Writing Agent Relay Workflows

Overview

The @relayflows/core workflow system orchestrates multiple AI agents (Claude, Codex, Gemini, Aider, Goose) through typed DAG-based workflows. Workflows can be written in TypeScript (preferred), Python, or YAML.

Language preference: TypeScript > Python > YAML. Use TypeScript unless the project is Python-only or a simple config-driven workflow suits YAML.

Pattern selection: Do not default to dag blindly. If the job needs a different swarm/workflow type, consult the choosing-swarm-patterns skill when available and select the pattern that best matches the coordination problem.

When to Use

• Building multi-agent workflows with step dependencies
• Orchestrating different AI CLIs (claude, codex, gemini, aider, goose)
• Creating DAG, pipeline, fan-out, or other swarm patterns
• Needing verification gates, retries, or step output chaining
• Designing product-contract workflows where failing checks should route to agents for repair instead of stopping the run
• Dynamic channel management: agents joining/leaving/muting channels mid-workflow

Non-Negotiable Workflow Checklist

Every generated workflow should satisfy this checklist before it is considered complete:

1. Start with a deterministic, resumable preflight for repository state, credentials, and declared write scope.
2. Pick the coordination shape deliberately: Conversation for non-trivial coordination, Pipeline only for linear one-shot handoffs.
3. Use repairable validation gates: capture red output with failOnError: false, hand it to a repair owner, then rerun the same check.
4. Run the mandatory fresh-eyes loops in order: Claude review/fix/final review/final fix, then Codex review/fix/final review/final fix.
5. Require review fixers to add or update appropriate tests, fixtures, assertions, or deterministic proofs for testable findings.
6. Run final deterministic acceptance after the Codex loop and before commit, PR creation, or handoff.
7. If a real blocker remains, write BLOCKED_NO_COMMIT with exact evidence and skip commit/PR creation instead of crashing the workflow.
8. If the workflow owns shipping, model branch, commit, push, PR creation, and PR URL verification as explicit deterministic steps.

Default Principle: Workflows Repair Before They Fail

The point of an agent team workflow is not to discover a red gate and stop. The point is to capture the failure, route it to the right agent, fix it, and continue toward a shippable result. Author non-trivial workflows as repairable systems:

1. Run deterministic checks as evidence-capturing gates with captureOutput: true.
2. Prefer failOnError: false for intermediate validation gates so the workflow can pass the output to a repair agent.
3. Add a repair step immediately after each red-prone gate. The repair agent reads {{steps.<gate>.output}}, fixes source/tests/config, reruns the same command locally, and exits only after the gate is green or the blocker is external.
4. Keep final acceptance deterministic, but still put an agent repair step before commit/PR creation. If the repair budget is exhausted or a true external blocker remains, write a blocked artifact and skip commit/PR creation; do not let the workflow end as FAILED.
5. Use .reliable() or .repairable() on SDK versions that support it, especially for product-contract workflows. As of AgentWorkforce/relay#827, retry-mode workflows with agents are repair-aware by default, repair agents run before retrying malformed/failed agent steps, and the SDK covers DAG, pipeline, fan-out, worktree-backed, deterministic-only, and agent-plus-gate shapes.

Avoid hard-stop gates (failOnError: true with no repair step) in workflows that are supposed to be self-healing. Even cheap preconditions such as missing credentials, wrong repository, or an unsafe dirty worktree should normally write a clear BLOCKED_* artifact and exit cleanly. For implementation, build, test, lint, schema, artifact, and review failures, model the fix path in the workflow.

Mandatory Fresh-Eyes Review Loops

Every workflow must include two comprehensive fresh-eyes review/fix loops before final acceptance, commit, PR creation, or handoff: first Claude, then Codex. This applies even to small workflows and even when deterministic tests pass. Tests prove commands passed; the fresh-eyes loops make independent agents read the actual resulting files and artifacts as if they did not author them.

The required shape is:

1. claude-review: Claude reads the spec, repo rules, changed files, artifacts, test evidence, and final diff. It must produce a durable review artifact with either actionable findings or an explicit NO_ISSUES_FOUND verdict.
2. claude-fix: a fixer repairs every valid Claude finding, adds or updates appropriate tests/proofs for the fix, reruns the relevant checks, and records what changed. If the review found no issues, it records that no fix was needed.
3. claude-review-final: Claude reviews the post-fix state from scratch. It must not rely on the first review or the fixer's summary.
4. claude-fix-final: if the final Claude review still finds issues, fix them, add or update appropriate tests/proofs, and rerun the checks. If anything cannot be fixed, write a BLOCKED_NO_COMMIT artifact with exact evidence.
5. codex-review: Codex starts after the Claude loop and reviews the post-Claude-fix state from scratch.
6. codex-fix: a fixer repairs every valid Codex finding, adds or updates appropriate tests/proofs for the fix, reruns relevant checks, and records what changed.
7. codex-review-final: Codex reviews the post-fix state from scratch.
8. codex-fix-final: if the final Codex review still finds issues, fix them, add or update appropriate tests/proofs, and rerun checks. If anything cannot be fixed, write BLOCKED_NO_COMMIT.
9. Final acceptance/commit/PR steps depend on the post-Codex-fix review path, not directly on implementation, tests, or the Claude loop.

Because WorkflowBuilder DAGs do not provide an unbounded dynamic while loop, model this as explicit bounded review/fix loops plus a final signoff gate. Inside each fix step, instruct the agent to keep iterating locally: review the finding, edit, add or update appropriate regression tests/proofs, rerun targeted checks, review its own fix, and repeat until that round has no remaining valid issues. For high-risk workflows, add more unrolled review/fix rounds or split the reviews into focused reviewers by subsystem.

Use Claude first and Codex second unless one of those CLIs is unavailable in the target environment. If one is unavailable, write that limitation into the workflow artifact and keep the remaining review loop mandatory.

Review artifacts should use a consistent schema so later steps can act on them deterministically:

verdict: FINDINGS | NO_ISSUES_FOUND | BLOCKED
finding_id: short stable id
severity: blocker | high | medium | low
file: path/to/file
issue: what is wrong
fix_required: concrete change needed
test_required: test, fixture, assertion, or proof command needed
status: open | fixed | wontfix | blocked
evidence: commands run, file paths, or blocker details

Use NO_ISSUES_FOUND only when there are no actionable findings. Use BLOCKED only when the blocker is external or unsafe to resolve inside the workflow.

Choose Your Coordination Style — Conversation vs Pipeline

Before writing the workflow, decide _how the agents will coordinate_. The relay primitive supports two very different shapes, and picking the wrong one wastes the most valuable thing the SDK gives you.

Default to Conversation for any non-trivial work. Pipeline DAGs are simpler to reason about but they do not exercise the relay primitive — they are a Unix pipe with extra steps. If you would happily write the same task as a single shell pipeline, pipeline-shape is fine. Otherwise, you almost certainly want a Conversation shape.

