forgedock 1.3.0 → 1.4.0

package/bin/engine.mjs

@@ -5,7 +5,7 @@
  */
 import { fileURLToPath } from "node:url";
 import { appendEvent, readLog, deriveState, rewriteLog } from "./engine/runlog.mjs";
-import { pickPhase, TERMINAL_REASONS } from "./engine/phases.mjs";
+import { pickPhase, TERMINAL_REASONS, issueSnapshot } from "./engine/phases.mjs";
 import { reconcileState } from "./engine/reconcile.mjs";
 import { makeProjector } from "./engine/projector.mjs";
 import { VALID_BACKENDS } from "./runner.mjs";
@@ -14,6 +14,12 @@ import { VALID_BACKENDS } from "./runner.mjs";
 // diagnostics without duplicating the retry budget constant.
 export const DEFAULT_MAX_ATTEMPTS = 3;
 
+// forge#2239: lease TTL and renewal-interval defaults. Exported/overridable
+// (mirrors DEFAULT_MAX_ATTEMPTS) so tests can exercise "phase outlives the
+// TTL, lease gets renewed" without waiting on real 10-minute wall-clock time.
+export const DEFAULT_LEASE_TTL_MS = 600000;
+export const DEFAULT_LEASE_RENEW_INTERVAL_MS = 240000;
+
 /**
  * @param {object} opts
  * @param {string} [opts.backend] - "cli" | "api" | "auto" (forge#2028). Forwarded
@@ -23,6 +29,22 @@ export const DEFAULT_MAX_ATTEMPTS = 3;
  * @param {string} [opts.model] - Model id (forge#2028). Forwarded to every phase's
  *   `runner()` call when supplied; only applies on the "api" backend. Omit to keep
  *   runner.mjs's default (`FORGEDOCK_MODEL` env or its built-in default).
+ * @param {number} [opts.leaseTtlMs] - forge#2239: how long a claimed lease is
+ *   valid for. Defaults to DEFAULT_LEASE_TTL_MS. Overridable for tests.
+ * @param {number} [opts.leaseRenewIntervalMs] - forge#2239: how often the lease
+ *   is re-written while an unsatisfied phase's runner is executing. Defaults to
+ *   DEFAULT_LEASE_RENEW_INTERVAL_MS. Overridable for tests.
+ * @param {(event: {event: string, phase: string, status?: string, detail?: string}) => void} [opts.onProgress] -
+ *   forge#2240: optional phase-boundary observer. Called with
+ *   `{event: "phase_enter", phase}` right before a phase's runner is about to
+ *   execute (i.e. the phase was NOT already satisfied on reconcile), and with
+ *   `{event: "phase_exit", phase, status: "committed"|"blocked", detail?}`
+ *   once that phase's outcome is known. Defaults to a no-op so every existing
+ *   caller/test is unaffected. Deliberately engine.mjs's ONLY new surface for
+ *   this issue — no `console.log`/stdout write is added here; printing is the
+ *   CLI layer's job (bin/engine-cli.mjs), preserving the io-injection/testability
+ *   convention this module already follows. Invocations are wrapped so a
+ *   throwing callback can never crash a run.
  */
 export async function runIssue(opts) {
   const { issue, dir, agentId, lane = "staging", io, runner,
@@ -32,7 +54,30 @@ export async function runIssue(opts) {
           // call (forge#2028 / MAT-3). Left undefined by default so existing
           // callers keep runner.mjs's own "auto" ladder default unchanged —
           // this is purely additive pass-through, not a new default.
