Merge pull request #2075 from open-gsd/fix/2009-load-failed-capability-injects-blocking-

fix(#2009): load-failed capability gates fail open with a loud warning
2026-07-07 23:26:12 -04:00
parent 3f82929cc2 5aa7771dd3
commit a467dad065
10 changed files with 263 additions and 113 deletions
--- a/.changeset/2009-load-failed-capability-fail-open.md
+++ b/.changeset/2009-load-failed-capability-fail-open.md
@@ -0,0 +1,5 @@
+---
+type: Fixed
+pr: 2075
+---
+**Load-failed capability gates now fail open with a loud warning instead of blocking the whole project** — when an installed overlay (third-party) capability failed to load (e.g. an incompatible `engines.gsd` range) but had declared a `gate`-kind loop hook, the loop resolver injected a blocking synthetic gate (`blocking:true`, `onError:halt`) at every point where that capability declared a gate. A single incompatible capability therefore halted every `ship:pre` and `verify:post` in the project — unrelated to what the gate would have checked, and with no remediation surfaced. The resolver now injects no gate and instead emits a loud warning — to stderr and in the `loop render-hooks` envelope's `warnings` array — naming the load-failure reason and the exact `gsd capability remove <id>` remediation, and the loop proceeds (fail open). The capability id embedded in that remediation is validated against the canonical id shape first, so a malformed overlay directory name cannot inject shell metacharacters into the surfaced command. The loader still records `_overlay.blockedGates`; only the consequence changes from block to warn. `step`/`contribution` overlays were already skip-open. (#2009)
--- a/docs/ARCHITECTURE.md
+++ b/docs/ARCHITECTURE.md
@@ -381,7 +381,7 @@ Node.js CLI utility (`gsd-tools.cjs`) with domain modules split across `gsd-core
 | `profile-pipeline.cjs` | User behavioral profiling data pipeline, session file scanning                                      |
 | `profile-output.cjs`       | Profile rendering, USER-PROFILE.md and dev-preferences.md generation                                |
 | `loop-host-contract.cjs`   | Generated Loop Host Contract — 12 loop points, per-step agent roles, and core artifacts; emitted by `scripts/gen-loop-host-contract.cjs` from workflow markers (ADR-894 §3); consumed by `gen-capability-registry.cjs` |
-| `capability-loader.cjs`    | Runtime registry overlay loader (ADR-1244 D2) — `loadRegistry({ includeInstalled })` composes the frozen first-party registry with a validated installed overlay of third-party capability manifests read from global `$GSD_HOME/.gsd/capabilities/` and project `<projectRoot>/.gsd/capabilities/`; first-party always wins; load-time `engines.gsd` re-gate skips incompatible overlays with a warning; gate-kind hooks on skipped capabilities fail CLOSED |
+| `capability-loader.cjs`    | Runtime registry overlay loader (ADR-1244 D2) — `loadRegistry({ includeInstalled })` composes the frozen first-party registry with a validated installed overlay of third-party capability manifests read from global `$GSD_HOME/.gsd/capabilities/` and project `<projectRoot>/.gsd/capabilities/`; first-party always wins; load-time `engines.gsd` re-gate skips incompatible overlays with a warning; gate-kind hooks on skipped capabilities fail OPEN — no gate is injected; a loud warning (stderr + envelope `warnings`) names the load failure and the `gsd capability remove <id>` remediation (#2009) |
 | `capability-registry.cjs`  | Generated central Capability Registry — role-partitioned index of all co-located capability declarations; emitted by `scripts/gen-capability-registry.cjs` (ADR-894 §5) |
 | `loop-resolver.cjs`        | Loop Extension Point resolver — ADR-857 phase 3c registry-consuming query; consumes resolved Capability State, filters `byLoopPoint` by capability enablement plus config activation, renders active hooks as markdown, emits `{ point, activeHooks, rendered }` envelope; `gsd-tools loop render-hooks <point> [--config-dir <path>]` |
 | `capability-state.cjs`     | Unified capability-state resolver — ADR-857 phase 4b/6; composes install profile, runtime surface, and config activation into one per-capability view consumed by workflow hook rendering; pure `resolveCapabilityState`, reusable `resolveCapabilityRuntimeState`, I/O `cmdCapabilityState`, and convenience predicate `isCapabilityActive(capId, cwd)`; `gsd-tools capability state [--config-dir <path>]` emits `{ runtimeConfigDir, capabilities[] }` where each entry carries `enabled` (installed && surfaced) and `active` (enabled && configActivation via the capability's `activationKey`; absent key → active===enabled) |
--- a/docs/CONFIGURATION.md
+++ b/docs/CONFIGURATION.md
@@ -770,9 +770,9 @@ Installed overlay capabilities are merged via the same `buildRegistry` pipeline

 Each overlay manifest may declare an `engines.gsd` semver range. At load time GSD evaluates this range against the running GSD version. An overlay that does not satisfy the range is **skipped with a warning** — it is never loaded and never crashes the loop. Manifests without an `engines.gsd` field are accepted unconditionally.

