fix(agenstra): ticket automation reliability

apps/agenstra/backend-agent-controller/src/migrations/1775400000000_RenameAutomationCompletionSignal.ts

@@ -0,0 +1,19 @@
+import { MigrationInterface, QueryRunner } from 'typeorm';
+
+export class RenameAutomationCompletionSignal1775400000000 implements MigrationInterface {
+  name = 'RenameAutomationCompletionSignal1775400000000';
+
+  public async up(queryRunner: QueryRunner): Promise<void> {
+    await queryRunner.query(`
+      ALTER TABLE "ticket_automation_run"
+      RENAME COLUMN "completion_marker_seen" TO "completion_signal_seen"
+    `);
+  }
+
+  public async down(queryRunner: QueryRunner): Promise<void> {
+    await queryRunner.query(`
+      ALTER TABLE "ticket_automation_run"
+      RENAME COLUMN "completion_signal_seen" TO "completion_marker_seen"
+    `);
+  }
+}

apps/agenstra/backend-agent-controller/src/migrations/1775500000000_AddPreferredModelsForChatAndAutomation.ts

@@ -0,0 +1,29 @@
+import { MigrationInterface, QueryRunner } from 'typeorm';
+
+/**
+ * Persist explicit OpenCode model (`provider/model`) for ticket chat/AI and automation.
+ * OpenCode no longer has an auto model; callers must pass a concrete selection.
+ */
+export class AddPreferredModelsForChatAndAutomation1775500000000 implements MigrationInterface {
+  name = 'AddPreferredModelsForChatAndAutomation1775500000000';
+
+  public async up(queryRunner: QueryRunner): Promise<void> {
+    await queryRunner.query(`
+      ALTER TABLE "tickets"
+      ADD COLUMN IF NOT EXISTS "preferred_chat_model" varchar(256) NULL
+    `);
+    await queryRunner.query(`
+      ALTER TABLE "ticket_automation"
+      ADD COLUMN IF NOT EXISTS "preferred_model" varchar(256) NULL
+    `);
+  }
+
+  public async down(queryRunner: QueryRunner): Promise<void> {
+    await queryRunner.query(`
+      ALTER TABLE "ticket_automation" DROP COLUMN IF EXISTS "preferred_model"
+    `);
+    await queryRunner.query(`
+      ALTER TABLE "tickets" DROP COLUMN IF EXISTS "preferred_chat_model"
+    `);
+  }
+}

apps/agenstra/backend-agent-controller/src/queue/processors/controller-jobs.processor.ts

@@ -331,7 +331,13 @@ export class ControllerJobsProcessor extends WorkerHost {
   }
 
   private async runOpencodeConfigSyncCoordinator(): Promise<void> {
-    const targetIds = await this.opencodeConfigSyncTargets.findPendingTargetIds(getOpencodeConfigSyncBatchSize());
+    const batchSize = getOpencodeConfigSyncBatchSize();
+    // Prefer already-pending/failed targets so reconcile rollouts cannot starve retries.
+    const pending = await this.opencodeConfigSyncTargets.findPendingTargetIds(batchSize);
+    const remaining = Math.max(0, batchSize - pending.length);
+    const reconciled =
+      remaining > 0 ? await this.opencodeConfigSyncTargets.reconcileOutdatedSyncedTargets(remaining) : [];
+    const targetIds = [...new Set([...pending, ...reconciled])].slice(0, batchSize);
 