The two shapes can mix within one workflow: pipeline-style deterministic preflight → conversation in the middle → pipeline-style commit-and-PR at the end. See Quick Reference (Conversation) below and [Common Patterns → Interactive Team](#interactive-team-lead--workers-on-shared-channel) for the canonical recipe.

A blunt rule of thumb: if your workflow only uses agent steps with preset: 'worker' chained by {{steps.X.output}}, you are not using the relay — you are using claude -p | codex exec. That may still be the right answer; just make it a deliberate choice.

Quick Reference (Pipeline shape)

Use this when steps are linear, well-specified, and need no agent-to-agent feedback. For anything with iteration, review, or coordination, jump to Quick Reference (Conversation shape) below.

>

Note: examples use ESM import syntax, but workflow execution is always wrapped in an async function. See Failure Prevention → Do not use raw top-level `await` before copy-pasting into CJS or executor-generated files.

import { workflow } from '@relayflows/core';

async function runWorkflow() {
  const result = await workflow('my-workflow')
    .description('What this workflow does')
    .pattern('dag') // or 'pipeline', 'fan-out', etc.
    .channel('wf-my-workflow') // dedicated channel (auto-generated if omitted)
    .maxConcurrency(3)
    .timeout(3_600_000) // global timeout (ms)
    .repairable()

    .agent('lead', { cli: 'claude', role: 'Architect', retries: 2 })
    .agent('worker', { cli: 'codex', role: 'Implementer', retries: 2 })
    .agent('claude-reviewer', {
      cli: 'claude',
      role: 'First-pass fresh-eyes reviewer',
      retries: 1,
      preset: 'reviewer',
    })
    .agent('claude-fixer', { cli: 'claude', role: 'First-pass review-finding fixer', retries: 2 })
    .agent('codex-reviewer', {
      cli: 'codex',
      role: 'Second-pass fresh-eyes reviewer',
      retries: 1,
      preset: 'reviewer',
    })
    .agent('codex-fixer', { cli: 'codex', role: 'Review-finding fixer', retries: 2 })

    .step('preflight', {
      type: 'deterministic',
      command: 'git rev-parse --show-toplevel >/dev/null && echo PREFLIGHT_OK',
      captureOutput: true,
      failOnError: true,
    })
    .step('plan', {
      agent: 'lead',
      dependsOn: ['preflight'],
      task: `Analyze the codebase and produce a plan.`,
      retries: 2,
      verification: { type: 'output_contains', value: 'PLAN_COMPLETE' },
    })
    .step('implement', {
      agent: 'worker',
      task: `Implement based on this plan:\n{{steps.plan.output}}`,
      dependsOn: ['plan'],
      verification: { type: 'exit_code' },
    })
    .step('claude-review', {
      agent: 'claude-reviewer',
      dependsOn: ['implement'],
      task: `Fresh-eyes review the completed workflow output. Read the actual files, diff, repo rules, and available evidence.
Write findings to .workflow-artifacts/my-workflow/claude-review.md.
If there are no actionable issues, write NO_ISSUES_FOUND.`,
      verification: { type: 'exit_code' },
    })
    .step('claude-fix', {
      agent: 'claude-fixer',
      dependsOn: ['claude-review'],
      task: `Read .workflow-artifacts/my-workflow/claude-review.md.
Fix every valid issue, add or update appropriate tests/proofs for the fix, rerun relevant checks, and update .workflow-artifacts/my-workflow/claude-fix.md.
If the review says NO_ISSUES_FOUND, record that no fix was needed.`,
      verification: { type: 'exit_code' },
    })
    .step('claude-review-final', {
      agent: 'claude-reviewer',
      dependsOn: ['claude-fix'],
      task: `Fresh-eyes review the post-fix state from scratch. Do not rely on the prior review or fix summary.
Write .workflow-artifacts/my-workflow/claude-review-final.md with either actionable findings or NO_ISSUES_FOUND.`,
      verification: { type: 'exit_code' },
    })
    .step('claude-fix-final', {
      agent: 'claude-fixer',
      dependsOn: ['claude-review-final'],
      task: `If .workflow-artifacts/my-workflow/claude-review-final.md contains findings, fix them, add or update appropriate tests/proofs, and rerun relevant checks.
If no fix is possible, write .workflow-artifacts/my-workflow/BLOCKED_NO_COMMIT.md with exact evidence.
If it says NO_ISSUES_FOUND, record Claude review signoff.`,
      verification: { type: 'exit_code' },
    })
    .step('codex-review', {
      agent: 'codex-reviewer',
      dependsOn: ['claude-fix-final'],
      task: `Second-pass fresh-eyes review of the post-Claude-fix state. Read the actual files, diff, repo rules, and available evidence.
Write findings to .workflow-artifacts/my-workflow/codex-review.md.
If there are no actionable issues, write NO_ISSUES_FOUND.`,
      verification: { type: 'exit_code' },
    })
    .step('codex-fix', {
      agent: 'codex-fixer',
      dependsOn: ['codex-review'],
      task: `Read .workflow-artifacts/my-workflow/codex-review.md.
Fix every valid issue, add or update appropriate tests/proofs for the fix, rerun relevant checks, and update .workflow-artifacts/my-workflow/codex-fix.md.
If the review says NO_ISSUES_FOUND, record that no fix was needed.`,
      verification: { type: 'exit_code' },
    })
    .step('codex-review-final', {
      agent: 'codex-reviewer',
      dependsOn: ['codex-fix'],
      task: `Fresh-eyes review the post-Codex-fix state from scratch. Do not rely on the prior review or fix summary.
Write .workflow-artifacts/my-workflow/codex-review-final.md with either actionable findings or NO_ISSUES_FOUND.`,
      verification: { type: 'exit_code' },
    })
    .step('codex-fix-final', {
      agent: 'codex-fixer',
      dependsOn: ['codex-review-final'],
      task: `If .workflow-artifacts/my-workflow/codex-review-final.md contains findings, fix them, add or update appropriate tests/proofs, and rerun relevant checks.
If no fix is possible, write .workflow-artifacts/my-workflow/BLOCKED_NO_COMMIT.md with exact evidence.
If it says NO_ISSUES_FOUND, record final review signoff.`,
      verification: { type: 'exit_code' },
    })
    .step('acceptance-after-review', {
      type: 'deterministic',
      dependsOn: ['codex-fix-final'],
      command: 'test ! -f .workflow-artifacts/my-workflow/BLOCKED_NO_COMMIT.md && echo ACCEPTANCE_OK',
      captureOutput: true,
      failOnError: true,
    })

    .onError('retry', { maxRetries: 2, retryDelayMs: 10_000 })
    .run({ cwd: process.cwd() });

  console.log('Result:', result.status);
}

runWorkflow().catch((error) => {
  console.error(error);
  process.exit(1);
});

Quick Reference (Conversation shape)

Use this for any non-trivial work — peer review, multi-file edits, cross-agent feedback, dynamic re-planning. Lead and workers spawn in parallel on a shared channel and self-organize via messages. The relay primitive does the coordinating; verification gates downstream of the lead close the workflow.

