zelari-code 1.28.0 → 1.29.0

package/dist/cli/kraken/executor.js

@@ -6,6 +6,9 @@
  * (`runTentacle`) to execute a validated `TaskGraph`:
  *   - repeatedly forms a parallel-safe "wave" of ready nodes (bounded by
  *     ZELARI_KRAKEN_MAX_PARALLEL) and runs it concurrently;
+ *   - hands each node the conclusions of its completed dependencies
+ *     (`buildUpstreamContext`), so a dep edge carries information and not just
+ *     ordering;
  *   - retries a failed node up to `node.maxRetries`, then spawns a `fix`
  *     node (bounded by a global fix budget) that inherits the failed
  *     node's deps/scope/acceptance; a node that still fails is left
@@ -31,7 +34,7 @@
  *
  * @since v0.10.x — Kraken graph engine (F3)
  */
-import { getReadyNodes, isSettled, isConverged, failedNodeIds, countByStatus, selectParallelWave, } from '@zelari/core';
+import { getReadyNodes, isSettled, isConverged, failedNodeIds, countByStatus, canRunParallel, parseVerifyVerdict, } from '@zelari/core';
 import { runTentacle, } from './tentacle.js';
 import { mergeKrakenWorktree, } from '../tools/krakenWorktree.js';
 import { appendKrakenRadio } from '../tools/krakenRadio.js';
@@ -51,12 +54,58 @@ export function resolveMaxParallel(env = process.env) {
     const n = Number.parseInt(raw, 10);
     return Number.isFinite(n) && n > 0 ? n : DEFAULT_MAX_PARALLEL;
 }
-export function resolveFixBudget(env = process.env) {
+/**
+ * How many `fix` nodes this run may spawn.
+ *
+ * A flat budget of 3 is a graph-size-blind number: on a 20-node graph it is
+ * spent on the first three failures and every later failure goes terminal,
+ * cascade-skipping its dependents — the larger the graph, the less repair it
+ * gets, which is backwards. Scale with the node count, keeping 3 as the floor
+ * for small graphs. An explicit env value still wins outright.
+ */
+export function resolveFixBudget(env = process.env, nodeCount = 0) {
     const raw = env.ZELARI_KRAKEN_FIX_BUDGET;
+    if (raw === undefined || raw === '') {
+        return Math.max(DEFAULT_FIX_BUDGET, Math.ceil(nodeCount / 2));
+    }
+    const n = Number.parseInt(raw, 10);
+    return Number.isFinite(n) && n >= 0
+        ? n
+        : Math.max(DEFAULT_FIX_BUDGET, Math.ceil(nodeCount / 2));
+}
+/**
+ * Default rework rounds per writer when its `verify` returns FAIL.
+ *
+ * One. A rework round is a full re-run of a writer plus a fresh verification,
+ * so the cost is roughly a doubling of that branch; and the second opinion of
+ * a model that just judged its own sibling's work has sharply diminishing
+ * value. Raise it deliberately (and with a wall-clock budget set) rather than
+ * by default.
+ */
+export const DEFAULT_MAX_REVIEW_ROUNDS = 1;
+export function resolveMaxReviewRounds(env = process.env) {
+    const raw = env.ZELARI_KRAKEN_MAX_REVIEW_ROUNDS;
     if (raw === undefined || raw === '')
-        return DEFAULT_FIX_BUDGET;
+        return DEFAULT_MAX_REVIEW_ROUNDS;
     const n = Number.parseInt(raw, 10);
-    return Number.isFinite(n) && n >= 0 ? n : DEFAULT_FIX_BUDGET;
+    return Number.isFinite(n) && n >= 0 ? n : DEFAULT_MAX_REVIEW_ROUNDS;
+}
+/**
+ * Wall-clock bound on the WHOLE graph run (ms); 0 (the default) disables it.
+ *
+ * Per-node timeouts bound each tentacle but say nothing about the total: a
+ * wide graph, plus retries, plus fix nodes, plus rework rounds, can run far
+ * longer than any single node's budget without anything noticing. On expiry
+ * the run takes the ordinary cancellation path, so it still settles and still
+ * prints its digest.
+ */
+export const DEFAULT_GRAPH_TIMEOUT_MS = 0;
+export function resolveGraphTimeoutMs(env = process.env) {
+    const raw = env.ZELARI_KRAKEN_GRAPH_TIMEOUT_MS;
+    if (raw === undefined || raw === '')
+        return DEFAULT_GRAPH_TIMEOUT_MS;
+    const n = Number.parseInt(raw, 10);
+    return Number.isFinite(n) && n >= 0 ? n : DEFAULT_GRAPH_TIMEOUT_MS;
 }
 /**
  * Default wall-clock bound per tentacle run (ms). The graph executor calls
@@ -143,6 +192,85 @@ function defaultChecksExists(cwd) {
         return false;
     }
 }
+/**
+ * Per-dependency cap on injected upstream text. Mirrors `MAX_PRIOR_CHARS` in
+ * the council path (`packages/core/src/agents/councilApi.ts`), for the same
+ * reason: one verbose tentacle must not be able to saturate every downstream
+ * one.
