@polderlabs/bizar 10.7.2 → 10.10.0

package/.claude/hooks/worker-suggest.mjs

@@ -4,10 +4,10 @@
  *
  * Bizar Background Workers — UserPromptSubmit hook.
  *
- * Runs on every user prompt. Calls `dispatch()` from cli/worker-dispatcher.mjs
- * to find Bizar skills/agents that should be suggested for the prompt text.
- * Emits suggestions to stderr (stdout is reserved for Claude Code's hook
- * protocol). Always exits 0 — this is a suggestion, not a gate.
+ * Runs on every user prompt. Injects the mandatory Bizar delegation policy,
+ * then calls `dispatch()` from cli/worker-dispatcher.mjs to add specialized
+ * skill/agent suggestions. Emits suggestions to stderr (stdout is reserved for
+ * Claude Code's hook protocol).
  *
  * Claude Code stdin shape (UserPromptSubmit):
  *   {
@@ -15,7 +15,7 @@
  *     "transcript_path": "...",
  *     "cwd": "...",
  *     "hook_event_name": "UserPromptSubmit",
- *     "user_prompt": "raw prompt text"
+ *     "prompt": "raw prompt text"
  *   }
  *
  * Claude Code stdout shape (hookSpecificOutput.additionalContext):
@@ -42,7 +42,7 @@ process.stdin.on('end', () => {
     input = {};
   }
 
-  const prompt = String(input.user_prompt || '').trim();
+  const prompt = String(input.prompt ?? input.user_prompt ?? '').trim();
 
   // Empty prompts get the silent treatment — the user has not yet
   // committed any intent.
@@ -57,6 +57,14 @@ process.stdin.on('end', () => {
     return;
   }
 
+  const routePolicy = [
+    'Mandatory Bizar routing policy:',
+    '- If this is the primary session, use the Agent tool to delegate the request to custom Bizar agent @mike before doing task analysis or implementation.',
+    '- @mike must route trivial work to @brenda and non-trivial work through the configured research, plan, implementation, review, and verification agents.',
+    '- Do not implement directly in the primary session.',
+    '- If you are already running as a Bizar custom agent, follow your assigned role and do not recursively dispatch yourself.',
+  ].join('\n');
+
   let suggestions;
   try {
     suggestions = dispatch(prompt, { maxSuggestions: 3 });
@@ -69,7 +77,7 @@ process.stdin.on('end', () => {
     process.stdout.write(JSON.stringify({
       hookSpecificOutput: {
         hookEventName: 'UserPromptSubmit',
-        additionalContext: '',
+        additionalContext: routePolicy,
       },
     }) + '\n');
     process.exit(0);
@@ -85,17 +93,17 @@ process.stdin.on('end', () => {
     );
   }
 
-  // Build additionalContext for the model so the next turn is primed.
-  let note;
-  if (suggestions.length === 0) {
-    note = '';
-  } else {
+  // Build additionalContext for the model so Bizar routing is mandatory even
+  // when no specialized worker pattern matches.
+  let note = routePolicy;
+  if (suggestions.length > 0) {
     const lines = suggestions.map((s) => {
       const skillPart = s.skill ? `, skill=${s.skill}` : '';
       const agentPart = s.agent ? `, agent=${s.agent}` : '';
       return `- ${s.workerId} (weight=${s.weight}${skillPart}${agentPart}) matched "${s.matchedPattern}"`;
     });
-    note =
+    note +=
+      '\n\n' +
       'Bizar workers suggest the following skills/agents for this prompt:\n' +
       lines.join('\n');
   }

