Q129: a run with failed actions it never redid does not count as completed

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/main/database/migrations/092-run-action-targets.ts

@@ -0,0 +1,7 @@
+// What each journaled action acted on; rows from before stay null, meaning any later success of the same tool redoes them.
+export const runActionTargetsSchema = `
+  alter table run_actions add column target text;
+
+  insert into schema_migrations(version, name, applied_at)
+  values (92, 'run_action_targets', datetime('now'));
+`;

src/main/database/schema.ts

@@ -90,6 +90,7 @@ import { hostAccessNoteSchema } from "./migrations/088-host-access-note";
 import { fileVersionsSchema } from "./migrations/089-file-versions";
 import { erasureReceiptsSchema } from "./migrations/090-erasure-receipts";
 import { requestPurposeSchema } from "./migrations/091-request-purpose";
+import { runActionTargetsSchema } from "./migrations/092-run-action-targets";
 
 interface Migration {
   version: number;
@@ -189,6 +190,7 @@ const MIGRATIONS: Migration[] = [
   { version: 89, name: "file_versions", sql: fileVersionsSchema },
   { version: 90, name: "erasure_receipts", sql: erasureReceiptsSchema },
   { version: 91, name: "request_purpose", sql: requestPurposeSchema },
+  { version: 92, name: "run_action_targets", sql: runActionTargetsSchema },
 ];
 
 const SCHEMA_VERSION = MIGRATIONS[MIGRATIONS.length - 1].version;

src/main/services/agent-run-loop.ts

@@ -105,19 +105,24 @@ export async function runAgentLoop(deps: AgentRunLoopDeps, input: AgentRunLoopIn
     lastText = turn.text.trim() || lastText;
     messages.push({ role: "assistant", text: turn.text, toolCalls: turn.toolCalls, ...(turn.replay ? { replay: turn.replay } : {}) });
     const results: Array<{ toolCallId: string; toolName: string; output: string }> = [];
+    // A call refused before it ran is still journaled, so a run that never redoes it cannot pass as completed.
+    const refuse = (call: ToolCall, output: string) => {
+      input.onToolCompleted?.(call, output);
+      results.push({ toolCallId: call.id, toolName: call.name, output });
+    };
     for (const [index, call] of turn.toolCalls.entries()) {
       throwIfRunStopped();
       if (index >= MAX_TOOL_CALLS_PER_TURN) {
-        results.push({ toolCallId: call.id, toolName: call.name, output: `ERROR: not run: this turn asked for ${turn.toolCalls.length} tool calls; the limit is ${MAX_TOOL_CALLS_PER_TURN}. Repeat the ones you still need.` });
+        refuse(call, `ERROR: not run: this turn asked for ${turn.toolCalls.length} tool calls; the limit is ${MAX_TOOL_CALLS_PER_TURN}. Repeat the ones you still need.`);
         continue;
       }
       if (call.argumentsError) {
         console.warn(`[run-loop] ${call.name} not run: arguments ${call.argumentsError}`);
-        results.push({ toolCallId: call.id, toolName: call.name, output: unreadableArguments(call) });
+        refuse(call, unreadableArguments(call));
         continue;
       }
       if (call.name === "complete_goal" && goalAudits >= MAX_GOAL_AUDITS_PER_SESSION) {
-        results.push({ toolCallId: call.id, toolName: call.name, output: `ERROR: this session already ran ${MAX_GOAL_AUDITS_PER_SESSION} goal audits. Close the gaps they named, then try again next session.` });
+        refuse(call, `ERROR: this session already ran ${MAX_GOAL_AUDITS_PER_SESSION} goal audits. Close the gaps they named, then try again next session.`);
         continue;
       }
       if (call.name === "complete_goal") goalAudits += 1;

src/main/services/run-journal.ts

@@ -1,4 +1,5 @@
 import type { DatabaseSync } from "node:sqlite";
+import { posix } from "node:path";
 import { createId } from "../../shared/ids";
 import type { ToolCall } from "./provider-tool-turn";
 
