peaks-loop 4.0.0-beta.3 → 4.0.0-beta.31

package/dist/cli/commands/dispatch-commands.js

@@ -0,0 +1,380 @@
+/**
+ * `peaks sub-agent dispatch <role> ...` — slice 2026-06-07-sub-agent-context-governance.
+ *
+ * Pulled out of `sub-agent-commands.ts` (slice 2026-06-23-audit-p0-split) to
+ * honor the 800-line file cap (Karpathy #2 Simplicity First). The single
+ * `dispatch` action lives here; the `--from-dag` sibling was further split
+ * into `dispatch-from-dag.ts` (slice 2026-06-23-audit-3rd #7) because the
+ * two paths share no logic and the `--from-dag` codepath loads three heavy
+ * modules on first call (slice 9 perf) that the warm-path single-dispatch
+ * never touches.
+ *
+ * Skill-first / CLI-auxiliary red line (PB-4 / AC-19/20): this command is
+ * a primitive that the peaks-code / peaks-rd / peaks-qa SKILL.md compose.
+ * Users do NOT invoke it directly; the --help text and dispatch
+ * envelope's `nextActions` reinforce the point.
+ */
+import { existsSync } from 'node:fs';
+import { randomUUID } from 'node:crypto';
+import { fail, getErrorMessage, ok } from 'peaks-loop-shared/result';
+import { addJsonOption, printResult } from '../cli-helpers.js';
+import { detectInstalledIde } from '../../services/ide/ide-detector.js';
+import { getAdapter } from '../../services/ide/ide-registry.js';
+import { SubAgentNotSupportedError } from '../../services/dispatch/sub-agent-dispatcher.js';
+import { emitObservabilityEvent, OBSERVABILITY_SUBAGENT_ROLES } from '../../services/observability/observability-service.js';
+import { noteDispatched, BATCH_LIMIT } from '../../services/dispatch/batch-counter.js';
+import { writeInitialDispatchRecord } from '../../services/dispatch/dispatch-record-writer.js';
+import { evaluatePromptSize } from '../../services/context/context-guard.js';
+import { getCurrentSessionId } from '../../services/skills/skill-presence-service.js';
+import { buildArtifactMeta, buildContextImpact } from '../../services/context/artifact-meta.js';
+import { assertSafeArtifactPath } from 'peaks-loop-shared-channel';
+import { compressPrompt } from '../../services/context/headroom-client.js';
+import { resolveHeadroomOptions } from '../../services/context/headroom-prefs.js';
+import { loadPreferences } from '../../services/preferences/preferences-service.js';
+import { DEFAULT_PREFERENCES } from '../../services/preferences/preferences-types.js';
+import { writeLogEntry } from '../../services/log/logger.js';
+import { HEADROOM_MODES, PROMPT_LIMIT_BYTES, RECOMMENDED_ROLES, validateRole } from './sub-agent-shared.js';
+import { runDispatchFromDag } from './dispatch-from-dag.js';
+import { TEST_TOOL_DETECTION_BLOCK, formatTestToolDetection } from '../../services/dispatch/test-tool-detection.js';
+import { MemoryPreflightService } from '../../services/context/memory-preflight-service.js';
+import { buildDispatchSystemPrompt } from '../../services/context/build-dispatch-system-prompt.js';
+export function registerDispatchCommand(parent, io) {
+    addJsonOption(parent
+        .command('dispatch')
+        .description('Build an IDE-specific tool-call descriptor for a sub-agent dispatch. ' +
+        'Dry-run by design; the LLM executes the returned toolCall in its own ' +
+        'environment. Flags: --write-artifact (G7), --use-headroom (G7.7), ' +
+        '--force (G9 CLI 兜底). ' +
+        'See skills/peaks-code/references/sub-agent-dispatch.md for the ' +
+        'orchestrator contract.')
+        .argument('<role>', 'sub-agent role (e.g. rd | qa | ui | txt | qa-business | qa-business-api)')
+        // 2.7.0 slice-dag-dispatcher MVP: --prompt is required ONLY when --from-dag is NOT
+        // supplied. Previously this was `.requiredOption('--prompt')`, which blocked
+        // `dispatch --from-dag <file>` calls because commander.js validates
+        // `.requiredOption` before the action handler runs. The mutual-exclusion
+        // check is enforced below in the action body (--prompt XOR --from-dag).