-          backend, model } = opts;
+          backend, model,
+          leaseTtlMs = DEFAULT_LEASE_TTL_MS,
+          leaseRenewIntervalMs = DEFAULT_LEASE_RENEW_INTERVAL_MS,
+          onProgress = () => {} } = opts;
+
+  // forge#2240: defensive wrapper — a caller's onProgress must never be able
+  // to crash an otherwise-healthy run. A plain try/catch only guards a
+  // *synchronous* throw — if onProgress is (or becomes) async and its
+  // returned promise rejects, that rejection is never awaited/caught here,
+  // producing an unhandled promise rejection that terminates the whole
+  // Node process (Node >=15 default behavior) well after runIssue() itself
+  // has already resolved. That is strictly worse than the silent-hang bug
+  // this issue fixes. Detect a thenable return value and attach a no-op
+  // .catch() to it (fire-and-forget — onProgress is not awaited either way,
+  // consistent with its documented synchronous-observer contract) so a
+  // rejection can never escape as unhandled. (review finding, #2240)
+  const emitProgress = (event) => {
+    try {
+      const result = onProgress(event);
+      if (result && typeof result.then === "function") {
+        result.catch(() => { /* best-effort observer, never fatal */ });
+      }
+    } catch { /* best-effort observer, never fatal */ }
+  };
 
   // forge#2054: validate `backend` before anything else — before state is
   // read/written and before the phase/retry loop begins below. An invalid
@@ -53,6 +98,55 @@ export async function runIssue(opts) {
     );
   }
 
+  // forge#2313: validate the lease-timing relationship before anything else —
+  // same placement/rationale as the INVALID_BACKEND check above (before state
+  // is read/written, before the phase loop begins). `leaseTtlMs` controls how
+  // long a claimed lease is valid for; `leaseRenewIntervalMs` controls how
+  // often the in-flight heartbeat (forge#2239) re-writes it. If the renew
+  // interval is >= the TTL, the first renewal tick fires at or after the
+  // moment the previously-written lease already expired — reopening the exact
+  // "phase outlives its lease" gap #2239 closed, because a genuinely-alive
+  // run would publish (or be caught holding) an expired `lease.until` in that
+  // window, which both the I3 concurrency guard above and the stall-recovery
+  // scanner (commands/orchestrate/phase-3-dependency.md) read directly. Both
+  // parameters are overridable-for-tests options with no current production
+  // caller forwarding them (bin/engine-cli.mjs's runFromCli() only forwards
+  // `backend`/`model`), so this only guards a future caller/config typo —
+  // reject rather than silently clamp, matching INVALID_BACKEND's precedent.
+  //
+  // forge#2329: guard for finite numbers BEFORE the relational check below.
+  // Every relational comparison against NaN evaluates to false in JS, so a
+  // NaN in either parameter silently passes `leaseRenewIntervalMs >=
+  // leaseTtlMs` (both `NaN >= x` and `x >= NaN` are false) — reopening the
+  // exact gap this whole guard exists to close, just via a different bad
+  // input. Infinity/-Infinity are "numbers" per `typeof` but are equally
+  // nonsensical as a lease duration/interval and are not reliably caught by
+  // the relational check either. Non-number types (strings, null, objects)
+  // are coerced by `>=` instead of rejected, which can silently poison
+  // downstream arithmetic (`now() + leaseTtlMs`) with string concatenation
+  // instead of numeric addition. Same placement precedent as the relational
+  // check itself: synchronous, before any state I/O or timer creation, so it
+  // cannot interact with the async lease-renewal heartbeat or the
+  // crash-and-relaunch write-then-throw architecture (bin/tests/engine-crash.test.mjs).
+  for (const [name, value] of [["leaseTtlMs", leaseTtlMs], ["leaseRenewIntervalMs", leaseRenewIntervalMs]]) {
+    if (typeof value !== "number" || !Number.isFinite(value)) {
+      throw Object.assign(
+        new Error(
+          `Invalid lease config: ${name} must be a finite number, got ${typeof value === "number" ? value : `${typeof value} (${JSON.stringify(value)})`}.`,
+        ),
+        { code: "INVALID_LEASE_CONFIG" },
+      );
+    }
+  }
+  if (leaseRenewIntervalMs >= leaseTtlMs) {
+    throw Object.assign(
+      new Error(
+        `Invalid lease config: leaseRenewIntervalMs (${leaseRenewIntervalMs}) must be less than leaseTtlMs (${leaseTtlMs}), or a live run's lease can expire before the next renewal.`,
+      ),
+      { code: "INVALID_LEASE_CONFIG" },
+    );
+  }
+
   const projector = makeProjector(io);
 
   // 1. Load + reconcile (GitHub wins).