+ */
+export const MAX_UPSTREAM_CHARS_PER_DEP = 2800;
+/** Total cap across all deps, so a wide fan-in cannot saturate a sub-agent. */
+export const MAX_UPSTREAM_CHARS_TOTAL = 8000;
+/**
+ * Render the conclusions of a node's already-completed dependencies as a
+ * prompt section for the sub-agent about to run.
+ *
+ * Without this the dependency edges were pure ordering constraints: a tentacle
+ * received only the prompt the PLANNER wrote before anything had executed, so
+ * an `explore` node's findings were computed, stored on the node, and then
+ * thrown away — the `general` nodes it "fed" started from zero and re-derived
+ * the same context (or guessed). The auto-injected `verify` node was worse
+ * still: it knew only the label of the work it was supposed to check.
+ *
+ * Direct dependencies only, deliberately — the transitive closure of a wide
+ * DAG would blow the sub-agent's context, and each hop's own conclusion is
+ * expected to carry forward what mattered.
+ */
+export function buildUpstreamContext(graph, node) {
+    const parts = [];
+    const omitted = [];
+    let budget = MAX_UPSTREAM_CHARS_TOTAL;
+    for (const depId of node.deps) {
+        const dep = graph.nodes.get(depId);
+        if (!dep || dep.status !== 'done')
+            continue;
+        const raw = (dep.result ?? '').trim();
+        if (!raw)
+            continue;
+        const cap = Math.min(MAX_UPSTREAM_CHARS_PER_DEP, budget);
+        if (cap <= 0) {
+            omitted.push(dep.label);
+            continue;
+        }
+        const body = raw.length > cap
+            ? `${raw.slice(0, cap)}\n… [truncated ${raw.length}→${cap} chars]`
+            : raw;
+        budget -= Math.min(raw.length, cap);
+        const scope = dep.scope && dep.scope.length > 0 ? `, scope: ${dep.scope.join(', ')}` : '';
+        parts.push(`### ${dep.label} (${dep.kind}${scope})\n${body}`);
+    }
+    if (parts.length === 0)
+        return '';
+    const lines = [
+        '',
+        '## Context from completed upstream tasks',
+        'Results reported by the tasks this one depends on. Treat them as ' +
+            'hypotheses — prefer the actual files on disk where they conflict.',
+        '',
+        ...parts,
+    ];
+    if (omitted.length > 0) {
+        lines.push('', `(omitted for context budget: ${omitted.join(', ')})`);
+    }
+    return lines.join('\n');
+}
+/**
+ * Tool budget per node kind. Writers do real, multi-file work and were being
+ * run at the same `medium` budget as a read-only lookup; readers stay tight
+ * because a wide budget just invites them to dump the repository.
+ */
+export function thoroughnessForKind(kind) {
+    return kind === 'general' || kind === 'fix' ? 'deep' : 'medium';
+}
+/** First non-empty line of a block of text, capped — for one-line markers. */
+function firstLine(text, maxChars = 160) {
+    const line = text.trim().split('\n').find((l) => l.trim() !== '')?.trim() ?? '';
+    return line.length > maxChars ? `${line.slice(0, maxChars)}…` : line;
+}
+/** The tentacle kind a node runs as (`fix`/`merge` are driven as `general`). */
+function agentForNode(node) {
+    return node.kind === 'explore' || node.kind === 'verify' ? node.kind : 'general';
+}
 export class KrakenGraphExecutor {
     deps;
     parentCwd;
@@ -152,12 +280,44 @@ export class KrakenGraphExecutor {
     /** Explicit all-kinds override; when undefined the budget is per-kind. */
     nodeTimeoutMs;
     cancelGraceMs;
+    /** Explicit override; when undefined the budget scales with the graph size. */
+    fixBudgetOption;
     fixBudgetRemaining;
+    maxReviewRounds;
+    graphTimeoutMs;
     worldModelGateOverride;
     runTentacleFn;
     mergeFn;
     backtestFn;
     nodeRunState = new Map();
+    /** fix node id → id of the failed node it was spawned to repair. */
+    repairs = new Map();
+    /**
+     * rework node id → id of the writer whose work it is redoing. A rework is a
+     * `fix` node, but unlike a repair it must NOT create a worktree of its own:
+     * it edits the writer's existing one (see {@link spawnReworkPair}).
+     */
+    reworks = new Map();
+    /** lineage root writer id → rework rounds already spent on that lineage. */
+    reviewRounds = new Map();
+    /**
+     * rework node id → the ORIGINAL writer its lineage started from.