+        .option('--prompt <text>', 'the prompt to send to the sub-agent (required unless --from-dag is provided)')
+        .option('--prompt-length <bytes>', 'DOGFOOD ONLY: synthesize a prompt of this size (overrides --prompt content for size only; content is "x" repeated)')
+        .option('--request-id <rid>', 'the same <rid> used by peaks request init')
+        .option('--session-id <sid>', 'override active session id (default: resolve from .peaks/_runtime/session.json; falls back to PEAKS_SESSION_ID env var; final fallback "unknown-sid")')
+        .option('--project <path>', 'target project root (defaults to cwd)')
+        .option('--batch-id <uuid>', 'batch id for the dispatch (default: auto-generated UUID)')
+        .option('--write-artifact <path>', 'G7: register an artifact file at <path>; CLI computes sha256 + size + writes ArtifactMeta to the dispatch record')
+        .option('--use-headroom', 'G7.7/G9: compress the prompt via headroom-ai before dispatch (opt-in; falls back to G7 metadata-only if headroom unavailable)')
+        .option('--headroom-mode <mode>', `G7.7: headroom mode (${HEADROOM_MODES.join(' | ')}); default balanced`)
+        .option('--force', 'G9: override the 80% hard reject threshold at CLI (NOT allowed at hook layer per RL-30 strict)')
+        .option('--from-dag <file>', '2.7.0 slice-dag-dispatcher MVP: read a SliceDag JSON file, dispatch one sub-agent per node in topological order; --batch-id overrides the auto-generated batch id (mutually exclusive with <role>)')).action(async (role, options) => {
+        const asJson = options.json === true;
+        // 2.7.0 slice-dag-dispatcher MVP: --from-dag short-circuits the single
+        // sub-agent path and runs the full DAG plan via `dag-orchestrator`.
+        if (typeof options.fromDag === 'string' && options.fromDag.length > 0) {
+            await runDispatchFromDag(role, options, asJson, io);
+            return;
+        }
+        const validation = validateRole(role);
+        if (validation !== null) {
+            printResult(io, fail('sub-agent.dispatch', 'INVALID_ROLE', validation, { role, toolCall: null, dispatchRecordPath: null }, [
+                'Use a non-empty role string with no control characters.',
+                `Recommended: ${RECOMMENDED_ROLES}.`
+            ]), asJson);
+            process.exitCode = 1;
+            return;
+        }
+        // Slice 3 (on-demand-ecc) D-012: the `agent` role was removed in
+        // 4.0.0-beta.11 — there is no longer a subprocess path for it
+        // (the upstream ECC v2.0.0 ships no `ecc` binary). This guard
+        // sits AFTER role validation but BEFORE the missing-prompt
+        // check so an action-path dispatch with a valid prompt still
+        // returns a clear ROLE_REMOVED envelope + exit 1. Note that
+        // Commander short-circuits `--help` BEFORE `.action()` fires,
+        // so `peaks sub-agent dispatch agent --help` continues to
+        // exit 0 with the help text — that is intentional, not a bug.
+        if (role === 'agent') {
+            printResult(io, fail('sub-agent.dispatch', 'ROLE_REMOVED', 'The agent role was removed in Slice 3', { role, reason: 'role-removed-in-slice-3', toolCall: null, dispatchRecordPath: null }, []), asJson);
+            process.exitCode = 1;
+            return;
+        }
+        if (!options.prompt || options.prompt.length === 0) {
+            printResult(io, fail('sub-agent.dispatch', 'MISSING_PROMPT', '--prompt is required when --from-dag is not provided', { role, toolCall: null, dispatchRecordPath: null }, [
+                'Re-run with either:',
+                '  • `--prompt <text>` for single-role dispatch, OR',
+                '  • `--from-dag <file>` for DAG-aware multi-slice dispatch (no --prompt needed; the per-slice prompt is generated from the DAG nodes).'
+            ]), asJson);
+            process.exitCode = 1;
+            return;
+        }
+        // DOGFOOD ONLY: --prompt-length overrides the actual prompt content with
+        // a synthetic prompt of the given size in bytes. The original --prompt
+        // is still required (commander needs it). This avoids ARG_MAX limits
+        // on Windows when the dogfood prompt is > 200KB.