@@ -72,9 +166,6 @@ export async function runIssue(opts) {
   if (!state) {
     state = freshState(issue, lane);
     appendEvent(dir, issue, { event: "RUN_START", issue, run: state.run, lane });
-    await projector.writeState(issue, state);
-  } else if (action === "remirror") {
-    await projector.writeState(issue, state);
   } else if (action === "hydrate") {
     // GitHub is ahead of (or the only source of) local state. Rebuild the
     // local run-log to match the remote compact index so downstream
@@ -83,8 +174,20 @@ export async function runIssue(opts) {
     // and regress committed/issue/v (C2).
     rewriteLog(dir, issue, eventsFromIndex(state));
     state = deriveState(readLog(dir, issue));
-    await projector.writeState(issue, state);
   }
+  // forge#2239: claim the lease unconditionally here — before the phase loop
+  // starts, for EVERY reconcile action ("fresh", "remirror", "hydrate", and
+  // "local"). Previously only "fresh"/"remirror"/"hydrate" wrote any state at
+  // all at this point, and none of them claimed a real (non-null) lease — the
+  // lease was only ever written retroactively after a phase's PHASE_COMMIT
+  // (see the write inside the loop below), leaving the entire first phase
+  // (and, for the "local" action, every phase before the first commit)
+  // publishing lease:null. That is indistinguishable from a dead run to both
+  // the I3 concurrency guard above and the stall-recovery scan in
+  // commands/orchestrate/phase-3-dependency.md. This write MUST stay after the I3 guard
+  // check above (commit 541a3e5 fixed the reverse ordering as a real bug —
+  // do not reintroduce it).
+  await projector.writeState(issue, { ...state, lease: { by: agentId, until: now() + leaseTtlMs } });
 
   // 2. Drive phases until terminal.
   //
@@ -98,6 +201,66 @@ export async function runIssue(opts) {
   // the real branch from the `FORGE:BUILDER` comment instead.
   let phase;
   while ((phase = pickPhase(state))) {
+    // forge#2352: state-vs-GitHub divergence guard. Every phase's own
+    // `entryCondition` only ever checked `state.committed` (local run-log
+    // progress) — never the issue's live GitHub state/labels — so a phase
+    // could run to completion against an issue that was independently closed
+    // (e.g. by a human, or by the investigate phase's own marker path) or
+    // labeled `workflow:invalid`/`needs-human` after this run's local state
+    // was last derived. `close` is deliberately exempt: it is the one phase
+    // whose entire job IS to read and act on this exact state (see its
+    // reconcile/detectOutcome in bin/engine/phases.mjs), so guarding it here
+    // would be redundant and could race with its own read of the same data.
+    //
+    // Cost: one extra `gh issue view` per loop iteration (skipped only for
+    // `close`, which already pays this cost itself) — bounded by the phase
+    // count (currently 6), not the retry budget inside a single phase.
+    //
+    // Fail-open on a snapshot error (`!snap.ok`): a transient `gh`/network
+    // failure here must not block a healthy run any more than the identical
+    // fail-open behavior in `close`'s own `reconcile` (bin/engine/phases.mjs)
+    // or the `reconcile` try/catch just below in this same loop.
+    if (phase.id !== "close") {
+      let snap;
+      try {
+        snap = await issueSnapshot(issue, io);
+      } catch {
+        snap = { ok: false, state: null, labels: [] };
+      }
+      if (snap.ok) {
+        // `workflow:invalid`, or CLOSED without `workflow:merged`, means the
+        // issue is dead — nothing a further phase does can be consequential.
+        // Reuses the existing "invalid" terminal reason (TERMINAL_REASONS)
+        // rather than inventing a new one; #2353 makes the same choice for
+        // `close`'s own CLOSED-not-merged case.
