlLines.push(`  - ${tool}`);
     }
   }
+  if (hasRouting) {
+    yamlLines.push("modelRouting:");
+    yamlLines.push(`  default: ${agent.modelRouting.default}`);
+    if (agent.modelRouting.escalateTo) {
+      yamlLines.push(`  escalateTo: ${agent.modelRouting.escalateTo}`);
+    }
+    if (agent.modelRouting.escalateWhen) {
+      yamlLines.push(`  escalateWhen: ${agent.modelRouting.escalateWhen}`);
+    }
+  }
   if (hasExtra) {
     for (const [key, value] of Object.entries(agent.extraFrontmatter)) {
       if (Array.isArray(value)) {
@@ -3350,6 +3509,93 @@ var init_regularization = __esm({
   }
 });
 
+// src/evolve/targeting.ts
+function mutationToAspect(mutation) {
+  switch (mutation.type) {
+    case "update_section": {
+      const id = mutation.sectionId;
+      if (id === "conventions" || id === "gotchas" || id === "debugging" || id === "git") return "conventions";
+      if (id === "commands" || id === "custom-key-commands") return "commands";
+      if (id === "verification") return "verification";
+      if (id === "architecture") return "architecture";
+      return "general";
+    }
+    case "add_section": {
+      const id = mutation.section.id;
+      if (id === "conventions" || id === "gotchas" || id === "debugging" || id === "git") return "conventions";
+      if (id === "commands" || id === "custom-key-commands") return "commands";
+      if (id === "verification") return "verification";
+      if (id === "architecture") return "architecture";
+      return "general";
+    }
+    case "remove_section":
+    case "reorder_section":
+      return "general";
+    case "add_command":
+    case "update_command":
+    case "remove_command":
+      return "commands";
+    case "add_rule":
+    case "update_rule":
+    case "remove_rule":
+      return "rules";
+    case "add_agent":
+    case "update_agent":
+    case "remove_agent":
+      return "agents";
+    case "add_mcp_server":
+    case "remove_mcp_server":
+      return "mcp";
+    case "update_settings":
+      return "settings";
+    case "raw_text":
+      return "general";
+  }
+}
+function mutationsToAspects(mutations) {
+  const aspects = /* @__PURE__ */ new Set();
+  for (const m of mutations) {
+    aspects.add(mutationToAspect(m));
+  }
+  return aspects;
+}
+function taskDependsOnAspects(task) {
+  const aspects = TEMPLATE_ASPECTS[task.template];
+  return new Set(aspects ?? ["general"]);
+}
+function shouldReEvaluate(task, changedAspects) {
+  if (changedAspects.has("general")) return true;
+  if (changedAspects.size === 0) return false;
+  const taskAspects = taskDependsOnAspects(task);
+  if (taskAspects.has("general")) return true;
+  for (const aspect of taskAspects) {
+    if (changedAspects.has(aspect)) return true;
+  }
+  return false;
+}
+function filterTasksByAspects(tasks, changedAspects) {
+  return tasks.filter((t) => shouldReEvaluate(t, changedAspects));
+}
+var TEMPLATE_ASPECTS;
+var init_targeting = __esm({
+  "src/evolve/targeting.ts"() {
+    "use strict";
+    TEMPLATE_ASPECTS = {
+      "convention-adherence": ["conventions", "rules"],
+      "workflow-compliance": ["commands", "verification"],
+      "rule-compliance": ["rules"],
+      "intent-routing": ["settings"],