     for (const targetId of targetIds) {
       await enqueueUnitJob({

apps/agenstra/frontend-agent-console/src/i18n/messages.de.xlf

@@ -3127,13 +3127,17 @@
         <source>Agent provider error</source>
         <target>Agent-Provider-Fehler</target>
       </trans-unit>
+      <trans-unit id="featureTicketsBoard-runFailureAgentNoCompletionStatus" datatype="html">
+        <source>No completion status</source>
+        <target>Kein Abschlussstatus</target>
+      </trans-unit>
       <trans-unit id="featureTicketsBoard-runFailureAgentNoCompletionMarker" datatype="html">
         <source>No completion marker</source>
         <target>Kein Abschluss-Marker</target>
       </trans-unit>
       <trans-unit id="featureTicketsBoard-runFailureMarkerWithoutVerify" datatype="html">
-        <source>Completion marker without verify profile</source>
-        <target>Abschluss-Marker ohne Verifizierungsprofil</target>
+        <source>Completion without verify profile</source>
+        <target>Abschluss ohne Verify-Profil</target>
       </trans-unit>
       <trans-unit id="featureTicketsBoard-runFailureVerifyCommandFailed" datatype="html">
         <source>Verify command failed</source>
@@ -3307,10 +3311,30 @@
         <source> Agent for chat / AI </source>
         <target>Agent für Chat / KI</target>
       </trans-unit>
+      <trans-unit id="featureTicketsBoard-modelLabel" datatype="html">
+        <source> Model for chat / AI </source>
+        <target>Modell für Chat / KI</target>
+      </trans-unit>
+      <trans-unit id="featureTicketsBoard-modelPlaceholder" datatype="html">
+        <source> {{ ticketAiModelsLoading() ? 'Loading models…' : 'Select agent to load models…' }} </source>
+        <target> {{ ticketAiModelsLoading() ? 'Modelle werden geladen…' : 'Agent wählen, um Modelle zu laden…' }} </target>
+      </trans-unit>
       <trans-unit id="featureTicketsBoard-agentPlaceholder" datatype="html">
         <source>Select agent…</source>
         <target>Agent wählen…</target>
       </trans-unit>
+      <trans-unit id="featureTicketsBoard-automationPreferredModel" datatype="html">
+        <source>Model for automation</source>
+        <target>Modell für Automatisierung</target>
+      </trans-unit>
+      <trans-unit id="featureTicketsBoard-automationPreferredModelHint" datatype="html">
+        <source> Required when automation is enabled. OpenCode has no auto model selection. </source>
+        <target> Erforderlich, wenn die Automatisierung aktiviert ist. OpenCode hat keine Auto-Modellauswahl. </target>
+      </trans-unit>
+      <trans-unit id="featureTicketsBoard-automationModelPlaceholder" datatype="html">
+        <source> {{ ticketAiModelsLoading() ? 'Loading models…' : 'Select an agent to load models…' }} </source>
+        <target> {{ ticketAiModelsLoading() ? 'Modelle werden geladen…' : 'Agent wählen, um Modelle zu laden…' }} </target>
+      </trans-unit>
       <trans-unit id="featureTicketsBoard-acceptanceLabel" datatype="html">
         <source> Acceptance <x id="START_TAG_SPAN" ctype="x-span" equiv-text="&lt;span class=&quot;text-muted small d-block fw-normal&quot; [attr.aria-label]=&quot;percentage + &apos;%&apos;&quot;
                         &gt;"/><x id="INTERPOLATION" equiv-text="{{ percentage | number: &apos;1.0-0&apos; }}"/>%<x id="CLOSE_TAG_SPAN" ctype="x-span" equiv-text="&lt;/span

apps/agenstra/frontend-agent-console/src/i18n/messages.xlf

@@ -2953,11 +2953,14 @@
       <trans-unit id="featureTicketsBoard-runFailureAgentProviderError" datatype="html">
         <source>Agent provider error</source>
         </trans-unit>
+      <trans-unit id="featureTicketsBoard-runFailureAgentNoCompletionStatus" datatype="html">
+        <source>No completion status</source>
+        </trans-unit>
       <trans-unit id="featureTicketsBoard-runFailureAgentNoCompletionMarker" datatype="html">
         <source>No completion marker</source>
         </trans-unit>
       <trans-unit id="featureTicketsBoard-runFailureMarkerWithoutVerify" datatype="html">
-        <source>Completion marker without verify profile</source>
+        <source>Completion without verify profile</source>
         </trans-unit>
       <trans-unit id="featureTicketsBoard-runFailureVerifyCommandFailed" datatype="html">
         <source>Verify command failed</source>
@@ -3088,9 +3091,24 @@
       <trans-unit id="featureTicketsBoard-agentLabel" datatype="html">
         <source> Agent for chat / AI </source>
         </trans-unit>
+      <trans-unit id="featureTicketsBoard-modelLabel" datatype="html">
+        <source> Model for chat / AI </source>
+        </trans-unit>
+      <trans-unit id="featureTicketsBoard-modelPlaceholder" datatype="html">
+        <source> {{ ticketAiModelsLoading() ? 'Loading models…' : 'Select agent to load models…' }} </source>
+        </trans-unit>
       <trans-unit id="featureTicketsBoard-agentPlaceholder" datatype="html">
         <source>Select agent…</source>
         </trans-unit>
+      <trans-unit id="featureTicketsBoard-automationPreferredModel" datatype="html">
+        <source>Model for automation</source>
+        </trans-unit>
+      <trans-unit id="featureTicketsBoard-automationPreferredModelHint" datatype="html">
+        <source> Required when automation is enabled. OpenCode has no auto model selection. </source>
+        </trans-unit>
+      <trans-unit id="featureTicketsBoard-automationModelPlaceholder" datatype="html">
+        <source> {{ ticketAiModelsLoading() ? 'Loading models…' : 'Select an agent to load models…' }} </source>
+        </trans-unit>
       <trans-unit id="featureTicketsBoard-acceptanceLabel" datatype="html">
         <source> Acceptance <x id="START_TAG_SPAN" ctype="x-span" equiv-text="&lt;span class=&quot;text-muted small d-block fw-normal&quot; [attr.aria-label]=&quot;percentage + &apos;%&apos;&quot;
                         &gt;"/><x id="INTERPOLATION" equiv-text="{{ percentage | number: &apos;1.0-0&apos; }}"/>%<x id="CLOSE_TAG_SPAN" ctype="x-span" equiv-text="&lt;/span

docs/agenstra/ai-agents/agent-client-protocol.md

@@ -42,7 +42,7 @@ Built-in provider:
   - **User chats** — `-chat-{uuid}` where `{uuid}` is the chat session id (one ACP session per user-created thread)
   - **Reserved / hidden** — not listed in agent `chats` / UI: `-prompt-enhance`, `-ticket-body`, and `-ticket-auto-*` (for example `-ticket-auto-pre`, `-ticket-auto-loop`, `-ticket-auto-commit-msg`)
     After an API restart, the manager opens a new stdio transport and calls `session/load` when the container id still matches. That covers primary, user, and background sessions the same way in-memory reuse already did within a process.
-- **Permissions:** `session/request_permission` is auto-approved when `ACP_AUTO_APPROVE` is not `false` (default for headless agents)
+- **Permissions:** `session/request_permission` is auto-approved when `ACP_AUTO_APPROVE` is not `false` (default for headless agents). **OpenCode ticket automation does not use this flag** — reserved `-ticket-auto-*` sessions use OpenCode session permission rulesets and runtime auto-reply instead (see [Ticket automation](../features/ticket-automation.md)).
 
 ## Configuration
 

docs/agenstra/features/ticket-automation.md

@@ -38,18 +38,29 @@ There is no separate “start run” HTTP call for this path: eligible tickets a
    - `new_per_run`: ephemeral branch `automation/run/{first 8 chars of run UUID}` per run.
    - `force_new_automation_branch_next_run`: with `reuse_per_ticket`, forces one ephemeral branch for the next run, then clears the flag.
 4. **Optional pre-improve** If autonomy `pre_improve_ticket` is true, one chat turn asks the agent to clarify the ticket only (no implementation).
-5. **Implementation loop** Up to `max_iterations` remote chat turns (default `20`). The agent must include the exact completion marker `AGENSTRA_AUTOMATION_COMPLETE` in assistant text to exit the loop successfully.
+5. **Implementation loop** Up to `max_iterations` remote chat turns (default `20`). Each turn uses an OpenCode automation session with session-scoped allow-all permissions and structured `json_schema` turn status. The loop exits successfully only when structured output reports `status: "complete"`.
 6. **Verification** If a `verifier_profile` with `commands` is configured, those commands run in the agent workspace (bounded timeout); any non-zero exit fails the run.
 7. **Finalize** Stage changes, generate a Conventional-Commits-style subject (with fallback), `git commit` and `git push` unless the tree is already clean.
 8. **Success** Run status `succeeded`, ticket status set to **`prototype`**, failure counters cleared, lease released, board and activity events emitted.
 
+### Unattended OpenCode sessions
+
+Reserved resume-session suffixes (`-ticket-auto-pre`, `-ticket-auto-loop`, `-ticket-auto-commit-msg`) create **hidden ephemeral** OpenCode sessions that:
+
+- Apply a session-scoped permission ruleset (`allow` for all actions) so interactive ask/deny from the worker UI config does not block the run.
+- Auto-reply residual permission asks (`always`) and questions on those sessions only; interactive chat suffixes are unchanged.
+- Force the platform agent `agenstra-automation` and skill `agenstra-ticket-automation` (injected at config sync so UI overlays cannot remove them). The skill documents the structured-status protocol; completion detection uses structured output only.
+- Keep transcripts out of the user’s primary chat history; run progress still surfaces via automation chat cards / board events.
+
+Ticket approval and agent autonomy prerequisites still apply. Only reserved automation sessions bypass interactive permission prompts.
+
 Remote chat turns and commit-message generation are recorded in usage statistics as `autonomous_ticket_run_turn` and `autonomous_ticket_commit_message` interaction kinds (see [Usage statistics](./usage-statistics.md)).
 
 ## Failures and retries
 
 Failures map to a terminal run status (`failed`, `timed_out`, `escalated`, `cancelled`) and may adjust ticket status. When policy says **requeue**, `next_retry_at` is set to approximately **one minute** ahead so the scheduler does not hot-loop; `consecutive_failure_count` increments.
 
-Examples (not exhaustive): missing completion marker or budget-related timeout tend to move the ticket back toward `todo` and requeue; human escalation stops requeue; lease contention skips starting a run but schedules a retry. Product policy is centralized in code (`routeAutomationFailure` in the agent-controller library); adjust there when behavior changes.
+Examples (not exhaustive): missing structured completion status or budget-related timeout tend to move the ticket back toward `todo` and requeue; human escalation stops requeue; lease contention skips starting a run but schedules a retry. Product policy is centralized in code (`routeAutomationFailure` in the agent-controller library); adjust there when behavior changes.
 
 ## HTTP and realtime
 

graph/graph.json

@@ -1,6 +1,6 @@
 {
   "version": 1,
-  "generatedAt": "2026-10-04T15:47:00.617Z",
+  "generatedAt": "2026-10-05T19:13:17.421Z",
   "nodes": [
     {
       "id": "project:@forepath/test/mounted-plugin-fixture",
@@ -60821,6 +60821,11 @@
       "to": "file:libs/domains/agenstra/backend/feature-agent-controller/src/lib/services/ticket-automation.service.ts",
       "type": "injects"
     },
+    {
+      "from": "file:libs/domains/agenstra/backend/feature-agent-controller/src/lib/services/autonomous-run-orchestrator.service.ts",
+      "to": "file:libs/domains/agenstra/backend/feature-agent-controller/src/lib/services/knowledge-tree.service.ts",
+      "type": "injects"
+    },
     {
       "from": "file:libs/domains/agenstra/backend/feature-agent-controller/src/lib/services/client-agent-autonomy.service.ts",
       "to": "file:libs/domains/agenstra/backend/feature-agent-controller/src/lib/repositories/clients.repository.ts",

libs/domains/agenstra/backend/feature-agent-controller/docs/autonomous-ticket-decision-tree.mmd

@@ -26,9 +26,9 @@ flowchart TD
     workspacePrep -- "No" --> failProvider["Fail run: AGENT_PROVIDER_ERROR"]
     workspacePrep -- "Yes" --> loopStart["Agent loop (maxIterations)"]
 
-    loopStart --> markerSeen{"Completion marker seen?"}
-    markerSeen -- "No, iterations exhausted" --> failNoMarker["Fail run: AGENT_NO_COMPLETION_MARKER (timed_out)"]
-    markerSeen -- "Yes" --> verifierConfigured{"Verifier commands configured?"}
+    loopStart --> signalSeen{"Structured status complete?"}
+    signalSeen -- "No, iterations exhausted" --> failNoSignal["Fail run: AGENT_NO_COMPLETION_STATUS (timed_out)"]
+    signalSeen -- "Yes" --> verifierConfigured{"Verifier commands configured?"}
 
     verifierConfigured -- "No" --> verifySkipped["Skip verify"]
     verifierConfigured -- "Yes" --> verifyResult{"All verifier commands exit 0?"}

libs/domains/agenstra/backend/feature-agent-controller/docs/sequence-autonomous-ticket.mmd

@@ -21,15 +21,15 @@ sequenceDiagram
     end
     Note over Orch,Vcs: Default strategy reuses `automation/ticket-{ticketId}`; `new_per_run` or one-shot force uses ephemeral `automation/{runPrefix}`.
     loop maxIterations
-        Orch->>Chat: sendChatSync(sync, ephemeral, contextInjection)
+        Orch->>Chat: sendChatSync(sync, ephemeral, expectAutomationTurnStatus)
         Chat->>Mgr: chat
-        Mgr->>Prov: sendMessage(continue, resumeSessionSuffix)
-        Prov-->>Mgr: NDJSON / text
-        Mgr-->>Chat: chatMessage
-        Chat-->>Orch: assistant text
+        Mgr->>Prov: sendMessage(continue, resumeSessionSuffix, format json_schema)
+        Prov-->>Mgr: structured status + text
+        Mgr-->>Chat: chatMessage (final result)
+        Chat-->>Orch: RemoteChatSyncResult (text, turnStatus)
         Orch->>DB: append run step
     end
-    Note over Orch,Chat: contextInjection carries includeWorkspace, environmentIds, and autoEnrichmentEnabled.
+    Note over Orch,Chat: Loop exits only when turnStatus is complete. Automation sessions use allow-all + auto-reply.
     Orch->>Vcs: runVerifierCommands
     Orch->>Vcs: getStatus; if dirty: stage all
     Orch->>Chat: sendChatSync (conventional commit subject; autonomous_ticket_commit_message)

