@kylecheng3146/agent-ops 0.5.2 → 0.5.3

package/dist/packages/cli/src/commands/review.js

@@ -71,7 +71,9 @@ async function taskContext(options) {
     }
     return {
         taskId: record.task.id,
-        title: record.task.title,
+        title: record.task.intent === undefined
+            ? record.task.title
+            : `${record.task.title}\nModification intent: ${record.task.intent}`,
         active: record.status === "active",
         policyConfigHash: record.policyConfigHash,
         evidence: record.evidence,
@@ -80,6 +82,51 @@ async function taskContext(options) {
         allCriteriaReviewed: criteria.length === record.task.criteria.length
     };
 }
+async function treeContext(options) {
+    if (options.tasks === undefined || options.taskId === undefined) {
+        throw new AgentOpsError("REVIEW_TREE_TASK_REQUIRED", "Tree review requires a parent task.");
+    }
+    const all = await options.tasks.list();
+    const root = all.find(({ task }) => task.id === options.taskId);
+    if (root === undefined || root.status === "archived") {
+        throw new AgentOpsError("REVIEW_TREE_TASK_REQUIRED", "The selected review tree does not exist.");
+    }
+    const selected = new Set([root.task.id]);
+    for (let changed = true; changed;) {
+        changed = false;
+        for (const record of all) {
+            if (record.status !== "archived" && record.task.parentTaskId !== undefined &&
+                selected.has(record.task.parentTaskId) && !selected.has(record.task.id)) {
+                selected.add(record.task.id);
+                changed = true;
+            }
+        }
+    }
+    const treeTasks = all.filter(({ task }) => selected.has(task.id));
+    const owners = new Map();
+    const evidence = {};
+    const criteria = treeTasks.flatMap((record) => record.task.criteria.map((criterion) => {
+        const id = `${record.task.id}:${criterion.id}`;
+        owners.set(id, { taskId: record.task.id, criterionId: criterion.id });
+        evidence[id] = record.evidence[criterion.id] ?? [];
+        return { id, description: `${record.task.title}: ${criterion.description}`, verifierIds: [...criterion.verifierIds] };
+    }));
+    if (treeTasks.some(({ policyConfigHash }) => policyConfigHash !== root.policyConfigHash)) {
+        throw new AgentOpsError("REVIEW_TREE_POLICY_MISMATCH", "Tree tasks must share one verifier policy.");
+    }
+    return {
+        taskId: root.task.id,
+        title: [root.task.title, ...treeTasks.map(({ task }) => `Task ${task.id}${task.parentTaskId === undefined ? "" : ` under ${task.parentTaskId}`}: ${task.title}\nModification intent: ${task.intent ?? task.title}`)].join("\n\n"),
+        active: treeTasks.every(({ status }) => status === "active"),
+        policyConfigHash: root.policyConfigHash,
+        evidence,
+        failureFingerprint: treeTasks.find(({ failureFingerprint }) => failureFingerprint !== null)?.failureFingerprint ?? null,
+        criteria,
+        allCriteriaReviewed: true,
+        treeTasks,
+        criterionOwners: owners
+    };
+}
 function newestEvidence(values) {
     return [...values].sort((left, right) => {
         const leftTime = Date.parse(left.evidence.finishedAt);
@@ -111,8 +158,9 @@ async function currentEvidence(options, context, criterionId, commandId, configH
             continue;
         }
         const evidence = validation.value;
-        if (evidence.taskId !== context.taskId ||
-            evidence.criterionId !== criterionId ||
+        const owner = context.criterionOwners?.get(criterionId);
+        if (evidence.taskId !== (owner?.taskId ?? context.taskId) ||
+            evidence.criterionId !== (owner?.criterionId ?? criterionId) ||
             evidence.commandId !== commandId ||
             evidence.configHash !== configHash ||
             evidence.sourceFingerprint !== sourceFingerprint) {
@@ -307,6 +355,12 @@ async function reusableReview(options, sourceFingerprint, context) {
     if (attestation === null || attestation.taskId !== context?.taskId) {
         return null;
     }
+    if (context?.treeTasks !== undefined &&
+        (attestation.tree?.rootTaskId !== context.taskId ||
+            JSON.stringify(attestation.tree.taskIds) !== JSON.stringify(context.treeTasks.map(({ task }) => task.id)) ||
+            JSON.stringify(attestation.tree.criterionIds) !== JSON.stringify(context.criteria.map(({ id }) => id)))) {
+        return null;
+    }
     // The task's own policy baseline must still be the active one. This is
     // checked before this point too; repeating it here keeps the reuse path
     // safe wherever it is called from.