+      "add-feature": ["general"],
+      "fix-bug": ["general"],
+      "refactor": ["architecture", "conventions"],
+      "test-writing": ["verification", "commands"],
+      "config-change": ["settings", "mcp"],
+      "documentation": ["general"],
+      "persistence-completion": ["commands", "verification"]
+    };
+  }
+});
+
 // src/evolve/loop.ts
 import fs25 from "fs/promises";
 import path25 from "path";
@@ -3382,6 +3628,7 @@ async function evolve(workspacePath, tasks, kairnConfig, evolveConfig, onProgres
       }
     }
   }
+  let lastChangedAspects = null;
   let rngState = evolveConfig.rngSeed ?? 42;
   const rng = () => {
     rngState = rngState * 1664525 + 1013904223 & 4294967295;
@@ -3428,6 +3675,22 @@ async function evolve(workspacePath, tasks, kairnConfig, evolveConfig, onProgres
           tasksToRun.push(task);
         }
       }
+      if (lastChangedAspects !== null) {
+        const targetedTasks = filterTasksByAspects(tasksToRun, lastChangedAspects);
+        const skippedByTargeting = tasksToRun.filter((t) => !targetedTasks.includes(t));
+        for (const task of skippedByTargeting) {
+          const prev = prevLog.taskResults[task.id];
+          const prevVal = prev ? prev.score ?? (prev.pass ? 100 : 0) : 0;
+          carriedScores[task.id] = { pass: prevVal >= 50, score: prevVal };
+          onProgress?.({
+            type: "task-skipped",
+            iteration: iter,
+            taskId: task.id,
+            message: `Skipped ${task.id} (unaffected by mutations)`
+          });
+        }
+        tasksToRun = targetedTasks;
+      }
       const sampleSize = evolveConfig.evalSampleSize;
       if (sampleSize > 0 && sampleSize < tasksToRun.length) {
         let sampled;
@@ -3583,6 +3846,13 @@ async function evolve(workspacePath, tasks, kairnConfig, evolveConfig, onProgres
           }
           const nextIterDir2 = path25.join(workspacePath, "iterations", (iter + 1).toString());
           await applyMutations(bestHarnessPath, nextIterDir2, rollbackProposal.mutations);
+          try {
+            const rollbackIR = await parseHarness(bestHarnessPath);
+            const irMuts = translateMutations(rollbackProposal.mutations, rollbackIR);
+            lastChangedAspects = mutationsToAspects(irMuts);
+          } catch {
+            lastChangedAspects = null;
+          }
           onProgress?.({
             type: "mutations-applied",
             iteration: iter,
@@ -3687,8 +3957,16 @@ async function evolve(workspacePath, tasks, kairnConfig, evolveConfig, onProgres
         proposal.mutations
       );
       diffPatch = mutationResult.diffPatch;
+      try {
+        const currentIR = await parseHarness(harnessPath);
+        const irMuts = translateMutations(proposal.mutations, currentIR);
+        lastChangedAspects = mutationsToAspects(irMuts);
+      } catch {
+        lastChangedAspects = null;
+      }
     } catch {
       await copyDir(harnessPath, path25.join(nextIterDir, "harness"));
+      lastChangedAspects = null;
     }
     onProgress?.({
       type: "mutations-applied",
@@ -3787,6 +4065,8 @@ var init_loop = __esm({
     init_sampling();
     init_regularization();
     init_parser();
+    init_translate();
+    init_targeting();
   }
 });
 
