crewhaus 0.1.4 → 0.1.6

package/dist/index.js

@@ -0,0 +1,2362 @@
+#!/usr/bin/env bun
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { basename, dirname, join, relative, resolve } from "node:path";
+import { SpecParseError, compile, lower } from "@crewhaus/compiler";
+import { buildContextBundle, discoverRoots } from "@crewhaus/context-bundle";
+import { CrewhausError } from "@crewhaus/errors";
+import { loadDataset } from "@crewhaus/eval-dataset";
+import { parseGradersConfig } from "@crewhaus/eval-grader";
+import { optimizeSpec } from "@crewhaus/eval-optimizer-orchestrator";
+import { diffReports, loadRun, renderReport } from "@crewhaus/eval-report";
+import { runEval as runEvalLib } from "@crewhaus/eval-runner";
+import { loadHooks } from "@crewhaus/hooks-engine";
+import { ArgParseError, parseArgs, } from "@crewhaus/infra-utils";
+import { createLogger } from "@crewhaus/logging";
+import { McpHost } from "@crewhaus/mcp-host";
+import { BUILTIN_DEFAULT_RULES, PermissionConfigError, parsePermissionsConfig, tagRules, } from "@crewhaus/permission-engine";
+import { runChatLoop } from "@crewhaus/runtime-core";
+import { createSessionStore } from "@crewhaus/session-store";
+import { createSkillTool, discoverSkills } from "@crewhaus/skills-registry";
+import { loadCommands } from "@crewhaus/slash-commands";
+import { parseSpec } from "@crewhaus/spec";
+import { spawnSubAgent } from "@crewhaus/sub-agent-spawner";
+import { ToolCatalog } from "@crewhaus/tool-catalog";
+import { registerMcpServer } from "@crewhaus/tool-mcp";
+import { createTaskTool } from "@crewhaus/tool-task";
+import { TraceEventBus } from "@crewhaus/trace-event-bus";
+// Model-aware doctor credential checks (provider parsed from the cwd spec's
+// agent.model via the model-router grammar), in a side-effect-free module so
+// it is unit-testable (this entry file runs an argv switch on import).
+import { buildCredentialChecks, extractSpecModel } from "./doctor-checks";
+// FR-006 — Pillar 3 sink-side egress-matcher selection (the substring/semantic
+// selector lowered to ir.security.egressMatcher). Side-effect-free + lazily
+// imports the optional semantic package only when "semantic" is selected, so
+// the default path pulls in no embedding dependency.
+import { DEFAULT_EGRESS_EMBEDDER_MODEL, InvalidEgressMatcherChoiceError, createEgressMatcher, resolveEgressMatcherChoice, } from "./egress-matcher";
+// FR-004 — Pillar 3 intent-gate judge + durable audit-sink resolution, in a
+// side-effect-free module so it is unit-testable (this entry file runs an
+// argv switch on import).
+import { InvalidJudgeChoiceError, createJustificationJudge, openJustificationAuditSink, resolveJudgeChoice, } from "./justification-gate";
+// FR-002 — Pillar 3 sink-side scope gate, shared by `compile --strict` and
+// `doctor --philosophy-alignment`. Kept in a side-effect-free module so it is
+// unit-testable (this entry file runs an argv switch on import).
+import { auditSpecToolNames, auditToolScopes, collectToolNames } from "./scope-audit";
+/**
+ * crewhaus — slice-scope CLI.
+ * Subcommands:
+ *   compile <spec.yaml> -o <out-dir>     parse → IR → emit bundle to disk
+ *   compile <spec.yaml> --emit-ir        parse → IR → print IR JSON (no codegen)
+ *   run <spec.yaml> [--model <model>]    compile in-memory and execute the agent
+ *   init [name]                          scaffold a new crewhaus.yaml
+ *   doctor                               check environment health
+ *
+ * Future (per catalog F4 spec-cli): deploy, eval, watch.
+ *
+ * User-facing messages (status, errors) go directly to stdout/stderr for clean
+ * UX. The logger is for diagnostic events visible only when CREWHAUS_LOG_LEVEL
+ * is set to debug (or CREWHAUS_LOG=json for machine-readable traces).
+ */
+const logger = createLogger({ bindings: { app: "crewhaus" } });
+const COMPILE_SCHEMA = {
+    flags: [
+        { name: "out", short: "o", takesValue: true },
+        { name: "emit-ir", takesValue: false },
+        // FR-002 — the strict scope gate (fail the build when an outward-reaching
+        // tool is left at a non-"external" scope) now runs by DEFAULT on every
+        // compile. `--strict` is retained as an accepted no-op so existing
+        // invocations and CI scripts keep working; the gate fires with or without
+        // it (see whitepaper §6).
+        { name: "strict", takesValue: false },
+        // FR-002 — explicit opt-out for users who knowingly bypass the gate (e.g.
+        // an outward sink whose external scope is verified out of band). Either
+        // spelling disables the gate; both are accepted so neither reach is a
+        // surprise. Without one of these flags an unmarked outward/io-capable tool
+        // FAILS the compile.
+        { name: "allow-unmarked-sinks", takesValue: false },
+        { name: "no-strict-scope", takesValue: false },
+        { name: "help", short: "h" },
+    ],
+};
+const RUN_SCHEMA = {
+    flags: [
+        { name: "model", takesValue: true },
+        { name: "permission-mode", takesValue: true },
+        { name: "resume", takesValue: true },
+        { name: "continue", takesValue: false },
+        { name: "prompt", takesValue: true },
+        // FR-004 — select the Pillar 3 intent-gate judge for this run.
+        // rule-based (default) | claude. Overrides the spec's
+        // security.justification.judge when supplied.
+        { name: "justification-judge", takesValue: true },
+        // FR-004 — disable the durable `permission_justification_evaluated`
+        // audit-log record (on by default for `run`). The ephemeral trace-bus
+        // event is unaffected; this only skips opening `.crewhaus/audit`.
+        { name: "no-justification-audit", takesValue: false },
+        // FR-006 — select the Pillar 3 sink-side egress matcher for this run.
+        // substring (default) | semantic. Overrides the spec's
+        // security.egressMatcher when supplied.
+        { name: "egress-matcher", takesValue: true },
+        // FR-006 — embedder model for the "semantic" egress matcher (the
+        // @crewhaus/embedder prefix grammar, e.g. openai/text-embedding-3-small,
+        // mock/deterministic). Flag > CREWHAUS_EGRESS_EMBEDDER env > default.
+        { name: "egress-embedder", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const SESSION_ID_REGEX = /^sess_[0-9a-f]{16}$/;
+const VALID_PERMISSION_MODES = ["default", "plan", "auto", "bypass"];
+function isValidPermissionMode(s) {
+    return VALID_PERMISSION_MODES.includes(s);
+}
+const INIT_SCHEMA = {
+    flags: [{ name: "help", short: "h" }],
+};
+const DOCTOR_SCHEMA = {
+    flags: [
+        { name: "philosophy-alignment", takesValue: false },
+        // Process-liveness only — exit 0 fast, no credential/spec checks. The
+        // probe target for container HEALTHCHECKs and k8s exec probes.
+        { name: "liveness", takesValue: false },
+        { name: "help", short: "h" },
+    ],
+};
+const CONTEXT_SCHEMA = {
+    flags: [
+        { name: "bundle", takesValue: false },
+        { name: "out", short: "o", takesValue: true },
+        { name: "factory-root", takesValue: true },
+        { name: "docs-root", takesValue: true },
+        { name: "demos-root", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const OPTIMIZE_SCHEMA = {
+    flags: [
+        { name: "dataset", takesValue: true },
+        { name: "graders", takesValue: true },
+        { name: "mutator", takesValue: true },
+        { name: "iterations", takesValue: true },
+        { name: "seed", takesValue: true },
+        { name: "improvement-threshold", takesValue: true },
+        // FR-003 — dollar ceiling for model-driven runs (composes with --iterations).
+        { name: "budget-usd", takesValue: true },
+        { name: "write-back", takesValue: false },
+        { name: "out", short: "o", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const EVAL_SCHEMA = {
+    flags: [
+        { name: "dataset", takesValue: true },
+        { name: "graders", takesValue: true },
+        { name: "judge-model", takesValue: true },
+        { name: "concurrency", takesValue: true },
+        { name: "seed", takesValue: true },
+        { name: "out", short: "o", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const EVAL_REPORT_SCHEMA = {
+    flags: [
+        { name: "out", short: "o", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const COST_SUMMARY_SCHEMA = {
+    flags: [
+        { name: "session", takesValue: true },
+        { name: "tenant", takesValue: true },
+        { name: "format", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const SANDBOX_SCHEMA = {
+    flags: [
+        { name: "probe", takesValue: false },
+        { name: "format", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const COMPLIANCE_SCHEMA = {
+    flags: [
+        { name: "framework", takesValue: true },
+        { name: "control", takesValue: true },
+        { name: "period", takesValue: true },
+        { name: "audit-dir", takesValue: true },
+        { name: "out-dir", takesValue: true },
+        { name: "signing-key-env", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const SECRETS_SCHEMA = {
+    flags: [
+        { name: "backend", takesValue: true },
+        { name: "root-dir", takesValue: true },
+        { name: "value", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const SPEC_SCHEMA = {
+    flags: [
+        { name: "root-dir", takesValue: true },
+        { name: "tenant", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const DEPLOY_SCHEMA = {
+    flags: [
+        { name: "root-dir", takesValue: true },
+        { name: "tenant", takesValue: true },
+        { name: "actor", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const MIGRATE_SCHEMA = {
+    flags: [
+        { name: "root-dir", takesValue: true },
+        { name: "from", takesValue: true },
+        { name: "to", takesValue: true },
+        { name: "dry-run" },
+        { name: "help", short: "h" },
+    ],
+};
+const BUILD_IMAGE_SCHEMA = {
+    flags: [
+        { name: "tag", takesValue: true },
+        { name: "platform", takesValue: true },
+        { name: "push" },
+        { name: "help", short: "h" },
+    ],
+};
+const CLOUD_SCHEMA = {
+    flags: [
+        { name: "provider", takesValue: true },
+        { name: "region", takesValue: true },
+        { name: "tier", takesValue: true },
+        { name: "image-tag", takesValue: true },
+        { name: "working-dir", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+const FEDERATION_SCHEMA = {
+    flags: [
+        { name: "srv-domain", takesValue: true },
+        { name: "format", takesValue: true },
+        { name: "help", short: "h" },
+    ],
+};
+function usageText() {
+    return [
+        "usage: crewhaus <subcommand> [args]",
+        "",
+        "subcommands:",
+        "  compile <spec.yaml> -o <out-dir>     compile a spec to a runnable bundle",
+        "                                       (fails if an outward tool is left non-external — FR-002)",
+        "                  [--allow-unmarked-sinks]  opt out of the external-sink scope gate",
+        "  compile <spec.yaml> --emit-ir        print the lowered IR as JSON (debug)",
+        "  run <spec.yaml> [--model <model>]    compile in-memory and execute the agent",
+        "                  [--resume <id>]      resume a specific session (cli targets only)",
+        "                  [--continue]         resume the most-recent session (cli targets only)",
+        "                  [--prompt <text>]    initial user prompt (browser targets; defaults to stdin)",
+        "                  [--justification-judge rule-based|claude]  Pillar 3 intent-gate judge (FR-004)",
+        "                  [--egress-matcher substring|semantic]  Pillar 3 sink-side matcher (FR-006)",
+        "                  [--egress-embedder <model>]  embedder for --egress-matcher semantic",
+        "  eval <spec.yaml> --dataset <data>    run the agent against a dataset and grade",
+        "       --graders <graders.yaml>       (deterministic graders + LLM-as-judge)",
+        "       [--judge-model <model>] [--concurrency N] [--seed N] -o <out-dir>",
+        "  eval-report diff <prev> <new>        compare two eval runs and emit a diff report",
+        "       [-o <out-dir>]",
+        "  optimize <spec.yaml> --dataset <data> --graders <graders.yaml>",
+        "       [--mutator rule-based|claude] [--iterations N] [--seed N]",
+        "       [--budget-usd N]                     stop a model-driven run before it exceeds $N (FR-003)",
+        "       [--write-back] [-o <out-dir>]        active eval-driven optimization (Pillar 2)",
+        "  init [name]                          scaffold a new crewhaus.yaml",
+        "  doctor                               check environment health",
+        "  context --bundle [-o <file>]         emit a single-markdown orientation manifest",
+        "       [--factory-root <p>] [--docs-root <p>] [--demos-root <p>]",
+        "  cost-summary --session <id>          summarize cost_accrual events for a session",
+        "  secrets doctor                       list known secrets via the configured backend",
+        "  secrets rotate <name> [--value V]    rotate a named secret (file backend)",
+        "  spec put|list|get|pin|alias ...      versioned spec storage (Section 28 spec-registry)",
+        "  deploy promote|rollback ...          re-pin a spec for an environment (Section 28)",
+        "  migrate-all --from N --to N          batch-migrate every spec in the registry",
+        "  build-image <target> --tag <tag>     build the docker image for a target shape (Section 32)",
+        "       [--platform <p>] [--push]",
+        "  cloud deploy --provider <p>          deploy a managed CrewHaus cluster (Section 32)",
+        "       --region <r> [--tier <t>] [--image-tag <tag>]",
+        "  cloud teardown --provider <p>        tear down a managed cluster",
+        "       --region <r>",
+        "  federation discover <deployment>     resolve a federated peer's endpoint + cert fingerprint (Section 34)",
+        "       [--srv-domain <d>] [--format json|yaml]",
+        "  sandbox doctor [--probe]             list registered sandbox images + healthcheck status (Section 36)",
+        "       [--format json|table]",
+        "  compliance evidence                  collect SOC 2 / ISO 27001 / HIPAA evidence (Section 39)",
+        "       --framework <id> [--control <id>] --period <p> [--audit-dir <d>]",
+        "       [--out-dir <d>] [--signing-key-env <ENV>]",
+        "  version                              print the CLI version (also: --version, -v)",
+        "",
+    ].join("\n");
+}
+/** No subcommand given (or a parse error) — usage to stderr, exit 1. */
+function usage() {
+    process.stderr.write(usageText());
+    process.exit(1);
+}
+/**
+ * Explicit `-h`/`--help` at the top level — usage to stdout, exit 0. Help was
+ * requested, so it is not an error; this matches every subcommand's own
+ * `--help` (stdout, exit 0) and lets `crewhaus --help` work in `set -e` health
+ * checks instead of looking like a broken CLI.
