fix(engine): drain deliveries when agents are released

.agentworkforce/trajectories/compacted/compact_whqlsrrkhceb_2026-09-20.json

@@ -0,0 +1,52 @@
+{
+  "id": "compact_x03t9bwpqklx",
+  "version": 1,
+  "type": "compacted",
+  "compactedAt": "2026-09-20T17:28:10.496Z",
+  "sourceTrajectories": [
+    "traj_pcfd4ebl1pam"
+  ],
+  "dateRange": {
+    "start": "2026-09-20T17:13:27.676Z",
+    "end": "2026-09-20T17:27:21.837Z"
+  },
+  "summary": {
+    "totalDecisions": 2,
+    "totalEvents": 4,
+    "uniqueAgents": [
+      "default"
+    ]
+  },
+  "decisionGroups": [
+    {
+      "category": "other",
+      "decisions": [
+        {
+          "question": "Require atomic writes for every irreversible release and share node completion implementation",
+          "chosen": "Require atomic writes for every irreversible release and share node completion implementation",
+          "reasoning": "Sequential fallback can permanently settle deliveries before a later tombstone failure; legacy node completion also committed lifecycle writes separately even on capable adapters.",
+          "fromTrajectory": "traj_pcfd4ebl1pam"
+        }
+      ]
+    },
+    {
+      "category": "api",
+      "decisions": [
+        {
+          "question": "Address all three PR threads",
+          "chosen": "Address all three PR threads",
+          "reasoning": "Devin requires fail-closed local releases; CodeRabbit independently identifies sequential legacy node completion and requests naming both active delivery states in release notes.",
+          "fromTrajectory": "traj_pcfd4ebl1pam"
+        }
+      ]
+    }
+  ],
+  "keyLearnings": [],
+  "keyFindings": [
+    "Engine typecheck and lint passed under Node 22.23.2.",
+    "Focused release and capacity tests: 101 passed, including 21 new refusal/rollback regressions.",
+    "Full engine suite: 1176 tests across 100 files passed with --maxWorkers=2. Initial default-parallel run had one unrelated retention CLI timeout; no timeout or dependency changes were committed."
+  ],
+  "filesAffected": [],
+  "commits": []
+}

.agentworkforce/trajectories/compacted/compact_whqlsrrkhceb_2026-09-20.md

@@ -0,0 +1,23 @@
+# Trajectory Compaction: Sep 20, 2026 - Sep 20, 2026
+
+## Summary
+- Sessions: 1
+- Decisions: 2
+- Events: 4
+- Agents: default
+- Files: 0
+- Commits: 0
+
+## Other
+- Require atomic writes for every irreversible release and share node completion implementation -> Require atomic writes for every irreversible release and share node completion implementation (traj_pcfd4ebl1pam)
+
+## Api
+- Address all three PR threads -> Address all three PR threads (traj_pcfd4ebl1pam)
+
+## Key Learnings
+- None
+
+## Key Findings
+- Engine typecheck and lint passed under Node 22.23.2.
+- Focused release and capacity tests: 101 passed, including 21 new refusal/rollback regressions.
+- Full engine suite: 1176 tests across 100 files passed with `--maxWorkers=2`. Initial default-parallel run had one unrelated retention CLI timeout; no timeout or dependency changes were committed.

.agentworkforce/trajectories/completed/2026-09/traj_2t44mh839vla/summary.md

@@ -0,0 +1,32 @@
+# Trajectory: Drain workspace delivery capacity when agents are released
+
+> **Status:** ✅ Completed
+> **Confidence:** 93%
+> **Started:** September 20, 2026 at 09:53 AM
+> **Completed:** September 20, 2026 at 10:00 AM
+
+---
+
+## Summary
+
+Dead-lettered active delivery rows during irreversible agent release across direct, local, guarded node, and legacy node lifecycle paths; added release-path and expired-unswept workspace-cap regressions; full engine suite, typecheck, and lint pass.
+
+**Approach:** Standard approach
+
+---
+
+## Key Decisions
+
+### Fix capacity recovery in the Relaycast engine release lifecycle
+- **Chose:** Fix capacity recovery in the Relaycast engine release lifecycle
+- **Reasoning:** Release stops future fan-out but queued/delivered rows for the tombstoned recipient remain active and continue consuming the workspace cap until TTL. The engine owns both the lifecycle mutation and active-depth accounting, so it can atomically dead-letter those rows on irreversible delete_agent releases across every path.
+
+---
+
+## Chapters
+
+### 1. Work
+*Agent: default*
+
+- Fix capacity recovery in the Relaycast engine release lifecycle: Fix capacity recovery in the Relaycast engine release lifecycle
+- Release-time settlement now covers direct deletion, local reaping, guarded node completion, and legacy node completion. Full engine suite passed (1,154 tests), and a workspace-cap regression proves expired unswept rows are already excluded from admission.

