diff --git a/CHANGELOG.md b/CHANGELOG.md
index 9514711..6387c99 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,7 +4,9 @@ All notable changes to OLP land here. Per `CLAUDE.md` release_kit overlay, this
 
 ## Unreleased
 
-(empty — Phase 2 entries land here once Phase 2 opens)
+### D38 — maxConcurrent runtime enforcement (issue #1)
+
+- **Spawn lifecycle gate** — `hints.maxConcurrent` is now enforced at runtime per ADR 0002 Amendment 6: `lib/providers/index.mjs` exports a per-provider `tryAcquireSpawn` / `releaseSpawn` / `getActiveSpawnCount` semaphore; `server.mjs` gates both the buffered and streaming spawn call sites in `handleChatCompletions` with a try/finally release. Saturation surfaces as `ProviderError(CONCURRENCY_LIMIT)` which the fallback engine treats as a hard trigger (ADR 0004 Amendment 4) — the chain advances to the next hop instead of queueing. If the entire chain is saturated, the user receives a chain-exhausted error via the existing exhaustion path. Closes #1. Queue+timeout deferred (see ADR 0002 Amendment 6 § Design choice). Test count 431 → 447.
 
 ## v0.1.0 — 2026-05-24
 
diff --git a/docs/adr/0002-plugin-architecture.md b/docs/adr/0002-plugin-architecture.md
index 58e19b0..2eff2d2 100644
--- a/docs/adr/0002-plugin-architecture.md
+++ b/docs/adr/0002-plugin-architecture.md
@@ -7,6 +7,30 @@
 
 ## Amendments
 
+### Amendment 6 — 2026-05-24: `maxConcurrent` runtime enforcement landed (D38, issue #1)
+
+- **Finding:** Amendment 1 (2026-05-23) ratified `maxSpawnTimeMs` into the Provider contract but explicitly noted that `hints.maxConcurrent` remained **declarative-only at v0.1** — type-validated at startup in `lib/providers/base.mjs` (`validateProvider` requires it to be a non-negative integer) but unenforced at runtime (no semaphore / in-flight counter / spawn queue in `server.mjs`). Cold-audit catch from D11 (commit `f659e29`): the diff-review reviewer grep-verified that the original ADR draft's claim "Enforced by the spawn-concurrency guard in `server.mjs`" was false. GitHub issue #1 was filed to track the gap. D38 closes that gap.
+- **Change (D38):**
+  - Add a per-provider in-flight semaphore in `lib/providers/index.mjs` exporting three primitives plus a constant:
+    - `tryAcquireSpawn(providerName, maxConcurrent)` — atomic check-then-increment; returns `true` on success, `false` if at limit. Atomicity rests on the JS single-threaded invariant — the read and write are synchronous with NO `await` between them. A future async refactor MUST preserve this.
+    - `releaseSpawn(providerName)` — decrement; throws if the count would go negative (defensive bug guard for missing acquire / double release).
+    - `getActiveSpawnCount(providerName)` — returns current in-flight count; exported for diagnostics and tests (server.mjs uses it to populate the `activeSpawns` field on a synthesised `CONCURRENCY_LIMIT` error). `/health` integration deferred — when surfaced there it will land at `providers.status.<name>.activeSpawns`; not wired at D38.
+    - `DEFAULT_MAX_CONCURRENT_SPAWNS = 4` — defense-in-depth fallback when a plugin path bypasses `validateProvider` and passes undefined/null/NaN. The value matches the v0.1 plugin defaults (anthropic / codex / mistral all declare `hints.maxConcurrent: 4`).
+  - Wire the gate at both `provider.spawn(...)` call sites in `server.mjs handleChatCompletions`:
+    - **Buffered path** (inside `executeHopFn → collectAllChunks`): `tryAcquireSpawn` runs before `provider.spawn(...)`. On failure, synthesise `ProviderError(CONCURRENCY_LIMIT)` with `providerName` / `maxConcurrent` / `activeSpawns` fields for diagnostics and re-throw — the fallback engine treats it as a hard trigger (see ADR 0004 Amendment 4) and advances to the next chain hop. On success, the spawn drain loop runs inside a `try { … } finally { releaseSpawn(...) }` so the slot releases on every exit path (success, error, D16 SPAWN_FAILED salvage return, unexpected throw).
+    - **Streaming path** (single-hop real-SSE, `chain.length === 1` cache-miss branch): acquire happens BEFORE the streaming branch entry. If acquire fails, the branch is skipped and the request falls through to the buffered path — that path's own gate re-attempts acquire; a single-hop chain at maxConcurrent has no other hop to advance to, so the request surfaces a chain-exhausted error via `executeWithFallback`'s exhaustion path. If acquire succeeds, the existing streaming try/catch gains a `finally { releaseSpawn(streamProvider) }` so the slot releases on stop-chunk completion, generator exhaustion, abort, or any exception path.
+  - Add `CONCURRENCY_LIMIT` to `PROVIDER_ERROR_CODES` in `lib/providers/base.mjs` so the synthesised error type-checks with the existing closed enum. (Note: `CONCURRENCY_LIMIT` is synthesised by the orchestration layer, NOT thrown by provider plugins themselves — the code is in the enum for type consistency with the fallback engine's `HARD_TRIGGER_CODES` lookup.)
+  - Update the `maxConcurrent` description in § Decision (Provider contract hints) below — remove the "Declarative hint only at v0.1" caveat and add the implementation reference.
+- **Update to § Decision § Provider contract hints (`maxConcurrent`):** replace the v0.1 caveat with: "`maxConcurrent` — integer; maximum simultaneous spawn count OLP will allow for this provider. The value is type-validated at startup (`lib/providers/base.mjs validateProvider`) and enforced at runtime by `tryAcquireSpawn` / `releaseSpawn` in `lib/providers/index.mjs`. Saturation surfaces as `ProviderError(CONCURRENCY_LIMIT)` (per `PROVIDER_ERROR_CODES`, `lib/providers/base.mjs`), which the fallback engine treats as a hard trigger per ADR 0004 Amendment 4 — the chain advances to the next hop. If the entire chain is saturated, the user receives a chain-exhausted error via the existing `executeWithFallback` exhaustion path."
+- **Design choice — immediate-advancement vs. queue+timeout:** D38 implements immediate-advancement through the fallback chain. Rationale:
+  1. The fallback chain exists precisely for this kind of overflow — saturation on the primary hop is a natural fit for the existing advancement mechanism.
+  2. Queue+timeout introduces head-of-line blocking risk (a stuck/slow spawn blocks queued waiters) and adds a new timeout config surface (`hints.maxConcurrentWaitMs`?) that the contract does not currently have.
+  3. Immediate-advancement gives fail-fast latency and matches the OLP multi-provider proxy philosophy (the user has spread their quota across providers explicitly so saturation should reach an alternate provider as fast as possible).
+  4. Queue+timeout is **deferred to a future iteration** if real usage shows demand. Track via a follow-up issue if the design pressure surfaces.
+- **Authority:** ALIGNMENT.md Rule 1 (Cite First) — internal authority is ADR 0002 (this ADR) + ADR 0004 (which adds CONCURRENCY_LIMIT to the hard-trigger taxonomy in its Amendment 4). No provider CLI doc cited because this change is internal to the orchestration layer; no provider plugin code changes (anthropic / codex / mistral already declare `hints.maxConcurrent` correctly per validateProvider).
+- **Tests:** Suite 18 in `test-features.mjs` — 16 tests covering: `PROVIDER_ERROR_CODES` membership, `evaluateHardTriggers(CONCURRENCY_LIMIT)` returns true, semaphore unit behaviour (acquire / release / count / reset), saturation rejection, defensive coercion of non-integer maxConcurrent, double-release throws, HTTP-level concurrent-request peak-in-flight assertion (5 requests against maxConcurrent:2 → peak == 2), buffered-path counter release, streaming-path counter release, fallback advancement to secondary on saturated primary.
+- **Procedural mechanism:** CC 开发铁律 v1.6 § 10.x (D38 issue #1 implementation; fresh-context opus reviewer to follow this implementation per Iron Rule 10).
+
 ### Amendment 5 — 2026-05-24: Correct § Decision filesystem layout — `vibe.mjs` → `mistral.mjs` (D36 #5)
 
