@deksden-com/dd-flow-cli 0.9.0-beta.63 → 0.9.0-beta.64

package/dist/services/work-registry.js

@@ -91,15 +91,17 @@ export function assertWorkLifecycleOwner(context, input) {
     throw new AppError("lifecycle_caller_mismatch", "Work lifecycle command was issued by a Session that does not own the running Work", 1, payload);
 }
 export function createChildWork(context, input) {
-    const parent = requireWork(context, input.parentWorkId);
-    assertRunMutationAllowed(context, parent.project_id, parent.run_id);
-    if (parent.status !== "running")
-        throw new AppError("invalid_work_state", "Child Work requires a running parent", 2, { parent_work_id: parent.work_id, status: parent.status });
-    const id = nextWorkId(context, parent.project_id, input.slug);
-    const now = context.now();
-    context.db.run(`INSERT INTO works (${workColumns}) VALUES (?, ?, ?, ?, ?, ?, ?, ?, '[]', 'created', NULL, ?, NULL, ?, NULL)`, [id, parent.project_id, parent.run_id, parent.work_id, input.task, input.launchPolicy, input.resultSchema ?? null, input.payload ? JSON.stringify(input.payload) : null, now, now]);
-    refreshRunWorkProjection(context, parent.project_id, parent.run_id);
-    return requireWork(context, id);
+    return context.db.writeTransaction(() => {
+        const parent = requireWork(context, input.parentWorkId);
+        assertRunMutationAllowed(context, parent.project_id, parent.run_id);
+        if (parent.status !== "running")
+            throw new AppError("invalid_work_state", "Child Work requires a running parent", 2, { parent_work_id: parent.work_id, status: parent.status });
+        const id = nextWorkId(context, parent.project_id, input.slug);
+        const now = context.now();
+        context.db.run(`INSERT INTO works (${workColumns}) VALUES (?, ?, ?, ?, ?, ?, ?, ?, '[]', 'created', NULL, ?, NULL, ?, NULL)`, [id, parent.project_id, parent.run_id, parent.work_id, input.task, input.launchPolicy, input.resultSchema ?? null, input.payload ? JSON.stringify(input.payload) : null, now, now]);
+        refreshRunWorkProjection(context, parent.project_id, parent.run_id);
+        return requireWork(context, id);
+    });
 }
 /** Validate a proposed batch before PLAN accepts it, without registering Work. */
 export function validateWorkBatchFile(file) {
@@ -122,10 +124,6 @@ export function validateWorkBatchFile(file) {
 }
 export function addWorkBatch(context, input) {
     ensureWorkRegistry(context);
-    const parent = requireWork(context, input.parentWorkId);
-    assertRunMutationAllowed(context, parent.project_id, parent.run_id);
-    if (parent.status !== "running")
-        throw new AppError("invalid_work_state", "Batch parent must be running", 2, { work_id: parent.work_id, status: parent.status });
     const parsed = readJson(input.file);
     if (!Array.isArray(parsed.works) || parsed.works.length === 0)
         throw new AppError("validation", "work batch requires non-empty works", 2);
@@ -133,18 +131,21 @@ export function addWorkBatch(context, input) {
     const keys = new Set(items.map((item) => item.key));
     if (keys.size !== items.length)
         throw new AppError("validation", "work batch keys must be unique", 2);
-    const existing = context.db.all(`SELECT ${workColumns} FROM works WHERE project_id = ? AND run_id = ?`, [parent.project_id, parent.run_id]);
-    const existingIds = new Set(existing.map((row) => row.work_id));
-    for (const item of items)
-        for (const dependency of item.depends_on ?? [])
-            if (!keys.has(dependency) && !existingIds.has(dependency))
-                throw new AppError("validation", "work dependency is unknown", 2, { key: item.key, dependency });
-    for (const item of items)
-        if (item.parent && !keys.has(item.parent) && !existingIds.has(item.parent))
-            throw new AppError("validation", "work parent is unknown", 2, { key: item.key, parent: item.parent });
-    const now = context.now();
-    context.db.exec("SAVEPOINT work_batch");
-    try {
+    return context.db.writeTransaction(() => {
+        const parent = requireWork(context, input.parentWorkId);
+        assertRunMutationAllowed(context, parent.project_id, parent.run_id);
+        if (parent.status !== "running")
+            throw new AppError("invalid_work_state", "Batch parent must be running", 2, { work_id: parent.work_id, status: parent.status });
+        const existing = context.db.all(`SELECT ${workColumns} FROM works WHERE project_id = ? AND run_id = ?`, [parent.project_id, parent.run_id]);
+        const existingIds = new Set(existing.map((row) => row.work_id));
+        for (const item of items)
+            for (const dependency of item.depends_on ?? [])
+                if (!keys.has(dependency) && !existingIds.has(dependency))
+                    throw new AppError("validation", "work dependency is unknown", 2, { key: item.key, dependency });
+        for (const item of items)
+            if (item.parent && !keys.has(item.parent) && !existingIds.has(item.parent))