@@ -355,9 +409,19 @@ async function persistReviewEvidence(options, result, sourceFingerprint, context
             reason: "adversarial-review-missing"
         }
         : result;
-    let reportArtifact;
+    const taskIds = complete && context?.treeTasks !== undefined
+        ? context.treeTasks.map(({ task }) => task.id)
+        : [context?.taskId];
+    const tree = context?.treeTasks === undefined ? undefined : {
+        rootTaskId: context.taskId,
+        taskIds: context.treeTasks.map(({ task }) => task.id),
+        criterionIds: context.criteria.map(({ id }) => id)
+    };
+    const reportArtifacts = [];
     try {
-        reportArtifact = await saveReviewReportArtifact(options.root, artifactResult, sourceFingerprint, context?.taskId);
+        for (const taskId of taskIds) {
+            reportArtifacts.push({ taskId, path: await saveReviewReportArtifact(options.root, artifactResult, sourceFingerprint, taskId, complete ? tree : undefined) });
+        }
     }
     catch {
         return {
@@ -375,21 +439,24 @@ async function persistReviewEvidence(options, result, sourceFingerprint, context
     try {
         const attempts = result.attempts;
         const plannedTargets = result.plannedTargets;
-        await saveReviewAttestation(options.root, {
-            schemaVersion: 2,
-            ...(context === undefined ? {} : { taskId: context.taskId }),
-            harness: result.harness,
-            status: "PASS",
-            sourceFingerprint,
-            reviewTargets: [plannedTargets[0], plannedTargets[1]],
-            reviewSessionIds: [attempts[0].sessionId, attempts[1].sessionId],
-            sessionIsolation: "fresh",
-            primaryReportDigest: reviewReportDigest(result.report),
-            adversarialReportDigest: reviewReportDigest(result.adversarial.report),
-            reportArtifact,
-            ...(result.hostTarget === undefined ? {} : { hostTarget: result.hostTarget }),
-            createdAt: new Date().toISOString()
-        });
+        for (const artifact of reportArtifacts) {
+            await saveReviewAttestation(options.root, {
+                schemaVersion: 2,
+                ...(artifact.taskId === undefined ? {} : { taskId: artifact.taskId }),
+                ...(tree === undefined ? {} : { tree }),
+                harness: result.harness,
+                status: "PASS",
+                sourceFingerprint,
+                reviewTargets: [plannedTargets[0], plannedTargets[1]],
+                reviewSessionIds: [attempts[0].sessionId, attempts[1].sessionId],
+                sessionIsolation: "fresh",
+                primaryReportDigest: reviewReportDigest(result.report),
+                adversarialReportDigest: reviewReportDigest(result.adversarial.report),
+                reportArtifact: artifact.path,
+                ...(result.hostTarget === undefined ? {} : { hostTarget: result.hostTarget }),
+                createdAt: new Date().toISOString()
+            });
+        }
     }
     catch {
         return {
@@ -432,7 +499,9 @@ export async function runReviewCommand(options) {
             reason: targetPlan.reason
         });
     }
-    const context = await taskContext(options);
+    const context = options.args.tree === true
+        ? await treeContext(options)
+        : await taskContext(options);
     const evidenceRequirements = (options.args.evidence ?? []).map((value) => {
         const separator = value.indexOf("=");
         return {
@@ -505,7 +574,7 @@ export async function runReviewCommand(options) {
             // source was reviewed; it says nothing about the policy in force now or
             // about verification that has failed since, and both of those are
             // decided above.
-            const reused = options.args.rerun
+            const reused = options.args.rerun || context?.treeTasks !== undefined
                 ? null
                 : await reusableReview(options, sourceFingerprint, context);
             if (reused !== null) {
@@ -518,13 +587,18 @@ export async function runReviewCommand(options) {
                     })
                 });
             }
-            await invalidateReviewAttestation(options.root, sourceFingerprint, context?.taskId);
+            for (const taskId of context?.treeTasks?.map(({ task }) => task.id) ?? [context?.taskId]) {
+                await invalidateReviewAttestation(options.root, sourceFingerprint, taskId);
+            }
             pendingInvalidation = false;
         }
     }
     const criteria = [
         ...(context?.criteria ?? GENERIC_CRITERIA)
     ];
+    if (context?.treeTasks !== undefined && criteria.length > 128) {
+        return notRunEnvelope({ ...resultBase, status: "NOT_RUN", reason: "scope-too-large" });
+    }
     let packet;
     try {
         packet = buildReviewPacket({
@@ -632,12 +706,14 @@ export async function runReviewCommand(options) {
         finalResult.status === "PASS" &&
         context.active &&
         finalResult.results !== undefined) {
-        await options.tasks.recordEvidence(context.taskId, Object.fromEntries(finalResult.results.map((item) => [
-            item.criterionId,
-            item.evidence.map((reference) => `review:${finalResult.harness}:${reference}`)
-        ])), 
-        // What `worktree finish` completes this task against.