-### Gate-kind fail-closed policy
+### Gate-kind fail-open policy (#2009)

-If a skipped overlay capability declared a `gate`-kind loop hook, the loop resolver **injects a blocking gate** at that hook point (fail CLOSED). Skipped capabilities whose hooks are `step` or `contribution` kind skip open — the loop proceeds without them.
+If a skipped or load-failed overlay capability (for example, one whose `engines.gsd` range is incompatible) declared a `gate`-kind loop hook, the loop resolver does **not** inject a gate at that hook point (fail OPEN): the loop proceeds. Instead it emits a loud warning — to stderr and in the `loop render-hooks` envelope's `warnings` array — naming the load-failure reason and the exact remediation, `gsd capability remove <id>`, so the operator is loudly told how to clear it. Skipped capabilities whose hooks are `step` or `contribution` kind skip open too, as before — the loop proceeds without them.

 ### Overlay config federation

--- a/docs/INVENTORY.md
+++ b/docs/INVENTORY.md
@@ -407,7 +407,7 @@ Full listing: `gsd-core/bin/lib/*.cjs`.
 | `capability-lock.cjs` | Shared cross-process lock primitive (#1459 finding 4) — the SINGLE hardened lockfile protocol used by BOTH capability-lifecycle (`.gsd/capabilities/.lock`) and capability-consent (`.consent.lock`); exports `acquireLock(lockPath, opts?)`/`releaseLock(handle)` with pid + process-start-time liveness identity, a hard deadman, and token+inode owner-safe release — NEVER stale-steals a verified-live same-host holder, reclaims only a provably-dead/unverifiable holder, never deadlocks; `opts.maxAttempts`/`opts.waitForFresh` let the consent store serialize genuinely-contended writers; `_setLockProbes`/`_resetLockProbes` are test seams |
 | `capability-ledger.cjs` | Per-runtime install ledger (ADR-1244 D4) — atomic read/write of `.gsd-capabilities.json` recording `{ id, version, source, integrity, files[], sharedEdits[] }` per installed capability; exports `readLedger`/`writeLedger`/`recordInstall`/`removeEntry`/`reconcile` (orphan detection)/`readSmallRegularFile` (utf8) + `readSmallRegularFileBuffer` (raw bytes, the byte-exact consent-hash reader, #1459 finding 1); atomic commit point and reconciliation basis for Phase-4 upgrade/remove |
 | `capability-lifecycle.cjs` | Capability lifecycle orchestration (ADR-1244 Phase 4, D5+D6) — composes the source resolver + ledger + trust gate into `installCapability`/`upgradeCapability`/`removeCapability`/`reconcileCapabilities`; ledger write is the commit point; upgrade is atomic stage-then-swap (old set aside, new swapped in, ledger committed, backup dropped) with deterministic crash recovery (`reconcileCapabilities` rolls forward/back to a fully-old-or-fully-new state); remove surgically strips only marker-stamped (`_gsdCapability`) shared-config entries, preserving user hand-edits; never executes capability code |
-| `capability-loader.cjs` | Runtime Capability Registry overlay (ADR-1244 D2) — `loadRegistry({ includeInstalled })` composes the frozen first-party registry with a validated installed overlay read from `$GSD_HOME/.gsd/capabilities/<id>/` (global) and `<projectRoot>/.gsd/capabilities/<id>/` (project); first-party-wins on id/skill/agent/config collisions, reserved-namespace rejection, load-time `engines.gsd` re-gate (skip-with-warning), and gate-kind fail-closed via `_overlay.blockedGates`; composes through the canonical `buildRegistry` so derived views never drift |
+| `capability-loader.cjs` | Runtime Capability Registry overlay (ADR-1244 D2) — `loadRegistry({ includeInstalled })` composes the frozen first-party registry with a validated installed overlay read from `$GSD_HOME/.gsd/capabilities/<id>/` (global) and `<projectRoot>/.gsd/capabilities/<id>/` (project); first-party-wins on id/skill/agent/config collisions, reserved-namespace rejection, load-time `engines.gsd` re-gate (skip-with-warning), and gate-kind fail-open via `_overlay.blockedGates` — a loud warning (stderr + envelope `warnings`) naming the load failure and `gsd capability remove <id>` remediation; no gate injected (#2009); composes through the canonical `buildRegistry` so derived views never drift |
 | `capability-registry.cjs` | Generated central Capability Registry — role-partitioned index of all co-located capability declarations (`capabilities/<id>/capability.json`); emitted by `scripts/gen-capability-registry.cjs --write` (ADR-894 §5) |
 | `capability-source.cjs` | Capability source resolver (ADR-1244 D3) — `resolveCapabilitySource(spec, opts)` fetches and stages a capability from local path, git (https/ssh/git transports only), npm pack (no lifecycle scripts), tarball (sha512 integrity verify before extraction), or registry (stub); tar-slip/symlink rejection; atomic staging to `$GSD_HOME/.gsd/capabilities/<id>/`; no capability code executes during install |
 | `capability-state.cjs` | Unified capability-state resolver (ADR-857 phase 4b/6) — composes install profile, runtime surface, and config activation into one per-capability view consumed by workflow hook rendering; exports pure `resolveCapabilityState`, reusable `resolveCapabilityRuntimeState`, and I/O handler `cmdCapabilityState`; command surface: `gsd-tools capability state [--config-dir <path>]` emitting `{ runtimeConfigDir, capabilities[] }` |
--- a/docs/README.md
+++ b/docs/README.md
@@ -72,7 +72,7 @@ Language versions: [English](README.md) · [Português (pt-BR)](pt-BR/README.md)
 - [Multi-agent orchestration](explanation/multi-agent-orchestration.md) — how subagents are spawned, scoped, and coordinated
 - [Security model](explanation/security-model.md) — trust boundaries, permissions, and safe automation
 - [The capability trust model](explanation/capability-trust-model.md) — why third-party capabilities are gated by consent + integrity + reversibility, not a sandbox
- [How overlay capabilities compose](explanation/capability-overlay-model.md) — why first-party always wins and how the loader resolves precedence, conflicts, and fail-closed gates
+- [How overlay capabilities compose](explanation/capability-overlay-model.md) — why first-party always wins and how the loader resolves precedence, conflicts, and fail-open load-failure warnings
 - [Architecture](ARCHITECTURE.md) — system architecture, agent model, and data flow
 - [Discuss modes](workflow-discuss-mode.md) — assumptions mode vs interview mode for `/gsd-discuss-phase`
 - [Context monitoring](context-monitor.md) — context window monitoring hook architecture
--- a/docs/explanation/capability-overlay-model.md
+++ b/docs/explanation/capability-overlay-model.md
@@ -181,9 +181,10 @@ candidate is processed. A single broken overlay cannot poison the rest of the se