+     *
+     * The budget has to be per lineage, not per node: a rework is itself a
+     * writer, so counting rounds against the node being reworked reset the
+     * counter every round and the graph chained rework → verify → rework
+     * forever, terminating only on the scheduler's iteration cap.
+     */
+    reviewLineage = new Map();
+    /** Verify verdicts left unresolved when the run ends. */
+    unresolved = [];
+    /** Live cancellation handles for the tentacles currently running. */
+    nodeControllers = new Map();
+    /** Wall-clock duration of each node's last run, by node id. */
+    durationsMs = new Map();
+    signal;
+    /** Set once the run has been cancelled: stops admission, retries and fixes. */
+    aborted = false;
     fixCounter = 0;
     constructor(opts) {
         this.deps = opts.taskToolDeps;
@@ -167,19 +327,69 @@ export class KrakenGraphExecutor {
         this.maxParallel = opts.maxParallel ?? resolveMaxParallel();
         this.nodeTimeoutMs = opts.nodeTimeoutMs;
         this.cancelGraceMs = opts.cancelGraceMs;
+        this.fixBudgetOption = opts.fixBudget;
+        // Provisional: `execute()` re-resolves it once the graph size is known.
         this.fixBudgetRemaining = opts.fixBudget ?? resolveFixBudget();
+        this.maxReviewRounds = opts.maxReviewRounds ?? resolveMaxReviewRounds();
+        this.graphTimeoutMs = opts.graphTimeoutMs ?? resolveGraphTimeoutMs();
         this.worldModelGateOverride = opts.worldModelGate;
+        this.signal = opts.signal;
         this.runTentacleFn = opts.runTentacleFn ?? runTentacle;
         this.mergeFn = opts.mergeFn ?? mergeKrakenWorktree;
         this.backtestFn = opts.backtestFn ?? runBacktest;
     }
     /** Execute the graph in place (mutates node statuses) until it settles. */
     async execute(graph) {
+        // Now that the graph is known, size the repair budget to it (unless the
+        // caller pinned one). Done here rather than in the constructor because the
+        // node count is the whole input to the decision.
+        if (this.fixBudgetOption === undefined) {
+            this.fixBudgetRemaining = resolveFixBudget(process.env, graph.nodes.size);
+        }
+        // Cancel eagerly rather than at the next loop turn: without a listener the
+        // scheduler would sit in `Promise.race` until some node finished on its
+        // own, which for a stuck writer is the whole point of cancelling.
+        const onAbort = () => this.cancelRun();
+        if (this.signal) {
+            if (this.signal.aborted)
+                this.aborted = true;
+            else
+                this.signal.addEventListener('abort', onAbort, { once: true });
+        }
+        // Whole-run wall-clock bound. Reuses the cancellation path rather than
+        // inventing a second way to stop, so the run still settles and still
+        // reports.
+        let graphTimer;
+        if (this.graphTimeoutMs > 0) {
+            graphTimer = setTimeout(() => {
+                this.radio('graph_failed', {
+                    description: 'graph executor',
+                    detail: `graph exceeded its ${this.graphTimeoutMs}ms wall-clock budget — cancelling`,
+                    ok: false,
+                });
+                this.cancelRun();
+            }, this.graphTimeoutMs);
+            // Do not hold the process open just to fire a cancellation.
+            graphTimer.unref?.();
+        }
+        try {
+            return await this.schedule(graph);
+        }
+        finally {
+            if (graphTimer)
+                clearTimeout(graphTimer);
+            this.signal?.removeEventListener('abort', onAbort);
+        }
+    }
+    /** The scheduling loop proper. See {@link execute} for the cancellation wrapper. */
+    async schedule(graph) {
         // Anti-explosion / self-inflicted-hang guard: cap total loop iterations
         // at a generous multiple of the node count instead of trusting the
         // graph to always shrink monotonically (fix-node spawns grow it).
         const maxIterations = Math.max(64, graph.nodes.size * 8);
         let iterations = 0;
+        /** Node id → its running tentacle. One entry per in-flight node. */
+        const inFlight = new Map();
         startKrakenGraphLive(graph);
         while (!isSettled(graph)) {
             iterations += 1;
@@ -191,30 +401,66 @@ export class KrakenGraphExecutor {
                 });
                 break;
             }
-            const ready = getReadyNodes(graph);
-            if (ready.length === 0) {
-                // Nothing is ready but the graph hasn't settled: some pending nodes
-                // are permanently blocked by a failed/skipped dependency. Cascade
-                // skip them so the loop can terminate.
-                const skippedAny = this.skipBlockedNodes(graph);
-                if (!skippedAny) {
-                    // Nothing ready and nothing to skip — should not happen for a
-                    // validated DAG, but bail out rather than spin forever.