 - **Finding:** Issue #5 (D36) — § Decision filesystem layout (around line 47 of the original ADR) listed the Mistral provider plugin as `vibe.mjs` (named after the CLI binary `vibe`). The shipped file at `lib/providers/mistral.mjs` (D8) is named after the provider key, matching the established convention from the other two plugins: `anthropic.mjs` (provider key `anthropic`, CLI `claude`) and `codex.mjs` (provider key `openai`, CLI `codex`). The ADR's `vibe.mjs` entry was a drafting-time placeholder that did not get corrected when D8 landed `lib/providers/mistral.mjs`.
@@ -94,7 +118,7 @@ Every provider plugin exports an object conforming to:
 - `hints: { requiresTTY, concurrentSpawnSafe, maxConcurrent, maxSpawnTimeMs, cacheable }` — fingerprint, concurrency, timeout, and cache hints:
   - `requiresTTY` — boolean; whether the provider CLI requires a TTY to produce non-interactive output (e.g., some CLIs suppress JSON output unless forced with a flag or a TTY is present).
   - `concurrentSpawnSafe` — boolean; whether the provider CLI is safe to spawn concurrently under the same auth context without rate-limit or session collisions.
-  - `maxConcurrent` — integer; maximum simultaneous spawn count OLP will allow for this provider. **Declarative hint only at v0.1**: the value is type-validated at startup (`lib/providers/base.mjs`) but no runtime enforcement (semaphore / in-flight counter / spawn queue) is wired in `server.mjs` yet. Tracking issue to be filed for a follow-up that lands the runtime guard.
+  - `maxConcurrent` — integer; maximum simultaneous spawn count OLP will allow for this provider. The value is type-validated at startup (`lib/providers/base.mjs validateProvider`) and **enforced at runtime** by `tryAcquireSpawn` / `releaseSpawn` in `lib/providers/index.mjs` (D38 — see Amendment 6). Saturation surfaces as `ProviderError(CONCURRENCY_LIMIT)`, which the fallback engine treats as a hard trigger per ADR 0004 Amendment 4 — the chain advances to the next hop. If the entire chain is saturated, the user receives a chain-exhausted error via the existing `executeWithFallback` exhaustion path. (Pre-D38 caveat removed; tracking issue #1 closed by Amendment 6.)
   - `maxSpawnTimeMs` — optional integer, milliseconds; maximum wall-clock time OLP allows for a single provider spawn before treating it as a hard fallback trigger. Defaults to `600000` (10 minutes) if absent. Enforcement lives inside each provider plugin's spawn drain loop (`_spawnAndStream`), which uses a `setTimeout` / `proc.kill` / `reject` pattern to throw `ProviderError(SPAWN_TIMEOUT)`; the fallback engine then treats this error as a hard trigger (ADR 0004 § Trigger taxonomy — Hard triggers bullet 4). The engine itself does not run the timer loop; it only acts on the thrown error.
   - `cacheable` — optional boolean, default `true`; if explicitly set to `false`, the provider opts out of OLP's response cache entirely. `executeHopFn` skips `cacheStore.getOrCompute` and calls `collectAllChunks` directly; no cache read or write occurs for any request to this provider. Omitting the field is equivalent to `cacheable: true`. See ADR 0005 § "Cache write conditions" item 3 and Amendment 3 above. (D23)
 
diff --git a/docs/adr/0004-fallback-engine.md b/docs/adr/0004-fallback-engine.md
index f1a9daf..a27e922 100644
--- a/docs/adr/0004-fallback-engine.md
+++ b/docs/adr/0004-fallback-engine.md
@@ -7,6 +7,23 @@
 
 ## Amendments
 
+### Amendment 4 — 2026-05-24: Add `CONCURRENCY_LIMIT` to v0.1 hard-trigger code taxonomy (D38, issue #1)
+
+- **Finding:** ADR 0002 Amendment 1 (2026-05-23) ratified `hints.maxConcurrent` into the Provider contract as **declarative-only at v0.1** — no runtime enforcement. GitHub issue #1 tracked the gap. D38 lands runtime enforcement (see ADR 0002 Amendment 6 for the implementation details and design rationale). Once a saturation event occurs, the orchestration layer must communicate "this hop is at capacity — advance the chain" to the fallback engine using a code that fits the existing hard-trigger taxonomy in `evaluateHardTriggers`.
+- **Change (D38):** Add `CONCURRENCY_LIMIT` to both:
+  - `PROVIDER_ERROR_CODES` in `lib/providers/base.mjs` (closed enum used by `ProviderError`).
+  - `HARD_TRIGGER_CODES` in `lib/fallback/engine.mjs` — value `true` so `evaluateHardTriggers(ProviderError CONCURRENCY_LIMIT)` returns `true` and `classifyTrigger` returns `'hard'`. The chain advances to the next hop. The synthesised error carries diagnostic fields (`providerName`, `maxConcurrent`, `activeSpawns`) which surface in the existing `fallback_hard_trigger` log event via the `error.message` field.
+- **v0.1 live hard-trigger codes after this amendment (5 codes):** `SPAWN_FAILED`, `CLI_NOT_FOUND`, `SPAWN_TIMEOUT`, `CONCURRENCY_LIMIT`, plus the explicit non-trigger `AUTH_MISSING:false`. Pre-D38 list (per Amendment 3) was 4 codes (`SPAWN_FAILED`, `CLI_NOT_FOUND`, `SPAWN_TIMEOUT` as triggers + `AUTH_MISSING:false`).
+- **Synthesis vs. plugin-thrown:** Unlike the other live hard-trigger codes, `CONCURRENCY_LIMIT` is **NOT thrown by provider plugins themselves**. It is synthesised by `server.mjs handleChatCompletions` when `tryAcquireSpawn(provider, hints.maxConcurrent)` returns false. The code lives in `PROVIDER_ERROR_CODES` for type consistency with the closed enum that `HARD_TRIGGER_CODES` keys on; the orchestration layer is the only callsite that throws it. A future provider plugin that gains its own internal concurrency limit (e.g., a CLI that returns a specific exit code on rate-limit) could thrown this code too; the enum is forward-compatible.
+- **Design choice — immediate-advancement vs. queue+timeout:** Re-stating from ADR 0002 Amendment 6 because this ADR governs the trigger taxonomy that surfaces the decision to the user: saturation is treated as a **hard trigger** (chain advances immediately) rather than as a **soft trigger** (would gate before spawn but would not advance after spawn attempt) or as a queueable condition (would block + timeout). The hard-trigger framing matches "the primary hop refused to serve this request; advance" semantics. Queue+timeout would require a NEW trigger category outside the existing taxonomy (hard / soft / deterministic-deferred / cost-aware-deferred) and is deferred per ADR 0002 Amendment 6 rationale.
+- **First-chunk safety:** `tryAcquireSpawn` runs **before** `provider.spawn(...)` and before any bytes are written to the response. A `CONCURRENCY_LIMIT` rejection therefore satisfies the first-chunk rule trivially — zero bytes have been emitted to the client. Fallback is safe.
+- **Authority:** ADR 0002 Amendment 6 (runtime enforcement implementation); GitHub issue #1 (tracking).
+- **Tests:** Suite 18 in `test-features.mjs` — see ADR 0002 Amendment 6 § Tests for the full list. Specifically for this ADR: tests 18a (PROVIDER_ERROR_CODES membership), 18b (`evaluateHardTriggers(CONCURRENCY_LIMIT) === true`), 18c (AUTH_MISSING regression guard — D38 did not flip it), 18k (chain advances to fallback hop on saturated primary).
+- **v1.x re-evaluation triggers:**
+  - If a future plugin gains a CLI-level concurrency response that should NOT be a hard trigger (e.g., "soft limit hit, retry after backoff") — file a follow-up to add a new code (e.g., `CONCURRENCY_BACKOFF`) rather than reclassifying `CONCURRENCY_LIMIT`.
+  - If queue+timeout becomes desirable (real usage shows fail-fast advancement is too aggressive for certain workloads), file an amendment to this ADR adding queue semantics as a NEW trigger category — do not reclassify CONCURRENCY_LIMIT into the existing taxonomy.
+- **Procedural mechanism:** CC 开发铁律 v1.6 § 10.x (D38 issue #1 implementation; fresh-context opus reviewer to follow per Iron Rule 10).
+
 ### Amendment 3 — 2026-05-24: Narrow v0.1 hard-trigger code taxonomy (D34 F7)
 