+        const isDead = snap.labels.includes("workflow:invalid") ||
+          (snap.state === "CLOSED" && !snap.labels.includes("workflow:merged"));
+        if (isDead) {
+          const detail = `issue ${issue} is ${snap.state === "CLOSED" ? "closed" : "open"} ` +
+            `with divergent state before phase ${phase.id} (labels: ${snap.labels.join(", ") || "none"})`;
+          return await terminate(state, "invalid", detail);
+        }
+        // `needs-human` is a PAUSE, not a death sentence — a human may still
+        // resolve it, and /orchestrate's classify_predecessor_state() already
+        // treats this label as GATED (not FAILED), so terminating here with
+        // the same reason composes with that existing contract instead of
+        // conflating "paused pending a human" with "dead". Deliberately does
+        // NOT check `workflow:invalid`/CLOSED here — those are handled above.
+        if (snap.labels.includes("needs-human")) {
+          const detail = `issue ${issue} carries needs-human — pausing before phase ${phase.id}`;
+          return await terminate(state, "needs-human", detail);
+        }
+      }
+    }
+    // forge#2321: tracks whether `phase_enter` was actually emitted for this
+    // loop iteration. Re-declared `false` on every iteration (never hoisted
+    // above the loop) so a reconcile-satisfied phase can never inherit a
+    // stale `true` from a previous iteration's actually-run phase. Only the
+    // committed-exit emission below needs this guard — the blocked-exit
+    // emission is unreachable from the `reconciled.satisfied` branch, since
+    // that branch always yields `status: "committed"` (see below), never
+    // "blocked".
+    let phaseEntered = false;
     let reconciled;
     try {
       reconciled = phase.reconcile ? await phase.reconcile(state, io) : { satisfied: false };
@@ -111,18 +274,164 @@ export async function runIssue(opts) {
       outcome = { status: "committed", outputs: reconciled.outputs || {} };
     } else {
       if (reconciled.outputs?.pr) state.pr = reconciled.outputs.pr;
-      outcome = await runPhaseWithRetry(phase, state, { io, runner, dir, issue, commandsDir, maxAttempts, backend, model });
+      // forge#2240: the phase is actually about to run (not a resume no-op)
+      // — this is the one point in the loop where "entering phase X" is true.
+      phaseEntered = true;
+      emitProgress({ event: "phase_enter", phase: phase.id });
+      // forge#2239: renew the lease immediately before running this phase's
+      // runner, then keep renewing it on a heartbeat for as long as the
+      // runner is in flight. A single phase can legitimately run longer than
+      // leaseTtlMs — without renewal, a healthy, still-running phase would
+      // eventually publish an *expired* (not just null) lease and get reaped
+      // by stall recovery even though the run is fine. The interval is
+      // `.unref()`'d so it can never keep the process alive on its own, and
+      // is always cleared in `finally` so it cannot outlive this phase's
+      // execution — covers the committed path, the blocked path, and the
+      // engine-error fail-fast throw caught just below.
+      await projector.writeState(issue, { ...state, lease: { by: agentId, until: now() + leaseTtlMs } });
+      // forge#2239 (review finding): `projector.writeState` is a plain
+      // read-body/edit-body round trip with no CAS — whichever `gh issue
+      // edit` call actually lands last on GitHub wins, regardless of dispatch
+      // order. A fire-and-forget renewal write left in flight when this
+      // phase finishes could land AFTER the post-commit write or even after
+      // terminate()'s `lease: null` write, resurrecting a phantom lease on an
+      // already-terminated run. Track the most recently dispatched renewal's
+      // promise and await it in `finally` (below) before letting the loop
+      // proceed — this guarantees no renewal write is still in flight when
+      // control moves on to the next write (commit or terminate).