 ---

-## The one place where skipping is dangerous: gates
+## The one place where a skip must be loud: gates

-Skipping a broken overlay is the safe default for most surfaces — but not for *gates*.
+Skipping a broken overlay is the safe default for every surface — including gates, though
+gates get special treatment.

 A capability's loop hooks come in three kinds:

@@ -195,21 +196,28 @@ For steps and contributions, skipping a capability means the loop simply proceed
 **without** that addition. That is *fail-open*, and it is correct: the loop is missing an
 optional step, not doing something unsafe.

-A gate is the opposite. The whole purpose of a gate is to *stop* the loop when a
-condition is not met — a deploy gate, a house-style verification gate, a safety check. If
-GSD skipped a broken gate-declaring capability and proceeded, it would behave exactly as
-if the gate had *passed* — silently waving through the very thing the gate existed to
-block. That is a fail-open on a security-relevant control, and it is unacceptable.
+A gate looks different at first glance. The whole purpose of a gate is to *stop* the loop
+when a condition is not met — a deploy gate, a house-style verification gate, a safety
+check. Silently skipping a broken gate-declaring capability and proceeding as if the gate
+had *passed* would wave through the very thing the gate existed to block, with no signal
+to the operator at all.

-So composition treats gates asymmetrically from steps and contributions. When a
-capability that declares a gate is skipped, GSD records its gate points in
-`_overlay.incompatibleGateCapIds` and `_overlay.blockedGates`, and the loop resolver
-**injects a synthetic blocking gate** at each of those extension points. The loop
-**fails closed**: rather than proceed as if the gate passed, it halts with a message
-naming the skipped capability and why its gate could not be evaluated.
+So, per the maintainer decision on [#2009](https://github.com/open-gsd/gsd-core/issues/2009),
+composition treats gates like steps and contributions for control flow — the loop always
+proceeds — but never silently. When a capability that declares a gate is skipped, GSD
+records its gate points in `_overlay.incompatibleGateCapIds` and `_overlay.blockedGates`,
+and the loop resolver **injects no gate** at each of those extension points. The loop
+**fails open**, but loudly: it emits a warning through two channels — stderr (the channel
+host workflows/agents see when they run `gsd_run loop render-hooks <point>`) and the
+`loop render-hooks` JSON envelope's top-level `warnings` array. The warning names the
+skipped capability, why it could not be loaded (for example, an incompatible
+`engines.gsd` range), and the exact remediation — `gsd capability remove <id>` — so the
+operator sees the missing control on every pass through the loop until they act on it,
+instead of the loop halting project-wide over a single incompatible overlay.

 The discriminator is therefore *not* "is this overlay broken?" but "what does failing
-to load it mean?" — and for a gate, failing to load it means you must not proceed.
+to load it mean?" — and for a gate, failing to load it means the operator must be told,
+unmistakably, until they resolve it.

 ---

@@ -230,10 +238,12 @@ details make this safe rather than merely convenient:
  behind a path that a runtime dispatcher might `require()` a command module from.
 - Every dropped overlay's **gates are recorded as blocked** — using the same extraction
  as the per-candidate path — so a gate-declaring overlay that vanishes in the fallback
-  still **fails closed**, never open.
+  still **surfaces a loud warning** (stderr + envelope `warnings`) at its gate points
+  rather than vanishing silently (#2009).

 The principle is the same at every layer: when GSD cannot compose an overlay, it removes
-the overlay's *additions* but never weakens a *control*.
+the overlay's *additions* but never silences a *control* — a missing gate always
+surfaces, even though, per #2009, it no longer blocks the loop.