 - **Finding:** Round-6 cold-audit F7 (P2) — `PROVIDER_ERROR_CODES` in `lib/providers/base.mjs` and `HARD_TRIGGER_CODES` in `lib/fallback/engine.mjs` both listed `QUOTA_EXHAUSTED` and `RATE_LIMITED` as live hard-trigger codes. No v0.1 plugin emits either code. The Anthropic, Codex, and Mistral plugins all use `claude -p`, `codex exec --json`, and `vibe --prompt` respectively — none parse the underlying-API HTTP response status code or surface a structured quota/rate error; they only throw `SPAWN_FAILED`, `SPAWN_TIMEOUT`, `CLI_NOT_FOUND`, or `AUTH_MISSING`. The two `Hard triggers` bullets in § Trigger taxonomy ("HTTP 5xx from provider's underlying API" and "HTTP 4xx quota exhaustion") are therefore unreachable through the `ProviderError` code path at v0.1.
@@ -15,7 +32,7 @@
   - `QUOTA_EXHAUSTED` and `RATE_LIMITED` removed from `PROVIDER_ERROR_CODES` (base.mjs) and `HARD_TRIGGER_CODES` (engine.mjs). Dead code removal.
   - A comment block added in `evaluateHardTriggers` labeling the HTTP-status branches as "forward-compat reserved — v0.1 plugins never attach statusCode."
   - Test coverage: the two unit tests for `evaluateHardTriggers: ProviderError QUOTA_EXHAUSTED/RATE_LIMITED → fires` are removed (tombstoned with a removal comment). All other hard-trigger tests that used these codes as convenient test vectors are rewritten to use `SPAWN_FAILED` / `SPAWN_TIMEOUT`.
-- **v0.1 live hard-trigger codes after this amendment:** `SPAWN_FAILED`, `CLI_NOT_FOUND`, `SPAWN_TIMEOUT`. `AUTH_MISSING` remains in the table as an explicit `false` entry (deliberate non-trigger per ADR 0004 § "No fallback for client-side errors" analogue).
+- **v0.1 live hard-trigger codes after this amendment:** `SPAWN_FAILED`, `CLI_NOT_FOUND`, `SPAWN_TIMEOUT`. `AUTH_MISSING` remains in the table as an explicit `false` entry (deliberate non-trigger per ADR 0004 § "No fallback for client-side errors" analogue). **Subsequently extended by Amendment 4 (D38) — `CONCURRENCY_LIMIT` added as a 4th true entry; the v0.1 live-codes list as of D38 is `SPAWN_FAILED`, `CLI_NOT_FOUND`, `SPAWN_TIMEOUT`, `CONCURRENCY_LIMIT` plus the explicit `AUTH_MISSING:false` non-trigger entry.**
 - **v1.x re-activation path:** When a plugin gains HTTP-status parsing (e.g., an Anthropic plugin variant that makes direct Messages API calls rather than spawning `claude -p`), add the plugin-layer HTTP parsing, re-add `QUOTA_EXHAUSTED` and `RATE_LIMITED` to both tables, and amend this entry. The `evaluateHardTriggers` HTTP-status branches will then activate naturally with no further engine changes.
 - **Procedural mechanism:** CC 开发铁律 v1.6 § 10.x (Round-6 Cold Audit).
 
diff --git a/lib/fallback/engine.mjs b/lib/fallback/engine.mjs
index ad0cf3d..e68c20c 100644
--- a/lib/fallback/engine.mjs
+++ b/lib/fallback/engine.mjs
@@ -27,11 +27,16 @@ import { computeIRRequestHash } from '../cache/keys.mjs';
 /**
  * Maps ProviderError codes to hard-trigger decisions.
  *
- * v0.1 live codes (per ADR 0004 Amendment 3, D34 F7):
- *   - SPAWN_FAILED    → hard trigger (provider CLI failed)
- *   - CLI_NOT_FOUND   → hard trigger (binary missing)
- *   - AUTH_MISSING    → NOT a hard trigger (user-config failure; user must fix)
- *   - SPAWN_TIMEOUT   → hard trigger (per ADR 0004 § Trigger taxonomy bullet 4)
+ * v0.1 live codes (per ADR 0004 Amendment 3 (D34 F7) + Amendment 4 (D38)):
+ *   - SPAWN_FAILED       → hard trigger (provider CLI failed)
+ *   - CLI_NOT_FOUND      → hard trigger (binary missing)
+ *   - AUTH_MISSING       → NOT a hard trigger (user-config failure; user must fix)
+ *   - SPAWN_TIMEOUT      → hard trigger (per ADR 0004 § Trigger taxonomy bullet 4)
+ *   - CONCURRENCY_LIMIT  → hard trigger (D38 / issue #1, ADR 0004 Amendment 4):
+ *                          synthesized by server.mjs when a provider is at its
+ *                          hints.maxConcurrent in-flight limit. The chain
+ *                          advances immediately to the next hop instead of
+ *                          queueing — design rationale per ADR 0004 Amendment 4.
  *
  * QUOTA_EXHAUSTED and RATE_LIMITED removed (D34 F7 / ADR 0004 Amendment 3):
  *   no v0.1 plugin parses underlying-API HTTP status codes, so these codes
@@ -44,8 +49,9 @@ import { computeIRRequestHash } from '../cache/keys.mjs';
 const HARD_TRIGGER_CODES = {
   SPAWN_FAILED: true,
   CLI_NOT_FOUND: true,
-  AUTH_MISSING: false,   // user config problem — never fall over (ADR 0004 § Decision)
-  SPAWN_TIMEOUT: true,   // ADR 0004 § Trigger taxonomy bullet 4: spawn timeout is a hard trigger
+  AUTH_MISSING: false,      // user config problem — never fall over (ADR 0004 § Decision)
+  SPAWN_TIMEOUT: true,      // ADR 0004 § Trigger taxonomy bullet 4: spawn timeout is a hard trigger
+  CONCURRENCY_LIMIT: true,  // ADR 0004 Amendment 4 (D38, issue #1): saturation → advance chain
 };
 
 /**
diff --git a/lib/providers/base.mjs b/lib/providers/base.mjs
index a33d4ae..bccbed6 100644
--- a/lib/providers/base.mjs
+++ b/lib/providers/base.mjs
@@ -141,8 +141,15 @@ export function validateProvider(p) {
 /**
  * Error codes surfaced by provider plugins.
  *
- * v0.1 live codes (per ADR 0004 Amendment 3, D34 F7):
- *   SPAWN_FAILED, CLI_NOT_FOUND, AUTH_MISSING, SPAWN_TIMEOUT
+ * v0.1 live codes (per ADR 0004 Amendment 3, D34 F7 and Amendment 4, D38):
+ *   SPAWN_FAILED, CLI_NOT_FOUND, AUTH_MISSING, SPAWN_TIMEOUT, CONCURRENCY_LIMIT
+ *
+ * CONCURRENCY_LIMIT (D38, issue #1): synthesized by the orchestration layer
+ * (NOT thrown by provider plugins themselves) when a spawn is attempted
+ * against a provider already at its hints.maxConcurrent limit. The fallback
+ * engine treats this as a hard trigger so the chain advances to the next hop
+ * rather than queueing. See ADR 0002 Amendment 6 (runtime enforcement) and
+ * ADR 0004 Amendment 4 (CONCURRENCY_LIMIT in hard-trigger taxonomy).
  *
  * QUOTA_EXHAUSTED and RATE_LIMITED were removed (D34 F7): no plugin parses
  * underlying-API HTTP status codes at v0.1, so these codes are never emitted.
@@ -153,7 +160,8 @@ export const PROVIDER_ERROR_CODES = /** @type {const} */ ([
   'AUTH_MISSING',
   'CLI_NOT_FOUND',
   'SPAWN_FAILED',
-  'SPAWN_TIMEOUT',  // ADR 0004 § Trigger taxonomy bullet 4: spawn timeout is a hard trigger
+  'SPAWN_TIMEOUT',     // ADR 0004 § Trigger taxonomy bullet 4: spawn timeout is a hard trigger
+  'CONCURRENCY_LIMIT', // ADR 0002 Amendment 6 / ADR 0004 Amendment 4 (D38, issue #1)
 ]);
 