@@ -10,6 +11,15 @@ const READ_ONLY_TOOLS = new Set([
 const JOURNAL_LINES = 40;
 const FIELD_CHARS = 160;
 const DECLINED_LINES = 5;
+/** The arguments that name what an action acted on (retry controls such as confirm_not_done excluded); a tool not listed is identified by all of its arguments. */
+const TARGET_ARGUMENTS: Record<string, string[]> = {
+  write_deliverable: ["path"], create_document: ["title", "doc_type"], update_document: ["doc_id"], create_task: ["title"],
+  update_task_status: ["task_id"], hire_agent: ["name"], grant_connector: ["agent_name", "toolkit"], revoke_connector: ["agent_name", "toolkit"],
+  grant_skill: ["agent_name", "skill"], revoke_skill: ["agent_name", "skill"], record_knowledge: ["topic"], revise_knowledge: ["id"],
+  update_goal: ["goal_id"], complete_goal: [], run_command: ["command"], composio_run_action: ["action", "account", "arguments"],
+};
+/** A result line that reports failure: a refused or broken call, or a command that exited non-zero. */
+const FAILED_RESULT = (row: string) => `(${row}.result_excerpt like 'ERROR%' or (${row}.result_excerpt like 'exit code %' and ${row}.result_excerpt <> 'exit code 0'))`;
 /** What a held deliverable would become, over a `held_deliverables` row alias: one file, one existing document, or a new document by type and title. */
 const heldTarget = (row: string) => `(case ${row}.kind when 'file' then ${row}.workspace_root || '/' || ${row}.relative_path when 'document_update' then ${row}.document_id else ${row}.doc_type || ':' || ${row}.title end)`;
 
@@ -17,10 +27,41 @@ const heldTarget = (row: string) => `(case ${row}.kind when 'file' then ${row}.w
 export function recordRunAction(db: DatabaseSync, input: { orgId: string; runId: string; taskId: string; call: ToolCall; output: string }): void {
   if (READ_ONLY_TOOLS.has(input.call.name)) return;
   db.prepare(`
-    insert into run_actions(id, org_id, run_id, task_id, tool_name, arguments_excerpt, result_excerpt, created_at)
-    values (?, ?, ?, ?, ?, ?, ?, ?)
+    insert into run_actions(id, org_id, run_id, task_id, tool_name, arguments_excerpt, result_excerpt, target, created_at)
+    values (?, ?, ?, ?, ?, ?, ?, ?, ?)
   `).run(createId("rac"), input.orgId, input.runId, input.taskId, input.call.name,
-    clip(JSON.stringify(input.call.arguments)), clip(input.output.split("\n")[0] ?? ""), new Date().toISOString());
+    clip(JSON.stringify(input.call.arguments)), clip(input.output.split("\n")[0] ?? ""), actionTarget(input.call), new Date().toISOString());
+}
+
+/** What the call acted on, or null when its arguments could not be read. */
+function actionTarget(call: ToolCall): string | null {
+  if (call.argumentsError) return null;
+  const keys = TARGET_ARGUMENTS[call.name];
+  if (!keys) return JSON.stringify(canonical(call.arguments));
+  return JSON.stringify(keys.map((key) => targetValue(key, call.arguments[key])));
+}
+
+/** An argument as the tools read it: text trimmed, paths in one spelling, nested values in one key order. */
+function targetValue(key: string, value: unknown): unknown {
+  if (typeof value !== "string") return canonical(value ?? null);
+  return key === "path" ? posix.normalize(value.trim().replaceAll("\\", "/")) : value.trim();
+}
+
+/** Why a session that ended normally still failed: actions that failed and were not later done successfully on the same target (any target when unknown). */
+export function unresolvedActionFailure(db: DatabaseSync, runId: string): string | null {
+  const failed = db.prepare(`
+    select a.tool_name, a.result_excerpt from run_actions a
+    where a.run_id = ? and ${FAILED_RESULT("a")}
+      and not exists (
+        select 1 from run_actions later
+        where later.run_id = a.run_id and later.tool_name = a.tool_name and later.rowid > a.rowid and not ${FAILED_RESULT("later")}
+          and (a.target is null or later.target = a.target)
+      )
+    order by a.rowid
+  `).all(runId) as Array<{ tool_name: string; result_excerpt: string }>;
+  if (failed.length === 0) return null;
+  const list = failed.map((action) => `${action.tool_name} (${action.result_excerpt})`).join("; ");
+  return `The session ended with ${failed.length} failed action${failed.length === 1 ? "" : "s"} it did not redo: ${list}`;
 }
 