+ */
+function help() {
+    process.stdout.write(usageText());
+    process.exit(0);
+}
+function die(message) {
+    process.stderr.write(`crewhaus: ${message}\n`);
+    process.exit(1);
+}
+function printVersion() {
+    if (typeof CREWHAUS_EMBEDDED_VERSION === "string") {
+        process.stdout.write(`${CREWHAUS_EMBEDDED_VERSION}\n`);
+        return;
+    }
+    // The package ships src/ directly (bin → src/index.ts) and tsc -b also
+    // emits dist/, so resolve package.json relative to this module — one level
+    // up lands on apps/cli/package.json from either tree, and on
+    // node_modules/@crewhaus/cli/package.json when installed.
+    let version;
+    try {
+        version = JSON.parse(readFileSync(new URL("../package.json", import.meta.url), "utf-8")).version;
+    }
+    catch {
+        die("could not locate package.json to determine the version");
+    }
+    process.stdout.write(`${version}\n`);
+}
+function parseFor(rest, schema) {
+    try {
+        return parseArgs(rest, schema);
+    }
+    catch (err) {
+        if (err instanceof ArgParseError)
+            die(err.message);
+        throw err;
+    }
+}
+async function runCompile(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus compile <spec.yaml> [-o <out-dir>] [--emit-ir]\n" +
+            "                        [--allow-unmarked-sinks]\n" +
+            "  --emit-ir  Skip code emission; print the lowered IR as JSON to\n" +
+            "             stdout (or to <out-dir>/ir.json when -o is set).\n" +
+            "\n" +
+            "  FR-002 — the strict scope gate runs by DEFAULT: the build fails\n" +
+            "  (exit 1) if any I/O-capable tool the spec uses is left at a\n" +
+            '  non-"external" scope. It flags:\n' +
+            "    (a) a resolvable built-in whose declared io-capability or\n" +
+            "        outward name disagrees with its scope;\n" +
+            "    (b) an outward-by-name sink — any mcp__* tool or known outward\n" +
+            '        built-in — that cannot be resolved to a scope:"external"\n' +
+            "        tool offline (its egress scope is unverifiable at compile\n" +
+            "        time, so the gate refuses it).\n" +
+            "  A non-outward custom tool whose name carries no signal is left to\n" +
+            "  the live `doctor --philosophy-alignment` audit / runtime egress\n" +
+            "  fabric.\n" +
+            "\n" +
+            "  --allow-unmarked-sinks   Opt out of the gate for this compile (alias:\n" +
+            "             --no-strict-scope). Use only when you knowingly bypass it.\n" +
+            "  --strict   Accepted no-op — the gate is already on by default.\n");
+        return;
+    }
+    const specPath = args.positional[0];
+    const outDir = args.flags["out"];
+    const emitIr = args.flags["emit-ir"] === true;
+    // FR-002 — the strict scope gate is DEFAULT-ON. It runs unless the user
+    // explicitly opts out with --allow-unmarked-sinks (or its alias
+    // --no-strict-scope). `--strict` is now a no-op kept for back-compat.
+    const allowUnmarkedSinks = args.flags["allow-unmarked-sinks"] === true || args.flags["no-strict-scope"] === true;
+    const strict = !allowUnmarkedSinks;
+    if (typeof specPath !== "string")
+        die("missing <spec.yaml>");
+    if (!emitIr && typeof outDir !== "string")
+        die("missing -o <out-dir>");
+    const absSpec = resolve(specPath);
+    logger.debug("compile.start", { spec: absSpec, out: outDir, emitIr, strict, allowUnmarkedSinks });
+    let yamlText;
+    try {
+        yamlText = readFileSync(absSpec, "utf-8");
+    }
+    catch (err) {
+        die(`could not read ${absSpec}: ${err.message}`);
+    }
+    // FR-006 — `security.egressMatcher: semantic` is now emitted into the
+    // standalone cli bundle by `@crewhaus/target-cli` (it constructs
+    // `@crewhaus/egress-matcher-semantic` with an injected embedder and threads
+    // it into the bundle's `runChatLoop({ egressMatcher })`), so a compiled
+    // artifact honours the selection WITHOUT the `run` path. No compile-time
+    // warning is needed anymore — emission replaced the warn-only shim.
+    // FR-002 — strict scope gate, now DEFAULT-ON. Lower the spec (pure), resolve
+    // every referenced built-in tool name to its RegisteredTool, and audit scopes
+    // BEFORE any emission. Runs for both --emit-ir and bundle modes so the gate
+    // can't be sidestepped by mode choice. `lower()` only carries tool NAMES, not
+    // scope, so resolution via loadToolMap() is what surfaces the real
+    // RegisteredTool.scope. Skipped only when the user opts out explicitly with
+    // --allow-unmarked-sinks (strict === false).
+    if (strict) {
+        await runStrictScopeGate(yamlText);
+    }
+    if (emitIr) {
+        let ir;
+        try {
+            ir = lower(parseSpec(yamlText));
+        }
+        catch (err) {
+            // parseSpec throws SpecParseError; lower() can throw CompilerError (e.g. a
+            // malformed credential env-ref). Both extend CrewhausError — route the
+            // family through die() for a clean one-liner instead of a raw stack trace.
+            if (err instanceof CrewhausError)
+                die(err.message);
+            throw err;
+        }
+        const json = `${JSON.stringify(ir, null, 2)}\n`;
+        if (typeof outDir === "string") {
+            const absOut = resolve(outDir);
+            mkdirSync(absOut, { recursive: true });
+            const irPath = join(absOut, "ir.json");
+            writeFileSync(irPath, json);
+            process.stdout.write(`wrote ${irPath}\n`);
+        }
+        else {
+            process.stdout.write(json);
+        }
+        logger.debug("compile.emit-ir.success", { out: outDir ?? "stdout" });
+        return;
+    }
+    const absOut = resolve(outDir);
+    let bundle;
+    try {
+        bundle = compile(yamlText);
+    }
+    catch (err) {
+        // compile() runs parse → lower → emit. parseSpec throws SpecParseError;
+        // each target emitter throws its own TargetEmitError (e.g. an unresolvable
+        // tool name once the default-on scope gate has been bypassed with
+        // --allow-unmarked-sinks). Both — like every structured failure in this
+        // pipeline — extend CrewhausError, so route the whole family through die()
+        // for a clean one-line error + exit 1 instead of letting the emitter crash
+        // escape as an uncaught stack trace. A non-CrewhausError (a genuine bug)
+        // still propagates with its full stack for debugging.
+        if (err instanceof CrewhausError) {
+            die(err.message);
+        }
+        throw err;
+    }
+    mkdirSync(absOut, { recursive: true });
+    for (const file of bundle.files) {
+        const fullPath = join(absOut, file.path);
+        mkdirSync(dirname(fullPath), { recursive: true });
+        writeFileSync(fullPath, file.content);
+        process.stdout.write(`wrote ${fullPath}\n`);
+    }
+    process.stdout.write(`compiled bundle (${bundle.files.length} file(s)) → ${absOut}\n`);
+    logger.debug("compile.success", { files: bundle.files.length, out: absOut });
+}
+/**
+ * FR-002 — the `compile --strict` enforcement body. Lowers the spec, resolves
+ * the tool names it references to their RegisteredTools, and runs the shared
+ * `auditToolScopes`. On any finding, writes a `[strict]` diagnostic per
+ * tool to stderr and exits 1 BEFORE emitting. Shares the exact audit with
+ * `crewhaus doctor --philosophy-alignment`.
+ */
+async function runStrictScopeGate(yamlText) {
+    let ir;
+    try {
+        ir = lower(parseSpec(yamlText));
+    }
+    catch (err) {
+        // SpecParseError (parse) and CompilerError (lower, e.g. a malformed
+        // credential env-ref) both extend CrewhausError — render as a clean die()
+        // one-liner rather than an uncaught stack trace.
+        if (err instanceof CrewhausError)
+            die(err.message);
+        throw err;
+    }
+    const toolNames = collectToolNames(ir);
+    if (toolNames.length === 0)
+        return;
+    const toolMap = await loadToolMap();
+    // `loadToolMap` is keyed by the camelCase spec key (`webSearch`), which is how
+    // a top-level `tools:` list names its tools. But a sub-agent `tools:` list
+    // names tools by their REGISTERED name (PascalCase, e.g. `WebSearch`) — the
+    // runtime child-catalog filter (`buildChildCatalog` in tool-task) matches on
+    // `RegisteredTool.name`, not the spec key. `collectToolNames` gathers BOTH
+    // forms from the IR, so also index the map by registered name; otherwise an
+    // outward sub-agent sink like `WebSearch` fails to resolve and is wrongly
+    // flagged as an unverifiable external sink (it resolves to a built-in whose
+    // scope IS statically "external"). A genuinely unknown name (`mcp__*`, a typo)
+    // still resolves to undefined under both keys and is still gated.
+    const byRegisteredName = {};
+    for (const tool of Object.values(toolMap))
+        byRegisteredName[tool.name] = tool;
+    // Shared, name-independent audit (see scope-audit.ts):
+    //   - resolvable built-ins are checked by capability/outward-name vs scope;
+    //   - an outward-by-name sink we CANNOT resolve to a scope:"external" tool
+    //     offline (a custom outward name, or any `mcp__*` dynamic sink) is a
+    //     finding, because --strict refuses to emit a bundle that reaches a
+    //     sink whose external scope it cannot verify at compile time.
+    const findings = auditSpecToolNames(toolNames, (name) => toolMap[name] ?? byRegisteredName[name]);
+    if (findings.length > 0) {
+        for (const f of findings) {
+            process.stderr.write(`crewhaus: [strict] tool "${f.toolName}" ${f.reason}\n`);
+        }
+        process.stderr.write(`crewhaus: [strict] ${findings.length} scope finding(s) — refusing to emit. Set scope: "external" on each tool above, or pass --allow-unmarked-sinks to bypass the gate.\n`);
+        process.exit(1);
+    }
+}
+function runInit(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus init [name]\n");
+        return;
+    }
+    const nameArg = args.positional[0];
+    const targetDir = typeof nameArg === "string" ? resolve(nameArg) : process.cwd();
+    const specName = typeof nameArg === "string" ? nameArg : basename(targetDir);
+    const targetFile = join(targetDir, "crewhaus.yaml");
+    if (existsSync(targetFile)) {
+        die(`${targetFile} already exists — refusing to overwrite`);
+    }
+    mkdirSync(targetDir, { recursive: true });
+    const yamlText = `name: ${specName}
+target: cli
+agent:
+  model: claude-opus-4-7
+  instructions: |
+    You are a helpful assistant. Replace these instructions with your
+    agent's actual behavior, persona, and constraints.
+`;
+    writeFileSync(targetFile, yamlText);
+    process.stdout.write(`wrote ${targetFile}\n`);
+    // The runtime resolves the spec and the `.crewhaus/` session store from
+    // the current working directory, so guide the user to run from inside
+    // the harness directory (where crewhaus.yaml lives), not from here.
+    const rel = relative(process.cwd(), targetDir);
+    const cd = rel === "" ? "" : `cd ${rel} && `;
+    process.stdout.write(`next: ${cd}crewhaus run crewhaus.yaml\n`);
+    logger.debug("init.success", { target: targetFile });
+}
+/**
+ * Built-in tool name → RegisteredTool, populated lazily so that subcommands
+ * which don't need tools (init, doctor) don't pay the import cost. Mirror of
+ * `BUILTIN_TOOL_MAP` in packages/target-cli/src/index.ts — keep them in sync.
