phoebe-agent 0.0.0 → 0.1.0

package/bootstrap/crash-loop.ts

@@ -0,0 +1,391 @@
+// Crash-loop fallback for `phoebe boot` — the guard on a bad engine ref.
+//
+// The reconcile watch (reconcile.ts) will happily follow the tracked branch onto
+// any commit someone pushes, including one that dies on startup. Without a guard
+// that is an unattended container crash-looping on an unattended repo. So boot
+// keeps a small record of which engine commits actually *ran*: a SHA that dies
+// fast, repeatedly, is quarantined, and boot pins back to the last SHA that ran
+// healthily until the tracked branch moves past the bad one.
+//
+// This module is that record and the decisions over it — pure folds plus a JSON
+// file — so the whole ladder (first crash → threshold → fallback → recovery) is
+// tested without spawning an engine. The wiring (when a run ends, what to
+// materialize, what to log) lives in boot.ts; the loop that calls it is
+// reconcile.ts.
+//
+// The record lives in the engine's state dir (`paths.stateDir`, default
+// `/data/state` — a named volume) rather than beside the engine clone, so a
+// quarantine survives a container restart *and* a wiped engine volume: the whole
+// point is to remember across the restart the crash itself causes.
+//
+// Only a github source with a moving branch is guarded. A local mount has no
+// commit to pin, and a pinned ref means pinning — boot must not silently serve
+// different code than the operator asked for (see boot.ts's eligibility check).
+//
+// This is the only crash-loop policy there is: the engine's own self-update and
+// the shell supervisor that mirrored a copy of these decisions are gone (#44).
+
+import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { dirname, join } from "node:path";
+
+/**
+ * Consecutive fast crashes of one engine SHA before boot pins back to the
+ * last-good one. Three is enough to rule out a one-off (a flaky network on the
+ * engine's first fetch, a busy host) while still recovering in minutes.
+ */
+export const CRASH_LOOP_THRESHOLD = 3;
+
+/**
+ * How long an engine run must survive to count as healthy. A commit that cannot
+ * boot dies well inside this; a long-lived engine that later hits a runtime
+ * error does not, and pinning to older code would not help it.
+ */
+export const HEALTHY_RUN_MS = 60_000;
+
+/** State dir used when the mounted config names none (matches the engine's default). */
+export const DEFAULT_STATE_DIR = "/data/state";
+
+/** Filename of the crash-loop record inside the state dir. */
+export const CRASH_LOOP_STATE_FILE = "engine-crash-loop.json";
+
+/** Crash-loop bookkeeping, persisted across container restarts on the state volume. */
+export type CrashLoopState = {
+  /** SHA that last ran healthily — the fallback target. */
+  lastGoodSha: string | null;
+  /** SHA currently accumulating fast-crash counts (or quarantined as bad). */
+  failingSha: string | null;
+  /** Consecutive fast crashes recorded for `failingSha`. */
+  failureCount: number;
+};
+
+/** Nothing known yet: no proven commit, no quarantine. */
+export const INITIAL_CRASH_LOOP_STATE: CrashLoopState = {
+  lastGoodSha: null,
+  failingSha: null,
+  failureCount: 0,
+};
+
+/** A finished engine run, as the fallback policy sees it. */
+export type RunOutcome = {
+  /** The engine commit that was running. */
+  sha: string;
+  /** Its exit code; null when a signal killed it. */
+  exitCode: number | null;
+  /** How long it lived. */
+  elapsedMs: number;
+  /**
+   * Whether boot ended this run — a reconcile drain or a container stop — rather
+   * than the engine exiting of its own accord. Load-bearing: a run boot cut
+   * short proves nothing either way.