@@ -3917,7 +4197,8 @@ ${userMessage}`;
   const response = await callLLM(proposerConfig, fullMessage, {
     systemPrompt,
     maxTokens: 8192,
-    jsonMode: true
+    jsonMode: true,
+    cacheControl: true
   });
   const proposal = parseProposerResponse(response);
   return {
@@ -4629,7 +4910,7 @@ You must output a JSON object matching the SkeletonSpec schema.
 - MCP servers: maximum 6. Prefer fewer.
 - Skills: maximum 3. Only include directly relevant ones.
 - Agents: maximum 5. Orchestration pipeline (/develop) agents.
-- Hooks: maximum 4 (auto-format, block-destructive, PostCompact, plus one contextual).
+- Hooks: maximum 5 (auto-format, block-destructive, PostCompact, memory-persistence, plus one contextual).
 
 If the workflow doesn't clearly need a tool, DO NOT include it.
 Each MCP server costs 500-2000 tokens of context window.
@@ -4752,6 +5033,12 @@ At the start of every session, before doing ANY work:
 
 This saves 2-5 exploratory turns. Never ask "what files are here?" \u2014 look first.
 
+## Sprint Contract
+
+Before implementing, confirm acceptance criteria exist in docs/SPRINT.md.
+Each criterion must be numbered, testable, and independently verifiable.
+After implementing, verify EACH criterion individually. Do not mark done until all pass.
+
 ## Completion Standards
 
 Never mark a task "done" without running the Completion Verification checklist.
@@ -4781,6 +5068,7 @@ Do not add generic filler. Every line must be specific to the user's workflow.
 15. "Engineering Standards", "Tool Usage Policy", and "Code Philosophy" sections in CLAUDE.md
 16. A "First Turn Protocol" section in CLAUDE.md (orient before working: pwd, ls, git status, check relevant runtimes, read task files)
 17. A "Completion Standards" section in CLAUDE.md (never mark done without verifying: requirements met, tests passing, no debug artifacts, reviewed from 3 perspectives)
+18. A "Sprint Contract" section in CLAUDE.md (confirm acceptance criteria exist before implementing, verify each criterion after)
 
 ## Shell-Integrated Commands
 
@@ -4837,17 +5125,19 @@ Only generate scoped rules when the workflow involves multiple code domains.
 
 Generate hooks in settings.json based on project type:
 
-**All code projects** \u2014 block destructive commands:
+**All code projects** \u2014 block destructive commands, credential leaks, injection, and network exfiltration:
 \`\`\`json
 {
   "hooks": {
-    "PreToolUse": [{
-      "matcher": "Bash",
-      "hooks": [{
-        "type": "command",
-        "command": "CMD=$(cat | jq -r '.tool_input.command // empty') && echo \\"$CMD\\" | grep -qiE 'rm\\\\s+-rf\\\\s+/|DROP\\\\s+TABLE|curl.*\\\\|\\\\s*sh' && echo 'Blocked destructive command' >&2 && exit 2 || true"
-      }]
-    }]
+    "PreToolUse": [
+      {
+        "matcher": "Bash",
+        "hooks": [{
+          "type": "command",
+          "command": "CMD=$(cat | jq -r '.tool_input.command // empty') && echo \\"$CMD\\" | grep -qiE 'rm\\\\s+-rf\\\\s+/|DROP\\\\s+(TABLE|DATABASE)|curl.*\\\\|\\\\s*sh|:(){ :|:& };:|git\\\\s+push.*--force(?!-with-lease)|ch(mod|own).*-R\\\\s+/|npm\\\\s+publish(?!.*--dry-run)|(api[_-]?key|secret|token|password)\\\\s*[:=]|AKIA[0-9A-Z]{16}|BEGIN.*PRIVATE\\\\s+KEY|;\\\\s*(DROP|DELETE|ALTER|TRUNCATE)\\\\s+|\\\\.\\\\./\\\\.\\\\./\\\\.\\\\./|nc\\\\s+.*-e|/dev/tcp/|bash\\\\s+-i|curl.*-d.*@|wget.*--post-file' && echo 'Blocked dangerous command' >&2 && exit 2 || true"
+        }]
+      }
+    ]
   }
 }
 \`\`\`
@@ -4889,6 +5179,23 @@ All projects should include a PostCompact hook to restore context after compacti
 
 Merge this into the settings hooks alongside the PreToolUse and PostToolUse hooks.
 