import { workflow } from '@relayflows/core';
import { ClaudeModels, CodexModels } from '@agent-relay/config';

async function runWorkflow() {
  const result = await workflow('my-workflow')
    .description('Multi-file change with peer review')
    .pattern('dag')
    .channel('wf-my-feature') // dedicated channel — agents share it
    .maxConcurrency(4)
    .timeout(3_600_000)
    .repairable()

    // Interactive agents — no preset, they live on the channel
    .agent('lead', {
      cli: 'claude',
      model: ClaudeModels.OPUS,
      role: 'Architect + reviewer. Plans, assigns, reviews, posts feedback.',
      retries: 1,
    })
    .agent('impl-a', {
      cli: 'codex',
      model: CodexModels.GPT_5_4,
      role: 'Implementer. Listens on channel for assignments and feedback.',
      retries: 2,
    })
    .agent('impl-b', {
      cli: 'codex',
      model: CodexModels.GPT_5_4,
      role: 'Implementer. Listens on channel for assignments and feedback.',
      retries: 2,
    })
    .agent('claude-reviewer', {
      cli: 'claude',
      model: ClaudeModels.OPUS,
      preset: 'reviewer',
      role: 'First-pass fresh-eyes reviewer. Reads the final diff and artifacts from scratch.',
      retries: 1,
    })
    .agent('claude-fixer', {
      cli: 'claude',
      model: ClaudeModels.SONNET,
      role: 'First-pass review-finding fixer. Repairs valid findings, adds tests/proofs, and reruns checks.',
      retries: 2,
    })
    .agent('codex-reviewer', {
      cli: 'codex',
      model: CodexModels.GPT_5_4,
      preset: 'reviewer',
      role: 'Second-pass fresh-eyes reviewer. Reviews the post-Claude-fix state from scratch.',
      retries: 1,
    })
    .agent('codex-fixer', {
      cli: 'codex',
      model: CodexModels.GPT_5_4,
      role: 'Review-finding fixer. Repairs valid findings, adds tests/proofs, and reruns checks.',
      retries: 2,
    })

    // Deterministic context — pre-reads files once, posts to the channel for everyone
    .step('preflight', {
      type: 'deterministic',
      command: 'git rev-parse --show-toplevel >/dev/null && echo PREFLIGHT_OK',
      captureOutput: true,
      failOnError: true,
    })
    .step('context', {
      type: 'deterministic',
      dependsOn: ['preflight'],
      command: 'git ls-files src/',
      captureOutput: true,
    })

    // Lead and workers all depend on `context` — they start CONCURRENTLY.
    // They coordinate over #wf-my-feature, not via {{steps.X.output}}.
    .step('lead-coordinate', {
      agent: 'lead',
      dependsOn: ['context'],
      task: `You are the lead on #wf-my-feature. Workers: impl-a, impl-b.
Post the plan. Assign files. Review their PRs/diffs. Post feedback in-channel.
Workers iterate based on your feedback. Exit when both files pass review.`,
    })
    .step('impl-a-work', {
      agent: 'impl-a',
      dependsOn: ['context'], // SAME dep as lead → starts in parallel, no deadlock
      task: `You are impl-a on #wf-my-feature. Wait for the lead's plan.
Implement your assigned file. Post a completion message. Address feedback.`,
    })
    .step('impl-b-work', {
      agent: 'impl-b',
      dependsOn: ['context'], // SAME dep as lead
      task: `You are impl-b on #wf-my-feature. Wait for the lead's plan.
Implement your assigned file. Post a completion message. Address feedback.`,
    })

    // Downstream gates on the lead — lead exits when satisfied.
    // Capture failures, then hand them to an agent for repair.
    .step('verify', {
      type: 'deterministic',
      dependsOn: ['lead-coordinate'],
      command: 'npm run typecheck && npm test 2>&1',
      captureOutput: true,
      failOnError: false,
    })
    .step('repair-verify', {
      agent: 'lead',
      dependsOn: ['verify'],
      task: `If verification passed, summarize evidence.
If it failed, use this output to assign and fix issues, then rerun the command until green:
{{steps.verify.output}}`,
      verification: { type: 'exit_code' },
    })
    .step('verify-final', {
      type: 'deterministic',
      dependsOn: ['repair-verify'],
      command: 'npm run typecheck && npm test 2>&1',
      captureOutput: true,
      failOnError: false,
    })
    .step('claude-review', {
      agent: 'claude-reviewer',
      dependsOn: ['verify-final'],
      task: `First-pass fresh-eyes review of the post-implementation state.
Read the actual changed files, git diff, repo instructions, task spec, and verification output:
{{steps.verify-final.output}}

Write .workflow-artifacts/my-feature/claude-review.md with:
- actionable findings, each with file paths and required fix
- or NO_ISSUES_FOUND if there are no remaining issues`,
      verification: { type: 'exit_code' },
    })
    .step('claude-fix', {
      agent: 'claude-fixer',
      dependsOn: ['claude-review'],
      task: `Read .workflow-artifacts/my-feature/claude-review.md.
If there are findings, fix every valid one and add or update appropriate tests/proofs. After each fix, rerun the relevant check and review the changed files again.
Keep iterating locally until this round has no remaining valid issues.
Write .workflow-artifacts/my-feature/claude-fix.md with fixes and commands run.
If the review says NO_ISSUES_FOUND, write that no fix was needed.`,
      verification: { type: 'exit_code' },
    })
    .step('claude-review-final', {
      agent: 'claude-reviewer',
      dependsOn: ['claude-fix'],
      task: `Perform a fresh post-fix review from scratch. Do not rely on previous review text or the fixer's summary.
Read files, diff, repo rules, task spec, and evidence. Write .workflow-artifacts/my-feature/claude-review-final.md.
Use NO_ISSUES_FOUND only if there are no actionable issues left.`,
      verification: { type: 'exit_code' },
    })
    .step('claude-fix-final', {
      agent: 'claude-fixer',
      dependsOn: ['claude-review-final'],
      task: `If the final Claude review found issues, fix them, add or update appropriate tests/proofs, and rerun the relevant checks until green.
If no fix is possible, write .workflow-artifacts/my-feature/BLOCKED_NO_COMMIT.md with exact evidence and do not commit.
If the final review says NO_ISSUES_FOUND, record signoff in .workflow-artifacts/my-feature/claude-signoff.md.`,
      verification: { type: 'exit_code' },
    })
    .step('verify-after-claude-review', {
      type: 'deterministic',
      dependsOn: ['claude-fix-final'],
      command:
        'test ! -f .workflow-artifacts/my-feature/BLOCKED_NO_COMMIT.md && npm run typecheck && npm test 2>&1',
      captureOutput: true,
      failOnError: false,
    })
    .step('codex-review', {
      agent: 'codex-reviewer',
      dependsOn: ['verify-after-claude-review'],
      task: `Second-pass fresh-eyes review of the post-Claude-fix state.
Read the actual changed files, git diff, repo instructions, task spec, and verification output:
{{steps.verify-after-claude-review.output}}