+            if (this.aborted && inFlight.size === 0)
+                break;
+            // Admit everything that can start alongside what is already running,
+            // up to the concurrency cap. A cancelled run admits nothing.
+            const admitted = this.aborted ? [] : this.admit(graph, inFlight);
+            if (admitted.length > 0) {
+                // Mark and publish the whole admission BEFORE starting any of it: an
+                // async function body runs synchronously up to its first await, so
+                // starting a tentacle first would let it observe stale counts. (The
+                // StatusBar's "n↑" never appeared at all while this was sampled only
+                // after the work had already settled.)
+                for (const node of admitted)
+                    node.status = 'running';
+                updateKrakenGraphLive(graph);
+                for (const node of admitted)
+                    inFlight.set(node.id, this.runNodeSafely(node, graph));
+            }
+            if (inFlight.size === 0) {
+                // Nothing running and nothing admissible: the remaining pending nodes
+                // are permanently blocked by a failed/skipped dependency (or by the
+                // run being cancelled). Cascade skip them so the loop can terminate.
+                if (!this.skipBlockedNodes(graph)) {
+                    // Nothing ready, nothing running, nothing to skip — should not
+                    // happen for a validated DAG, but bail out rather than spin forever.
                     break;
                 }
                 continue;
             }
-            const wave = selectParallelWave(ready).slice(0, this.maxParallel);
-            for (const node of wave)
-                node.status = 'running';
-            const results = await Promise.all(wave.map((node) => this.runNode(node, graph)));
-            for (let i = 0; i < wave.length; i++) {
-                this.applyResult(graph, wave[i], results[i]);
-            }
+            // Settle ONE node at a time. Waiting for a whole wave meant a 15-minute
+            // writer held back every explore that became ready a second later, and
+            // left the concurrency budget idle for the duration.
+            const { id, res } = await Promise.race(inFlight.values());
+            inFlight.delete(id);
+            const settledNode = graph.nodes.get(id);
+            if (settledNode)
+                this.applyResult(graph, settledNode, res);
             updateKrakenGraphLive(graph);
         }
+        // The loop can break with work still in flight (iteration cap, or a
+        // cancellation whose tentacles have not unwound yet). Let those settle
+        // rather than returning a summary while they are still writing, and
+        // record what they produced.
+        if (inFlight.size > 0) {
+            for (const { id, res } of await Promise.all(inFlight.values())) {
+                const node = graph.nodes.get(id);
+                if (node)
+                    this.applyResult(graph, node, res);
+            }
+            inFlight.clear();
+        }
+        // A cancelled run leaves nodes that never started: `skipped` says exactly
+        // that, and lets the graph settle so a summary can be returned.
+        if (this.aborted) {
+            for (const n of graph.nodes.values()) {
+                if (n.status === 'pending')
+                    n.status = 'skipped';
+            }
+        }
         let backtest;
-        const converged = isConverged(graph);
+        const converged = !this.aborted && isConverged(graph);
         if (converged) {
             const gateOn = this.worldModelGateOverride ?? isWorldModelGateEnabled(this.parentCwd);
             if (gateOn) {
@@ -229,45 +475,148 @@ export class KrakenGraphExecutor {
         else {
             this.radio('graph_failed', {
                 description: 'graph executor',
-                detail: `failed nodes: ${failedNodeIds(graph).join(', ') || 'none'}`,
+                detail: this.aborted
+                    ? 'cancelled by caller'
+                    : `failed nodes: ${failedNodeIds(graph).join(', ') || 'none'}`,
                 ok: false,
             });
         }
         endKrakenGraphLive(graph, converged);
         // Cross-run memory: let the next planning pass see where this one stopped
         // instead of replanning the whole goal blind.
-        await saveGraphSnapshot(this.parentCwd, toGraphSnapshot(graph, { goal: this.goal ?? graph.id, converged }));
+        await saveGraphSnapshot(this.parentCwd, toGraphSnapshot(graph, {
+            goal: this.goal ?? graph.id,
+            converged,
+            unresolvedFindings: this.unresolved,
+        }));
         return {
             graph,
             converged,
             failedNodeIds: failedNodeIds(graph),
             counts: countByStatus(graph),
+            durationsMs: Object.fromEntries(this.durationsMs),
+            cancelled: this.aborted,
+            unresolvedFindings: [...this.unresolved],
             ...(backtest ? { backtest } : {}),
         };
     }
+    /**
+     * Stop the run: no further admissions, and every tentacle currently running
+     * is told to unwind. Each node's own timeout/grace machinery then resolves
+     * it, so `execute()` settles instead of leaving orphans behind.