+For long-running sessions (>2 hours or >3 compactions), prefer "Full Reset" over re-inject:
+replace the prompt-type PostCompact hook with a command-type hook that pipes CLAUDE.md + SPRINT.md + DECISIONS.md content directly into additionalContext.
+
+## Memory Persistence Hooks
+
+For projects with multi-session workflows, include SessionStart/End hooks that persist context to \`.claude/memory.json\`:
+- **SessionEnd:** Save recent decisions, sprint status, and known gotchas to \`.claude/memory.json\`
+- **SessionStart:** Load \`.claude/memory.json\` and inject as additionalContext
+
+This ensures accumulated project knowledge survives session boundaries.
+
+## Persistence Routing Hook
+
+For code projects with autonomy level 3+, include a \`UserPromptSubmit\` hook that routes complex tasks through \`/project:persist\`.
+Generate \`.claude/hooks/persist-router.mjs\` (the routing logic) and add a \`UserPromptSubmit\` hook entry in settings.json that invokes it.
+Also add \`"persistence_routing": "auto"\` (level 3-4) or \`"persistence_routing": "manual"\` (level 1-2) to settings.json.
+
 ## For Code Projects, Additionally Include
 
 - \`/project:plan\` command (plan before coding)
@@ -4898,7 +5205,7 @@ Merge this into the settings hooks alongside the PreToolUse and PostToolUse hook
 - \`/project:status\` command (live git status, recent commits, SPRINT.md overview using ! prefix)
 - \`/project:fix\` command (takes $ARGUMENTS as issue number, plans fix, implements, tests, commits)
 - \`/project:sprint\` command (define acceptance criteria before coding, writes to docs/SPRINT.md)
-- \`/project:develop\` command (full development pipeline \u2014 orchestrates @architect \u2192 @planner \u2192 @implementer \u2192 @verifier \u2192 @fixer \u2192 @grill \u2192 @doc-updater through spec, plan, TDD implement, review, and doc update phases). MUST include a Phase 7 "Completion Gate" that runs a Completion Verification checklist before marking the feature done: re-read original requirements, confirm each is met with evidence, run test suite + lint/typecheck, review git diff for unexpected changes or debug artifacts, answer 3 perspective questions (test engineer, code reviewer, requesting user). If ANY check fails, loop back to fix before completing.
+- \`/project:develop\` command (full development pipeline \u2014 orchestrates @architect \u2192 @planner \u2192 @implementer \u2192 @verifier \u2192 @fixer \u2192 @grill \u2192 @doc-updater through spec, plan, TDD implement, review, and doc update phases). Phase 4 (Verify) MUST validate EACH acceptance criterion from docs/SPRINT.md individually, reporting PASS/FAIL per item as a contract scorecard. MUST include a Phase 7 "Completion Gate" that runs a Completion Verification checklist before marking the feature done: re-read original requirements, confirm each is met with evidence, run test suite + lint/typecheck, review git diff for unexpected changes or debug artifacts, answer 3 perspective questions (test engineer, code reviewer, requesting user). If ANY check fails, loop back to fix before completing.
 - A TDD skill using the 3-phase isolation pattern (RED \u2192 GREEN \u2192 REFACTOR):
   - RED: Write failing test only. Verify it FAILS.
   - GREEN: Write MINIMUM code to pass. Nothing extra.
@@ -4908,16 +5215,31 @@ Merge this into the settings hooks alongside the PreToolUse and PostToolUse hook
   - \`@qa-orchestrator\` (sonnet) \u2014 delegates to linter and e2e-tester, compiles QA report
   - \`@linter\` (haiku) \u2014 runs formatters, linters, security scanners
   - \`@e2e-tester\` (sonnet, only when Playwright is in tools) \u2014 browser-based QA via Playwright
-- Development pipeline agents (used by /project:develop):
-  - \`@architect\` (opus) \u2014 conducts spec interview with user, writes confirmed spec to docs/SPRINT.md
-  - \`@planner\` (opus) \u2014 reads spec and codebase, creates step-by-step implementation plan in docs/PLAN.md
-  - \`@implementer\` (sonnet) \u2014 TDD-focused implementation, writes failing tests then minimum code to pass
-  - \`@fixer\` (sonnet) \u2014 targeted bug fixing from verifier/review feedback
-  - \`@doc-updater\` (haiku) \u2014 extracts decisions and learnings from completed work, updates docs/DECISIONS.md and docs/LEARNINGS.md
-- \`/project:spec\` command (interview-based spec creation \u2014 asks 5-8 questions one at a time, writes structured spec to docs/SPRINT.md, does NOT start coding until confirmed)