Write .workflow-artifacts/my-feature/codex-review.md with:
- actionable findings, each with file paths and required fix
- or NO_ISSUES_FOUND if there are no remaining issues`,
      verification: { type: 'exit_code' },
    })
    .step('codex-fix', {
      agent: 'codex-fixer',
      dependsOn: ['codex-review'],
      task: `Read .workflow-artifacts/my-feature/codex-review.md.
If there are findings, fix every valid one and add or update appropriate tests/proofs. After each fix, rerun the relevant check and review the changed files again.
Keep iterating locally until this round has no remaining valid issues.
Write .workflow-artifacts/my-feature/codex-fix.md with fixes and commands run.
If the review says NO_ISSUES_FOUND, write that no fix was needed.`,
      verification: { type: 'exit_code' },
    })
    .step('codex-review-final', {
      agent: 'codex-reviewer',
      dependsOn: ['codex-fix'],
      task: `Perform a fresh post-Codex-fix review from scratch. Do not rely on previous review text or the fixer's summary.
Read files, diff, repo rules, task spec, and evidence. Write .workflow-artifacts/my-feature/codex-review-final.md.
Use NO_ISSUES_FOUND only if there are no actionable issues left.`,
      verification: { type: 'exit_code' },
    })
    .step('codex-fix-final', {
      agent: 'codex-fixer',
      dependsOn: ['codex-review-final'],
      task: `If the final Codex review found issues, fix them, add or update appropriate tests/proofs, and rerun the relevant checks until green.
If no fix is possible, write .workflow-artifacts/my-feature/BLOCKED_NO_COMMIT.md with exact evidence and do not commit.
If the final review says NO_ISSUES_FOUND, record signoff in .workflow-artifacts/my-feature/codex-signoff.md.`,
      verification: { type: 'exit_code' },
    })
    .step('verify-after-review', {
      type: 'deterministic',
      dependsOn: ['codex-fix-final'],
      command:
        'test ! -f .workflow-artifacts/my-feature/BLOCKED_NO_COMMIT.md && npm run typecheck && npm test 2>&1',
      captureOutput: true,
      failOnError: true,
    })

    .onError('retry', { maxRetries: 2, retryDelayMs: 10_000 })
    .run({ cwd: process.cwd() });

  console.log('Result:', result.status);
}

runWorkflow().catch((error) => {
  console.error(error);
  process.exit(1);
});

What this exercises that pipeline-shape does not:

• Ambient awareness — workers see each other's completion messages and start dependent work without the lead relaying.
• Lead-as-reviewer — the lead reads actual files between rounds and posts diff-aware feedback in chat. One agent does coordination + review; no separate reviewer step.
• Iterative correction — when the lead pings _"impl-a, the type on line 42 is wrong"_, impl-a fixes and re-posts. No new step, no re-spawn, no {{output}} chaining.

Critical workflow rules for this shape:

1. Lead and workers MUST share the same dependsOn (e.g., both depend on context). If a worker depends on the lead, you have a deadlock — the lead is waiting for worker output, the worker is waiting for the lead step to "complete."
2. Drop preset: 'worker' on the implementer agents — interactive mode is what lets them receive channel messages via PTY injection.
3. Downstream gates depend on the lead step, not the workers. The lead exits when it's satisfied; that's the workflow's signal that implementation is ready for repairable deterministic checks.
4. Use a dedicated .channel('wf-...') so the team is isolated from other workflows and the global general channel.

See Common Patterns → Interactive Team for production notes from real runs and decision criteria for picking this shape over one-shot DAG.

Default For Serious Implementation: Shadowed Squad Review Loop

When a workflow is expected to produce production-quality code, generated workflows, runtime behavior, or shared execution contracts, use a structured squad-review-loop unless the task is clearly small enough for the lighter shape.

The default unit is a 2-3 agent squad:

• implementer: owns a tight file/subsystem scope and writes the change
• shadow reviewer: follows the implementer in real time, checks drift against the spec, and leaves feedback early
• optional validation owner: owns tests, dry-run proof, or fixture coverage when that is a separate deliverable

Encode the loop explicitly:

1. Deterministically read the spec, AGENTS.md / CLAUDE.md, workflow standards, recent local docs, and declared file targets.
2. Lead splits work into bounded squads with non-overlapping ownership.
3. Squads run in parallel. The shadow reads actual files and channel updates, then posts feedback while the implementer is still active.
4. Each implementer writes a self-reflection artifact before external review. It must answer: what changed, what spec items are satisfied, what tests/proofs ran, what risks remain, and how the work follows repo rules.
5. A fresh self-review agent reads the post-implementation files, recent local conventions, AGENTS.md / CLAUDE.md, and related rules. It should not rely on the implementer's summary.
6. The implementer gets that feedback and performs a repair pass.
7. Deterministic gates run with captured output. Red output goes to a repair owner, then the same gate reruns.
8. Run the mandatory fresh-eyes review loops in sequence: Claude reviews the actual final diff and artifacts, a fixer repairs findings and hardens them with appropriate tests/proofs, Claude reviews the post-fix state again, then Codex repeats the same cycle from scratch over the post-Claude-fix state.
9. Optional extra reviewers can be added for high-stakes work, but they do not replace the sequential Claude-then-Codex loops.
10. Final signoff only happens after post-Codex-fix review and final deterministic gates prove the spec is complete, or a blocker artifact explains why it cannot be completed.

For small doc/spec workflows, a lead + author + the mandatory Claude-then-Codex review/fix loops is enough. For serious implementation workflows, do not collapse implementer self-reflection, shadow review, independent review, final dual review, and repair into one vague "review" step.

Critical TypeScript rules:

1. Check the project's package.json for "type": "module" — if ESM, use import; if CJS, use require(). In both cases, wrap execution in an async function instead of raw top-level await.
2. agent-relay local run <file.ts> executes the file as a standalone subprocess — it does NOT inspect exports. The file MUST call .run().
3. Use .run({ cwd: process.cwd() }) — createWorkflowRenderer does not exist
4. For dry-run validation, call .run({ dryRun: true, cwd: process.cwd() }) or runWorkflow(path, { dryRun: true }) from TypeScript. Use agent-relay local run <file> for execution.

⚡ Parallelism — Design for Speed

This is the most important design consideration. Sequential workflows waste hours. Always design for maximum parallelism.