 /** What earlier sessions (including an interrupted attempt of this run) already did on the task, and where the last one failed. */
@@ -58,3 +99,10 @@ export function runJournalLines(db: DatabaseSync, taskId: string): string[] {
 function clip(text: string): string {
   return text.length > FIELD_CHARS ? `${text.slice(0, FIELD_CHARS)}...` : text;
 }
+
+/** The same JSON value with object keys sorted, so equal arguments serialize equally. */
+function canonical(value: unknown): unknown {
+  if (Array.isArray(value)) return value.map(canonical);
+  if (value === null || typeof value !== "object") return value;
+  return Object.fromEntries(Object.keys(value).sort().map((key) => [key, canonical((value as Record<string, unknown>)[key])]));
+}

src/main/services/task-execution-service.ts

@@ -11,7 +11,7 @@ import { SkillService } from "./skill-service";
 import { runAgentLoop, runTools } from "./agent-run-loop";
 import type { RunModel } from "./skill-runtime";
 import { combineMidRunPolls, createGoalChangePoller } from "./goal-change-poller";
-import { recordRunAction, runJournalLines } from "./run-journal";
+import { recordRunAction, unresolvedActionFailure, runJournalLines } from "./run-journal";
 import { classifyProviderFailure } from "./provider-failure";
 import { runStopReason, SESSION_TIME_LIMIT_MESSAGE, SESSION_TIME_LIMIT_MS, withRunControl } from "./run-control";
 import type { AgentToolContext } from "./agent-tools";
@@ -145,7 +145,7 @@ export class TaskExecutionService {
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
       console.error(`[task-run] run ${runId} failed before execution: ${message}`);
-      this.failRun(orgId, runId, null, message, false);
+      this.failRun(orgId, runId, null, message, null);
       this.audit.record({
         orgId,
         entityType: "task_run",
@@ -277,15 +277,18 @@ export class TaskExecutionService {
         });
         if (delivered) this.notifyDataChanged();
       }
-      this.completeRun(orgId, runId, result.lastRequestId, result.hitStepLimit);
-      if (task.origin === null) recordRunMemory(this.db, { orgId, runId, projectId: task.project_id, taskTitle: task.title, outcome: "completed", text: result.finalText });
+      // Real work whose writes or external actions failed and were not redone has not succeeded, whatever the model said last.
+      const actionFailure = result.hitStepLimit || task.origin !== null ? null : unresolvedActionFailure(this.db, runId);
+      if (actionFailure) this.failRun(orgId, runId, result.lastRequestId, actionFailure, null);
+      else this.completeRun(orgId, runId, result.lastRequestId, result.hitStepLimit);
+      if (task.origin === null) recordRunMemory(this.db, { orgId, runId, projectId: task.project_id, taskTitle: task.title, outcome: actionFailure ? "failed" : "completed", text: actionFailure ?? result.finalText });
       this.audit.record({
         orgId,
         entityType: "task_run",
         entityId: runId,
         actorType: "system",
         actorId: null,
-        action: "completed",
+        action: actionFailure ? "failed" : "completed",
         payload: {
           taskId: task.id,
           agentId: agent.id,
@@ -315,7 +318,7 @@ export class TaskExecutionService {
         throw new Error(stopReason);
       }
       const failure = classifyProviderFailure(error, Date.now());
-      this.failRun(orgId, runId, providerRequestId, message, failure.transient);
+      this.failRun(orgId, runId, providerRequestId, message, failure.transient ? "transient" : null);
       if (failure.transient) this.cooldowns.park(provider, message, failure.retryAfterMs);
       else if (task.origin === null) recordRunMemory(this.db, { orgId, runId, projectId: task.project_id, taskTitle: task.title, outcome: "failed", text: message });
       this.audit.record({
@@ -434,12 +437,7 @@ export class TaskExecutionService {
     if (!checkpoint) this.closeChatWakeTask(orgId, runId);
   }
 