+- A "Model Selection" section in generated agents:
+  \`\`\`
+  ## Model Selection (all agents)
+  - Haiku: simple file edits, linting, formatting, doc updates (<50 lines changed)
+  - Sonnet: implementation, testing, debugging, code review (50-500 lines)
+  - Opus: architecture decisions, spec writing, complex refactors (>500 lines or cross-cutting)
+  Default: Sonnet. Only escalate to Opus when the task involves multi-file architecture or ambiguous requirements.
+  \`\`\`
+- Development pipeline agents (used by /project:develop). Each agent should include a modelRouting field in its YAML frontmatter:
+  - \`@architect\` (default: opus) \u2014 conducts spec interview with user, writes confirmed spec to docs/SPRINT.md with numbered acceptance criteria. Your spec is a CONTRACT \u2014 the verifier will check every criterion. Vague criteria = guaranteed rework.
+  - \`@planner\` (default: sonnet, escalate to opus for cross-cutting changes) \u2014 reads spec and codebase, creates step-by-step implementation plan in docs/PLAN.md
+  - \`@implementer\` (default: sonnet, escalate to opus for cross-cutting changes) \u2014 TDD-focused implementation, writes failing tests then minimum code to pass
+  - \`@fixer\` (default: sonnet, use haiku for single-file fixes) \u2014 targeted bug fixing from verifier/review feedback
+  - \`@doc-updater\` (default: haiku) \u2014 extracts decisions and learnings from completed work, updates docs/DECISIONS.md and docs/LEARNINGS.md
+- \`/project:spec\` command (interview-based spec creation \u2014 asks 5-8 questions one at a time, writes structured spec to docs/SPRINT.md with ## Acceptance Criteria containing 3-8 numbered, testable conditions. Each criterion must be independently verifiable. Does NOT start coding until confirmed)
 - \`/project:prove\` command (runs tests, shows git diff vs main, rates confidence HIGH/MEDIUM/LOW with evidence)
 - \`/project:grill\` command (adversarial code review \u2014 challenges each change with "why this approach?", "what if X input?", rates BLOCKER/SHOULD-FIX/NITPICK, blocks until BLOCKERs resolved)
 - \`/project:reset\` command (reads DECISIONS.md and LEARNINGS.md, proposes clean restart, stashes current work, implements elegant solution)
+- \`/project:persist\` command (persistent execution loop \u2014 reads acceptance criteria from docs/SPRINT.md, works criterion-by-criterion with structured progress tracking in .claude/progress.json, auto-retries on verification failure up to 3 times per criterion, delegates to @grill for review gate before completion, resumes from progress.json if session was interrupted). The command protocol:
+  1. Load or initialize .claude/progress.json from docs/SPRINT.md numbered acceptance criteria
+  2. For each incomplete criterion: implement, run verification (build/test/typecheck/lint), mark PASSED or retry (max 3 attempts per criterion, mark BLOCKED after 3 failures)
+  3. After all criteria attempted: if any BLOCKED report which and why; if all PASSED proceed to review gate
+  4. Review gate: delegate to @grill for adversarial review; fix blockers if found (max 1 fix cycle)
+  5. Persist state: write final progress.json; include progress summary in memory.json for session resume
+  Resume protocol: when progress.json exists, skip PASSED criteria, resume from first non-PASSED criterion, carry forward failure notes from prior attempts.
 
 ## For Research Projects, Additionally Include
 
@@ -4962,7 +5284,7 @@ Return ONLY valid JSON matching this structure:
 \`\`\`json
 {
   "claude_md": "Full CLAUDE.md content (under 150 lines)",
-  "commands": { "help": "...", "develop": "...", "status": "...", "fix": "...", "sprint": "...", "spec": "...", "prove": "...", "grill": "...", "reset": "..." },
+  "commands": { "help": "...", "develop": "...", "status": "...", "fix": "...", "sprint": "...", "spec": "...", "prove": "...", "grill": "...", "reset": "...", "persist": "..." },
   "rules": { "continuity": "...", "security": "..." },
   "agents": { "architect": "...", "planner": "...", "implementer": "...", "fixer": "...", "doc-updater": "...", "qa-orchestrator": "...", "linter": "...", "e2e-tester": "..." },
   "skills": { "skill-name/SKILL": "..." },
@@ -5067,6 +5389,12 @@ At the start of every session, before doing ANY work:
 
 This saves 2-5 exploratory turns. Never ask "what files are here?" \u2014 look first.
 