+     */
+    cancelRun() {
+        if (this.aborted)
+            return;
+        this.aborted = true;
+        this.radio('graph_failed', {
+            description: 'graph executor',
+            detail: `cancelling ${this.nodeControllers.size} running tentacle(s)`,
+            ok: false,
+        });
+        for (const controller of this.nodeControllers.values())
+            controller.abort();
+    }
+    /**
+     * Pick the ready nodes that may start right now: parallel-safe against every
+     * node already running AND against each other, within the concurrency cap.
+     *
+     * Unlike a wave-at-a-time scheduler this is called on every completion, so a
+     * node becomes eligible the moment its blocker settles instead of waiting
+     * for the slowest member of some earlier batch.
+     */
+    admit(graph, inFlight) {
+        const capacity = this.maxParallel - inFlight.size;
+        if (capacity <= 0)
+            return [];
+        const running = [];
+        for (const id of inFlight.keys()) {
+            const n = graph.nodes.get(id);
+            if (n)
+                running.push(n);
+        }
+        const admitted = [];
+        for (const node of getReadyNodes(graph)) {
+            if (admitted.length >= capacity)
+                break;
+            const safe = running.every((r) => canRunParallel(r, node)) &&
+                admitted.every((a) => canRunParallel(a, node));
+            if (safe)
+                admitted.push(node);
+        }
+        return admitted;
+    }
+    /**
+     * Run one node, tagging the result with its id and converting an unexpected
+     * throw into a node failure. The scheduler races these promises, so a
+     * rejection would abandon every other in-flight tentacle mid-write; one
+     * failed node that the retry/fix machinery can reason about is strictly
+     * better than an aborted graph.
+     */
+    runNodeSafely(node, graph) {
+        return this.runNode(node, graph).then((res) => ({ id: node.id, res }), (err) => ({
+            id: node.id,
+            res: {
+                ok: false,
+                agent: agentForNode(node),
+                error: `tentacle threw: ${err instanceof Error ? err.message : String(err)}`,
+                cancelled: true,
+            },
+        }));
+    }
     /** Run one node: dispatch to the merge handler for `merge` nodes, else a tentacle. */
     async runNode(node, graph) {
+        const startedAt = Date.now();
+        try {
+            return await this.runNodeInner(node, graph);
+        }
+        finally {
+            this.durationsMs.set(node.id, Date.now() - startedAt);
+            this.nodeControllers.delete(node.id);
+        }
+    }
+    async runNodeInner(node, graph) {
         this.radio('node_start', { description: node.label, agent: node.kind });
         if (node.kind === 'merge') {
             return this.runMergeNode(node, graph);
         }
-        const usesWorktree = node.kind === 'general' || node.kind === 'fix';
+        // A rework edits the worktree its writer already produced (see
+        // spawnReworkPair) — it must not open a second one on the same scope.
+        const isRework = this.reworks.has(node.id);
+        const usesWorktree = (node.kind === 'general' || node.kind === 'fix') && !isRework;
         const agent = node.kind === 'fix' ? 'general' : node.kind;
         const controller = new AbortController();
+        // Registered so a cancelled run can reach in and stop this tentacle
+        // instead of waiting for it to finish on its own.
+        this.nodeControllers.set(node.id, controller);
+        if (this.aborted)
+            controller.abort();
+        // Hand the sub-agent what its dependencies actually concluded — the whole
+        // point of the dep edge (see buildUpstreamContext).
+        const upstream = buildUpstreamContext(graph, node);
+        // A verify inspects the tree its writer wrote in; a rework edits that same
+        // tree. Both resolve to the worktree recorded behind their deps.
+        const inheritedCwd = node.kind === 'verify' || isRework
+            ? this.inheritedWorktreeCwdFor(node, graph)
+            : undefined;
         const res = await this.withNodeTimeout(this.runTentacleFn({
-            deps: this.deps,
+            // `allowWorktree: false` is what actually stops a rework from opening
+            // its own worktree: creation is driven by the agent kind ('general')
+            // inside runTentacle, not by anything the executor passes per-call.
+            deps: isRework ? { ...this.deps, allowWorktree: false } : this.deps,
             args: {
                 description: node.label,
-                prompt: node.prompt,
+                prompt: upstream ? `${node.prompt}\n${upstream}` : node.prompt,
                 scope: node.scope,
                 acceptance: node.acceptance,
             },
             agent,
-            thoroughness: 'medium',
+            thoroughness: thoroughnessForKind(node.kind),
             parentCwd: this.parentCwd,
+            ...(inheritedCwd ? { cwdOverride: inheritedCwd } : {}),
             sessionId: this.sessionId,
             // Defer merge for writers so the executor controls merge ordering
-            // (Correction 4); explore/verify never create a worktree.