-  /**
-   * A chat-wake task exists only to carry its reply run — the reply is the
-   * whole deliverable. The run completing therefore closes the task; left
-   * open it would strand forever (a task with a completed run is never
-   * enqueued again) and pile zombie "Reply to ..." rows onto the board.
-   */
+  /** A chat-wake task only carries its reply run, so the run completing closes it; left open it would strand forever. */
   private closeChatWakeTask(orgId: string, runId: string): void {
     const run = this.db.prepare("select r.task_id, t.origin from task_runs r join tasks t on t.id = r.task_id where r.id = ? and r.org_id = ?")
       .get(runId, orgId) as { task_id: string; origin: string | null } | undefined;
@@ -479,12 +477,12 @@ export class TaskExecutionService {
     return runId;
   }
 
-  private failRun(orgId: string, runId: string, providerRequestId: string | null, errorMessage: string, transient: boolean): void {
+  private failRun(orgId: string, runId: string, providerRequestId: string | null, errorMessage: string, failureKind: "transient" | null): void {
     this.db.prepare(`
       update task_runs
       set state = ?, provider_request_id = coalesce(?, provider_request_id), completed_at = ?, error_message = ?, failure_kind = ?
       where id = ? and org_id = ?
-    `).run(TASK_RUN_STATE.FAILED, providerRequestId, new Date().toISOString(), errorMessage, transient ? "transient" : null, runId, orgId);
+    `).run(TASK_RUN_STATE.FAILED, providerRequestId, new Date().toISOString(), errorMessage, failureKind, runId, orgId);
   }
 
   private resolveSkillRuntimeSnapshot(orgId: string, agentId: string, runId: string, maxTokens: number, taskText: string, runModel: RunModel, houseStyle: boolean): SkillRuntimeSnapshotRow | null {
@@ -498,6 +496,4 @@ export class TaskExecutionService {
     if (!snapshot) throw new Error("Skill runtime snapshot was not found.");
     return snapshot;
   }
-
-
 }