 ---

@@ -262,16 +272,20 @@ The overlay model rests on a few rules applied consistently:
 - **First-party always wins** every collision — id, skill/agent stem, config key,
  command family, reserved prefix. An overlay can only add, never override.
 - A bad overlay is **skipped, not crashed** — the loop always gets a usable registry.
- Skipping **fails open** for steps and contributions (a missing optional addition) but
-  **fails closed** for gates (a missing control must block, not pass).
+- Skipping **fails open** for steps and contributions (a missing optional addition) and,
+  per [#2009](https://github.com/open-gsd/gsd-core/issues/2009), **fails open** for gates
+  too (a missing control, no gate injected) — but loudly, via a warning (stderr + the
+  envelope's `warnings` array) that names the load failure and its
+  `gsd capability remove <id>` remediation.
 - A whole-set compose failure **falls back to first-party**, clearing command roots and
-  still blocking dropped gates.
+  still surfacing dropped gates as loud warnings.
 - One canonical builder materialises both first-party and overlay views, so an accepted
  overlay has true parity with a shipped capability.

 Every one of these choices answers the same question — *what does it mean if this
-composition step fails?* — and resolves it in favour of first-party authority and a
-fail-closed security posture.
+composition step fails?* — and resolves it in favour of first-party authority and,
+per #2009, a loud fail-open posture: never silent, never a project-wide halt over a
+single incompatible overlay.

 ---

--- a/src/capability-loader.cts
+++ b/src/capability-loader.cts
@@ -17,10 +17,12 @@
 *     `gsd-core-` / `anthropic-` id prefix) is rejected.
 *   - Load-time re-gate (default-resilient): an overlay that fails validation or
 *     whose `engines.gsd` does not satisfy the running GSD version is SKIPPED
- *     with a warning — it never crashes the loop. EXCEPTION (per-hook-kind
- *     policy): a skipped capability that declares a `gate` is recorded in
- *     `_overlay.incompatibleGateCapIds` so the loop resolver can fail CLOSED for
- *     that gate rather than silently proceeding as if it had passed.
+ *     with a warning — it never crashes the loop. A skipped capability that
+ *     declares a `gate` is additionally recorded in
+ *     `_overlay.incompatibleGateCapIds` / `_overlay.blockedGates` so the loop
+ *     resolver can surface a loud fail-OPEN advisory for that gate (#2009): the
+ *     un-evaluable gate is skipped (not enforced) with a remediation message,
+ *     rather than silently vanishing.
 *
 * The merged registry is materialized by the canonical `buildRegistry`
 * (re-exported from the generator, which ships) over a cap-map reconstructed
@@ -845,12 +847,12 @@ export function loadRegistry(options: LoadRegistryOptions = {}): Registry {
    // to load. Clear the map (the first-party base never lists overlay commandRoots — first-party
    // command modules ship in bin/lib/, not via _overlay.commandRoots).
    meta.commandRoots = {};
-    // #1461 OVL-2 fail-CLOSED on compose failure (HIGH): the fallback DROPS every accepted overlay,
-    // so any accepted overlay that DECLARED a gate would have its gate silently vanish → a blocking
-    // gate FAILS OPEN, violating ADR-1244 (a skipped capability declaring a gate must FAIL CLOSED).
+    // #1461 OVL-2 (HIGH): on compose failure the fallback DROPS every accepted overlay, so any
+    // accepted overlay that DECLARED a gate would have its gate silently vanish with no trace.
    // Record each dropped gate-declaring overlay's gate as blocked using the SAME extraction the
-    // per-candidate `skip()` closure uses (gatePointsOf), so loop-resolver injects the synthetic
-    // blocking gate at each declared point exactly as it would for a per-candidate skip.
+    // per-candidate `skip()` closure uses (gatePointsOf), so loop-resolver surfaces the loud
+    // fail-OPEN advisory (#2009) at each declared point exactly as it would for a per-candidate
+    // skip — the gate does not silently disappear.
    for (const cap of overlayCaps) {
      const gatePoints = gatePointsOf(cap);
      if (gatePoints.length === 0) continue;
--- a/src/loop-resolver.cts
+++ b/src/loop-resolver.cts
@@ -454,6 +454,27 @@ function renderLoopHooks(resolved: ResolveLoopHooksResult): string {
 * Missing <capId> value → coreError + non-zero exit.
 * Unknown/inactive capId → `false` (not an error).
 */
+// #2009: a capability id surfaced inside the runnable `gsd capability remove <id>`
+// remediation must match the canonical kebab-case id shape (identical to
+// capability-consent.cts / capability-ledger.cts) before it is embedded — a raw
+// overlay directory name is attacker-controlled and can carry shell/markdown
+// metacharacters (backticks, ';', '|', '$()'). An id that fails this check is
+// withheld and no runnable command is rendered for it.
+const LOAD_FAIL_CAP_ID_RE = /^[a-z][a-z0-9-]*$/;
+
+// #2009: neutralize control chars, newlines, and backticks from a third-party
+// load-failure reason so a malicious manifest cannot break out of the warning
+// line or inject markdown / prompt content into the surfaced message.
+function sanitizeLoadFailReason(reason: unknown): string {
+  const cleaned = String(reason)
+    // Strip C0 control chars, DEL, and backticks; collapse remaining whitespace.
+    .replace(/[\x00-\x1F\x7F`]/g, ' ')
+    .replace(/\s+/g, " ")
+    .trim()
+    .slice(0, 300);
+  return cleaned || '(no reason given)';
+}
+
 function cmdLoopRenderHooks(
  cwd: string,
  point: string,
@@ -519,26 +540,55 @@ function cmdLoopRenderHooks(
    return;
  }