+      // forge#2348 (review finding): the above join-tracking scheme only
+      // ever holds ONE promise in `pendingRenewal` — it protects against a
+      // single in-flight write landing late, but has no backpressure against
+      // a SECOND renewal tick firing while the first write is still in
+      // flight (write round trip > leaseRenewIntervalMs). Without a guard,
+      // `pendingRenewal` would simply be overwritten with the newer write's
+      // promise, silently orphaning the earlier one from every join point
+      // (both this closure's own `finally` below and the #2338 catch-path
+      // join) — the earlier write could then land on GitHub after a later
+      // write, or after the terminate()/commit write joins only the latest
+      // promise and proceeds. Guarding here so at most one renewal write is
+      // ever in flight keeps the existing single-variable join points
+      // correct and sufficient: skip this tick entirely if a previous
+      // renewal write hasn't settled yet, and clear `pendingRenewal` back to
+      // null once it does (success or failure) so the guard never
+      // permanently wedges future renewals.
+      let pendingRenewal = null;
+      const renewLease = () => {
+        if (pendingRenewal) return;
+        pendingRenewal = projector.writeState(issue, { ...state, lease: { by: agentId, until: now() + leaseTtlMs } }).catch(() => {
+          // Best-effort: a transient renewal failure must not crash the run.
+          // The next scheduled renewal (or the post-commit write) will retry.
+        }).finally(() => {
+          pendingRenewal = null;
+        });
+      };
+      const renewTimer = setInterval(renewLease, leaseRenewIntervalMs);
+      if (typeof renewTimer.unref === "function") renewTimer.unref();
+      try {
+        outcome = await runPhaseWithRetry(phase, state, { io, runner, dir, issue, commandsDir, maxAttempts, backend, model });
+      } catch (e) {
+        // forge#2261: NO_API_KEY/NO_SDK/CLI_BACKEND_FAILED are fail-fast
+        // rethrown by runPhaseWithRetry() (see its own catch below) instead
+        // of being retried — but until now that throw was never caught here,
+        // so it propagated all the way out of runIssue() uncaught (through
+        // bin/engine-cli.mjs's runFromCli(), which also has no try/catch)
+        // and only landed in bin/forgedock.mjs's outermost `run-issue` case,
+        // which just prints the message to stderr and exits 1 — no terminal
+        // state or label was ever written, leaving the issue stuck on
+        // whatever workflow label it already had. Catch it here instead and
+        // reach a clean terminal state: "engine-error" is deliberately NOT
+        // "needs-human" — this is the engine/tool breaking, not a genuine
+        // human-judgment block (see #2244/#2261). Any other thrown error is
+        // a true unexpected crash and keeps propagating unchanged.
+        if (e.code === "NO_API_KEY" || e.code === "NO_SDK" || e.code === "CLI_BACKEND_FAILED") {
+          // forge#2241: when the runner attached a session-limit reset time
+          // (bin/runner.mjs's extractSessionLimitResetTime(), only ever set
+          // for a genuine session-limit CLI_BACKEND_FAILED — never
+          // fabricated), append it so the terminal state is legible without
+          // reading raw logs. Purely additive: the base detail string is
+          // unchanged, and this appends nothing when e.resetAt is absent.
+          const resetSuffix = e.resetAt ? ` (resets: ${e.resetAt})` : "";
+          const detail = `phase ${phase.id}: ${e.code} - ${e.message}${resetSuffix}`;
+          // forge#2240 (review finding): this fail-fast path previously left
+          // phase_exit unreported — a caller tailing progress output would see
+          // "→ phase X started" and then nothing, dangling exactly on the
+          // phase that actually failed. Report it as blocked before
+          // terminating so the progress trail stays complete on this path too.
+          emitProgress({ event: "phase_exit", phase: phase.id, status: "blocked", detail });
+          // forge#2338 (review finding): `return await terminate(...)` here
+          // sits INSIDE this catch block, so per try/catch/finally semantics
+          // its expression is fully evaluated — including terminate()'s own
+          // `lease: null` write completing — BEFORE the `finally` block below
+          // ever runs `clearInterval`/`await pendingRenewal`. A renewal write
+          // already dispatched before this throw can therefore land on
+          // GitHub AFTER terminate()'s `lease: null` write, resurrecting a
+          // phantom lease on an already-terminated run — reopening the exact
+          // race #2239 was written to close, just on this one call site the
+          // #2239 fix (07b3b8a) didn't cover. Explicitly join here, before
+          // calling terminate(), rather than relying on `finally` (too
+          // late). The `finally` block's identical calls remain below as a
+          // harmless idempotent safety net for the `throw e;` path.