tests/integration/run-action-failures.test.ts

@@ -0,0 +1,128 @@
+import { randomUUID } from "node:crypto";
+import { mkdirSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { describe, expect, test } from "vitest";
+import { DatabaseConnection } from "../../src/main/database/connection";
+import { OnboardingService } from "../../src/main/services/onboarding-service";
+import { ProviderService } from "../../src/main/services/provider-service";
+import { providerAdapters } from "../../src/main/services/provider-adapters";
+import { recordRunAction, unresolvedActionFailure } from "../../src/main/services/run-journal";
+import { TaskExecutionService } from "../../src/main/services/task-execution-service";
+import { claimNextQueuedRun } from "../../src/main/app/task-queue-worker";
+import type { AgentTurnResult, ProviderToolTurnInput, ToolCall } from "../../src/main/services/provider-tool-turn";
+import { answeringModelCheck } from "../helpers/model-check";
+
+const testVault = {
+  encryptSecret: (secret: string) => Buffer.from(secret, "utf8"),
+  decryptSecret: (ciphertext: Buffer) => ciphertext.toString("utf8"),
+};
+
+/** A company whose kickoff run makes the given tool calls, one turn each, then ends with a final answer. */
+function company(calls: ToolCall[]) {
+  let turns = 0;
+  const turn = async (_input: ProviderToolTurnInput): Promise<AgentTurnResult> => {
+    const call = calls[turns];
+    turns += 1;
+    return call ? { text: "Working.", toolCalls: [call], rawResponseJson: "{}" } : { text: "All done.", toolCalls: [], rawResponseJson: "{}" };
+  };
+  const connection = new DatabaseConnection(join(tmpdir(), `ac-action-failures-${randomUUID()}.db`));
+  const providers = new ProviderService(connection.db, providerAdapters, testVault, answeringModelCheck(turn));
+  const onboarding = new OnboardingService({ db: connection.db, saveProviderCredential: (input) => providers.saveCredential(input) });
+  const execution = new TaskExecutionService(connection.db, providers);
+  const workspace = join(tmpdir(), `ac-action-failures-ws-${randomUUID()}`);
+  mkdirSync(workspace, { recursive: true });
+  let orgId = "";
+  connection.transaction(() => {
+    orgId = onboarding.bootstrap({
+      companyName: "FlowPilot", companyDescription: "CSM analytics.", employeeName: "Yang", workspacePath: workspace,
+      executionMode: "ai_ceo", autoApproveHires: true, reviewDeliverables: false,
+      goalTitle: "Ship the MVP.", goalDescription: "A working product.",
+      providerKey: "openai", providerSecret: "sk-test-key", providerModel: "gpt-5.4",
+    }).orgId;
+  });
+  return { connection, execution, orgId, runId: claimNextQueuedRun(connection)!.id };
+}
+
+const failingWrite: ToolCall = { id: "c1", name: "write_deliverable", arguments: { path: "../outside.md", content: "x" } };
+
+describe("a run whose actions failed", () => {
+  test("does not count as completed when a failed action was never redone", async () => {
+    const h = company([failingWrite]);
+    await h.execution.executeRun(h.orgId, h.runId);
+    const run = h.connection.db.prepare("select state, failure_kind, error_message from task_runs where id = ?").get(h.runId) as { state: string; failure_kind: string; error_message: string };
+    // Not a passing provider problem, so it counts against the task's retry budget like any failure.
+    expect([run.state, run.failure_kind]).toEqual(["failed", null]);
+    expect(run.error_message).toContain("The session ended with 1 failed action it did not redo: write_deliverable");
+    h.connection.close();
+  });
+
+  test("completes when every failed action was later done successfully", async () => {
+    // The tool trims titles, so " Plan " and "Plan" name the same document.
+    const broken: ToolCall = { id: "c1", name: "create_document", arguments: { title: " Plan ", doc_type: "plan" } };
+    const fixed: ToolCall = { id: "c2", name: "create_document", arguments: { title: "Plan", doc_type: "plan", content: "The plan." } };
+    const h = company([broken, fixed]);
+    await h.execution.executeRun(h.orgId, h.runId);
+    expect(h.connection.db.prepare("select state, failure_kind from task_runs where id = ?").get(h.runId)).toEqual({ state: "completed", failure_kind: null });
+    h.connection.close();
+  });
+
+  test("a success on a different target does not redo the failed one", async () => {
+    const broken: ToolCall = { id: "c1", name: "create_document", arguments: { title: "Required plan", doc_type: "plan" } };