 export class ProviderError extends Error {
diff --git a/lib/providers/index.mjs b/lib/providers/index.mjs
index d831216..3ae8196 100644
--- a/lib/providers/index.mjs
+++ b/lib/providers/index.mjs
@@ -209,3 +209,133 @@ export function getProviderByName(loadedProviders, name) {
 export function listAllProviderNames() {
   return STATIC_REGISTRY.map(p => p.name);
 }
+
+// ── Concurrency semaphore (D38, issue #1) ─────────────────────────────────
+//
+// Authority: ADR 0002 Amendment 6 (maxConcurrent runtime enforcement landed in D38)
+//           and ADR 0004 Amendment 4 (CONCURRENCY_LIMIT added to hard-trigger taxonomy).
+//
+// Per-provider in-flight spawn counter. The orchestration layer (server.mjs
+// handleChatCompletions) calls tryAcquireSpawn() before provider.spawn() and
+// releaseSpawn() after spawn lifecycle completion. On saturation, the caller
+// synthesises a ProviderError(CONCURRENCY_LIMIT) which the fallback engine
+// treats as a hard trigger — the chain advances to the next hop. If the
+// entire chain is saturated, the user receives a chain-exhausted error via
+// the existing executeWithFallback exhaustion path.
+//
+// Design decision (deliberate): immediate-advancement via fallback, NOT
+// queue+timeout. Rationale (per D38 issue #1 design discussion):
+//   1. The fallback chain exists precisely for this kind of overflow.
+//   2. A queue introduces head-of-line blocking + a new timeout config surface.
+//   3. Immediate-advancement gives fail-fast latency, matching the OLP
+//      multi-provider proxy philosophy.
+//   4. Queue+timeout is deferred — track via a future issue if real usage
+//      shows need.
+//
+// **Atomicity invariant**: JavaScript is single-threaded; the
+// read-then-write pair inside tryAcquireSpawn() executes synchronously with
+// NO `await` between the check and the increment. This is the only reason
+// the semaphore is correct without a Mutex. A future async refactor MUST
+// preserve this — do NOT introduce an `await` between the limit check and
+// the count update or the semaphore loses its mutual-exclusion guarantee
+// (two callers could each read count=limit-1 before either increments).
+//
+// Module-level state: lives for the process lifetime; tests that need
+// isolation should call __resetSpawnCounters() in their teardown.
+//
+// @type {Map<string, number>} provider name → current in-flight spawn count
+const _activeSpawns = new Map();
+
+/**
+ * Default cap for tryAcquireSpawn when a plugin omits hints.maxConcurrent.
+ *
+ * validateProvider in base.mjs requires hints.maxConcurrent to be a
+ * non-negative integer at startup, so a missing value should not happen in
+ * production. This default is defense-in-depth for callers that pass a
+ * stripped-down provider stub (e.g., in tests) or future plugin paths that
+ * bypass validation. The value (4) matches the v0.1 plugin defaults
+ * (anthropic / codex / mistral all declare hints.maxConcurrent: 4).
+ */
+export const DEFAULT_MAX_CONCURRENT_SPAWNS = 4;
+
+/**
+ * Atomically attempts to reserve a spawn slot for `providerName`.
+ *
+ * If the current in-flight count is below `maxConcurrent`, increments the
+ * counter and returns true. Otherwise returns false WITHOUT incrementing —
+ * the caller is responsible for surfacing the saturation as a
+ * ProviderError(CONCURRENCY_LIMIT) for the fallback engine to consume.
+ *
+ * Atomicity: the check and the increment happen in a single synchronous
+ * block with no `await` in between. See the module-level invariant comment
+ * above for why this is sufficient.
+ *
+ * @param {string} providerName — provider key (e.g. 'anthropic')
+ * @param {number} [maxConcurrent=DEFAULT_MAX_CONCURRENT_SPAWNS] — limit from hints.maxConcurrent
+ * @returns {boolean} true if a slot was acquired, false if at limit
+ */
+export function tryAcquireSpawn(providerName, maxConcurrent = DEFAULT_MAX_CONCURRENT_SPAWNS) {
+  // Defensive: coerce undefined/null/non-integer to the default. validateProvider
+  // already enforces this at startup; this guards future plugin paths that
+  // bypass validation.
+  const limit = (typeof maxConcurrent === 'number' && Number.isInteger(maxConcurrent) && maxConcurrent >= 0)
+    ? maxConcurrent
+    : DEFAULT_MAX_CONCURRENT_SPAWNS;
+
+  const current = _activeSpawns.get(providerName) ?? 0;
+  // Atomic check-then-increment (no `await` between read and write).
+  if (current >= limit) {
+    return false;
+  }
+  _activeSpawns.set(providerName, current + 1);
+  return true;
+}
+
+/**
+ * Releases a spawn slot for `providerName`. Must be called exactly once per
+ * successful tryAcquireSpawn() call, regardless of whether the spawn succeeded
+ * or threw. The caller in server.mjs uses a try/finally pattern to guarantee
+ * the release fires on every exit path (success, error, abort, streaming end).
+ *
+ * Throws if the count would go negative — that indicates a bug (a release
+ * without a matching acquire, or a double-release). The throw is loud on
+ * purpose so the bug surfaces in tests rather than silently corrupting the
+ * counter for future requests.
+ *
+ * @param {string} providerName — provider key (e.g. 'anthropic')
+ * @throws {Error} if no slot is currently held for providerName
+ */
+export function releaseSpawn(providerName) {
+  const current = _activeSpawns.get(providerName) ?? 0;
+  if (current <= 0) {
+    throw new Error(
+      `releaseSpawn(${providerName}): counter would go negative — release without matching acquire (or double-release)`,
+    );
+  }
+  const next = current - 1;
+  if (next === 0) {
+    _activeSpawns.delete(providerName);
+  } else {
+    _activeSpawns.set(providerName, next);
+  }
+}
+
+/**
+ * Returns the current in-flight spawn count for `providerName`. Used by
+ * /health, diagnostics, and tests that need to assert peak concurrency.
+ *
+ * @param {string} providerName
+ * @returns {number} non-negative integer; 0 if no spawns in flight
+ */
+export function getActiveSpawnCount(providerName) {
+  return _activeSpawns.get(providerName) ?? 0;
+}
+
+/**
+ * @internal — test seam: reset all in-flight spawn counters to zero. Used by
+ * test teardown to ensure a clean state across suites. Production code MUST
+ * NOT call this — it bypasses the acquire/release pairing invariant.
+ */
+export function __resetSpawnCounters() {
+  _activeSpawns.clear();
+}
diff --git a/server.mjs b/server.mjs
index 51bbf65..e55ef13 100644
--- a/server.mjs
+++ b/server.mjs
@@ -29,7 +29,16 @@ import {
   generateRequestId,
   SSE_DONE,
 } from './lib/ir/ir-to-openai.mjs';
-import { loadProviders, listAllProviderNames, getAliasMap, getModelCreated } from './lib/providers/index.mjs';
+import {
+  loadProviders,
+  listAllProviderNames,
+  getAliasMap,
+  getModelCreated,
+  tryAcquireSpawn,
+  releaseSpawn,
+  getActiveSpawnCount,
+  DEFAULT_MAX_CONCURRENT_SPAWNS,
+} from './lib/providers/index.mjs';
 import { ProviderError } from './lib/providers/base.mjs';
 import { computeCacheKey, hasCacheControl, extractCacheControlMarkers } from './lib/cache/keys.mjs';
 import { CacheStore } from './lib/cache/store.mjs';
@@ -503,43 +512,80 @@ async function handleChatCompletions(req, res) {
     // conditions item 1 requires no truncation). A non-enumerable __truncated marker
     // is set on the returned array so executeHopFn can evict the cache entry below.
     async function collectAllChunks() {
+      // D38 (issue #1): maxConcurrent runtime enforcement.
+      // Authority: ADR 0002 Amendment 6 + ADR 0004 Amendment 4.
+      //
+      // Try to acquire a spawn slot for this provider BEFORE invoking spawn().
+      // If at limit, synthesise ProviderError(CONCURRENCY_LIMIT) which the
+      // fallback engine treats as a hard trigger — the chain advances to the
+      // next hop. validateProvider guarantees hints.maxConcurrent is present;
+      // DEFAULT_MAX_CONCURRENT_SPAWNS is a defense-in-depth fallback.
+      const maxConcurrent = hopProviderPlugin.hints?.maxConcurrent ?? DEFAULT_MAX_CONCURRENT_SPAWNS;
+      if (!tryAcquireSpawn(hopProvider, maxConcurrent)) {
+        const concurrencyErr = new ProviderError(
+          `provider ${hopProvider} at maxConcurrent (${maxConcurrent}) — advancing to next hop`,
+          'CONCURRENCY_LIMIT',
+        );
+        concurrencyErr.providerName = hopProvider;
+        concurrencyErr.maxConcurrent = maxConcurrent;
+        // activeSpawns reflects the live counter at the rejection moment,
+        // queried directly. Since tryAcquireSpawn returned false the value
+        // equals maxConcurrent — read it explicitly for diagnostic clarity
+        // rather than echoing the limit (avoids future-reader confusion).
+        concurrencyErr.activeSpawns = getActiveSpawnCount(hopProvider);
+        throw concurrencyErr;
+      }
+
       const chunks = [];
+      // try/finally: releaseSpawn MUST fire on every exit path — success
+      // (return at end), spawn throw (caught and re-thrown below), or the
+      // D16 truncation-salvage return. The finally is the only mechanism
+      // that guarantees release across all three.
       try {
-        for await (const irChunk of hopProviderPlugin.spawn(irReq, authContext)) {
-          // D16: check error chunks BEFORE pushing — preserves the invariant that
-          // chunks array contains only delta/stop chunks. Without this, the catch
-          // block's `chunks.length > 0` would mistake a single error chunk for
-          // "usable content streamed" (Case B) and synthesize a stop + return,
-          // sending an empty body to the client when the correct behavior is to
-          // re-throw and let the fallback engine advance the chain.
-          if (irChunk.type === 'error') {
-            throw new ProviderError(
-              irChunk.error ?? 'Provider emitted error chunk',
-              'SPAWN_FAILED',
-            );
+        try {
+          for await (const irChunk of hopProviderPlugin.spawn(irReq, authContext)) {
+            // D16: check error chunks BEFORE pushing — preserves the invariant that
+            // chunks array contains only delta/stop chunks. Without this, the catch
+            // block's `chunks.length > 0` would mistake a single error chunk for
+            // "usable content streamed" (Case B) and synthesize a stop + return,
+            // sending an empty body to the client when the correct behavior is to
+            // re-throw and let the fallback engine advance the chain.
+            if (irChunk.type === 'error') {
+              throw new ProviderError(
+                irChunk.error ?? 'Provider emitted error chunk',
+                'SPAWN_FAILED',
+              );
+            }
+            chunks.push(irChunk);
+            if (irChunk.type === 'stop') break;
           }
-          chunks.push(irChunk);
-          if (irChunk.type === 'stop') break;
-        }
-      } catch (spawnErr) {
-        if (spawnErr instanceof ProviderError && spawnErr.code === 'SPAWN_FAILED' && chunks.length > 0) {
-          // Case B (ADR 0004 Amendment 1): provider emitted usable chunks then exited
-          // non-zero. Synthesize a truncated stop and surface the partial response.
-          chunks.push({ type: 'stop', finish_reason: 'length' });
-          logEvent('warn', 'spawn_failed_after_usable_chunks', {
-            chunks_count: chunks.length - 1, // exclude the synthesized stop
-            provider: hopProvider,
-            model: hopModel,
-          });
-          // Mark as truncated so the caller can evict this entry from cache.
-          Object.defineProperty(chunks, '__truncated', { value: true, enumerable: false });
-          return chunks;
+        } catch (spawnErr) {
+          if (spawnErr instanceof ProviderError && spawnErr.code === 'SPAWN_FAILED' && chunks.length > 0) {
+            // Case B (ADR 0004 Amendment 1): provider emitted usable chunks then exited
+            // non-zero. Synthesize a truncated stop and surface the partial response.
+            chunks.push({ type: 'stop', finish_reason: 'length' });
+            logEvent('warn', 'spawn_failed_after_usable_chunks', {
+              chunks_count: chunks.length - 1, // exclude the synthesized stop
+              provider: hopProvider,
+              model: hopModel,
+            });
+            // Mark as truncated so the caller can evict this entry from cache.
+            Object.defineProperty(chunks, '__truncated', { value: true, enumerable: false });
+            return chunks;
+          }
+          // Case A (SPAWN_FAILED with no chunks) or any other error: re-throw.
+          // Fallback engine fires hard trigger and advances chain as before.
+          throw spawnErr;
         }
-        // Case A (SPAWN_FAILED with no chunks) or any other error: re-throw.
-        // Fallback engine fires hard trigger and advances chain as before.
-        throw spawnErr;
+        return chunks;
+      } finally {
+        // D38: spawn lifecycle ended (drain completed, stop chunk received,
+        // SPAWN_FAILED salvage returned, or any unexpected throw). Release
+        // the slot so the next caller can acquire it. Single-threaded JS
+        // guarantees no other caller has incremented this provider's count
+        // between our tryAcquireSpawn() above and this releaseSpawn().
+        releaseSpawn(hopProvider);
       }
-      return chunks;
     }
 