.agentworkforce/trajectories/completed/2026-09/traj_2t44mh839vla/trajectory.json

@@ -0,0 +1,73 @@
+{
+  "id": "traj_2t44mh839vla",
+  "version": 1,
+  "task": {
+    "title": "Drain workspace delivery capacity when agents are released"
+  },
+  "status": "completed",
+  "startedAt": "2026-09-20T16:53:13.768Z",
+  "completedAt": "2026-09-20T17:00:33.145Z",
+  "agents": [
+    {
+      "name": "default",
+      "role": "lead",
+      "joinedAt": "2026-09-20T16:54:06.278Z"
+    }
+  ],
+  "chapters": [
+    {
+      "id": "chap_xydn9gw6mt5u",
+      "title": "Work",
+      "agentName": "default",
+      "startedAt": "2026-09-20T16:54:06.278Z",
+      "endedAt": "2026-09-20T17:00:33.145Z",
+      "events": [
+        {
+          "ts": 1789923246278,
+          "type": "decision",
+          "content": "Fix capacity recovery in the Relaycast engine release lifecycle: Fix capacity recovery in the Relaycast engine release lifecycle",
+          "raw": {
+            "question": "Fix capacity recovery in the Relaycast engine release lifecycle",
+            "chosen": "Fix capacity recovery in the Relaycast engine release lifecycle",
+            "alternatives": [],
+            "reasoning": "Release stops future fan-out but queued/delivered rows for the tombstoned recipient remain active and continue consuming the workspace cap until TTL. The engine owns both the lifecycle mutation and active-depth accounting, so it can atomically dead-letter those rows on irreversible delete_agent releases across every path."
+          },
+          "significance": "high"
+        },
+        {
+          "ts": 1789923595490,
+          "type": "reflection",
+          "content": "Release-time settlement now covers direct deletion, local reaping, guarded node completion, and legacy node completion. Full engine suite passed (1,154 tests), and a workspace-cap regression proves expired unswept rows are already excluded from admission.",
+          "raw": {
+            "focalPoints": [
+              "release atomicity",
+              "capacity accounting",
+              "regression coverage"
+            ],
+            "confidence": 0.9
+          },
+          "significance": "high",
+          "tags": [
+            "focal:release atomicity",
+            "focal:capacity accounting",
+            "focal:regression coverage",
+            "confidence:0.9"
+          ]
+        }
+      ]
+    }
+  ],
+  "retrospective": {
+    "summary": "Dead-lettered active delivery rows during irreversible agent release across direct, local, guarded node, and legacy node lifecycle paths; added release-path and expired-unswept workspace-cap regressions; full engine suite, typecheck, and lint pass.",
+    "approach": "Standard approach",
+    "confidence": 0.93
+  },
+  "commits": [],
+  "filesChanged": [],
+  "projectId": "AgentWorkforce/relaycast",
+  "tags": [],
+  "_trace": {
+    "startRef": "6414a59e619459fa64640937f4c4634c207191f2",
+    "endRef": "6414a59e619459fa64640937f4c4634c207191f2"
+  }
+}

CHANGELOG.md

@@ -16,7 +16,11 @@ This project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 Packages without a separate changelog are covered by the cross-package notes below.
 
-## [Unreleased]
+## [Unreleased - Patch]
+
+### Fixed
+
+- Permanently releasing an agent now dead-letters its active `queued` and `delivered` deliveries immediately, restoring workspace messaging capacity instead of waiting for delivery TTL expiry.
 
 ## [8.11.5] - 2026-09-20
 

README.md

@@ -701,7 +701,11 @@ live host when one exists. If the host is absent or offline, a normal release
 fails explicitly with `agent_host_unavailable` instead of creating an ownerless
 pending invocation. A `delete_agent` request can be completed locally in that
 case: Relaycast deactivates bindings, frees the live name, and removes its
-implicit direct node. Cleanup callers that retain the issued agent token can
+implicit direct node. Irreversible releases and agent deletion require a database
+transaction or atomic batch; adapters without either capability are refused
+before identity, membership, or queued deliveries change. Successful cleanup
+dead-letters the released agent's active deliveries in the same atomic write.
+Cleanup callers that retain the issued agent token can
 send its SHA-256 hash as `expected_token_hash`; Relaycast then rejects a stale
 release with `agent_release_generation_conflict` before dispatch or completion,
 so a same-name takeover is left untouched.

openapi.yaml

@@ -2584,7 +2584,10 @@ paths:
 
     delete:
       summary: Delete agent
-      description: Delete an agent from the workspace
+      description: >-
+        Tombstone an agent, remove memberships, and dead-letter its active deliveries
+        in one atomic write. Database adapters without transaction or atomic batch
+        support are refused before these changes.
       tags:
         - Agents
       security:
@@ -2926,7 +2929,10 @@ paths:
         release fails explicitly with `503 agent_host_unavailable`; it never
         creates an ownerless pending invocation. With `delete_agent`, the engine
         can reap the database record directly and returns a completed invocation,
-        deleting the agent and any implicit direct node.
+        tombstoning the agent and deleting any implicit direct node. Irreversible
+        release removes memberships and dead-letters active deliveries in the
+        same atomic write. Database adapters without transaction or atomic batch
+        support are refused before these changes.
       tags:
         - Agents
       security:

packages/engine/CHANGELOG.md

@@ -7,7 +7,11 @@ See the [root changelog](../../CHANGELOG.md) for cross-package release highlight
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [Unreleased]
+## [Unreleased - Patch]
+
+### Fixed
+
+- Irreversible agent release paths atomically dead-letter that recipient's active deliveries with `recipient agent released`, so tombstoned identities cannot hold workspace delivery-depth capacity until TTL. Release and deletion refuse adapters without atomic writes before changing identity, membership, or deliveries.
 