+   */
+  requestedStop: boolean;
+};
+
+/** What a finished run proved about the commit it was running. */
+export type RunVerdict =
+  /** The commit works — it lived past the healthy window, or finished its own work. */
+  | "healthy"
+  /** The commit dies on startup: the thing the fallback exists to catch. */
+  | "crash"
+  /** Nothing was proved, so the record must not move. */
+  | "inconclusive";
+
+/**
+ * Judge a finished run. The three-way answer matters: a run that neither proved
+ * nor disproved its commit has to leave the record alone, and reading it as
+ * "healthy" would be the worse mistake — a container stop landing mid-crash-loop
+ * would promote the crashing commit to last-good and disarm the fallback for
+ * good.
+ */
+export function judgeRun(run: RunOutcome, healthyMs: number = HEALTHY_RUN_MS): RunVerdict {
+  // Living past the window proves the commit boots, however the run then ended.
+  if (run.elapsedMs >= healthyMs) return "healthy";
+  // Boot pulled the plug before the window was up: no verdict either way.
+  if (run.requestedStop) return "inconclusive";
+  // Exiting 0 unprompted means the engine finished what it was asked to do
+  // (`--run-once`), which it could only do by booting.
+  if (run.exitCode === 0) return "healthy";
+  // A signal from outside — an OOM kill, a `docker kill` — says nothing about
+  // the code either.
+  if (run.exitCode === null) return "inconclusive";
+  return "crash";
+}
+
+/**
+ * Fold a finished run into the record. A healthy run becomes the new last-good,
+ * but a quarantine on a *different* SHA is preserved — the healthy run is
+ * usually the fallback itself, and clearing the record there would send the next
+ * launch straight back into the commit it is avoiding. A fast crash counts
+ * against its own SHA, starting over whenever the failing SHA moves — except
+ * against an *active* quarantine, which outlives it.
+ */
+export function recordRun(
+  state: CrashLoopState,
+  run: RunOutcome,
+  opts: { healthyMs?: number; threshold?: number } = {},
+): CrashLoopState {
+  const threshold = opts.threshold ?? CRASH_LOOP_THRESHOLD;
+  switch (judgeRun(run, opts.healthyMs ?? HEALTHY_RUN_MS)) {
+    case "inconclusive":
+      return state;
+    case "healthy": {
+      const stillQuarantining = state.failingSha !== null && state.failingSha !== run.sha;
+      return {
+        lastGoodSha: run.sha,
+        failingSha: stillQuarantining ? state.failingSha : null,
+        failureCount: stillQuarantining ? state.failureCount : 0,
+      };
+    }
+    case "crash": {
+      if (state.failingSha === run.sha) return { ...state, failureCount: state.failureCount + 1 };
+      // The fallback crashing too does not exonerate the commit it is standing
+      // in for: the quarantine has to hold until the tracked ref moves past the
+      // bad SHA, so a crash of some other commit must not overwrite it.
+      if (state.failingSha !== null && state.failureCount >= threshold) return state;
+      return { lastGoodSha: state.lastGoodSha, failingSha: run.sha, failureCount: 1 };
+    }
+  }
+}
+
+/**
+ * Is there a better commit to relaunch onto than the one that just crashed?
+ * Answers boot's question after a fast crash: retry (the fallback is reachable,
+ * even if the threshold is not met yet) or give up and let the container exit.
+ * Without this, a first-ever bad ref — or a fallback that crashes too — would
+ * relaunch forever instead of failing where an operator can see it.
+ */
+export function hasFallbackTarget(state: CrashLoopState, crashedSha: string): boolean {
+  return state.lastGoodSha !== null && state.lastGoodSha !== crashedSha;
+}
+
+/**
+ * The commit to run *instead of* `targetSha` (the tracked ref's current tip), or
+ * null to run the target as normal. Non-null only once the target has
+ * fast-crashed `threshold` times and a different known-good commit exists — so
+ * the fallback lapses by itself the moment the branch advances past the bad SHA,
+ * which is exactly "fallback persists until the ref moves on".