+                throw new AppError("validation", "work parent is unknown", 2, { key: item.key, parent: item.parent });
+        const now = context.now();
         const allocated = nextWorkIds(context, parent.project_id, items.map((item) => item.key));
         const ids = new Map(items.map((item, index) => [item.key, allocated[index]]));
         const resolve = (value) => ids.get(value) ?? value;
@@ -157,14 +158,8 @@ export function addWorkBatch(context, input) {
         for (const item of proposed)
             appendFlowRunTimelineEvent(context, parent.project_id, parent.run_id, { type: "work_materialized", work_id: item.id, parent_work_id: item.parentId, depends_on: item.dependencies, launch_policy: item.launch_policy ?? "reuse_allowed" });
         refreshRunWorkProjection(context, parent.project_id, parent.run_id);
-        context.db.exec("RELEASE work_batch");
         return { ok: true, work_ids: Object.fromEntries(ids) };
-    }
-    catch (error) {
-        context.db.exec("ROLLBACK TO work_batch");
-        context.db.exec("RELEASE work_batch");
-        throw error;
-    }
+    });
 }
 export function listWorks(context, input) {
     ensureWorkRegistry(context);
@@ -216,6 +211,7 @@ function receiptIntentPath(resultPath, sessionId) {
 function receiptFileHash(file) {
     return fs.existsSync(file) ? crypto.createHash("sha256").update(fs.readFileSync(file)).digest("hex") : null;
 }
+function sha256(contents) { return crypto.createHash("sha256").update(contents).digest("hex"); }
 function publishReceiptFile(file, contents) {
     const temporary = `${file}.tmp-${crypto.randomUUID()}`;
     try {
@@ -240,9 +236,44 @@ function publishReceiptFile(file, contents) {
         fs.rmSync(temporary, { force: true });
     }
 }
+/** Rebuild prompt/context projections from a committed Work-start receipt. */
+export function materializePendingWorkStarts(context, projectId, runId) {
+    const runHome = requireRunHome(requireRun(context, projectId, runId));
+    context.db.writeTransaction(() => {
+        const rows = context.db.all("SELECT ws.id, ws.work_id, ws.prompt_path, ws.start_receipt_json FROM work_sessions ws JOIN works w ON w.work_id = ws.work_id WHERE w.project_id = ? AND w.run_id = ? AND ws.status = 'running' AND ws.start_receipt_json IS NOT NULL", [projectId, runId]);
+        for (const row of rows) {
+            // The same coordinator Session can move on to a different stage packet.
+            // A start receipt is not authority to overwrite that later packet.
+            let receipt;
+            try {
+                receipt = JSON.parse(row.start_receipt_json);
+            }
+            catch {
+                throw new AppError("work_start_receipt_invalid", "Committed Work start receipt is not valid JSON", 1, { work_id: row.work_id });
+            }
+            const originalPromptPath = path.join(runHome, "works", row.work_id, ...(receipt.generation && receipt.generation > 0 ? ["starts", row.id] : []), "prompt.md");
+            if (path.resolve(row.prompt_path) !== path.resolve(originalPromptPath))
+                continue;
+            if (receipt.version !== 1 || typeof receipt.prompt !== "string" || typeof receipt.context !== "string" || receipt.prompt_sha256 !== sha256(receipt.prompt) || receipt.context_sha256 !== sha256(receipt.context))
+                throw new AppError("work_start_receipt_invalid", "Committed Work start receipt does not match its retained bytes", 1, { work_id: row.work_id });
+            publishStartReceiptFile(row.prompt_path, receipt.prompt, String(receipt.prompt_sha256));
+            publishStartReceiptFile(path.join(path.dirname(row.prompt_path), "context.json"), receipt.context, String(receipt.context_sha256));
+        }
+    });
+}
+function publishStartReceiptFile(file, contents, expectedHash) {
+    const actual = receiptFileHash(file);
+    if (actual === expectedHash)
+        return;
+    if (actual !== null) {
+        throw new AppError("work_start_projection_conflict", "Work start packet differs from its committed receipt; preserve it for recovery", 1, { path: file, expected_sha256: expectedHash, actual_sha256: actual });
+    }
+    fs.mkdirSync(path.dirname(file), { recursive: true });
+    publishReceiptFile(file, contents);
+}
 /** Replay only a prepared publication whose exact receipt was committed in SQLite. */
 export function materializePendingWorkResults(context, projectId, runId) {
-    context.db.exec("BEGIN IMMEDIATE");
+    context.db.beginWriteTransaction();
     try {
         const rows = context.db.all(`SELECT w.work_id, w.result, w.status, ws.id, ws.result_path FROM works w JOIN work_sessions ws ON ws.work_id = w.work_id AND ws.status = w.status WHERE w.project_id = ? AND w.run_id = ? AND w.parent_work_id IS NOT NULL AND w.result IS NOT NULL AND w.status IN ('completed', 'failed', 'cancelled') AND ws.result_path IS NOT NULL`, [projectId, runId]);
         for (const row of rows) {
@@ -280,33 +311,37 @@ export function materializePendingWorkResults(context, projectId, runId) {