Cross-Workflow Parallelism: Wave Planning

When a project has multiple workflows, group independent ones into parallel waves:

# BAD — sequential (14 hours for 27 workflows at ~30 min each)
agent-relay local run workflows/34-sst-wiring.ts
agent-relay local run workflows/35-env-config.ts
agent-relay local run workflows/36-loading-states.ts
# ... one at a time

# GOOD — parallel waves (3-4 hours for 27 workflows)
# Wave 1: independent infra (parallel)
agent-relay local run workflows/34-sst-wiring.ts &
agent-relay local run workflows/35-env-config.ts &
agent-relay local run workflows/36-loading-states.ts &
agent-relay local run workflows/37-responsive.ts &
wait
git add -A && git commit -m "Wave 1"

# Wave 2: testing (parallel — independent test suites)
agent-relay local run workflows/40-unit-tests.ts &
agent-relay local run workflows/41-integration-tests.ts &
agent-relay local run workflows/42-e2e-tests.ts &
wait
git add -A && git commit -m "Wave 2"

Wave Planning Heuristics

Two workflows can run in parallel if they don't have write-write or write-read file conflicts:

Declare File Scope for Planning

Help wave planners (human or automated) understand what each workflow touches:

workflow('48-comparison-mode')
  .packages(['web', 'core']) // monorepo packages touched
  .isolatedFrom(['49-feedback-system']) // explicitly safe to parallelize
  .requiresBefore(['46-admin-dashboard']); // explicit ordering constraint

Within-Workflow Parallelism

Use shared dependsOn to fan out independent sub-tasks:

// BAD — unnecessary sequential chain
.step('fix-component-a', { agent: 'worker', dependsOn: ['review'] })
.step('fix-component-b', { agent: 'worker', dependsOn: ['fix-component-a'] })  // why wait?

// GOOD — parallel fan-out, merge at the end
.step('fix-component-a', { agent: 'impl-1', dependsOn: ['review'] })
.step('fix-component-b', { agent: 'impl-2', dependsOn: ['review'] })  // same dep = parallel
.step('verify-all', { agent: 'reviewer', dependsOn: ['fix-component-a', 'fix-component-b'] })

Impact

Real-world example (Relayed — 60 workflows):

• Sequential: ~30 min × 60 = 30 hours
• Parallel waves (4-6 per wave): ~12 waves × 35 min = ~7 hours (4x faster)
• Aggressive parallelism (8-way): ~4 hours (7.5x faster)

Failure Prevention

These workflow files are easy to break in ways that only appear mid-run. Follow these rules when authoring or editing workflow .ts files.

1. Do not use raw top-level await

Executor-driven workflow files may be run through a tsx/esbuild path that behaves like CJS. Raw top-level await can fail with:

• Top-level await is currently not supported with the "cjs" output format

Always wrap execution like this:

async function runWorkflow() {
  const result = await workflow('my-workflow')
    // ...
    .run({ cwd: process.cwd() });

  console.log('Workflow status:', result.status);
}

runWorkflow().catch((error) => {
  console.error(error);
  process.exit(1);
});

Do not end workflow files with bare top-level await workflow(...).run(...).

1b. Make commit and PR boundaries explicit

Workflows do not get a PR for free just because they pass validation. If the intended deliverable is a branch, commit, push, or GitHub PR, the workflow itself must own that boundary explicitly and document the expected file scope.

Use this pattern only when the workflow is supposed to own repository delivery:

1. Preflight the git state and fail on unexpected staged changes.
2. Create or verify the intended branch.
3. Run implementation, repairable validation, the mandatory sequential Claude-then-Codex review/fix loops, and final acceptance gates.
4. Stage only the declared target files and review/signoff artifacts.
5. Commit with a deterministic message.
6. Push the branch.
7. Open the PR from a deterministic step, either with git/gh locally or a script that uses GitHubClient from @relayflows/github-primitive.
8. Verify the PR URL/state deterministically and write it into the final signoff artifact.

Do not hide commit/PR work in agent prose. Model it as deterministic steps whenever possible. For local workflows, use git and gh after a preflight that proves gh auth status works. For adapter-based GitHub operations, use a deterministic script that imports GitHubClient from @relayflows/github-primitive. The downstream acceptance gate must still verify the PR exists before signoff, and any PR creation failure should route to a repair step before the workflow stops.

If commit or PR creation is intentionally outside the workflow, say that directly in the workflow description and signoff so the operator knows to do it after completion.

2. Avoid raw fenced code blocks inside workflow task template literals

Raw triple-backtick code fences inside large inline task: \...\`template strings are fragile and can break outer TypeScript parsing, especially when they contain language tags likeswiftordiff`.

Preferred options, in order:

1. Avoid inline fenced examples entirely
2. Move larger examples to referenced files
3. Use plain indented examples instead of fenced blocks
4. If fenced blocks must exist inside generated inner code, escape them consistently and syntax-check the outer workflow file afterward

2b. Standard preflight template for resumable workflows

Every non-trivial workflow should start with a deterministic preflight step that validates the environment before any agent runs. A workflow that fails mid-DAG and gets re-run (or resumed via --start-from) will re-execute preflight, so preflight must tolerate the partial state left behind by the previous run — specifically, dirty files that the workflow itself is expected to edit.

The battle-tested template:

.step('preflight', {
  type: 'deterministic',
  command: [
    'set -e',
    'BRANCH=$(git rev-parse --abbrev-ref HEAD)',
    'echo "branch: $BRANCH"',
    'if [ "$BRANCH" != "fix/your-branch-name" ]; then echo "ERROR: wrong branch"; exit 1; fi',
    // Files the workflow is allowed to find dirty on entry:
    //   - package-lock.json: npm install is idempotent and often touches it
    //   - every file the workflow's edit steps will rewrite: a prior partial
    //     run may have left them dirty, and the edit step will rewrite
    //     them cleanly before commit
    // Everything else is unexpected drift and must fail preflight.
    'ALLOWED_DIRTY="package-lock.json|path/to/file1\\\\.ts|path/to/file2\\\\.ts"',
    'DIRTY=$(git diff --name-only | grep -vE "^(${ALLOWED_DIRTY})$" || true)',
    'if [ -n "$DIRTY" ]; then echo "ERROR: unexpected tracked drift:"; echo "$DIRTY"; exit 1; fi',
    'if ! git diff --cached --quiet; then echo "ERROR: staging area is dirty"; git diff --cached --stat; exit 1; fi',
    'gh auth status >/dev/null 2>&1 || (echo "ERROR: gh CLI not authenticated"; exit 1)',
    'echo PREFLIGHT_OK',
  ].join(' && '),
  captureOutput: true,
  failOnError: true,
}),

Rules baked into this template:

• Always include `package-lock.json` in ALLOWED_DIRTY. Both npm install and npm ci can touch it idempotently.
• Include every file the workflow's edit steps will rewrite. The commit step uses explicit git add <path> (never git add -A), so allowing these files to be dirty on entry is safe — unrelated drift in other files still fails preflight.
• Escape dots in regex paths: setup\.ts not setup.ts. In a JS template literal this means four backslashes: "setup\\\\.ts".
• Use `grep -vE "^(...)$"` for full-line match. Substring matches bleed across unrelated files (e.g., setup.ts would also match packages/core/src/bootstrap/setup.ts).
• Append `|| true` to the grep. Without it, an empty result triggers set -e and the whole preflight fails before the if can even run.
• Check the staging area separately. A dirty index is different from a dirty working tree and both must be clean (modulo allow-list).
• Check `gh auth status` early if downstream GitHub operations will use the local transport. Failing on auth at the end of a long DAG is painful.