-        scope?.mode === "base" ? scope.resolvedBase : null);
+        for (const record of context.treeTasks ?? [await options.tasks.status({ taskId: context.taskId })]) {
+            await options.tasks.recordEvidence(record.task.id, Object.fromEntries(record.task.criteria.map((criterion) => {
+                const id = context.treeTasks === undefined
+                    ? criterion.id : `${record.task.id}:${criterion.id}`;
+                const result = finalResult.results?.find((item) => item.criterionId === id);
+                return [criterion.id, (result?.evidence ?? []).map((reference) => `review:${finalResult.harness}:${reference}`)];
+            })), scope?.mode === "base" ? scope.resolvedBase : null);
+        }
     }
     const message = finalResult.status === "PASS"
         ? "Independent review passed."

package/dist/packages/cli/src/commands/task.js

@@ -92,6 +92,7 @@ export async function runTaskCommand(options) {
                 : await options.sessionWorktree(sessionId);
             const record = await (worktree?.service ?? options.service).create({
                 title: options.args.title,
+                ...(options.args.intent === undefined ? {} : { intent: options.args.intent }),
                 criteria: (options.args.criteria ?? []).map(parseCriterion),
                 ...(options.policyConfigHash === undefined
                     ? {}

package/dist/runtime/src/install/harness.js

@@ -322,7 +322,7 @@ export function managedRules(descriptor, context) {
         lines.push("For every change:", "", "1. Define two to five mechanically verifiable acceptance criteria.", "2. Inspect the smallest relevant scope and preserve unrelated changes.", "3. Apply the smallest safe change.", "4. Run evidence-producing verification for every criterion.", "5. Obtain complete independent review before claiming completion, via", "   `agent-ops review --task <task-id> --yes` (or the CLI's equivalent invocation).", "   Never call a", "   review-target CLI (agy, codex, claude) directly — direct calls skip", "   the enforced read-only sandbox flags and can hang or fail on command", "   permission prompts.", "   Start this command through the trusted outer host runner as its first", "   and only invocation. On Codex, the outer tool request must use", "   `sandbox_permissions: \"require_escalated\"` so network and loopback", "   are available; in that same request, unset", "   `CODEX_SANDBOX_NETWORK_DISABLED` and set `AGENT_OPS_HOST` when three", "   targets are configured. The Codex shell form is", "   `env -u CODEX_SANDBOX_NETWORK_DISABLED AGENT_OPS_HOST=<current-host> agent-ops review --task <task-id> --yes`.", "   Keep it a plain command line: to keep the output add `--output <file>`", "   (the JSON envelope is written there and one line is printed) instead of", "   a shell redirect, which the host can only allow by matching the whole", "   command string, task id included.", "   Do not run it once in the restricted sandbox and retry; changing the", "   variable inside that sandbox is not elevation.", "   Review runs exactly two fresh sessions: with three configured targets,", "   set `AGENT_OPS_HOST` to the current CLI id so it is excluded; with two", "   targets both run in configured order; with one target it runs twice.", "   The first session is necessary; a PASS then requires an adversarial", "   re-check carrying the first report. A first FAIL or NOT_RUN stops.", "", "Treat `.agent-ops/config.json` as verifier authority. Discovery output is", "only a proposal until a user confirms it. Repository commands require an", "exact matching trust record. Confirmed project init/update grants it", "automatically when verification commands are configured.", "");
     }
     if (context.capabilities.includes("task")) {
-        lines.push("Split a change that exceeds five acceptance criteria into subtasks:", "`agent-ops task create --parent <task-id>` records one, and", "`agent-ops task status --parent <task-id>` lists them. Each subtask", "carries its own criteria, verification, and independent review;", "completing one never completes its parent.", "", "To review every active subtask of a parent together, run", "`agent-ops batch --parent <task-id> --yes` instead of one review per", "subtask. It needs what review needs: start it through the trusted outer", "host runner as its first and only invocation, with the same Codex", "elevation, and keep `--output <file>` instead of a shell redirect. It", "verifies, one task at a time, each task that lacks fresh evidence, and", "reviews a few tasks at once. Do not commit or edit the worktree while it", "runs: a moved HEAD voids every verify and review in the batch.", "", "Delegate a subtask whose scope spans unknown files, broad scans, or an", "independent workstream instead of doing it inline. Name the bounded", "scope, the artifact to return, and the acceptance criteria it serves.", subagentInstruction(descriptor), "Read every returned artifact back yourself: integration, verification,", "and review stay with the coordinator.", "", "One writer at a time. `agent-ops verify` and `agent-ops review`", "fingerprint the working tree and report UNKNOWN when anything else", "writes it during the run, so delegate edits to one subagent at a time,", "or give each writer its own Git worktree and merge before verifying.", "Read-only delegation has no such limit and may run in parallel.", "", "Parallel conversations in one repository each need their own worktree.", "When `.agent-ops/config.json` sets `worktree.mode` to `auto`, run", "`agent-ops task create` from the main checkout before your first edit:", "it creates this session's worktree and puts the task there (`agent-ops", "worktree add <name> --session <session-id>` does the same without a", "task). Work only inside the printed path", "(Claude Code: EnterWorktree with that path; otherwise use it as every", "command's working directory). Commit, then run verify and review for", "each task with the printed `--base`; do not run `task complete` in the", "worktree. Leave it (Claude Code: ExitWorktree with action keep) and run", "`agent-ops worktree finish <name>` from the main checkout: it completes", "every task in the worktree, subtasks first, then merges. A finish that", "reports an incomplete task names it; verify or review that task again", "in the worktree and finish again. A finish that reports a conflict is", "resolved once, as a new task whose criteria cover both intents; if its", "verification or review fails, reset as instructed and report.", "");
+        lines.push("Split a change that exceeds five acceptance criteria into subtasks:", "`agent-ops task create --parent <task-id>` records one, and", "`agent-ops task status --parent <task-id>` lists them. Each subtask", "carries its own criteria and verification; final review covers all", "tasks together when their worktrees are integrated;", "completing one never completes its parent.", "", "To review every active subtask of a parent together, run", "`agent-ops batch --parent <task-id> --yes` instead of one review per", "subtask. It needs what review needs: start it through the trusted outer", "host runner as its first and only invocation, with the same Codex", "elevation, and keep `--output <file>` instead of a shell redirect. It", "verifies, one task at a time, each task that lacks fresh evidence, and", "reviews a few tasks at once. Do not commit or edit the worktree while it", "runs: a moved HEAD voids every verify and review in the batch.", "For subagents in separate worktrees, use `task advance` instead.", "", "Delegate a subtask whose scope spans unknown files, broad scans, or an", "independent workstream instead of doing it inline. Name the bounded", "scope, the artifact to return, and the acceptance criteria it serves.", subagentInstruction(descriptor), "Read every returned artifact back yourself: integration, verification,", "and review stay with the coordinator.", "", "One writer at a time. `agent-ops verify` and `agent-ops review`", "fingerprint the working tree and report UNKNOWN when anything else", "writes it during the run, so delegate edits to one subagent at a time,", "or give each writer its own Git worktree and merge before verifying.", "Read-only delegation has no such limit and may run in parallel.", "", "Parallel conversations in one repository each need their own worktree.", "When `.agent-ops/config.json` sets `worktree.mode` to `auto`, run", "`agent-ops task create` from the main checkout before your first edit:", "it creates this session's worktree and puts the task there (`agent-ops", "worktree add <name> --session <session-id>` does the same without a", "task). Work only inside the printed path", "(Claude Code: EnterWorktree with that path; otherwise use it as every", "command's working directory). Commit, then run verify and review for", "each task with the printed `--base`; do not run `task complete` in the", "worktree. Leave it (Claude Code: ExitWorktree with action keep) and run", "`agent-ops worktree finish <name>` from the main checkout: it completes", "every task in the worktree, subtasks first, then merges. A finish that", "reports an incomplete task names it; verify or review that task again", "in the worktree and finish again. A finish that reports a conflict is", "resolved once, as a new task whose criteria cover both intents; if its", "verification or review fails, reset as instructed and report.", "", "For a session with writing subagents, record each child task's complete", "modification intent with `task create --intent <text>` before editing.", "Each child commits and runs local `verify --task` in its own worktree.", "From the main checkout run `agent-ops task advance --task <parent-id>", "--session <session-id> --yes` through the trusted outer host runner,", "unsetting `CODEX_SANDBOX_NETWORK_DISABLED` on Codex. Advance integrates", "child commits and intent into the coordinator worktree, verifies every", "task on the final candidate, then runs one two-round tree review.", "Only `NOT_RUN / scope-too-large` falls back to per-task two-round review;", "a substantive FAIL must be fixed. After PASS, advance writes a local", "receipt under the main checkout's Git common directory and calls finish.", "If the target moves, it retries the final gate once. An unresolved", "failure preserves worktrees and cannot be reported as finished.", "");
     }
     if (context.capabilities.includes("lifecycle-summary")) {
         lines.push("Advisory lifecycle summaries and local logs are informational. Advisory", "failures must remain fail-open and cannot become verification evidence.", "");

package/dist/runtime/src/install/preauth.js

@@ -24,11 +24,13 @@ export const CLAUDE_PREAUTH_ALLOW = [
     "Bash(npx lint-staged *)"
 ];
 /**
- * review and batch talk to the reviewer's API, which the sandbox withholds;
+ * review, batch and task advance talk to the reviewer's API, which the sandbox withholds;
  * worktree commit and lint-staged write the main .git, which a worktree's
  * sandbox cannot.