+          clearInterval(renewTimer);
+          if (pendingRenewal) await pendingRenewal;
+          return await terminate(state, "engine-error", detail);
+        }
+        throw e;
+      } finally {
+        clearInterval(renewTimer);
+        // Join any renewal write already dispatched before this phase's
+        // outcome is allowed to reach a commit/terminate write — closes the
+        // TOCTOU window above. `pendingRenewal` already swallows its own
+        // rejection (see renewLease()), so this await never throws.
+        if (pendingRenewal) await pendingRenewal;
+      }
     }
 
-    if (outcome.status === "blocked") return await terminate(state, "needs-human", outcome.detail);
+    if (outcome.status === "blocked") {
+      emitProgress({ event: "phase_exit", phase: phase.id, status: "blocked", detail: outcome.detail });
+      return await terminate(state, outcome.reason || "needs-human", outcome.detail);
+    }
 
     // committed
-    appendEvent(dir, issue, { event: "PHASE_COMMIT", phase: phase.id, outputs: outcome.outputs || {} });
+    // forge#2321: only report a phase_exit if a matching phase_enter was
+    // actually emitted for this phase this iteration. A reconcile-satisfied
+    // phase (short-circuited above without ever entering the `else` branch)
+    // never "started" this run — emitting phase_exit for it produced a
+    // dangling exit with no preceding enter (bin/engine-cli.mjs prints
+    // "✓ phase X committed" with no prior "→ phase X started" line).
+    // Suppressing the unmatched exit is correct here rather than fabricating
+    // a synthetic phase_enter, since the phase's runner genuinely never ran.
+    if (phaseEntered) emitProgress({ event: "phase_exit", phase: phase.id, status: "committed" });
+    // forge#2377: `outcome.usage` is populated by runPhaseWithRetry() below
+    // from the injected runner()'s (== bin/runner.mjs's runCommand()) return
+    // value — null when the backend doesn't report usage (CLI backend today)
+    // or when detectOutcome() never ran (this call site is only reached on
+    // the "committed" path, so that case doesn't apply here). Kept as a
+    // sibling of `outputs` rather than nested inside it, since `outputs` is
+    // owned by phase.detectOutcome() (bin/engine/phases.mjs).
+    appendEvent(dir, issue, { event: "PHASE_COMMIT", phase: phase.id, outputs: outcome.outputs || {}, usage: outcome.usage ?? null });
     state = deriveState(readLog(dir, issue));
     if (outcome.terminalReason) state.terminalReason = outcome.terminalReason;
-    await projector.writeState(issue, { ...state, lease: { by: agentId, until: now() + 600000 } });
+    await projector.writeState(issue, { ...state, lease: { by: agentId, until: now() + leaseTtlMs } });
 
-    if (outcome.terminalReason && TERMINAL_REASONS.includes(outcome.terminalReason))
+    // forge#2379: the "investigate" phase reporting terminalReason "decomposed"
+    // is a HANDOFF to the "decompose" phase (bin/engine/phases.mjs), not a
+    // dead end — decompose is what actually dispatches work-on/decompose
+    // (sub-issue fan-out, FORGE:DECOMPOSED posting). Before this change,
+    // "decomposed" being a member of TERMINAL_REASONS meant the run
+    // terminated the instant investigate reported it, and decompose's own
+    // entryCondition (state.terminalReason === "decomposed") was never
+    // actually reachable through pickPhase. Narrowly exempting exactly this
+    // phase/reason combination — and no other phase, no other reason — lets
+    // the loop continue to pickPhase, which now correctly selects "decompose"
+    // next (investigate.isTerminalAfter was narrowed to only "invalid" in
+    // the same change, so it no longer independently forces termination
+    // here either). Once the "decompose" phase itself later reports
+    // terminalReason "decomposed" (re-affirming it after FORGE:DECOMPOSED:COMPLETE
+    // is seen), phase.id is "decompose" — not "investigate" — so this
+    // exemption does not apply and the normal terminate() path below fires,
+    // ending the run for real.