+        if (typeof options.promptLength === 'string' && options.promptLength.length > 0) {
+            const len = Number.parseInt(options.promptLength, 10);
+            if (Number.isInteger(len) && len > 0) {
+                options.prompt = 'x'.repeat(len);
+            }
+        }
+        if (options.prompt.length + TEST_TOOL_DETECTION_BLOCK.length > PROMPT_LIMIT_BYTES) {
+            printResult(io, fail('sub-agent.dispatch', 'PROMPT_TOO_LARGE', `prompt exceeds ${PROMPT_LIMIT_BYTES} bytes (got ${options.prompt.length})`, { role, toolCall: null, dispatchRecordPath: null }, [
+                'Truncate the prompt or split into multiple dispatches.',
+                'Pass --force to override the 80% threshold at CLI (NOT allowed at hook layer).'
+            ]), asJson);
+            process.exitCode = 1;
+            return;
+        }
+        // G9 CLI 兜底 — evaluate prompt size against the threshold table.
+        const decision = evaluatePromptSize(options.prompt.length, { force: options.force === true });
+        if (!decision.allow) {
+            printResult(io, fail('sub-agent.dispatch', decision.code, `prompt size ${options.prompt.length} bytes exceeds threshold (tier=${decision.evaluation.tier}, ratio=${decision.evaluation.ratio.toFixed(3)})`, {
+                role,
+                toolCall: null,
+                dispatchRecordPath: null
+            }, [
+                decision.suggest ?? 'Trim prompt or pass --force to override at CLI.',
+                'PreToolUse hook layer will still reject regardless of --force (RL-30 strict).'
+            ]), asJson);
+            process.exitCode = 1;
+            return;
+        }
+        try {
+            const projectRoot = options.project ?? process.cwd();
+            // Slice 2026-06-26-unknown-sid-fallback-fix: when --session-id is not
+            // passed, auto-resolve the active peaks session id from
+            // `.peaks/_runtime/session.json` (or PEAKS_SESSION_ID env var) so
+            // dispatch records land in `.peaks/_sub_agents/<real-sid>/` instead
+            // of the `unknown-sid` fallback. The unknown-sid branch is preserved
+            // as the last-resort so callers without a bound session (e.g. an
+            // ad-hoc dispatch in a fresh tree) still get a deterministic path.
+            const sid = options.sessionId
+                ?? process.env.PEAKS_SESSION_ID
+                ?? getCurrentSessionId(projectRoot)
+                ?? 'unknown-sid';
+            const rid = options.requestId ?? 'unknown-rid';
+            const batchId = options.batchId ?? randomUUID();
+            // G7.7 / G9: resolve headroom options from preferences + CLI overrides.
+            // Preferences hard-block when headroom.enabled=false (returns HEADROOM_DISABLED_BY_PREFERENCE).
+            // loadPreferences can throw on schema mismatch; we fall back to defaults to avoid
+            // breaking the dispatch on a stale preferences.json file.
+            let projectPrefs = DEFAULT_PREFERENCES;
+            let headroomPrefs = DEFAULT_PREFERENCES.headroom;
+            try {
+                projectPrefs = loadPreferences(projectRoot);
+                headroomPrefs = projectPrefs.headroom;
+            }
+            catch { // TODO(g2): legacy silent catch — grace: 1 minor release (v2.14.0)
+                // Keep default preferences; the user can re-run with explicit --headroom-mode
+                // if they want to override the fallback.
+            }
+            const headroomResolved = resolveHeadroomOptions(headroomPrefs, {
+                useHeadroom: options.useHeadroom === true,
+                ...(options.headroomMode !== undefined ? { headroomMode: options.headroomMode } : {})
+            });
+            if (headroomResolved.blocked !== null) {
+                printResult(io, fail('sub-agent.dispatch', headroomResolved.blocked, `headroom integration is disabled in preferences (headroom.enabled=false); pass --headroom-mode and update preferences first, or run without --use-headroom`, {
+                    role,
+                    toolCall: null,
+                    dispatchRecordPath: null
+                }, [
+                    'Edit .peaks/preferences.json: set headroom.enabled = true (per-touchpoint mode is headroom.perTouchpoint.subAgentDispatch).',
+                    'Or re-run without --use-headroom to dispatch without compression.'
+                ]), asJson);
+                process.exitCode = 1;
+                return;
+            }
+            const ide = detectInstalledIde(projectRoot) ?? 'claude-code';
+            const adapter = getAdapter(ide);
+            if (!adapter.subAgentDispatcher.supportsRole(role)) {
+                printResult(io, fail('sub-agent.dispatch', 'IDE_NOT_SUPPORTED', `IDE ${ide} does not support role "${role}"`, { role, toolCall: null, dispatchRecordPath: null }, [
+                    'Switch to a registered IDE (e.g. claude-code) or pick a role the current IDE supports.'