+    const other: ToolCall = { id: "c2", name: "create_document", arguments: { title: "Other report", doc_type: "report", content: "Body." } };
+    const h = company([broken, other]);
+    await h.execution.executeRun(h.orgId, h.runId);
+    const run = h.connection.db.prepare("select state, error_message from task_runs where id = ?").get(h.runId) as { state: string; error_message: string };
+    expect(run.state).toBe("failed");
+    expect(run.error_message).toContain("1 failed action it did not redo: create_document");
+    h.connection.close();
+  });
+
+  test("a call whose arguments could not be read counts as a failed action", async () => {
+    const h = company([{ id: "c1", name: "write_deliverable", arguments: {}, argumentsError: "not valid JSON (Unexpected end): {\"path\":" }]);
+    await h.execution.executeRun(h.orgId, h.runId);
+    const run = h.connection.db.prepare("select state, error_message from task_runs where id = ?").get(h.runId) as { state: string; error_message: string };
+    expect(run.state).toBe("failed");
+    expect(run.error_message).toContain("1 failed action it did not redo: write_deliverable (ERROR: not run: the arguments");
+    h.connection.close();
+  });
+
+  test("a command that exited non-zero is a failure until the same command exits zero", () => {
+    const h = company([]);
+    const task = h.connection.db.prepare("select task_id from task_runs where id = ?").get(h.runId) as { task_id: string };
+    const run = (command: string, output: string) => recordRunAction(h.connection.db, { orgId: h.orgId, runId: h.runId, taskId: task.task_id, call: { id: command, name: "run_command", arguments: { command } }, output });
+    run("pnpm test", "exit code 1\n3 failed");
+    run("git status", "exit code 0\nclean");
+    expect(unresolvedActionFailure(h.connection.db, h.runId)).toBe("The session ended with 1 failed action it did not redo: run_command (exit code 1)");
+    run("pnpm test", "exit code 0\nall passed");
+    expect(unresolvedActionFailure(h.connection.db, h.runId)).toBeNull();
+    h.connection.close();
+  });
+
+  test("a connector action on other arguments does not redo the failed one", () => {
+    const h = company([]);
+    const task = h.connection.db.prepare("select task_id from task_runs where id = ?").get(h.runId) as { task_id: string };
+    const update = (issue: number, output: string) => recordRunAction(h.connection.db, {
+      orgId: h.orgId, runId: h.runId, taskId: task.task_id, output,
+      call: { id: `i${issue}`, name: "composio_run_action", arguments: { action: "GITHUB_UPDATE_ISSUE", arguments: { issue_number: issue } } },
+    });
+    update(1, "ERROR: GITHUB_UPDATE_ISSUE failed: 502");
+    recordRunAction(h.connection.db, {
+      orgId: h.orgId, runId: h.runId, taskId: task.task_id, output: "ERROR: reply lost",
+      call: { id: "multi", name: "composio_run_action", arguments: { action: "GITHUB_UPDATE_ISSUE", arguments: { issue_number: 3, state: "closed" } } },
+    });
+    recordRunAction(h.connection.db, {
+      orgId: h.orgId, runId: h.runId, taskId: task.task_id, output: "{\"successful\":true}",
+      call: { id: "multi2", name: "composio_run_action", arguments: { action: "GITHUB_UPDATE_ISSUE", arguments: { state: "closed", issue_number: 3 } } },
+    });
+    update(2, "{\"successful\":true}");
+    expect(unresolvedActionFailure(h.connection.db, h.runId)).toBe("The session ended with 1 failed action it did not redo: composio_run_action (ERROR: GITHUB_UPDATE_ISSUE failed: 502)");
+    // A retry that confirms the lost reply, with the same input in another key order, redoes issue 1.
+    recordRunAction(h.connection.db, {
+      orgId: h.orgId, runId: h.runId, taskId: task.task_id, output: "{\"successful\":true}",
+      call: { id: "retry", name: "composio_run_action", arguments: { arguments: { issue_number: 1 }, confirm_not_done: true, action: "GITHUB_UPDATE_ISSUE" } },
+    });
+    expect(unresolvedActionFailure(h.connection.db, h.runId)).toBeNull();
+    h.connection.close();
+  });
+});

tests/unit/domain-services.test.ts

@@ -35,7 +35,7 @@ describe("domain services", () => {
     const connection = createConnection();
 
     const rows = connection.db.prepare("select version from schema_migrations order by version asc").all() as Array<{ version: number }>;
-    expect(rows.map((row) => row.version)).toEqual([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91]);
+    expect(rows.map((row) => row.version)).toEqual([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92]);
     connection.close();
   });