     // D23: cacheable opt-out check (ADR 0002 Amendment 3 + ADR 0005 Amendment 3).
@@ -626,7 +672,24 @@ async function handleChatCompletions(req, res) {
   //
   // On success: write chunks to res AND cache so subsequent identical requests
   // hit the burst-replay path.
+  // D38 (issue #1): pre-acquire a concurrency slot for the streaming path so
+  // saturation here behaves identically to saturation on the buffered path.
+  // If acquire fails, fall through (skip this branch) — the buffered path
+  // below also gates via tryAcquireSpawn and will surface a chain-exhausted
+  // error through executeWithFallback (correct behaviour: a single-hop chain
+  // at maxConcurrent has no other hop to advance to).
+  //
+  // Acquired here, released in the `finally` below. The gate intentionally
+  // lives BEFORE the streaming-branch entry check so the buffered fallthrough
+  // can re-attempt acquire from a clean slate.
+  let streamingAcquired = false;
   if (ir.stream && chain.length === 1 && !bypassCacheForFirstHop && !preCheckHit && cacheableForFirstHop) {
+    const candidatePlugin = loadedProviders.get(chain[0].provider);
+    const candidateMax = candidatePlugin?.hints?.maxConcurrent ?? DEFAULT_MAX_CONCURRENT_SPAWNS;
+    streamingAcquired = candidatePlugin ? tryAcquireSpawn(chain[0].provider, candidateMax) : false;
+  }
+
+  if (ir.stream && chain.length === 1 && !bypassCacheForFirstHop && !preCheckHit && cacheableForFirstHop && streamingAcquired) {
     const streamProvider = chain[0].provider;
     const streamModel = chain[0].model;
     const streamCacheKey = computeCacheKey(streamProvider, streamModel, ir);
@@ -634,6 +697,8 @@ async function handleChatCompletions(req, res) {
 