  */
 export const CLAUDE_PREAUTH_UNSANDBOXED = [
+    "agent-ops task advance",
+    "agent-ops task advance *",
     "agent-ops review",
     "agent-ops review *",
     "agent-ops batch",
@@ -110,6 +112,13 @@ export const CODEX_RULES_CONTENT = `${[
                 match: `${env} batch --parent task-x --yes --output batch.json`,
                 notMatch: `${env} trust grant --scope project --yes`
             }),
+            rule({
+                withHost,
+                command: ["task", "advance"],
+                justification: "agent-ops task advance runs the final reviewer chain and finish outside the sandbox.",
+                match: `${env} task advance --task task-x --session session-x --yes`,
+                notMatch: `${env} trust grant --scope project --yes`
+            }),
             rule({
                 withHost,
                 command: ["doctor", "--check-auth"],

package/dist/runtime/src/parallel/finish.js

@@ -5,6 +5,8 @@ import { AgentOpsError, resolveContainedPath } from "../fs/paths.js";
 import { currentGateFingerprint } from "../hooks/completion-gate.js";
 import { collectChangeSurface } from "../verify/change-surface.js";
 import { aggregateVerificationStatus, executeConfiguredCommand } from "../verify/command-executor.js";
+import { prepareFinishReceipt } from "./receipt.js";
+import { listWorktrees } from "./manage.js";
 import { assertWorktreeName, NOTES_REF, noteSessionLine, parseWorktreeListPorcelain, readWorktreeRecord, removeCheckout, resolveCheckouts, worktreePath } from "./service.js";
 const LOCK_NAME = "agent-ops-finish.lock";
 const LOCK_POLL_MS = 2_000;
@@ -184,6 +186,7 @@ function noteText(record, tasks) {
             const indent = "  ".repeat(depth);
             return [
                 `${indent}agent-ops task ${task.task.id}: ${task.task.title}`,
+                ...(task.task.intent === undefined ? [] : [`${indent}intent: ${task.task.intent}`]),
                 `${indent}criteria:`,
                 ...task.task.criteria.map(({ id, description }) => `${indent}- ${id}: ${description}`)
             ];
@@ -206,6 +209,30 @@ export async function finishWorktree(deps, options) {
         throw finishError("WORKTREE_NOT_FOUND", `No agent-ops worktree named ${name}; see agent-ops worktree list.`);
     }
     return await withFinishLock(deps, commonDir, async () => {
+        const sessionChildren = (await listWorktrees(deps, mainRoot))
+            .map(({ record: child }) => child)
+            .filter((child) => child.sessionId === record.sessionId && child.agentId !== undefined);
+        if (options.finalProof === undefined && sessionChildren.length > 0) {
+            throw finishError("WORKTREE_CHILDREN_REQUIRE_ADVANCE", "This session has child worktrees; run task advance for integrated final evidence before finish.");
+        }
+        if (options.finalProof !== undefined &&
+            (sessionChildren.length !== options.finalProof.children.length ||
+                sessionChildren.some((child) => !options.finalProof.children.some(({ name }) => name === child.name)))) {
+            throw finishError("WORKTREE_CHILDREN_CHANGED", "The child worktree set changed after final review.");
+        }
+        const assertPinnedChild = async (child) => {
+            const expected = options.finalProof?.children.find(({ name }) => name === child.name)?.commit;
+            const head = await git(deps, child.path, ["rev-parse", "HEAD"], "WORKTREE_CHILDREN_CHANGED", `Cannot read child ${child.name}.`);
+            if (expected === undefined || head !== expected ||
+                (await collectChangeSurface(runner(deps, child.path))).paths.length > 0) {
+                throw finishError("WORKTREE_CHILDREN_CHANGED", `Child ${child.name} changed after integration; preserve its worktree.`);
+            }
+            return head;
+        };
+        if (options.finalProof !== undefined) {
+            for (const child of sessionChildren)
+                await assertPinnedChild(child);
+        }
         const mainRunner = runner(deps, mainRoot);
         const worktreeRunner = runner(deps, record.path);
         const onBranch = (await deps.git(mainRoot, ["symbolic-ref", "--short", "-q", "HEAD"])).stdout.trim();