+    const isDecomposeHandoff = phase.id === "investigate" && outcome.terminalReason === "decomposed";
+    if (outcome.terminalReason && TERMINAL_REASONS.includes(outcome.terminalReason) && !isDecomposeHandoff)
       return await terminate(state, outcome.terminalReason);
     if (phase.isTerminalAfter && phase.isTerminalAfter(state))
       return await terminate(state, state.terminalReason || "merged");
@@ -134,16 +443,41 @@ export async function runIssue(opts) {
     const final = { ...deriveState(readLog(dir, issue)), terminal: true, terminalReason: reason, lease: null };
     await projector.writeState(issue, final);
     if (reason === "needs-human") await projector.setLabel(issue, "needs-human");
+    // forge#2261: a distinct label for engine/tool-level failures (broken CLI
+    // invocation, exhausted retries where the runner itself never once
+    // succeeded) — NOT needs-human. /orchestrate's classify_predecessor_state()
+    // (commands/orchestrate/phase-4-execution.md) must not treat this as GATED:
+    // there is no human decision pending, just a tool that needs to be re-run.
+    else if (reason === "engine-error") await projector.setLabel(issue, "workflow:engine-error");
     return { terminalReason: reason, detail };
   }
 }
 
 async function runPhaseWithRetry(phase, state, ctx) {
   const { io, runner, dir, issue, commandsDir, maxAttempts, backend, model } = ctx;
+  // forge#2261: true only if EVERY attempt failed by the runner itself
+  // throwing (never once reached phase.detectOutcome()). This is the signal
+  // that distinguishes an engine/tool crash (the tool never even produced a
+  // result to evaluate) from a genuine content-level block (the tool ran
+  // fine, but the phase's own completion criteria weren't met — e.g. an
+  // unmerged PR, an unresolved branch, or a fixed-point zero-commits case).
+  // Flips to false the instant any attempt's runner() call succeeds, even if
+  // that attempt's detectOutcome() itself reports failure.
+  let allAttemptsThrew = true;
+  // forge#2377: the last successful attempt's `usage` (from runner()'s ==
+  // runCommand()'s resolved value — {input_tokens, output_tokens,
+  // cache_creation_input_tokens, cache_read_input_tokens} on the API
+  // backend, or null on the CLI backend / when the field is absent). Reset
+  // is unnecessary since a thrown attempt never reaches the assignment
+  // below — `lastUsage` simply stays whatever the previous successful
+  // attempt (if any) set it to, which is correct: it always reflects the
+  // most recent attempt that actually produced a result.
+  let lastUsage = null;
   for (let attempt = 1; attempt <= maxAttempts; attempt++) {
     appendEvent(dir, issue, { event: "PHASE_START", phase: phase.id, attempt });
+    let result;
     try {
-      await runner({
+      result = await runner({
         commandsDir, commandName: phase.command, args: [String(issue)],
         // Only forwarded when explicitly provided — omitting them preserves
         // runner.mjs's existing default ("auto" backend / DEFAULT_MODEL).
@@ -154,13 +488,34 @@ async function runPhaseWithRetry(phase, state, ctx) {
       // A missing API key / SDK is a config error, not a transient phase
       // failure — surface it distinctly instead of burning retries and
       // escalating to needs-human as if the LLM run itself misbehaved.