+ */
+async function loadToolMap() {
+    const [fs, bash, todo, web, image, fetchPkg, imageGen, docIngest, codegraph] = await Promise.all([
+        import("@crewhaus/tool-fs"),
+        import("@crewhaus/tool-bash"),
+        import("@crewhaus/tool-todo"),
+        import("@crewhaus/tool-web"),
+        import("@crewhaus/tool-image"),
+        import("@crewhaus/tool-fetch"),
+        import("@crewhaus/tool-image-generation"),
+        import("@crewhaus/tool-document-ingest"),
+        import("@crewhaus/tool-codegraph"),
+    ]);
+    return {
+        read: fs.read,
+        write: fs.write,
+        edit: fs.edit,
+        glob: fs.glob,
+        grep: fs.grep,
+        bash: bash.bash,
+        todoWrite: todo.todoWrite,
+        webFetch: web.webFetch,
+        webSearch: web.webSearch,
+        readImage: image.readImage,
+        fetch: fetchPkg.fetch,
+        imageGenerate: imageGen.imageGenerate,
+        ingestDocument: docIngest.ingestDocument,
+        // Pillar 2 — AST-aware code intelligence (recipe 54).
+        codegraphSearch: codegraph.codegraphSearch,
+        codegraphCallers: codegraph.codegraphCallers,
+        codegraphCallees: codegraph.codegraphCallees,
+        codegraphImpact: codegraph.codegraphImpact,
+    };
+}
+/**
+ * Section 14 — apply per-tool config from the IR's `toolConfigs` map by
+ * calling each tool's registration function. Mirror of the codegen-emitted
+ * init calls in target-cli/target-channel-bot. Keep in sync.
+ */
+async function applyToolConfigs(toolNames, toolConfigs) {
+    const used = new Set(toolNames);
+    if (used.has("fetch") && toolConfigs["fetch"] !== undefined) {
+        const { registerFetchConfig } = await import("@crewhaus/tool-fetch");
+        registerFetchConfig(toolConfigs["fetch"]);
+    }
+    if (used.has("webFetch") && toolConfigs["webFetch"] !== undefined) {
+        const { registerWebFetchConfig } = await import("@crewhaus/tool-web");
+        registerWebFetchConfig(toolConfigs["webFetch"]);
+    }
+}
+/**
+ * Build the layered permission rule set from the IR (yaml source) and an
+ * optional `.crewhaus/settings.json` (settings source). The flag and hooks
+ * sources are placeholders for future sections (no rules yet, just modes).
+ *
+ * All non-flag config goes through `parsePermissionsConfig` which rejects
+ * `mode: bypass` (defense in depth on top of the spec parser's check).
+ */
+function buildRuleSet(yamlRules, cwd) {
+    let settings = [];
+    const settingsPath = join(cwd, ".crewhaus", "settings.json");
+    if (existsSync(settingsPath)) {
+        let raw;
+        try {
+            raw = JSON.parse(readFileSync(settingsPath, "utf-8"));
+        }
+        catch (err) {
+            die(`failed to parse ${settingsPath}: ${err.message}`);
+        }
+        // Section 11 introduced top-level keys (`hooks`) into settings.json.
+        // Only parse the `permissions` sub-object — never the bare root —
+        // so hooks/skills/slash-command keys don't trip the strict permission
+        // validator. If the file has no `permissions` block, treat as empty.
+        const root = raw.permissions;
+        if (root !== undefined) {
+            try {
+                const parsed = parsePermissionsConfig(root, "settings");
+                settings = tagRules(parsed.rules, "settings");
+            }
+            catch (err) {
+                if (err instanceof PermissionConfigError)
+                    die(err.message);
+                throw err;
+            }
+        }
+    }
+    return {
+        flag: [],
+        settings,
+        yaml: tagRules(yamlRules, "yaml"),
+        hooks: [],
+        builtin: BUILTIN_DEFAULT_RULES,
+    };
+}
+/**
+ * FR-004 — resolve the Pillar 3 intent-gate judge for a `run`. Precedence:
+ * `--justification-judge` flag > spec `security.justification.judge` >
+ * `"rule-based"`. Returns `undefined` for the rule-based path so the
+ * caller omits `justificationJudge` from `runChatLoop` and runtime-core
+ * falls back to `ruleBasedJustificationJudge` (the documented default for
+ * tests/offline runs). For `"claude"` it lazily imports the adapter +
+ * `@crewhaus/justification-judge-claude` (matching the run path's existing
+ * lazy-import style) so the model-backed judge code only loads when
+ * actually selected. An unknown value `die()`s.
+ */
+async function resolveJustificationJudge(args, securityJustification) {
+    const flag = args.flags["justification-judge"];
+    const flagValue = typeof flag === "string" ? flag : undefined;
+    let choice;
+    try {
+        choice = resolveJudgeChoice(flagValue, securityJustification);
+    }
+    catch (err) {
+        if (err instanceof InvalidJudgeChoiceError)
+            die(err.message);
+        throw err;
+    }
+    return createJustificationJudge(choice, securityJustification?.model);
+}
+/**
+ * FR-006 — resolve the Pillar 3 sink-side egress matcher for a `run`.
+ * Precedence: `--egress-matcher` flag > spec `security.egressMatcher` (lowered
+ * to `ir.security.egressMatcher`) > `"substring"`. Returns `undefined` for the
+ * substring path so the caller omits `egressMatcher` from `runChatLoop` and
+ * runtime-core stays on the built-in `substringMatcher` (the default egress
+ * path then pulls in no embedding dependency). For `"semantic"` it lazily
+ * imports `@crewhaus/embedder` + `@crewhaus/egress-matcher-semantic` and
+ * constructs a `SemanticEgressMatcher`; the embedder model comes from
+ * `--egress-embedder` > `CREWHAUS_EGRESS_EMBEDDER` env >
+ * `DEFAULT_EGRESS_EMBEDDER_MODEL`. An unknown matcher value `die()`s.
+ */
+async function resolveEgressMatcher(args, irEgressMatcher) {
+    const flag = args.flags["egress-matcher"];
+    const flagValue = typeof flag === "string" ? flag : undefined;
+    let choice;
+    try {
+        choice = resolveEgressMatcherChoice(flagValue, irEgressMatcher);
+    }
+    catch (err) {
+        if (err instanceof InvalidEgressMatcherChoiceError)
+            die(err.message);
+        throw err;
+    }
+    const embedderFlag = args.flags["egress-embedder"];
+    const embedderModel = (typeof embedderFlag === "string" ? embedderFlag : undefined) ??
+        process.env["CREWHAUS_EGRESS_EMBEDDER"] ??
+        DEFAULT_EGRESS_EMBEDDER_MODEL;
+    return createEgressMatcher(choice, { embedderModel });
+}
+async function runRun(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus run <spec.yaml> [--model <model>] [--permission-mode <default|plan|auto|bypass>] [--resume <sessionId> | --continue] [--prompt <text>] [--justification-judge rule-based|claude] [--egress-matcher substring|semantic] [--egress-embedder <model>]\n" +
+            "  --model accepts the full router grammar: claude-* (Anthropic), openai/<m>, gemini/<m>, bedrock/<id> (geo prefixes tolerated), local/<m>@<url>\n");
+        return;
+    }
+    const specPath = args.positional[0];
+    if (typeof specPath !== "string")
+        die("missing <spec.yaml>");
+    const absSpec = resolve(specPath);
+    logger.debug("run.start", { spec: absSpec });
+    let yamlText;
+    try {
+        yamlText = readFileSync(absSpec, "utf-8");
+    }
+    catch (err) {
+        die(`could not read ${absSpec}: ${err.message}`);
+    }
+    let ir;
+    try {
+        ir = lower(parseSpec(yamlText));
+    }
+    catch (err) {
+        if (err instanceof SpecParseError)
+            die(err.message);
+        throw err;
+    }
+    if (ir.target === "cli")
+        return runRunCli(args, ir, specPath);
+    if (ir.target === "browser")
+        return runRunBrowser(args, ir);
+    die(`crewhaus run supports target: cli or browser (got "${ir.target}"). Other target shapes are compile-only — see PACKAGES.md.`);
+}
+/**
+ * cli-target run path. Multi-turn interactive REPL, session-store backed,
+ * loads hooks/skills/slash-commands/sub-agents from the user's workspace,
+ * and wires every spec-declared MCP server.
+ */
+async function runRunCli(args, ir, specPath) {
+    const resumeFlag = args.flags["resume"];
+    const continueFlag = args.flags["continue"] === true;
+    if (typeof resumeFlag === "string" && continueFlag) {
+        die("--resume and --continue are mutually exclusive");
+    }
+    let resumeId;
+    if (typeof resumeFlag === "string") {
+        if (!SESSION_ID_REGEX.test(resumeFlag)) {
+            die(`invalid --resume sessionId "${resumeFlag}" — expected sess_<16 hex>`);
+        }
+        resumeId = resumeFlag;
+    }
+    // --continue resolves to the most-recently-updated session for this
+    // spec's name, scoped to the current working directory's session
+    // store. session-store's list() returns sessions sorted by updatedAt
+    // descending, with a TTL-based eviction sweep as a side effect.
+    if (continueFlag) {
+        const store = createSessionStore();
+        const sessions = await store.list();
+        const match = sessions.find((s) => s.name === ir.name);
+        if (match === undefined) {
+            const absSpec = resolve(specPath);
+            die(`no prior session for spec "${ir.name}" in ${process.cwd()}/.crewhaus/sessions/. Sessions are stored under the directory you run from — start one from the harness directory with: cd ${dirname(absSpec)} && crewhaus run ${basename(absSpec)}`);
+        }
+        resumeId = match.id;
+        process.stdout.write(`[continue] resuming session ${match.id} (last updated ${match.updatedAt})\n`);
+    }
+    let tools = [];
+    if (ir.tools.length > 0) {
+        // Section 14 — apply per-tool config (e.g. registerFetchConfig) before
+        // loading the tools so first-call execution sees the registered config.
+        await applyToolConfigs(ir.tools, ir.toolConfigs);
+        const toolMap = await loadToolMap();
+        tools = ir.tools.map((name) => {
+            const tool = toolMap[name];
+            if (!tool) {
+                const known = Object.keys(toolMap).sort().join(", ");
+                die(`unknown tool "${name}" — known tools: ${known}`);
+            }
+            return tool;
+        });
+    }
+    // Section 9 — connect to declared MCP servers and register their remote
+    // tools alongside the built-ins. Mirror of the codegen path in
+    // @crewhaus/target-cli (renderMcpServers); keep them in sync.
+    let mcpHost;
+    if (Object.keys(ir.mcp_servers).length > 0) {
+        const host = new McpHost({ logger });
+        mcpHost = host;
+        for (const [name, cfg] of Object.entries(ir.mcp_servers)) {
+            host.addServer(name, cfg);
+        }
+        const tempCatalog = new ToolCatalog();
+        for (const t of tools)
+            tempCatalog.register(t);
+        await Promise.all(Object.keys(ir.mcp_servers).map((name) => registerMcpServer(host, name, tempCatalog, {
+            onRegister: ({ fullName }) => process.stdout.write(`[mcp] registered ${fullName}\n`),
+        })));
+        tools = tempCatalog.list().slice();
+    }
+    const modelOverride = args.flags["model"];
+    const model = typeof modelOverride === "string" ? modelOverride : ir.agent.model;
+    // Permission mode resolution: CLI flag > spec > "default".
+    // bypass is reachable only via the flag (the spec parser has already
+    // rejected `mode: bypass`).
+    const flagMode = args.flags["permission-mode"];
+    let permissionMode;
+    if (typeof flagMode === "string") {
+        if (!isValidPermissionMode(flagMode)) {
+            die(`invalid --permission-mode "${flagMode}" — allowed: ${VALID_PERMISSION_MODES.join(", ")}`);
+        }
+        permissionMode = flagMode;
+    }
+    else if (ir.permissions.mode !== undefined) {
+        permissionMode = ir.permissions.mode;
+    }
+    else {
+        permissionMode = "default";
+    }
+    const permissionRules = buildRuleSet(ir.permissions.rules, process.cwd());
+    // FR-004 — resolve the Pillar 3 intent-gate judge (flag > spec > rule-based).
+    // undefined leaves runtime-core on `ruleBasedJustificationJudge`.
+    const justificationJudge = await resolveJustificationJudge(args, ir.security?.justification);
+    // FR-004 — open the durable audit sink the intent gate appends a
+    // `permission_justification_evaluated` record to. On by default for `run`
+    // (rooted at .crewhaus/audit); `--no-justification-audit` skips it. undefined
+    // leaves runtime-core writing only the ephemeral trace-bus event.
+    const justificationAuditSink = await openJustificationAuditSink({
+        cwd: process.cwd(),
+        enabled: args.flags["no-justification-audit"] !== true,
+    });
+    // FR-006 — resolve the Pillar 3 sink-side egress matcher (flag > spec >
+    // substring). undefined leaves runtime-core on the built-in
+    // `substringMatcher`; `"semantic"` constructs the optional
+    // `@crewhaus/egress-matcher-semantic` with an injected embedder. The
+    // placement (IR-wired, every external sink) and the warn/block policy are
+    // unchanged — only *how* lineage matches are detected.
+    const egressMatcher = await resolveEgressMatcher(args, ir.security?.egressMatcher);
+    // Section 11 — discover hooks, skills, and slash commands from the user's
+    // workspace. Hooks come from `~/.crewhaus/settings.json` + `<cwd>/.crewhaus/settings.json`;
+    // skills from `~/.crewhaus/skills/*/SKILL.md` + project-equivalent; slash
+    // commands from `<cwd>/.crewhaus/commands/*.md`. When skills are present,
+    // a synthetic `Skill(name)` tool is appended to the tool list so the
+    // model can lazily fetch each skill's body.