+                ]), asJson);
+                process.exitCode = 1;
+                return;
+            }
+            // G7.7 headroom compress (opt-in). If headroom fails or is unavailable,
+            // fall back to the original prompt + emit warning.
+            // Slice 2026-07-22-orchestrator-memory-preflight (Task 5): prepend the
+            // memory preflight block (or silently skip when unavailable) via the
+            // pure-function builder, BEFORE the headroom-ai compress step so the
+            // compressor sees the augmented payload.
+            const preflightService = new MemoryPreflightService(projectRoot, projectPrefs);
+            const memoryBlock = await preflightService.fetchBlock(role);
+            const memoryAugmentedBody = buildDispatchSystemPrompt({
+                taskTitle: role,
+                taskBody: options.prompt,
+                memoryBlock,
+            });
+            let effectivePrompt = `${formatTestToolDetection()}\n\n${memoryAugmentedBody}`;
+            let headroomCompressed = false;
+            let headroomResult = null;
+            const warnings = [...decision.warnings];
+            if (headroomResolved.mode !== null) {
+                headroomResult = await compressPrompt(effectivePrompt, headroomResolved.mode);
+                if (headroomResult.warning !== null) {
+                    warnings.push(headroomResult.warning);
+                }
+                if (headroomResult.compressed && headroomResult.compressedPrompt !== null) {
+                    effectivePrompt = headroomResult.compressedPrompt;
+                    headroomCompressed = true;
+                }
+            }
+            let toolCall;
+            try {
+                toolCall = adapter.subAgentDispatcher.buildToolCall({ role, prompt: effectivePrompt, requestId: rid, sessionId: sid });
+                // Slice C of v2.11.1 — observability hook #2/7. Fire-and-forget
+                // per PRD Q4 (full-auto must never fail-loud). The
+                // synchronous emit returns {written:false} on disk-full; we
+                // deliberately swallow the result so dispatch contract is
+                // unchanged. role is only included when it matches the schema's
+                // known sub-agent role set; otherwise it's omitted (non-standard
+                // roles like 'qa-business' would otherwise drop the event
+                // through schema rejection).
+                const KNOWN_ROLES = new Set(OBSERVABILITY_SUBAGENT_ROLES);
+                const knownRole = KNOWN_ROLES.has(role) ? role : null;
+                emitObservabilityEvent({
+                    schemaVersion: 1,
+                    ts: new Date().toISOString(),
+                    sessionId: sid,
+                    category: 'dispatch',
+                    ...(knownRole !== null ? { role: knownRole } : {}),
+                    detail: {
+                        requestId: rid,
+                        ide: adapter.subAgentDispatcher.label,
+                        promptBytes: effectivePrompt.length,
+                        headroomCompressed
+                    }
+                }, { projectRoot });
+            }
+            catch (error) {
+                if (error instanceof SubAgentNotSupportedError) {
+                    printResult(io, fail('sub-agent.dispatch', 'IDE_NOT_SUPPORTED', error.message, { role, toolCall: null, dispatchRecordPath: null }, [
+                        'Switch IDE or pick a role the current IDE supports.'
+                    ]), asJson);
+                    process.exitCode = 1;
+                    return;
+                }
+                throw error;
+            }
+            // G7 — optional --write-artifact: build ArtifactMeta, attach to record.
+            let artifactMeta = null;
+            if (typeof options.writeArtifact === 'string' && options.writeArtifact.length > 0) {
+                try {
+                    assertSafeArtifactPath(options.writeArtifact, projectRoot);
+                    if (!existsSync(options.writeArtifact)) {
+                        warnings.push('ARTIFACT_NOT_FOUND');
+                    }
+                    else {
+                        artifactMeta = buildArtifactMeta({
+                            path: options.writeArtifact,
+                            rid,
+                            role,
+                            idx: 1, // single dispatch, idx=1
+                            summary: null
+                        });
+                    }
+                }
+                catch (err) {
+                    warnings.push(`ARTIFACT_PATH_INVALID: ${getErrorMessage(err)}`);
+                }
+            }
+            const { path: dispatchRecordPath } = writeInitialDispatchRecord({
+                projectRoot,
+                sessionId: sid,
+                requestId: rid,
+                role,
+                prompt: effectivePrompt,
+                toolCall,
+                batchId
+            });
+            const counter = noteDispatched(projectRoot, sid, batchId);
+            if (counter.warning) {
+                warnings.push(counter.warning.message);
+            }
+            const contextImpact = buildContextImpact({
+                promptSize: effectivePrompt.length,
+                artifactSizes: artifactMeta ? [artifactMeta.size] : []
+            });
+            const nextActions = [
+                'Tool call is dry-run; LLM must execute the tool to actually dispatch the sub-agent.',
+                'After dispatching, the sub-agent should call `peaks sub-agent heartbeat --record ' + dispatchRecordPath + '` periodically.'