     }
 }
 export function mutateWorkDeps(context, input) {
-    const work = requireWork(context, input.workId);
-    if (input.action === "list")
+    if (input.action === "list") {
+        const work = requireWork(context, input.workId);
         return { work_id: work.work_id, depends_on: parseDependencies(work) };
-    assertRunMutationAllowed(context, work.project_id, work.run_id);
-    if (work.status !== "created")
-        throw new AppError("invalid_work_state", "Dependencies change only while Work is created", 2);
-    const dependencies = new Set(parseDependencies(work));
-    if (input.action === "clear")
-        dependencies.clear();
-    for (const id of input.on ?? []) {
-        const dependency = requireWork(context, id);
-        if (dependency.run_id !== work.run_id)
-            throw new AppError("validation", "Dependencies must stay in one RUN", 2);
-        if (input.action === "add")
-            dependencies.add(dependency.work_id);
-        else
-            dependencies.delete(dependency.work_id);
-    }
-    if (dependencies.has(work.work_id))
-        throw new AppError("validation", "Work cannot depend on itself", 2);
-    assertNoCycles(context.db.all(`SELECT ${workColumns} FROM works WHERE project_id = ? AND run_id = ?`, [work.project_id, work.run_id]).map((row) => ({ id: row.work_id, dependencies: row.work_id === work.work_id ? [...dependencies] : parseDependencies(row) })));
-    context.db.run("UPDATE works SET depends_on_json = ?, updated_at = ? WHERE work_id = ?", [JSON.stringify([...dependencies]), context.now(), work.work_id]);
-    refreshRunWorkProjection(context, work.project_id, work.run_id);
-    return { work_id: work.work_id, depends_on: [...dependencies] };
+    }
+    return context.db.writeTransaction(() => {
+        const work = requireWork(context, input.workId);
+        assertRunMutationAllowed(context, work.project_id, work.run_id);
+        if (work.status !== "created")
+            throw new AppError("invalid_work_state", "Dependencies change only while Work is created", 2);
+        const dependencies = new Set(parseDependencies(work));
+        if (input.action === "clear")
+            dependencies.clear();
+        for (const id of input.on ?? []) {
+            const dependency = requireWork(context, id);
+            if (dependency.run_id !== work.run_id)
+                throw new AppError("validation", "Dependencies must stay in one RUN", 2);
+            if (input.action === "add")
+                dependencies.add(dependency.work_id);
+            else
+                dependencies.delete(dependency.work_id);
+        }
+        if (dependencies.has(work.work_id))
+            throw new AppError("validation", "Work cannot depend on itself", 2);
+        assertNoCycles(context.db.all(`SELECT ${workColumns} FROM works WHERE project_id = ? AND run_id = ?`, [work.project_id, work.run_id]).map((row) => ({ id: row.work_id, dependencies: row.work_id === work.work_id ? [...dependencies] : parseDependencies(row) })));
+        context.db.run("UPDATE works SET depends_on_json = ?, updated_at = ? WHERE work_id = ?", [JSON.stringify([...dependencies]), context.now(), work.work_id]);
+        refreshRunWorkProjection(context, work.project_id, work.run_id);
+        return { work_id: work.work_id, depends_on: [...dependencies] };
+    });
 }
-export function deleteWork(context, id) { const work = requireWork(context, id); assertRunMutationAllowed(context, work.project_id, work.run_id); const canonical = work.work_id; if (work.status !== "created" || work.started_at || context.db.get("SELECT 1 FROM works WHERE parent_work_id = ? LIMIT 1", [canonical]) || context.db.get("SELECT 1 FROM works WHERE depends_on_json LIKE ? LIMIT 1", [`%${canonical}%`]))
-    throw new AppError("invalid_work_state", "Only an unstarted unreferenced Work may be deleted", 2); context.db.run("DELETE FROM works WHERE work_id = ?", [canonical]); refreshRunWorkProjection(context, work.project_id, work.run_id); return { ok: true, deleted: canonical }; }
+export function deleteWork(context, id) { return context.db.writeTransaction(() => { const work = requireWork(context, id); assertRunMutationAllowed(context, work.project_id, work.run_id); const canonical = work.work_id; if (work.status !== "created" || work.started_at || context.db.get("SELECT 1 FROM works WHERE parent_work_id = ? LIMIT 1", [canonical]) || context.db.get("SELECT 1 FROM works WHERE depends_on_json LIKE ? LIMIT 1", [`%${canonical}%`]))
+    throw new AppError("invalid_work_state", "Only an unstarted unreferenced Work may be deleted", 2); context.db.run("DELETE FROM works WHERE work_id = ?", [canonical]); refreshRunWorkProjection(context, work.project_id, work.run_id); return { ok: true, deleted: canonical }; }); }
 export function shortWorkId(id) { return /^WRK-\d{3,}(?:-|$)/.exec(id)?.[0]?.replace(/-$/, "") ?? id; }
 /**
  * Return the engine-bound command, never a PATH-dependent `dd-flow` token.