-  // ── ADR-1244 D2 fail-closed gate injection ────────────────────────────────────
-  // For every skipped overlay capability that declared a gate at this point,
-  // inject a synthetic BLOCKING gate into the resolved output so the loop HALTS
-  // rather than silently proceeding as if the gate had passed. step/contribution
-  // overlays that were skipped are left open (skip-open is correct for them).
+  // ── ADR-1244 D2: load-failed capability gates FAIL OPEN with a loud warning ────
+  // Decision (#2009): a capability that failed to LOAD must not block the loop.
+  // The prior behavior injected a BLOCKING synthetic gate (blocking:true,
+  // onError:'halt') at every point where the skipped cap declared a gate, so a
+  // single incompatible capability halted every ship:pre / verify:post
+  // project-wide for a load error unrelated to what the gate checked — with no
+  // remediation surfaced. We now fail OPEN: no gate is injected (the loop proceeds
+  // and `--active-cap <failed-cap>` correctly reports it inactive), and a loud
+  // warning is emitted instead — to STDERR (which the operator, or the agent
+  // running the command, actually sees regardless of how the host workflow
+  // consumes stdout) AND in the envelope's `warnings` channel for structured
+  // consumers. The warning names the load reason and the exact
+  // `gsd capability remove <id>` remediation so the operator can clear the broken
+  // capability. blockedGates is still recorded by the loader; only the consequence
+  // changes from block to warn. step/contribution overlays were already skip-open.
+  //
+  // The gate injection was dropped rather than made non-blocking because no host
+  // workflow generically surfaces an arbitrary gate's message at ship:pre /
+  // verify:post (consumers dispatch on specific capIds / ref.skills), and the
+  // generic gate consumers expect an object-shaped `check`, not a prose string —
+  // so an injected advisory gate would be silently dropped or mis-dispatched. A
+  // stderr warning is the channel that is actually surfaced. (See #2009 review.)
  const overlayMeta = (registry as { _overlay?: { blockedGates?: Array<{ point: string; capId: string; reason: string }> } })['_overlay'];
+  const loadFailWarnings: string[] = [];
  if (overlayMeta && Array.isArray(overlayMeta.blockedGates)) {
    for (const blocked of overlayMeta.blockedGates) {
-      if (blocked.point === point) {
-        const syntheticGate: ActiveHook = {
-          capId: blocked.capId,
-          kind: 'gate',
-          blocking: true,
-          onError: 'halt',
-          check: `capability "${blocked.capId}" was skipped at load (${blocked.reason}); its gate at ${point} cannot be evaluated — failing closed`,
-        };
-        resolved.activeHooks.push(syntheticGate);
-      }
+      if (blocked.point !== point) continue;
+      // Security (#2009 review): capId/reason come from a third-party manifest or
+      // directory name. Validate capId before embedding it in the runnable
+      // remediation command; withhold it (no runnable command) if it is not a
+      // canonical id. Strip control chars/backticks from reason.
+      const idValid = LOAD_FAIL_CAP_ID_RE.test(String(blocked.capId));
+      const capLabel = idValid
+        ? `"${blocked.capId}"`
+        : 'with an invalid id (withheld) under .gsd/capabilities/';
+      const remediation = idValid
+        ? `Run \`gsd capability remove ${blocked.capId}\` to remove it, or fix the load error.`
+        : 'Remove the offending capability directory under .gsd/capabilities/, or fix the load error.';
+      loadFailWarnings.push(
+        `capability ${capLabel} failed to load (${sanitizeLoadFailReason(blocked.reason)}); ` +
+        `its gate at ${point} is SKIPPED and NOT enforced (failing open). ${remediation}`,
+      );
    }
  }
+  // Emit loudly to stderr in EVERY output mode (including --active-cap), so a
+  // skipped gate is never silently invisible to the operator/agent.
+  for (const w of loadFailWarnings) {
+    process.stderr.write(`gsd: warning — ${w}\n`);
+  }

  // --active-cap mode: print exactly 'true' or 'false' with no envelope
  if (activeCapId !== undefined) {
@@ -558,8 +608,12 @@ function cmdLoopRenderHooks(
    activeHooks: resolved.activeHooks,
    rendered,
  };
-  if (state.warnings && state.warnings.length > 0) {
-    envelope.warnings = state.warnings;
+  // Surface capability-state warnings and the #2009 load-failure fail-open
+  // warnings together in the structured `warnings` channel (in addition to the
+  // stderr emission above, which is the channel host workflows actually see).
+  const combinedWarnings = [...(state.warnings || []), ...loadFailWarnings];
+  if (combinedWarnings.length > 0) {
+    envelope.warnings = combinedWarnings;
  }

  coreOutput(envelope, raw);
--- a/tests/loop-render-hooks.test.cjs
+++ b/tests/loop-render-hooks.test.cjs
@@ -1039,79 +1039,136 @@ describe('Phase 4 regression: capabilityStatesById gates on active (not enabled)
  });
 });

-// ─── ADR-1244 D2 fail-closed gate injection ────────────────────────────────────
+// ─── ADR-1244 D2: load-failed capability gates FAIL OPEN with a loud warning (#2009) ──