+    const cwd = process.cwd();
+    const [hooks, skills, slashCommands] = await Promise.all([
+        loadHooks({ cwd }),
+        discoverSkills({ cwd }),
+        loadCommands({ cwd }),
+    ]);
+    if (skills.length > 0) {
+        tools.push(createSkillTool(skills));
+        process.stdout.write(`[skills] ${skills.length} available: ${skills.map((s) => s.name).join(", ")}\n`);
+    }
+    if (hooks.length > 0)
+        process.stdout.write(`[hooks] ${hooks.length} loaded\n`);
+    if (slashCommands.size > 0) {
+        process.stdout.write(`[slash] ${slashCommands.size} commands: ${[...slashCommands.keys()].join(", ")}\n`);
+    }
+    // Section 13 — when the IR carries inline sub-agent definitions, build the
+    // registry, register the Task tool, and inject `spawnSubAgent` so the
+    // runtime can populate the bridge for framework-aware tools.
+    let subAgents;
+    if (ir.subAgents.length > 0) {
+        subAgents = new Map(ir.subAgents.map((d) => [
+            d.name,
+            {
+                name: d.name,
+                description: d.description,
+                instructions: d.instructions,
+                tools: d.tools,
+                ...(d.model !== undefined ? { model: d.model } : {}),
+                permissions: d.permissions,
+                inherit_bypass: d.inheritBypass,
+            },
+        ]));
+        tools.push(createTaskTool({ subAgents }));
+        process.stdout.write(`[sub-agents] ${subAgents.size} available: ${[...subAgents.keys()].join(", ")}\n`);
+    }
+    try {
+        await runChatLoop({
+            model,
+            instructions: ir.agent.instructions,
+            tools,
+            permissionMode,
+            permissionRules,
+            sessionName: ir.name,
+            sessionTarget: ir.target,
+            hooks,
+            skills,
+            slashCommands,
+            ...(subAgents !== undefined ? { subAgents, spawnSubAgent } : {}),
+            ...(ir.target === "cli" && ir.agent.maxTokens !== undefined
+                ? { maxTokens: ir.agent.maxTokens }
+                : {}),
+            ...(resumeId !== undefined ? { resume: { sessionId: resumeId } } : {}),
+            ...(justificationJudge !== undefined ? { justificationJudge } : {}),
+            ...(justificationAuditSink !== undefined ? { justificationAuditSink } : {}),
+            ...(egressMatcher !== undefined ? { egressMatcher } : {}),
+        });
+    }
+    finally {
+        if (mcpHost)
+            await mcpHost.disconnectAll();
+    }
+}
+/**
+ * browser-target run path. Single-turn: read one prompt (from --prompt or
+ * stdin), drive a chromium session via @crewhaus/computer-use-driver,
+ * disconnect cleanly. Mirrors the codegen template in
+ * packages/target-browser-driver/src/index.ts — keep them in sync.
+ *
+ * V0 deliberately omits MCP, hooks, skills, slash commands, and sub-agents:
+ * runChatLoop runs with `singleTurn: true`, so multi-turn-only features
+ * aren't load-bearing here. The cli path keeps them. The browser bundle
+ * emitter (`@crewhaus/target-browser-driver`) honors ir.tools +
+ * ir.toolConfigs and similarly skips mcp_servers / compaction — extend
+ * both in lockstep if that ever changes.
+ */
+async function runRunBrowser(args, ir) {
+    if (args.flags["resume"] !== undefined || args.flags["continue"] === true) {
+        die("--resume and --continue are not supported for target: browser (single-turn)");
+    }
+    const promptFlag = args.flags["prompt"];
+    let prompt;
+    if (typeof promptFlag === "string" && promptFlag.length > 0) {
+        prompt = promptFlag;
+    }
+    else if (process.stdin.isTTY) {
+        die("no prompt — pass --prompt <text> or pipe input on stdin");
+    }
+    else {
+        prompt = (await readAllStdin()).trim();
+        if (prompt.length === 0) {
+            die("no prompt — pass --prompt <text> or pipe non-empty input on stdin");
+        }
+    }
+    let tools = [];
+    if (ir.tools.length > 0) {
+        await applyToolConfigs(ir.tools, ir.toolConfigs);
+        const toolMap = await loadToolMap();
+        tools = ir.tools.map((name) => {
+            const tool = toolMap[name];
+            if (!tool) {
+                const known = Object.keys(toolMap).sort().join(", ");
+                die(`unknown tool "${name}" — known tools: ${known}`);
+            }
+            return tool;
+        });
+    }
+    const modelOverride = args.flags["model"];
+    const model = typeof modelOverride === "string" ? modelOverride : ir.agent.model;
+    const flagMode = args.flags["permission-mode"];
+    let permissionMode;
+    if (typeof flagMode === "string") {
+        if (!isValidPermissionMode(flagMode)) {
+            die(`invalid --permission-mode "${flagMode}" — allowed: ${VALID_PERMISSION_MODES.join(", ")}`);
+        }
+        permissionMode = flagMode;
+    }
+    else if (ir.permissions.mode !== undefined) {
+        permissionMode = ir.permissions.mode;
+    }
+    else {
+        permissionMode = "default";
+    }
+    const permissionRules = buildRuleSet(ir.permissions.rules, process.cwd());
+    // FR-004 — honour --justification-judge on the browser one-shot path too.
+    // The browser spec shape carries no `security` block, so this is
+    // flag-only (flag > rule-based); both run paths thread the same judge
+    // into the same gate inside runChatLoop.
+    const justificationJudge = await resolveJustificationJudge(args, undefined);
+    // FR-006 — honour --egress-matcher on the browser one-shot path too. The
+    // browser spec shape carries no `security` block, so this is flag-only
+    // (flag > substring); both run paths thread the same matcher into the same
+    // egress check inside runChatLoop.
+    const egressMatcher = await resolveEgressMatcher(args, undefined);
+    // FR-004 — same durable audit sink as the cli path, so a justification-gated
+    // browser tool also writes the `permission_justification_evaluated` record.
+    const justificationAuditSink = await openJustificationAuditSink({
+        cwd: process.cwd(),
+        enabled: args.flags["no-justification-audit"] !== true,
+    });
+    // Lazy-import the browser-runtime packages so cli/init/doctor invocations
+    // don't pay the playwright + computer-use-driver load cost.
+    const [{ createDriver }, navigate, screenCapture, mouseKeyboard, visionGrounding] = await Promise.all([
+        import("@crewhaus/computer-use-driver"),
+        import("@crewhaus/tool-navigate"),
+        import("@crewhaus/tool-screen-capture"),
+        import("@crewhaus/tool-mouse-keyboard"),
+        import("@crewhaus/tool-vision-grounding"),
+    ]);
+    const driver = createDriver({
+        backend: ir.driver.backend,
+        viewport: ir.driver.viewport,
+    });
+    emitEvent({ kind: "browser_start", backend: ir.driver.backend });
+    await driver.connect();
+    try {
+        if (ir.driver.startUrl !== undefined) {
+            await driver.goto(ir.driver.startUrl);
+            emitEvent({ kind: "navigated", url: ir.driver.startUrl });
+        }
+        const navigateTool = navigate.createNavigateTool({ driver });
+        const screenshotTool = screenCapture.createScreenshotTool({ driver });
+        const mk = mouseKeyboard.createAllMouseKeyboardTools({ driver });
+        const findElement = visionGrounding.createFindElementTool({
+            driver,
+            model: ir.groundingModel,
+        });
+        const allTools = [
+            navigateTool,
+            screenshotTool,
+            mk.click,
+            mk.type,
+            mk.key,
+            mk.scroll,
+            findElement,
+            ...tools,
+        ];
+        const finalText = await runChatLoop({
+            model,
+            instructions: ir.agent.instructions,
+            tools: allTools,
+            permissionMode,
+            permissionRules,
+            sessionName: ir.name,
+            sessionTarget: "browser",
+            singleTurn: true,
+            seedMessages: [{ role: "user", content: prompt }],
+            ...(justificationJudge !== undefined ? { justificationJudge } : {}),
+            ...(justificationAuditSink !== undefined ? { justificationAuditSink } : {}),
+            ...(egressMatcher !== undefined ? { egressMatcher } : {}),
+            installSigintHandler: false,
+            maxTokens: 4096,
+        });
+        emitEvent({ kind: "browser_done", finalText });
+    }
+    finally {
+        await driver.disconnect();
+    }
+}
+async function readAllStdin() {
+    const chunks = [];
+    for await (const chunk of process.stdin)
+        chunks.push(chunk);
+    return Buffer.concat(chunks).toString("utf-8");
+}
+function emitEvent(event) {
+    process.stdout.write(`${JSON.stringify(event)}\n`);
+}
+function checkBunVersion(version) {
+    const parts = version.split(".");
+    const major = Number.parseInt(parts[0] ?? "", 10);
+    const minor = Number.parseInt(parts[1] ?? "", 10);
+    if (Number.isNaN(major) || Number.isNaN(minor)) {
+        return { pass: false, reason: `unparseable version "${version}"` };
+    }
+    const ok = major > 1 || (major === 1 && minor >= 2);
+    return ok ? { pass: true } : { pass: false, reason: `bun ${version} is below minimum 1.2.0` };
+}
+function runContext(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus context --bundle [-o <file>]\n" +
+            "  Emits a single-markdown manifest of the spec schema, recipe index,\n" +
+            "  module catalog headings, and getting-started guide. Designed for\n" +
+            "  piping into an agent's system prompt or for caching as the answer\n" +
+            '  to "give me CrewHaus context."\n' +
+            "\n" +
+            "  --factory-root <p>   override factory checkout (default: $CREWHAUS_FACTORY_ROOT\n" +
+            "                       or sibling-walk from cwd)\n" +
+            "  --docs-root <p>      override docs checkout\n" +
+            "  --demos-root <p>     override demos checkout\n");
+        return;
+    }
+    if (args.flags["bundle"] !== true) {
+        die("missing --bundle (the only mode currently supported)");
+    }
+    const factoryOverride = args.flags["factory-root"];
+    const docsOverride = args.flags["docs-root"];
+    const demosOverride = args.flags["demos-root"];
+    const env = {
+        ...process.env,
+        ...(typeof factoryOverride === "string" ? { CREWHAUS_FACTORY_ROOT: factoryOverride } : {}),
+        ...(typeof docsOverride === "string" ? { CREWHAUS_DOCS_ROOT: docsOverride } : {}),
+        ...(typeof demosOverride === "string" ? { CREWHAUS_DEMOS_ROOT: demosOverride } : {}),
+    };
+    let roots;
+    try {
+        roots = discoverRoots({ env });
+    }
+    catch (err) {
+        die(err.message);
+    }
+    const bundle = buildContextBundle({
+        factoryRoot: roots.factoryRoot,
+        docsRoot: roots.docsRoot,
+        demosRoot: roots.demosRoot,
+    });
+    const outFile = args.flags["out"];
+    if (typeof outFile === "string") {
+        const abs = resolve(outFile);
+        const dir = dirname(abs);
+        if (!existsSync(dir))
+            mkdirSync(dir, { recursive: true });
+        writeFileSync(abs, bundle.markdown);
+        process.stderr.write(`wrote bundle: ${abs} (${bundle.markdown.length} chars, ${bundle.sources.length} sources)\n`);
+        return;
+    }
+    process.stdout.write(bundle.markdown);
+}
+async function runDoctor(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus doctor [--philosophy-alignment] [--liveness]\n" +
+            "  --philosophy-alignment   audit the codebase + examples against the three architectural pillars\n" +
+            "  --liveness               process-liveness probe: exit 0 immediately, no credential or\n" +
+            "                           spec checks (for container HEALTHCHECKs / k8s exec probes)\n" +
+            "\n" +
+            "  Credential checks are model-aware: doctor parses the cwd crewhaus.yaml's\n" +
+            "  agent.model (claude-*, openai/*, gemini/*, bedrock/*, local/<m>@<url>) and\n" +
+            "  checks the matching provider's env; other providers report informationally.\n");
+        return;
+    }
+    // --liveness: pure process-liveness for container/k8s probes. The doctor
+    // binary booting far enough to parse argv IS the signal — no credential or
+    // spec checks (a probe must not flap on missing keys or a spec-less image).
+    if (args.flags["liveness"]) {
+        process.stdout.write("ok\n");
+        process.exit(0);
+    }
+    if (args.flags["philosophy-alignment"]) {
+        await runDoctorPhilosophyAlignment();
+        return;
+    }
+    const checks = [];
+    // Model-aware provider credentials: read the cwd spec's agent.model (when
+    // present + parseable) and check the MATCHING provider's env. Falls back to
+    // the legacy Anthropic-first check when no model is extractable. See
+    // doctor-checks.ts for the full policy (bedrock/local are informational).