     if (!streamPlugin) {
       // Provider disappeared between chain build and here (edge case).
+      // Release the slot we acquired above so the counter stays balanced.
+      releaseSpawn(streamProvider);
       return sendError(res, 503, `Provider ${streamProvider} is not enabled`, 'no_enabled_provider',
         olpErrorHeaders({ startMs, model: ir.model }));
     }
@@ -783,6 +848,13 @@ async function handleChatCompletions(req, res) {
           res.end();
         }
       }
+    } finally {
+      // D38 (issue #1): streaming spawn lifecycle ended — drain completed,
+      // stop chunk seen, generator exhausted without stop, or any catch
+      // path returned via res.end(). Release the slot acquired before
+      // entering the streaming branch. The finally fires on every JS exit
+      // path including the `return;` statements inside the try/catch body.
+      releaseSpawn(streamProvider);
     }
     return;
   }
diff --git a/test-features.mjs b/test-features.mjs
index b1d2fa5..b26da5e 100644
--- a/test-features.mjs
+++ b/test-features.mjs
@@ -28,7 +28,20 @@ import {
   SSE_DONE,
 } from './lib/ir/ir-to-openai.mjs';
 import { validateProvider, ProviderError, withTimeout } from './lib/providers/base.mjs';
-import { loadProviders, getProviderForModel, getProviderByName, listAllProviderNames, getAliasMap, getModelCreated, REGISTRY_BOOTSTRAP_CREATED } from './lib/providers/index.mjs';
+import {
+  loadProviders,
+  getProviderForModel,
+  getProviderByName,
+  listAllProviderNames,
+  getAliasMap,
+  getModelCreated,
+  REGISTRY_BOOTSTRAP_CREATED,
+  tryAcquireSpawn,
+  releaseSpawn,
+  getActiveSpawnCount,
+  __resetSpawnCounters,
+  DEFAULT_MAX_CONCURRENT_SPAWNS,
+} from './lib/providers/index.mjs';
 import anthropic, {
   irToAnthropic,
   anthropicChunkToIR,
@@ -8347,3 +8360,481 @@ describe('D33 round-5 cold-audit cleanup', () => {
 });
 