+ */
+export function fallbackSha(
+  targetSha: string,
+  state: CrashLoopState,
+  threshold: number = CRASH_LOOP_THRESHOLD,
+): string | null {
+  if (state.failingSha !== targetSha || state.failureCount < threshold) return null;
+  return hasFallbackTarget(state, targetSha) ? state.lastGoodSha : null;
+}
+
+/**
+ * Where the engine keeps its persistent state, read straight from the mounted
+ * config. The bootstrapper resolves this itself — it writes the record before
+ * the engine (which owns the config schema) has even been materialized — so a
+ * missing or malformed `paths.stateDir` falls back to the engine's own default
+ * rather than being trusted or rejected.
+ */
+export function readStateDir(config: Record<string, unknown>): string {
+  const paths = config["paths"];
+  if (paths === null || typeof paths !== "object") return DEFAULT_STATE_DIR;
+  const stateDir = (paths as Record<string, unknown>)["stateDir"];
+  return typeof stateDir === "string" && stateDir.length > 0 ? stateDir : DEFAULT_STATE_DIR;
+}
+
+/** The crash-loop record's path inside a state dir. */
+export function crashLoopStatePath(stateDir: string): string {
+  return join(stateDir, CRASH_LOOP_STATE_FILE);
+}
+
+function isCrashLoopState(value: unknown): value is CrashLoopState {
+  if (value === null || typeof value !== "object") return false;
+  const state = value as Record<string, unknown>;
+  return (
+    (state["lastGoodSha"] === null || typeof state["lastGoodSha"] === "string") &&
+    (state["failingSha"] === null || typeof state["failingSha"] === "string") &&
+    typeof state["failureCount"] === "number"
+  );
+}
+
+/**
+ * Read the record, or start fresh. Anything unreadable — no file yet, a
+ * half-written one from a container killed mid-write, a hand-edited mess —
+ * degrades to "nothing known" rather than failing boot: the cost is losing one
+ * fallback target, and the alternative is a bootstrapper that a bad file bricks.
+ */
+export function readCrashLoopState(path: string): CrashLoopState {
+  try {
+    const parsed: unknown = JSON.parse(readFileSync(path, "utf8"));
+    if (!isCrashLoopState(parsed)) return INITIAL_CRASH_LOOP_STATE;
+    return {
+      lastGoodSha: parsed.lastGoodSha,
+      failingSha: parsed.failingSha,
+      failureCount: parsed.failureCount,
+    };
+  } catch {
+    return INITIAL_CRASH_LOOP_STATE;
+  }
+}
+
+/**
+ * Persist the record, creating the state dir if the volume is empty. Throws on a
+ * genuinely unwritable state dir; the guard turns that into an event, since
+ * losing the fallback is better than refusing to run the engine.
+ */
+export function writeCrashLoopState(path: string, state: CrashLoopState): void {
+  mkdirSync(dirname(path), { recursive: true });
+  writeFileSync(path, `${JSON.stringify(state, null, 2)}\n`);
+}
+
+/**
+ * Everything the guard decides that is worth telling an operator about. Emitted
+ * rather than logged so the policy stays testable and boot.ts owns the wording —
+ * a container silently serving older code than its config asks for is exactly
+ * the confusing situation these lines exist to prevent.
+ */
+export type CrashGuardEvent =
+  | {
+      kind: "crash";
+      sha: string;
+      exitCode: number | null;
+      elapsedMs: number;
+      failureCount: number;
+      threshold: number;
+    }
+  /** A run proved its commit; it is now the fallback target. */
+  | { kind: "last-good"; sha: string }
+  /** This launch is pinned to `lastGoodSha` instead of the tracked ref's tip. */
+  | { kind: "fallback"; quarantinedSha: string; lastGoodSha: string; failureCount: number }
+  /** The fallback crashed too. The quarantine holds; boot is out of options. */
+  | {
+      kind: "fallback-crashed";
+      sha: string;
+      quarantinedSha: string;
+      exitCode: number | null;
+      elapsedMs: number;
+    }
+  /** The tracked ref moved past the quarantined commit; the pin lapses. */
+  | { kind: "recovered"; quarantinedSha: string; sha: string }
+  | { kind: "persist-failed"; path: string; error: unknown };
+
+/**
+ * The crash-loop guard boot supervises with: the persisted record, held in
+ * memory for the life of the container and written through on every change.