 ## [8.11.5] - 2026-09-20
 

packages/engine/src/__tests__/conformance/agentLifecycle.test.ts

@@ -694,6 +694,7 @@ describe('agent presence and release lifecycle', () => {
   it('reaps a hostless agent that has already spoken', async () => {
     const ws = await createWorkspace(stack.app, 'hostless-agent-delete-with-history');
     const target = await registerAgent(stack.app, ws.workspaceKey, 'talkative-agent');
+    const sender = await registerAgent(stack.app, ws.workspaceKey, 'hostless-release-sender');
     const nodeId = `node_direct_${target.agentId}`;
 
     // Every agent worth reaping has history. Four FKs reference agents.id
@@ -710,6 +711,15 @@ describe('agent presence and release lifecycle', () => {
       body: JSON.stringify({ text: 'i have said something' }),
     });
     expect(posted.status).toBe(201);
+    const queued = await stack.app.request('/v1/channels/general/messages', {
+      method: 'POST',
+      headers: {
+        'content-type': 'application/json',
+        authorization: `Bearer ${sender.token}`,
+      },
+      body: JSON.stringify({ text: 'this delivery must be settled by local release' }),
+    });
+    expect(queued.status).toBe(201);
 
     await stack.runtime.deps.db
       .update(agents)
@@ -745,6 +755,17 @@ describe('agent presence and release lifecycle', () => {
         eq(actionInvocations.actionName, 'release'),
       ));
     expect(invocation.status).toBe('completed');
+    expect(
+      await stack.runtime.deps.db
+        .select({ status: deliveries.status, error: deliveries.error })
+        .from(deliveries)
+        .where(eq(deliveries.agentId, target.agentId)),
+    ).toEqual([
+      expect.objectContaining({
+        status: 'dead_lettered',
+        error: 'recipient agent released',
+      }),
+    ]);
   });
 
   it('refuses to register into the reserved released-agent namespace', async () => {

packages/engine/src/__tests__/conformance/deleteAgentRoute.test.ts

@@ -1,7 +1,7 @@
 import { afterEach, beforeEach, describe, expect, it } from 'vitest';
 import { and, eq } from 'drizzle-orm';
 import { createWorkspace, makeNodeStack, registerAgent, type TestStack } from './harness.js';
-import { agents, channelMembers, messages } from '../../db/schema.js';
+import { agents, channelMembers, deliveries, messages } from '../../db/schema.js';
 
 /**
  * `DELETE /v1/agents/:name` -> `agentEngine.deleteAgent` was the last release
@@ -48,7 +48,16 @@ describe('DELETE /v1/agents/:name preserves attributed history', () => {
   it('removes an agent that has authored messages, keeping the attribution', async () => {
     const ws = await createWorkspace(stack.app, 'route-delete-with-history');
     const target = await registerAgent(stack.app, ws.workspaceKey, 'talkative');
+    const sender = await registerAgent(stack.app, ws.workspaceKey, 'sender');
     await post(target.token, 'this message must keep its author');
+    await post(sender.token, 'this queued delivery must stop consuming capacity');
+
+    expect(
+      await stack.runtime.deps.db
+        .select({ status: deliveries.status })
+        .from(deliveries)
+        .where(eq(deliveries.agentId, target.agentId)),
+    ).toContainEqual({ status: 'queued' });
 
     const res = await removeAgent(ws.workspaceKey, target.name);
     expect(res.status).toBeLessThan(300);
@@ -61,6 +70,27 @@ describe('DELETE /v1/agents/:name preserves attributed history', () => {
         .where(and(eq(agents.workspaceId, ws.workspaceId), eq(agents.name, target.name))),
     ).toHaveLength(0);
 
+    // A tombstoned recipient can never ACK its old queue. Settle those rows
+    // immediately so they stop consuming the workspace delivery-depth cap.
+    expect(
+      await stack.runtime.deps.db
+        .select({
+          status: deliveries.status,
+          error: deliveries.error,
+          retryable: deliveries.retryable,
+          deadLetteredAt: deliveries.deadLetteredAt,
+        })
+        .from(deliveries)
+        .where(eq(deliveries.agentId, target.agentId)),
+    ).toEqual([
+      expect.objectContaining({
+        status: 'dead_lettered',
+        error: 'recipient agent released',
+        retryable: false,
+        deadLetteredAt: expect.any(Date),
+      }),
+    ]);
+
     // The row survives as a tombstone so history keeps its author.
     const [tombstone] = await stack.runtime.deps.db
       .select({ name: agents.name, status: agents.status })