 });
+
+// ── Suite 18: D38 — maxConcurrent runtime enforcement (issue #1) ──────────
+//
+// Authority:
+//   - ADR 0002 Amendment 6 (D38): maxConcurrent runtime enforcement landed.
+//   - ADR 0004 Amendment 4 (D38): CONCURRENCY_LIMIT added to hard-trigger taxonomy.
+//   - GitHub issue #1 (maxConcurrent: declarative-only at v0.1).
+//
+// Tests:
+//   18a: CONCURRENCY_LIMIT is in PROVIDER_ERROR_CODES.
+//   18b: CONCURRENCY_LIMIT is a hard trigger (evaluateHardTriggers true).
+//   18c: classifyTrigger returns 'hard' for CONCURRENCY_LIMIT.
+//   18d: tryAcquireSpawn / releaseSpawn / getActiveSpawnCount unit behaviour.
+//   18e: tryAcquireSpawn returns false at the limit.
+//   18f: releaseSpawn throws when called without a matching acquire.
+//   18g: DEFAULT_MAX_CONCURRENT_SPAWNS applies when maxConcurrent omitted.
+//   18h: __resetSpawnCounters clears all in-flight counts.
+//   18i: HTTP — 5 concurrent buffered requests against a maxConcurrent:2 mock →
+//        peak in-flight == 2; requests 3-5 hit chain-exhausted error.
+//   18j: HTTP — counter releases after request completes (single-hop, sequential).
+//   18k: Fallback — saturated primary (maxConcurrent:1) → secondary serves the 2nd.
+//   18l: HTTP streaming — counter releases at end of stream, not at start.
+
+describe('D38 — maxConcurrent runtime enforcement (Suite 18)', () => {
+
+  // ── 18a: PROVIDER_ERROR_CODES contains CONCURRENCY_LIMIT ──────────────────
+
+  it('18a: CONCURRENCY_LIMIT is in PROVIDER_ERROR_CODES', () => {
+    assert.ok(
+      PROVIDER_ERROR_CODES.includes('CONCURRENCY_LIMIT'),
+      'CONCURRENCY_LIMIT must be in PROVIDER_ERROR_CODES per ADR 0004 Amendment 4',
+    );
+  });
+
+  // ── 18b: CONCURRENCY_LIMIT triggers hard fallback ─────────────────────────
+
+  it('18b: evaluateHardTriggers(ProviderError CONCURRENCY_LIMIT) → true', () => {
+    const err = new ProviderError('provider stub at maxConcurrent (2)', 'CONCURRENCY_LIMIT');
+    assert.equal(
+      evaluateHardTriggers(err), true,
+      'CONCURRENCY_LIMIT must be classified as a hard trigger',
+    );
+  });
+
+  // ── 18c: AUTH_MISSING regression guard ───────────────────────────────────
+  // CONCURRENCY_LIMIT hard-trigger classification is covered by 18b;
+  // 18c guards that D38 did NOT accidentally flip AUTH_MISSING (a soft signal
+  // for v0.1 — see ADR 0004 Amendment 2) to hard alongside CONCURRENCY_LIMIT.
+
+  it('18c: evaluateHardTriggers still false for AUTH_MISSING (regression guard)', () => {
+    // Regression guard: D38 didn't accidentally flip AUTH_MISSING to a hard trigger.
+    const err = new ProviderError('auth missing', 'AUTH_MISSING');
+    assert.equal(evaluateHardTriggers(err), false);
+  });
+
+  // ── 18d: semaphore unit tests ─────────────────────────────────────────────
+
+  describe('18d — tryAcquireSpawn / releaseSpawn / getActiveSpawnCount unit', () => {
+
+    // Use a unique provider name to avoid colliding with other tests' state.
+    // __resetSpawnCounters() runs after each unit test to keep them isolated.
+
+    it('18d.1: tryAcquireSpawn increments count from 0 → 1', () => {
+      __resetSpawnCounters();
+      assert.equal(getActiveSpawnCount('unit-d1'), 0);
+      const ok = tryAcquireSpawn('unit-d1', 4);
+      assert.equal(ok, true, 'first acquire should succeed');
+      assert.equal(getActiveSpawnCount('unit-d1'), 1);
+      __resetSpawnCounters();
+    });
+
+    it('18d.2: releaseSpawn decrements count back to 0', () => {
+      __resetSpawnCounters();
+      tryAcquireSpawn('unit-d2', 4);
+      assert.equal(getActiveSpawnCount('unit-d2'), 1);
+      releaseSpawn('unit-d2');
+      assert.equal(getActiveSpawnCount('unit-d2'), 0);
+      __resetSpawnCounters();
+    });
+
+    it('18d.3: each provider has independent counters', () => {
+      __resetSpawnCounters();
+      tryAcquireSpawn('unit-d3-a', 4);
+      tryAcquireSpawn('unit-d3-a', 4);
+      tryAcquireSpawn('unit-d3-b', 4);
+      assert.equal(getActiveSpawnCount('unit-d3-a'), 2);
+      assert.equal(getActiveSpawnCount('unit-d3-b'), 1);
+      __resetSpawnCounters();
+    });
+
+    it('18d.4: counter is removed from map when it reaches 0', () => {
+      __resetSpawnCounters();
+      tryAcquireSpawn('unit-d4', 4);
+      releaseSpawn('unit-d4');
+      // After release-to-zero, getActiveSpawnCount still returns 0 (Map.delete'd).
+      assert.equal(getActiveSpawnCount('unit-d4'), 0);
+      __resetSpawnCounters();
+    });
+  });
+
+  // ── 18e: tryAcquireSpawn returns false at the limit ───────────────────────
+
+  it('18e: tryAcquireSpawn returns false when count == maxConcurrent', () => {
+    __resetSpawnCounters();
+    assert.equal(tryAcquireSpawn('unit-e', 2), true);
+    assert.equal(tryAcquireSpawn('unit-e', 2), true);
+    // At the limit now — third acquire must fail without incrementing.
+    assert.equal(tryAcquireSpawn('unit-e', 2), false,
+      'acquire must fail when current count equals limit');
+    // Counter must not have incremented on the failed acquire.
+    assert.equal(getActiveSpawnCount('unit-e'), 2,
+      'failed acquire must not increment the counter');
+    __resetSpawnCounters();
+  });
+
+  // ── 18f: releaseSpawn without acquire throws ──────────────────────────────
+
+  it('18f: releaseSpawn without matching acquire throws', () => {
+    __resetSpawnCounters();
+    assert.throws(
+      () => releaseSpawn('unit-f-never-acquired'),
+      /releaseSpawn.*counter would go negative/,
+      'release without acquire must throw to surface the bug loudly',
+    );
+    __resetSpawnCounters();
+  });
+
+  // ── 18g: DEFAULT_MAX_CONCURRENT_SPAWNS fallback ───────────────────────────
+
+  it('18g: DEFAULT_MAX_CONCURRENT_SPAWNS applies when maxConcurrent argument omitted', () => {
+    __resetSpawnCounters();
+    assert.ok(Number.isInteger(DEFAULT_MAX_CONCURRENT_SPAWNS) && DEFAULT_MAX_CONCURRENT_SPAWNS > 0,
+      'DEFAULT_MAX_CONCURRENT_SPAWNS must be a positive integer');
+    // Acquire up to the default limit without arg — all should succeed.
+    for (let i = 0; i < DEFAULT_MAX_CONCURRENT_SPAWNS; i++) {
+      assert.equal(tryAcquireSpawn('unit-g'), true, `acquire #${i + 1} should succeed`);
+    }
+    // Next acquire (still no arg) should fail at the default boundary.
+    assert.equal(tryAcquireSpawn('unit-g'), false,
+      `acquire #${DEFAULT_MAX_CONCURRENT_SPAWNS + 1} should fail at default limit`);
+    __resetSpawnCounters();
+  });
+
+  it('18g.2: non-integer maxConcurrent coerces to DEFAULT_MAX_CONCURRENT_SPAWNS', () => {
+    __resetSpawnCounters();
+    // Defensive coercion: a plugin path that bypasses validateProvider could
+    // produce undefined/null/NaN. The gate must not blow up; it must use the default.
+    assert.equal(tryAcquireSpawn('unit-g2', undefined), true);
+    assert.equal(tryAcquireSpawn('unit-g2', null), true);
+    assert.equal(tryAcquireSpawn('unit-g2', NaN), true);
+    assert.equal(tryAcquireSpawn('unit-g2', 'four'), true);
+    // After 4 acquires (matching default), 5th should fail.
+    assert.equal(getActiveSpawnCount('unit-g2'), 4);
+    assert.equal(tryAcquireSpawn('unit-g2', undefined), false);
+    __resetSpawnCounters();
+  });
+
+  // ── 18h: __resetSpawnCounters clears state ────────────────────────────────
+
+  it('18h: __resetSpawnCounters clears all in-flight counts', () => {
+    tryAcquireSpawn('unit-h-1', 4);
+    tryAcquireSpawn('unit-h-2', 4);
+    assert.equal(getActiveSpawnCount('unit-h-1'), 1);
+    assert.equal(getActiveSpawnCount('unit-h-2'), 1);
+    __resetSpawnCounters();
+    assert.equal(getActiveSpawnCount('unit-h-1'), 0);
+    assert.equal(getActiveSpawnCount('unit-h-2'), 0);
+  });
+
+  // ────────────────────────────────────────────────────────────────────────────
+  // HTTP integration tests
+  // ────────────────────────────────────────────────────────────────────────────
+  //
+  // Strategy: install a controllable mock spawn on the anthropic plugin that
+  // exposes a Promise-based "release gate" — the mock yields no data until
+  // the test explicitly resolves the gate. This lets us hold multiple spawns
+  // in-flight at the same time deterministically (no timing races).
+  //
+  // For the maxConcurrent:N tests we override anthropic.hints.maxConcurrent
+  // for the duration of the test, then restore. ProviderError synthesised by
+  // the gate carries CONCURRENCY_LIMIT and the fallback engine treats it as
+  // a hard trigger. With a single-hop chain the chain exhausts and the
+  // client receives a 502/non-200.
+
+  describe('18 HTTP integration', () => {
+    let server18;
+    let port18;
+    let savedToken18;
+    let savedMaxConcurrent;
+
+    before(async () => {
+      savedToken18 = process.env.CLAUDE_CODE_OAUTH_TOKEN;
+      process.env.CLAUDE_CODE_OAUTH_TOKEN = 'fake-token-suite-18';
+      __setProvidersEnabled({ anthropic: true });
+      __resetSpawnCounters();
+
+      const { createOlpServer: s18, __clearCache: cc18 } = await import('./server.mjs');
+      cc18();
+      server18 = s18();
+      await new Promise((resolve, reject) => {
+        server18.listen(0, '127.0.0.1', resolve);
+        server18.once('error', reject);
+      });
+      port18 = server18.address().port;
+    });
+
+    after(async () => {
+      __resetProvidersEnabled();
+      __resetSpawnImpl();
+      __resetSpawnCounters();
+      if (savedToken18 !== undefined) {
+        process.env.CLAUDE_CODE_OAUTH_TOKEN = savedToken18;
+      } else {
+        delete process.env.CLAUDE_CODE_OAUTH_TOKEN;
+      }
+      if (!server18) return;
+      return new Promise(r => server18.close(r));
+    });
+
+    /**
+     * Makes a controllable mock spawn that completes only when the given
+     * gate Promise resolves. Records peak in-flight via `peakRef`.
+     */
+    function makeGatedMockSpawn(gatePromise, content, peakRef, activeRef) {
+      return function gatedSpawn(_bin, _args, _opts) {
+        const proc = new EventEmitter();
+        proc.stdout = new EventEmitter();
+        proc.stderr = new EventEmitter();
+        proc.stdin = {
+          write: () => {},
+          end: async () => {
+            // Track this spawn as in-flight from the moment stdin closes.
+            activeRef.count++;