package/.claude/hooks/worktree-bootstrap.mjs

@@ -0,0 +1,64 @@
+#!/usr/bin/env node
+// SubagentStart — initialize an isolated editing worktree without sharing
+// mutable build output or task state with sibling agents.
+
+import { spawnSync } from 'node:child_process';
+import { existsSync } from 'node:fs';
+import { join, resolve } from 'node:path';
+
+let raw = '';
+process.stdin.setEncoding('utf8');
+process.stdin.on('data', (chunk) => { raw += chunk; });
+process.stdin.on('end', () => {
+  let input = {};
+  try { input = JSON.parse(raw || '{}'); } catch { input = {}; }
+
+  const cwd = resolve(String(input.cwd || process.cwd()));
+  const projectDir = resolve(process.env.CLAUDE_PROJECT_DIR || cwd);
+  const setup = join(projectDir, 'scripts', 'worktree-setup.sh');
+  if (!existsSync(setup)) {
+    process.stdout.write('{}\n');
+    return;
+  }
+
+  const top = spawnSync('git', ['rev-parse', '--show-toplevel'], {
+    cwd,
+    encoding: 'utf8',
+  });
+  const list = spawnSync('git', ['worktree', 'list', '--porcelain'], {
+    cwd,
+    encoding: 'utf8',
+  });
+  if (top.status !== 0 || list.status !== 0) {
+    process.stdout.write('{}\n');
+    return;
+  }
+
+  const worktree = resolve(top.stdout.trim());
+  const main = /^worktree (.+)$/m.exec(list.stdout)?.[1];
+  if (!main || resolve(main) === worktree) {
+    process.stdout.write('{}\n');
+    return;
+  }
+
+  const result = spawnSync('bash', [setup, worktree], {
+    cwd: projectDir,
+    encoding: 'utf8',
+  });
+  if (result.status !== 0) {
+    const reason = (result.stderr || result.stdout || 'unknown error').trim();
+    process.stdout.write(JSON.stringify({
+      continue: false,
+      stopReason: `Bizar worktree bootstrap failed: ${reason}`,
+    }) + '\n');
+    return;
+  }
+
+  process.stdout.write(JSON.stringify({
+    hookSpecificOutput: {
+      hookEventName: 'SubagentStart',
+      additionalContext:
+        'Bizar: this editing agent is isolated in a worktree; claim a bizar task scope before modifying shared project paths.',
+    },
+  }) + '\n');
+});

package/.claude/model-router.json

@@ -0,0 +1,57 @@
+{
+  "$schema": "https://bizar.dev/schema/model-router.v1.json",
+  "version": "11.0.0",
+  "endpoint": "http://localhost:20128/v1",
+  "comment": "Single local router. Set BIZAR_MODEL_ROUTER_URL or ANTHROPIC_BASE_URL to override. Models discovered live from /v1/models.",
+  "tiers": {
+    "premium": {
+      "models": ["cx/gpt-5.6-sol"],
+      "purpose": "Complex debugging, planning work, UI design language, deep review — heaviest reasoning, highest cost."
+    },
+    "high": {
+      "models": ["cx/gpt-5.6-terra"],
+      "purpose": "Complex implementation, routing, architecture — high reasoning at lower cost than sol."
+    },
+    "mid-design": {
+      "models": ["cx/gpt-5.6-luna"],
+      "purpose": "Design critique alt — visually-oriented mid-tier."
+    },
+    "default": {
+      "models": ["bizar/MiniMax-M3"],
+      "purpose": "Everyday default — fast general-purpose reasoning."
+    },
+    "mid": {
+      "models": ["bizar/MiniMax-M2.7"],
+      "purpose": "Mid-complexity implementation, research — slightly cheaper than default."
+    },
+    "budget": {
+      "models": ["bizar/MiniMax-M2.5", "oc/deepseek-v4-flash-free", "oc/mimo-v2.5-free"],
+      "purpose": "Trivial lookups, mechanical edits, clarifying questions — lowest cost."
+    }
+  },
+  "agents": {
+    "kevin": { "model": "bizar/MiniMax-M3", "tier": "default", "rationale": "Browser driving is mechanical + reactive." },
+    "brad":         { "model": "cx/gpt-5.6-sol",  "tier": "premium", "rationale": "UI design language needs the highest reasoning." },
+    "linda":       { "model": "cx/gpt-5.6-sol",  "tier": "premium", "rationale": "Plan review + audit; weakest reasoning is the bottleneck." },
+    "susan":         { "model": "bizar/MiniMax-M3", "tier": "default", "rationale": "Read-only Q&A — fast default." },
+    "brenda":      { "model": "bizar/MiniMax-M3", "tier": "default", "rationale": "Mechanical edits and .bizar/ maintenance." },
+    "steve":        { "model": "bizar/MiniMax-M3", "tier": "default", "rationale": "Git/GitHub ops — command templates." },
+    "greg":         { "model": "bizar/MiniMax-M3", "tier": "default", "rationale": "Codebase research — sheer volume, cheap model wins." },
+    "mike":          { "model": "cx/gpt-5.6-terra","tier": "high",    "rationale": "Routing + decomposition; high reasoning required." },
+    "pam":         { "model": "oc/deepseek-v4-flash-free", "tier": "budget", "rationale": "Single-shot tiny tasks — budget tier." },
+    "oscar": { "model": "bizar/MiniMax-M3", "tier": "default", "rationale": "Semantic code search — default reasoning." },
+    "todd":          { "model": "bizar/MiniMax-M2.7","tier": "mid",     "rationale": "Mid-complexity implementation; cheaper than default." },
+    "karen":           { "model": "cx/gpt-5.6-terra","tier": "high",    "rationale": "Top-tier implementation; high reasoning." },
+    "carl":        { "model": "cx/gpt-5.6-sol",  "tier": "premium", "rationale": "Last-resort debug — strongest reasoning." },
+    "janet":           { "model": "oc/mimo-v2.5-free", "tier": "budget", "rationale": "Clarifying question; budget tier." },
+    "paul":            { "model": "cx/gpt-5.6-sol",  "tier": "premium", "rationale": "Default planner for every non-trivial session. Planning quality is the bottleneck for downstream impl cost — premium reasoning wins." },
+    "ria":             { "model": "cx/gpt-5.6-sol",  "tier": "premium", "rationale": "UI/UX design craft — typography, spacing, motion, anti-slop. Same reasoning budget as @brad." }
+  },
+  "policies": {
+    "gpt_only_for": ["complex debugging", "planning work", "default planning", "UI design language", "UI/UX craft", "deep review", "routing/decomposition"],
+    "gpt_never_for": ["research", "implementation", "everyday tasks", "Q&A", "mechanical edits"],
+    "fallback_chain": ["cx/gpt-5.6-sol", "cx/gpt-5.6-terra", "cx/gpt-5.6-luna", "bizar/MiniMax-M3", "bizar/MiniMax-M2.7", "bizar/MiniMax-M2.5", "oc/deepseek-v4-flash-free", "oc/mimo-v2.5-free"],
+    "escalate_on_failure": true,
+    "cost_ceiling_per_session_usd": 5.0
+  }
+}

package/.claude/settings.json

@@ -1,167 +1,119 @@
 {
   "$schema": "https://json.schemastore.org/claude-code-settings.json",
   "mcpServers": {
-    "bizar": {
-      "command": "npx",
-      "args": ["-y", "@polderlabs/bizar-sdk", "mcp"],
-      "env": {
-        "BIZAR_HOME": "${HOME}/.bizar_home",
-        "BIZAR_MEMORY_VAULT": "${HOME}/.bizar_home/memory"
-      }
-    },
-    "semble": {
-      "command": "semble",
-      "args": ["mcp"]
-    },
-    "kevin": {
-      "command": "kevin",
-      "args": ["mcp"]
-    }
+    "bizar": { "command": "npx", "args": ["-y", "@polderlabs/bizar-sdk", "mcp"], "env": { "BIZAR_HOME": "${HOME}/.config/bizar" } },
+    "semble": { "command": "semble", "args": ["mcp"] },
+    "agent-browser": { "command": "agent-browser", "args": ["mcp"] }
   },
   "permissions": {
+    "defaultMode": "acceptEdits",
     "allow": [
-      "mcp__bizar__memory_read",
-      "mcp__bizar__memory_list",
-      "mcp__bizar__memory_search",
-      "mcp__bizar__memory_write",
-      "mcp__bizar__plan_action",
-      "mcp__bizar__open_kb",
-      "mcp__bizar__loop_list",
-      "mcp__bizar__loop_status",
-      "mcp__bizar__loop_start",
-      "mcp__bizar__loop_stop",
-      "mcp__bizar__graph_query",
-      "mcp__bizar__graph_path",
-      "mcp__bizar__danger_check",
-      "Read(~/.bizar_memory/**)",
-      "mcp__semble__*",
-      "mcp__agent-browser__*"
+      "mcp__bizar__plan_action", "mcp__bizar__loop_list", "mcp__bizar__loop_status",
+      "mcp__bizar__loop_start", "mcp__bizar__loop_stop", "mcp__bizar__graph_query",
+      "mcp__bizar__graph_path", "mcp__bizar__list_instincts", "mcp__bizar__list_decisions",
+      "mcp__semble__*", "mcp__agent-browser__*"
+    ],
+    "ask": [
+      "Bash(git commit *)", "Bash(git push *)", "Bash(gh pr create *)", "Bash(gh pr edit *)",
+      "Bash(gh pr merge *)", "Bash(gh release *)", "Bash(npm publish *)", "Bash(bun publish *)",
+      "Bash(vercel deploy *)", "Bash(wrangler deploy *)"
     ],
     "deny": [
-      "Read(./.env)",
-      "Read(./secrets/**)",
-      "Bash(rm -rf /)",
-      "Bash(sudo *)"
+      "Read(./.env)", "Read(./.env.*)", "Read(./secrets/**)",
+      "Bash(rm -rf /)", "Bash(sudo *)", "Bash(git push --force *)", "Bash(git push -f *)",
+      "Bash(git rebase *)", "Write(./node_modules/**)"
     ]
   },
+  "autoMode": {
+    "environment": [
+      "$defaults",
+      "Source control: the current repository and its configured remotes.",
+      "Development scope: local files, tests, builds, and declared dependencies in the current repository."
+    ],
+    "allow": [
+      "$defaults",
+      "Local reversible edits and test/build commands within the current repository.",
+      "Installing dependencies already declared by repository manifests without changing those manifests.",
+      "Read-only Git, GitHub, documentation, and package-registry operations."
+    ],
+    "soft_deny": [
+      "$defaults",
+      "Commits, pushes, pull-request mutations, releases, package publication, or deployments without the configured human approval hook.",
+      "Force push, rebase, remote branch deletion, or other history rewriting.",
+      "Production or shared-infrastructure writes, credential changes, public exposure, or irreversible local destruction."
+    ]
+  },
+  "attribution": { "commit": "", "pr": "" },
+  "worktree": {
+    "baseRef": "head",
+    "cleanupPeriodDays": 7
+  },
+  "enableWorkflows": true,
+  "alwaysThinkingEnabled": true,
+  "autoDreamEnabled": true,
+  "showThinkingSummaries": true,
   "env": {
-    "BIZAR_HOME": "${HOME}/.bizar_home",
-    "BIZAR_MEMORY_VAULT": "${HOME}/.bizar_home/memory",
-    "BIZAR_LOOPS_DIR": "${HOME}/.bizar_loops",
+    "BIZAR_HOME": "${HOME}/.config/bizar",
     "ANTHROPIC_BASE_URL": "http://localhost:20128/v1",
     "BIZAR_MODEL_ROUTER_URL": "http://localhost:20128/v1"
   },
   "hooks": {
     "PreToolUse": [
-      {
-        "matcher": "Write|Edit|MultiEdit",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node /home/drb0rk/projects/BizarHarness/.claude/hooks/pretooluse-editwrite.mjs",
-            "timeout": 30
-          }
-        ]
-      },
-      {
-        "matcher": "Bash",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node /home/drb0rk/projects/BizarHarness/.claude/hooks/pretooluse-bash.mjs",
-            "timeout": 30
-          }
-        ]
-      }
+      { "matcher": "Write|Edit|MultiEdit", "hooks": [
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/pretooluse-editwrite.mjs\"", "timeout": 30 },
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/path-ownership-guard.mjs\"", "timeout": 30 },
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/content-style-guard.mjs\"", "timeout": 30 }
+      ] },
+      { "matcher": "Bash", "hooks": [
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/pretooluse-bash.mjs\"", "timeout": 30 },
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/git-workflow-guard.mjs\"", "timeout": 30 },
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/content-style-guard.mjs\"", "timeout": 30 },
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/simplify-guard.mjs\"", "timeout": 30 }
+      ] }
     ],
     "PostToolUse": [
-      {
-        "matcher": "Edit|Write|MultiEdit",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node /home/drb0rk/projects/BizarHarness/.claude/hooks/posttooluse-editwrite.mjs",
-            "timeout": 30
-          }
-        ]
-      },
-      {
-        "matcher": "Bash",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "bash /home/drb0rk/projects/BizarHarness/.claude/hooks/auto-instinct.sh",
-            "timeout": 15
-          }
-        ]
-      },
-      {
-        "matcher": "Bash(git push*)",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "bash /home/drb0rk/projects/BizarHarness/.claude/hooks/post-merge-audit.sh",
-            "timeout": 60
-          }
-        ]
-      },
-      {
-        "matcher": "EnterWorktree",
-        "hooks": [
-          {
-            "type": "command",
-            "command": "bash /home/drb0rk/projects/BizarHarness/scripts/worktree-setup.sh",
-            "timeout": 60
-          }
-        ]
-      }
+      { "matcher": "Edit|Write|MultiEdit", "hooks": [
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/posttooluse-editwrite.mjs\"", "timeout": 30 }
+      ] },
+      { "matcher": "Bash", "hooks": [
+        { "type": "command", "command": "bash \"$CLAUDE_PROJECT_DIR/.claude/hooks/auto-instinct.sh\"", "timeout": 15 }
+      ] },
+      { "matcher": "Skill", "hooks": [
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/simplify-guard.mjs\"", "timeout": 30 }
+      ] }
     ],
     "UserPromptSubmit": [
-      {
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node /home/drb0rk/projects/BizarHarness/.claude/hooks/worker-suggest.mjs",
-            "timeout": 10
-          }
-        ]
-      },
-      {
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node /home/drb0rk/projects/BizarHarness/.claude/hooks/thinking-route.mjs",
-            "timeout": 10
-          }
-        ]
-      }
+      { "hooks": [{ "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/control-inbox.mjs\"", "timeout": 10 }] },
+      { "hooks": [{ "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/worker-suggest.mjs\"", "timeout": 10 }] },
+      { "hooks": [{ "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/thinking-route.mjs\"", "timeout": 10 }] },
+      { "hooks": [{ "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/telemetry.mjs\"", "timeout": 10 }] }
     ],
     "SessionStart": [
-      {
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node /home/drb0rk/projects/BizarHarness/.claude/hooks/sessionstart-prime.mjs",
-            "timeout": 15
-          }
-        ]
-      }
+      { "hooks": [
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/control-inbox.mjs\"", "timeout": 10 },
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/sessionstart-prime.mjs\"", "timeout": 15 },
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/telemetry.mjs\"", "timeout": 10 }
+      ] }
     ],
     "SessionEnd": [
-      {
-        "hooks": [
-          {
-            "type": "command",
-            "command": "node /home/drb0rk/projects/BizarHarness/.claude/hooks/sessionend-recall.mjs",
-            "timeout": 15
-          },
-          {
-            "type": "command",
-            "command": "node /home/drb0rk/projects/BizarHarness/.claude/hooks/learning-extract.mjs",
-            "timeout": 15
-          }
-        ]
-      }
+      { "hooks": [
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/sessionend-recall.mjs\"", "timeout": 15 },
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/learning-extract.mjs\"", "timeout": 15 }
+      ] }
+    ],
+    "PreCompact": [
+      { "matcher": "*", "hooks": [{ "type": "command", "command": "bash \"$CLAUDE_PROJECT_DIR/.claude/hooks/precompact-priorities.sh\"", "timeout": 15 }] }
+    ],
+    "SubagentStart": [
+      { "hooks": [
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/agent-grounding.mjs\"", "timeout": 15 }
+      ] },
+      { "matcher": "^(linda|karen|carl|qa-reviewer|principal-engineer|debug-specialist)$", "hooks": [
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/advisor-context.mjs\"", "timeout": 15 }
+      ] },
+      { "matcher": "^(brad|carl|pam|brenda|karen|todd|ria)$", "hooks": [
+        { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/.claude/hooks/worktree-bootstrap.mjs\"", "timeout": 30 }
+      ] }
     ]
   }
 }

package/AGENTS.md

@@ -0,0 +1,95 @@
+# AGENTS.md — Bizar Harness
+
+Bizar Harness is a Claude Code-native, guarded-autonomy harness. It ships project and user-level agents, skills, slash commands, hooks, an MCP server, CLI utilities, and verification scripts. It has no embedded web control plane and no Bizar note-vault subsystem; OpenKan can consume the explicit `bizar control` CLI boundary as an optional external UI.
+
+If you are an agent: read this file, read `PROGRESS.md`, inspect `feature_list.json`, then run `make check` before changing code.
+
+## Commands
+
+```sh
+make setup                    # install Claude Code and project dependencies
+make check                    # TypeScript gate
+make test                     # retained SDK, CLI, hook, and harness tests
+make e2e                      # real SDK/MCP/Claude Code integration smoke test
+make check-arch               # architectural and removed-surface checks
+make verify-removed-surfaces  # prove deleted subsystems are absent
+make verify-repo-structure    # prove tracked/package paths are clean
+make clean-check              # debug-artifact/static hygiene gate
+make vcr                      # feature-ledger reality ratio
+make session-start            # lifecycle compatibility target
+make session-end              # lifecycle compatibility target
+```
+
+## Hard constraints
+
+- **MUST** update `PROGRESS.md` before and after each logical code change.
+- **MUST** keep WIP=1 in `feature_list.json`.
+- **MUST** keep one logical operation per commit and keep its docs in the same commit.
+- **MUST** run targeted tests, then `make check`; run `make e2e` for cross-component changes.
+- **MUST** verify evidence before claiming completion.
+- **MUST NOT** commit `console.log`, `debugger`, `.only()`, credentials, generated secrets, or runtime logs.
+- **MUST NOT** use a persistent Claude daemon. Claude Code and the Agent SDK run in-process; background work uses Claude Code's Agent tool.
+- **MUST NOT** bypass human approval for commits, pushes, pull-request mutations, releases, package publication, deployments, production/shared-infrastructure writes, credential changes, public exposure, or irreversible destruction.
+- **MUST NOT** rebase or force-push under the default project policy.
+
+## Execution model
+
+Clear, local, reversible work proceeds autonomously: inspect, edit, test, and iterate without permission handoffs. The project defaults to `acceptEdits`; eligible operators may opt into Claude Code Auto mode. PreToolUse hooks still deny prohibited actions and escalate externally visible or irreversible actions with `permissionDecision: "ask"`.
+
+Every non-empty primary request must use Claude Code's Agent tool to enter the
+Bizar agent pipeline through `office-manager` (`@mike`) before task analysis or
+implementation. Mike routes trivial work to `@brenda` and non-trivial work
+through the phased team. A session already running as a Bizar custom agent
+follows its assigned role and does not recursively dispatch itself.
+
+Every shipped agent has `WebSearch` access. Before proposing, explaining,
+troubleshooting, or implementing behavior from an external API, library,
+framework, CLI, configuration format, or version-sensitive dependency, the
+agent must WebSearch for current official documentation and WebFetch the exact
+relevant page. Guess-and-try integration work is prohibited. When official
+documentation is unavailable or ambiguous, inspect authoritative source code
+and report the evidence gap.
+
+For non-trivial requests, `office-manager` coordinates three ordered phases:
+
+1. Research: `greg` (`research-analyst.md`) plus an implementation-context specialist.
+2. Plan: `planner` drafts; `qa-reviewer` challenges assumptions and test shape.
+3. Implement: engineering agents edit and test; review and verification follow before `commit-staged` asks for the final human commit confirmation.
+
+Use Claude Code's native Agent tool for Bizar routing on every request.
+Additional parallel fan-out remains bounded to scopes that materially improve
+speed, quality, or safety.
+
+## Architecture
+
+- `.claude/agents/` — Claude Code subagent definitions.
+- `config/skills/` — canonical skills; `.claude/skills/` is the verified project mirror.
+- `.claude/commands/` — user-invoked workflows.
+- `.claude/hooks/` + `.claude/settings.json` — safety, routing, lifecycle, telemetry, compaction, reviewer-context, simplify, and HITL gates.
+- `packages/sdk/` — typed autonomy primitives and the nine-tool stdio MCP surface: plans, loops, graph queries, instincts, and decisions.
+- `cli/` — install/provision, audit, validation, backup, cost/claim/task, OpenKan control, sandbox, and repair utilities.
+- `scripts/` + `.harness/` + `templates/` — verification, feature/eval state, audit output, and reusable contracts.
+
+The harness has no embedded browser/server UI layer or local web editor.
+`bizar control` is a machine-readable subprocess boundary for optional OpenKan
+integration; OpenKan owns HTTP, WebSocket, and presentation concerns. Session
+handoff, control inbox, and learning logs are bounded operational records for
+autonomy; they are not a general note vault, semantic search service, or
+knowledge-base API.
+
+## State and evidence
+
+- `PROGRESS.md` — current objective, evidence, next actions, blockers.
+- `feature_list.json` — WIP=1 feature state.
+- `DECISIONS.md` and `docs/decisions/` — current architecture decisions.
+- `.harness/evals/` — feature evaluation records.
+- `~/.config/bizar/telemetry/` — local correlation and rejected-action feedback.
+- `.bizar/session-state.json` — bounded SessionEnd→SessionStart handoff.
+
+## Definition of done
+
+1. Behavior is implemented with a regression test.
+2. Documentation and feature state describe the actual code.
+3. `make verify-removed-surfaces`, `make verify-repo-structure`, `make check-arch`, `make test`, `make e2e`, `make clean-check`, and `make check` pass as applicable.
+4. `PROGRESS.md` records fresh evidence and no required work remains.
+5. `/simplify` reviews the staged diff before the approval-gated commit.