-describe('ADR-1244 D2: fail-closed gate injection for skipped overlay caps with gates', () => {
-  // Verifies that cmdLoopRenderHooks injects a BLOCKING synthetic gate at the
-  // declared point when an overlay capability that declares a gate is skipped at
-  // load time due to an incompatible engines.gsd version constraint.
-  //
-  // Fixture: overlay cap declares a gate at execute:wave:post with engines.gsd: ">=99.0.0"
-  // → loadRegistry skips it → records it in _overlay.blockedGates
-  // → cmdLoopRenderHooks injects a blocking=true, onError=halt gate at execute:wave:post
+describe('ADR-1244 D2: load-failed capability gates fail OPEN with a loud warning (#2009)', () => {
+  // Decision (#2009): a capability that fails to LOAD must not block the loop.
+  // cmdLoopRenderHooks injects NO gate (the loop proceeds — fail open) and emits a
+  // loud warning naming the load reason and the exact `gsd capability remove <id>`
+  // remediation, both to STDERR (the channel host workflows actually surface) and
+  // in the envelope's `warnings` array. Previously it injected a blocking=true,
+  // onError=halt gate that halted every declared point project-wide.

-  test('skipped gate-kind overlay cap → BLOCKING synthetic gate at its declared point', (t) => {
-    const overlayHome = fs.mkdtempSync(path.join(os.tmpdir(), 'loop-fail-closed-'));
-    t.after(() => cleanup(overlayHome));
-
-    // Write an overlay capability that:
-    // - declares a gate at execute:wave:post
-    // - has engines.gsd: ">=99.0.0" (incompatible → will be skipped at load)
-    const capId = 'fail-closed-gate-cap';
+  // Helper: write an overlay cap with an incompatible engines.gsd (→ skipped at load)
+  // that declares gate(s) at the given point(s). `capId` may be an invalid id to
+  // exercise the sanitization path.
+  function writeSkippedGateCap(overlayHome, capId, gatePoints) {
    const capDir = path.join(overlayHome, '.gsd', 'capabilities', capId);
    fs.mkdirSync(capDir, { recursive: true });
    const capManifest = {
      id: capId,
      role: 'feature',
      version: '1.0.0',
-      title: 'Fail Closed Gate Cap',
-      description: 'ADR-1244 D2 fail-closed wiring test',
+      title: 'Load-Failed Gate Cap',
+      description: 'ADR-1244 D2 fail-open wiring test',
      tier: 'standard',
      requires: [],
-      engines: { gsd: '>=99.0.0' },  // intentionally incompatible → always skipped
+      engines: { gsd: '>=99.0.0' },  // intentionally incompatible → always skipped at load
      runtimeCompat: { supported: ['*'], unsupported: [] },
      skills: [], agents: [], hooks: [], config: {}, steps: [], contributions: [],
-      gates: [{ point: 'execute:wave:post', check: 'always-pass', blocking: true, onError: 'halt' }],
+      gates: gatePoints.map((point) => ({ point, check: { query: 'always-pass' }, blocking: true, onError: 'halt' })),
    };
    fs.writeFileSync(path.join(capDir, 'capability.json'), JSON.stringify(capManifest), 'utf8');
+  }

-    // Invoke gsd-tools via subprocess so stdout is the real fd-1 (io.cjs writes via writeSync).
-    // Set GSD_HOME to the overlay home so loadRegistry picks up the incompatible cap.
+  function renderHooks(overlayHome, point, extraArgs = []) {
    const result = spawnSync(
      process.execPath,
-      [GSD_TOOLS, 'loop', 'render-hooks', 'execute:wave:post', '--cwd', overlayHome],
-      {
-        cwd: ROOT,
-        encoding: 'utf8',
-        env: { ...process.env, GSD_HOME: overlayHome },
-      },
+      [GSD_TOOLS, 'loop', 'render-hooks', point, '--cwd', overlayHome, ...extraArgs],
+      { cwd: ROOT, encoding: 'utf8', env: { ...process.env, GSD_HOME: overlayHome } },
    );
+    assert.strictEqual(result.status, 0, `Expected exit 0 at ${point}. stderr: ` + (result.stderr || ''));
+    return result;
+  }

-    assert.strictEqual(result.status, 0, 'Expected exit 0. stderr: ' + (result.stderr || ''));
-
-    let envelope;
+  function parseEnvelope(result) {
    try {
-      envelope = JSON.parse(result.stdout.trim());
+      return JSON.parse(result.stdout.trim());
    } catch {
      assert.fail('loop render-hooks output must be valid JSON; got: ' + result.stdout.slice(0, 300));
    }
+  }