+            if (activeRef.count > peakRef.peak) peakRef.peak = activeRef.count;
+            // Wait for the gate to resolve before emitting any data.
+            await gatePromise;
+            // Decrement on exit.
+            setImmediate(() => {
+              proc.stdout.emit('data', Buffer.from(content));
+              proc.stdout.emit('end');
+              proc.stderr.emit('end');
+              proc.emit('close', 0, null);
+              activeRef.count--;
+            });
+          },
+        };
+        proc.killed = false;
+        proc.kill = () => {};
+        return proc;
+      };
+    }
+
+    // ── 18i: 5 concurrent requests, maxConcurrent:2 ────────────────────────
+
+    it('18i: 5 concurrent buffered requests with maxConcurrent:2 → peak in-flight == 2', async () => {
+      // Override anthropic's maxConcurrent for this test.
+      savedMaxConcurrent = anthropic.hints.maxConcurrent;
+      anthropic.hints.maxConcurrent = 2;
+
+      const { __clearCache: cc18i } = await import('./server.mjs');
+      cc18i();
+
+      // Gate that we'll resolve to let mock spawns complete.
+      let releaseGate;
+      const gatePromise = new Promise(resolve => { releaseGate = resolve; });
+      const peakRef = { peak: 0 };
+      const activeRef = { count: 0 };
+      __setSpawnImpl(makeGatedMockSpawn(gatePromise, 'concurrency-test-response', peakRef, activeRef));
+
+      try {
+        // Fire 5 concurrent buffered requests with DIFFERENT message content
+        // so cache lookups don't collapse them.
+        const reqs = [];
+        for (let i = 0; i < 5; i++) {
+          reqs.push(fetch({
+            port: port18,
+            method: 'POST',
+            path: '/v1/chat/completions',
+            body: {
+              model: 'claude-sonnet-4-6',
+              messages: [{ role: 'user', content: `concurrency-test-msg-${i}-${Date.now()}` }],
+            },
+          }));
+        }
+
+        // Give the server a moment to receive all requests and attempt acquires
+        // (the requests that succeed acquire will wait on gatePromise; the
+        // ones that fail acquire will return immediately with an error).
+        // The wait must be long enough for all 5 to either enter the spawn or
+        // be rejected by the gate, but short enough to keep the test fast.
+        await new Promise(r => setTimeout(r, 100));
+
+        // Peak in-flight count must equal exactly 2 (== maxConcurrent).
+        assert.equal(peakRef.peak, 2,
+          `peak in-flight must be exactly 2 (== maxConcurrent), observed ${peakRef.peak}`);
+        // Also assert via the exported getActiveSpawnCount helper (covers a
+        // different path — the module-level _activeSpawns map vs the mock-side
+        // counter). Both must agree.
+        assert.equal(getActiveSpawnCount('anthropic'), 2,
+          'module-level active-spawn counter must report 2 in-flight');
+
+        // Release the gate so the 2 in-flight spawns can complete.
+        releaseGate();
+
+        // Wait for all 5 requests to settle.
+        const results = await Promise.all(reqs);
+
+        // 2 successes (requests that acquired), 3 errors (CONCURRENCY_LIMIT).
+        const successes = results.filter(r => r.status === 200);
+        const failures = results.filter(r => r.status !== 200);
+        assert.equal(successes.length, 2,
+          `expected exactly 2 successful responses, got ${successes.length} (statuses: ${results.map(r => r.status).join(', ')})`);
+        assert.equal(failures.length, 3,
+          `expected exactly 3 failed responses (chain-exhausted on CONCURRENCY_LIMIT), got ${failures.length}`);
+      } finally {
+        // Restore original maxConcurrent.
+        anthropic.hints.maxConcurrent = savedMaxConcurrent;
+        __resetSpawnImpl();
+        __resetSpawnCounters();
+      }
+    });
+
+    // ── 18j: counter releases after request completes ─────────────────────
+
+    it('18j: counter releases to 0 after buffered request completes', async () => {
+      savedMaxConcurrent = anthropic.hints.maxConcurrent;
+      anthropic.hints.maxConcurrent = 2;
+      const { __clearCache: cc18j } = await import('./server.mjs');
+      cc18j();
+
+      // Plain mock that completes immediately.
+      __setSpawnImpl(makeMockSpawn(['release-test-response']));
+
+      try {
+        // Fire request 1 and wait for completion.
+        const r1 = await fetch({
+          port: port18,
+          method: 'POST',
+          path: '/v1/chat/completions',
+          body: {
+            model: 'claude-sonnet-4-6',
+            messages: [{ role: 'user', content: `release-test-msg-1-${Date.now()}` }],
+          },
+        });
+        assert.equal(r1.status, 200);
+        // Counter must be back to 0 after the request completes.
+        assert.equal(getActiveSpawnCount('anthropic'), 0,
+          'counter must release to 0 after request completes');
+
+        // Fire request 2 (different content for cache miss) — must also succeed,
+        // proving the slot is reusable.
+        const r2 = await fetch({
+          port: port18,
+          method: 'POST',
+          path: '/v1/chat/completions',
+          body: {
+            model: 'claude-sonnet-4-6',
+            messages: [{ role: 'user', content: `release-test-msg-2-${Date.now()}` }],
+          },
+        });
+        assert.equal(r2.status, 200);
+        assert.equal(getActiveSpawnCount('anthropic'), 0,
+          'counter must release after second request too');
+      } finally {
+        anthropic.hints.maxConcurrent = savedMaxConcurrent;
+        __resetSpawnImpl();
+        __resetSpawnCounters();
+      }
+    });
+
+    // ── 18l: streaming counter release ────────────────────────────────────
+
+    it('18l: streaming request — counter releases at end of stream', async () => {
+      savedMaxConcurrent = anthropic.hints.maxConcurrent;
+      anthropic.hints.maxConcurrent = 2;
+      const { __clearCache: cc18l } = await import('./server.mjs');
+      cc18l();
+
+      __setSpawnImpl(makeMockSpawn(['stream-release-content']));
+
+      try {
+        // Make a streaming request and consume the full stream.
+        const port = port18;
+        const body = await new Promise((resolve, reject) => {
+          const req = httpRequest({
+            hostname: '127.0.0.1',
+            port,
+            method: 'POST',
+            path: '/v1/chat/completions',
+            headers: { 'Content-Type': 'application/json' },
+          }, res => {
+            let data = '';
+            res.on('data', c => { data += c; });
+            res.on('end', () => resolve(data));
+            res.on('error', reject);
+          });
+          req.on('error', reject);
+          req.write(JSON.stringify({
+            model: 'claude-sonnet-4-6',
+            messages: [{ role: 'user', content: `stream-release-${Date.now()}` }],
+            stream: true,
+          }));
+          req.end();
+        });
+        assert.ok(body.includes('data:'), 'streaming body must contain SSE data lines');
+
+        // After the stream completes (res.on('end') resolved), the counter must
+        // be back to 0. Tests the streaming-branch finally{} fired.
+        assert.equal(getActiveSpawnCount('anthropic'), 0,
+          'counter must release after streaming response completes');
+      } finally {
+        anthropic.hints.maxConcurrent = savedMaxConcurrent;
+        __resetSpawnImpl();
+        __resetSpawnCounters();
+      }
+    });
+  });
+
+  // ── 18k: fallback test — saturated primary advances to secondary ──────────
+  //
+  // Uses a 2-hop chain executed via executeWithFallback (unit-level, not HTTP).
+  // Primary has maxConcurrent:1 and is already saturated by a synthetic
+  // pre-acquire; second request must advance to the fallback hop.
+
+  describe('18k — fallback advances on saturation', () => {
+    after(() => __resetSpawnCounters());
+
+    it('18k: saturated primary (maxConcurrent:1) → fallback hop serves the request', async () => {
+      // Pre-saturate the 'anthropic' counter (1/1) so the next acquire fails.
+      __resetSpawnCounters();
+      tryAcquireSpawn('anthropic', 1);
+      assert.equal(getActiveSpawnCount('anthropic'), 1);
+
+      // Build a 2-hop chain: anthropic (saturated) → openai (will serve).
+      // executeHopFn for this test calls tryAcquireSpawn itself, mirroring
+      // the production gate semantics. On failure: throws ProviderError
+      // CONCURRENCY_LIMIT; fallback engine advances to next hop.
+      const chain = [
+        { provider: 'anthropic', model: 'claude-sonnet-4-6' },
+        { provider: 'openai', model: 'gpt-5.5' },
+      ];
+      const ir = makeIR({ model: 'claude-sonnet-4-6', stream: false });
+
+      async function testHopFn(hopProvider, hopModel, _ir) {
+        const max = hopProvider === 'anthropic' ? 1 : 4;
+        if (!tryAcquireSpawn(hopProvider, max)) {
+          const err = new ProviderError(
+            `provider ${hopProvider} at maxConcurrent (${max})`,
+            'CONCURRENCY_LIMIT',
+          );
+          err.providerName = hopProvider;
+          err.maxConcurrent = max;
+          err.activeSpawns = max;
+          throw err;
+        }
+        try {
+          // Yield a single stop chunk so the hop "succeeds".
+          return [{ type: 'delta', content: `served-by-${hopProvider}` }, { type: 'stop', finish_reason: 'stop' }];
+        } finally {
+          releaseSpawn(hopProvider);
+        }
+      }
+
+      const result = await executeWithFallback(chain, ir, testHopFn, { logEvent: () => {} });
+      assert.ok(result.chunks !== null, 'fallback should have produced chunks');
+      assert.equal(result.fallbackHops, 1, 'must have advanced to second hop');
+      assert.equal(result.providerUsed, 'openai',
+        'secondary (openai) must serve when primary (anthropic) is saturated');
+      const content = result.chunks.filter(c => c.type === 'delta').map(c => c.content).join('');
+      assert.ok(content.includes('served-by-openai'), 'content must come from fallback hop');
+
+      // Cleanup: release the synthetic pre-saturation slot.
+      releaseSpawn('anthropic');
+      assert.equal(getActiveSpawnCount('anthropic'), 0);
+    });
+  });
+
+});