+    const specPath = join(process.cwd(), "crewhaus.yaml");
+    let specModel;
+    if (existsSync(specPath)) {
+        try {
+            specModel = extractSpecModel(readFileSync(specPath, "utf-8"));
+        }
+        catch {
+            specModel = undefined;
+        }
+    }
+    checks.push(...buildCredentialChecks(specModel, process.env));
+    const bunCheck = checkBunVersion(Bun.version);
+    checks.push({
+        label: `Bun runtime (${Bun.version})`,
+        pass: bunCheck.pass,
+        reason: bunCheck.reason,
+    });
+    checks.push({
+        label: "crewhaus.yaml in cwd",
+        pass: existsSync(specPath),
+        reason: existsSync(specPath) ? undefined : `not found at ${specPath} — run \`crewhaus init\``,
+    });
+    for (const c of checks) {
+        if (c.warn && c.pass) {
+            process.stdout.write(`~ ${c.label}: ${c.reason ?? "informational"}\n`);
+        }
+        else if (c.pass) {
+            process.stdout.write(`✓ ${c.label}\n`);
+        }
+        else {
+            process.stdout.write(`✗ ${c.label}: ${c.reason ?? "failed"}\n`);
+        }
+    }
+    const allPass = checks.every((c) => c.pass);
+    process.stdout.write(allPass ? "\nall checks passed.\n" : "\nsome checks failed.\n");
+    process.exit(allPass ? 0 : 1);
+}
+/**
+ * Workstream D — `crewhaus doctor --philosophy-alignment`. Audits the
+ * repo against the three architectural pillars (compiler-as-protagonist,
+ * eval-is-active, security-is-fabric). Exits 0 on green, 1 on any
+ * finding. Designed to be runnable in CI as a soft gate.
+ */
+async function runDoctorPhilosophyAlignment() {
+    const findings = [];
+    // Pillar 1 — compiler-as-protagonist. The IR-discriminated-union is
+    // the contract; the architecture doc must reference the IR variants.
+    // Canonical docs live off-repo at github.com/crewhaus/docs; we look
+    // for a sibling checkout (../docs/) in the developer workspace and
+    // warn-skip when absent rather than failing the audit.
+    const archDocPath = resolve(process.cwd(), "..", "docs", "COMPILER-ARCHITECTURE.md");
+    if (existsSync(archDocPath)) {
+        const content = readFileSync(archDocPath, "utf8");
+        const referencesIrVariants = content.includes("IrV0") && content.includes("IrPipelineV0") && content.includes("IrGraphV0");
+        findings.push({
+            label: "Pillar 1 — ../docs/COMPILER-ARCHITECTURE.md references IR variants",
+            pass: referencesIrVariants,
+            reason: referencesIrVariants
+                ? undefined
+                : "doc exists but does not enumerate the IR-discriminated-union variants",
+        });
+    }
+    else {
+        findings.push({
+            label: "Pillar 1 — COMPILER-ARCHITECTURE.md (sibling checkout)",
+            pass: true,
+            warn: true,
+            reason: "sibling ../docs not cloned; canonical docs at github.com/crewhaus/docs — clone alongside factory/ to enable this check",
+        });
+    }
+    // Pillar 2 — eval is active. The eval-optimizer-orchestrator must
+    // be in the workspace and the optimize CLI subcommand must exist.
+    const orchestratorPkg = join(process.cwd(), "packages", "eval-optimizer-orchestrator", "package.json");
+    findings.push({
+        label: "Pillar 2 — eval-optimizer-orchestrator package present",
+        pass: existsSync(orchestratorPkg),
+        reason: existsSync(orchestratorPkg)
+            ? undefined
+            : "Workstream B did not land — install or rebuild",
+    });
+    const specPatchPkg = join(process.cwd(), "packages", "spec-patch", "package.json");
+    findings.push({
+        label: "Pillar 2 — spec-patch package present",
+        pass: existsSync(specPatchPkg),
+        reason: existsSync(specPatchPkg) ? undefined : "Workstream B did not land",
+    });
+    // Pillar 3 — security is a fabric. The boundary-classifier package
+    // must exist and be referenced by the canonical boundary sites
+    // (tool-mcp, sub-agent-spawner, skills-registry, compaction-autocompact,
+    // federation-router, channel-adapter-base).
+    const boundaryPkg = join(process.cwd(), "packages", "boundary-classifier", "package.json");
+    findings.push({
+        label: "Pillar 3 — boundary-classifier package present",
+        pass: existsSync(boundaryPkg),
+        reason: existsSync(boundaryPkg) ? undefined : "Workstream C did not land",
+    });
+    const boundarySites = [
+        { name: "tool-mcp", path: "packages/tool-mcp/src/index.ts" },
+        { name: "sub-agent-spawner", path: "packages/sub-agent-spawner/src/index.ts" },
+        { name: "skills-registry", path: "packages/skills-registry/src/index.ts" },
+        { name: "compaction-autocompact", path: "packages/compaction-autocompact/src/index.ts" },
+        { name: "federation-router", path: "packages/federation-router/src/index.ts" },
+        { name: "channel-adapter-base", path: "packages/channel-adapter-base/src/index.ts" },
+    ];
+    for (const site of boundarySites) {
+        const filePath = join(process.cwd(), site.path);
+        if (!existsSync(filePath)) {
+            findings.push({
+                label: `Pillar 3 — ${site.name} source present`,
+                pass: false,
+                reason: `${site.path} not found`,
+            });
+            continue;
+        }
+        const body = readFileSync(filePath, "utf8");
+        const classifies = body.includes("classifyBoundary") || body.includes("boundary-classifier");
+        findings.push({
+            label: `Pillar 3 — ${site.name} calls classifyBoundary`,
+            pass: classifies,
+            reason: classifies
+                ? undefined
+                : `no reference to classifyBoundary in ${site.path} — security regression`,
+        });
+    }
+    // FR-002 — Pillar 3 sink-side. Audit the full built-in tool map: every
+    // outward-reaching built-in must carry scope: "external" so the egress
+    // classifier fires on it. Uses the SAME `auditToolScopes` helper as
+    // `compile --strict` (acceptance: the two paths share one implementation).
+    // Importing the tool packages is heavier than the rest of doctor, so it's
+    // confined to this --philosophy-alignment branch.
+    const toolMap = await loadToolMap();
+    const scopeFindings = auditToolScopes(Object.values(toolMap));
+    if (scopeFindings.length === 0) {
+        findings.push({
+            label: 'Pillar 3 — all built-in outward tools scope:"external"',
+            pass: true,
+        });
+    }
+    else {
+        for (const f of scopeFindings) {
+            findings.push({
+                label: `Pillar 3 — tool "${f.toolName}" outward but scope!="external"`,
+                pass: false,
+                reason: f.reason,
+            });
+        }
+    }
+    // Contributor compass exists at project root. AGENTS.md is the canonical
+    // vendor-neutral convention (agents.md); CLAUDE.md is accepted as a fallback
+    // for repos still on the Claude Code-specific naming.
+    const agentsmd = join(process.cwd(), "AGENTS.md");
+    const claudemd = join(process.cwd(), "CLAUDE.md");
+    const hasContributorCompass = existsSync(agentsmd) || existsSync(claudemd);
+    findings.push({
+        label: "Contributor doc — AGENTS.md (or CLAUDE.md) at project root",
+        pass: hasContributorCompass,
+        reason: hasContributorCompass ? undefined : "missing — contributors will drift",
+    });
+    for (const f of findings) {
+        if (f.warn && f.pass) {
+            process.stdout.write(`~ ${f.label}: ${f.reason ?? "skipped"}\n`);
+        }
+        else if (f.pass) {
+            process.stdout.write(`✓ ${f.label}\n`);
+        }
+        else {
+            process.stdout.write(`✗ ${f.label}: ${f.reason ?? "failed"}\n`);
+        }
+    }
+    const allPass = findings.every((f) => f.pass);
+    process.stdout.write(allPass
+        ? "\nphilosophy alignment: green. All three pillars intact.\n"
+        : `\nphilosophy alignment: ${findings.filter((f) => !f.pass).length} finding(s). See [/AGENTS.md](AGENTS.md) for invariants.\n`);
+    process.exit(allPass ? 0 : 1);
+}
+/**
+ * Workstream B — `crewhaus optimize <spec> --dataset <data> --graders
+ * <graders.yaml> [--mutator rule-based|claude] [--iterations N]
+ * [--budget-usd N] [--write-back] [-o <out-dir>]`. Closes the
+ * active-optimisation loop.
+ *
+ * Loads the spec + dataset + graders, builds a fitness function that
+ * re-runs `eval-runner` for every candidate prompt, delegates to
+ * `optimizeSpec`, and prints the resulting score delta + spend summary +
+ * patch path. FR-003: `--budget-usd N` bounds a model-driven run by a
+ * dollar ceiling (the orchestrator stops before a mutation call that
+ * would exceed it); it composes with `--iterations` (first bound wins)
+ * and is inert on rule-based runs (which make no model calls → $0). A
+ * `TraceEventBus` is threaded into `optimizeSpec` so each call's
+ * `cost_accrual` and the terminal aggregate spend summary land on the
+ * standard observability bus on this real path (CREWHAUS_TRACE_COST=1
+ * echoes them to stderr); the run-total $ also prints to stdout.
+ */
+async function runOptimize(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus optimize <spec.yaml> --dataset <data> --graders <graders.yaml> " +
+            "[--mutator rule-based|claude] [--iterations N] [--seed N] " +
+            "[--improvement-threshold F] [--budget-usd N] [--write-back] [-o <out-dir>]\n");
+        return;
+    }
+    const specPath = args.positional[0];
+    if (typeof specPath !== "string")
+        die("missing <spec.yaml>");
+    const datasetPath = args.flags["dataset"];
+    const gradersPath = args.flags["graders"];
+    if (typeof datasetPath !== "string")
+        die("missing --dataset <data>");
+    if (typeof gradersPath !== "string")
+        die("missing --graders <graders.yaml>");
+    const iterationsFlag = args.flags["iterations"];
+    const seedFlag = args.flags["seed"];
+    const thresholdFlag = args.flags["improvement-threshold"];
+    const iterations = typeof iterationsFlag === "string" ? Number.parseInt(iterationsFlag, 10) : 10;
+    const seed = typeof seedFlag === "string" ? Number.parseInt(seedFlag, 10) : 0xcafe;
+    const improvementThreshold = typeof thresholdFlag === "string" ? Number.parseFloat(thresholdFlag) : 0.01;
+    if (Number.isNaN(iterations) || iterations < 1) {
+        die(`invalid --iterations "${iterationsFlag}" — must be positive integer`);
+    }
+    // FR-003 — optional dollar budget for model-driven runs. Omit → today's
+    // behaviour (iterations cap only). On a rule-based run the gate is inert
+    // (no model calls → $0), so passing it is harmless.
+    const budgetFlag = args.flags["budget-usd"];
+    let budgetUsd;
+    if (typeof budgetFlag === "string") {
+        budgetUsd = Number.parseFloat(budgetFlag);
+        if (Number.isNaN(budgetUsd) || budgetUsd <= 0) {
+            die(`invalid --budget-usd "${budgetFlag}" — must be a positive number`);
+        }
+    }
+    const writeBack = args.flags["write-back"] === true;
+    const outDirArg = args.flags["out"];
+    const absSpec = resolve(specPath);
+    const runId = `opt_${Date.now().toString(16)}`;
+    const outDir = typeof outDirArg === "string"
+        ? resolve(outDirArg)
+        : resolve(join(".crewhaus", "optimize", runId));
+    let gradersYaml;
+    try {
+        gradersYaml = readFileSync(resolve(gradersPath), "utf-8");
+    }
+    catch (err) {
+        die(`could not read ${gradersPath}: ${err.message}`);
+    }
+    const { compiled } = parseGradersConfig(gradersYaml);
+    const dataset = await loadDataset(resolve(datasetPath));
+    // Materialize the dataset once — we'll re-iterate per fitness call.
+    const samples = [];
+    for await (const s of dataset.samples) {
+        samples.push({
+            id: s.id,
+            input: s.input,
+            ...(s.expected_output !== undefined ? { expected_output: s.expected_output } : {}),
+        });
+    }
+    if (samples.length === 0)
+        die(`dataset "${dataset.name}" yielded zero samples`);
+    // Train/dev split: 70/30 deterministic split by sample id ordering.
+    const splitIdx = Math.max(1, Math.floor(samples.length * 0.7));
+    const trainSet = samples.slice(0, splitIdx);
+    const devSet = samples.slice(splitIdx);
+    if (devSet.length === 0) {
+        die(`dataset has ${samples.length} samples — need at least 2 (70/30 split needs a dev split)`);
+    }
+    // Fitness fn: patch the spec with the candidate prompt, lower to IR,
+    // run eval-runner, return passRate. Each call is one full eval pass.
+    const fitness = async (prompt) => {
+        const yamlText = readFileSync(absSpec, "utf-8");
+        // Re-parse to capture spec.target without depending on the
+        // orchestrator's extractCurrentPrompt internals.
+        const parsedTarget = parseSpec(yamlText).target;
+        // Build a patch and apply it in-memory (no disk write — fitness is
+        // pure with respect to the source file).