+ */
+export type CrashGuard = {
+  /**
+   * The commit to run instead of `targetSha` (the tracked ref's tip), or null to
+   * run the target. Called once per launch, before the engine is spawned.
+   */
+  fallbackFor: (targetSha: string) => string | null;
+  /** Fold a finished run into the record and persist it. */
+  record: (run: RunOutcome) => void;
+  /**
+   * A commit is still running, and has been for `elapsedMs`. Once that passes
+   * the healthy window the commit has proved itself, so it is banked as the
+   * fallback target *while it runs* — an engine that has been up for a month and
+   * is then killed by a host reboot would otherwise have left no record at all,
+   * and the bad commit waiting on the branch would have nothing to fall back to.
+   */
+  noteAlive: (sha: string, elapsedMs: number) => void;
+  /**
+   * Is relaunching after this run worth it? Only for a crash, and only when a
+   * *different* known-good commit exists to end up on — otherwise boot lets the
+   * container exit rather than loop on the same broken commit forever. The
+   * answer does not depend on whether `record` ran first: folding a crash in
+   * never changes which commit is last-good.
+   */
+  shouldRetry: (run: RunOutcome) => boolean;
+};
+
+export function createCrashGuard(deps: {
+  /** The record's path, resolved from the mounted config at boot. */
+  statePath: string;
+  onEvent?: (event: CrashGuardEvent) => void;
+  threshold?: number;
+  healthyMs?: number;
+}): CrashGuard {
+  const threshold = deps.threshold ?? CRASH_LOOP_THRESHOLD;
+  const healthyMs = deps.healthyMs ?? HEALTHY_RUN_MS;
+  const options = { threshold, healthyMs };
+  const emit = (event: CrashGuardEvent) => deps.onEvent?.(event);
+  let state = readCrashLoopState(deps.statePath);
+
+  const persist = (next: CrashLoopState) => {
+    state = next;
+    try {
+      writeCrashLoopState(deps.statePath, next);
+    } catch (error) {
+      // In-memory bookkeeping still guards this container's life; only the
+      // across-restart half is lost.
+      emit({ kind: "persist-failed", path: deps.statePath, error });
+    }
+  };
+
+  return {
+    fallbackFor(targetSha) {
+      const pin = fallbackSha(targetSha, state, threshold);
+      if (pin !== null) {
+        emit({
+          kind: "fallback",
+          quarantinedSha: targetSha,
+          lastGoodSha: pin,
+          failureCount: state.failureCount,
+        });
+        return pin;
+      }
+      if (state.failingSha !== null && state.failingSha !== targetSha) {
+        // The tracked ref has moved on, so whatever we knew about the old commit
+        // is history. Dropping it here (rather than leaving it for the next
+        // crash to overwrite) is what makes the fallback lapse exactly once.
+        const quarantinedSha = state.failingSha;
+        const wasQuarantined = state.failureCount >= threshold;
+        persist({ lastGoodSha: state.lastGoodSha, failingSha: null, failureCount: 0 });
+        if (wasQuarantined) emit({ kind: "recovered", quarantinedSha, sha: targetSha });
+      }
+      return null;
+    },
+
+    record(run) {
+      const before = state;
+      const verdict = judgeRun(run, healthyMs);
+      persist(recordRun(before, run, options));
+
+      if (verdict === "inconclusive") return;
+      if (verdict === "healthy") {
+        if (state.lastGoodSha !== before.lastGoodSha) emit({ kind: "last-good", sha: run.sha });
+        return;
+      }
+      if (
+        before.failingSha !== null &&
+        before.failingSha !== run.sha &&
+        before.failureCount >= threshold
+      ) {
+        // A crash the quarantine swallowed — this is the fallback itself dying,
+        // which reads very differently in a log to the tracked ref dying.