+            ];
+            if (counter.warning) {
+                nextActions.push(`Batch is over the RL-1 limit (${BATCH_LIMIT}); consider splitting into multiple batches.`);
+            }
+            if (headroomResult && headroomResult.warning === 'HEADROOM_UNAVAILABLE') {
+                nextActions.push('Headroom daemon unavailable; dispatched with G7 metadata-only fallback.');
+            }
+            printResult(io, ok('sub-agent.dispatch', {
+                // Slice 2026-06-23-audit-4th #E1: every CLI envelope carries
+                // an envelopeVersion marker so consumers can detect contract
+                // changes (the previous #4 dropped `data.prompt` silently).
+                envelopeVersion: '2.2.0',
+                role,
+                ide: adapter.subAgentDispatcher.label,
+                // Slice 2026-06-23-audit-3rd #4: do NOT echo `prompt` in stdout.
+                // Prompts can carry user content (sometimes test credentials /
+                // internal URLs) that has no business landing in shell history,
+                // log aggregators, or tmux scrollback. The dispatch record on
+                // disk (gitignored under .peaks/_sub_agents/) keeps the prompt
+                // for the sub-agent to read; CLI stdout stays metadata-only.
+                // Surface promptSize + originalPromptSize so the LLM-side
+                // runner can still reason about headroom without seeing the
+                // content.
+                originalPromptSize: options.prompt.length,
+                promptSize: effectivePrompt.length,
+                toolCall,
+                dispatchRecordPath,
+                batchId,
+                dispatchedInBatch: counter.count,
+                headroomCompressed,
+                headroomResult: headroomResult
+                    ? {
+                        mode: headroomResult.mode,
+                        compressed: headroomResult.compressed,
+                        compressionRatio: headroomResult.compressionRatio,
+                        tokensSaved: headroomResult.tokensSaved,
+                        warning: headroomResult.warning
+                    }
+                    : null,
+                forcedAt: decision.forcedAt,
+                contextImpact,
+                artifactMetas: artifactMeta ? [artifactMeta] : []
+            }, warnings, nextActions), asJson);
+            // Slice 2026-06-23-audit-4th #B1: structured log on success path.
+            // Best-effort: writeLogEntry swallows its own errors (logger.ts:155-159),
+            // so a full disk or missing ~/.peaks/logs/ dir never blocks the dispatch.
+            try {
+                writeLogEntry({
+                    ts: new Date().toISOString(),
+                    level: 'info',
+                    command: 'sub-agent.dispatch',
+                    msg: 'dispatched',
+                    sessionId: sid,
+                    batchId,
+                    data: {
+                        rid,
+                        role,
+                        batchId,
+                        dispatchedInBatch: counter.count,
+                        headroomCompressed,
+                        forcedAt: decision.forcedAt
+                    }
+                });
+            }
+            catch { // TODO(g2): legacy silent catch — grace: 1 minor release (v2.14.0)
+                /* best-effort */
+            }
+        }
+        catch (error) {
+            printResult(io, fail('sub-agent.dispatch', 'DISPATCH_ERROR', getErrorMessage(error), { role, toolCall: null, dispatchRecordPath: null }, [
+                'See error message; if you are dispatching from a SKILL.md, the LLM should retry with a smaller prompt or pick a different role.'
+            ]), asJson);
+            process.exitCode = 1;
+        }
+    });
+}
+/**
+ * 2.7.0 slice-dag-dispatcher MVP — see `dispatch-from-dag.ts`.
+ * The function was pulled out of this file in slice 2026-06-23-audit-3rd
+ * #7 to honor the 800-line file cap and isolate the three heavy
+ * module loads (slice-dag / dag-orchestrator / contract-store) to the
+ * --from-dag codepath only.
+ */

package/dist/cli/commands/dispatch-from-dag.js

@@ -0,0 +1,251 @@
+/**
+ * `peaks sub-agent dispatch --from-dag <file>` — slice 2026-06-23-audit-3rd #7.
