agent-relay 2.3.13 → 2.3.15

package/packages/sdk/src/examples/workflows/README.md

@@ -0,0 +1,156 @@
+# Super-Powered Ralph Loops
+
+These workflows are the agent-relay take on the [Ralph Wiggum technique](https://github.com/mikeyobrien/ralph-orchestrator) — autonomous AI agents looping on a codebase until work is done. Where vanilla ralph runs one agent at a time, these workflows run **squads of specialists** with real quality gates, multi-model assignment, and observable coordination.
+
+## What makes these different
+
+| Vanilla Ralph                  | agent-relay Ralph                                                   |
+| ------------------------------ | ------------------------------------------------------------------- |
+| 1 agent per loop               | 2–10 agents per loop                                                |
+| Single model                   | Multi-model (opus plans, sonnet leads, codex builds, gemini audits) |
+| Agent checks its own work      | Independent reviewer with no confirmation bias                      |
+| Polling for completion signals | DAG-based wave scheduling with verification gates                   |
+| No parallelism                 | Workers implement features simultaneously                           |
+| One retry strategy             | Per-step `maxIterations` + global `errorHandling`                   |
+| No observability               | Trajectories, Relaycast channels, step output chaining              |
+
+---
+
+## `ralph-tdd.yaml` — Test-Driven Loop
+
+**Best for:** Projects with an existing test framework where correctness is the priority.
+
+```
+architect writes failing tests
+         ↓
+builder implements to pass them (maxIterations: 3)
+         ↓
+npm test / pytest / go test  ←── deterministic gate
+         ↓
+arch-review ──┐
+              ├── both must PASS (consensus: unanimous)
+sec-review  ──┘
+         ↓
+address feedback (maxIterations: 2)
+         ↓
+commit + record learnings → next story
+```
+
+**Key agents:** 3 (architect/sonnet, builder/codex, reviewer/opus)
+**Key feature:** Tests are written BEFORE implementation. Reviewer is separate from architect — no confirmation bias.
+
+```bash
+PRD_PATH=my-prd.json QUALITY_CMD="npm test" \
+  node packages/sdk/dist/workflows/cli.js ralph-tdd.yaml
+```
+
+---
+
+## `ralph-swarm.yaml` — Parallel Implementation Squad
+
+**Best for:** Large backlogs where work can be decomposed into independent tasks.
+
+```
+tech-lead decomposes PRD into 5 atomic tasks
+         ↓
+worker-1 ──┐
+worker-2   │
+worker-3   ├── all 5 implement in parallel
+worker-4   │
+worker-5 ──┘
+         ↓
+npm test / tsc / lint  ←── gate
+         ↓
+fix-failures (if gate red)
+         ↓
+review-correctness ──┐
+review-architecture  ├── all 3 must PASS
+review-security    ──┘
+         ↓
+consensus → address rework → commit → loop
+```
+
+**Key agents:** 10 (opus tech-lead, 5 codex workers, 2 claude reviewers, gemini security auditor)
+**Key feature:** 5x parallelism. Three independent reviewers. Gemini on security catches what Claude misses.
+
+```bash
+PRD_PATH=my-prd.json \
+  node packages/sdk/dist/workflows/cli.js ralph-swarm.yaml
+```
+
+---
+
+## `ralph-overnight.yaml` — 24-Hour Autonomous Session
+
+**Best for:** Long sessions where you want to drop a repo before bed and wake up to PRs.