+        const { applySpecPatch } = await import("@crewhaus/spec-patch");
+        const { yaml: patchedYaml } = applySpecPatch(yamlText, {
+            target: parsedTarget,
+            path: ["agent", "instructions"],
+            op: "replace",
+            value: prompt,
+        });
+        let ir;
+        try {
+            ir = lower(parseSpec(patchedYaml));
+        }
+        catch (err) {
+            if (err instanceof SpecParseError) {
+                process.stderr.write("[optimize] candidate compiled invalid spec, skipping\n");
+                return 0;
+            }
+            throw err;
+        }
+        if (ir.target !== "cli") {
+            die(`crewhaus optimize v0 only supports target: cli (got "${ir.target}")`);
+        }
+        const summary = await runEvalLib({
+            ir,
+            dataset: { name: dataset.name, samples: makeAsyncIterable(devSet) },
+            compiledGraders: compiled,
+            opts: {
+                outDir: join(outDir, "evals", `${prompt.length}_${ir.agent.instructions.length}`),
+                concurrency: 4,
+                seed,
+            },
+        });
+        return summary.aggregates.passRate;
+    };
+    const mutator = args.flags["mutator"];
+    let mutatorImpl;
+    if (mutator === "claude") {
+        const { createClaudeMutationProvider } = await import("@crewhaus/prompt-optimizer-claude");
+        const { resolveModel } = await import("@crewhaus/model-router");
+        const ir = lower(parseSpec(readFileSync(absSpec, "utf-8")));
+        const mutatorModel = ir.target === "cli" ? ir.agent.model : "claude-sonnet-4-5";
+        // Resolve via the model-router so non-Anthropic specs drive their own
+        // provider: the resolved adapter + STRIPPED wire modelId replace the old
+        // hardcoded createAnthropicAdapter() + verbatim prefixed string (which
+        // made `--mutator claude` a silent no-op for openai/gemini/bedrock/local
+        // specs — every mutation call failed and the provider fell back).
+        const resolution = await resolveModel(mutatorModel);
+        mutatorImpl = createClaudeMutationProvider({
+            adapter: resolution.adapter,
+            model: resolution.modelId,
+        });
+    }
+    else if (mutator !== undefined && mutator !== "rule-based") {
+        die(`unknown --mutator "${mutator}" — supported: rule-based, claude`);
+    }
+    process.stdout.write(`[optimize] runId=${runId} spec=${specPath} dataset=${dataset.name} ` +
+        `(${trainSet.length} train / ${devSet.length} dev) iterations=${iterations} ` +
+        `mutator=${mutator ?? "rule-based"}\n`);
+    // FR-003 — thread a real trace bus into the optimize run so the per-call
+    // `cost_accrual` events AND the terminal aggregate summary land on the
+    // standard observability bus on the actual `crewhaus optimize` path — not
+    // only when a unit test injects a bus. The spend total still prints to
+    // stdout below (unchanged default UX); set CREWHAUS_TRACE_COST=1 to also
+    // echo each bus cost event to stderr for live observability.
+    const traceBus = new TraceEventBus({ runId, sessionId: runId });
+    if (process.env["CREWHAUS_TRACE_COST"] === "1") {
+        traceBus.subscribe((e) => {
+            if (e.kind !== "cost_accrual")
+                return;
+            const ev = e;
+            const label = ev.summary === true ? "cost-total" : "cost-call";
+            process.stderr.write(`[optimize] ${label} provider=${ev.provider} model=${ev.modelId} ` +
+                `in=${ev.inputTokens} out=${ev.outputTokens} micros=${ev.costUsdMicros}\n`);
+        });
+    }
+    const result = await optimizeSpec({
+        specPath: absSpec,
+        fitness,
+        trainSet,
+        devSet,
+        iterations,
+        seed,
+        improvementThreshold,
+        outDir,
+        writeBack,
+        runId,
+        traceBus,
+        ...(mutatorImpl !== undefined ? { mutator: mutatorImpl } : {}),
+        ...(budgetUsd !== undefined ? { budgetUsd } : {}),
+    });
+    process.stdout.write(`[optimize] score: ${result.scoreBefore.toFixed(3)} → ${result.scoreAfter.toFixed(3)} ` +
+        `(Δ ${result.improvement >= 0 ? "+" : ""}${result.improvement.toFixed(3)})\n`);
+    // FR-003 — spend summary: total $ + model-call count, and whether the
+    // dollar budget (not the iterations cap) ended the run.
+    const spendStop = result.stoppedReason === "budget-reached"
+        ? ` (stopped: budget reached, $${budgetUsd?.toFixed(2)} cap)\n`
+        : "\n";
+    process.stdout.write(`[optimize] spend: ${result.spend.totalUsd} over ${result.spend.perIteration.length} model call(s)${spendStop}`);
+    process.stdout.write(`[optimize] patch: ${join(result.outDir, "patch.json")}\n`);
+    if (result.applied) {
+        if (result.writtenTo) {
+            process.stdout.write(`[optimize] wrote patched YAML to ${result.writtenTo}\n`);
+        }
+        else {
+            process.stdout.write(`[optimize] patch ready (improvement ≥ ${improvementThreshold}). Re-run with --write-back to apply.\n`);
+        }
+    }
+    else {
+        process.stdout.write(`[optimize] no improvement above threshold ${improvementThreshold}; source untouched.\n`);
+    }
+}
+function makeAsyncIterable(items) {
+    return {
+        async *[Symbol.asyncIterator]() {
+            for (const item of items)
+                yield item;
+        },
+    };
+}
+async function runEvalSubcommand(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus eval <spec.yaml> --dataset <data> --graders <graders.yaml> " +
+            "[--judge-model <model>] [--concurrency N] [--seed N] -o <out-dir>\n" +
+            "  --judge-model accepts the full router grammar (claude-*, openai/<m>, gemini/<m>,\n" +
+            "  bedrock/<id>, local/<m>@<url>); the default judge claude-sonnet-4-5 requires\n" +
+            "  Anthropic credentials (ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY)\n");
+        return;
+    }
+    const specPath = args.positional[0];
+    if (typeof specPath !== "string")
+        die("missing <spec.yaml>");
+    const datasetPath = args.flags["dataset"];
+    const gradersPath = args.flags["graders"];
+    const outDirArg = args.flags["out"];
+    if (typeof datasetPath !== "string")
+        die("missing --dataset <data>");
+    if (typeof gradersPath !== "string")
+        die("missing --graders <graders.yaml>");
+    if (typeof outDirArg !== "string")
+        die("missing -o <out-dir>");
+    const concurrencyFlag = args.flags["concurrency"];
+    const seedFlag = args.flags["seed"];
+    const judgeModelFlag = args.flags["judge-model"];
+    const concurrency = typeof concurrencyFlag === "string" ? Number.parseInt(concurrencyFlag, 10) : undefined;
+    const seed = typeof seedFlag === "string" ? Number.parseInt(seedFlag, 10) : undefined;
+    if (concurrency !== undefined && (Number.isNaN(concurrency) || concurrency < 1)) {
+        die(`invalid --concurrency "${concurrencyFlag}" — must be positive integer`);
+    }
+    if (seed !== undefined && Number.isNaN(seed)) {
+        die(`invalid --seed "${seedFlag}" — must be integer`);
+    }
+    const absSpec = resolve(specPath);
+    let yamlText;
+    try {
+        yamlText = readFileSync(absSpec, "utf-8");
+    }
+    catch (err) {
+        die(`could not read ${absSpec}: ${err.message}`);
+    }
+    let ir;
+    try {
+        ir = lower(parseSpec(yamlText));
+    }
+    catch (err) {
+        if (err instanceof SpecParseError)
+            die(err.message);
+        throw err;
+    }
+    if (ir.target !== "cli") {
+        die(`crewhaus eval only supports target: cli (got "${ir.target}")`);
+    }
+    let gradersYaml;
+    try {
+        gradersYaml = readFileSync(resolve(gradersPath), "utf-8");
+    }
+    catch (err) {
+        die(`could not read ${gradersPath}: ${err.message}`);
+    }
+    const { compiled } = parseGradersConfig(gradersYaml);
+    const dataset = await loadDataset(resolve(datasetPath));
+    const absOut = resolve(outDirArg);
+    process.stdout.write(`[eval] running ${dataset.name}: ${compiled.length} graders → ${absOut}\n`);
+    const summary = await runEvalLib({
+        ir,
+        dataset,
+        compiledGraders: compiled,
+        opts: {
+            outDir: absOut,
+            ...(concurrency !== undefined ? { concurrency } : {}),
+            ...(seed !== undefined ? { seed } : {}),
+            ...(typeof judgeModelFlag === "string" ? { judgeModel: judgeModelFlag } : {}),
+        },
+    });
+    // Render report
+    const loaded = await loadRun(absOut);
+    const rendered = renderReport(loaded);
+    writeFileSync(join(absOut, "index.html"), rendered.html);
+    process.stdout.write(`[eval] runId=${summary.runId} pass_rate=${(summary.aggregates.passRate * 100).toFixed(1)}% ` +
+        `mean_score=${summary.aggregates.meanScore.toFixed(3)} ` +
+        `errors=${summary.aggregates.errorCount} ` +
+        `tokens=${summary.aggregates.totalTokens.input}/${summary.aggregates.totalTokens.output}\n`);
+    process.stdout.write(`[eval] report: ${join(absOut, "index.html")}\n`);
+}
+async function runEvalReport(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus eval-report diff <prev> <new> [-o <out-dir>]\n");
+        return;
+    }
+    const action = args.positional[0];
+    if (action !== "diff")
+        die(`eval-report: unknown action "${action ?? ""}" — supported: diff`);
+    const prev = args.positional[1];
+    const next = args.positional[2];
+    if (typeof prev !== "string" || typeof next !== "string") {
+        die("eval-report diff: missing <prev> <new>");
+    }
+    const outArg = args.flags["out"];
+    const prevLoaded = await loadRun(prev);
+    const nextLoaded = await loadRun(next);
+    const result = diffReports(prevLoaded, nextLoaded);
+    if (typeof outArg === "string") {
+        const absOut = resolve(outArg);
+        mkdirSync(absOut, { recursive: true });
+        writeFileSync(join(absOut, "index.html"), result.html);
+        writeFileSync(join(absOut, "diff.json"), result.json);
+        process.stdout.write(`[eval-report] diff: ${join(absOut, "index.html")}\n`);
+    }
+    else {
+        process.stdout.write(result.html);
+    }
+    process.stdout.write(`[eval-report] regressions=${result.diff.regressions.length} ` +
+        `recoveries=${result.diff.recoveries.length} ` +
+        `score_shifts=${result.diff.scoreShifts.length} ` +
+        `unchanged=${result.diff.unchanged}\n`);
+}
+/**
+ * Section 27 — `crewhaus cost-summary [--session <id>] [--tenant <id>]
+ * [--format json|text]`. Reads `cost_accrual` events out of an `event-log`
+ * (or aggregates the per-day audit-log records) and prints a USD summary.
+ *
+ * v0 ships with the per-session readout — pass `--session <id>` to read
+ * the JSONL transcript at `.crewhaus/sessions/<id>.jsonl` and aggregate
+ * the cost_accrual events embedded in there. Tenant aggregation lands in
+ * §31's studio-server cost dashboard.
+ */
+async function runCostSummary(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus cost-summary --session <id> [--format json|text]\n");
+        return;
+    }
+    const session = args.flags["session"];
+    const format = args.flags["format"] ?? "text";
+    if (typeof session !== "string")
+        die("missing --session <id>");
+    const sessionFile = join(process.cwd(), ".crewhaus", "sessions", `${session}.jsonl`);
+    if (!existsSync(sessionFile)) {
+        die(`session log not found at ${sessionFile}`);
+    }
+    const lines = readFileSync(sessionFile, "utf-8")
+        .split("\n")
+        .filter((l) => l !== "");
+    let totalMicros = 0;
+    const byProvider = {};
+    let count = 0;
+    for (const raw of lines) {
+        let parsed;
+        try {
+            parsed = JSON.parse(raw);
+        }
+        catch {
+            continue;
+        }
+        if (parsed &&
+            typeof parsed === "object" &&
+            "kind" in parsed &&
+            parsed.kind === "cost_accrual") {
+            // event-log writes a `{ ts, version, kind, payload }` envelope, so the
+            // cost fields live under `payload`. Fall back to top-level fields so a
+            // hand-written/flat cost_accrual line still aggregates.
+            const e = parsed;
+            const provider = e.payload?.provider ?? e.provider;
+            const micros = e.payload?.costUsdMicros ?? e.costUsdMicros;
+            if (typeof provider === "string" && typeof micros === "number") {
+                totalMicros += micros;
+                byProvider[provider] = (byProvider[provider] ?? 0) + micros;
+                count++;
+            }
+        }
+    }
+    const totalDollars = totalMicros / 1_000_000;
+    if (format === "json") {
+        process.stdout.write(`${JSON.stringify({ session, count, totalUsdMicros: totalMicros, byProvider })}\n`);
+    }
+    else {
+        process.stdout.write(`session: ${session}\n`);
+        process.stdout.write(`accrual events: ${count}\n`);
+        process.stdout.write(`total: $${totalDollars.toFixed(4)}\n`);
+        for (const [p, m] of Object.entries(byProvider)) {
+            process.stdout.write(`  ${p}: $${(m / 1_000_000).toFixed(4)}\n`);
+        }
+    }
+}
+/**
+ * Section 27 — `crewhaus secrets <action> <name> [opts]`. Two actions:
+ *   doctor                 list configured secrets and report missing
+ *   rotate <name>          rotate the named secret (file backend)
+ */
+async function runSecrets(args, action) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage:\n" +
+            "  crewhaus secrets doctor [--backend env-var|file] [--root-dir <dir>]\n" +
+            "  crewhaus secrets rotate <name> [--value <new-value>] [--backend ...]\n");
+        return;
+    }
+    const backendIdFlag = args.flags["backend"];
+    const backendId = typeof backendIdFlag === "string" ? backendIdFlag : "env-var";
+    const rootDirFlag = args.flags["root-dir"];
+    const rootDir = typeof rootDirFlag === "string" ? rootDirFlag : join(process.cwd(), ".crewhaus", "secrets");
+    const { createSecrets, createEnvVarBackend, createFileBackend } = await import("@crewhaus/secrets-manager");
+    // The `vault` backend exists in @crewhaus/secrets-manager but is not wired
+    // into the CLI (it needs an address/auth story the CLI does not yet expose).