+            // (Correction 4); explore/verify/rework never create a worktree.
             deferMerge: usesWorktree,
             graphId: graph.id,
             nodeId: node.id,
@@ -342,21 +691,74 @@ export class KrakenGraphExecutor {
         };
     }
     /**
-     * Sequentially merge every dep node's deferred worktree (in dep order) into
-     * parent HEAD. Deps without a recorded worktree handle (worktree isolation
-     * disabled, or a read-only node) are a no-op. On conflict the branch is
-     * kept and the conflict is surfaced in the merge node's error — remaining
-     * deps still attempt to merge (independent branches shouldn't be blocked
-     * by one conflict).
+     * The worktree a node should run in, inherited from the writer behind it.
+     *
+     * For a `verify`: verification happens BEFORE the merge node, so when its
+     * writer worked in an isolated worktree the changes are not in the parent
+     * tree yet — a verify tentacle pointed at `parentCwd` was inspecting a tree
+     * that provably did not contain the work it was asked to check, and reported
+     * it missing.
+     *
+     * For a rework: the same tree, for the stronger reason that writing anywhere
+     * else would strand the round on a second branch.
+     *
+     * Returns undefined when there is no single tree: no worktrees (isolation
+     * disabled — the writers edited the parent tree directly), or several
+     * distinct ones, in which case no single cwd is correct and the parent tree
+     * is the honest default.
+     */
+    inheritedWorktreeCwdFor(node, graph) {
+        const paths = new Set(this.collectWorktreeSources(node, graph).map((s) => s.handle.path));
+        return paths.size === 1 ? [...paths][0] : undefined;
+    }
+    /**
+     * Resolve the deferred worktrees produced behind a node's dependencies, in
+     * ancestors-first order. Used to decide what a `merge` node must merge, and
+     * which tree a `verify` node should actually inspect.
+     *
+     * A merge node's direct deps are NOT the writers: `buildGraphFromPlan`
+     * injects a `verify` node after every `general` node and points the merge at
+     * those verifies, while worktree handles are recorded against the writer
+     * node ids. Looking only at direct deps therefore found nothing to merge and
+     * silently reported success while every tentacle's work stayed stranded on
+     * its branch. Walk up through non-writer deps until the writers are found.
+     *
+     * Post-order so a writer that depends on another writer merges after it (the
+     * later branch was cut from a HEAD that already contained the earlier work).
+     * `merge` nodes terminate the walk: another merge already owns its subtree.
+     */
+    collectWorktreeSources(node, graph) {
+        const out = [];
+        const seen = new Set();
+        const visit = (id) => {
+            if (seen.has(id))
+                return;
+            seen.add(id);
+            const n = graph.nodes.get(id);
+            if (!n || n.kind === 'merge')
+                return;
+            for (const dep of n.deps)
+                visit(dep);
+            const handle = this.nodeRunState.get(id)?.worktreeHandle;
+            if (handle)
+                out.push({ id, handle });
+        };
+        for (const depId of node.deps)
+            visit(depId);
+        return out;
+    }
+    /**
+     * Sequentially merge every deferred worktree this node covers (in
+     * ancestors-first order) into parent HEAD. Nodes without a recorded worktree
+     * handle (worktree isolation disabled, or a read-only node) are a no-op. On
+     * conflict the branch is kept and the conflict is surfaced in the merge
+     * node's error — remaining sources still attempt to merge (independent
+     * branches shouldn't be blocked by one conflict).
      */
     async runMergeNode(node, graph) {
         const conflicts = [];
         const merged = [];
-        for (const depId of node.deps) {
-            const state = this.nodeRunState.get(depId);
-            const handle = state?.worktreeHandle;
-            if (!handle)
-                continue; // nothing to merge for this dep
+        for (const { id: depId, handle } of this.collectWorktreeSources(node, graph)) {
             let result;
             try {
                 result = await this.mergeFn(handle, {
@@ -376,6 +778,10 @@ export class KrakenGraphExecutor {
                 conflicts.push(`${depId}: ${result.message}`);
             }
             else {
+                // The branch is merged (and, unless KEEP is set, the worktree is gone).
+                // Drop the handle so a second merge node covering the same ancestor
+                // cannot try to merge a branch that no longer exists.
+                this.nodeRunState.set(depId, { worktreeHandle: null });
                 merged.push(depId);
             }
         }
@@ -403,9 +809,26 @@ export class KrakenGraphExecutor {
             node.status = 'done';
             node.result = res.result;
             this.radio('node_end', { description: node.label, agent: node.kind, ok: true });
+            this.reconcileRepairedNode(graph, node);
+            // A verify that RAN successfully is not the same thing as work that
+            // PASSED. Read what it actually concluded.
+            if (node.kind === 'verify')
+                this.applyVerifyVerdict(graph, node);
             return;
         }
         node.error = res.error;
+        // A cancelled run must not retry or spawn repairs: the caller asked for
+        // the graph to stop, and re-spawning work is the opposite of that.