+## Sprint Contract
+
+Before implementing, confirm acceptance criteria exist in docs/SPRINT.md.
+Each criterion must be numbered, testable, and independently verifiable.
+After implementing, verify EACH criterion individually. Do not mark done until all pass.
+
 ## Completion Standards
 
 Never mark a task "done" without running the Completion Verification checklist.
@@ -5096,6 +5424,7 @@ Do not add generic filler. Every line must be specific to the user's workflow.
 15. "Engineering Standards", "Tool Usage Policy", and "Code Philosophy" sections in CLAUDE.md
 16. A "First Turn Protocol" section in CLAUDE.md (orient before working: pwd, ls, git status, check relevant runtimes, read task files)
 17. A "Completion Standards" section in CLAUDE.md (never mark done without verifying: requirements met, tests passing, no debug artifacts, reviewed from 3 perspectives)
+18. A "Sprint Contract" section in CLAUDE.md (confirm acceptance criteria exist before implementing, verify each criterion after)
 
 ## Tool Selection Rules
 
@@ -5114,7 +5443,7 @@ Do not add generic filler. Every line must be specific to the user's workflow.
 - Skills: maximum 3. Only include directly relevant ones.
 - Agents: maximum 5. Orchestration pipeline (/develop) agents.
 - Commands: no limit (loaded on demand, zero context cost).
-- Hooks: maximum 4 (auto-format, block-destructive, PostCompact, plus one contextual).
+- Hooks: maximum 5 (auto-format, block-destructive, PostCompact, memory-persistence, plus one contextual).
 