+    // Fail loudly rather than silently degrading a security-sensitive flag to the
+    // env-var backend — a silent fallback would, e.g., make `secrets doctor`
+    // cheerfully list process env var names as if they were Vault secrets.
+    let backend;
+    if (backendId === "env-var") {
+        backend = createEnvVarBackend();
+    }
+    else if (backendId === "file") {
+        backend = createFileBackend({ rootDir });
+    }
+    else if (backendId === "vault") {
+        die("vault backend is not wired into the CLI in this build — construct it programmatically via createVaultBackend(), or use --backend env-var|file");
+    }
+    else {
+        die(`unknown secrets backend "${backendId}" (expected: env-var | file)`);
+    }
+    const secrets = createSecrets({ backend });
+    if (action === "doctor") {
+        const known = (await backend.list?.()) ?? [];
+        process.stdout.write(`backend: ${backend.id}\n`);
+        process.stdout.write(`known: ${known.length} secret(s)\n`);
+        for (const n of known)
+            process.stdout.write(`  - ${n}\n`);
+        return;
+    }
+    if (action === "rotate") {
+        const name = args.positional[0];
+        if (typeof name !== "string")
+            die("missing <name>");
+        const value = args.flags["value"];
+        const newValue = await secrets.rotate(name, {
+            ...(typeof value === "string" ? { newValue: value } : {}),
+        });
+        process.stdout.write(`rotated ${name} (${newValue.length} chars) via ${backend.id} backend\n`);
+        return;
+    }
+    die(`unknown secrets action "${action}" (expected: doctor | rotate)`);
+}
+/**
+ * Section 28 — `crewhaus spec <action> ...` subcommands wrap the
+ * spec-registry. Actions: put / get / list / pin / alias.
+ */
+async function runSpec(args, action) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage:\n" +
+            "  crewhaus spec put <name> <version> <spec.yaml> [--root-dir <dir>]\n" +
+            "  crewhaus spec list <name>                                   list versions\n" +
+            "  crewhaus spec get <name> <version>                          print yaml\n" +
+            "  crewhaus spec pin <name> <env> <version> [--tenant <id>]   pin env → version\n" +
+            "  crewhaus spec alias <name> <env> [--tenant <id>]            resolve env → version\n");
+        return;
+    }
+    const rootDirFlag = args.flags["root-dir"];
+    const rootDir = typeof rootDirFlag === "string" ? rootDirFlag : join(process.cwd(), ".crewhaus", "specs");
+    const { createFileBackedRegistry } = await import("@crewhaus/spec-registry");
+    const reg = createFileBackedRegistry({ rootDir });
+    const tenantFlag = args.flags["tenant"];
+    const tenantId = typeof tenantFlag === "string" ? tenantFlag : undefined;
+    if (action === "put") {
+        const name = args.positional[0];
+        const version = args.positional[1];
+        const filePath = args.positional[2];
+        if (typeof name !== "string")
+            die("missing <name>");
+        if (typeof version !== "string")
+            die("missing <version>");
+        if (typeof filePath !== "string")
+            die("missing <spec.yaml>");
+        const yaml = readFileSync(resolve(filePath), "utf-8");
+        await reg.put(name, version, yaml);
+        process.stdout.write(`stored ${name}@${version} (${yaml.length} bytes)\n`);
+        return;
+    }
+    if (action === "list") {
+        const name = args.positional[0];
+        if (typeof name !== "string")
+            die("missing <name>");
+        const versions = await reg.list(name);
+        for (const v of versions)
+            process.stdout.write(`${v}\n`);
+        return;
+    }
+    if (action === "get") {
+        const name = args.positional[0];
+        const version = args.positional[1];
+        if (typeof name !== "string")
+            die("missing <name>");
+        if (typeof version !== "string")
+            die("missing <version>");
+        process.stdout.write(await reg.get(name, version));
+        return;
+    }
+    if (action === "pin") {
+        const name = args.positional[0];
+        const env = args.positional[1];
+        const version = args.positional[2];
+        if (typeof name !== "string")
+            die("missing <name>");
+        if (typeof env !== "string")
+            die("missing <env>");
+        if (typeof version !== "string")
+            die("missing <version>");
+        if (tenantId !== undefined) {
+            await reg.pinForTenant(tenantId, name, env, version);
+            process.stdout.write(`pinned tenant=${tenantId} ${name} ${env} → ${version}\n`);
+        }
+        else {
+            await reg.pin(name, env, version);
+            process.stdout.write(`pinned ${name} ${env} → ${version}\n`);
+        }
+        return;
+    }
+    if (action === "alias") {
+        const name = args.positional[0];
+        const env = args.positional[1];
+        if (typeof name !== "string")
+            die("missing <name>");
+        if (typeof env !== "string")
+            die("missing <env>");
+        const v = tenantId !== undefined
+            ? await reg.aliasForTenant(tenantId, name, env)
+            : await reg.aliasFor(name, env);
+        if (!v)
+            die(`no pin for ${name} ${env}`);
+        process.stdout.write(`${v}\n`);
+        return;
+    }
+    die(`unknown spec action "${action}" (expected: put | list | get | pin | alias)`);
+}
+/**
+ * Section 28 — `crewhaus deploy <action> ...` subcommands wrap the
+ * deployment-controller. Actions: promote / rollback.
+ */
+async function runDeploy(args, action) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage:\n" +
+            "  crewhaus deploy promote <name> <fromEnv> <toEnv>  copy env pin\n" +
+            "  crewhaus deploy rollback <name> <env> <version>   re-pin env to version\n");
+        return;
+    }
+    const rootDirFlag = args.flags["root-dir"];
+    const rootDir = typeof rootDirFlag === "string" ? rootDirFlag : join(process.cwd(), ".crewhaus", "specs");
+    const auditDir = join(process.cwd(), ".crewhaus", "audit");
+    const { createFileBackedRegistry } = await import("@crewhaus/spec-registry");
+    const { openAuditLog } = await import("@crewhaus/audit-log");
+    const { createDeploymentController } = await import("@crewhaus/deployment-controller");
+    const reg = createFileBackedRegistry({ rootDir });
+    const audit = await openAuditLog({ rootDir: auditDir });
+    const tenantFlag = args.flags["tenant"];
+    const actorFlag = args.flags["actor"];
+    const tenantId = typeof tenantFlag === "string" ? tenantFlag : undefined;
+    const actor = typeof actorFlag === "string" ? actorFlag : undefined;
+    const ctrl = createDeploymentController({
+        registry: reg,
+        auditLog: audit,
+        ...(tenantId !== undefined ? { tenantId } : {}),
+        ...(actor !== undefined ? { actor } : {}),
+    });
+    if (action === "promote") {
+        const name = args.positional[0];
+        const fromEnv = args.positional[1];
+        const toEnv = args.positional[2];
+        if (typeof name !== "string")
+            die("missing <name>");
+        if (typeof fromEnv !== "string")
+            die("missing <fromEnv>");
+        if (typeof toEnv !== "string")
+            die("missing <toEnv>");
+        const rec = await ctrl.promote(name, fromEnv, toEnv);
+        process.stdout.write(`promoted ${name} ${fromEnv} → ${toEnv} (now pinned to ${rec.toVersion})\n`);
+        return;
+    }
+    if (action === "rollback") {
+        const name = args.positional[0];
+        const env = args.positional[1];
+        const version = args.positional[2];
+        if (typeof name !== "string")
+            die("missing <name>");
+        if (typeof env !== "string")
+            die("missing <env>");
+        if (typeof version !== "string")
+            die("missing <version>");
+        const rec = await ctrl.rollback(name, env, version);
+        process.stdout.write(`rolled back ${name} ${env} → ${version} (was ${rec.fromVersion ?? "unset"})\n`);
+        return;
+    }
+    die(`unknown deploy action "${action}" (expected: promote | rollback)`);
+}
+/**
+ * Section 28 — `crewhaus migrate-all --from <ver> --to <ver> [--dry-run]`.
+ * Walks every spec in the registry and applies the migration chain.
+ */
+async function runMigrateAll(args) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage: crewhaus migrate-all --from <ver> --to <ver> [--dry-run] [--root-dir <dir>]\n");
+        return;
+    }
+    const fromFlag = args.flags["from"];
+    const toFlag = args.flags["to"];
+    if (typeof fromFlag !== "string")
+        die("missing --from <ver>");
+    if (typeof toFlag !== "string")
+        die("missing --to <ver>");
+    const fromVersion = Number.parseInt(fromFlag, 10);
+    const toVersion = Number.parseInt(toFlag, 10);
+    if (Number.isNaN(fromVersion) || Number.isNaN(toVersion)) {
+        die("--from / --to must be integers");
+    }
+    const rootDirFlag = args.flags["root-dir"];
+    const rootDir = typeof rootDirFlag === "string" ? rootDirFlag : join(process.cwd(), ".crewhaus", "specs");
+    const dryRun = args.flags["dry-run"] === true;
+    const { createFileBackedRegistry } = await import("@crewhaus/spec-registry");
+    const { createDefaultEngine } = await import("@crewhaus/migration-engine");
+    const { migrateAll } = await import("@crewhaus/migration-runner");
+    const result = await migrateAll({
+        registry: createFileBackedRegistry({ rootDir }),
+        engine: createDefaultEngine(),
+        fromVersion,
+        toVersion,
+        dryRun,
+    });
+    for (const item of result.plan) {
+        const arrow = item.newVersion ? ` → ${item.newVersion}` : "";
+        const err = item.error ? `   ERROR: ${item.error}` : "";
+        process.stdout.write(`${item.action.padEnd(15)} ${item.name}@${item.latestVersion}${arrow}${err}\n`);
+    }
+    const dryNote = dryRun ? " (dry-run)" : "";
+    process.stdout.write(`[migrate-all] migrated=${result.migrated} skipped=${result.skipped} failed=${result.failed}${dryNote}\n`);
+    if (result.failed > 0)
+        process.exit(1);
+}
+/**
+ * Section 32 — `crewhaus build-image <target> --tag <tag> [--platform <p>] [--push]`.
+ * Wraps `docker buildx build` for the per-target Dockerfiles in
+ * @crewhaus/docker-images.
+ */
+async function runBuildImage(rest) {
+    const positional = [];
+    const flags = new Map();
+    for (let i = 0; i < rest.length; i++) {
+        const a = rest[i];
+        if (a === undefined)
+            continue;
+        if (a === "--help" || a === "-h") {
+            process.stdout.write("usage: crewhaus build-image <target> --tag <tag> [--platform <p>] [--push]\n");
+            return;
+        }
+        if (a === "--push") {
+            flags.set("push", true);
+            continue;
+        }
+        if (a === "--tag" || a === "--platform") {
+            const v = rest[i + 1];
+            if (typeof v !== "string")
+                die(`${a} requires a value`);
+            flags.set(a.slice(2), v);
+            i++;
+            continue;
+        }
+        positional.push(a);
+    }
+    const target = positional[0];
+    if (typeof target !== "string")
+        die("missing <target> (one of cli, workflow, channel, ...)");
+    const tag = flags.get("tag");
+    if (typeof tag !== "string")
+        die("missing --tag <tag>");
+    const platform = flags.get("platform");
+    const push = flags.get("push") === true;
+    const { buildImage, isTargetShape } = await import("@crewhaus/docker-images");
+    if (!isTargetShape(target)) {
+        die(`unknown target shape: ${target}`);
+    }
+    try {
+        const result = await buildImage({
+            target,
+            tag,
+            platform: typeof platform === "string" ? platform : undefined,
+            push,
+        });
+        process.stdout.write(`built crewhaus/${result.target}:${result.tag}\n`);
+    }
+    catch (err) {
+        die(`build-image: ${err.message}`);
+    }
+}
+/**
+ * Section 32 — `crewhaus cloud deploy|teardown --provider <p> --region <r>`.
+ * Composite recipe: terraform-up + helm-chart + kustomize overlay.
+ */
+async function runCloud(args, action) {
+    if (args.flags["help"]) {
+        process.stdout.write(`usage: crewhaus cloud ${action} --provider <aws|gcp|azure|aws-localstack> --region <r> [--tier <dev|default|production>] [--image-tag <tag>]\n`);
+        return;
+    }
+    const provider = args.flags["provider"];
+    const region = args.flags["region"];
+    if (typeof provider !== "string")
+        die("missing --provider");
+    if (typeof region !== "string")
+        die("missing --region");
+    const tierFlag = args.flags["tier"];
+    const imageTagFlag = args.flags["image-tag"];
+    const workingDirFlag = args.flags["working-dir"];
+    const cloudMod = await import("@crewhaus/crewhaus-cloud");
+    if (!cloudMod.isCloudProvider(provider)) {
+        die(`unknown provider: ${provider} (allowed: ${cloudMod.listProviders().join(", ")})`);
+    }
+    const config = {
+        ...cloudMod.defaultCloudConfig(provider, region),
+        ...(typeof tierFlag === "string" ? { tier: tierFlag } : {}),
+        ...(typeof imageTagFlag === "string" ? { imageTag: imageTagFlag } : {}),
+    };
+    if (action === "deploy") {
+        const result = await cloudMod.deployCloud({
+            config,
+            workingDir: typeof workingDirFlag === "string" ? workingDirFlag : undefined,
+        });
+        process.stdout.write(`${cloudMod.summariseDeploy(result)}\n`);
+        return;
+    }
+    if (action === "teardown") {
+        await cloudMod.teardownCloud({
+            config,
+            workingDir: typeof workingDirFlag === "string" ? workingDirFlag : undefined,
+        });
+        process.stdout.write(`teardown complete for ${config.clusterName}\n`);
+        return;
+    }
+    die(`unknown cloud action "${action}" (expected: deploy | teardown)`);
+}
+/**
+ * Section 34 — `crewhaus federation discover <deployment> [--srv-domain <d>] [--format json|yaml]`.
+ * Resolves a peer's endpoint + supportedShapes + publicKeyFingerprint via
+ * .well-known/crewhaus.json, optionally seeded by a DNS SRV lookup.
+ */
+async function runFederation(rest) {
+    const positional = [];
+    const flags = new Map();
+    for (let i = 0; i < rest.length; i++) {
+        const a = rest[i];
+        if (a === undefined)
+            continue;
+        if (a === "--help" || a === "-h") {
+            process.stdout.write("usage: crewhaus federation discover <deployment> [--srv-domain <d>] [--format json|yaml]\n");
+            return;
+        }
+        if (a === "--srv-domain" || a === "--format") {
+            const v = rest[i + 1];
+            if (typeof v !== "string")
+                die(`${a} requires a value`);
+            flags.set(a.slice(2), v);
+            i++;
+            continue;
+        }
+        positional.push(a);
+    }
+    const deployment = positional[0];
+    if (typeof deployment !== "string")
+        die("missing <deployment>");
+    const { discoverDeployment } = await import("@crewhaus/federation-discovery");
+    try {
+        const config = {};
+        const srv = flags.get("srv-domain");
+        if (typeof srv === "string")
+            config.srvDomain = srv;
+        const record = await discoverDeployment(deployment, config);
+        const format = flags.get("format") ?? "json";
+        if (format === "yaml") {
+            const yaml = [
+                `endpoint: ${record.endpoint}`,
+                `version: ${record.version}`,
+                "supportedShapes:",
+                ...record.supportedShapes.map((s) => `  - ${s}`),
+                `publicKeyFingerprint: ${record.publicKeyFingerprint}`,
+            ].join("\n");
+            process.stdout.write(`${yaml}\n`);
+        }
+        else {
+            process.stdout.write(`${JSON.stringify(record, null, 2)}\n`);
+        }
+    }
+    catch (err) {
+        die(`federation discover: ${err.message}`);
+    }
+}
+/**
+ * Section 39 — `crewhaus compliance evidence` collects SOC 2 / ISO 27001 /
+ * HIPAA evidence bundles by walking an audit-log root and writing signed
+ * bundles to `.crewhaus/compliance/<framework>/<controlId>/<period>.json`.
+ *
+ * Usage:
+ *   crewhaus compliance evidence --framework soc2 --period 2026-Q2 \
+ *     [--control CC6.1] [--audit-dir <d>] [--out-dir <d>] \
+ *     [--signing-key-env CREWHAUS_COMPLIANCE_SIGNING_KEY]
+ */
+async function runCompliance(args, action) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage:\n" +
+            "  crewhaus compliance evidence --framework <id> [--control <id>] --period <p>\n" +
+            "    [--audit-dir <d>] [--out-dir <d>] [--signing-key-env <ENV>]\n");
+        return;
+    }
+    if (action !== "evidence") {
+        die(`unknown compliance action "${action}" (expected: evidence)`);
+    }
+    const frameworkFlag = args.flags["framework"];
+    const periodFlag = args.flags["period"];
+    if (typeof frameworkFlag !== "string")
+        die("--framework <id> is required");
+    if (typeof periodFlag !== "string")
+        die("--period <p> is required");
+    const auditDirFlag = args.flags["audit-dir"];
+    const outDirFlag = args.flags["out-dir"];
+    const auditDir = typeof auditDirFlag === "string" ? auditDirFlag : join(process.cwd(), ".crewhaus", "audit");
+    const outDir = typeof outDirFlag === "string" ? outDirFlag : join(process.cwd(), ".crewhaus", "compliance");
+    const signingKeyEnv = args.flags["signing-key-env"];
+    const signingKey = typeof signingKeyEnv === "string" ? (process.env[signingKeyEnv] ?? undefined) : undefined;
+    if (typeof signingKeyEnv === "string" && signingKey === undefined) {
+        die(`signing key env var "${signingKeyEnv}" is not set`);
+    }
+    const { createComplianceCollector } = await import("@crewhaus/compliance-controls");
+    const auditLog = await import("@crewhaus/audit-log");
+    const auditSource = {
+        async *read() {
+            // Use the verify shape to walk every record without re-running hash chain
+            // logic — auditLog.openAuditLog().read() is the supported public API.
+            const log = await auditLog.openAuditLog({ rootDir: auditDir });
+            for await (const r of log.read()) {
+                yield r;
+            }
+        },
+    };
+    const collector = createComplianceCollector({ auditSource, outputDir: outDir });
+    const controlFlag = args.flags["control"];
+    let bundles;
+    if (typeof controlFlag === "string") {
+        const single = await collector.collect(frameworkFlag, controlFlag, {
+            period: periodFlag,
+            ...(signingKey !== undefined ? { signingKey } : {}),
+        });
+        bundles = [single];
+    }
+    else {
+        bundles = await collector.collectAll(frameworkFlag, {
+            period: periodFlag,
+            ...(signingKey !== undefined ? { signingKey } : {}),
+        });
+    }
+    for (const b of bundles) {
+        const path = collector.writeBundle(b);
+        process.stdout.write(`${b.frameworkId}/${b.controlId} ${periodFlag}: ${b.recordCount} records → ${path}\n`);
+    }
+}
+/**
+ * Section 36 — `crewhaus sandbox <action>`. Single action today: `doctor`.
+ *   doctor              list registered sandbox images + healthcheck status
+ *
+ * `--probe` runs each registered image's healthcheck argv via the configured
+ * sandbox backend (docker / podman / noop). Without `--probe` the command
+ * just snapshots the in-process registry — useful for verifying that the
+ * polyglot images are wired up after a fresh `crewhaus install`.
+ */
+async function runSandbox(args, action) {
+    if (args.flags["help"]) {
+        process.stdout.write("usage:\n" + "  crewhaus sandbox doctor [--probe] [--format json|table]\n");
+        return;
+    }
+    if (action !== "doctor")
+        die(`unknown sandbox action "${action}" (expected: doctor)`);
+    const formatFlag = args.flags["format"];
+    const format = typeof formatFlag === "string" ? formatFlag : "table";
+    if (format !== "json" && format !== "table") {
+        die(`--format must be "json" or "table" (got "${format}")`);
+    }
+    const wantProbe = args.flags["probe"] === true;
+    const reg = await import("@crewhaus/sandbox-image-registry");
+    // Touch the registry so the §18 trio bootstraps.
+    reg.listSandboxImages();
+    let statuses = reg.snapshotImageStatuses();
+    if (wantProbe) {
+        const sandboxMod = await import("@crewhaus/sandbox");
+        const sandbox = sandboxMod.createSandbox({
+            allowedImages: reg.listAllowedImageRefs(),
+        });
+        statuses = await reg.runHealthchecks(async (entry) => {
+            const result = await sandbox.exec({
+                image: entry.image,
+                argv: [...entry.healthcheck.command],
+                timeoutMs: entry.healthcheck.timeoutMs ?? 5_000,
+            });
+            return { exitCode: result.exitCode, stderr: result.stderr };
+        });
+        await sandbox.close();
+    }
+    if (format === "json") {
+        process.stdout.write(`${JSON.stringify(statuses, null, 2)}\n`);
+        return;
+    }
+    process.stdout.write("registered sandbox images:\n");
+    const summaries = reg.listSandboxImages();
+    const byId = new Map(statuses.map((s) => [s.id, s]));
+    for (const e of summaries) {
+        const s = byId.get(e.id);
+        const mark = s?.healthy ? "✓" : wantProbe ? "✗" : "•";
+        const tail = s?.healthy
+            ? `last healthy ${s.lastHealthyAt}`
+            : (s?.lastError ?? (wantProbe ? "no probe result" : "not yet probed"));
+        process.stdout.write(`  ${mark} ${e.id.padEnd(10)} ${e.image.padEnd(28)} ${tail}\n`);
+    }
+}
+const argv = process.argv.slice(2);
+const subcommand = argv[0] ?? "";
+const rest = argv.slice(1);
+switch (subcommand) {
+    case "compile":
+        await runCompile(parseFor(rest, COMPILE_SCHEMA));
+        break;
+    case "init":
+        runInit(parseFor(rest, INIT_SCHEMA));
+        break;
+    case "run":
+        // Mirror runCompile's policy: every structured failure in the run
+        // pipeline (ConfigError from the model-router, ProviderAuthError from
+        // an adapter, RuntimeError, …) extends CrewhausError — route the
+        // family through die() for a clean one-line error + exit 1 instead of
+        // a raw stack trace. A non-CrewhausError (a genuine bug) still
+        // propagates with its full stack for debugging.
+        try {
+            await runRun(parseFor(rest, RUN_SCHEMA));
+        }
+        catch (err) {
+            if (err instanceof CrewhausError)
+                die(err.message);
+            throw err;
+        }
+        break;
+    case "eval":
+        await runEvalSubcommand(parseFor(rest, EVAL_SCHEMA));
+        break;
+    case "eval-report":
+        await runEvalReport(parseFor(rest, EVAL_REPORT_SCHEMA));
+        break;
+    case "optimize":
+        await runOptimize(parseFor(rest, OPTIMIZE_SCHEMA));
+        break;
+    case "doctor":
+        await runDoctor(parseFor(rest, DOCTOR_SCHEMA));
+        break;
+    case "context":
+        runContext(parseFor(rest, CONTEXT_SCHEMA));
+        break;
+    case "cost-summary":
+        await runCostSummary(parseFor(rest, COST_SUMMARY_SCHEMA));
+        break;
+    case "secrets": {
+        const action = rest[0] ?? "";
+        if (action !== "doctor" && action !== "rotate") {
+            die(`secrets action must be "doctor" or "rotate" (got "${action}")`);
+        }
+        await runSecrets(parseFor(rest.slice(1), SECRETS_SCHEMA), action);
+        break;
+    }
+    case "spec": {
+        const action = rest[0] ?? "";
+        if (!["put", "list", "get", "pin", "alias"].includes(action)) {
+            die(`spec action must be one of: put, list, get, pin, alias (got "${action}")`);
+        }
+        await runSpec(parseFor(rest.slice(1), SPEC_SCHEMA), action);
+        break;
+    }
+    case "deploy": {
+        const action = rest[0] ?? "";
+        if (action !== "promote" && action !== "rollback") {
+            die(`deploy action must be "promote" or "rollback" (got "${action}")`);
+        }
+        await runDeploy(parseFor(rest.slice(1), DEPLOY_SCHEMA), action);
+        break;
+    }
+    case "migrate-all":
+        await runMigrateAll(parseFor(rest, MIGRATE_SCHEMA));
+        break;
+    case "build-image":
+        await runBuildImage(rest);
+        break;
+    case "cloud": {
+        const action = rest[0] ?? "";
+        if (action !== "deploy" && action !== "teardown") {
+            die(`cloud action must be "deploy" or "teardown" (got "${action}")`);
+        }
+        await runCloud(parseFor(rest.slice(1), CLOUD_SCHEMA), action);
+        break;
+    }
+    case "federation": {
+        const action = rest[0] ?? "";
+        if (action !== "discover") {
+            die(`federation action must be "discover" (got "${action}")`);
+        }
+        await runFederation(rest.slice(1));
+        break;
+    }
+    case "sandbox": {
+        const action = rest[0] ?? "";
+        if (action !== "doctor") {
+            die(`sandbox action must be "doctor" (got "${action}")`);
+        }
+        await runSandbox(parseFor(rest.slice(1), SANDBOX_SCHEMA), action);
+        break;
+    }
+    case "compliance": {
+        const action = rest[0] ?? "";
+        if (action !== "evidence") {
+            die(`compliance action must be "evidence" (got "${action}")`);
+        }
+        await runCompliance(parseFor(rest.slice(1), COMPLIANCE_SCHEMA), action);
+        break;
+    }
+    case "version":
+    case "-v":
+    case "--version":
+        printVersion();
+        break;
+    case "-h":
+    case "--help":
+        help();
+        break;
+    case "":
+        usage();
+        break;
+    default:
+        die(`unknown subcommand: ${subcommand}`);
+}