@@ -329,6 +364,15 @@ export function assertWorkLaunchReady(context, id) {
     return work;
 }
 export function startWork(context, id, input) {
+    if (input.hookEventId && !context.env.DD_FLOW_CURRENT_INVOCATION) {
+        const retainedWork = requireWork(context, id);
+        const retainedRun = requireRun(context, retainedWork.project_id, retainedWork.run_id);
+        if (input.projectRoot && resolveProjectRoot(input.projectRoot) !== resolveProjectRoot(retainedRun.project_root))
+            throw new AppError("project_mismatch", "Work start belongs to another project", 1);
+        const reply = recoverCommittedWorkStart(context, null, input.hookEventId, recoveryGuard(context, retainedWork.project_id, retainedWork.run_id)?.generation ?? 0, retainedWork.work_id);
+        if (reply)
+            return reply;
+    }
     const work = assertWorkLaunchReady(context, id);
     const run = requireRun(context, work.project_id, work.run_id);
     if (input.projectRoot && resolveProjectRoot(input.projectRoot) !== resolveProjectRoot(run.project_root))
@@ -345,16 +389,9 @@ export function startWork(context, id, input) {
             hookEventId = findRecentMatchingHookEvent(context, { projectId: work.project_id, matchKey: workStartMatchKey(work.work_id, run.project_root, recoveryId), errorCode: "trusted_session_binding_required", operation: "work start" }).eventKey;
         }
     }
-    const identity = claimWorkStartHookEvent(context, { projectId: work.project_id, eventKey: hookEventId, workId: work.work_id, projectRoot: run.project_root, ...(recoveryId ? { recoveryId } : {}) });
-    let started;
-    try {
-        started = startBoundWork(context, work, run, identity, hookEventId);
-    }
-    catch (error) {
-        context.db.run("UPDATE hook_events SET status = 'observed', claimed_at = NULL WHERE project_id = ? AND event_key = ? AND status = 'claimed'", [work.project_id, hookEventId]);
-        throw error;
-    }
-    appendFlowRunTimelineEvent(context, work.project_id, work.run_id, { type: "work_started", work_id: work.work_id, session_id: identity.sessionId, agent_id: identity.agentId });
+    const eventKey = hookEventId;
+    const started = startBoundWork(context, work, run, () => claimWorkStartHookEvent(context, { projectId: work.project_id, eventKey, workId: work.work_id, projectRoot: run.project_root, ...(recoveryId ? { recoveryId } : {}) }), eventKey);
+    appendFlowRunTimelineEvent(context, work.project_id, work.run_id, { type: "work_started", work_id: work.work_id, work_session_id: started.work_session_id });
     return started;
 }
 /** Binds the already-running coordinator of a stage to its trusted hook Session. */
@@ -378,7 +415,7 @@ export function bindRunningWorkSession(context, input) {
                 throw new AppError("handoff_session_mismatch", "The active Work is already bound to a different Session", 1, { work_id: work.work_id });
             const now = context.now();
             const id = `WS-${crypto.randomUUID()}`;
-            context.db.exec("BEGIN IMMEDIATE");
+            context.db.beginWriteTransaction();
             try {
                 context.db.run("UPDATE work_sessions SET status = 'completed', completed_at = ?, updated_at = ? WHERE id = ?", [now, now, active.id]);
                 const stillRunning = context.db.get("SELECT 1 FROM work_sessions WHERE session_id = ? AND status = 'running' LIMIT 1", [active.session_id]);
@@ -400,7 +437,7 @@ export function bindRunningWorkSession(context, input) {
     }
     const now = context.now();
     const id = `WS-${crypto.randomUUID()}`;
-    context.db.exec("BEGIN IMMEDIATE");
+    context.db.beginWriteTransaction();
     try {
         bindSession(context, work, run, identity, now);
         context.db.run("INSERT INTO work_sessions (id, work_id, session_id, hook_event_id, status, prompt_path, result_path, created_at, updated_at, completed_at) VALUES (?, ?, ?, ?, 'running', ?, ?, ?, ?, NULL)", [id, work.work_id, identity.sessionId, input.hookEventId, input.promptPath, input.resultPath ?? null, now, now]);
@@ -445,65 +482,105 @@ export function startStageCoordinatorWork(context, input) {
     const run = requireRun(context, work.project_id, work.run_id);
     if (input.projectRoot && resolveProjectRoot(input.projectRoot) !== resolveProjectRoot(run.project_root))
         throw new AppError("project_mismatch", "stage coordinator project root does not match its RUN", 1, { work_id: work.work_id });
-    const identity = claimStageStartHookEvent(context, { projectId: work.project_id, eventKey: input.hookEventId, runId: work.run_id, stage: input.stage, projectRoot: run.project_root, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}) });
-    const binding = startBoundWork(context, work, run, identity, input.hookEventId);
-    const sessionId = String(binding.session_binding.session_id ?? "");
-    context.db.run("UPDATE sessions SET current_stage = ?, updated_at = ? WHERE project_id = ? AND session_id = ?", [input.stage, context.now(), work.project_id, sessionId]);
+    const binding = startBoundWork(context, work, run, () => claimStageStartHookEvent(context, { projectId: work.project_id, eventKey: input.hookEventId, runId: work.run_id, stage: input.stage, projectRoot: run.project_root, ...(input.contextSha256 ? { contextSha256: input.contextSha256 } : {}) }), input.hookEventId, input.stage);
     refreshRunSessionProjection(context, work.project_id, work.run_id);
     return binding;
 }
-/** Stage entry points call this only after their own trusted hook claim. */
-export function startBoundWork(context, work, run, identity, hookEventId) {
-    assertRecoveryHookCurrent(context, work.project_id, work.run_id, identity);
+/** Claim, command authority and binding commit together; only packet projection follows. */
+export function startBoundWork(context, work, run, claimIdentity, hookEventId, stage) {
     const generation = assertRunMutationAllowed(context, work.project_id, work.run_id);
-    if (work.status !== "created")
-        throw new AppError("invalid_work_state", "Work is not created", 2, { work_id: work.work_id, status: work.status });
-    if (fs.existsSync(path.join(requireRunHome(run), "works", work.work_id, ".retry.json")))
-        throw new AppError("work_retry_incomplete", "Complete the interrupted Work retry before starting its next attempt", 1, { work_id: work.work_id });
-    const now = context.now();
-    const linkId = `WS-${crypto.randomUUID()}`;
     const directory = path.join(requireRunHome(run), "works", work.work_id);
-    fs.mkdirSync(directory, { recursive: true });
-    const promptPath = path.join(directory, "prompt.md");
-    const resultPath = path.join(directory, "result.json");
-    const contextPath = path.join(directory, "context.json");
-    const dependencyResults = parseDependencies(work).map((dependency) => context.db.get("SELECT work_id, result FROM works WHERE work_id = ?", [dependency])).filter(Boolean);
-    const recovery = recoveryPromptForWork(context, work.work_id);
-    let prompt = renderWorkerPrompt(context, work, run, dependencyResults);
-    if (recovery)
-        prompt = `${recoveryPrompt(recovery)}\n\n${prompt}`;
-    const payload = parsePayload(work);
-    const readOnly = payload?.read_only === true;
-    const workContext = { schema_id: "dd-flow/work-context@1", run_id: work.run_id, work_id: work.work_id, parent_work_id: work.parent_work_id, project_root: run.project_root, workspace_root: run.workspace_root, run_root: requireRunHome(run), depends_on: parseDependencies(work), launch_policy: work.launch_policy, result_schema: work.result_schema, read_only: readOnly, ...(recovery ? { recovery } : {}), ...(readOnly ? { workspace_fingerprint: workspaceFingerprint(run.workspace_root) } : {}) };
-    const token = crypto.randomUUID();
-    const promptCandidate = `${promptPath}.${token}.tmp`;
-    const contextCandidate = `${contextPath}.${token}.tmp`;
-    fs.writeFileSync(promptCandidate, prompt);
-    fs.writeFileSync(contextCandidate, `${JSON.stringify(workContext, null, 2)}\n`);
-    context.db.exec("BEGIN IMMEDIATE");
+    const linkId = `WS-${crypto.randomUUID()}`;
+    // Recovery can reopen the same Work without retiring its historical packet.