Never use `git diff --quiet` alone as your "clean tree" check. It fails on any dirty file, including the ones the workflow is expected to rewrite, which causes false failures on every resume / re-run.

2c. Picking the right .join() for multi-line shell commands

When a command: field is a JS array that gets joined into a shell command string, the join delimiter determines what kinds of content the array can contain.

`.join(' && ')` — use when every element is a self-contained shell statement. Each element becomes independent and the next one runs only if the previous succeeded. Works for linear scripts with set -e.

command: [
  'set -e',
  'HITS=$(grep -c diag src/cli/commands/setup.ts || true)',
  'if [ "$HITS" -lt 6 ]; then echo "FAIL"; exit 1; fi',
  'echo OK',
].join(' && '),

`.join('\n')` — use when array elements must be part of a larger compound statement that spans multiple physical lines:

• heredocs (cat <<EOF ... EOF)
• multi-line if / while / for bodies
• shell functions defined inline

&& is a command separator. It cannot appear between a heredoc's opening line and its body, between a for and its body, or inside an if's consequent block. Joining such content with && produces a shell syntax error.

Never mix heredocs with `&&` joining. The most common failure mode:

// ❌ BROKEN — heredoc body gets && inserted between each line
command: [
  'set -e',
  'cat > /tmp/f <<EOF',
  'line 1',
  'line 2',
  'EOF',
  'next-command',
].join(' && '),

Results in set -e && cat > /tmp/f <<EOF && line 1 && line 2 && EOF && next-command — a shell syntax error because && cannot appear inside a heredoc body. Use .join('\n') or (better) sidestep the heredoc entirely.

The `printf` + `mktemp` alternative — use this for commit messages, raw-CLI fallback PR bodies, and any other multi-line file content. It avoids heredocs altogether and composes with .join(' && '):

command: [
  'set -e',
  'BODY=$(mktemp)',
  // Each line of the file is a separate printf argument. No heredoc,
  // no shell metacharacter hazards, no command-substitution nesting.
  'printf "%s\\n" "## Summary" "" "body line 1" "body line 2" > "$BODY"',
  'gh pr create --title "..." --body-file "$BODY"',
  'rm -f "$BODY"',
].join(' && '),

This pattern is specifically recommended over git commit -m "$(cat <<'EOF' ... EOF)" and raw gh pr create --body "$(cat <<'BODY' ... BODY)". Nesting a heredoc inside $(...) forces the shell to match a closing paren across many lines of unparsed body text, and any stray parenthesis in the body text can silently break the match. --body-file + mktemp + printf is immune to that entire class of bug when a workflow uses raw gh commands.

2d. Template-literal escape sequences are processed once before the string is rendered

If your file generates code as a giant template literal (the pattern used by packages/core/src/bootstrap/script-generator.ts in cloud), every backslash in that template gets processed by JavaScript before the string is returned. This silently breaks regexes and escape sequences that are meant to appear in the _generated_ output.

Specifically:

• \s is not a recognized string escape → the backslash is stripped → \s renders as a literal s
• \b _is_ a recognized string escape (backspace, U+0008) → \b renders as a backspace character in the output
• \n, \t, \r, \\, \0, \uXXXX, \xXX all get resolved at template time

The footgun: the outer TypeScript compiles cleanly, the rendered code parses and runs, and the regex/escape just never matches what the author intended. See AgentWorkforce/cloud#113 for the exact incident (hasConfigExport = /^export\s+.../m silently became /^exports+.../m in the generated bootstrap, making every TS workflow fall through to the standalone-script fallback).

Guidelines:

1. If you want a regex pattern that survives the template-literal pass unchanged, double every backslash in the source: \\s, \\b, \\n (the \\ renders to \ in the output, producing a correct regex at runtime).
2. If you want to write a long string-literal newline into the output, '\\n' in the template renders to '\n' in the output, which the runtime JS interprets as a newline. Using a literal '\n' would render an actual newline into the JS source — visually messy and sometimes surprising.
3. If you add anything non-trivial to a generator file that returns a big template literal, add a unit test that calls the generator with canonical inputs and asserts something about the rendered output — either exact string matches or, for regexes, eval/construct the regex and test it against known samples. See tests/orchestrator/script-generator.test.ts in cloud for prior art.

Task-prompt workaround: for agent-relay workflow _task prompts_ (where the contents go into a template literal but the inner content is plain text for an LLM), it's often cleaner to build the string as an array and .join('\n') at the boundary. That sidesteps the "does this backslash survive?" question entirely — no backslashes in the source, no processing to reason about. Several workflows in cloud/workflows/ use this pattern (see the sage migration PRs).

3. Keep final verification boring and deterministic

Final verification should validate real outputs with simple, portable shell commands. If checking for multiple symbols, use extended regex explicitly:

grep -Eq "foo|bar|baz" file.ts

Do not rely on basic grep alternation like:

grep -c "foo\|bar\|baz" file.ts

That can silently misbehave and create fake failures even when the generated code is correct.

4. Separate durable outputs from execution exhaust

Commit:

• generated product code
• migrations
• tests
• docs
• workflow-definition fixes

Do not commit by default:

• .logs/
• transient executor output
• retry artifacts
• temporary step-output files

5. Prefer Codex for implementation-heavy roles and dual review loops

Default team split for workflow-authored agent roles:

• lead / implementer / writer / fixer → codex
• first mandatory fresh-eyes review loop → claude
• second mandatory fresh-eyes review loop → codex

Use Claude as the primary implementer only when there is a specific reason. Use only one reviewer CLI only when the target environment cannot run the other, and record that limitation in the workflow artifact.

6. Be explicit about shell requirements

If executor scripts use Bash-only features such as associative arrays, require modern Bash explicitly. On macOS, prefer a known-good Bash path when needed, for example:

/opt/homebrew/bin/bash workflows/your-workflow/execute.sh --wave 2

7. Make resume semantics explicit

Document clearly whether the executor supports:

• full-run continuation
• --wave
• --workflow
• --resume

Do not assume users will infer the behavior. In particular, --wave N should be understood as "run only this wave" unless the executor explicitly chains onward.

7a. --resume vs --start-from when fixing a buggy step

When a workflow fails at step X and you want to re-run it after editing the workflow file, the flag choice matters:

Rule: if you edited the workflow file to fix the failing step, use --start-from <failing-step> --previous-run-id <id>, not --resume <id>. --resume pulls the entire workflow config from the run's DB record and replays it — your edits to the workflow file are ignored, and the step re-runs with its original (broken) definition.

This is counterintuitive because "resume" sounds like "pick up where you left off with whatever I just changed." It does not. It picks up where you left off with the stored config from when the run first started.