@@ -231,37 +258,60 @@ export async function finishWorktree(deps, options) {
             throw finishError("WORKTREE_DIRTY", `Commit or discard the uncommitted changes in ${record.path} first.`);
         }
         const target = await git(deps, mainRoot, ["rev-parse", "--verify", `${record.targetBranch}^{commit}`], "WORKTREE_TARGET_MISSING", `${record.targetBranch} does not resolve to a commit.`);
+        if (options.finalProof !== undefined && target !== options.finalProof.target) {
+            throw finishError("WORKTREE_TARGET_MOVED", "The target branch moved after final review; rerun the final evidence gate.");
+        }
         const ahead = Number(await git(deps, record.path, ["rev-list", "--count", `${target}..HEAD`], "WORKTREE_LOG_FAILED", "Git could not compare the branch with its target."));
         if (ahead === 0) {
             throw finishError("WORKTREE_NOTHING_TO_MERGE", `${record.branch} has no commits beyond ${record.targetBranch}; remove the worktree instead.`);
         }
         const worktreeConfig = await deps.loadConfig(record.path);
-        // The whole branch merges, so every task in the worktree completes first,
-        // subtasks before their parents, each over the range its review covered.
-        // Before any rebase: rebasing rewrites the commits those fingerprints name.
+        const headBeforeCompletion = await git(deps, record.path, ["rev-parse", "HEAD"], "WORKTREE_LOG_FAILED", "Git could not read HEAD.");
+        if (options.finalProof !== undefined) {
+            if (headBeforeCompletion !== options.finalProof.head) {
+                throw finishError("WORKTREE_SOURCE_MOVED", "The candidate changed after final review.");
+            }
+            if ((await deps.git(record.path, ["merge-base", "--is-ancestor", target, "HEAD"])).exitCode !== 0) {
+                throw finishError("WORKTREE_TARGET_MOVED", "The final candidate is not based on the target; rerun the final evidence gate.");
+            }
+        }
+        // Direct finish completes before merge. A final-proof finish checks all
+        // evidence first, then completes only after the target's fast-forward:
+        // a target that moves before that point can still be re-reviewed once.
         const tree = taskTree((await deps.tasks(record.path).list()).filter(({ status }) => status !== "archived"));
         if (tree.length === 0) {
             throw finishError("WORKTREE_TASK_INCOMPLETE", "The worktree has no task; create one, verify and review it before finishing.");
         }
-        for (const { task } of [...tree].sort((left, right) => right.depth - left.depth)) {
-            if (task.status === "complete")
-                continue;
-            const base = task.reviewBase ?? record.base;
-            try {
-                await deps.tasks(record.path, base).complete(task.task.id, {});
+        const completeTasks = async () => {
+            for (const { task } of [...tree].sort((left, right) => right.depth - left.depth)) {
+                if (task.status === "complete")
+                    continue;
+                const base = options.finalProof === undefined ? task.reviewBase ?? record.base : target;
+                try {
+                    await deps.tasks(record.path, base).complete(task.task.id, {});
+                }
+                catch (error) {
+                    throw finishError("WORKTREE_TASK_INCOMPLETE", `Task ${task.task.id} (${task.task.title}) could not be completed against ${base}: ${error instanceof Error ? error.message : String(error)} ` +
+                        "A task another session left here is not this session's to archive: its owner finishes or archives it, and finish never does.");
+                }
             }
-            catch (error) {
-                throw finishError("WORKTREE_TASK_INCOMPLETE", `Task ${task.task.id} (${task.task.title}) could not be completed against ${base}: ${error instanceof Error ? error.message : String(error)} ` +
-                    "A task another session left here is not this session's to archive: its owner finishes or archives it, and finish never does.");