+    // Give each new-generation start its own immutable publication destination.
+    const packetDirectory = generation > 0 ? path.join(directory, "starts", linkId) : directory;
+    const promptPath = path.join(packetDirectory, "prompt.md"), resultPath = path.join(directory, "result.json");
+    // The expensive filesystem observation is not performed under the writer lock.
+    const readOnly = parsePayload(work)?.read_only === true;
+    const fingerprint = readOnly ? workspaceFingerprint(run.workspace_root) : undefined;
+    let binding;
     try {
-        assertRunMutationAllowed(context, work.project_id, work.run_id, generation);
-        bindSession(context, work, run, identity, now);
-        if (work.parent_work_id && work.launch_policy === "reuse_allowed")
-            context.db.run("UPDATE work_sessions SET status = 'completed', completed_at = ?, updated_at = ? WHERE session_id = ? AND status = 'running' AND work_id <> ?", [now, now, identity.sessionId, work.work_id]);
-        const claimed = context.db.run("UPDATE works SET status = 'running', started_at = COALESCE(started_at, ?), updated_at = ? WHERE work_id = ? AND status = 'created'", [now, now, work.work_id]);
-        if (claimed.changes !== 1)
-            throw new AppError("conflict", "Work was claimed concurrently", 1, { work_id: work.work_id });
-        context.db.run(`INSERT INTO work_sessions (id, work_id, session_id, hook_event_id, status, prompt_path, result_path, created_at, updated_at, completed_at) VALUES (?, ?, ?, ?, 'running', ?, ?, ?, ?, NULL)`, [linkId, work.work_id, identity.sessionId, hookEventId, promptPath, resultPath, now, now]);
-        fs.renameSync(promptCandidate, promptPath);
-        fs.renameSync(contextCandidate, contextPath);
-        if (recovery)
-            activateRecoveryWork(context, work.work_id);
-        context.db.exec("COMMIT");
+        binding = context.db.writeTransaction(() => {
+            const current = requireWork(context, work.work_id);
+            assertRunMutationAllowed(context, current.project_id, current.run_id, generation);
+            if (current.status !== "created")
+                throw new AppError("conflict", "Work was claimed concurrently", 1, { work_id: current.work_id, status: current.status });
+            if (Object.keys(current).some(key => current[key] !== work[key]))
+                throw new AppError("work_not_ready", "Work inputs changed while preparing its start packet", 2, { work_id: current.work_id });
+            if (!isReady(context, current))
+                throw new AppError("work_not_ready", "Work dependencies are not ready", 2, { work_id: current.work_id, blockers: readinessBlockers(context, current) });
+            if (fs.existsSync(path.join(directory, ".retry.json")))
+                throw new AppError("work_retry_incomplete", "Complete the interrupted Work retry before starting its next attempt", 1, { work_id: work.work_id });
+            const identity = claimIdentity();
+            assertRecoveryHookCurrent(context, current.project_id, current.run_id, identity);
+            const dependencyResults = parseDependencies(current).map(dependency => context.db.get("SELECT work_id, result FROM works WHERE work_id = ?", [dependency])).filter(Boolean);
+            const recovery = recoveryPromptForWork(context, current.work_id);
+            let prompt = renderWorkerPrompt(context, current, run, dependencyResults);
+            if (recovery)
+                prompt = `${recoveryPrompt(recovery)}\n\n${prompt}`;
+            const workContext = { schema_id: "dd-flow/work-context@1", run_id: current.run_id, work_id: current.work_id, parent_work_id: current.parent_work_id, project_root: run.project_root, workspace_root: run.workspace_root, run_root: requireRunHome(run), depends_on: parseDependencies(current), launch_policy: current.launch_policy, result_schema: current.result_schema, read_only: readOnly, ...(recovery ? { recovery } : {}), ...(readOnly ? { workspace_fingerprint: fingerprint } : {}) };
+            const contextBytes = `${JSON.stringify(workContext, null, 2)}\n`;
+            const response = { ok: true, work_id: current.work_id, work_session_id: linkId, worker_prompt_markdown: prompt, prompt_path: promptPath, ...(recovery ? { recovery } : {}), session_binding: { source: "PreToolUse", session: { harness_id: identity.harness, session_id: identity.nativeSessionId } } };
+            const receipt = JSON.stringify({ version: 1, prompt, context: contextBytes, prompt_sha256: sha256(prompt), context_sha256: sha256(contextBytes), response, response_sha256: sha256(JSON.stringify(response)), generation, invocation_id: context.env.DD_FLOW_CURRENT_INVOCATION ?? null });
+            const now = context.now();
+            bindSession(context, current, run, identity, now);
+            if (current.parent_work_id && current.launch_policy === "reuse_allowed")
+                context.db.run("UPDATE work_sessions SET status = 'completed', completed_at = ?, updated_at = ? WHERE session_id = ? AND status = 'running' AND work_id <> ?", [now, now, identity.sessionId, current.work_id]);
+            if (context.db.run("UPDATE works SET status = 'running', started_at = COALESCE(started_at, ?), updated_at = ? WHERE work_id = ? AND status = 'created'", [now, now, current.work_id]).changes !== 1)
+                throw new AppError("conflict", "Work was claimed concurrently", 1, { work_id: current.work_id });
+            context.db.run("INSERT INTO work_sessions (id, work_id, session_id, hook_event_id, status, prompt_path, result_path, start_receipt_json, created_at, updated_at, completed_at) VALUES (?, ?, ?, ?, 'running', ?, ?, ?, ?, ?, NULL)", [linkId, current.work_id, identity.sessionId, hookEventId, promptPath, resultPath, receipt, now, now]);
+            if (stage)