-    // The synthetic blocking gate must be present in activeHooks
-    const syntheticGate = Array.isArray(envelope.activeHooks)
-      ? envelope.activeHooks.find((h) => h.capId === capId && h.kind === 'gate')
+  test('load-failed gate-kind overlay cap → NO gate injected (fail open) + loud warning on stderr and in envelope', (t) => {
+    const overlayHome = fs.mkdtempSync(path.join(os.tmpdir(), 'loop-fail-open-'));
+    t.after(() => cleanup(overlayHome));
+
+    const capId = 'load-failed-gate-cap';
+    writeSkippedGateCap(overlayHome, capId, ['execute:wave:post']);
+
+    const result = renderHooks(overlayHome, 'execute:wave:post');
+    const envelope = parseEnvelope(result);
+
+    // AC2 / AC-a: fail OPEN — NO gate (blocking or otherwise) is injected for the
+    // load-failed cap, so the loop proceeds.
+    const anyGate = Array.isArray(envelope.activeHooks)
+      ? envelope.activeHooks.find((h) => h.capId === capId)
      : undefined;
-    assert.ok(
-      syntheticGate !== undefined,
-      `activeHooks must contain a synthetic gate attributed to ${capId} (fail-closed injection). ` +
-      'Got: ' + JSON.stringify(envelope.activeHooks),
+    assert.strictEqual(
+      anyGate, undefined,
+      'no hook must be injected for a load-failed cap (fail open). Got: ' + JSON.stringify(envelope.activeHooks),
    );
-    assert.strictEqual(syntheticGate.blocking, true, 'synthetic gate must be blocking=true');
-    assert.strictEqual(syntheticGate.onError, 'halt', 'synthetic gate must have onError=halt');

-    // The rendered markdown must also reference the gate cap
+    // AC-b: a loud warning is surfaced in the envelope `warnings` channel...
+    const warnEnv = (envelope.warnings || []).find((w) => w.includes(capId));
+    assert.ok(warnEnv, 'envelope.warnings must name the load-failed cap. Got: ' + JSON.stringify(envelope.warnings));
+    // AC1: ...carrying the exact remediation, and making clear the gate is not enforced.
    assert.ok(
-      typeof envelope.rendered === 'string' && envelope.rendered.includes(capId),
-      'rendered output must reference the fail-closed gate cap. Got: ' + envelope.rendered,
+      warnEnv.includes(`gsd capability remove ${capId}`),
+      'warning must include the `gsd capability remove <id>` remediation. Got: ' + warnEnv,
+    );
+    assert.match(warnEnv, /skipped|not enforced|failing open/i, 'warning must say the gate is not enforced. Got: ' + warnEnv);
+
+    // ...and ALSO to stderr (the channel host workflows actually see).
+    assert.ok(
+      result.stderr.includes(`gsd capability remove ${capId}`),
+      'stderr must carry the loud fail-open warning with remediation. Got: ' + result.stderr,
+    );
+  });
+
+  test('load-failed cap declaring gates at ship:pre AND verify:post → neither point blocks (project can ship & verify)', (t) => {
+    const overlayHome = fs.mkdtempSync(path.join(os.tmpdir(), 'loop-fail-open-2pt-'));
+    t.after(() => cleanup(overlayHome));
+
+    const capId = 'load-failed-two-point-cap';
+    writeSkippedGateCap(overlayHome, capId, ['ship:pre', 'verify:post']);
+
+    for (const point of ['ship:pre', 'verify:post']) {
+      const result = renderHooks(overlayHome, point);
+      const envelope = parseEnvelope(result);
+      const injected = envelope.activeHooks.find((h) => h.capId === capId);
+      assert.strictEqual(
+        injected, undefined,
+        `${point} must NOT inject any hook for a load-failed cap (fail open). Got: ` + JSON.stringify(envelope.activeHooks),
+      );
+      const warn = (envelope.warnings || []).find((w) => w.includes(`gsd capability remove ${capId}`));
+      assert.ok(warn, `${point} must surface a fail-open warning with remediation. Got: ` + JSON.stringify(envelope.warnings));
+    }
+  });
+
+  test('security: an invalid (non-kebab) capability id is withheld — no runnable remove command is rendered (#2009 review)', (t) => {
+    const overlayHome = fs.mkdtempSync(path.join(os.tmpdir(), 'loop-fail-open-evil-'));
+    t.after(() => cleanup(overlayHome));
+
+    // A directory name that is a valid POSIX filename but NOT a valid capability
+    // id, and contains a shell metacharacter. It must never appear inside a
+    // runnable `gsd capability remove <id>` command in the surfaced warning.
+    const evilId = 'Bad;Cap';
+    writeSkippedGateCap(overlayHome, evilId, ['ship:pre']);
+
+    const result = renderHooks(overlayHome, 'ship:pre');
+    const envelope = parseEnvelope(result);
+
+    const combined = (envelope.warnings || []).join('\n') + '\n' + result.stderr;
+    // The raw metacharacter id must not be embedded in a runnable remove command.
+    assert.ok(
+      !combined.includes(`gsd capability remove ${evilId}`),
+      'invalid capability id must NOT be placed in a runnable remove command. Got: ' + combined,
+    );
+    // A warning is still surfaced (fail-open is loud), using the withheld-id form.
+    assert.match(
+      combined, /invalid id \(withheld\)/,
+      'an invalid id must be reported via the withheld-id placeholder. Got: ' + combined,
    );
  });
 });
--- a/tests/perf-316-state-lock-buffer-alloc.test.cjs
+++ b/tests/perf-316-state-lock-buffer-alloc.test.cjs
@@ -57,9 +57,13 @@ const fs = require('fs');
 // Write pid to lock file so acquireStateLock sees a live pid and retries.
 fs.writeFileSync(workerData.lockPath, String(process.pid));
 parentPort.postMessage({ pid: process.pid });
-// Synchronous sleep — blocks this worker thread for holdMs ms.
-const buf = new Int32Array(new SharedArrayBuffer(4));
-Atomics.wait(buf, 0, 0, workerData.holdMs);
+// Hold the lock until the writer signals it has made its first FAILED lock
+// attempt (deterministic contention), or until holdMs elapses as a safety cap.
+// Atomics.wait blocks this thread while releaseFlag[0] === 0. A fixed timer here
+// was racy: on a slow/loaded runner the writer's spawn+init could exceed the
+// timer, so the lock released before the writer ever contended (lockAttempts:1).
+const releaseFlag = new Int32Array(workerData.releaseSab);
+Atomics.wait(releaseFlag, 0, 0, workerData.holdMs);
 // Release the lock.
 try { fs.unlinkSync(workerData.lockPath); } catch { /* already gone */ }
 parentPort.postMessage({ done: true });
@@ -88,17 +92,29 @@ global.SharedArrayBuffer = StubSAB;
 // path — without this witness, a no-retry success would yield sabCount === 1
 // from BOTH pre-fix and post-fix code (the SAB is allocated unconditionally
 // post-fix, and exactly once for the single successful open pre-fix), giving
-// a false-pass against the bug. The 1000ms holdMs + 200ms SUT retry delay
-// guarantees >=4 attempts even on the slowest CI runners.
+// a false-pass against the bug. Contention is deterministic here: this worker
+// signals the holder to release only after its first failed attempt, so a
+// retry always occurs (lockAttempts >= 2) regardless of runner speed.
 const realOpenSync = fs.openSync.bind(fs);
+const releaseFlag = new Int32Array(workerData.releaseSab);
 let lockAttempts = 0;
 fs.openSync = function(filePath, flags, mode) {
-  if (typeof filePath === 'string' && filePath.endsWith('.lock') &&
+  const isLockCreate = typeof filePath === 'string' && filePath.endsWith('.lock') &&
      typeof flags === 'number' &&
-      (flags & fs.constants.O_CREAT) && (flags & fs.constants.O_EXCL)) {
-    lockAttempts++;
+      (flags & fs.constants.O_CREAT) && (flags & fs.constants.O_EXCL);
+  if (isLockCreate) lockAttempts++;
+  try {
+    return realOpenSync(filePath, flags, mode);
+  } catch (e) {
+    // On the FIRST failed atomic-create (lock is held → contention proven),
+    // signal the holder worker to release so the next retry succeeds. Makes the
+    // retry path deterministic regardless of worker-spawn latency.
+    if (isLockCreate && lockAttempts === 1) {
+      Atomics.store(releaseFlag, 0, 1);
+      Atomics.notify(releaseFlag, 0);
+    }
+    throw e;
  }
-  return realOpenSync(filePath, flags, mode);
 };