libs/domains/agenstra/backend/feature-agent-controller/spec/openapi.yaml

@@ -7660,6 +7660,11 @@ components:
           type: string
           format: uuid
           description: Preferred workspace agent for chat/AI when viewing this ticket (agent-manager agent id).
+        preferredChatModel:
+          type: string
+          maxLength: 256
+          nullable: true
+          description: Preferred OpenCode provider/model for chat/AI on this ticket (e.g. anthropic/claude-sonnet-4).
         automationEligible:
           type: boolean
           description: Whether autonomous prototyping is enabled for this ticket (ticket_automation.eligible).
@@ -7745,6 +7750,11 @@ components:
           type: string
           format: uuid
           description: Persist preferred workspace agent for chat/AI; send null to clear.
+        preferredChatModel:
+          type: string
+          maxLength: 256
+          nullable: true
+          description: Persist preferred OpenCode provider/model for chat/AI; send null to clear.
     TicketCommentResponseDto:
       type: object
       required: [id, ticketId, body, createdAt]
@@ -7829,6 +7839,11 @@ components:
           items:
             type: string
             format: uuid
+        preferredModel:
+          type: string
+          maxLength: 256
+          nullable: true
+          description: OpenCode provider/model for autonomous turns; required when eligible is true.
         includeWorkspaceContext:
           type: boolean
         contextEnvironmentIds:
@@ -7862,6 +7877,7 @@ components:
         - ticketId
         - eligible
         - allowedAgentIds
+        - preferredModel
         - includeWorkspaceContext
         - contextEnvironmentIds
         - autoEnrichmentEnabled
@@ -7888,6 +7904,11 @@ components:
           items:
             type: string
             format: uuid
+        preferredModel:
+          type: string
+          maxLength: 256
+          nullable: true
+          description: OpenCode provider/model used for autonomous turns on this ticket.
         includeWorkspaceContext:
           type: boolean
         contextEnvironmentIds:
@@ -7978,7 +7999,7 @@ components:
         - finishedAt
         - updatedAt
         - iterationCount
-        - completionMarkerSeen
+        - completionSignalSeen
         - verificationPassed
         - failureCode
         - summary
@@ -8021,7 +8042,7 @@ components:
           format: date-time
         iterationCount:
           type: integer
-        completionMarkerSeen:
+        completionSignalSeen:
           type: boolean
         verificationPassed:
           type: boolean
@@ -8030,8 +8051,8 @@ components:
           description: >
             Present when a run ends unsuccessfully. Includes values such as verify_command_failed,
             commit_failed (git commit after successful verification), push_failed (git push after that commit),
-            agent_no_completion_marker, marker_without_verify, agent_provider_error, vcs_dirty_workspace,
-            and others defined by the controller.
+            agent_no_completion_status, agent_no_completion_marker (legacy), marker_without_verify, agent_provider_error,
+            vcs_dirty_workspace, and others defined by the controller.
         summary:
           type: object
           additionalProperties: true

libs/domains/agenstra/backend/feature-agent-controller/src/lib/dto/ticket-automation/ticket-automation-response.dto.ts

@@ -4,6 +4,7 @@ export class TicketAutomationResponseDto {
   ticketId!: string;
   eligible!: boolean;
   allowedAgentIds!: string[];
+  preferredModel!: string | null;
   includeWorkspaceContext!: boolean;
   contextEnvironmentIds!: string[];
   autoEnrichmentEnabled!: boolean;

libs/domains/agenstra/backend/feature-agent-controller/src/lib/dto/ticket-automation/ticket-automation-run-response.dto.ts

@@ -25,7 +25,7 @@ export class TicketAutomationRunResponseDto {
   finishedAt!: Date | null;
   updatedAt!: Date;
   iterationCount!: number;
-  completionMarkerSeen!: boolean;
+  completionSignalSeen!: boolean;
   verificationPassed!: boolean | null;
   failureCode!: string | null;
   summary!: Record<string, unknown> | null;