+                context.db.run("UPDATE sessions SET current_stage = ?, updated_at = ? WHERE project_id = ? AND session_id = ?", [stage, now, current.project_id, identity.sessionId]);
+            if (recovery)
+                activateRecoveryWork(context, current.work_id);
+            return response;
+        });
     }
     catch (error) {
-        context.db.exec("ROLLBACK");
-        fs.rmSync(promptCandidate, { force: true });
-        fs.rmSync(contextCandidate, { force: true });
+        // This transaction publishes only SQL authority, never files or provider calls.
+        if (error instanceof AppError && error.code === "storage_write_failed" && error.details.transaction_rolled_back === true)
+            error.details.effect = "no_effect";
+        if (error instanceof AppError && error.details.effect === "no_effect" && !context.env.DD_FLOW_CURRENT_INVOCATION) {
+            let retry = false;
+            try {
+                retry = context.db.writeTransaction(() => {
+                    if (context.db.get("SELECT 1 FROM hook_events WHERE project_id = ? AND json_extract(outcome_json, '$.work_id') = ? AND json_extract(outcome_json, '$.generation') = ? AND json_extract(outcome_json, '$.storage_retry') = 1", [work.project_id, work.work_id, generation]))
+                        return false;
+                    return context.db.run("UPDATE hook_events SET outcome_json = ? WHERE project_id = ? AND event_key = ? AND status = 'observed' AND outcome_json IS NULL", [JSON.stringify({ work_id: work.work_id, run_id: work.run_id, generation, effect: "no_effect", disposition: "retrying", storage_retry: 1, deadline: Date.now() + 60000 }), work.project_id, hookEventId]).changes === 1;
+                });
+            }
+            catch (persistence) {
+                error.details.retry_persistence_error = String(persistence);
+            }
+            if (retry)
+                return startBoundWork(context, work, run, claimIdentity, hookEventId, stage);
+        }
         throw error;
     }
-    refreshRunWorkProjection(context, work.project_id, work.run_id);
-    return { ok: true, work_id: work.work_id, work_session_id: linkId, worker_prompt_markdown: prompt, prompt_path: promptPath, ...(recovery ? { recovery } : {}), session_binding: { source: "PreToolUse", session: { harness_id: identity.harness, session_id: identity.nativeSessionId } } };
+    try {
+        materializePendingWorkStarts(context, work.project_id, work.run_id);
+        refreshRunWorkProjection(context, work.project_id, work.run_id);
+    }
+    catch (error) {
+        if (error instanceof AppError) {
+            error.details.effect = "committed";
+            throw error;
+        }
+        throw new AppError("work_start_publication_failed", error instanceof Error ? error.message : String(error), 1, { effect: "committed", work_id: work.work_id, hook_event_id: hookEventId, cause: { code: error?.code ?? null } });
+    }
+    return binding;
+}
+/** Read the exact committed reply, never bind another Session or regenerate a packet. */
+export function recoverCommittedWorkStart(context, invocationId, eventKey, generation, workId) {
+    const row = context.db.get("SELECT ws.start_receipt_json, w.project_id, w.run_id FROM work_sessions ws JOIN works w ON w.work_id = ws.work_id WHERE ws.start_receipt_json IS NOT NULL AND ws.hook_event_id = ? AND json_extract(ws.start_receipt_json, '$.invocation_id') IS ? AND (? IS NULL OR w.work_id = ?)", [eventKey, invocationId, workId ?? null, workId ?? null]);
+    if (!row)
+        return undefined;
+    const receipt = JSON.parse(row.start_receipt_json);
+    if (receipt.generation !== generation || !receipt.response || receipt.response_sha256 !== sha256(JSON.stringify(receipt.response)))
+        throw new AppError("work_start_receipt_invalid", "Committed reply identity or checksum differs", 1, { invocation_id: invocationId });
+    assertRunMutationAllowed(context, row.project_id, row.run_id, generation);
+    materializePendingWorkStarts(context, row.project_id, row.run_id);
+    return receipt.response;
 }
 /** Root coordinators have no worker checks: settle them inside the stage transaction. */
 export function finishStageCoordinatorWork(context, id, result) {
@@ -532,7 +609,7 @@ export function retryWork(context, id, reason) {
     const directory = path.join(requireRunHome(run), "works", work.work_id);
     const journal = path.join(directory, ".retry.json");
     let archive = "";
-    context.db.exec("BEGIN IMMEDIATE");
+    context.db.beginWriteTransaction();
     try {
         assertRunMutationAllowed(context, work.project_id, work.run_id);
         const current = requireWork(context, work.work_id);
@@ -637,7 +714,7 @@ async function settle(context, id, status, result, progress, retry) {
         }
     }
     const now = context.now();
-    context.db.exec("BEGIN IMMEDIATE");
+    context.db.beginWriteTransaction();
     try {
         assertWorkSettlementAllowed(context, work.project_id, work.run_id, work.work_id);
         if (link && !context.db.get("SELECT 1 FROM work_sessions WHERE id = ? AND status = 'running'", [link.id]))

package/dist/storage/database.js

@@ -147,6 +147,7 @@ export function ensureVnextWorkStorage(db) {
       status TEXT NOT NULL,
       prompt_path TEXT NOT NULL,
       result_path TEXT,
+      start_receipt_json TEXT,
       created_at TEXT NOT NULL,
       updated_at TEXT NOT NULL,
       completed_at TEXT,
@@ -196,6 +197,13 @@ export function ensureVnextWorkStorage(db) {
     );
     CREATE INDEX IF NOT EXISTS idx_check_receipts_run ON check_receipts(project_id, run_id, started_at);
   `);
+    // A committed Work start is authoritative even if publishing its prompt
+    // files is interrupted after commit.  Keep the exact bytes for repair.