-      if (e.code === "NO_API_KEY" || e.code === "NO_SDK") throw e;
-      appendEvent(dir, issue, { event: "PHASE_FAILED", phase: phase.id, attempt, reason: e.message });
+      //
+      // forge#2259: a non-zero exit from the nested `claude` CLI
+      // (bin/runner.mjs's runCliBackend(), which sets err.code =
+      // "CLI_BACKEND_FAILED" on any non-zero `result.status`) is likewise
+      // not a transient phase failure — same argv/cwd/session state
+      // reproduces the identical crash on every attempt. Retrying it
+      // `maxAttempts` times burns the entire retry budget (observed live in
+      // #2244: 3 nested-CLI invocations, zero PRs/commits) on a failure that
+      // was never going to succeed. Fail fast instead, mirroring the
+      // NO_API_KEY/NO_SDK precedent (commit 570cb10 / #2054 established this
+      // exact dispatch pattern for INVALID_BACKEND). This does not assume
+      // *why* the CLI exited 1 (the quota/session-limit theory in #2244 is
+      // unproven) — a deterministic tool crash should not be retried as if
+      // transient regardless of its root cause.
+      if (e.code === "NO_API_KEY" || e.code === "NO_SDK" || e.code === "CLI_BACKEND_FAILED") throw e;
+      // forge#2377: no `usage` field here — the runner threw, so no result
+      // (and therefore no usage data) was ever produced for this attempt.
+      // Do not fabricate a value; omitting the field (rather than a stale
+      // `lastUsage` from a prior attempt) keeps this event's usage
+      // trustworthy as "usage actually observed on this attempt".
+      appendEvent(dir, issue, { event: "PHASE_FAILED", phase: phase.id, attempt, reason: e.message, maxAttempts });
       continue;
     }
+    allAttemptsThrew = false;
+    lastUsage = result?.usage ?? null;
     const outcome = await phase.detectOutcome(state, io);
-    if (outcome.status === "committed" || outcome.status === "blocked") return outcome;
-    appendEvent(dir, issue, { event: "PHASE_FAILED", phase: phase.id, attempt, reason: outcome.detail });
+    if (outcome.status === "committed" || outcome.status === "blocked") return { ...outcome, usage: lastUsage };
+    appendEvent(dir, issue, { event: "PHASE_FAILED", phase: phase.id, attempt, reason: outcome.detail, maxAttempts, usage: lastUsage });
     // forge#2176: a phase's detectOutcome can mark a failure as a known,
     // state-derived fixed point — re-running the phase's runner is
     // guaranteed to reproduce the identical failure (e.g. the build phase's
@@ -172,11 +527,28 @@ async function runPhaseWithRetry(phase, state, ctx) {
     // for every phase that doesn't opt in — investigate/context/architect/
     // review/close — is unchanged, preserving transient-failure retries).
     if (outcome.retryable === false) {
-      return { status: "blocked", detail: outcome.detail };
+      return { status: "blocked", detail: outcome.detail, usage: lastUsage };
     }
   }
   // Exhausted transient retries → escalate (spec §7).
-  return { status: "blocked", detail: `phase ${phase.id} failed after ${maxAttempts} attempts` };
+  // forge#2261: if the runner itself threw on every single attempt (it never
+  // once produced a result for detectOutcome() to evaluate), this is an
+  // engine/tool failure, not a content-level judgment call — tag the outcome
+  // so runIssue()'s terminate() writes a distinct label instead of
+  // needs-human. If at least one attempt reached detectOutcome() (the tool
+  // ran, the phase just isn't done), this stays the existing untagged shape,
+  // which runIssue() defaults to "needs-human" — unchanged behavior for every
+  // genuine content-level block (unmerged PR, unresolved branch, a transient
+  // git error inside commitsAhead(), etc.).
+  if (allAttemptsThrew) {
+    return {
+      status: "blocked",
+      detail: `phase ${phase.id} failed after ${maxAttempts} attempts (runner threw every attempt)`,
+      reason: "engine-error",
+      usage: null,
+    };
+  }
+  return { status: "blocked", detail: `phase ${phase.id} failed after ${maxAttempts} attempts`, usage: lastUsage };
 }
 
 /**