When to use each:

• Transient failure (network hiccup, rate limit, flaky agent), no code edits: --resume <id> is fine, fast, and correct.
• You edited the workflow file (any step definition, any prompt, any verify gate): always --start-from <failing-step> --previous-run-id <id>. Everything upstream of the failing step loads from cache, the fresh file supplies the fixed definition, and downstream steps run as normal.

If the runner complains that --start-from can't find cached outputs for the previous run id, fall back to a clean from-scratch run. The workflow's preflight should be forgiving enough (see §2b "Standard preflight template") that a from-scratch re-run succeeds even when a prior partial run left files dirty.

8. Syntax-check workflow files after editing

After editing workflow .ts files, run a lightweight syntax check before launching a large batch run. This is especially important if the workflow contains:

• large inline task template literals
• embedded code examples
• escaped backticks
• wrapper changes around workflow execution

9. Factor repo-specific setup into a shared helper

If multiple workflows in the same repo need the same boilerplate before any agent touches code (branch checkout, npm install, workspace-package prebuild, language toolchain init, etc.), do not copy-paste those steps into every workflow. Put them in workflows/lib/<repo>-setup.ts and import from there.

Why it matters: without a shared helper, the first workflow that needs a new prerequisite step (e.g. npm run build:platform because a workspace package's package.json points types at dist/) adds it locally, and every other workflow silently misses it. In a fresh cloud sandbox that means agents hit Cannot find module '@cloud/platform' during typecheck and paper over it with ad-hoc external-modules.d.ts shims or as GetObjectCommandOutput casts scattered across unrelated files. Those workarounds sync back down with the patch and pollute the PR.

Pattern:

// workflows/lib/cloud-repo-setup.ts
export interface CloudRepoSetupOptions {
  branch: string;
  committerName?: string;
  extraSetupCommands?: string[];
  skipWorkspaceBuild?: boolean;
}

export function applyCloudRepoSetup<T>(wf: T, opts: CloudRepoSetupOptions): T {
  // adds two steps: setup-branch, install-deps
  // install-deps runs: npm install + workspace prebuilds (build:platform, build:core, etc.)
  // ...
}

Consumer workflows break the builder chain once and call through:

const baseWf = workflow(NAME)
  .description(...)
  .pattern('dag')
  .agent(...)
  .agent(...);

const wf = applyCloudRepoSetup(baseWf, {
  branch: BRANCH,
  committerName: 'My Workflow Bot',
});

await wf
  .step('read-spec', { dependsOn: ['install-deps'], ... })
  ...
  .run(...);

Rules:

• The helper lives in the consumer repo, not in the SDK. Different customer repos have different languages, package managers, and build graphs — @agent-relay/sdk should stay agnostic.
• Pre-build any workspace package whose package.json main/types point at a generated dist/. Fresh sandboxes don't have that dist/ yet, and agents will invent workarounds rather than run the build. See the @cloud/platform case above.
• Every install step includes --legacy-peer-deps --no-audit --no-fund 2>&1 | tail -10 (or equivalent noise-trimming) because full install output blows past captureOutput size limits.
• Document the helper in the repo's CLAUDE.md / AGENTS.md so new workflow authors (and agents writing workflows) discover it.

End-to-End Bug Fix Workflows

For bug-fix or reliability workflows, do not stop at unit or integration tests. The workflow should explicitly prove that the original user-visible problem is fixed.

Required phases for fix workflows

1. Capture the original failure

• Reproduce the bug first in a deterministic or evidence-capturing step
• Save exact commands, logs, status codes, or screenshots/artifacts

1. State the acceptance contract

• Define the exact end-to-end success criteria before implementation
• Include the real entrypoint a user would run

1. Implement the fix
2. Rebuild / reinstall from scratch

• Do not trust dirty local state
• Prefer a clean environment when install/bootstrap behavior is involved

1. Run targeted regression checks

• Unit/integration tests are helpful but not sufficient by themselves

1. Run a full end-to-end validation

• Use the real CLI / API / install path
• Prefer a clean environment (Docker, sandbox, cloud workspace, Daytona, etc.) for install/runtime issues

1. Compare before vs after evidence

• Show that the original failure no longer occurs

1. Record residual risks

• Call out what was not covered

1. Ship the result as a PR

• Open the pull request from the workflow itself with deterministic git/GitHub steps
• See Shipping the Result - Open a PR below
• A workflow that fixes a bug and stops short of the PR has only done half the loop

Clean-environment validation guidance

When the bug involves install, bootstrap, PATH/shims, auth, brokers, background services, OS-specific packaging, or first-run UX, add a second workflow (or second phase) that validates the fix in a fresh environment.

Preferred order of proving environments:

1. disposable sandbox / cloud workspace
2. Docker / containerized environment
3. fresh local shell with isolated paths

Meta-workflow guidance

If the right proving environment is unclear, first write a meta-workflow that:

• compares candidate validation environments
• defines the acceptance contract
• chooses the best swarm pattern
• then authors the final fix/validation workflow

This is often better than jumping straight to implementation.

Shipping the Result - Open a PR

A workflow whose final artifact is "a clean working tree on a sandbox you'll throw away" has not shipped anything. End every code-changing workflow by opening a pull request, and do it from inside the workflow. Don't tell the operator to follow up with gh pr create; make the workflow's own last step the PR when the workflow owns shipping.

Current @relayflows/core does not provide createGitHubStep. Use one of these current surfaces:

The minimal local "open a PR" recipe

import { workflow } from '@relayflows/core';

const BRANCH = `agent-relay/run-${Date.now()}`;

async function runWorkflow() {
  await workflow('feature-x')
    // ... your real implementation, repair, review loops, and final acceptance ...
    .step('write-marker', {
      type: 'deterministic',
      command: `echo "fix landed at $(date -u)" >> CHANGELOG.md`,
    })
    .step('create-branch', {
      type: 'deterministic',
      dependsOn: ['write-marker'],
      command: `git switch -c ${BRANCH}`,
    })
    .step('commit-change', {
      type: 'deterministic',
      dependsOn: ['create-branch'],
      command: 'git add CHANGELOG.md && git commit -m "chore: changelog entry"',
    })
    .step('push-branch', {
      type: 'deterministic',
      dependsOn: ['commit-change'],
      command: `git push -u origin ${BRANCH}`,
    })
    .step('open-pr', {
      type: 'deterministic',
      dependsOn: ['push-branch'],
      command: `gh pr create --base main --head ${BRANCH} --title "feat: ship feature X" --body-file .workflow-artifacts/feature-x/pr-body.md`,
      verification: { type: 'pr_url', value: 'AgentWorkforce/cloud' },
    })
    .run({ cwd: process.cwd() });
}

runWorkflow().catch((error) => {
  console.error(error);
  process.exit(1);
});

For non-local GitHub operations, create a deterministic script that imports GitHubClient from @relayflows/github-primitive and calls the client methods for createBranch, createFile or updateFile, createPR, and getPR.