+    db.writeTransaction(() => {
+        const columns = db.all("PRAGMA table_info(work_sessions)");
+        if (!columns.some((column) => column.name === "start_receipt_json"))
+            db.execSchema("ALTER TABLE work_sessions ADD COLUMN start_receipt_json TEXT");
+    });
 }
 /** Durable intent for a provider-root external Work.  It is deliberately
  * separate from `work_sessions`: only a trusted hook may bind that Session. */
@@ -265,16 +273,7 @@ export function getDatabase(ddFlowHome, mode = "initialize") {
     if (mode === "read_existing")
         db.exec("PRAGMA query_only = ON");
     db.exec("PRAGMA foreign_keys = ON");
-    return {
-        path: dbPath,
-        writable: mode !== "read_existing",
-        close: () => db.close?.(),
-        exec: (sql) => db.exec(sql),
-        execSchema: (sql) => executeStoreSchema(db, sql),
-        run: (sql, params = []) => db.prepare(sql).run(...params),
-        get: (sql, params = []) => db.prepare(sql).get(...params),
-        all: (sql, params = []) => db.prepare(sql).all(...params)
-    };
+    return databaseFacade(dbPath, mode !== "read_existing", db);
 }
 /**
  * Opens the small host-wide resource registry.  It is intentionally separate
@@ -351,16 +350,10 @@ export function getResourceDatabase(resourceHome) {
         db.close();
         throw error;
     }
-    const database = {
-        path: dbPath,
-        writable: true,
-        close: () => { resourceDatabases.delete(dbPath); db.close?.(); },
-        exec: (sql) => db.exec(sql),
-        execSchema: (sql) => executeStoreSchema(db, sql),
-        run: (sql, params = []) => db.prepare(sql).run(...params),
-        get: (sql, params = []) => db.prepare(sql).get(...params),
-        all: (sql, params = []) => db.prepare(sql).all(...params)
-    };
+    const database = databaseFacade(dbPath, true, db, () => {
+        resourceDatabases.delete(dbPath);
+        db.close?.();
+    });
     resourceDatabases.set(dbPath, database);
     return database;
 }
@@ -370,11 +363,108 @@ function emptyReadOnlyDatabase(dbPath) {
         writable: false,
         exec: () => { throw new AppError("read_only_storage", "Read-only access cannot write absent dd-flow storage", 1, { path: dbPath }); },
         execSchema: () => { throw new AppError("read_only_storage", "Read-only access cannot write absent dd-flow storage", 1, { path: dbPath }); },
+        beginWriteTransaction: () => { throw new AppError("read_only_storage", "Read-only access cannot start a write transaction", 1, { path: dbPath }); },
+        writeTransaction: () => { throw new AppError("read_only_storage", "Read-only access cannot start a write transaction", 1, { path: dbPath }); },
         run: () => { throw new AppError("read_only_storage", "Read-only access cannot write absent dd-flow storage", 1, { path: dbPath }); },
         get: () => undefined,
         all: () => []
     };
 }
+function databaseFacade(dbPath, writable, db, close = () => db.close?.()) {
+    let reserved = false;
+    let poisoned = false;
+    let savepoint = 0;
+    const assertUsable = () => { if (poisoned)
+        throw new AppError("write_transaction_unresolved", "Transaction cleanup failed; reopen the connection before continuing", 1); };
+    const beginWriteTransaction = () => {
+        assertUsable();
+        if (!writable)
+            throw new AppError("read_only_storage", "Read-only access cannot start a write transaction", 1, { path: dbPath });
+        db.exec("BEGIN IMMEDIATE");
+        reserved = true;
+    };
+    const rollback = (sql, original) => {
+        try {
+            db.exec(sql);
+        }
+        catch (cleanup) {
+            poisoned = true;
+            if (original instanceof AppError)
+                throw new AppError(original.code, original.message, original.exitCode, { ...original.details, transaction_cleanup_error: errorDetails(cleanup) });
+            throw new AppError("write_transaction_failed", "Write transaction failed and SQLite cleanup also failed", 1, { cause: errorDetails(original), transaction_cleanup_error: errorDetails(cleanup) });
+        }
+        throw original;
+    };
+    const writeTransaction = (operation) => {
+        assertUsable();
+        if (!writable)
+            throw new AppError("read_only_storage", "Read-only access cannot start a write transaction", 1, { path: dbPath });
+        if (db.isTransaction && !reserved)
+            throw new AppError("write_transaction_unowned", "Cannot write inside an unowned transaction or read snapshot", 1);
+        const nested = db.isTransaction;
+        const marker = `dd_flow_write_${++savepoint}`;
+        try {