+        emit({
+          kind: "fallback-crashed",
+          sha: run.sha,
+          quarantinedSha: before.failingSha,
+          exitCode: run.exitCode,
+          elapsedMs: run.elapsedMs,
+        });
+        return;
+      }
+      emit({
+        kind: "crash",
+        sha: run.sha,
+        exitCode: run.exitCode,
+        elapsedMs: run.elapsedMs,
+        failureCount: state.failureCount,
+        threshold,
+      });
+    },
+
+    noteAlive(sha, elapsedMs) {
+      if (elapsedMs < healthyMs || state.lastGoodSha === sha) return;
+      persist(recordRun(state, { sha, exitCode: null, elapsedMs, requestedStop: false }, options));
+      emit({ kind: "last-good", sha });
+    },
+
+    shouldRetry(run) {
+      return judgeRun(run, healthyMs) === "crash" && hasFallbackTarget(state, run.sha);
+    },
+  };
+}

package/bootstrap/define-config.ts

@@ -0,0 +1,20 @@
+// `defineConfig` — the identity typing helper consumers import from
+// `phoebe-agent` to author their `phoebe.config.ts`:
+//
+// ```ts
+// import { defineConfig } from "phoebe-agent";
+// export default defineConfig({ repoSlug: "...", ... });
+// ```
+//
+// It lives in the bootstrapper (the published package surface) rather than the
+// engine because the bootstrapper is what a consumer installs. The config shape
+// itself (`PhoebeUserConfig`, including the bootstrapper-only `engine` field) is
+// owned by the engine's config schema; this is only the typing helper. The
+// value is never read at runtime beyond being forwarded — the whole benefit is
+// editor autocomplete and a compile-time check that only known fields appear.
+
+import type { PhoebeUserConfig } from "../src/config-schema.ts";
+
+export function defineConfig(config: PhoebeUserConfig): PhoebeUserConfig {
+  return config;
+}

package/bootstrap/engine-source.ts

@@ -0,0 +1,83 @@
+// The bootstrapper's minimal engine-source reader.
+//
+// The published `phoebe-agent` package is a thin bootstrapper: it materializes
+// the engine (`src/`) from a git ref or a local mount, then runs it. The only
+// part of the consumer's mounted `phoebe.config.ts` the bootstrapper needs is
+// the `engine` field — the full config schema stays the engine's concern
+// (config-schema.ts + resolveConfig). This module is that minimal reader: it
+// takes the `engine` field (or a loaded config carrying it) and returns the
+// resolved source with defaults applied, and nothing else.
+//
+// The user-facing field type (`EngineSourceField`) lives with the rest of the
+// config schema in the engine so there is one source of truth for the config
+// shape; this module owns only the resolution + the defaults.
+
+import type { EngineSourceField } from "../src/config-schema.ts";
+
+/** Repo the engine is cloned from when `source: "github"` omits `repo`. */
+export const DEFAULT_ENGINE_REPO = "JesusFilm/phoebe";
+/** Ref the engine is checked out at when `source: "github"` omits `ref`. */
+export const DEFAULT_ENGINE_REF = "main";
+
+/**
+ * The engine source with every default applied — what the bootstrapper acts on.
+ * `github` always carries a concrete `ref` and `repo`; `local` carries nothing
+ * (the engine is read from its mount).
+ */
+export type ResolvedEngineSource =
+  | { source: "github"; ref: string; repo: string }
+  | { source: "local" };
+
+/**
+ * Resolve the optional `engine` field into a concrete source. An omitted field
+ * (or an omitted `ref`/`repo` on a github source) falls back to the shipped
+ * defaults: github, ref `main`, repo `JesusFilm/phoebe`.