+
+```
+product-manager reads backlog, prioritizes, assigns to squads
+         ↓
+tech-lead plans architecture, creates feature branch
+         ↓
+squad-alpha (lead + 2 codex builders) ──┐
+                                         ├── parallel on different files
+squad-beta  (lead + 2 codex builders) ──┘
+         ↓
+full CI gate (lint + tsc + npm test)
+         ↓
+qa-engineer writes integration tests ──┐
+                                        ├── parallel
+security-auditor (gemini) audits     ──┘
+         ↓
+tech-lead reviews → fix if needed
+         ↓
+git push → gh pr create (draft)
+         ↓
+PM writes session-log.txt morning summary
+→ mark stories complete → next batch
+```
+
+**Key agents:** 10 (PM/opus, tech-lead/sonnet, 2 squad leads/sonnet, 4 codex builders, QA/sonnet, security/gemini-pro)
+**Key feature:** True hierarchical org. PM → Tech Lead → Squads. Morning report written to `session-log.txt`. Runs safely overnight with `errorHandling: continue` (one story failure doesn't kill the session).
+
+```bash
+BACKLOG_PATH=my-backlog.json SESSION_LOG=session-log.txt \
+  node packages/sdk/dist/workflows/cli.js ralph-overnight.yaml
+```
+
+---
+
+## Running any of these
+
+```bash
+# Preview the execution plan without spending tokens
+DRY_RUN=1 node packages/sdk/dist/workflows/cli.js ralph-tdd.yaml
+
+# Run for real
+node packages/sdk/dist/workflows/cli.js ralph-tdd.yaml
+
+# Watch Relaycast for live agent communication
+# → open https://relaycast.dev and join the workflow's channel
+```
+
+## PRD / Backlog format
+
+These workflows expect a JSON file with your stories:
+
+```json
+{
+  "branchName": "feature/my-project",
+  "userStories": [
+    {
+      "id": "US-001",
+      "title": "User can log in with email",
+      "description": "Implement email/password authentication",
+      "acceptanceCriteria": [
+        "POST /auth/login returns 200 with JWT on valid credentials",
+        "Returns 401 on invalid credentials",
+        "Passwords are bcrypt hashed"
+      ],
+      "passes": false
+    }
+  ]
+}
+```
+
+For `ralph-overnight.yaml`, use `backlog.json` with `priority` and `effort` fields.

package/packages/sdk/src/examples/workflows/ralph-overnight.yaml

@@ -0,0 +1,421 @@
+version: '1.0'
+name: ralph-overnight
+description: >
+  Autonomous 24-hour dev session. A product manager reads the backlog, a tech
+  lead plans the architecture, parallel squads implement features, a CI runner
+  gates every batch, and the cycle repeats until the backlog is empty or
+  morning arrives. Drop it on a repo before bed, wake up to merged features.
+
+  What this showcases vs vanilla ralph:
+    - Hierarchical orchestration: PM → Tech Lead → Implementation Squads
+    - Mixed CLI team: claude thinks, codex builds, gemini audits
+    - Integrated CI: actual git branch, PR creation, CI status polling
+    - Progressive learning: each loop iteration reads prior session notes
+    - Human checkpoint: if stuck 3x on same story, posts a question to Slack
+      (or a file) and pauses for human input before continuing
+    - Session summary written at the end for morning review
+
+swarm:
+  pattern: hierarchical
+  maxConcurrency: 5
+  timeoutMs: 86400000 # 24 hours — true overnight session
+  channel: ralph-overnight
+  idleNudge:
+    nudgeAfterMs: 600000 # 10 min — these agents think deeply
+    escalateAfterMs: 600000
+    maxNudges: 3
+
+agents:
+  # ── Strategic layer ────────────────────────────────────────────────────────
+
+  - name: product-manager
+    cli: claude
+    channels: [ralph-overnight, strategy]
+    role: >
+      Reads the backlog, prioritizes stories by impact/effort, plans the
+      session's scope, and writes the morning summary report.
+    constraints:
+      model: opus
+
+  - name: tech-lead
+    cli: claude
+    channels: [ralph-overnight, engineering]
+    role: >
+      Translates PM priorities into technical tasks, resolves architectural
+      decisions, monitors squad progress, unblocks stuck agents, and gates PRs.
+    constraints:
+      model: sonnet
+
+  # ── Implementation squads ──────────────────────────────────────────────────
+
+  - name: squad-alpha-lead
+    cli: claude
+    channels: [squad-alpha, ralph-overnight]
+    role: 'Leads squad alpha. Coordinates two codex builders on #squad-alpha.'