 If the workflow doesn't clearly need a tool, DO NOT include it.
 Each MCP server costs 500-2000 tokens of context window.
@@ -5143,7 +5472,8 @@ Return ONLY valid JSON matching this structure:
     },
     "commands": {
       "help": "markdown content for /project:help",
-      "develop": "markdown content for /project:develop"
+      "develop": "markdown content for /project:develop",
+      "persist": "markdown content for /project:persist"
     },
     "rules": {
       "continuity": "markdown content for continuity rule",
@@ -6220,6 +6550,9 @@ function applyAutonomyLevel(spec) {
   const agents = spec.harness.agents ?? {};
   const docs = spec.harness.docs ?? {};
   const settings = spec.harness.settings ?? {};
+  if (!("persistence_routing" in settings)) {
+    settings.persistence_routing = level >= 3 ? "auto" : "manual";
+  }
   if (level >= 1) {
     if (!("tour" in commands)) {
       commands.tour = TOUR_COMMAND;
@@ -6291,6 +6624,85 @@ var ENV_LOADER_HOOK = {
     command: 'if [ -f .env ] && [ -n "$CLAUDE_ENV_FILE" ]; then grep -v "^#" .env | grep -v "^$" | grep "=" >> "$CLAUDE_ENV_FILE"; fi'
   }]
 };
+var PERSIST_ROUTER_TEMPLATE = `import { readFileSync } from 'fs';
+
+const input = JSON.parse(readFileSync('/dev/stdin', 'utf8'));
+const prompt = (input.prompt ?? '').trim();
+
+// Pass-through patterns (fast exit)
+const PASSTHROUGH = /^(what|how|why|where|when|can you|does|is |show me|find |search |list |\\/project:)/i;
+const SINGLE_FILE = /^(edit|fix the typo|update the comment|change the|rename) .{3,60}$/i;
+
+if (PASSTHROUGH.test(prompt) || SINGLE_FILE.test(prompt) || prompt.length < 20) {
+  process.stdout.write(JSON.stringify({ continue: true }));
+  process.exit(0);
+}
+
+// Check config for routing mode
+let routingMode = 'auto';
+try {
+  const settings = JSON.parse(readFileSync('.claude/settings.json', 'utf8'));
+  routingMode = settings.persistence_routing ?? 'auto';
+} catch { /* default to auto */ }
+
+if (routingMode === 'off') {
+  process.stdout.write(JSON.stringify({ continue: true }));
+  process.exit(0);
+}
+
+// Complexity signals
+const signals = [];
+
+if (/\\b(then|after that|and also|next|finally|step \\d|first .* then)\\b/i.test(prompt)) {
+  signals.push('multi-step');
+}
+if (/\\b(add|implement|build|create|integrate|set up)\\b.*\\b(feature|auth|api|endpoint|page|component|module|service|database|migration)\\b/i.test(prompt)) {
+  signals.push('feature-scope');
+}
+if (/\\b(migrate|convert|replace|upgrade|refactor|rewrite|restructure)\\b/i.test(prompt)) {
+  signals.push('refactor-scope');
+}
+if (/\\b(when .* happens|steps to reproduce|broken|crash|regression|fails when)\\b/i.test(prompt)) {
+  signals.push('bug-with-repro');
+}
+if (/\\b(persist|keep working|don't stop|until done|until .* pass)\\b/i.test(prompt)) {
+  signals.push('explicit');
+}
+if (prompt.split(/\\s+/).length > 50) {
+  signals.push('long-prompt');
+}
+
+const shouldRoute = routingMode === 'manual'
+  ? signals.includes('explicit')
+  : signals.length >= 2 || signals.includes('explicit');
+
+if (shouldRoute) {
+  process.stdout.write(JSON.stringify({
+    continue: true,
+    hookSpecificOutput: {
+      hookEventName: 'UserPromptSubmit',
+      additionalContext: [
+        'PERSISTENCE ROUTING: This task has complexity signals (' + signals.join(', ') + ').',
+        'Execute this using the /project:persist workflow:',
+        '1. Ensure acceptance criteria exist in docs/SPRINT.md (create from this prompt if needed)',
+        '2. Initialize .claude/progress.json',
+        '3. Work criterion-by-criterion until all pass',
+        '4. Run review gate before marking complete',
+      ].join('\\n'),
+    },
+  }));
+} else {
+  process.stdout.write(JSON.stringify({ continue: true }));
+}
+`;
+var PERSIST_ROUTER_HOOK = {
+  matcher: "",
+  hooks: [{
+    type: "command",
+    command: 'node "$CLAUDE_PROJECT_DIR/.claude/hooks/persist-router.mjs"',
+    timeout: 5
+  }]
+};
 function resolveSettings(spec, options) {
   const settings = spec.harness.settings;
   const base = settings && Object.keys(settings).length > 0 ? { ...settings } : {};
@@ -6304,6 +6716,13 @@ function resolveSettings(spec, options) {
     hooks.SessionStart = sessionStart;
     base.hooks = hooks;
   }
+  if (isCodeProject(spec) && (spec.autonomy_level ?? 1) >= 3) {
+    const hooks = base.hooks ?? {};