+ */
+export function resolveEngineSource(field: EngineSourceField | undefined): ResolvedEngineSource {
+  if (field === undefined || field.source === "github") {
+    return {
+      source: "github",
+      ref: field?.ref ?? DEFAULT_ENGINE_REF,
+      repo: field?.repo ?? DEFAULT_ENGINE_REPO,
+    };
+  }
+  return { source: "local" };
+}
+
+/**
+ * Whether an arbitrary value is a well-formed `engine` field. The bootstrapper
+ * is the only reader of `engine` — the engine drops it (`resolveConfig`) rather
+ * than validating it — so a malformed value has to be rejected here or it never
+ * is. An unknown `source` must not fall through to `local`, and a non-string
+ * `ref`/`repo` must not escape into a `ResolvedEngineSource` typed as strings.
+ */
+function isEngineSourceField(value: unknown): value is EngineSourceField {
+  if (value === null || typeof value !== "object") return false;
+  const field = value as Record<string, unknown>;
+  if (field["source"] === "local") return true;
+  return (
+    field["source"] === "github" &&
+    (field["ref"] === undefined || typeof field["ref"] === "string") &&
+    (field["repo"] === undefined || typeof field["repo"] === "string")
+  );
+}
+
+/**
+ * Extract the resolved engine source from a loaded consumer config, ignoring
+ * every other field. This is the whole of the bootstrapper's interest in the
+ * config; the engine reads (and validates) the rest once it is materialized and
+ * run. The config arrives as a dynamically-imported module value, so the param
+ * is an arbitrary record — `engine`, if present, is validated before use so a
+ * typo (bad `source`, numeric `ref`/`repo`) fails loudly instead of silently
+ * resolving to the wrong source.
+ */
+export function readEngineSource(config: Record<string, unknown>): ResolvedEngineSource {
+  const field = config["engine"];
+  if (field !== undefined && !isEngineSourceField(field)) {
+    throw new Error(
+      `phoebe.config.ts \`engine\` must be { source: "github", ref?, repo? } or ` +
+        `{ source: "local" } with string ref/repo (got ${JSON.stringify(field)}).`,
+    );
+  }
+  return resolveEngineSource(field);
+}

package/bootstrap/github-engine.ts

@@ -0,0 +1,169 @@
+// Materialize the Phoebe engine from GitHub for `engine: { source: "github" }`.
+//
+// The published package is a thin bootstrapper; when the config selects a github
+// source, `phoebe boot` runs the engine straight from a git checkout rather than
+// the bundled copy. First boot clones the engine repo into a persistent dir;
+// every boot fetches the configured ref and checks it out, so a branch tracks
+// its tip while a pinned SHA or tag pins an exact commit. The clone lives on a
+// persistent volume (keep `PHOEBE_ENGINE_DIR` on one) so later boots fetch into
+// the existing clone instead of re-cloning.
+//
+// Auth reuses `GH_TOKEN` (the same token every `gh` call uses), embedded as an
+// `x-access-token` HTTPS credential. The token is passed explicitly to each
+// clone/fetch and scrubbed from the persisted `origin` URL, so a rotated token
+// keeps working and no secret is written to the volume's git config.
+//
+// The git mechanics reuse the engine's injectable GitRunner seam
+// (src/git-model.ts) so this materializer is unit-tested without running git.
+//
+// This module also owns the *ref watch* the reconcile loop polls (#42): asking
+// the remote where the configured ref points now, so `phoebe boot` can notice
+// the tracked branch has advanced past the commit the running engine was checked
+// out at. Only a moving branch is watched — a pinned SHA or tag is inert.