+            if (nested)
+                db.exec(`SAVEPOINT ${marker}`);
+            else
+                beginWriteTransaction();
+        }
+        catch (error) {
+            throw new AppError("storage_write_failed", "Cannot reserve SQLite writer", 1, { cause: errorDetails(error), transaction_phase: "begin", transaction_rolled_back: true });
+        }
+        let phase = "body";
+        try {
+            const value = operation();
+            if (value && typeof value.then === "function") {
+                void Promise.resolve(value).catch(() => { });
+                throw new AppError("write_transaction_async", "Write transaction callback must be synchronous", 1);
+            }
+            phase = "commit";
+            if (nested)
+                db.exec(`RELEASE ${marker}`);
+            else
+                db.exec("COMMIT");
+            return value;
+        }
+        catch (error) {
+            try {
+                return rollback(nested ? `ROLLBACK TO ${marker}; RELEASE ${marker}` : "ROLLBACK", error);
+            }
+            catch (failure) {
+                if (failure === error && !(error instanceof AppError) && typeof error === "object" && error !== null && "code" in error && String(error.code).startsWith("ERR_SQLITE"))
+                    throw new AppError("storage_write_failed", "SQLite write failed", 1, { cause: errorDetails(error), transaction_phase: phase, transaction_rolled_back: true });
+                throw failure;
+            }
+        }
+        finally {
+            if (!db.isTransaction)
+                reserved = false;
+        }
+    };
+    return {
+        path: dbPath,
+        writable,
+        close,
+        exec: (sql) => { assertUsable(); try {
+            db.exec(sql);
+        }
+        finally {
+            if (!db.isTransaction)
+                reserved = false;
+        } },
+        execSchema: (sql) => { assertUsable(); if (db.isTransaction && !reserved)
+            throw new AppError("write_transaction_unowned", "Cannot change schema inside an unowned transaction", 1); executeStoreSchema(db, sql); },
+        beginWriteTransaction,
+        writeTransaction,
+        run: (sql, params = []) => { assertUsable(); return db.prepare(sql).run(...params); },
+        get: (sql, params = []) => db.prepare(sql).get(...params),
+        all: (sql, params = []) => db.prepare(sql).all(...params)
+    };
+}
+function errorDetails(error) {
+    const value = error;
+    return { message: String(value?.message ?? error), ...(typeof value?.code === "string" ? { code: value.code } : {}), ...(typeof value?.errcode === "number" ? { sqlite_extended_code: value.errcode } : {}) };
+}
 function configureDatabase(db, mode) {
     if (mode !== "read_existing") {
         // Concurrent first openers can collide while changing journal mode, even
@@ -1026,6 +1116,7 @@ function migrate(db) {
     ensureColumn(db, "runs", "layout_version", "ALTER TABLE runs ADD COLUMN layout_version TEXT");
     ensureColumn(db, "runs", "artifact_root_kind", "ALTER TABLE runs ADD COLUMN artifact_root_kind TEXT");
     ensureColumn(db, "works", "payload_json", "ALTER TABLE works ADD COLUMN payload_json TEXT");
+    ensureColumn(db, "work_sessions", "start_receipt_json", "ALTER TABLE work_sessions ADD COLUMN start_receipt_json TEXT");
     ensureColumn(db, "check_receipts", "declaration_id", "ALTER TABLE check_receipts ADD COLUMN declaration_id TEXT NOT NULL DEFAULT 'CHK-LEGACY'");
     ensureColumn(db, "check_receipts", "workspace_fingerprint", "ALTER TABLE check_receipts ADD COLUMN workspace_fingerprint TEXT NOT NULL DEFAULT ''");
     ensureColumn(db, "check_receipts", "artifacts_json", "ALTER TABLE check_receipts ADD COLUMN artifacts_json TEXT NOT NULL DEFAULT '[]'");
@@ -1055,6 +1146,7 @@ function migrate(db) {
     ensureColumn(db, "hook_events", "parent_session_id", "ALTER TABLE hook_events ADD COLUMN parent_session_id TEXT");
     ensureColumn(db, "hook_events", "daemon_id", "ALTER TABLE hook_events ADD COLUMN daemon_id TEXT");
     ensureColumn(db, "hook_events", "cwd", "ALTER TABLE hook_events ADD COLUMN cwd TEXT");
+    ensureColumn(db, "hook_events", "outcome_json", "ALTER TABLE hook_events ADD COLUMN outcome_json TEXT");
     ensureColumn(db, "harness_usage_snapshots", "tool_calls", "ALTER TABLE harness_usage_snapshots ADD COLUMN tool_calls INTEGER");
     ensureColumn(db, "harness_usage_snapshots", "tool_failures", "ALTER TABLE harness_usage_snapshots ADD COLUMN tool_failures INTEGER");
     ensureColumn(db, "harness_usage_snapshots", "tool_by_name_json", "ALTER TABLE harness_usage_snapshots ADD COLUMN tool_by_name_json TEXT");