+        if (this.aborted) {
+            node.status = 'error';
+            this.radio('node_end', {
+                description: node.label,
+                agent: node.kind,
+                detail: res.error,
+                ok: false,
+            });
+            return;
+        }
         // A run we could not confirm has stopped may still be writing this node's
         // scope. Re-spawning would put two tentacles in the same directory — the
         // failure mode that produced duplicate parallel implementations of the
@@ -455,6 +878,201 @@ export class KrakenGraphExecutor {
             ok: false,
         });
     }
+    /**
+     * A `fix` node just completed the work its failed predecessor could not.
+     * That unit of work IS done — but the predecessor was left terminally
+     * `error`, and since `isConverged` requires every node to be `done`/
+     * `skipped`, a fully repaired graph reported "did not converge" and listed
+     * the repaired node under `failedNodeIds`. The cross-run snapshot then told
+     * the next planner to redo work the fix had already completed.
+     *
+     * Marking it `done` has no scheduling effect (dependents were re-pointed at
+     * the fix when it was spawned) — it is purely how the run is reported. The
+     * original failure stays visible as the separate `fix: …` node and in the
+     * repaired node's result line.
+     */
+    reconcileRepairedNode(graph, fixNode) {
+        const failedId = this.repairs.get(fixNode.id);
+        if (!failedId)
+            return;
+        const failed = graph.nodes.get(failedId);
+        if (!failed || failed.status !== 'error')
+            return;
+        const original = failed.error ? ` (original failure: ${failed.error})` : '';
+        failed.status = 'done';
+        failed.result = `repaired by "${fixNode.label}"${original}${fixNode.result ? `: ${fixNode.result}` : ''}`;
+        // Clear the error: the snapshot lists a `done` node under "Already
+        // completed — do NOT redo this work", and a trailing error message there
+        // reads as if it still needs repair.
+        failed.error = undefined;
+        this.radio('node_end', {
+            description: failed.label,
+            agent: failed.kind,
+            detail: `repaired by ${fixNode.id}`,
+            ok: true,
+        });
+    }
+    /**
+     * Act on what a completed `verify` node concluded.
+     *
+     * The verify itself stays `done` either way — it did its job, and doing it
+     * well means being free to say "no". A FAIL instead sends the WRITER back
+     * through a bounded rework round.
+     *
+     * Without this the verdict text was never read: a verify that reported the
+     * work as wrong was recorded exactly like one that reported it correct, the
+     * graph converged over the defect, and the only iteration the engine could
+     * do was on execution failure. An `unknown` verdict (no parseable trailer)
+     * is deliberately non-blocking — a prompt drift must not be able to wedge
+     * every graph — but it is recorded, because a gate that has silently stopped
+     * working is worse than no gate.
+     */
+    applyVerifyVerdict(graph, verify) {
+        const { verdict, findings } = parseVerifyVerdict(verify.result);
+        if (verdict === 'pass')
+            return;
+        const writer = this.writerBehind(verify, graph);
+        if (!writer)
+            return; // nothing to send back to
+        if (verdict === 'unknown') {
+            this.unresolved.push({
+                nodeId: writer.id,
+                label: writer.label,
+                reason: 'unknown',
+                findings: findings || '(verify produced no parseable VERDICT line)',
+            });
+            return;
+        }
+        // A cancelled run must not spawn new work, and a lineage that has spent
+        // its rounds is done being reworked.
+        const root = this.reviewLineage.get(writer.id) ?? writer.id;
+        const spent = this.reviewRounds.get(root) ?? 0;
+        if (this.aborted || spent >= this.maxReviewRounds) {
+            this.unresolved.push({
+                nodeId: writer.id,
+                label: writer.label,
+                reason: 'fail',
+                findings,
+            });
+            // Keep the unresolved verdict attached to the node too, so the digest,
+            // the snapshot and any reader of the graph see it in the same place the
+            // repaired-node marker lives.
+            writer.result =
+                `${writer.result ?? ''}\n\n[accepted with unresolved verify findings from ` +
+                    `"${verify.label}"]${findings ? `: ${firstLine(findings)}` : ''}`.trim();
+            this.radio('node_end', {
+                description: writer.label,
+                agent: writer.kind,
+                detail: this.aborted
+                    ? 'verify FAIL left unresolved (run cancelled)'
+                    : `verify FAIL left unresolved (rework budget ${this.maxReviewRounds} spent)`,
+                ok: false,
+            });
+            return;
+        }
+        this.reviewRounds.set(root, spent + 1);
+        this.spawnReworkPair(graph, writer, verify, findings, root, spent + 1);
+    }
+    /**
+     * The writer whose work a `verify` node judged.
+     *
+     * Walks up through non-writer deps, the same shape `collectWorktreeSources`
+     * relies on: a verify's dep is normally its writer directly, but after a
+     * rework round the chain is writer → verify → rework → verify, and the
+     * rework (a `fix` node) is itself the writer to send back.