+
+import { existsSync, mkdirSync } from "node:fs";
+import { join } from "node:path";
+import { defaultGit, type GitRunner } from "../src/git-model.ts";
+
+/** The resolved github source the materializer acts on (defaults already applied). */
+export type GithubSource = { source: "github"; ref: string; repo: string };
+
+/** The persistent clone directory for a given engine repo, keyed by repo slug. */
+export function githubEngineDir(baseDir: string, repo: string): string {
+  return join(baseDir, "github", repo.replace(/[^a-zA-Z0-9]+/g, "-"));
+}
+
+/**
+ * The HTTPS URL git clones/fetches from. With a token it carries an
+ * `x-access-token` credential (GitHub's convention for PAT/app tokens); without
+ * one it is the plain URL, so public repos and anonymous local dev still work.
+ */
+export function buildAuthenticatedRepoUrl(repo: string, token: string | undefined): string {
+  if (!token) return `https://github.com/${repo}.git`;
+  return `https://x-access-token:${token}@github.com/${repo}.git`;
+}
+
+/**
+ * Ensure a git checkout of `repo` at `ref` exists under `baseDir` and return the
+ * engine-CLI path (`<clone>/src/cli.ts`) `phoebe boot` execs, plus the commit it
+ * landed on. Clones only when the clone is absent; otherwise fetches into the
+ * existing one. `ref` (branch, 40-char SHA, or tag) is fetched by name and
+ * checked out via `FETCH_HEAD`, so a branch lands on its current tip and a
+ * SHA/tag on that exact commit.
+ *
+ * The returned `sha` is what the reconcile ref-watch compares the remote tip
+ * against — it is the commit actually running, so a relaunch is triggered by the
+ * branch advancing rather than by anything the poll itself remembers.
+ *
+ * `git`/`exists` are injectable so the command sequence is unit-tested without a
+ * real repository or network.
+ */
+export function materializeGithubEngine(
+  source: GithubSource,
+  deps: {
+    baseDir: string;
+    token?: string | undefined;
+    git?: GitRunner;
+    exists?: (path: string) => boolean;
+  },
+): { entry: string; sha: string | null } {
+  const git = deps.git ?? defaultGit;
+  const exists = deps.exists ?? existsSync;
+  const dir = githubEngineDir(deps.baseDir, source.repo);
+  const authUrl = buildAuthenticatedRepoUrl(source.repo, deps.token);
+  const cleanUrl = buildAuthenticatedRepoUrl(source.repo, undefined);
+
+  let sha: string | null;
+  try {
+    if (!exists(join(dir, ".git"))) {
+      // Clone into a pre-created empty dir (git needs the leading dirs to exist).
+      // Inlined rather than reusing git-model's `ensureClone`: this has to inject
+      // `exists` (for testing) and interleave the token-scrub below, neither of
+      // which that helper exposes.
+      mkdirSync(dir, { recursive: true });
+      git(["clone", authUrl, dir], { stdio: "inherit" });
+      // Drop the token from the persisted `origin` URL — the fetch below
+      // re-supplies it explicitly, so a rotated token keeps working and the
+      // volume's git config never holds the secret. The token is still passed on
+      // git's argv (clone/fetch); that is acceptable here because `GH_TOKEN`
+      // already lives in the container env and boot runs before any agent child.
+      git(["-C", dir, "remote", "set-url", "origin", cleanUrl]);
+    }
+    // Fetch the ref by name (works for branch/tag/reachable-SHA on GitHub) and
+    // detach onto exactly what was fetched — latest for a branch, pinned for a
+    // SHA/tag. `--force` lets a moved tag/branch and a dirty tree update cleanly.
+    git(["-C", dir, "fetch", "--force", "--tags", authUrl, source.ref], { stdio: "inherit" });
+    git(["-C", dir, "checkout", "--force", "--detach", "FETCH_HEAD"], { stdio: "inherit" });
+    sha = git(["-C", dir, "rev-parse", "HEAD"]).trim() || null;
+  } catch (error) {
+    throw new Error(
+      `Failed to materialize the engine from ${source.repo}@${source.ref}: ` +
+        `${error instanceof Error ? error.message : String(error)}`,
+    );
+  }
+
+  return { entry: join(dir, "src", "cli.ts"), sha };
+}
+
+/** A 40-char hex ref is an exact commit — nothing for the ref-watch to track. */
+export function isPinnedSha(ref: string): boolean {
+  return /^[0-9a-f]{40}$/i.test(ref);
+}
+
+/**
+ * Bound on the reconcile poll's `ls-remote`. It runs on the bootstrapper's event
+ * loop (`execFileSync`), so a hung remote must not stall boot's SIGTERM latch or
+ * its child-exit handling — a timed-out poll surfaces as a sample error and is
+ * treated as "no change", not a crash.
+ */
+export const LS_REMOTE_TIMEOUT_MS = 30_000;
+
+/**
+ * Parse `git ls-remote` output into `{ sha, refName }` rows. Each line is
+ * `<sha>\t<refname>`; peeled tag lines (`refs/tags/v1^{}`) come through as
+ * ordinary rows and are simply never matched as branches.
+ */
+export function parseLsRemote(output: string): Array<{ sha: string; refName: string }> {
+  const rows: Array<{ sha: string; refName: string }> = [];
+  for (const line of output.split("\n")) {
+    const [sha, refName] = line.trim().split(/\s+/);
+    if (sha && refName) rows.push({ sha, refName });
+  }
+  return rows;
+}
+
+/**
+ * The tip `ref` currently points at, but only when `ref` names a **branch** —
+ * the one kind of ref that moves under a running engine. A tag (or a ref the
+ * remote does not have) yields null, which the reconcile loop reads as "nothing
+ * to watch": per #42 a pinned SHA/tag never triggers a relaunch, even if someone
+ * force-moves the tag.
+ *
+ * Accepts both the short form (`main`) and a fully-qualified `refs/heads/main`.
+ */
+export function branchShaFromLsRemote(output: string, ref: string): string | null {
+  const wanted = ref.startsWith("refs/") ? ref : `refs/heads/${ref}`;
+  if (!wanted.startsWith("refs/heads/")) return null;
+  return parseLsRemote(output).find((row) => row.refName === wanted)?.sha ?? null;
+}
+
+/**
+ * Ask the remote where the configured branch points now — the ref half of a
+ * reconcile poll, and deliberately the *only* network call it makes: `ls-remote`
+ * reads a ref without fetching any objects. Returns null when there is nothing
+ * to watch (pinned SHA — not even asked; tag; unknown ref).
+ *
+ * The token rides on the URL exactly as it does for clone/fetch (same tradeoff,
+ * documented above): `GH_TOKEN` is already in the container env and boot runs
+ * before any agent child.
+ */
+export function lsRemoteBranchSha(
+  source: GithubSource,
+  deps: { token?: string | undefined; git?: GitRunner } = {},
+): string | null {
+  if (isPinnedSha(source.ref)) return null;
+  const git = deps.git ?? defaultGit;
+  const url = buildAuthenticatedRepoUrl(source.repo, deps.token);
+  const output = git(["ls-remote", url, source.ref], { timeout: LS_REMOTE_TIMEOUT_MS });
+  return branchShaFromLsRemote(output, source.ref);
+}

package/bootstrap/index.mjs

@@ -0,0 +1,9 @@
+// Runtime surface of the `phoebe-agent` package — what a consumer's
+// `phoebe.config.ts` resolves when it imports from `phoebe-agent`. Node resolves
+// a bare package specifier into `node_modules`, where Node 24 won't type-strip
+// `.ts`, so this entry must be JS. It is trivially small: the only thing a
+// consumer calls at runtime is `defineConfig`, an identity helper (its typed
+// signature lives in index.ts, the package `types` entry). Everything else the
+// package exposes is types-only. The bootstrapper's real logic is TypeScript —
+// see bootstrap/cli.ts, reached via the bin launcher (bootstrap/bin.mjs).
+export const defineConfig = (config) => config;