+    const userPromptSubmit = hooks.UserPromptSubmit ?? [];
+    userPromptSubmit.push(PERSIST_ROUTER_HOOK);
+    hooks.UserPromptSubmit = userPromptSubmit;
+    base.hooks = hooks;
+  }
   const hasIntentHooks = spec.harness.hooks && Object.keys(spec.harness.hooks).length > 0;
   if (hasIntentHooks) {
     const hooks = base.hooks ?? {};
@@ -6395,6 +6814,9 @@ function buildFileMap(spec, options) {
       files.set(".claude/hooks/intent-log.jsonl", "");
     }
   }
+  if (isCodeProject(spec) && (spec.autonomy_level ?? 1) >= 3) {
+    files.set(".claude/hooks/persist-router.mjs", PERSIST_ROUTER_TEMPLATE);
+  }
   return files;
 }
 async function writeEnvironment(spec, targetDir, options) {
@@ -6465,6 +6887,11 @@ async function writeEnvironment(spec, targetDir, options) {
       written.push(".claude/hooks/intent-log.jsonl");
     }
   }
+  if (isCodeProject(spec) && (spec.autonomy_level ?? 1) >= 3) {
+    const p = path5.join(claudeDir, "hooks", "persist-router.mjs");
+    await writeFile(p, PERSIST_ROUTER_TEMPLATE);
+    written.push(".claude/hooks/persist-router.mjs");
+  }
   return written;
 }
 function summarizeSpec(spec, registry) {
@@ -8163,14 +8590,20 @@ var EVAL_TEMPLATES = {
     name: "Intent Routing",
     description: "Test that natural language prompts route to the correct workflow command via intent hooks",
     bestFor: ["feature-development", "full-stack", "api-building"]
+  },
+  "persistence-completion": {
+    id: "persistence-completion",
+    name: "Persistence Completion",
+    description: "Can the agent complete a multi-criterion task using the persistence loop?",
+    bestFor: ["feature-development", "full-stack", "api-building", "maintenance"]
   }
 };
 function selectTemplatesForWorkflow(workflowType) {
   const mapping = {
-    "feature-development": ["add-feature", "test-writing", "convention-adherence", "workflow-compliance", "intent-routing"],
-    "api-building": ["add-feature", "fix-bug", "test-writing", "convention-adherence"],
-    "full-stack": ["add-feature", "fix-bug", "test-writing", "convention-adherence"],
-    "maintenance": ["fix-bug", "refactor", "test-writing", "rule-compliance"],
+    "feature-development": ["add-feature", "test-writing", "convention-adherence", "workflow-compliance", "intent-routing", "persistence-completion"],
+    "api-building": ["add-feature", "fix-bug", "test-writing", "convention-adherence", "persistence-completion"],
+    "full-stack": ["add-feature", "fix-bug", "test-writing", "convention-adherence", "persistence-completion"],
+    "maintenance": ["fix-bug", "refactor", "test-writing", "rule-compliance", "persistence-completion"],
     "debugging": ["fix-bug", "test-writing", "rule-compliance"],
     "qa": ["fix-bug", "test-writing", "add-feature", "workflow-compliance"],
     "architecture": ["refactor", "test-writing", "config-change", "convention-adherence"],
@@ -8191,6 +8624,7 @@ IMPORTANT: For harness-aware templates (convention-adherence, workflow-complianc
 - convention-adherence: Task must require following specific conventions from CLAUDE.md (naming, file structure, patterns). Judge by whether output matches the conventions.
 - workflow-compliance: Task must require using project slash commands or workflow steps defined in .claude/commands/. Judge by whether the agent followed the defined workflow.
 - rule-compliance: Task must create a scenario where .claude/rules/ content is relevant. Judge by whether the agent respected all rules.
+- persistence-completion: Task MUST have 3+ acceptance criteria that require sequential implementation. The task description should be a realistic feature request \u2014 the agent must parse it into criteria. Judge by: (a) all criteria met (progress.json status: complete), (b) structured tracking used (progress.json exists with 3+ criteria), (c) tests pass, (d) review gate executed (progress.json review field present).
 
 These harness-aware tasks are critical \u2014 they test whether the .claude/ environment actually improves agent behavior.
 
@@ -8860,7 +9294,7 @@ async function applyEvolution(workspacePath, projectRoot, targetIteration, pbt)
 // src/commands/evolve.ts
 var DEFAULT_CONFIG = {
   model: "claude-sonnet-4-6",
-  proposerModel: "claude-opus-4-6",
+  proposerModel: "claude-sonnet-4-6",
   scorer: "pass-fail",
   maxIterations: 5,
   parallelTasks: 1,