+ *
+ * Pulled out of `dispatch-commands.ts` to honor the 800-line file cap
+ * (Karpathy #2 Simplicity First). The single-dispatch action stays in
+ * `dispatch-commands.ts`; the DAG-dispatch path lives here because the
+ * two paths share no logic — the warm-path single dispatch does NOT
+ * load slice-dag / dag-orchestrator / contract-store (slice 9 perf),
+ * while the --from-dag codepath loads all three on first call.
+ *
+ * 2.7.0 slice-dag-dispatcher MVP: read a SliceDag from a file and run
+ * it through `runDag`. The orchestrator's `runSlice` is a thin wrapper
+ * that emits the per-IDE `buildToolCall` envelope for each topological
+ * level.
+ *
+ * 1.2 MVP scope: this dispatches the FIRST topological level synchronously
+ * (returns the dispatch specs); subsequent levels are not surfaced in
+ * the CLI envelope because the LLM-side runner must actually execute
+ * each `buildToolCall` and write the resulting contract before level 2+
+ * can be safely auto-advanced. The LLM re-invokes
+ * `peaks sub-agent dispatch --from-dag <file> --batch-id <id>` after
+ * writing level-1 contracts, which causes runDag to plan level 2+ with
+ * the level-1 ancestor contracts spliced into their dispatch prompts.
+ *
+ * Internally, runDag DOES iterate all levels so its join-barrier +
+ * cancel-on-fail path is exercised end-to-end during the MVP emit; only
+ * the CLI envelope is filtered to level-1 toolCalls (see `firstLevelIds`).
+ */
+import { randomUUID } from 'node:crypto';
+import { printResult } from '../cli-helpers.js';
+import { fail, ok } from 'peaks-loop-shared/result';
+import { detectInstalledIde } from '../../services/ide/ide-detector.js';
+import { getAdapter } from '../../services/ide/ide-registry.js';
+import { getCurrentSessionId } from '../../services/skills/skill-presence-service.js';
+export async function runDispatchFromDag(role, options, asJson, io) {
+    if (!options.fromDag)
+        return;
+    const projectRoot = options.project ?? process.cwd();
+    // Slice 2026-06-26-unknown-sid-fallback-fix: see dispatch-commands.ts.
+    // Auto-resolve sid from .peaks/_runtime/session.json before falling back.
+    const sid = options.sessionId
+        ?? process.env.PEAKS_SESSION_ID
+        ?? getCurrentSessionId(projectRoot)
+        ?? 'unknown-sid';
+    const rid = options.requestId ?? 'unknown-rid';
+    const batchId = options.batchId ?? randomUUID();
+    // Slice 9 (dispatch CLI latency): the --from-dag codepath loads three
+    // heavy modules (slice-dag, dag-orchestrator, contract-store) that
+    // transitively import 10+ more services. They are NOT touched by the
+    // warm-path single-dispatch action, so we lazy-load them here only
+    // when --from-dag is actually requested. The dynamic-import promise
+    // resolves on first call, then ESM caches the module for subsequent
+    // calls in the same process.
+    const [{ readFileSync }, sliceDagMod, dagOrchestratorMod, contractStoreMod] = await Promise.all([
+        import('node:fs'),
+        import('../../services/dispatch/slice-dag.js'),
+        import('../../services/code/dag-orchestrator.js'),
+        import('../../services/dispatch/contract-store.js')
+    ]);
+    const { validateDag, topologicalLevels, isSliceComplexity } = sliceDagMod;
+    const { runLayeredDag } = dagOrchestratorMod;
+    const { listContracts, hashContract } = contractStoreMod;
+    let dag;
+    try {
+        const raw = readFileSync(options.fromDag, 'utf8');
+        const parsed = JSON.parse(raw);
+        validateDag(parsed);
+        dag = parsed;
+    }
+    catch (err) {
+        printResult(io, fail('sub-agent.dispatch', 'INVALID_DAG', `failed to read or validate DAG from ${options.fromDag}: ${err.message}`, { role, toolCall: null, dispatchRecordPath: null }, [
+            'Check the JSON file at the given path; the DAG must have {nodes, edges} and pass validateDag().'
+        ]), asJson);
+        process.exitCode = 1;
+        return;
+    }
+    // MVP (1.2): the orchestrator iterates topological levels + join
+    // barrier + cancel-on-fail end-to-end. The CLI's runner returns `done`
+    // for every leaf (so runDag's cancel-on-fail path is exercised when a
+    // leaf actually fails); the CLI envelope filters `emittedToolCalls` to
+    // the first level only (see `firstLevelIds`). The LLM-side runner
+    // re-invokes `--from-dag` for level 2+ after writing level-1
+    // contracts via `peaks contract write`.