Authoring rules for PR-shipping workflows

1. Open the PR from the workflow, not from the operator's shell. "Tell the user to run gh pr create" is a regression to a manual step the workflow could have done. The whole point of running this in cloud is that there is no operator's shell.
2. One PR per workflow, by default. A workflow that opens five PRs from one run is almost always wrong — humans review one PR at a time. If you genuinely need multiple, prefer a tracking issue + linked PRs, or split into separate workflows.
3. Branch name encodes the run. agent-relay/run-${runId} or agent-relay/${workflow-name}-${timestamp} so reviewers can tell the PR apart from other automation, and so reruns don't clash.
4. `draft: true` while iterating. Once the workflow is stable end-to-end, flip to draft: false.
5. Body is a real PR description. Summary + Test plan, generated from the workflow's own evidence (verification step output, diff stats, test run output). If you find yourself writing a placeholder body, the workflow isn't done — capture the real evidence in an earlier step and template it in.
6. Don't duplicate cloud `paths[]` push-back. If the diff lives in a tarballed paths[] mount and cloud push-back is configured, let cloud open that PR. Use explicit GitHub steps when you need a PR against a repo or branch outside the paths[] set, or when you want to add an extra PR.
7. PR creation failures route to repair. If PR creation errors (auth, permissions, branch conflict), capture the output and give a repair owner a chance to fix auth, branch state, labels, or body generation before stopping. A "successful" workflow that silently failed to open the PR is the worst-case outcome — the human thinks the work shipped.

Where this fits in the bug-fix phases

End-to-End Bug Fix Workflows lists "Ship the result as a PR" as phase 9. Concretely that means: after phase 7 (compare before/after evidence) succeeds, the workflow's next step opens a PR with that evidence templated into the body. The PR opening is the ship — there is no further manual step.

Key Concepts

Step Output Chaining

Use {{steps.STEP_NAME.output}} in a downstream step's task to inject the prior step's terminal output.

Mental model: this is a Unix pipe, not agent communication. {{steps.A.output}} flowing into step B is A | B — A is dead by the time B reads its stdout. There is no chat, no feedback, no addressing. If your workflow's coordination story is _only_ output chaining, you're using the relay as transport, not as a coordination layer. See [Choose Your Coordination Style](#choose-your-coordination-style--conversation-vs-pipeline) before defaulting to this.

Only chain output from clean sources:

• Deterministic steps (shell commands — always clean)
• Non-interactive agents (preset: 'worker' — clean stdout)

Never chain from interactive agents (cli: 'claude' without preset) — PTY output includes spinners, ANSI codes, and TUI chrome. Instead, have the agent write to a file, then read it in a deterministic step. (Or: don't use chaining at all — let the agents coordinate over the channel.)

Verification Gates

verification: { type: 'exit_code' }                        // preferred for code-editing steps
verification: { type: 'output_contains', value: 'DONE' }   // optional accelerator
verification: { type: 'file_exists', value: 'src/out.ts' } // deterministic file check
verification: { type: 'pr_url', value: 'owner/repo' }      // step must leave behind a PR

Only these five types are valid: exit_code, output_contains, file_exists, custom, pr_url. Invalid types are silently ignored and fall through to process-exit auto-pass.

Use `pr_url` for any step whose deliverable is a published change — opening a PR, merging a branch, publishing a package. It blocks the common failure mode where a worker produces green tests and posts OWNER_DECISION: COMPLETE but never actually opened a PR. Pair it with a deterministic PR step that prints the PR URL, or with an adapter script that uses GitHubClient from @relayflows/github-primitive. Pass <owner>/<repo> to require the URL belongs to a specific repository, or leave value: '' to accept any GitHub PR URL in the step output.

Verification token gotcha: If the token appears in the task text, the runner requires it twice in output (once from task echo, once from agent). Prefer exit_code for code-editing steps to avoid this.

DAG Dependencies

Steps with dependsOn wait for all listed steps. Steps with no dependencies start immediately. Steps sharing the same dependsOn run in parallel:

.step('fix-types',  { agent: 'worker', dependsOn: ['review'], ... })
.step('fix-tests',  { agent: 'worker', dependsOn: ['review'], ... })
.step('final',      { agent: 'lead',   dependsOn: ['fix-types', 'fix-tests'], ... })

Self-Termination

Do NOT add exit instructions to task strings. The runner handles this automatically.

For bounded Codex steps that must produce one artifact or a structured answer, use a non-interactive preset (preset: 'worker', reviewer, or analyst) instead of interactive PTY. This runs through one-shot subprocess mode (codex exec), so completion is the process lifecycle plus verification. Interactive Codex is for live channel coordination; it is weaker for "write one file then exit" loops because idle detection can see an auth or prompt-delivery failure as silence.

Step Completion Model

Steps complete through a multi-signal pipeline (highest priority first):

1. Deterministic verification — exit_code, file_exists, output_contains pass → immediate completion
2. Owner decision — OWNER_DECISION: COMPLETE|INCOMPLETE_RETRY|INCOMPLETE_FAIL
3. Evidence-based — channel signals, file artifacts, clean exit code
4. Marker fast-path — STEP_COMPLETE:<step-name> (optional accelerator)
5. Process-exit fallback — agent exits 0 with no signals → completes after grace period

Key principle: No single signal is mandatory. Describe the deliverable, not what to print.

Dynamic Channel Management

Agents can dynamically subscribe, unsubscribe, mute, and unmute channels after spawn. This eliminates the need for client-side channel filtering and manual peer fanout.

#### SDK API

// Subscribe an agent to additional channels post-spawn
relay.subscribe({ agent: 'security-auditor', channels: ['review-pr-456'] });

// Unsubscribe — agent leaves the channel entirely
relay.unsubscribe({ agent: 'security-auditor', channels: ['general'] });

// Mute — agent stays subscribed (history access) but messages are NOT injected into PTY
relay.mute({ agent: 'security-auditor', channel: 'review-pr-123' });

// Unmute — resume PTY injection
relay.unmute({ agent: 'security-auditor', channel: 'review-pr-123' });

Agent-level methods are also available:

const agent = await relay.claude.spawn({ name: 'auditor', channels: ['ch-a'] });
await agent.subscribe(['ch-b']); // now subscribed to ch-a and ch-b
await agent.mute('ch-a'); // ch-a messages silenced (still in history)
await agent.unmute('ch-a'); // ch-a messages resume
await agent.unsubscribe(['ch-b']); // leaves ch-b
console.log(agent.channels); // ['ch-a']
console.log(agent.mutedChannels); // []

#### Semantics

#### Events

relay.onChannelSubscribed = (agent, channels) => {
  /* ... */
};
relay.onChannelUnsubscribed = (agent, channels) => {
  /* ... */
};
relay.onChannelMuted = (agent, channel) => {
  /* ... */
};
relay.onChannelUnmuted = (agent, channel) => {
  /* ... */
};

#### When to Use in Workflows

• Multi-PR chat sessions: Agents focused on one PR can mute other PR channels to reduce noise
• Phase transitions: Subscribe agents to new channels as work progresses between phases
• Team isolation: Workers mute the