+        };
+        const validateGate = async () => {
+            const worktreeGate = await deps.gate(record.path, worktreeConfig);
+            const problem = await worktreeGate.validate(record.sessionId);
+            if (problem !== null) {
+                throw finishError("WORKTREE_TASK_INCOMPLETE", `The worktree's task is not ready to merge (${problem.code}). ${problem.remedy ?? ""}`.trim());
             }
-        }
-        const worktreeGate = await deps.gate(record.path, worktreeConfig);
-        const problem = await worktreeGate.validate(record.sessionId);
-        if (problem !== null) {
-            throw finishError("WORKTREE_TASK_INCOMPLETE", `The worktree's task is not ready to merge (${problem.code}). ${problem.remedy ?? ""}`.trim());
+        };
+        if (options.finalProof === undefined) {
+            await completeTasks();
+            await validateGate();
         }
         const head = await git(deps, record.path, ["rev-parse", "HEAD"], "WORKTREE_LOG_FAILED", "Git could not read HEAD.");
         const fastForward = (await deps.git(record.path, ["merge-base", "--is-ancestor", target, "HEAD"])).exitCode === 0;
+        if (options.finalProof !== undefined && !fastForward) {
+            throw finishError("WORKTREE_TARGET_MOVED", "The target branch moved after final review.");
+        }
         if (!fastForward) {
             const before = await patchHash(deps, record.path, target);
             const rebase = await deps.git(record.path, ["rebase", "--no-autostash", target]);
@@ -287,19 +337,64 @@ export async function finishWorktree(deps, options) {
                 throw finishError("WORKTREE_REVERIFY_FAILED", `After rebasing onto ${target}, verification failed: ${failed.join(", ")}. The branch was reset; fix it as a new task and finish again.`);
             }
         }
+        const receiptOptions = options.finalProof === undefined ? undefined : {
+            commonDir,
+            record,
+            candidateHead: head,
+            targetCommit: target,
+            expectedFingerprint: options.finalProof.sourceFingerprint,
+            children: options.finalProof.children,
+            config: worktreeConfig,
+            gitRunner: worktreeRunner
+        };
+        let receipt = receiptOptions === undefined ? undefined : await prepareFinishReceipt({
+            ...receiptOptions,
+            tasks: await deps.tasks(record.path).list().then((tasks) => tasks.filter(({ status }) => status !== "archived"))
+        });
         const mainConfig = await deps.loadConfig(mainRoot);
         const gateEnabled = mainConfig.features.completionGate.enabled;
         const before = gateEnabled ? await currentGateFingerprint(mainRoot, mainRunner) : "";
         let mergedHead;
         if (onTarget) {
-            await git(deps, mainRoot, ["merge", "--ff-only", record.branch], "WORKTREE_MERGE_FAILED", `Git could not fast-forward ${record.targetBranch}.`);
+            if (options.finalProof !== undefined &&
+                await git(deps, mainRoot, ["rev-parse", "HEAD"], "WORKTREE_LOG_FAILED", "Git could not read HEAD.") !== target) {
+                throw finishError("WORKTREE_TARGET_MOVED", "The target changed after the final proof was prepared.");
+            }
+            const merge = await deps.git(mainRoot, ["merge", "--ff-only", record.branch]);
+            if (merge.exitCode !== 0) {
+                if (options.finalProof !== undefined &&
+                    await git(deps, mainRoot, ["rev-parse", "HEAD"], "WORKTREE_LOG_FAILED", "Git could not read HEAD.") !== target) {
+                    throw finishError("WORKTREE_TARGET_MOVED", "The target moved during finish; rerun the final evidence gate.");
+                }
+                throw finishError("WORKTREE_MERGE_FAILED", `Git could not fast-forward ${record.targetBranch}. ${merge.stderr.trim()}`);
+            }
             mergedHead = await git(deps, mainRoot, ["rev-parse", "HEAD"], "WORKTREE_LOG_FAILED", "Git could not read HEAD.");
         }
         else {
             // Compare-and-swap on the target's value read above: a target that moved
             // since fails here instead of being overwritten.
             mergedHead = await git(deps, mainRoot, ["rev-parse", "--verify", `${record.branch}^{commit}`], "WORKTREE_LOG_FAILED", `Git could not read ${record.branch}.`);
-            await git(deps, mainRoot, ["update-ref", "-m", `agent-ops finish ${record.branch}`, `refs/heads/${record.targetBranch}`, mergedHead, target], "WORKTREE_MERGE_FAILED", `Git could not fast-forward ${record.targetBranch}.`);
+            const update = await deps.git(mainRoot, ["update-ref", "-m", `agent-ops finish ${record.branch}`, `refs/heads/${record.targetBranch}`, mergedHead, target]);
+            if (update.exitCode !== 0) {
+                if (options.finalProof !== undefined &&
+                    await git(deps, mainRoot, ["rev-parse", `${record.targetBranch}^{commit}`], "WORKTREE_LOG_FAILED", "Git could not read the target.") !== target) {
+                    throw finishError("WORKTREE_TARGET_MOVED", "The target moved during finish; rerun the final evidence gate.");
+                }
+                throw finishError("WORKTREE_MERGE_FAILED", `Git could not fast-forward ${record.targetBranch}. ${update.stderr.trim()}`);
+            }
+        }
+        if (receiptOptions !== undefined) {
+            try {
+                await completeTasks();
+                await validateGate();
+                receipt = await prepareFinishReceipt({
+                    ...receiptOptions,
+                    tasks: await deps.tasks(record.path).list().then((tasks) => tasks.filter(({ status }) => status !== "archived"))
+                });
+            }
+            catch (error) {
+                throw finishError("WORKTREE_FINISH_PARTIAL", `Code merged at ${mergedHead}; final task state or receipt failed: ${error instanceof Error ? error.message : String(error)}`);
+            }
         }
         // The merge happened; everything below is cleanup, reported, never undone.
         const warnings = [];
@@ -308,10 +403,15 @@ export async function finishWorktree(deps, options) {
                 await step();
             }
             catch (error) {
-                warnings.push(`${label}: ${error instanceof Error ? error.message : String(error)}`);
+                const message = `${label}: ${error instanceof Error ? error.message : String(error)}`;
+                if (options.finalProof !== undefined) {
+                    throw finishError("WORKTREE_FINISH_PARTIAL", `Code merged at ${mergedHead} with receipt ${receipt?.path}; ${message}`);
+                }
+                warnings.push(message);
             }
         };
-        await attempt("note", async () => await git(deps, mainRoot, ["notes", `--ref=${NOTES_REF}`, "add", "-f", "-m", noteText(record, tree), mergedHead], "WORKTREE_NOTE_FAILED", "Git could not write the agent-ops note."));
+        const finishNote = `${noteText(record, tree)}${receipt === undefined ? "" : `\nreceipt: ${receipt.path}\nreceipt-sha256: ${receipt.digest}`}`;
+        await attempt("note", async () => await git(deps, mainRoot, ["notes", `--ref=${NOTES_REF}`, "add", "-f", "-m", finishNote, mergedHead], "WORKTREE_NOTE_FAILED", "Git could not write the agent-ops note."));
         if (gateEnabled) {
             await attempt("gate", async () => {
                 const mainGate = await deps.gate(mainRoot, mainConfig);
@@ -321,10 +421,25 @@ export async function finishWorktree(deps, options) {
                 // this session began would leave it reading the merge as its own change.
                 // Only this worktree leaves the session; its other roots stay covered.
                 await mainGate.afterFinish(record.sessionId, record.path, after);
+                if (options.finalProof !== undefined) {
+                    for (const child of sessionChildren) {
+                        await mainGate.afterFinish(record.sessionId, child.path, after);
+                    }
+                }
             });
         }
         await attempt("trust", async () => await deps.trust.revoke(record.path, worktreeConfig));
-        await attempt("remove", async () => await removeCheckout(deps, record, true));
-        return { record, mergedHead, rebased: !fastForward, taskIds: tree.map(({ task }) => task.task.id), warnings };
+        if (options.finalProof !== undefined) {
+            for (const child of sessionChildren) {
+                await attempt(`child ${child.name}`, async () => {
+                    const pinnedHead = await assertPinnedChild(child);
+                    await deps.trust.revoke(child.path, await deps.loadConfig(child.path));
+                    await removeCheckout(deps, child, false, pinnedHead);
+                });
+            }
+        }
+        await attempt("remove", async () => await removeCheckout(deps, record, options.finalProof === undefined, options.finalProof?.head));
+        return { record, mergedHead, rebased: !fastForward, taskIds: tree.map(({ task }) => task.task.id), warnings,
+            ...(receipt === undefined ? {} : { receipt: receipt.path }) };
     });
 }