+    let levelArr;
+    try {
+        levelArr = topologicalLevels(dag);
+    }
+    catch (err) {
+        printResult(io, fail('sub-agent.dispatch', 'INVALID_DAG', `topologicalLevels failed for ${options.fromDag}: ${err.message}`, { role, toolCall: null, dispatchRecordPath: null }, [
+            'The DAG passed validateDag() but topologicalLevels threw (likely a cycle that slipped past validateDag, or a runtime invariant).',
+            'Inspect the DAG file with the editor and re-invoke dispatch. (No peaks scan dag CLI ships in 2.7.0; if you need a programmatic DAG validator, import validateDag / topologicalLevels from src/services/dispatch/slice-dag.ts directly.)'
+        ]), asJson);
+        process.exitCode = 1;
+        return;
+    }
+    // Read upstream contracts so downstream-level prompts (when re-invoked
+    // for level 2+) get auto-injected ancestors via `formatContractInjection`.
+    // At level-1 emit time, contracts is empty (no ancestors yet); the
+    // injection matters when the LLM re-invokes for level 2+ after writing
+    // level-1 contracts via `writeContract`.
+    const existingContracts = listContracts(projectRoot, sid);
+    const ide = detectInstalledIde(projectRoot) ?? 'claude-code';
+    const adapter = getAdapter(ide);
+    const dispatcher = adapter.subAgentDispatcher;
+    if (!dispatcher.supportsRole(role)) {
+        printResult(io, fail('sub-agent.dispatch', 'IDE_NOT_SUPPORTED', `IDE ${ide} does not support role "${role}"`, { role, toolCall: null, dispatchRecordPath: null }, [
+            'Switch to a registered IDE (e.g. claude-code) or pick a role the current IDE supports.'
+        ]), asJson);
+        process.exitCode = 1;
+        return;
+    }
+    // Track per-level emissions so the CLI envelope can expose ONLY the
+    // first topological level to the LLM (matching MVP 1.2 semantics: LLM
+    // executes level-1 toolCalls, writes contracts via `peaks contract
+    // write`, then re-invokes `--from-dag` for level 2+). Level-2+ leaves
+    // still flow through runDag's join-barrier — their toolCalls are
+    // generated internally so the orchestrator's happy path is exercised
+    // end-to-end, but they are NOT surfaced in the CLI envelope (the LLM
+    // sees them only after re-invoking with fresh level-1 contracts).
+    const firstLevelIds = new Set(levelArr[0] ?? []);
+    const emittedToolCalls = [];
+    const emittedSliceIds = [];
+    // CLI runner: emit the per-slice `buildToolCall` envelope, then return
+    // `done` with an empty `publicSurface` placeholder. The orchestrator
+    // treats the slice as "CLI-emitted, awaiting LLM-side execution" and
+    // advances to the next topological level (or breaks on failure). The
+    // placeholder surface is intentionally empty: the real public surface
+    // arrives later via `peaks contract write` from the LLM-side runner,
+    // and overwrites the CLI placeholder contract written by `noopWriter`.
+    // Why not `cancelled`? Returning `cancelled` here short-circuits
+    // runDag's join-barrier + cancel-on-fail path — the orchestrator
+    // breaks out after the first level with a cancellation, which hides
+    // the real cancel-on-fail semantics from the test suite (and from any
+    // future caller that wants to use `runDag` end-to-end from the CLI).
+    // Returning `done` exercises the full happy path; cancel-on-fail
+    // semantics will be covered when 1.3 lands real per-IDE await.
+    const cliRunner = async (spec) => {
+        const toolCall = dispatcher.buildToolCall({
+            role: spec.role,
+            prompt: spec.prompt,
+            requestId: rid,
+            sessionId: sid
+        });
+        if (firstLevelIds.has(spec.sliceId)) {
+            emittedToolCalls.push(toolCall);
+            emittedSliceIds.push(spec.sliceId);
+        }
+        return {
+            status: 'done',
+            publicSurface: {
+                exports: [],
+                types: [],
+                publicSignatures: []
+            }
+        };
+    };
+    // CLI noop writer — the orchestrator collects a `SliceContract` per
+    // emitted toolCall so downstream re-invocations (level 2+) can splice
+    // ancestor contracts via `formatContractInjection`. We don't write to
+    // disk here (the LLM-side runner does that via `peaks contract write`
+    // after each toolCall resolves), but the returned placeholder MUST be
+    // shape-valid: the `contractHash` field has to be a real SHA-256 hex
+    // string so formatContractInjection + downstream validators don't
+    // reject the placeholder. The placeholder hash is computed from the
+    // slice identity only (`sliceId|sessionId`) — it's deterministic,
+    // stable, and collision-free for the MVP run, but obviously does NOT
+    // represent the slice's actual public surface. The LLM-side runner's
+    // `peaks contract write` call will overwrite this with a content-based
+    // hash once the slice finishes.