+     */
+    writerBehind(verify, graph) {
+        const seen = new Set();
+        const visit = (id) => {
+            if (seen.has(id))
+                return undefined;
+            seen.add(id);
+            const n = graph.nodes.get(id);
+            if (!n || n.kind === 'merge')
+                return undefined;
+            if (n.kind === 'general' || n.kind === 'fix')
+                return n;
+            for (const dep of n.deps) {
+                const found = visit(dep);
+                if (found)
+                    return found;
+            }
+            return undefined;
+        };
+        for (const depId of verify.deps) {
+            const found = visit(depId);
+            if (found)
+                return found;
+        }
+        return undefined;
+    }
+    /**
+     * Send a writer's work back for one more round: a rework node carrying the
+     * verify's findings, plus a fresh verify to judge the result.
+     *
+     * The rework runs INSIDE the writer's worktree instead of creating one of
+     * its own. Two worktrees for one scope means two branches, and the merge
+     * node walks up to the writer — so a rework on its own branch would be
+     * merged never or twice, exactly the stranded-work failure the merge fix
+     * addressed. `allowWorktree: false` on this node's deps suppresses creation,
+     * and `cwdOverride` (resolved via {@link inheritedWorktreeCwdFor}) points it
+     * at the existing tree; the handle stays registered against the writer.
+     *
+     * Acyclicity is preserved by construction: both new nodes point only at
+     * nodes that already exist, and the rewiring moves an existing edge forward
+     * along the chain rather than back into it.
+     */
+    spawnReworkPair(graph, writer, verify, findings, root, round) {
+        // Named after the lineage root, so round 2 of g1 is `rework-g1-2` rather
+        // than a nested `rework-rework-g1-1-1`.
+        const reworkId = `rework-${root}-${round}`;
+        const reworkNode = {
+            id: reworkId,
+            kind: 'fix',
+            label: `rework: ${writer.label}`,
+            prompt: `A reviewer inspected this work on disk and REJECTED it. Address every finding below, ` +
+                `then leave the work in a state that satisfies the original task.\n\n` +
+                `## Original task\n${writer.prompt}\n\n` +
+                `## Reviewer findings (these are what must change)\n${findings || '(the reviewer reported FAIL without detail)'}`,
+            ...(writer.scope ? { scope: writer.scope } : {}),
+            ...(writer.acceptance ? { acceptance: writer.acceptance } : {}),
+            // The verify is already `done`, so the rework is immediately ready.
+            deps: [verify.id],
+            status: 'pending',
+            retryCount: 0,
+            maxRetries: 0,
+        };
+        graph.nodes.set(reworkId, reworkNode);
+        this.reworks.set(reworkId, writer.id);
+        this.reviewLineage.set(reworkId, root);
+        const reVerifyId = `verify-${reworkId}`;
+        const reVerifyNode = {
+            id: reVerifyId,
+            kind: 'verify',
+            label: `verify: ${writer.label} (rework ${round})`,
+            prompt: verify.prompt,
+            deps: [reworkId],
+            status: 'pending',
+            retryCount: 0,
+            maxRetries: verify.maxRetries,
+        };
+        graph.nodes.set(reVerifyId, reVerifyNode);
+        // Whatever waited on the old verify (typically the merge node) must now
+        // wait on the new one, or it would merge the branch mid-rework.
+        for (const other of graph.nodes.values()) {
+            if (other.id === reworkId || other.id === reVerifyId)
+                continue;
+            if (other.deps.includes(verify.id)) {
+                other.deps = other.deps.map((d) => (d === verify.id ? reVerifyId : d));
+            }
+        }
+        this.radio('node_fix', {
+            description: reworkNode.label,
+            agent: 'fix',
+            detail: `verify FAIL on "${writer.label}" — rework round ${round}/${this.maxReviewRounds}`,
+            ok: false,
+        });
+    }
     /**
      * Create a `fix` node that attempts to redo the failed node's work, wired
      * so downstream dependents of the failed node also wait on the fix.
@@ -477,9 +1095,12 @@ export class KrakenGraphExecutor {
             // no further retries — one fix attempt per failed node in v1
         };
         graph.nodes.set(fixId, fixNode);
+        this.repairs.set(fixId, failed.id);
         // Downstream dependents wait on the fix attempt INSTEAD of the failed
-        // node (not in addition to it) — the failed node never becomes `done`,
-        // so leaving it in `deps` would strand dependents forever.
+        // node (not in addition to it): the failed node is terminally `error` and
+        // will never run again, so leaving it in `deps` would strand them forever.
+        // (If the fix succeeds, `reconcileRepairedNode` marks the failed node
+        // `done` for reporting — after this rewiring, and with no effect on it.)
         for (const other of graph.nodes.values()) {
             if (other.id === fixId)
                 continue;