 // Delete cache entry to ensure a fresh require picks up the stubbed constructor.
@@ -160,17 +176,18 @@ describe('perf #316: acquireStateLock hoists sleep buffer — exactly one SAB pe
    { timeout: 8000 },
    async () => {
      // ── Worker A: hold the lock for 1000ms ─────────────────────────────────
-      // state.cjs retry delay = 200ms + 0-50ms jitter; 1000ms hold guarantees
-      // >=4 retries even on the slowest CI worker (~200ms spawn + 4 retry
-      // intervals ~1000ms ≈ hold duration). The lockAttempts assertion below
-      // proves the retry path was exercised end-to-end.
+      // The holder releases the lock only when the writer signals its first
+      // failed lock attempt (see WRITER_WORKER_CODE), so the writer is guaranteed
+      // to contend at least once regardless of worker-spawn latency. holdMs is now
+      // only a safety cap in case that signal never arrives.
      const holdMs = 1000;
+      const releaseSab = new SharedArrayBuffer(4);
      let resolveLockWritten;
      const lockWritten = new Promise((resolve) => { resolveLockWritten = resolve; });
      const holderDone = new Promise((resolve, reject) => {
        holderWorker = new Worker(HOLDER_WORKER_CODE, {
          eval: true,
-          workerData: { lockPath, holdMs },
+          workerData: { lockPath, holdMs, releaseSab },
        });
        holderWorker.on('message', (msg) => {
          if (msg.pid !== undefined) resolveLockWritten();
@@ -223,6 +240,7 @@ describe('perf #316: acquireStateLock hoists sleep buffer — exactly one SAB pe
              statePath,
              content: MINIMAL_STATE_MD,
              tmpDir,
+              releaseSab,
            },
          });
          writerWorker.on('message', resolve);
@@ -259,8 +277,8 @@ describe('perf #316: acquireStateLock hoists sleep buffer — exactly one SAB pe
      assert.ok(
        writeResult.lockAttempts >= 2,
        'SUT must have entered the retry path (>=1 failed lock attempt before success). ' +
-          'Got lockAttempts: ' + writeResult.lockAttempts + '. The 1000ms holdMs + 200ms ' +
-          'SUT retry delay guarantees >=2 attempts on any CI runner.'
+          'Got lockAttempts: ' + writeResult.lockAttempts + '. The holder releases only ' +
+          'after the writer signals its first failed attempt, so contention is deterministic.'
      );

      // THE KEY INVARIANT: