@robmclarty/volley 0.7.0 → 0.9.0

package/dist/index.js

@@ -1,10 +1,11 @@
 import { accessSync, constants, existsSync, mkdirSync, readFileSync, readdirSync, renameSync, rmSync, statSync, symlinkSync, writeFileSync } from "node:fs";
 import { basename, dirname, isAbsolute, join, relative, resolve, sep } from "node:path";
 import { fileURLToPath, pathToFileURL } from "node:url";
-import { randomUUID } from "node:crypto";
+import { createHmac, randomBytes, randomUUID } from "node:crypto";
 import { spawn, spawnSync } from "node:child_process";
 import { aborted_error, branch, claude_cli_reported, create_engine, fallback, loop, model_call, pipe, retry, run, sequence, step, throughput } from "fascicle";
-import { filesystem_logger } from "fascicle/adapters";
+import { homedir } from "node:os";
+import { filesystem_logger, tee_logger } from "fascicle/adapters";
 import { lookup } from "node:dns";
 import { Readability } from "@mozilla/readability";
 import ipaddr from "ipaddr.js";
@@ -13,6 +14,31 @@ import TurndownService from "turndown";
 import { Agent, buildConnector, interceptors, request } from "undici";
 import { z } from "zod";
 import pc from "picocolors";
+//#region src/types.ts
+function config_error(message) {
+	return Object.assign(new Error(message), { kind: "config_error" });
+}
+function check_error(message) {
+	return Object.assign(new Error(message), { kind: "check_error" });
+}
+/** Wraps a provider/schema failure with the phase and iteration it hit. */
+function phase_error(phase, iteration, cause) {
+	const detail = cause instanceof Error ? cause.message : String(cause);
+	return Object.assign(/* @__PURE__ */ new Error(`${phase} failed on iteration ${iteration}: ${detail}`), {
+		kind: "phase_error",
+		phase,
+		iteration,
+		cause
+	});
+}
+function error_kind(err) {
+	if (typeof err === "object" && err !== null && "kind" in err) {
+		const kind = err.kind;
+		return typeof kind === "string" ? kind : null;
+	}
+	return null;
+}
+//#endregion
 //#region src/changes.ts
 /**
 * What the builder changed, and whether it touched the gate.
@@ -117,6 +143,29 @@ function capture_baseline(root) {
 	const sha = head.stdout.trim();
 	return sha.length > 0 ? sha : null;
 }
+/**
+* The commit a critique's change is measured from: `rev` (`--baseline`)
+* resolved in `root`, so the change is everything that differs from it,
+* committed or not, untracked files included. With no revision named it's
+* HEAD, so the uncommitted and untracked edits are the change, and, as for a
+* run, null where there's no repository or no commit yet. A named revision
+* that isn't a commit there, or one named where there's no repository, is a
+* config error.
+*/
+function resolve_baseline(root, rev) {
+	if (rev === void 0) return capture_baseline(root);
+	if (!in_git_repo(root)) throw config_error(`--baseline needs the workspace to be a git repository: ${root}`);
+	const resolved = git$1(root, [
+		"rev-parse",
+		"--verify",
+		"--quiet",
+		"--end-of-options",
+		`${rev}^{commit}`
+	]);
+	const sha = resolved.stdout.trim();
+	if (resolved.status !== 0 || sha.length === 0) throw config_error(`--baseline names no commit in ${root}: ${rev}`);
+	return sha;
+}
 /** git's status letters, folded onto the four cases worth naming. Rename and
 * copy both record the destination path — where the content lives now. */
 function change_status(code) {
@@ -153,7 +202,8 @@ function parse_name_status(stdout) {
 function parse_paths(stdout) {
 	return stdout.split("\0").filter((path) => path.length > 0);
 }
-function excluded(path) {
+function excluded(path, excluded_dir) {
+	if (excluded_dir !== null && path.startsWith(`${excluded_dir}/`)) return true;
 	return EXCLUDED_PREFIXES.some((prefix) => path.startsWith(prefix));
 }
 /**
@@ -233,10 +283,11 @@ function collect_changes(options) {
 		path,
 		status: "added"
 	}));
+	const excluded_dir = options.excluded_dir ?? null;
 	const seen = /* @__PURE__ */ new Set();
 	const files = [];
 	for (const change of [...tracked, ...untracked]) {
-		if (excluded(change.path) || seen.has(change.path)) continue;
+		if (excluded(change.path, excluded_dir) || seen.has(change.path)) continue;
 		seen.add(change.path);
 		files.push({
 			...change,
@@ -278,29 +329,23 @@ function resolve_check_runner(check, workspace) {
 	return /(^|\s)checkride(\s|$)/.test(check.trim()) ? "checkride" : "command";
 }
 //#endregion
-//#region src/types.ts
-function config_error(message) {
-	return Object.assign(new Error(message), { kind: "config_error" });
-}
-function check_error(message) {
-	return Object.assign(new Error(message), { kind: "check_error" });
-}
-/** Wraps a provider/schema failure with the phase and iteration it hit. */
-function phase_error(phase, iteration, cause) {
-	const detail = cause instanceof Error ? cause.message : String(cause);
-	return Object.assign(/* @__PURE__ */ new Error(`${phase} failed on iteration ${iteration}: ${detail}`), {
-		kind: "phase_error",
-		phase,
-		iteration,
-		cause
-	});
-}
-function error_kind(err) {
-	if (typeof err === "object" && err !== null && "kind" in err) {
-		const kind = err.kind;
-		return typeof kind === "string" ? kind : null;
-	}
-	return null;
+//#region src/containment.ts
+/**
+* Whether volley is running inside its sandbox container. A leaf, so the
+* config resolution, the preflight, and the run-config signing can all ask
+* without importing one another.
+*/
+/** volley *detects* it is running inside its sandbox container rather than
+* starting the container itself. The signal is the `VOLLEY_CONTAINED=1` marker
+* volley's own image bakes in (`Dockerfile` `ENV`), so a `docker run <hardened
+* flags> <image> volley run …` is recognized as contained out of the box. Kept
+* env-only — no ambient `/.dockerenv` probe — so the safety gate is deterministic
+* and testable, and so a CI job that itself runs in an unrelated container does
+* not silently satisfy the gate; a hand-rolled sandbox must set the marker (or
+* the operator uses `--allow-unsandboxed-builder`). */
+function detect_containment(env) {
+	const value = env["VOLLEY_CONTAINED"];
+	return value === "1" || value === "true";
 }
 //#endregion
 //#region src/engine.ts
@@ -315,6 +360,13 @@ function error_kind(err) {
 * and 404s. */
 const DEFAULT_OLLAMA_URL = "http://localhost:11434";
 const DEFAULT_LMSTUDIO_URL = "http://localhost:1234/v1";
+/** OpenRouter's API root, the one fascicle's openrouter adapter defaults to. */
+const DEFAULT_OPENROUTER_URL = "https://openrouter.ai/api/v1";
+/** The variable an `openrouter` seat's API key is read from. */
+const OPENROUTER_KEY_VAR = "OPENROUTER_API_KEY";
+/** The variable a `claude_cli` seat's API key is read from, which the Claude
+* CLI also reads on its own. */
+const ANTHROPIC_KEY_VAR = "ANTHROPIC_API_KEY";
 /** Loopback authorities the container→host crossing rewrites. A base URL
 * pointing at one of these is local to wherever volley runs; contained, that is
 * *inside* the sandbox container, where the host LLM daemon is unreachable
@@ -361,9 +413,15 @@ const SEAT_URL_SOURCES = {
 		env: "VOLLEY_LMSTUDIO_URL",
 		default_url: DEFAULT_LMSTUDIO_URL,
 		normalize: strip_trailing_slashes
+	},
+	openrouter: {
+		env: "VOLLEY_OPENROUTER_URL",
+		default_url: DEFAULT_OPENROUTER_URL,
+		normalize: strip_trailing_slashes
 	}
 };
-/** The providers that take a seat URL (`--builder-url` / `--critic-url`). */
+/** The providers that take a seat URL (`--builder-url` / `--critic-url`), and
+* so get their endpoint probed by `--dry-run`. */
 const SEAT_URL_PROVIDERS = Object.keys(SEAT_URL_SOURCES);
 /** Whether a seat on `provider` reaches its model at a base URL volley
 * configures, and so takes a URL of its own. */
@@ -376,16 +434,25 @@ function present(value) {
 	return trimmed === "" ? null : trimmed;
 }
 /** The base URL a seat's model calls go to: the seat's own URL, else the
-* provider's env var (`VOLLEY_OLLAMA_URL` / `VOLLEY_LMSTUDIO_URL`), else its
-* default. A seat never falls back to the other seat's URL. The result is
-* normalized, then a loopback authority is crossed to the host LLM endpoint
-* when volley runs contained. Both seats can be loopback URLs on different
-* ports, and each crosses on its own port. */
+* provider's env var (`VOLLEY_OLLAMA_URL` / `VOLLEY_LMSTUDIO_URL` /
+* `VOLLEY_OPENROUTER_URL`), else its default. A seat never falls back to the
+* other seat's URL. The result is normalized, then a loopback authority is
+* crossed to the host LLM endpoint when volley runs contained. Both seats can
+* be loopback URLs on different ports, and each crosses on its own port. */
 function seat_base_url(provider, seat_url, env) {
 	const source = SEAT_URL_SOURCES[provider];
 	const raw = present(seat_url) ?? present(env[source.env]) ?? source.default_url;
 	return cross_to_host_gateway(source.normalize(raw), env);
 }
+/** The API key an `openrouter` seat sends, from `OPENROUTER_API_KEY`. An unset
+* or blank key is a config error naming the variable, which `resolve_config`
+* raises before any work, because fascicle's own "requires a non-empty
+* api_key" doesn't say where the key comes from. */
+function openrouter_api_key(env) {
+	const key = present(env[OPENROUTER_KEY_VAR]);
+	if (key === null) throw config_error(`${OPENROUTER_KEY_VAR} is not set: an openrouter seat sends it as its API key`);
+	return key;
+}
 function load_pricing_overrides(path) {
 	if (!existsSync(path)) throw config_error(`VOLLEY_PRICING_PATH not found: ${path}`);
 	try {
@@ -398,29 +465,44 @@ function load_pricing_overrides(path) {
 * peer (`ai-sdk-ollama` / `@ai-sdk/openai-compatible`) is loaded lazily by
 * fascicle only when the provider actually runs. The entry is complete, because
 * `Engine.with_providers` replaces a provider's whole entry: whatever this
-* returns is all a seat's derived engine knows about the provider.
+* returns is all a seat's derived engine knows about the provider, so an
+* `openrouter` entry carries its key and its transport along with its URL.
+*
+* `openrouter` runs on fascicle's native transport, its OpenAI-compatible core,
+* so no ai-sdk peer gets installed for it, and its `X-Title` attributes the
+* traffic to volley.
 *
-* Native transport is the deferred future bridge, not adopted here: volley stays
-* on the ai_sdk transport, which fascicle 0.14.0 already selects by default
-* (`transport?: 'ai_sdk' | 'native'` defaults to `'ai_sdk'`), so no field is
-* needed today. Flipping later is a one-line change — add `transport: 'native'`
-* to the ollama/lmstudio entry returned below, which both seats' engines then
-* share. No URL re-pointing is required: `seat_base_url` already yields the
-* daemon-root URL the native `/api/chat` transport wants (its adapter appends
-* `/api/chat` itself), which is exactly what the v0.3.1 base-URL fix produces.
+* For the local providers, native transport is the deferred future bridge, not
+* adopted here: they stay on the ai_sdk transport, which fascicle 0.14.0 already
+* selects by default (`transport?: 'ai_sdk' | 'native'` defaults to `'ai_sdk'`),
+* so no field is needed today. Flipping later is a one-line change — add
+* `transport: 'native'` to the ollama/lmstudio entry returned below, which both
+* seats' engines then share. No URL re-pointing is required: `seat_base_url`
+* already yields the daemon-root URL the native `/api/chat` transport wants
+* (its adapter appends `/api/chat` itself), which is exactly what the v0.3.1
+* base-URL fix produces.
 * Kept as a proven-once-then-reverted option so the transport stays a clean
 * second variable in the model-vs-transport comparison. Flipping it today would
 * also lose the reasoning: fascicle's native Ollama adapter drops
 * `message.thinking`, with `think` unset or true (robmclarty/fascicle#12), so a
 * thinking seat's reasoning would vanish from the terminal and the trajectory.
 *
-* Both providers' base URLs cross a loopback authority to `VOLLEY_MODEL_HOST` when
-* volley runs contained, so an in-container model client reaches the
+* Every provider's base URL crosses a loopback authority to `VOLLEY_MODEL_HOST`
+* when volley runs contained, so an in-container model client reaches the
 * host LLM endpoint at `host.docker.internal`; on the host / unsandboxed path the
 * override is unset and this is a no-op. */
 function seat_provider_config(provider, seat_url, env) {
 	const base_url = seat_base_url(provider, seat_url, env);
-	return provider === "ollama" ? { ollama: { base_url } } : { lmstudio: { base_url } };
+	switch (provider) {
+		case "ollama": return { ollama: { base_url } };
+		case "lmstudio": return { lmstudio: { base_url } };
+		case "openrouter": return { openrouter: {
+			api_key: openrouter_api_key(env),
+			base_url,
+			transport: "native",
+			x_title: "volley"
+		} };
+	}
 }
 /** Per-run engine. `claude_cli` is always configured with `options.workspace`
 * (the run's build root) as its session cwd; any URL-taking provider selected
@@ -440,7 +522,7 @@ function create_volley_engine(options) {
 		setting_sources: ["project", "local"],
 		auth_mode,
 		...binary !== void 0 && binary.length > 0 ? { binary } : {},
-		...auth_mode === "api_key" && env["ANTHROPIC_API_KEY"] !== void 0 ? { api_key: env["ANTHROPIC_API_KEY"] } : {}
+		...auth_mode === "api_key" && env[ANTHROPIC_KEY_VAR] !== void 0 ? { api_key: env[ANTHROPIC_KEY_VAR] } : {}
 	} };
 	if (takes_seat_url(options.critic_provider)) Object.assign(providers, seat_provider_config(options.critic_provider, options.critic_url ?? null, env));
 	if (takes_seat_url(options.builder_provider)) Object.assign(providers, seat_provider_config(options.builder_provider, options.builder_url ?? null, env));
@@ -479,6 +561,264 @@ async function dispose_seat_engines(engines) {
 	const distinct = engines.critic === engines.builder ? [engines.builder] : [engines.builder, engines.critic];
 	await Promise.all(distinct.map((engine) => engine.dispose()));
 }
+//#endregion
+//#region src/trust.ts
+/**
+* Which run configs a host-side resume may trust. A resume runs what a run's
+* `config.json` says: its check, a shell command, and its seats, permission
+* mode and prompt included. Inside the sandbox the builder runs arbitrary
+* commands with write access to its own run's state, so a contained run's
+* config is the builder's to rewrite, and resuming it on the host would run
+* whatever the builder left there.
+*
+* So volley on the host signs each config it writes, an HMAC under a key of
+* this machine's that never enters the container, and a host-side resume
+* trusts only a config whose signature holds. Inside the sandbox nothing is
+* signed, since the key isn't there and one made there would be in the
+* builder's reach, so a contained run's config never carries a signature the
+* host accepts, and the builder can't forge one without the key.
+*/
+/** Prefixed to what is signed, so a signature over a run config can't pass for
+* one over anything else under the same key. */
+const DOMAIN = "volley run config v1\n";
+/** Where this machine's key lives: `$XDG_CONFIG_HOME/volley/run-key`, else
+* `~/.config/volley/run-key`, outside every state root, so a container a run
+* root is mounted into still can't read it. A relative `XDG_CONFIG_HOME` is
+* ignored, as the XDG spec says. */
+function trust_key_path(env, home = homedir()) {
+	const xdg = env["XDG_CONFIG_HOME"];
+	const base = xdg !== void 0 && isAbsolute(xdg) ? xdg : join(home, ".config");
+	return join(base, "volley", "run-key");
+}
+function error_code(err) {
+	return err.code;
+}
+/** The key as stored, or null when there is none or it isn't one. */
+function read_key(path) {
+	try {
+		const hex = readFileSync(path, "utf8").trim();
+		return /^[0-9a-f]{64}$/.test(hex) ? Buffer.from(hex, "hex") : null;
+	} catch {
+		return null;
+	}
+}
+/** The key, created on first use: 32 random bytes in a file only its user can
+* read, in a directory only its user can enter. Two runs that start together
+* race to create it; the loser reads the winner's. */
+function machine_key(env) {
+	const path = trust_key_path(env);
+	const existing = read_key(path);
+	if (existing !== null) return existing;
+	try {
+		mkdirSync(dirname(path), {
+			recursive: true,
+			mode: 448
+		});
+		writeFileSync(path, `${randomBytes(32).toString("hex")}\n`, {
+			flag: "wx",
+			mode: 384
+		});
+	} catch (err) {
+		if (error_code(err) !== "EEXIST") return null;
+	}
+	return read_key(path);
+}
+function hmac(key, config_text) {
+	return createHmac("sha256", key).update(DOMAIN).update(config_text).digest("hex");
+}
+/** This machine's signature over a run's `config.json` text, or null where
+* there's none to give: inside the sandbox, or when the key can't be read or
+* created. */
+function sign_config(config_text, env = process.env) {
+	if (detect_containment(env)) return null;
+	const key = machine_key(env);
+	return key === null ? null : hmac(key, config_text);
+}
+//#endregion
+//#region src/workspace.ts
+/**
+* Workspace lifecycle: run-state initialization, stale-run
+* backup rotation, path helpers, and optional git checkpoints.
+*/
+/** A run id names the run's `volley/<run id>` branch and its state directory,
+* so it is held to what both take without quoting: letters, digits, `-`, and
+* `_`, starting with a letter or digit, so it never reads as an option and
+* never climbs out of the directory it names. */
+const RUN_ID_PATTERN = /^[A-Za-z0-9][A-Za-z0-9_-]*$/;
+/** The root each run's state sits under when `--state-dir` names none:
+* `$XDG_STATE_HOME/volley`, else `<home>/.local/state/volley`. A relative
+* `XDG_STATE_HOME` is ignored, as the XDG spec says. `home` is the host's by
+* default; the sandbox passes its container's to spell the same root inside. */
+function default_state_root(env, home = homedir()) {
+	const xdg = env["XDG_STATE_HOME"];
+	const base = xdg !== void 0 && isAbsolute(xdg) ? xdg : join(home, ".local", "state");
+	return join(base, "volley");
+}
+/** Where a run keeps its config, summary, trajectory, feedback, verdict, and
+* iterations: `<state_dir>/<run_id>/` under its state root, or, in the layout
+* older runs used (`state_dir` null), `<workspace>/.volley/`. */
+function run_state_dir(run) {
+	return run.state_dir === null ? join(run.workspace, ".volley") : join(run.state_dir, run.run_id);
+}
+function state_path(run, ...parts) {
+	return join(run_state_dir(run), ...parts);
+}
+function iteration_dir(run, iteration) {
+	return state_path(run, "iterations", String(iteration).padStart(3, "0"));
+}
+/** A state root that sits inside `root` (the build root), as the path from
+* there in git's `/`-separated vocabulary, or null when it sits outside. A
+* root inside the tree would otherwise pass for the builder's work: listed in
+* the changes the critic reads, and committed by every `--git` checkpoint.
+* The legacy `.volley/` (`state_dir` null) is excluded by name already. */
+function state_root_within(root, state_dir) {
+	if (state_dir === null) return null;
+	const path = relative(resolve(root), resolve(state_dir));
+	if (path === "" || path === ".." || path.startsWith(`..${sep}`) || isAbsolute(path)) return null;
+	return path.split(sep).join("/");
+}
+/** Filesystem-safe timestamp for `.bak.<stamp>` rotation names. Shared by the
+* run-state rotation here and the worktree rotation so both preserve
+* prior state under the same never-destroy naming. */
+function backup_stamp(date = /* @__PURE__ */ new Date()) {
+	return date.toISOString().replace(/[:.]/g, "-");
+}
+/** Fresh-run initialization: rotate a stale state directory aside (a prior
+* `.volley/` becomes `.volley.bak.<stamp>`), then create the directory tree.
+* Under a state root each run has a directory of its own, so only a reused
+* `--run-id` ever finds one to rotate. An empty one is used as it is: that is
+* what a contained run is handed when its run directory is the mount, which
+* can't be renamed. Resume runs skip rotation (`preserve: true`). */
+function initialize_run_state(run, options = {}) {
+	const dir = run_state_dir(run);
+	if (existsSync(dir) && options.preserve !== true && readdirSync(dir).length > 0) renameSync(dir, `${dir}.bak.${backup_stamp()}`);
+	if (run.state_dir !== null && !existsSync(run.state_dir)) mkdirSync(run.state_dir, {
+		recursive: true,
+		mode: 448
+	});
+	mkdirSync(join(dir, "iterations"), { recursive: true });
+}
+/** Beside a run's `config.json`: this machine's signature over it
+* (`./trust.ts`), which a host-side resume checks before it trusts the
+* config. */
+const CONFIG_SIGNATURE = "config.sig";
+/** Record the run's config, signed when volley runs on the host. Returns
+* whether it was signed: inside the sandbox it never is, and on the host only
+* when the machine's key can't be read or created. A signature a reused run
+* directory held from before is removed rather than left to vouch for this
+* config. */
+function write_resolved_config(config, env = process.env) {
+	const record = {
+		version: config.version,
+		run_id: config.run_id,
+		started_at: config.started_at,
+		mode: config.mode,
+		prompt: config.prompt,
+		criteria: config.criteria,
+		check: config.check,
+		check_resolved: config.check_resolved,
+		builder_model: config.builder_model,
+		builder_provider: config.builder_provider,
+		builder_url: config.builder_url,
+		builder_max_steps: config.builder_max_steps,
+		builder_max_tool_calls: config.builder_max_tool_calls,
+		critic_model: config.critic_model,
+		critic_provider: config.critic_provider,
+		critic_url: config.critic_url,
+		builder_permission_mode: config.builder_permission_mode,
+		critic_preset: config.critic_preset,
+		critic_prompt_path: config.critic_prompt_path,
+		max_iterations: config.max_iterations,
+		max_cost_usd: config.max_cost_usd,
+		git_checkpoints: config.git_checkpoints,
+		worktree: config.worktree,
+		discard_worktree: config.discard_worktree,
+		gate_paths: config.gate_paths,
+		fail_on_gate_edit: config.fail_on_gate_edit,
+		sandbox_image: config.sandbox_image,
+		state_dir: config.state_dir,
+		workspace: config.workspace
+	};
+	const text = `${JSON.stringify(record, null, 2)}\n`;
+	writeFileSync(state_path(config, "config.json"), text);
+	const signature = sign_config(text, env);
+	if (signature === null) {
+		rmSync(state_path(config, CONFIG_SIGNATURE), { force: true });
+		return false;
+	}
+	writeFileSync(state_path(config, CONFIG_SIGNATURE), `${signature}\n`);
+	return true;
+}
+/**
+* `--git`: commit everything under `root` after a phase (`git -C <root>`, run
+* here via the spawn `cwd`). `root` is the build root (`build_root`): the
+* workspace on the default path, or the per-run worktree under `--worktree`, so
+* checkpoints land on the worktree branch rather than the workspace's.
+* A checkpoint with nothing to commit is fine; a missing git binary or repo is
+* not. `excluded_dir` (from `state_root_within`) keeps a state root inside the
+* tree out of the commit; `:/` keeps the rest of the staging repo-wide, as a
+* bare `git add -A` is, and `literal` reads the path as written, wildcards and
+* all.
+*/
+function git_checkpoint(root, message, excluded_dir = null) {
+	if (!existsSync(join(root, ".git"))) throw config_error(`git checkpoints require a git repository: ${root}`);
+	const pathspec = excluded_dir === null ? [] : [
+		"--",
+		":/",
+		`:(exclude,literal)${excluded_dir}`
+	];
+	const add = spawnSync("git", [
+		"add",
+		"-A",
+		...pathspec
+	], { cwd: root });
+	if (add.error !== void 0 || add.status !== 0) throw check_error(`git add failed: ${add.stderr?.toString() ?? String(add.error)}`);
+	const commit = spawnSync("git", [
+		"commit",
+		"-m",
+		message,
+		"--no-verify"
+	], { cwd: root });
+	if (commit.error !== void 0) throw check_error(`git commit failed to start: ${String(commit.error)}`);
+	const out = `${commit.stdout?.toString() ?? ""}${commit.stderr?.toString() ?? ""}`;
+	if (commit.status !== 0 && !/nothing (?:added )?to commit/.test(out)) throw check_error(`git commit failed: ${out.trim()}`);
+}
+/**
+* Integration — bring a successful `--worktree --git` run's effects onto the
+* workspace branch. The per-phase checkpoints on `branch` are the raw
+* audit/replay trail; here they collapse into a single squash commit
+* (`git merge --squash` then `commit`) on the workspace's current branch. The
+* shared object store makes this a local, fetch-free merge, and — because the
+* workspace branch never moved during the run (checkpoints went to the phase
+* branch) — it is conflict-free by construction. A run whose phase branch has no
+* new commits integrates to a no-op. Only successful runs integrate; an
+* abandoned run's branch is discarded wholesale by the teardown trio instead.
+*/
+function integrate_worktree(workspace, branch, message) {
+	if (!existsSync(join(workspace, ".git"))) throw config_error(`--worktree integration requires a git repository: ${workspace}`);
+	const squash = spawnSync("git", [
+		"merge",
+		"--squash",
+		branch
+	], {
+		cwd: workspace,
+		encoding: "utf8"
+	});
+	if (squash.error !== void 0) throw check_error(`git merge --squash failed to start: ${String(squash.error)}`);
+	if (squash.status !== 0) throw check_error(`git merge --squash ${branch} failed: ${(squash.stderr || squash.stdout).trim()}`);
+	const commit = spawnSync("git", [
+		"commit",
+		"-m",
+		message,
+		"--no-verify"
+	], {
+		cwd: workspace,
+		encoding: "utf8"
+	});
+	if (commit.error !== void 0) throw check_error(`git commit failed to start: ${String(commit.error)}`);
+	const out = `${commit.stdout ?? ""}${commit.stderr ?? ""}`;
+	if (commit.status !== 0 && !out.includes("nothing to commit")) throw check_error(`git commit failed: ${out.trim()}`);
+}
 const DEFAULT_CRITIC_MODEL = "opus";
 const DEFAULT_CRITIC = "reviewer";
 const DEFAULT_BUILDER_PROVIDER = "claude_cli";
@@ -490,18 +830,24 @@ const DEFAULT_SANDBOX_IMAGE = "volley-sandbox:latest";
 const BUILDER_PROVIDERS = [
 	"claude_cli",
 	"ollama",
-	"lmstudio"
+	"lmstudio",
+	"openrouter"
+];
+/** Builder providers that drive their model through volley's own tool loop —
+* a real `bash` (write + exec), wherever the model runs. Refused unless volley
+* is running inside a container (the correct blast-radius default) or the
+* operator explicitly opts out; `claude_cli` has its own permission model and
+* is exempt. */
+const TOOL_LOOP_BUILDER_PROVIDERS = [
+	"ollama",
+	"lmstudio",
+	"openrouter"
 ];
-/** Builder providers that run a local model through volley's own tool loop —
-* a real `bash` (write + exec). Refused unless volley is running
-* inside a container (the correct blast-radius default) or the operator
-* explicitly opts out; `claude_cli` has its own permission model and is
-* exempt. */
-const LOCAL_BUILDER_PROVIDERS = ["ollama", "lmstudio"];
 const CRITIC_PROVIDERS = [
 	"claude_cli",
 	"ollama",
-	"lmstudio"
+	"lmstudio",
+	"openrouter"
 ];
 const CRITIC_PRESETS = [
 	"reviewer",
@@ -555,17 +901,49 @@ function resolve_critic(critic, cwd) {
 		critic_prompt_path: path
 	};
 }
-/** `--builder-max-steps` (a merged flag/config value) wins over the
-* `VOLLEY_BUILDER_MAX_STEPS` env var, which wins over the default. Local
-* builder only; ignored by the `claude_cli` arm. */
-function resolve_builder_max_steps(raw, env) {
-	if (raw.builder_max_steps !== void 0) return validate_builder_max_steps(raw.builder_max_steps, raw.builder_max_steps);
-	const env_value = env["VOLLEY_BUILDER_MAX_STEPS"];
-	if (env_value !== void 0 && env_value !== "") return validate_builder_max_steps(Number(env_value), env_value);
-	return 50;
-}
-function validate_builder_max_steps(value, original) {
-	if (!Number.isInteger(value) || value < 1) throw config_error(`--builder-max-steps / VOLLEY_BUILDER_MAX_STEPS must be a positive integer; got: ${String(original)}`);
+/** What only a builder, or the loop around it, uses, with the flag that sets
+* each: a critique has no use for any of them, so naming one is an error
+* rather than something quietly ignored. */
+const CRITIQUE_REFUSED = [
+	["builder_model", "--builder-model"],
+	["builder_provider", "--builder-provider"],
+	["builder_url", "--builder-url"],
+	["builder_max_steps", "--builder-max-steps"],
+	["builder_max_tool_calls", "--builder-max-tool-calls"],
+	["builder_permission_mode", "--builder-permission-mode"],
+	["allow_unsandboxed_builder", "--allow-unsandboxed-builder"],
+	["sandbox_image", "--sandbox-image"],
+	["max_iterations", "--max-iterations"],
+	["max_cost_usd", "--max-cost-usd"],
+	["git_checkpoints", "--git"],
+	["worktree", "--worktree"],
+	["discard_worktree", "--discard-worktree"],
+	["fail_on_gate_edit", "--fail-on-gate-edit"]
+];
+function refuse_builder_settings(raw) {
+	for (const [field, flag] of CRITIQUE_REFUSED) if (raw[field] !== void 0 && raw[field] !== false) throw config_error(`volley critique refuses ${flag} (${field}): a critique runs no builder and no loop`);
+}
+const BUILDER_MAX_STEPS = {
+	flag: "--builder-max-steps",
+	env_var: "VOLLEY_BUILDER_MAX_STEPS",
+	fallback: 50
+};
+const BUILDER_MAX_TOOL_CALLS = {
+	flag: "--builder-max-tool-calls",
+	env_var: "VOLLEY_BUILDER_MAX_TOOL_CALLS",
+	fallback: 1
+};
+/** A builder count's flag or config value wins over its env var, which wins
+* over the default, and either must be a positive integer. Local builder
+* only; ignored by the `claude_cli` arm. */
+function resolve_builder_count(value, env, count) {
+	if (value !== void 0) return positive_count(value, value, count);
+	const env_value = env[count.env_var];
+	if (env_value !== void 0 && env_value !== "") return positive_count(Number(env_value), env_value, count);
+	return count.fallback;
+}
+function positive_count(value, original, count) {
+	if (!Number.isInteger(value) || value < 1) throw config_error(`${count.flag} / ${count.env_var} must be a positive integer; got: ${String(original)}`);
 	return value;
 }
 /** `--sandbox-image` (a merged flag/config value) wins over the
@@ -580,7 +958,8 @@ function resolve_sandbox_image(raw, env) {
 }
 /** The `--allow-unsandboxed-builder` flag/config value wins; otherwise
 * `VOLLEY_ALLOW_UNSANDBOXED_BUILDER=1` (or `=true`) opts out. Any other value
-* (incl. `0`/`false`/unset) leaves the local builder refused unless contained. */
+* (incl. `0`/`false`/unset) leaves a tool-loop builder refused unless
+* contained. */
 function resolve_allow_unsandboxed_builder(raw, env) {
 	if (raw.allow_unsandboxed_builder === true) return true;
 	const env_value = env["VOLLEY_ALLOW_UNSANDBOXED_BUILDER"];
@@ -600,7 +979,7 @@ function resolve_seat_url(seat, value, provider) {
 	if (typeof value !== "string") throw config_error(`${setting} must be a URL string; got a ${typeof value}`);
 	const url = value.trim();
 	if (url === "") return null;
-	if (!takes_seat_url(provider)) throw config_error(`${setting} applies only to a ${seat} on ${SEAT_URL_PROVIDERS.join(" or ")}; ${seat}-provider is ${provider}`);
+	if (!takes_seat_url(provider)) throw config_error(`${setting} applies only to a ${seat} on ${new Intl.ListFormat("en", { type: "disjunction" }).format(SEAT_URL_PROVIDERS)}; ${seat}-provider is ${provider}`);
 	let parsed;
 	try {
 		parsed = new URL(url);
@@ -612,22 +991,46 @@ function resolve_seat_url(seat, value, provider) {
 	if (parsed.search !== "" || parsed.hash !== "") throw config_error(`${setting} must not carry a query or fragment; got: ${url}`);
 	return url;
 }
-/** volley *detects* it is running inside its sandbox container rather than
-* starting the container itself. The signal is the `VOLLEY_CONTAINED=1` marker
-* volley's own image bakes in (`Dockerfile` `ENV`), so a `docker run <hardened
-* flags> <image> volley run …` is recognized as contained out of the box. Kept
-* env-only — no ambient `/.dockerenv` probe — so the safety gate is deterministic
-* and testable, and so a CI job that itself runs in an unrelated container does
-* not silently satisfy the gate; a hand-rolled sandbox must set the marker (or
-* the operator uses `--allow-unsandboxed-builder`). */
-function detect_containment(env) {
-	const value = env["VOLLEY_CONTAINED"];
-	return value === "1" || value === "true";
+/** `--state-dir` (`state_dir`): the root each run's state sits under at
+* `<root>/<run_id>/`, resolved against `cwd` so a resumed run finds the same
+* place. Unset, or empty as a seat URL may be, it's the default root,
+* `$XDG_STATE_HOME/volley`, else `~/.local/state/volley`: outside the
+* workspace, where no check that reads every file can trip over the run's own
+* feedback. */
+function resolve_state_dir(value, cwd, env) {
+	if (value === void 0 || value === null) return default_state_root(env);
+	if (typeof value !== "string") throw config_error(`--state-dir (state_dir) must be a path string; got a ${typeof value}`);
+	const path = value.trim();
+	return path === "" ? default_state_root(env) : resolve(cwd, path);
+}
+function is_writable_dir(path) {
+	try {
+		if (!statSync(path).isDirectory()) return false;
+		accessSync(path, constants.W_OK);
+		return true;
+	} catch {
+		return false;
+	}
 }
-/** A caller's run id names the run's `volley/<run id>` branch, so it is held to
-* what a branch name accepts without quoting: letters, digits, `-`, and `_`,
-* starting with a letter or digit so it never reads as an option. */
-const RUN_ID_PATTERN = /^[A-Za-z0-9][A-Za-z0-9_-]*$/;
+/** The run's state directory must be one volley can write: the directory
+* itself when it's there, as a contained run's mount is, else the nearest
+* directory above it, where volley creates it. Checked before any spend,
+* because otherwise the run dies as it starts on a raw EROFS or EACCES. Inside
+* the sandbox the root filesystem is read-only, so that is what an unmounted
+* run directory looks like, and the message says what to mount. */
+function validate_run_state_dir(dir, contained) {
+	let existing = dir;
+	while (!existsSync(existing) && dirname(existing) !== existing) existing = dirname(existing);
+	if (is_writable_dir(existing)) return;
+	throw config_error(contained ? `run state directory not writable: ${dir}. Inside the sandbox, mount the run's own host directory there (sandbox_run_args run_state, or -v <root>/<run id>:${dir}), created on the host first, and pass volley the same --run-id; or point --state-dir at a writable mount` : `run state directory not writable: ${dir}. Point --state-dir at a directory volley can write`);
+}
+/** Whether a tool-loop builder's commands would reach the host with nothing in
+* between: volley isn't contained, and the operator hasn't opted out with
+* `--allow-unsandboxed-builder` or `VOLLEY_ALLOW_UNSANDBOXED_BUILDER`. */
+function unguarded_builder(raw, env) {
+	return !detect_containment(env) && !resolve_allow_unsandboxed_builder(raw, env);
+}
+/** A caller's run id, held to `RUN_ID_PATTERN`; none gets a fresh UUID. */
 function resolve_run_id(run_id) {
 	if (run_id === void 0) return randomUUID();
 	if (!RUN_ID_PATTERN.test(run_id)) throw config_error(`--run-id must be letters, digits, '-', and '_', starting with a letter or digit; got: ${run_id}`);
@@ -639,13 +1042,17 @@ function resolve_run_id(run_id) {
 function resolve_config(raw, options = {}) {
 	const cwd = options.cwd ?? process.cwd();
 	const env = options.env ?? process.env;
-	if (typeof raw.prompt !== "string" || raw.prompt.length === 0) throw config_error("missing required option: --prompt");
+	const critique = options.mode === "critique";
+	if (critique) refuse_builder_settings(raw);
+	const builder_env = critique ? {} : env;
+	const has_prompt = typeof raw.prompt === "string" && raw.prompt.length > 0;
+	if (!has_prompt && !critique) throw config_error("missing required option: --prompt");
 	if (typeof raw.workspace !== "string" || raw.workspace.length === 0) throw config_error("missing required option: --workspace");
 	if (typeof raw.criteria !== "string" || raw.criteria.length === 0) throw config_error("missing required option: --criteria");
 	const workspace = validate_workspace(resolve(cwd, raw.workspace));
-	const prompt = expand_at_file(raw.prompt, cwd);
+	const prompt = has_prompt ? expand_at_file(raw.prompt, cwd) : "";
 	const criteria = expand_at_file(raw.criteria, cwd);
-	const max_iterations = raw.max_iterations ?? 10;
+	const max_iterations = raw.max_iterations ?? (critique ? 1 : 10);
 	if (!Number.isInteger(max_iterations) || max_iterations < 1) throw config_error(`--max-iterations must be a positive integer; got: ${String(raw.max_iterations)}`);
 	const max_cost_usd = raw.max_cost_usd ?? null;
 	if (max_cost_usd !== null && !(typeof max_cost_usd === "number" && max_cost_usd > 0)) throw config_error(`--max-cost-usd must be a positive number; got: ${String(raw.max_cost_usd)}`);
@@ -654,16 +1061,18 @@ function resolve_config(raw, options = {}) {
 	const { critic_preset, critic_prompt_path } = resolve_critic(raw.critic ?? DEFAULT_CRITIC, cwd);
 	const builder_provider = raw.builder_provider ?? DEFAULT_BUILDER_PROVIDER;
 	if (!BUILDER_PROVIDERS.includes(builder_provider)) throw config_error(`--builder-provider must be one of ${BUILDER_PROVIDERS.join(", ")}; got: ${String(raw.builder_provider)}`);
-	const allow_unsandboxed_builder = resolve_allow_unsandboxed_builder(raw, env);
+	const allow_unsandboxed_builder = resolve_allow_unsandboxed_builder(raw, builder_env);
 	const contained = detect_containment(env);
-	if (LOCAL_BUILDER_PROVIDERS.includes(builder_provider) && !contained && !allow_unsandboxed_builder) throw config_error(`builder-provider '${builder_provider}' runs a local model with a real bash (write + exec) — refused unless contained. Launch volley inside its sandbox container (docker run <hardened flags> <image> volley run …, which sets VOLLEY_CONTAINED=1), or pass --allow-unsandboxed-builder (VOLLEY_ALLOW_UNSANDBOXED_BUILDER=1) to run uncontained on the host.`);
-	const builder_max_steps = resolve_builder_max_steps(raw, env);
+	if (TOOL_LOOP_BUILDER_PROVIDERS.includes(builder_provider) && options.uncontained_builder !== "ask" && unguarded_builder(raw, builder_env)) throw config_error(`builder-provider '${builder_provider}' runs volley's tool loop with a real bash (write + exec) — refused unless contained. Launch volley inside its sandbox container (docker run <hardened flags> <image> volley run …, which sets VOLLEY_CONTAINED=1), or pass --allow-unsandboxed-builder (VOLLEY_ALLOW_UNSANDBOXED_BUILDER=1) to run uncontained on the host.`);
+	const builder_max_steps = resolve_builder_count(raw.builder_max_steps, builder_env, BUILDER_MAX_STEPS);
+	const builder_max_tool_calls = resolve_builder_count(raw.builder_max_tool_calls, builder_env, BUILDER_MAX_TOOL_CALLS);
 	const critic_provider = raw.critic_provider ?? DEFAULT_CRITIC_PROVIDER;
 	if (!CRITIC_PROVIDERS.includes(critic_provider)) throw config_error(`--critic-provider must be one of ${CRITIC_PROVIDERS.join(", ")}; got: ${String(raw.critic_provider)}`);
 	const builder_url = resolve_seat_url("builder", raw.builder_url, builder_provider);
 	const critic_url = resolve_seat_url("critic", raw.critic_url, critic_provider);
+	if (builder_provider === "openrouter" || critic_provider === "openrouter") openrouter_api_key(env);
 	if (contained) {
-		const claude_roles = [builder_provider === "claude_cli" ? "builder" : null, critic_provider === "claude_cli" ? "critic" : null].filter((role) => role !== null);
+		const claude_roles = [builder_provider === "claude_cli" && !critique ? "builder" : null, critic_provider === "claude_cli" ? "critic" : null].filter((role) => role !== null);
 		const auth_mode = env["VOLLEY_AUTH_MODE"] ?? "auto";
 		if (claude_roles.length > 0 && auth_mode !== "api_key") throw config_error(`claude_cli (${claude_roles.join(", ")}) cannot use its subscription/OAuth token inside a container — it is mangled there. Set VOLLEY_AUTH_MODE=api_key (with ANTHROPIC_API_KEY) to drive Claude contained, or run claude_cli on the host uncontained.`);
 	}
@@ -674,13 +1083,17 @@ function resolve_config(raw, options = {}) {
 	if (discard_worktree && !worktree) throw config_error("--discard-worktree applies only to a --worktree run: there is no worktree to discard");
 	const gate_paths = raw.gate_paths ?? DEFAULT_GATE_PATTERNS;
 	if (!Array.isArray(gate_paths) || gate_paths.some((glob) => typeof glob !== "string")) throw config_error("gate_paths must be an array of glob strings");
-	const sandbox_image = resolve_sandbox_image(raw, env);
+	const sandbox_image = resolve_sandbox_image(raw, builder_env);
+	const state_dir = options.state_dir !== void 0 ? options.state_dir : resolve_state_dir(raw.state_dir, cwd, env);
+	if (state_dir === workspace) throw config_error(`--state-dir (state_dir) can't be the workspace itself: ${workspace}`);
 	const run_id = resolve_run_id(options.run_id);
+	if (state_dir !== null) validate_run_state_dir(join(state_dir, run_id), contained);
 	const check = raw.check ?? "auto";
 	return {
 		version: 2,
 		run_id,
 		started_at: options.started_at ?? (/* @__PURE__ */ new Date()).toISOString(),
+		mode: critique ? "critique" : "run",
 		prompt,
 		criteria,
 		check,
@@ -689,6 +1102,7 @@ function resolve_config(raw, options = {}) {
 		builder_provider,
 		builder_url,
 		builder_max_steps,
+		builder_max_tool_calls,
 		allow_unsandboxed_builder,
 		critic_model: raw.critic_model ?? DEFAULT_CRITIC_MODEL,
 		critic_provider,
@@ -704,6 +1118,7 @@ function resolve_config(raw, options = {}) {
 		gate_paths: [...gate_paths],
 		fail_on_gate_edit: raw.fail_on_gate_edit ?? false,
 		sandbox_image,
+		state_dir,
 		workspace,
 		verbose: raw.verbose ?? false,
 		quiet: raw.quiet ?? false,
@@ -782,8 +1197,9 @@ async function warn_small_local_context(provider, model, base_url, warn, abort)
 */
 const READ_FILE_MAX_BYTES = 2e5;
 const LIST_MAX_ENTRIES = 2e3;
-/** Directories never worth handing a critic; they bury signal and blow the
-* traversal budget. */
+/** Directories never worth handing a seat; they bury signal and blow the
+* traversal budget. The acp seats' `list_files` and `search` skip the same
+* names (`./acp_tools.ts`), so a seat sees one tree under acp and on the CLI. */
 const IGNORED_DIRS = /* @__PURE__ */ new Set([
 	"node_modules",
 	".git",
@@ -796,7 +1212,7 @@ const IGNORED_DIRS = /* @__PURE__ */ new Set([
 * escape. Returns the absolute path; throws on containment violation. The
 * `workspace` root passed here is the build root, which is the per-run git
 * worktree under `--worktree` (`build_root`), not necessarily the
-* control-plane workspace that holds `.volley/`. */
+* workspace the run was started on. */
 function contain(workspace, rel) {
 	const root = resolve(workspace);
 	const candidate = isAbsolute(rel) ? resolve(rel) : resolve(root, rel);
@@ -835,6 +1251,25 @@ function walk(root, start, limit) {
 	}
 	return out;
 }
+/** The files a read tool's optional `path` covers: the whole tree, the tree
+* under a directory, or the one file it names. A path that doesn't exist
+* throws, so the seat reads why rather than an empty result that looks like
+* "nothing matched". */
+function scoped_files(workspace, path, limit) {
+	const root = resolve(workspace);
+	if (path === void 0) return walk(root, root, limit);
+	const start = contain(workspace, path);
+	let is_dir;
+	try {
+		is_dir = statSync(start).isDirectory();
+	} catch (err) {
+		throw new Error(`no such file or directory: ${path}`, { cause: err });
+	}
+	return is_dir ? walk(root, start, limit) : [{
+		rel: relative(root, start),
+		abs: start
+	}];
+}
 function build_regex(pattern, flags) {
 	try {
 		return new RegExp(pattern, flags);
@@ -846,11 +1281,11 @@ function build_regex(pattern, flags) {
 const read_file_input = z.object({ path: z.string().describe("Workspace-relative path to the file to read.") });
 const search_files_input = z.object({
 	pattern: z.string().describe("JavaScript regular expression to match against each line."),
-	path: z.string().optional().describe("Optional workspace-relative subdirectory to limit the search to."),
+	path: z.string().optional().describe("Optional workspace-relative directory or file to limit the search to."),
 	ignore_case: z.boolean().optional().describe("Case-insensitive match. Default false.")
 });
 const list_files_input = z.object({
-	path: z.string().optional().describe("Optional workspace-relative subdirectory. Default: the workspace root."),
+	path: z.string().optional().describe("Optional workspace-relative directory or file. Default: the workspace root."),
 	contains: z.string().optional().describe("Optional substring filter applied to each relative path.")
 });
 function read_only_tools(workspace) {
@@ -871,9 +1306,8 @@ function read_only_tools(workspace) {
 			input_schema: search_files_input,
 			execute: (raw) => {
 				const input = search_files_input.parse(raw);
-				const start = input.path !== void 0 ? contain(workspace, input.path) : resolve(workspace);
 				const regex = build_regex(input.pattern, input.ignore_case === true ? "i" : "");
-				const files = walk(resolve(workspace), start, LIST_MAX_ENTRIES);
+				const files = scoped_files(workspace, input.path, LIST_MAX_ENTRIES);
 				const hits = [];
 				for (const file of files) {
 					if (hits.length >= 200) break;
@@ -901,8 +1335,7 @@ function read_only_tools(workspace) {
 			input_schema: list_files_input,
 			execute: (raw) => {
 				const input = list_files_input.parse(raw);
-				const start = input.path !== void 0 ? contain(workspace, input.path) : resolve(workspace);
-				let files = walk(resolve(workspace), start, LIST_MAX_ENTRIES).map((f) => f.rel);
+				let files = scoped_files(workspace, input.path, LIST_MAX_ENTRIES).map((f) => f.rel);
 				if (input.contains !== void 0) {
 					const needle = input.contains;
 					files = files.filter((f) => f.includes(needle));
@@ -1295,7 +1728,19 @@ function builder_tools(workspace, options = {}) {
 			max_bytes: options.fetch_max_bytes ?? FETCH_MAX_BYTES
 		})
 	};
-	const finish = {
+	return [
+		...read_only_tools(workspace),
+		write_file,
+		edit_file,
+		bash,
+		fetch,
+		finish_tool()
+	];
+}
+/** The terminal `finish` tool, on its own for a builder set that brings its
+* own file and command tools (`../acp_tools.ts`). */
+function finish_tool() {
+	return {
 		name: "finish",
 		description: "Declare the task complete. Calling this tool ends your turn — do not call it until the work is done and verified.",
 		input_schema: finish_input,
@@ -1304,14 +1749,6 @@ function builder_tools(workspace, options = {}) {
 			return `finished: ${finish_input.parse(raw).summary}`;
 		}
 	};
-	return [
-		...read_only_tools(workspace),
-		write_file,
-		edit_file,
-		bash,
-		fetch,
-		finish
-	];
 }
 //#endregion
 //#region src/cost.ts
@@ -1339,9 +1776,13 @@ function add_usage(a, b) {
 	}
 	return sum;
 }
+/** The providers whose calls carry the charge they report, which fascicle
+* records as the call's cost in place of a pricing table's estimate: the
+* Claude CLI's own total, and OpenRouter's per-response charge. */
+const REPORTING_PROVIDERS = /* @__PURE__ */ new Set(["claude_cli", "openrouter"]);
 function cost_source_of(result) {
 	if (result.cost === void 0) return "unknown";
-	return result.model_resolved.provider === "claude_cli" ? "provider_reported" : "engine_derived";
+	return REPORTING_PROVIDERS.has(result.model_resolved.provider) ? "provider_reported" : "engine_derived";
 }
 /** One phase's record from its generate result. The duration is fascicle's
 * call-level wall clock (`GenerateResult.timing`: every turn and the tools
@@ -1376,170 +1817,56 @@ function accumulate(state, role, result, model, duration_ms) {
 		iteration_cost_usd: state.iteration_cost_usd + cost,
 		builder_cost_usd: state.builder_cost_usd + (role === "builder" ? cost : 0),
 		critic_cost_usd: state.critic_cost_usd + (role === "critic" ? cost : 0),
-		cost_warned: state.cost_warned || record.cost_usd === null
-	};
-}
-function cost_cap_hit(config, state) {
-	return config.max_cost_usd !== null && state.total_cost_usd >= config.max_cost_usd;
-}
-//#endregion
-//#region src/prewarm.ts
-/**
-* Best-effort Ollama model pre-load. A cold multi-GB model can spend minutes
-* in load + first prompt eval — longer than the provider path's in-request
-* header timeout — so the first real `generate` dies with an opaque
-* `stream interrupted: fetch failed`. Ollama's documented load call (a
-* `/api/generate` POST with a model and no prompt) returns as soon as the
-* model is resident, so absorbing the load here keeps the real call's
-* time-to-first-byte inside its budget.
-*
-* Best-effort like the context probe: any failure (server down, unknown
-* model, timeout) is swallowed — the real generate call surfaces the error
-* with retry and reporting behind it. Ollama-only; LM Studio has no
-* equivalent load-without-generating endpoint.
-*/
-/** Generous ceiling on the load itself: a disk-cold 30B-class model takes
-* tens of seconds to minutes. Past this, give up and let the real call try. */
-const PREWARM_TIMEOUT_MS = 3e5;
-/** Load `model` into the Ollama server's memory (`base_url` is the server
-* root). Resolves when the model is resident, on any failure, or at the
-* timeout — never throws, never gates the phase. An already-resident model
-* returns immediately, so warming the same model for both roles is free.
-*
-* Load-bearing only on the ai_sdk transport: the cold-load death this
-* absorbs is the ai-sdk/undici in-request header timeout, not an Ollama
-* limit. On fascicle's native transport — the deferred future bridge, see
-* `seat_provider_config` in engine.ts — the per-turn call is a bare
-* `fetch(..., { signal })` with no default body timeout, so a cold model can
-* take as long as it needs and this pre-warm's necessity drops away (keep it
-* regardless: a resident model makes it a no-op). Native would also move the
-* `num_ctx`/`keep_alive` levers to per-call `provider_options.ollama`, rather
-* than the server-side Modelfile defaults volley probes today. */
-async function prewarm_ollama_model(base_url, model, abort) {
-	const signals = [AbortSignal.timeout(PREWARM_TIMEOUT_MS)];
-	if (abort !== void 0) signals.push(abort);
-	try {
-		await fetch(`${base_url.replace(/\/$/, "")}/api/generate`, {
-			method: "POST",
-			headers: { "content-type": "application/json" },
-			body: JSON.stringify({ model }),
-			signal: AbortSignal.any(signals)
-		});
-	} catch {}
-}
-//#endregion
-//#region src/workspace.ts
-/**
-* Workspace lifecycle: `.volley/` initialization, stale-run
-* backup rotation, path helpers, and optional git checkpoints.
-*/
-function volley_path(workspace, ...parts) {
-	return join(workspace, ".volley", ...parts);
-}
-function iteration_dir(workspace, iteration) {
-	return volley_path(workspace, "iterations", String(iteration).padStart(3, "0"));
-}
-/** Filesystem-safe timestamp for `.bak.<stamp>` rotation names. Shared by the
-* `.volley/` rotation here and the worktree rotation so both preserve
-* prior state under the same never-destroy naming. */
-function backup_stamp(date = /* @__PURE__ */ new Date()) {
-	return date.toISOString().replace(/[:.]/g, "-");
-}
-/** Fresh-run initialization: rotate a stale `.volley/` aside, then create the
-* directory tree. Resume runs skip rotation (`preserve: true`). */
-function initialize_workspace(workspace, options = {}) {
-	const dir = volley_path(workspace);
-	if (existsSync(dir) && options.preserve !== true) renameSync(dir, join(workspace, `.volley.bak.${backup_stamp()}`));
-	mkdirSync(volley_path(workspace, "iterations"), { recursive: true });
-}
-function write_resolved_config(config) {
-	const record = {
-		version: config.version,
-		run_id: config.run_id,
-		started_at: config.started_at,
-		prompt: config.prompt,
-		criteria: config.criteria,
-		check: config.check,
-		check_resolved: config.check_resolved,
-		builder_model: config.builder_model,
-		builder_provider: config.builder_provider,
-		builder_url: config.builder_url,
-		builder_max_steps: config.builder_max_steps,
-		critic_model: config.critic_model,
-		critic_provider: config.critic_provider,
-		critic_url: config.critic_url,
-		builder_permission_mode: config.builder_permission_mode,
-		critic_preset: config.critic_preset,
-		critic_prompt_path: config.critic_prompt_path,
-		max_iterations: config.max_iterations,
-		max_cost_usd: config.max_cost_usd,
-		git_checkpoints: config.git_checkpoints,
-		worktree: config.worktree,
-		discard_worktree: config.discard_worktree,
-		gate_paths: config.gate_paths,
-		fail_on_gate_edit: config.fail_on_gate_edit,
-		sandbox_image: config.sandbox_image,
-		workspace: config.workspace
-	};
-	writeFileSync(volley_path(config.workspace, "config.json"), `${JSON.stringify(record, null, 2)}\n`);
-}
-/**
-* `--git`: commit everything under `root` after a phase (`git -C <root>`, run
-* here via the spawn `cwd`). `root` is the build root (`build_root`): the
-* workspace on the default path, or the per-run worktree under `--worktree`, so
-* checkpoints land on the worktree branch rather than the workspace's.
-* A checkpoint with nothing to commit is fine; a missing git binary or repo is
-* not.
-*/
-function git_checkpoint(root, message) {
-	if (!existsSync(join(root, ".git"))) throw config_error(`git checkpoints require a git repository: ${root}`);
-	const add = spawnSync("git", ["add", "-A"], { cwd: root });
-	if (add.error !== void 0 || add.status !== 0) throw check_error(`git add failed: ${add.stderr?.toString() ?? String(add.error)}`);
-	const commit = spawnSync("git", [
-		"commit",
-		"-m",
-		message,
-		"--no-verify"
-	], { cwd: root });
-	if (commit.error !== void 0) throw check_error(`git commit failed to start: ${String(commit.error)}`);
-	const out = `${commit.stdout?.toString() ?? ""}${commit.stderr?.toString() ?? ""}`;
-	if (commit.status !== 0 && !out.includes("nothing to commit")) throw check_error(`git commit failed: ${out.trim()}`);
-}
-/**
-* Integration — bring a successful `--worktree --git` run's effects onto the
-* workspace branch. The per-phase checkpoints on `branch` are the raw
-* audit/replay trail; here they collapse into a single squash commit
-* (`git merge --squash` then `commit`) on the workspace's current branch. The
-* shared object store makes this a local, fetch-free merge, and — because the
-* workspace branch never moved during the run (checkpoints went to the phase
-* branch) — it is conflict-free by construction. A run whose phase branch has no
-* new commits integrates to a no-op. Only successful runs integrate; an
-* abandoned run's branch is discarded wholesale by the teardown trio instead.
-*/
-function integrate_worktree(workspace, branch, message) {
-	if (!existsSync(join(workspace, ".git"))) throw config_error(`--worktree integration requires a git repository: ${workspace}`);
-	const squash = spawnSync("git", [
-		"merge",
-		"--squash",
-		branch
-	], {
-		cwd: workspace,
-		encoding: "utf8"
-	});
-	if (squash.error !== void 0) throw check_error(`git merge --squash failed to start: ${String(squash.error)}`);
-	if (squash.status !== 0) throw check_error(`git merge --squash ${branch} failed: ${(squash.stderr || squash.stdout).trim()}`);
-	const commit = spawnSync("git", [
-		"commit",
-		"-m",
-		message,
-		"--no-verify"
-	], {
-		cwd: workspace,
-		encoding: "utf8"
-	});
-	if (commit.error !== void 0) throw check_error(`git commit failed to start: ${String(commit.error)}`);
-	const out = `${commit.stdout ?? ""}${commit.stderr ?? ""}`;
-	if (commit.status !== 0 && !out.includes("nothing to commit")) throw check_error(`git commit failed: ${out.trim()}`);
+		cost_warned: state.cost_warned || record.cost_usd === null
+	};
+}
+function cost_cap_hit(config, state) {
+	return config.max_cost_usd !== null && state.total_cost_usd >= config.max_cost_usd;
+}
+//#endregion
+//#region src/prewarm.ts
+/**
+* Best-effort Ollama model pre-load. A cold multi-GB model can spend minutes
+* in load + first prompt eval — longer than the provider path's in-request
+* header timeout — so the first real `generate` dies with an opaque
+* `stream interrupted: fetch failed`. Ollama's documented load call (a
+* `/api/generate` POST with a model and no prompt) returns as soon as the
+* model is resident, so absorbing the load here keeps the real call's
+* time-to-first-byte inside its budget.
+*
+* Best-effort like the context probe: any failure (server down, unknown
+* model, timeout) is swallowed — the real generate call surfaces the error
+* with retry and reporting behind it. Ollama-only; LM Studio has no
+* equivalent load-without-generating endpoint.
+*/
+/** Generous ceiling on the load itself: a disk-cold 30B-class model takes
+* tens of seconds to minutes. Past this, give up and let the real call try. */
+const PREWARM_TIMEOUT_MS = 3e5;
+/** Load `model` into the Ollama server's memory (`base_url` is the server
+* root). Resolves when the model is resident, on any failure, or at the
+* timeout — never throws, never gates the phase. An already-resident model
+* returns immediately, so warming the same model for both roles is free.
+*
+* Load-bearing only on the ai_sdk transport: the cold-load death this
+* absorbs is the ai-sdk/undici in-request header timeout, not an Ollama
+* limit. On fascicle's native transport — the deferred future bridge, see
+* `seat_provider_config` in engine.ts — the per-turn call is a bare
+* `fetch(..., { signal })` with no default body timeout, so a cold model can
+* take as long as it needs and this pre-warm's necessity drops away (keep it
+* regardless: a resident model makes it a no-op). Native would also move the
+* `num_ctx`/`keep_alive` levers to per-call `provider_options.ollama`, rather
+* than the server-side Modelfile defaults volley probes today. */
+async function prewarm_ollama_model(base_url, model, abort) {
+	const signals = [AbortSignal.timeout(PREWARM_TIMEOUT_MS)];
+	if (abort !== void 0) signals.push(abort);
+	try {
+		await fetch(`${base_url.replace(/\/$/, "")}/api/generate`, {
+			method: "POST",
+			headers: { "content-type": "application/json" },
+			body: JSON.stringify({ model }),
+			signal: AbortSignal.any(signals)
+		});
+	} catch {}
 }
 //#endregion
 //#region src/worktree.ts
@@ -1574,9 +1901,8 @@ function worktree_branch(run_id) {
 * when `--worktree` is on, else the workspace itself. This is the containment
 * root the builder/critic file tools resolve through `contain()` and the
 * `bash` tool's cwd — re-pointing it moves writes into the worktree without
-* disturbing the workspace. `.volley/` state, the resolved config, and the
-* deterministic check stay under the workspace (the control plane), so only the
-* model-driven effect surface moves.
+* disturbing the workspace. The check, the change report, and the checkpoints
+* follow it there; the run's state stays wherever `state_dir` puts it.
 */
 function build_root(workspace, worktree) {
 	return worktree ? worktree_path(workspace) : workspace;
@@ -1879,7 +2205,7 @@ async function with_worktree(options, body) {
 /**
 * Builder invocation: one call to the builder leaf per iteration is one
 * complete agentic session in the workspace — Claude Code under `claude_cli`,
-* or a local model driving volley's tools.
+* or a model driving volley's tools on any other provider, local or hosted.
 */
 const BUILDER_ALLOWED_TOOLS = [
 	"Read",
@@ -1901,7 +2227,7 @@ function compose_builder_system() {
 }
 /** Walk up from this module to the package root (works from src/ under tsx
 * and from dist/ in the published package, which ships src/builder/presets). */
-function presets_dir$1() {
+function builder_presets_dir() {
 	let dir = dirname(fileURLToPath(import.meta.url));
 	for (let depth = 0; depth < 6; depth += 1) {
 		const candidate = join(dir, "src", "builder", "presets");
@@ -1910,40 +2236,67 @@ function presets_dir$1() {
 	}
 	throw config_error("builder presets directory not found");
 }
-function compose_builder_system_local() {
-	return `${[
+/** The harness append each tool set's builder reads: one naming volley's own
+* tools, or one naming fascicle's workspace tools under `volley acp` and
+* saying what a call the user refused means. */
+const HARNESS_APPENDS = {
+	volley: "harness_append_local.md",
+	acp: "harness_append_acp.md"
+};
+/** Step 2 of "Work this way", as both harness appends word it for a builder
+* held to one tool call a turn. */
+const ONE_CALL_A_TURN = "2. Call one tool at a time and read its result before deciding the next\n   call.";
+/** Step 2 for a builder that may make several calls a turn: fascicle runs a
+* turn's calls in order, so a read can't race an edit, but each change or
+* command is still worth reading before the next. */
+function batched_calls(max_tool_calls) {
+	return [
+		`2. Batch independent reads into one turn, up to ${String(max_tool_calls)} calls, but`,
+		"   make one change or command per turn, and read its result before",
+		"   deciding the next call."
+	].join("\n");
+}
+function compose_builder_system_local(tool_set = "volley", max_tool_calls = 1) {
+	const identity = [
 		"You are the builder inside the volley harness: an autonomous agent",
 		"iterating on a workspace until it satisfies a task and its acceptance",
 		"criteria."
-	].join("\n")}\n\n${readFileSync(join(presets_dir$1(), "harness_append_local.md"), "utf8").trimEnd()}`;
+	].join("\n");
+	const append = readFileSync(join(builder_presets_dir(), HARNESS_APPENDS[tool_set]), "utf8");
+	return `${identity}\n\n${(max_tool_calls > 1 ? append.replace(ONE_CALL_A_TURN, batched_calls(max_tool_calls)) : append).trimEnd()}`;
 }
 /** The builder system prompt for the active provider: the `claude_cli`
 * prompt (implicit built-in tools, "the current working directory") for the
-* CLI builder, or the local prompt (volley's tool set, workspace-is-cwd, and
-* the harness-enforced `finish` stop) for a local model. Mirrors the critic's
-* `resolve_critic_prompt`. */
-function resolve_builder_system(config) {
-	return config.builder_provider === "claude_cli" ? compose_builder_system() : compose_builder_system_local();
+* CLI builder, which never takes an injected set, or the local prompt naming
+* the seat's tool set for every provider that runs volley's tool loop,
+* `openrouter` included. Mirrors the critic's `resolve_critic_prompt`. */
+function resolve_builder_system(config, tool_set) {
+	return config.builder_provider === "claude_cli" ? compose_builder_system() : compose_builder_system_local(tool_set, config.builder_max_tool_calls);
 }
 const BUILDER_TOOL_CALL_REPAIR_ATTEMPTS = 3;
 /** Per-provider tool wiring for the builder, mirroring `critic_tool_options`.
 * The `claude_cli` builder is confined at the CLI permission layer (allowlist +
 * permission mode) and brings its own built-in tools — this arm is unchanged
-* from v2. A local builder brings none, so volley supplies its whole
-* workspace tool surface plus the five per-call loop knobs and **no**
-* schema (it produces a workspace, not a verdict); per-call values win over
-* engine defaults. */
-function builder_tool_options(config, bash_executor) {
+* from v2, and takes no injected tools or approval handler. Every other
+* builder, local or `openrouter`, brings none, so volley supplies its whole
+* workspace tool surface (or the set the caller injected in its place, with
+* the handler that answers its approvals) plus the five per-call loop knobs
+* and **no** schema (it produces a workspace, not a verdict); per-call values
+* win over engine defaults. */
+function builder_tool_options(deps) {
+	const { config } = deps;
+	const bash_executor = deps.bash_executor ?? null;
 	if (config.builder_provider === "claude_cli") return { provider_options: { claude_cli: {
 		allowed_tools: [...BUILDER_ALLOWED_TOOLS],
 		extra_args: ["--permission-mode", config.builder_permission_mode]
 	} } };
 	return {
-		tools: builder_tools(build_root(config.workspace, config.worktree), bash_executor !== null ? { bash_executor } : {}),
+		tools: deps.tools === void 0 ? builder_tools(build_root(config.workspace, config.worktree), bash_executor !== null ? { bash_executor } : {}) : [...deps.tools],
+		...deps.on_tool_approval !== void 0 ? { on_tool_approval: deps.on_tool_approval } : {},
 		max_steps: config.builder_max_steps,
 		tool_error_policy: "feed_back",
 		tool_call_repair_attempts: BUILDER_TOOL_CALL_REPAIR_ATTEMPTS,
-		max_tool_calls_per_step: 1
+		max_tool_calls_per_step: config.builder_max_tool_calls
 	};
 }
 function compose_builder_prompt(input) {
@@ -1976,9 +2329,9 @@ function make_builder_step(deps) {
 		description: "model call: one agentic session in the build root",
 		provider: config.builder_provider,
 		model: config.builder_model,
-		system: resolve_builder_system(config),
+		system: resolve_builder_system(config, deps.tool_set),
 		on_chunk: deps.on_chunk,
-		...builder_tool_options(config, deps.bash_executor ?? null)
+		...builder_tool_options(deps)
 	});
 }
 async function run_builder(deps, state, ctx) {
@@ -2022,11 +2375,30 @@ function presets_dir() {
 	}
 	throw config_error("critic presets directory not found");
 }
-function resolve_critic_prompt(config) {
+/** The harness append that names the critic's reads. The CLI critic reads
+* via built-in Read/Grep/Glob and never takes an injected set; every other
+* critic, local or `openrouter`, reads via volley's read_file/search_files/
+* list_files, or under `volley acp` via fascicle's read_file/list_files/
+* search. */
+function harness_append_file(config, tool_set) {
+	if (config.critic_provider === "claude_cli") return "harness_append.md";
+	return tool_set === "acp" ? "harness_append_acp.md" : "harness_append_local.md";
+}
+function resolve_critic_prompt(config, tool_set = "volley") {
 	const base = config.critic_prompt_path !== null ? readFileSync(config.critic_prompt_path, "utf8") : readFileSync(join(presets_dir(), `${config.critic_preset}.md`), "utf8");
-	const append_file = config.critic_provider === "claude_cli" ? "harness_append.md" : "harness_append_local.md";
-	const append = readFileSync(join(presets_dir(), append_file), "utf8");
-	return `${base.trimEnd()}\n\n${append.trimEnd()}`;
+	const append = readFileSync(join(presets_dir(), harness_append_file(config, tool_set)), "utf8");
+	return `${base.trimEnd()}\n\n${place_iterations_dir(append, config).trimEnd()}`;
+}
+const ITERATIONS_DIR = "{{iterations_dir}}";
+/** Fill the harness append's line about prior iteration state with where the
+* run's `iterations/` sits, relative to the build root the critic reads from,
+* or drop the line when the run's state is outside that root: a confined
+* critic can't reach it, and naming it would only spend the critic's steps on
+* a path it can't open. */
+function place_iterations_dir(append, config) {
+	const path = relative(build_root(config.workspace, config.worktree), state_path(config, "iterations"));
+	const outside = path === ".." || path.startsWith(`..${sep}`) || isAbsolute(path);
+	return append.split("\n").filter((line) => !(outside && line.includes(ITERATIONS_DIR))).join("\n").replaceAll(ITERATIONS_DIR, `${path}/`);
 }
 /** One-letter status column, so a hundred-path list stays readable. */
 const STATUS_MARKS = {
@@ -2106,9 +2478,10 @@ function compose_critic_prompt(input) {
 //#endregion
 //#region src/critic/run.ts
 /**
-* Critic invocation: a read-only session (`claude_cli`, or a local model with
-* volley's read tools) returning a schema-validated structured verdict. The
-* harness — not the critic — writes `.volley/feedback.md` and `.volley/verdict`.
+* Critic invocation: a read-only session (`claude_cli`, or a local or
+* `openrouter` model with volley's read tools) returning a schema-validated
+* structured verdict. The
+* harness — not the critic — writes the run's `feedback.md` and `verdict`.
 */
 const CRITIC_ALLOWED_TOOLS = [
 	"Read",
@@ -2116,18 +2489,24 @@ const CRITIC_ALLOWED_TOOLS = [
 	"Glob"
 ];
 const CRITIC_DISALLOWED_TOOLS = "Write,Edit,MultiEdit,NotebookEdit,Bash";
+/** A critic that holds tools isn't constrain-decoded, so nothing stops it
+* answering null for the feedback or the unmet list, as GLM did on approving.
+* Both fields allow null and read it as empty, so that answer is a verdict
+* rather than a schema failure that ends the run. fascicle sends the schema's
+* input side, so the model sees each field as its type or null, still
+* required; `.describe()` comes last so the description stays on the field. */
 const verdict_schema = z.object({
 	verdict: z.enum(["approved", "changes_requested"]),
-	feedback: z.string().describe("Free-form markdown for the builder. Concrete and actionable. Will be passed verbatim into the next iteration."),
-	unmet_criteria: z.array(z.string()).describe("The specific acceptance criteria judged unmet, verbatim. Empty when approved.")
+	feedback: z.string().nullable().transform((feedback) => feedback ?? "").describe("Free-form markdown for the builder. Concrete and actionable. Will be passed verbatim into the next iteration."),
+	unmet_criteria: z.array(z.string()).nullable().transform((unmet) => unmet ?? []).describe("The specific acceptance criteria judged unmet, verbatim. Empty when approved.")
 });
 /** Read-only tool wiring per provider. The claude_cli critic is confined at
-* the CLI permission layer (allowlist + explicit disallow); a local-model
-* critic gets volley's own workspace-scoped read-only tools and no write
-* path at all.
+* the CLI permission layer (allowlist + explicit disallow); every other critic,
+* local or `openrouter`, gets volley's own workspace-scoped read-only tools (or
+* the set the caller injected in their place) and no write path at all.
 *
-* The two providers also enforce `verdict_schema` by two different paths, both
-* verified live. The claude_cli critic compiles the schema for
+* The providers also enforce `verdict_schema` by different paths, the first
+* two verified live. The claude_cli critic compiles the schema for
 * `claude --json-schema`; that only works because fascicle 0.14.0's
 * `compile_schema` strips the top-level `$schema`/`$id` that zod v4 stamps (the
 * CLI rejects them) — `live_smoke` asserts a structured verdict comes back, so a
@@ -2137,16 +2516,22 @@ const verdict_schema = z.object({
 * structurally guaranteed at decode time, with no prompt-parse-repair;
 * `live_local_builder` exercises that path. Unchanged in v3: on ai_sdk
 * constrained decode needs no wiring, and under the deferred native flip it
-* would move to `provider_options.ollama.format` (parked). */
-function critic_tool_options(config) {
+* would move to `provider_options.ollama.format` (parked). An `openrouter`
+* critic rides fascicle's native OpenAI-compatible core, which has no
+* constrained decode: the schema travels in the prompt, and the reply is parsed
+* and repaired. */
+function critic_tool_options(config, tools) {
 	if (config.critic_provider === "claude_cli") return { provider_options: { claude_cli: {
 		allowed_tools: [...CRITIC_ALLOWED_TOOLS],
 		extra_args: ["--disallowedTools", CRITIC_DISALLOWED_TOOLS]
 	} } };
-	return { tools: read_only_tools(build_root(config.workspace, config.worktree)) };
+	return {
+		tools: tools === void 0 ? read_only_tools(build_root(config.workspace, config.worktree)) : [...tools],
+		max_steps: 10
+	};
 }
-/** Is this critic failure the transient local-provider stream death the ladder
-* retries? Local providers only (the `claude_cli` path is proven and its
+/** Is this critic failure the transient provider stream death the ladder
+* retries? Every provider but `claude_cli` (that path is proven and its
 * retries cost real money), and fascicle's typed `provider_error` only (no
 * message string-matching). A user abort never gets here: `retry` and
 * `fallback` pass control-flow signals through untouched, and `retry` checks the
@@ -2155,6 +2540,23 @@ function is_retryable_critic_error(config, err) {
 	if (config.critic_provider === "claude_cli") return false;
 	return error_kind(err) === "provider_error";
 }
+/** Did a schema call end without a verdict because the tool loop ran out —
+* its step budget spent, or its last turn a tool call — rather than because
+* the response was cut off or blocked? fascicle throws
+* `incomplete_generation_error` for every such finish; only these two are
+* fixed by taking the tools away. Matched on `kind` and `finish_reason`. */
+function is_step_exhaustion(err) {
+	if (error_kind(err) !== "incomplete_generation_error") return false;
+	const finish = err.finish_reason;
+	return finish === "max_steps" || finish === "tool_calls";
+}
+/** Does this failure send a local critic to the tool-less rung? A stream death
+* the retry rung gave up on, or step exhaustion, which no retry would fix. A
+* `length` or `content_filter` finish is neither, and stays fatal. */
+function is_degradable_critic_error(config, err) {
+	if (config.critic_provider === "claude_cli") return false;
+	return error_kind(err) === "provider_error" || is_step_exhaustion(err);
+}
 /** fascicle's `provider_error.cause_kind` is `... | undefined`; fold the
 * missing case into `'unknown'` so a recorded retry always names a cause. */
 function retry_cause_kind_of(err) {
@@ -2173,7 +2575,8 @@ function finalize_critic(state, result, model, mark, duration_ms) {
 			...next.critic,
 			retries: mark.retries,
 			...mark.retry_cause_kind !== void 0 ? { retry_cause_kind: mark.retry_cause_kind } : {},
-			...mark.critic_degraded === true ? { critic_degraded: true } : {}
+			...mark.critic_degraded === true ? { critic_degraded: true } : {},
+			...mark.degrade_cause !== void 0 ? { degrade_cause: mark.degrade_cause } : {}
 		},
 		verdict: result.content.verdict,
 		feedback: result.content.feedback,
@@ -2203,7 +2606,7 @@ function critic_call_config(deps) {
 		engine: deps.engine,
 		provider: config.critic_provider,
 		model: config.critic_model,
-		system: resolve_critic_prompt(config),
+		system: resolve_critic_prompt(config, deps.tool_set),
 		schema: verdict_schema,
 		on_chunk: deps.on_chunk
 	};
@@ -2215,7 +2618,7 @@ function make_critic_tools_step(deps) {
 		...critic_call_config(deps),
 		id: "critic_tools",
 		description: "model call with read-only workspace tools",
-		...critic_tool_options(deps.config)
+		...critic_tool_options(deps.config, deps.tools)
 	});
 }
 /** Rung 2's leaf: the same critic with no tools at all, so nothing enters
@@ -2228,9 +2631,28 @@ function make_critic_toolless_step(deps) {
 		description: "model call judging from the check output and a file list"
 	});
 }
-/** The ladder's retry predicate for this run's critic provider. */
+const ladder_traces = /* @__PURE__ */ new WeakMap();
+/** The ladder's retry predicate for this run's critic provider: stream deaths
+* only. It also traces every failure it sees, so a degraded verdict can say how
+* many retries preceded it and why. A first attempt starts a fresh critique, so
+* it forgets the last one's retries. */
 function critic_retryable(config) {
-	return (err) => is_retryable_critic_error(config, err);
+	let retried_cause;
+	return (err, attempt) => {
+		if (attempt === 1) retried_cause = void 0;
+		if (typeof err === "object" && err !== null) ladder_traces.set(err, {
+			attempt,
+			retried_cause
+		});
+		const retryable = is_retryable_critic_error(config, err);
+		if (retryable) retried_cause = retry_cause_kind_of(err);
+		return retryable;
+	};
+}
+/** The ladder's fallback predicate: a stream death the retry rung gave up on,
+* or step exhaustion, which it never retries. */
+function critic_degradable(config) {
+	return (err) => is_degradable_critic_error(config, err);
 }
 /** Rung 1's projection: a verdict the tool-bearing critic produced, marked with
 * how many stream deaths it took and the last one's cause. */
@@ -2244,27 +2666,42 @@ function with_retries(outcome) {
 		}
 	};
 }
-/** Rung 2's answer: every retry was spent, and the verdict is marked degraded
-* so it never passes silently as a full critique. */
+/** Rung 2's answer, marked degraded so it never passes silently as a full
+* critique. Why it degraded, and the retries before it, are the ladder's to
+* stamp (`with_fallback_cause`), since the tool-less call can't see them. */
 function as_degraded(result) {
 	return {
 		result,
 		mark: {
-			retries: 1,
+			retries: 0,
 			retry_cause_kind: void 0,
 			critic_degraded: true
 		}
 	};
 }
 /** The ladder's projection: the tool-less backup cannot see why it ran, so the
-* cause comes from the last stream death the retry rung gave up on. */
+* mark comes from the error that sent it there.
+*
+* - A stream death the retry rung gave up on: every retry was spent, and the
+*   cause is that error's own.
+* - Step exhaustion, never retried: the retries are the stream deaths before
+*   it, usually none, and the cause is the last of those.
+*
+* An error the retry rung never traced reads as one that spent every retry. */
 function with_fallback_cause(outcome) {
 	if (outcome.source === "primary") return outcome.value;
+	const err = outcome.primary_error;
+	const trace = typeof err === "object" && err !== null ? ladder_traces.get(err) : void 0;
+	const retries = trace === void 0 ? 1 : trace.attempt - 1;
+	const exhausted = is_step_exhaustion(err);
+	const retry_cause_kind = exhausted ? trace?.retried_cause : retry_cause_kind_of(err);
 	return {
 		...outcome.value,
 		mark: {
 			...outcome.value.mark,
-			retry_cause_kind: retry_cause_kind_of(outcome.primary_error)
+			retries,
+			retry_cause_kind,
+			degrade_cause: exhausted ? "step_exhaustion" : "stream_error"
 		}
 	};
 }
@@ -2374,14 +2811,50 @@ function status_of(value, converged) {
 */
 const CHECK_ARTIFACT_MAX_CHARS = 8e3;
 const CHECKRIDE_SCHEMA_VERSION = 1;
+/** How long a check's output may stay open after its shell exits, while
+* something that escaped its process group still holds it, before the reads
+* stop so the check can finish. */
+const OUTPUT_GRACE_MS = 2e3;
+/** Kill a check's whole process group with SIGKILL. A test runner's workers,
+* and anything the check started in the background, sit in the shell's group,
+* so killing the shell alone would leave them running. Windows has no process
+* groups to signal, so there only the shell goes. */
+function kill_group(proc) {
+	if (proc.pid === void 0) return;
+	try {
+		if (process.platform === "win32") proc.kill("SIGKILL");
+		else process.kill(-proc.pid, "SIGKILL");
+	} catch {}
+}
+/** Run a check to its end in a process group of its own, capturing its
+* output. The abort kills the whole group and rejects with `aborted_error`,
+* and the shell's exit kills whatever it left behind, so a stopped run leaves
+* no test workers or servers running. */
 function spawn_capture(cmd, args, options) {
+	const { abort } = options;
+	if (abort?.aborted === true) return Promise.reject(new aborted_error("check aborted"));
 	return new Promise((resolve_promise, reject) => {
 		const proc = spawn(cmd, args, {
 			cwd: options.cwd,
-			...options.abort !== void 0 ? { signal: options.abort } : {}
+			...options.env !== void 0 ? { env: options.env } : {},
+			detached: process.platform !== "win32",
+			stdio: [
+				"ignore",
+				"pipe",
+				"pipe"
+			]
 		});
 		let stdout = "";
 		let stderr = "";
+		let grace;
+		const on_abort = () => {
+			kill_group(proc);
+			reject(new aborted_error("check aborted"));
+		};
+		const finish = () => {
+			clearTimeout(grace);
+			abort?.removeEventListener("abort", on_abort);
+		};
 		proc.stdout.on("data", (chunk) => {
 			stdout += chunk.toString();
 		});
@@ -2389,19 +2862,25 @@ function spawn_capture(cmd, args, options) {
 			stderr += chunk.toString();
 		});
 		proc.on("error", (err) => {
-			if (err.name === "AbortError") {
-				reject(new aborted_error("check aborted"));
-				return;
-			}
+			finish();
 			reject(check_error(`check command failed to start: ${err.message}`));
 		});
+		proc.on("exit", () => {
+			kill_group(proc);
+			grace = setTimeout(() => {
+				proc.stdout.destroy();
+				proc.stderr.destroy();
+			}, OUTPUT_GRACE_MS);
+		});
 		proc.on("close", (code) => {
+			finish();
 			resolve_promise({
 				exit_code: code,
 				stdout,
 				stderr
 			});
 		});
+		abort?.addEventListener("abort", on_abort, { once: true });
 	});
 }
 /** Best-effort parse of the summary contract. `pnpm exec` talks over the
@@ -2465,13 +2944,14 @@ function read_failing_artifacts(workspace, failing) {
 	});
 }
 async function run_checkride(opts) {
+	const { workspace, ...spawn_options } = opts;
 	const proc = await spawn_capture("pnpm", [
 		"exec",
 		"checkride",
 		"--json"
 	], {
-		cwd: opts.workspace,
-		...opts.abort !== void 0 ? { abort: opts.abort } : {}
+		...spawn_options,
+		cwd: workspace
 	});
 	if (proc.exit_code === 2) throw check_error(`checkride harness error: ${proc.stderr.trim()}`);
 	const { summary } = parse_summary_with_fallback(proc.stdout, opts.workspace);
@@ -2495,12 +2975,13 @@ async function run_checkride(opts) {
 * (exit 4), not an iterable failure.
 */
 async function run_command_check(opts) {
+	const { command, workspace, ...spawn_options } = opts;
 	const started = Date.now();
-	const proc = await spawn_capture("sh", ["-c", opts.command], {
-		cwd: opts.workspace,
-		...opts.abort !== void 0 ? { abort: opts.abort } : {}
+	const proc = await spawn_capture("sh", ["-c", command], {
+		...spawn_options,
+		cwd: workspace
 	});
-	if (proc.exit_code === 127 || proc.exit_code === 126) throw check_error(`check command failed to start (exit ${String(proc.exit_code)}): ${opts.command}`);
+	if (proc.exit_code === 127 || proc.exit_code === 126) throw check_error(`check command failed to start (exit ${String(proc.exit_code)}): ${command}`);
 	const log = `${proc.stdout}${proc.stderr}`;
 	return {
 		ran: true,
@@ -2524,6 +3005,63 @@ function skipped_check(reason) {
 		detail: []
 	};
 }
+/** What a rejected check says in place of its output. */
+const CHECK_REJECTED = "the check was rejected when it asked to run, so it never ran";
+/** A check the user rejected when it asked to run: recorded as a failed
+* check, with the refusal as its output, so a round can't converge on work
+* nothing checked, and the critic hears why there's no result. */
+function rejected_check(runner) {
+	return {
+		ran: true,
+		runner,
+		ok: false,
+		exit_code: null,
+		duration_ms: 0,
+		failing_slots: [],
+		detail: [],
+		log: CHECK_REJECTED
+	};
+}
+//#endregion
+//#region src/render/format.ts
+/**
+* Display formatting: glyph taxonomy, role colors, truncation.
+* All human output goes to stderr; these helpers only build strings.
+*/
+const GLYPH_TOOL = "🔧";
+const GLYPH_COST = "💰";
+const VERBOSE_TRUNCATE_CHARS = 4e3;
+function paint(role, text, color) {
+	if (!color) return text;
+	switch (role) {
+		case "builder": return pc.cyan(text);
+		case "critic": return pc.yellow(text);
+		case "check": return pc.green(text);
+		case "cost": return pc.magenta(text);
+		case "error": return pc.bold(pc.red(text));
+	}
+}
+function dim(text, color) {
+	return color ? pc.dim(text) : text;
+}
+function truncate(text, max, pointer) {
+	if (text.length <= max) return text;
+	return `${text.slice(0, max)}\n… [truncated at ${String(max)} chars; ${pointer}]`;
+}
+function format_usd(value) {
+	return value === null ? "$?" : `$${value.toFixed(3)}`;
+}
+/** One-line summary of a tool input for the default streaming mode. */
+function summarize_value(value, max = 120) {
+	let text;
+	try {
+		text = typeof value === "string" ? value : JSON.stringify(value) ?? "";
+	} catch {
+		text = String(value);
+	}
+	const flat = text.replace(/\s+/g, " ");
+	return flat.length > max ? `${flat.slice(0, max)}…` : flat;
+}
 //#endregion
 //#region src/iteration.ts
 /**
@@ -2534,9 +3072,9 @@ function skipped_check(reason) {
 function write_json(path, value) {
 	writeFileSync(path, `${JSON.stringify(value, null, 2)}\n`);
 }
-function persist_verdict_files(workspace, feedback, verdict) {
-	writeFileSync(volley_path(workspace, "feedback.md"), feedback.endsWith("\n") ? feedback : `${feedback}\n`);
-	writeFileSync(volley_path(workspace, "verdict"), `${verdict}\n`);
+function persist_verdict_files(config, feedback, verdict) {
+	writeFileSync(state_path(config, "feedback.md"), feedback.endsWith("\n") ? feedback : `${feedback}\n`);
+	writeFileSync(state_path(config, "verdict"), `${verdict}\n`);
 }
 function phase_summary(record) {
 	if (record === null) return null;
@@ -2555,17 +3093,18 @@ function phase_summary(record) {
 		...record.retries !== void 0 ? { retries: record.retries } : {},
 		...record.retry_cause_kind !== void 0 ? { retry_cause_kind: record.retry_cause_kind } : {},
 		...record.critic_degraded === true ? { critic_degraded: true } : {},
+		...record.degrade_cause !== void 0 ? { degrade_cause: record.degrade_cause } : {},
 		...record.throughput !== void 0 ? { throughput: record.throughput } : {}
 	};
 }
 /** Archive one completed iteration: verdict files, check artifacts, and the
 * per-iteration summary. */
 function archive_iteration(config, state) {
-	const dir = iteration_dir(config.workspace, state.iteration);
+	const dir = iteration_dir(config, state.iteration);
 	mkdirSync(join(dir, "check"), { recursive: true });
 	if (state.verdict !== null && state.feedback !== null) {
-		persist_verdict_files(config.workspace, state.feedback, state.verdict);
-		writeFileSync(join(dir, "feedback.md"), readFileSync(volley_path(config.workspace, "feedback.md")));
+		persist_verdict_files(config, state.feedback, state.verdict);
+		writeFileSync(join(dir, "feedback.md"), readFileSync(state_path(config, "feedback.md")));
 		writeFileSync(join(dir, "verdict"), `${state.verdict}\n`);
 	}
 	const check = state.check;
@@ -2614,13 +3153,15 @@ function run_result_from_state(config, state, status, completed_at) {
 		critic_cost_usd: state.critic_cost_usd,
 		check_duration_ms: state.check_duration_ms,
 		final_verdict: state.verdict,
-		salvaged_branch: null
+		salvaged_branch: null,
+		run_state_dir: run_state_dir(config),
+		mode: config.mode
 	};
 }
 //#endregion
 //#region src/summary.ts
 /**
-* Enriched run summary (`.volley/summary.json`): the `RunResult` projection plus
+* Enriched run summary (the run's `summary.json`): the `RunResult` projection plus
 * a `comparison` block that makes the two blessed examples' runs diff-able.
 * The block gathers the seven comparison metrics — transport,
 * iterations-to-converge, wall-clock, cost, verdict, the deterministic check
@@ -2635,22 +3176,27 @@ function run_result_from_state(config, state, status, completed_at) {
 * surviving branch on `salvaged_branch`, which teardown only names after the
 * final status is known.
 * Each write recomputes the comparison from what is currently archived under
-* `.volley/iterations/`, so the trajectory grows as the run does.
+* the run's `iterations/`, so the trajectory grows as the run does.
 */
 function transport_of(provider) {
-	return provider === "claude_cli" ? "claude_cli" : "ai_sdk";
+	switch (provider) {
+		case "claude_cli": return "claude_cli";
+		case "openrouter": return "native";
+		case "ollama":
+		case "lmstudio": return "ai_sdk";
+	}
 }
 /** Read the archived per-iteration summaries in order. A missing archive root
 * (a run that failed before the first `record`) or an unreadable entry yields an
 * empty/short list — the run summary still writes, just without that trajectory
 * detail. */
-function read_iteration_archives(workspace) {
-	const root = volley_path(workspace, "iterations");
+function read_iteration_archives(config) {
+	const root = state_path(config, "iterations");
 	if (!existsSync(root)) return [];
 	const dirs = readdirSync(root).filter((dir) => /^\d{3}$/.test(dir)).toSorted();
 	const archives = [];
 	for (const dir of dirs) {
-		const path = volley_path(workspace, "iterations", dir, "summary.json");
+		const path = state_path(config, "iterations", dir, "summary.json");
 		if (!existsSync(path)) continue;
 		try {
 			archives.push(JSON.parse(readFileSync(path, "utf8")));
@@ -2672,6 +3218,13 @@ function check_trajectory(archives) {
 function any_critic_degraded(archives) {
 	return archives.some((archive) => archive.critic?.critic_degraded === true);
 }
+/** Each distinct reason an archived critic verdict degraded. An archive from
+* before the cause was recorded reads as a stream error, the only way a verdict
+* could degrade then. */
+function critic_degrade_causes(archives) {
+	const causes = archives.filter((archive) => archive.critic?.critic_degraded === true).map((archive) => archive.critic?.degrade_cause ?? "stream_error");
+	return [...new Set(causes)];
+}
 /** Every gate path the run's builder touched, across all iterations: an edit in
 * iteration 1 that a later iteration reverted still happened, so the run-level
 * answer is the union rather than the last iteration's list. */
@@ -2697,12 +3250,12 @@ function salvage_stats(archives) {
 * archive for the check trajectory and salvage aggregate. Pure over its inputs
 * plus the on-disk archive; `write_run_summary` persists the result. */
 function build_run_summary(config, result) {
-	const archives = read_iteration_archives(config.workspace);
+	const archives = read_iteration_archives(config);
 	const wall_clock_ms = result.completed_at === null ? null : Date.parse(result.completed_at) - Date.parse(result.started_at);
 	return {
 		...result,
 		comparison: {
-			builder_transport: transport_of(config.builder_provider),
+			builder_transport: config.mode === "critique" ? null : transport_of(config.builder_provider),
 			critic_transport: transport_of(config.critic_provider),
 			iterations_to_converge: result.status === "success" ? result.iterations_completed : null,
 			wall_clock_ms,
@@ -2714,17 +3267,54 @@ function build_run_summary(config, result) {
 			gate_edits: all_gate_edits(archives),
 			unmet_criteria: archives.at(-1)?.unmet_criteria ?? [],
 			local_salvage: salvage_stats(archives),
-			critic_degraded: any_critic_degraded(archives)
+			critic_degraded: any_critic_degraded(archives),
+			critic_degrade_causes: critic_degrade_causes(archives)
 		}
 	};
 }
-/** Write the enriched summary to `.volley/summary.json` (the projection the
+/** Write the enriched summary to the run's `summary.json` (the projection the
 * examples' comparison is read from). */
 function write_run_summary(config, result) {
-	writeFileSync(volley_path(config.workspace, "summary.json"), `${JSON.stringify(build_run_summary(config, result), null, 2)}\n`);
+	writeFileSync(state_path(config, "summary.json"), `${JSON.stringify(build_run_summary(config, result), null, 2)}\n`);
 }
 //#endregion
 //#region src/phases.ts
+/**
+* The step bodies behind the nodes of `./flow.ts`, one function per node. Each
+* takes the loop's carry-state and returns the next one. The phase modules they
+* call (`./builder.ts`, `./check/`, `./critic/`) do the work; these bodies add
+* what surrounds it: the renderer lines an operator watches, change detection,
+* git checkpoints, the per-iteration archive, and the progress events a
+* trajectory reader follows the run by.
+*/
+/** The step id the check's card is recorded under: the check step's own, in
+* `./flow.ts`, so each round's check counts as a fresh run of it and opens a
+* card of its own. */
+const CHECK_STEP_ID = "check";
+/** The raw call id and tool name of the check's card. The name sits outside
+* the workspace tools, so nothing dresses the card as one of theirs. */
+const CHECK_CALL = {
+	id: "check",
+	name: "check"
+};
+function emit(ctx, event) {
+	ctx.emit(event);
+}
+/** Record one chunk of the check's card, the way `model_call` records a tool
+* call's, under a streamed run only, since nothing reads a card otherwise. */
+function record_check_chunk(ctx, chunk) {
+	if (ctx.streaming) ctx.trajectory.record({
+		kind: "model_chunk",
+		step_id: CHECK_STEP_ID,
+		chunk
+	});
+}
+/** The command line the check runs, or null when no check is configured. */
+function check_command(config) {
+	if (config.check_resolved === "checkride") return "pnpm exec checkride --json";
+	if (config.check_resolved === "command") return config.check;
+	return null;
+}
 /** How many gate edits a warning names before it summarises the rest. */
 const LISTED_GATE_EDITS = 5;
 /**
@@ -2740,20 +3330,53 @@ function report_changes(config, renderer, changes) {
 	if (changes.gate_edits.length === 0) return;
 	const hidden = changes.gate_edits.length - LISTED_GATE_EDITS;
 	const listed = changes.gate_edits.slice(0, LISTED_GATE_EDITS).join(", ");
-	renderer.warn(`gate edit: the builder changed ${String(changes.gate_edits.length)} file(s) that decide whether its work passes — ${listed}${hidden > 0 ? ` (+${String(hidden)} more)` : ""}. A green check proves less when the builder can edit the gate; the critic is told about this and the run summary records it.` + (config.fail_on_gate_edit ? " This run will end without success (--fail-on-gate-edit, exit 8)." : ""));
+	const [who, whose, edits] = config.mode === "critique" ? [
+		"the change under review touches",
+		"whether it passes",
+		"the change edits"
+	] : [
+		"the builder changed",
+		"whether its work passes",
+		"the builder can edit"
+	];
+	renderer.warn(`gate edit: ${who} ${String(changes.gate_edits.length)} file(s) that decide ${whose} — ${listed}${hidden > 0 ? ` (+${String(hidden)} more)` : ""}. A green check proves less when ${edits} the gate; the critic is told about this and the run summary records it.` + (config.fail_on_gate_edit ? " This run will end without success (--fail-on-gate-edit, exit 8)." : ""));
 }
-async function execute_check(config, abort) {
+/** Run the configured check, checkride or a command, with the environment
+* the caller set for it, else volley's own. */
+async function execute_check(deps, abort) {
+	const { config } = deps;
 	const root = build_root(config.workspace, config.worktree);
-	if (config.check_resolved === "checkride") return run_checkride({
-		workspace: root,
-		abort
-	});
-	if (config.check_resolved === "command") return run_command_check({
+	const options = {
+		abort,
+		...deps.check?.env !== void 0 ? { env: deps.check.env } : {}
+	};
+	return config.check_resolved === "checkride" ? run_checkride({
+		...options,
+		workspace: root
+	}) : run_command_check({
+		...options,
 		command: config.check,
-		workspace: root,
+		workspace: root
+	});
+}
+/** Whether the check may run `command`: yes, unless the caller set an
+* approval and it answers no. It asks on the check's card under the command
+* line as its tool name, which titles the prompt the way the card is titled,
+* and which names none of fascicle's gated tools, so the prompt offers Allow
+* and Reject alone, never an "always" that would cover the builder's commands
+* too. A stop while it asks ends the check as a stop, not a rejection. */
+async function check_allowed(deps, command, abort) {
+	const approve = deps.check?.on_approval;
+	if (approve === void 0) return true;
+	const allowed = await approve({
+		tool_name: command,
+		input: { command },
+		tool_call_id: CHECK_CALL.id,
+		step_index: 0,
 		abort
 	});
-	return skipped_check("none");
+	if (abort.aborted) throw new aborted_error("check aborted");
+	return allowed;
 }
 /** Open the next iteration: advance the counter and clear every per-iteration
 * field, so nothing from the last round leaks into this one. */
@@ -2770,11 +3393,31 @@ function next_iteration(s) {
 		unmet_criteria: []
 	};
 }
+/** What changed in the build root since the baseline, reported as it's
+* measured: the count, and a warning for any gate edit. A run measures the
+* builder's work this way after every build, and a critique measures the
+* change under review the same way, so the critic is shown one shape either
+* way. */
+function measure_changes(deps, root) {
+	const { config, renderer } = deps;
+	const changes = collect_changes({
+		root,
+		baseline: deps.baseline,
+		gate_patterns: config.gate_paths,
+		excluded_dir: state_root_within(root, config.state_dir)
+	});
+	report_changes(config, renderer, changes);
+	return changes;
+}
 /** `build`: one builder session edits the build root, then volley measures what
 * it changed and, under `--git`, checkpoints it. */
 async function build_phase(deps, builder, s, ctx) {
 	const { config, renderer } = deps;
 	const next = next_iteration(s);
+	emit(ctx, {
+		volley: "round_start",
+		round: next.iteration
+	});
 	renderer.phase_start(next.iteration, "builder");
 	const built = await run_builder({
 		builder,
@@ -2784,18 +3427,28 @@ async function build_phase(deps, builder, s, ctx) {
 	renderer.phase_end(built.iteration, "builder", true);
 	renderer.cost_line(built.iteration, "builder", built.builder?.cost_usd ?? null, built.total_cost_usd);
 	if (built.cost_warned && !s.cost_warned) renderer.warn("cost unavailable for this phase; recording null (run totals sum what is known)");
-	const changes = collect_changes({
-		root: build_root(config.workspace, config.worktree),
-		baseline: deps.baseline,
-		gate_patterns: config.gate_paths
-	});
-	report_changes(config, renderer, changes);
-	if (config.git_checkpoints) git_checkpoint(build_root(config.workspace, config.worktree), `volley iter ${String(built.iteration)}: build`);
+	const root = build_root(config.workspace, config.worktree);
+	const changes = measure_changes(deps, root);
+	if (config.git_checkpoints) git_checkpoint(root, `volley iter ${String(built.iteration)}: build`, state_root_within(root, config.state_dir));
 	return {
 		...built,
 		changes
 	};
 }
+/** `measure`: a critique's one round opens, and the change under review is
+* measured from the baseline, as a run measures the builder's work. */
+function measure_phase(deps, s, ctx) {
+	const next = next_iteration(s);
+	emit(ctx, {
+		volley: "round_start",
+		round: next.iteration
+	});
+	const root = build_root(deps.config.workspace, deps.config.worktree);
+	return {
+		...next,
+		changes: measure_changes(deps, root)
+	};
+}
 /** `skip_verify`: the build alone crossed the cost cap, so the check is recorded
 * as skipped and no critic is called — a doomed iteration spends nothing more.
 * The loop guard then halts the run on the cap. */
@@ -2806,20 +3459,79 @@ function skip_verify_phase(deps, s) {
 		check: skipped_check("cost_cap")
 	};
 }
-/** `check`: the deterministic gate over the tree the builder wrote. */
+/** Open the check's card on a streamed run: the call's start, then the
+* command as its input, the two chunks `model_call` records for a tool call
+* before it runs. */
+function open_check_card(ctx, command) {
+	record_check_chunk(ctx, {
+		kind: "tool_call_start",
+		...CHECK_CALL,
+		step_index: 0
+	});
+	record_check_chunk(ctx, {
+		kind: "tool_call_end",
+		id: CHECK_CALL.id,
+		input: { command },
+		step_index: 0
+	});
+}
+/** Close the check's card with its result, which the card's reader renders,
+* or with the error that stopped the check before it had one. */
+function close_check_card(ctx, outcome) {
+	record_check_chunk(ctx, "result" in outcome ? {
+		kind: "tool_result",
+		id: CHECK_CALL.id,
+		output: outcome.result,
+		step_index: 0
+	} : {
+		kind: "tool_result",
+		id: CHECK_CALL.id,
+		error: { message: outcome.error instanceof Error ? outcome.error.message : String(outcome.error) },
+		step_index: 0
+	});
+}
+/** `check`: the deterministic gate over the tree the builder wrote. A check
+* that runs a command emits its start and end, and on a streamed run also
+* records its card as a tool call, so a reader of the stream shows it running
+* while it runs. When the caller set an approval, the check asks on that card
+* first, and a rejected check closes it failed and is recorded as failed. */
 async function check_phase(deps, s, ctx) {
 	const { config, renderer } = deps;
 	renderer.phase_start(s.iteration, "check");
+	const command = check_command(config);
+	if (command === null) {
+		renderer.info("no deterministic check configured; loop is critic-gated only");
+		return {
+			...s,
+			check: skipped_check("none")
+		};
+	}
+	emit(ctx, {
+		volley: "check_start",
+		round: s.iteration,
+		command
+	});
+	open_check_card(ctx, command);
 	try {
-		const result = await execute_check(config, ctx.abort);
-		if (result.ran) renderer.phase_end(s.iteration, "check", result.ok, result.failing_slots.length > 0 ? `failing: ${result.failing_slots.join(", ")}` : void 0);
-		else renderer.info("no deterministic check configured; loop is critic-gated only");
+		const allowed = await check_allowed(deps, command, ctx.abort);
+		const result = allowed ? await execute_check(deps, ctx.abort) : rejected_check(config.check_resolved);
+		close_check_card(ctx, allowed ? { result } : { error: CHECK_REJECTED });
+		emit(ctx, {
+			volley: "check_end",
+			round: s.iteration,
+			ok: result.ok,
+			exit_code: result.exit_code,
+			failing_slots: result.failing_slots,
+			duration_ms: result.duration_ms
+		});
+		renderer.phase_end(s.iteration, "check", result.ok, !allowed ? "rejected" : result.failing_slots.length > 0 ? `failing: ${result.failing_slots.join(", ")}` : void 0);
 		return {
 			...s,
 			check: result,
 			check_duration_ms: s.check_duration_ms + result.duration_ms
 		};
 	} catch (err) {
+		close_check_card(ctx, { error: err });
 		if (error_kind(err) === "check_error") throw err;
 		throw phase_error("check", s.iteration, err);
 	}
@@ -2832,6 +3544,14 @@ async function critique_phase(deps, critic, s, ctx) {
 		critic,
 		config: deps.config
 	}, s, ctx);
+	if (critiqued.verdict !== null) emit(ctx, {
+		volley: "verdict",
+		round: critiqued.iteration,
+		verdict: critiqued.verdict,
+		feedback: critiqued.feedback ?? "",
+		unmet_criteria: critiqued.unmet_criteria,
+		degraded: critiqued.critic?.critic_degraded === true
+	});
 	renderer.phase_end(critiqued.iteration, "critic", critiqued.verdict === "approved", critiqued.verdict ?? void 0);
 	renderer.cost_line(critiqued.iteration, "critic", critiqued.critic?.cost_usd ?? null, critiqued.total_cost_usd);
 	return critiqued;
@@ -2842,7 +3562,10 @@ function record_phase(deps, s) {
 	const { config } = deps;
 	archive_iteration(config, s);
 	write_run_summary(config, run_result_from_state(config, s, "running", null));
-	if (config.git_checkpoints) git_checkpoint(build_root(config.workspace, config.worktree), `volley iter ${String(s.iteration)}: critique (${s.verdict ?? "skipped"})`);
+	if (config.git_checkpoints) {
+		const root = build_root(config.workspace, config.worktree);
+		git_checkpoint(root, `volley iter ${String(s.iteration)}: critique (${s.verdict ?? "skipped"})`, state_root_within(root, config.state_dir));
+	}
 	return s;
 }
 //#endregion
@@ -2854,6 +3577,11 @@ function record_phase(deps, s) {
 * around the flow (workspace setup, the worktree lifecycle, `run`, the final
 * summary) in `./orchestrator.ts`.
 *
+* A critique (`build_critique_flow`) is one round over a change already made,
+* so it has no builder, no loop, and no guard: `measure`, which opens the
+* round and measures the change from its baseline, then the same `check`,
+* `critique`, and `record` a run's round ends with.
+*
 * fascicle's `describe.diagram` draws the tree below from the flow itself, and
 * `test/unit/flow_diagram.test.ts` holds it to the code: change a row by changing
 * its step's `description`, then paste `pnpm diagram --prefix ' *   '` over it.
@@ -2867,58 +3595,70 @@ function record_phase(deps, s) {
 *   │  │  └─ else  sequence
 *   │  │     ├─ check                     the deterministic gate: checkride, a command, or none
 *   │  │     └─ critique                  the read-only critic's verdict and feedback
-*   │  │        └─ critic                 fallback: judge without tools if a local critic's stream keeps dying
-*   │  │           ├─ retry               one more try after a stream death, local providers only
+*   │  │        └─ critic                 fallback: judge without tools if a critic on any provider but claude_cli keeps dying or runs out of steps
+*   │  │           ├─ retry               one more try after a stream death, on any provider but claude_cli
 *   │  │           │  └─ critic_tools     model call with read-only workspace tools
 *   │  │           └─ pipe                mark the verdict degraded
 *   │  │              └─ critic_toolless  model call judging from the check output and a file list
 *   │  └─ record                          archive the iteration and rewrite the run summary
 *   └─ guard  gate                        stop on approval over a green check, the cost cap, or a gate edit
 */
-/** The critic arm: the degradation ladder. A local critic whose tool-bearing
-* stream dies is retried at once (the death is stochastic parser failure, not
-* load, so waiting buys nothing), then judged once without tools; anything the
-* ladder does not retry propagates from either rung untouched. */
+/** The critic arm: the degradation ladder. A critic on any provider but
+* `claude_cli` whose tool-bearing stream dies is retried at once (the death is
+* stochastic parser failure, not load, so waiting buys nothing), then judged
+* once without tools. One that runs out of steps skips the retry, since
+* another identical tool loop would only run out again, and goes straight to
+* the tool-less pass. Anything else propagates from either rung untouched. */
 function build_critic(deps) {
 	const retryable = critic_retryable(deps.config);
 	return fallback(retry(make_critic_tools_step(deps), {
-		description: "one more try after a stream death, local providers only",
+		description: "one more try after a stream death, on any provider but claude_cli",
 		max_attempts: 2,
 		backoff_ms: 0,
 		when: retryable,
 		project: with_retries
 	}), pipe(make_critic_toolless_step(deps), as_degraded, { description: "mark the verdict degraded" }), {
 		name: "critic",
-		description: "judge without tools if a local critic's stream keeps dying",
-		when: retryable,
+		description: "judge without tools if a critic on any provider but claude_cli keeps dying or runs out of steps",
+		when: critic_degradable(deps.config),
 		handoff: (prompt) => toolless_critic_prompt(deps.config, prompt),
 		project: with_fallback_cause
 	});
 }
+/** The steps a run and a critique share, built one way for both. */
+function check_step(deps) {
+	return step("check", (s, ctx) => check_phase(deps, s, ctx), { description: "the deterministic gate: checkride, a command, or none" });
+}
+function critique_step(deps, seat) {
+	const critic = build_critic({
+		...seat,
+		config: deps.config,
+		on_chunk: deps.renderer.critic_chunk
+	});
+	return step("critique", (s, ctx) => critique_phase(deps, critic, s, ctx), {
+		description: "the read-only critic's verdict and feedback",
+		arm: critic
+	});
+}
+function record_step(deps) {
+	return step("record", (s) => record_phase(deps, s), { description: "archive the iteration and rewrite the run summary" });
+}
 function build_flow(deps, resume_from) {
 	const { config, renderer } = deps;
 	const builder = make_builder_step({
-		engine: deps.engines.builder,
+		...deps.seats.builder,
 		config,
 		on_chunk: renderer.builder_chunk,
 		bash_executor: deps.bash_executor
 	});
-	const critic = build_critic({
-		engine: deps.engines.critic,
-		config,
-		on_chunk: renderer.critic_chunk
-	});
 	const build = step("build", (s, ctx) => build_phase(deps, builder, s, ctx), {
 		description: "open the iteration, run the builder, measure the change",
 		arm: builder
 	});
 	const skip_verify = step("skip_verify", (s) => skip_verify_phase(deps, s), { description: "record the check as skipped and call no critic" });
-	const check = step("check", (s, ctx) => check_phase(deps, s, ctx), { description: "the deterministic gate: checkride, a command, or none" });
-	const critique = step("critique", (s, ctx) => critique_phase(deps, critic, s, ctx), {
-		description: "the read-only critic's verdict and feedback",
-		arm: critic
-	});
-	const record = step("record", (s) => record_phase(deps, s), { description: "archive the iteration and rewrite the run summary" });
+	const check = check_step(deps);
+	const critique = critique_step(deps, deps.seats.critic);
+	const record = record_step(deps);
 	const verify = branch({
 		name: "verify",
 		description: "did the build alone cross the cost cap?",
@@ -2944,6 +3684,19 @@ function build_flow(deps, resume_from) {
 		max_rounds: remaining_iterations(config, resume_from)
 	});
 }
+/** A critique: one round over a change already made. `measure` opens it and
+* measures the change from the baseline, then the check, the critic, and the
+* record, as in a run's round. No loop and no guard: there is nothing to
+* iterate on, and the caller reads the outcome off the final state. */
+function build_critique_flow(deps) {
+	const measure = step("measure", (s, ctx) => measure_phase(deps, s, ctx), { description: "open the round and measure the change under review" });
+	return sequence([
+		measure,
+		check_step(deps),
+		critique_step(deps, deps.critic),
+		record_step(deps)
+	], { name: "critique_run" });
+}
 //#endregion
 //#region src/orchestrator.ts
 /**
@@ -2954,19 +3707,25 @@ function build_flow(deps, resume_from) {
 * substrate.
 *
 *   run_volley
-*   ├─ initialize the workspace (or preserve it, on resume)
+*   ├─ initialize the run's state (or preserve it, on resume)
 *   ├─ with_worktree                    off unless --worktree
 *   │  ├─ capture the change baseline
-*   │  ├─ run_loop                      run(build_flow(...)), then the final summary
+*   │  ├─ run_loop                      run the flow (streamed when a sink is given), then the final summary
 *   │  └─ integrate                     a successful --worktree --git run only
 *   └─ report a kept checkout, or a salvaged branch (re-stamping the summary)
+*
+*   run_critique
+*   ├─ resolve the baseline              --baseline, else HEAD; a bad revision exits 5
+*   ├─ initialize the run's state
+*   └─ run the critique flow             one round on the critic's engine alone, then the final summary
 */
 async function run_volley(config, deps, resume_from = null) {
 	const { renderer } = deps;
 	if (resume_from === null) {
-		initialize_workspace(config.workspace);
-		write_resolved_config(config);
-	} else initialize_workspace(config.workspace, { preserve: true });
+		initialize_run_state(config);
+		if (!write_resolved_config(config) && !detect_containment(process.env)) renderer.warn(`could not sign this run's config with this machine's key (${trust_key_path(process.env)}), so resuming it here will need --trust-config`);
+	} else initialize_run_state(config, { preserve: true });
+	renderer.info(`state: ${run_state_dir(config)}`);
 	report_worktree_fate(config, renderer);
 	const branch = worktree_branch(config.run_id);
 	const fate = worktree_fate(config);
@@ -3007,18 +3766,24 @@ async function run_loop(config, deps, resume_from, bash_executor, baseline) {
 		critic: deps.engine
 	};
 	const flow = build_flow({
-		engines,
+		seats: {
+			builder: {
+				...deps.seats?.builder,
+				engine: engines.builder
+			},
+			critic: {
+				...deps.seats?.critic,
+				engine: engines.critic
+			}
+		},
 		config,
 		renderer,
 		bash_executor,
-		baseline
+		baseline,
+		...deps.check !== void 0 ? { check: deps.check } : {}
 	}, resume_from);
 	try {
-		const { value, converged } = await run(flow, { resume_from }, {
-			trajectory: filesystem_logger({ output_path: volley_path(config.workspace, "trajectory.jsonl") }),
-			...deps.abort !== void 0 ? { abort: deps.abort } : {},
-			...deps.install_signal_handlers !== void 0 ? { install_signal_handlers: deps.install_signal_handlers } : {}
-		});
+		const { value, converged } = await execute_flow(config, deps, flow, { resume_from });
 		const result = run_result_from_state(config, value, status_of(value, converged), (/* @__PURE__ */ new Date()).toISOString());
 		write_run_summary(config, result);
 		return result;
@@ -3029,6 +3794,45 @@ async function run_loop(config, deps, resume_from, bash_executor, baseline) {
 		await dispose_seat_engines(engines);
 	}
 }
+/** Run the flow into its file trajectory and, when the caller passed one, its
+* own sink. Without a sink this is the one `run` it has always been. With one,
+* it goes through `run.stream`, because only a streamed run records each model
+* call's chunks as `model_chunk` events, which are what an outer consumer maps
+* to text, tool cards, and approval prompts. The sink is that consumer, so the
+* handle's own event stream is closed at once rather than left to buffer, and
+* the file skips the chunks, so it holds the same kinds of event either way.
+*
+* The file goes first in the tee because the first sink's span ids are the
+* ones the run hands out as parents, so the file's parent links stay whole.
+* The sink sees parent ids it never issued, which costs an ACP consumer
+* nothing: it counts a step's runs by the `id` on each `span_start`. */
+function execute_flow(config, deps, flow, input) {
+	const file = filesystem_logger({ output_path: state_path(config, "trajectory.jsonl") });
+	const options = {
+		...deps.abort !== void 0 ? { abort: deps.abort } : {},
+		...deps.install_signal_handlers !== void 0 ? { install_signal_handlers: deps.install_signal_handlers } : {}
+	};
+	if (deps.trajectory === void 0) return run(flow, input, {
+		...options,
+		trajectory: file
+	});
+	const handle = run.stream(flow, input, {
+		...options,
+		trajectory: tee_logger(without_model_chunks(file), deps.trajectory)
+	});
+	handle.events[Symbol.asyncIterator]().return?.();
+	return handle.result;
+}
+/** A logger that passes `inner` every event but a model call's streamed
+* chunks. */
+function without_model_chunks(inner) {
+	return {
+		...inner,
+		record: (event) => {
+			if (event.kind !== "model_chunk") inner.record(event);
+		}
+	};
+}
 /** Best-effort status stamp on the run summary when the loop threw; the
 * on-disk iteration state is already consistent (record runs per round). */
 function finalize_failed_run(config, err) {
@@ -3036,59 +3840,64 @@ function finalize_failed_run(config, err) {
 	const cause = kind === "phase_error" ? error_kind(err.cause) : null;
 	const status = kind === "aborted_error" || cause === "aborted_error" ? "interrupted" : "error";
 	try {
-		const summary_path = volley_path(config.workspace, "summary.json");
+		const summary_path = state_path(config, "summary.json");
 		write_run_summary(config, {
 			...existsSync(summary_path) ? JSON.parse(readFileSync(summary_path, "utf8")) : run_result_from_state(config, initial_state(), status, null),
+			run_state_dir: run_state_dir(config),
 			status,
 			completed_at: (/* @__PURE__ */ new Date()).toISOString()
 		});
 	} catch {}
 }
-//#endregion
-//#region src/render/format.ts
 /**
-* Display formatting: glyph taxonomy, role colors, truncation.
-* All human output goes to stderr; these helpers only build strings.
+* `volley critique`: the check and the critic, once, over a change already
+* made. No builder is constructed, and the engine serves the critic's seat
+* alone, so a critic can be measured on fixed, labeled changes without paying
+* for builder turns. It writes the state a one-round run does, with the mode
+* recorded, and ends `success` on approval over a green check, else
+* `rejected`.
 */
-const GLYPH_TOOL = "🔧";
-const GLYPH_COST = "💰";
-const VERBOSE_TRUNCATE_CHARS = 4e3;
-function paint(role, text, color) {
-	if (!color) return text;
-	switch (role) {
-		case "builder": return pc.cyan(text);
-		case "critic": return pc.yellow(text);
-		case "check": return pc.green(text);
-		case "cost": return pc.magenta(text);
-		case "error": return pc.bold(pc.red(text));
-	}
-}
-function dim(text, color) {
-	return color ? pc.dim(text) : text;
-}
-function truncate(text, max, pointer) {
-	if (text.length <= max) return text;
-	return `${text.slice(0, max)}\n… [truncated at ${String(max)} chars; ${pointer}]`;
-}
-function format_usd(value) {
-	return value === null ? "$?" : `$${value.toFixed(3)}`;
-}
-/** One-line summary of a tool input for the default streaming mode. */
-function summarize_value(value, max = 120) {
-	let text;
+async function run_critique(config, deps) {
+	if (config.mode !== "critique") throw config_error("run_critique needs a config resolved with mode 'critique'");
+	const { renderer } = deps;
+	const baseline = resolve_baseline(config.workspace, deps.baseline);
+	initialize_run_state(config);
+	write_resolved_config(config);
+	renderer.info(`state: ${run_state_dir(config)}`);
+	renderer.info(baseline === null ? "baseline: none (not a git repository, or no commit yet); the critic reviews the tree" : `baseline: ${baseline}`);
+	const engine = deps.engine ?? create_volley_engine({
+		workspace: config.workspace,
+		critic_provider: config.critic_provider,
+		critic_url: config.critic_url
+	});
+	const flow = build_critique_flow({
+		config,
+		renderer,
+		baseline,
+		critic: {
+			...deps.critic,
+			engine
+		},
+		...deps.check !== void 0 ? { check: deps.check } : {}
+	});
 	try {
-		text = typeof value === "string" ? value : JSON.stringify(value) ?? "";
-	} catch {
-		text = String(value);
+		const value = await execute_flow(config, deps, flow, initial_state());
+		const result = run_result_from_state(config, value, value.check?.ok === true && value.verdict === "approved" ? "success" : "rejected", (/* @__PURE__ */ new Date()).toISOString());
+		write_run_summary(config, result);
+		return result;
+	} catch (err) {
+		finalize_failed_run(config, err);
+		throw err;
+	} finally {
+		await engine.dispose();
 	}
-	const flat = text.replace(/\s+/g, " ");
-	return flat.length > max ? `${flat.slice(0, max)}…` : flat;
 }
 //#endregion
 //#region src/render/renderer.ts
 function create_renderer(options) {
 	const write = options.write ?? ((text) => process.stderr.write(text));
 	const { mode, color } = options;
+	const see_trajectory = `see ${options.trajectory_path ?? "the run's trajectory.jsonl"}`;
 	const silent = mode === "json";
 	const streaming = mode === "default" || mode === "verbose";
 	let mid_line = false;
@@ -3110,7 +3919,7 @@ function create_renderer(options) {
 				mid_line = !chunk.text.endsWith("\n");
 				return;
 			case "tool_call_end": {
-				const summary = mode === "verbose" ? truncate(JSON.stringify(chunk.input, null, 2) ?? "", VERBOSE_TRUNCATE_CHARS, "see .volley/trajectory.jsonl") : summarize_value(chunk.input);
+				const summary = mode === "verbose" ? truncate(JSON.stringify(chunk.input, null, 2) ?? "", VERBOSE_TRUNCATE_CHARS, see_trajectory) : summarize_value(chunk.input);
 				line(`${GLYPH_TOOL} ${paint(role, chunk.id, color)} ${summary}`);
 				return;
 			}
@@ -3118,7 +3927,7 @@ function create_renderer(options) {
 				const ok = chunk.error === void 0;
 				if (mode === "verbose") {
 					const body = ok ? summarize_output(chunk.output) : chunk.error?.message ?? "";
-					line(`${ok ? "✅" : "❌"} ${truncate(body, VERBOSE_TRUNCATE_CHARS, "see .volley/trajectory.jsonl")}`);
+					line(`${ok ? "✅" : "❌"} ${truncate(body, VERBOSE_TRUNCATE_CHARS, see_trajectory)}`);
 				} else line(`${ok ? "✅" : "❌"} ${ok ? "" : summarize_value(chunk.error?.message)}`);
 				return;
 			}
@@ -3166,6 +3975,7 @@ function create_renderer(options) {
 			line(`iterations: ${String(result.iterations_completed)} | verdict: ${result.final_verdict ?? "n/a"}`);
 			line(paint("cost", `${GLYPH_COST} total: ${format_usd(result.total_cost_usd)} (builder ${format_usd(result.builder_cost_usd)}, critic ${format_usd(result.critic_cost_usd)})`, color));
 			line(`tokens: ${String(result.total_usage.input_tokens)} in / ${String(result.total_usage.output_tokens)} out`);
+			line(`state: ${result.run_state_dir}`);
 		}
 	};
 }
@@ -3187,7 +3997,7 @@ function summarize_output(output) {
 * volley detects it is already contained. The create/`exec`/reap/teardown
 * lifecycle and the `docker exec` bash executor were retired; what
 * remains here is the *spec* — pure, daemon-free flag-builders the blessed
-* examples and the `--dry-run` preflight render into the `docker run` line,
+* examples and a parent harness render into the `docker run` line,
 * unit-testable without a daemon.
 *
 * Inside that container volley's `bash` is the local `host_bash_executor`
@@ -3212,6 +4022,13 @@ const SANDBOX_WORKDIR = "/workspace";
 * otherwise unset) and its `.cache` is a tmpfs so the read-only rootfs does not
 * brick tools that scribble there. */
 const SANDBOX_HOME = "/home/node";
+/** Where volley inside the container keeps each run's state when no
+* `--state-dir` names a root: the default root under the pinned `HOME`, since
+* the image sets no `XDG_STATE_HOME`. `sandbox_run_args`'s `run_state` binds one
+* run's host directory at that run's directory under it, so the state outlives
+* the `--rm` container. Exported so a harness can map a contained run's
+* `run_state_dir` back to the directory it mounted. */
+const SANDBOX_STATE_ROOT = default_state_root({}, SANDBOX_HOME);
 /** The host spelling of the Docker gateway (Linux-portable; Docker Desktop
 * already provides it). It is both the `--add-host` target for the allowlist
 * bridge and the value the example injects as `VOLLEY_MODEL_HOST` so the
@@ -3243,7 +4060,7 @@ function network_run_args(network, network_name) {
 /**
 * The hardened `docker run` argv that launches volley inside its container.
 * Pure and exported so the flag set is unit-testable without a
-* daemon and the blessed examples / `--dry-run` preflight can render it.
+* daemon and the blessed examples and a parent harness can render it.
 *
 * A one-shot, foreground `--rm` container — the whole run *is* this
 * container (an earlier design ran a detached `-d … sleep infinity` daemon volley
@@ -3255,6 +4072,25 @@ function network_run_args(network, network_name) {
 * the image entrypoint (the opt-in real-docker isolation test runs a raw `sh`
 * this way).
 *
+* `run_state` binds the run's own state directory on the host,
+* `<root>/<run_id>/`, at `<SANDBOX_STATE_ROOT>/<run_id>/`, where volley inside
+* keeps that run's state when no `--state-dir` names a root, so it outlives the
+* container at the path a run on the host would use. Only the one directory
+* is mounted, never the root: the builder inside runs arbitrary commands, so
+* it can rewrite its own run's state but can't read another run's trajectory
+* or plant a check in another run's config. Its own run's state is still the
+* builder's to rewrite, config.json included, so a host-side resume refuses
+* that config, which volley signs only outside the sandbox (`./trust.ts`), and
+* a directory holding a symlink: resume a contained run inside the container,
+* with the same mount. Create the
+* directory on the host first, empty (Docker on Linux makes a missing bind
+* source owned by root, which the container's user can't write), and pass the
+* same id to volley as `--run-id`. Keep it outside `build_root`, so the check
+* never reads it, and pass no `XDG_STATE_HOME` in `env`, which would move the
+* root off the mount. Omit it and no state is mounted: a contained run keeping
+* its state under the default root then stops before spending (exit 5),
+* since the read-only root filesystem leaves it nowhere to write.
+*
 * `store_volume` mounts the named volume as the pnpm store; omit it and no store
 * is mounted. A store is shared by every run that mounts it, so a caller that
 * installs dependencies in a separate container first can keep the builder's
@@ -3264,7 +4100,7 @@ function network_run_args(network, network_name) {
 * Never emits `--privileged` and refuses `--user 0`: volley must not hand
 * the sandbox root or full capabilities. Everything writable under the
 * `--read-only` rootfs is an explicit tmpfs or volume; the bind-mounted worktree
-* is the one host-visible write path.
+* and, when one is named, the state root are the host-visible write paths.
 */
 function sandbox_run_args(options) {
 	if (options.uid === 0) throw config_error("refusing to start the builder sandbox as root (uid 0)");
@@ -3273,6 +4109,7 @@ function sandbox_run_args(options) {
 	const entrypoint = options.entrypoint !== void 0 && options.entrypoint !== null ? ["--entrypoint", options.entrypoint] : [];
 	const env = (options.env ?? []).flatMap(([key, value]) => ["-e", `${key}=${value}`]);
 	const labels = (options.labels ?? []).flatMap(([key, value]) => ["--label", `${key}=${value}`]);
+	const state = options.run_state !== void 0 && options.run_state !== null ? ["-v", run_state_mount(options.run_state)] : [];
 	const store = options.store_volume !== void 0 && options.store_volume !== null ? ["-v", `${options.store_volume}:${SANDBOX_HOME}/.local/share/pnpm/store`] : [];
 	return [
 		"run",
@@ -3306,6 +4143,7 @@ function sandbox_run_args(options) {
 		...env,
 		"-v",
 		`${resolve(options.build_root)}:${SANDBOX_WORKDIR}`,
+		...state,
 		...store,
 		"-w",
 		SANDBOX_WORKDIR,
@@ -3314,6 +4152,13 @@ function sandbox_run_args(options) {
 		...options.command ?? []
 	];
 }
+/** The `-v` value for one run's state directory. The run id becomes a path on
+* both sides, so it is held to volley's own `--run-id` rule, which keeps it to
+* one directory name. */
+function run_state_mount(run_state) {
+	if (!RUN_ID_PATTERN.test(run_state.run_id)) throw config_error(`run_state.run_id must be letters, digits, '-', and '_', starting with a letter or digit; got: ${run_state.run_id}`);
+	return `${join(resolve(run_state.root), run_state.run_id)}:${SANDBOX_STATE_ROOT}/${run_state.run_id}`;
+}
 //#endregion
 //#region src/exit_codes.ts
 /**
@@ -3331,7 +4176,8 @@ const EXIT_INTERRUPTED = 130;
 function exit_code_for_status(status) {
 	switch (status) {
 		case "success": return 0;
-		case "budget_exhausted": return 2;
+		case "budget_exhausted":
+		case "rejected": return 2;
 		case "cost_cap_reached": return 7;
 		case "gate_edit_blocked": return 8;
 		case "interrupted": return 130;
@@ -3366,6 +4212,6 @@ function exit_code_for_error(err) {
 	return 3;
 }
 //#endregion
-export { EXIT_BUDGET_EXHAUSTED, EXIT_BUILDER_ERROR, EXIT_CHECK_ERROR, EXIT_CONFIG_ERROR, EXIT_COST_CAP, EXIT_CRITIC_ERROR, EXIT_INTERRUPTED, EXIT_SUCCESS, SANDBOX_NETWORK_NAME, SANDBOX_NETWORK_SUBNET, create_renderer, exit_code_for_error, exit_code_for_status, expand_at_file, gate, initial_state, load_config_file, network_run_args, resolve_config, run_volley, sandbox_run_args, status_of, verdict_schema };
+export { EXIT_BUDGET_EXHAUSTED, EXIT_BUILDER_ERROR, EXIT_CHECK_ERROR, EXIT_CONFIG_ERROR, EXIT_COST_CAP, EXIT_CRITIC_ERROR, EXIT_INTERRUPTED, EXIT_SUCCESS, SANDBOX_NETWORK_NAME, SANDBOX_NETWORK_SUBNET, SANDBOX_STATE_ROOT, create_renderer, default_state_root, exit_code_for_error, exit_code_for_status, expand_at_file, gate, initial_state, load_config_file, network_run_args, resolve_config, run_critique, run_volley, sandbox_run_args, status_of, verdict_schema };
 
 //# sourceMappingURL=index.js.map