+    const noopWriter = (sliceId, _publicSurface) => {
+        const partial = {
+            sliceId,
+            sessionId: sid,
+            exports: [],
+            types: [],
+            publicSignatures: []
+        };
+        return {
+            sliceId,
+            sessionId: sid,
+            completedAt: new Date(0).toISOString(), // epoch sentinel = "placeholder, not yet finished"
+            exports: [],
+            types: [],
+            publicSignatures: [],
+            contractHash: hashContract(partial)
+        };
+    };
+    // runLayeredDag IS in the path: validates the DAG, iterates all
+    // topological levels (foundation/upstreamSync slices prioritized within
+    // each level by `topologicalLevels`), runs the join barrier after
+    // each level, and triggers cancel-on-fail rollback if any leaf
+    // returns `failed` (or `cancelled`, once 1.3 lands). The CLI
+    // envelope below filters `emittedToolCalls` / `emittedSliceIds` to
+    // the first level — see the `firstLevelIds` filter inside `cliRunner`.
+    try {
+        await runLayeredDag(dag, {
+            projectRoot,
+            sessionId: sid,
+            existingContracts,
+            runSlice: cliRunner,
+            writeContractFn: noopWriter
+        });
+    }
+    catch (err) {
+        // runLayeredDag throws DagPlanError for plan-level failures (e.g.
+        // cycle slipping past validateDag). Surface it as INVALID_DAG so
+        // the CLI envelope has a clear failure code, not a generic
+        // DISPATCH_ERROR.
+        printResult(io, fail('sub-agent.dispatch', 'INVALID_DAG', `runLayeredDag failed for ${options.fromDag}: ${err.message}`, { role, toolCall: null, dispatchRecordPath: null }, [
+            'runLayeredDag threw DagPlanError; the DAG passed validateDag() but failed at topologicalLevels or contractStore dispatch.',
+            'Inspect the DAG file and re-run.'
+        ]), asJson);
+        process.exitCode = 1;
+        return;
+    }
+    printResult(io, ok('sub-agent.dispatch', {
+        // Slice 2026-06-23-audit-4th #E1: envelopeVersion marker
+        envelopeVersion: '2.1.0',
+        role,
+        ide: dispatcher.label,
+        fromDag: options.fromDag,
+        batchId,
+        dispatchCount: emittedSliceIds.length,
+        levelsTotal: levelArr.length,
+        firstLevel: emittedSliceIds,
+        toolCalls: emittedToolCalls,
+        existingContractCount: existingContracts.length,
+        // v2.15.0 follow-up — G12/G11/G2: per-slice metadata (foundation /
+        // upstreamSync / complexity) for downstream scheduling decisions.
+        sliceMeta: emittedSliceIds.map((id) => {
+            const node = dag.nodes.find((n) => n.id === id);
+            return {
+                id,
+                foundation: node?.foundation === true,
+                upstreamSync: node?.upstreamSync === true,
+                complexity: node?.complexity !== undefined && isSliceComplexity(node.complexity)
+                    ? node.complexity
+                    : null
+            };
+        }),
+        nextActions: [
+            'Execute each toolCall in your IDE; on completion, write the slice contract to .peaks/_runtime/<sid>/dispatch/contracts/<slice-id>.json.',
+            'Re-invoke `peaks sub-agent dispatch --from-dag <file>` with the same batch-id to advance to the next level once all current-level slices have written contracts.'
+        ]
+    }, [], [
+        'MVP (1.2) plans the first level via runLayeredDag orchestrator (cancel-on-fail path active); the LLM drives subsequent levels by re-invoking this command after contract writes.',
+        existingContracts.length > 0
+            ? `Injected ${existingContracts.length} upstream contract(s) into downstream-level prompts via formatContractInjection.`
+            : 'No upstream contracts found; first-level prompts have empty ancestor blocks.'
+    ]), asJson);
+}