+    constraints:
+      model: sonnet
+
+  - name: squad-alpha-builder-1
+    cli: codex
+    role: 'Implements BUILDER_1_SPEC from squad-alpha output via {{steps.squad-alpha.output}}.'
+    interactive: false
+    constraints:
+      model: gpt-5.3-codex
+
+  - name: squad-alpha-builder-2
+    cli: codex
+    role: 'Implements BUILDER_2_SPEC from squad-alpha output via {{steps.squad-alpha.output}}.'
+    interactive: false
+    constraints:
+      model: gpt-5.3-codex-spark
+
+  - name: squad-beta-lead
+    cli: claude
+    channels: [squad-beta, ralph-overnight]
+    role: 'Leads squad beta. Coordinates two codex builders on #squad-beta.'
+    constraints:
+      model: sonnet
+
+  - name: squad-beta-builder-1
+    cli: codex
+    role: 'Implements BUILDER_1_SPEC from squad-beta output via {{steps.squad-beta.output}}.'
+    interactive: false
+    constraints:
+      model: gpt-5.3-codex
+
+  - name: squad-beta-builder-2
+    cli: codex
+    role: 'Implements BUILDER_2_SPEC from squad-beta output via {{steps.squad-beta.output}}.'
+    interactive: false
+    constraints:
+      model: gpt-5.3-codex-spark
+
+  # ── Quality & audit ────────────────────────────────────────────────────────
+
+  - name: qa-engineer
+    cli: claude
+    channels: [ralph-overnight]
+    role: 'Writes integration tests for completed features. Runs the test suite.'
+    interactive: false
+    constraints:
+      model: sonnet
+
+  - name: security-auditor
+    cli: gemini
+    channels: [ralph-overnight]
+    role: 'Audits all changes for security issues before PR creation.'
+    interactive: false
+    constraints:
+      model: gemini-2.5-pro
+
+workflows:
+  - name: overnight-session
+    description: >
+      PM planning → tech architecture → parallel squads (alpha+beta) →
+      QA + security audit → CI gate → PR → advance backlog → loop.
+    onError: continue # don't abort entire session on one story failure
+    preflight:
+      - command: test -f "${BACKLOG_PATH:-backlog.json}"
+        description: 'Backlog file exists'
+      - command: git status --porcelain
+        failIf: non-empty
+        description: 'Clean working directory for overnight session'
+      - command: git rev-parse --abbrev-ref HEAD
+        description: 'Confirm current branch'
+        captureOutput: true
+
+    steps:
+      # ── Wave 1: PM reads backlog and plans the night's scope ─────────────
+
+      - name: pm-planning
+        type: agent
+        agent: product-manager
+        task: |
+          Read ${BACKLOG_PATH:-backlog.json} and ${SESSION_LOG:-session-log.txt}.
+
+          Select the highest-priority stories for a ~4-hour parallel session.
+          Assign to two squads working on DIFFERENT files (no conflicts):
+            Alpha squad: stories + file paths
+            Beta squad: stories + file paths
+          Include risk flags for any cross-squad dependencies.
+
+          Post SESSION_PLAN to #ralph-overnight, then output: PLANNING_COMPLETE
+        verification:
+          type: output_contains
+          value: PLANNING_COMPLETE
+        retries: 1
+
+      # ── Wave 2: Tech lead creates feature branches + architecture notes ───
+
+      - name: tech-architecture
+        type: agent
+        agent: tech-lead
+        dependsOn: [pm-planning]
+        task: |
+          PM plan: {{steps.pm-planning.output}}
+
+          1. Create feature branch: git checkout -b overnight/$(date +%Y%m%d)
+          2. For each story assigned to squads, write implementation notes:
+             - Which files to touch and why
+             - Any shared utilities to create first
+             - Test strategy (unit / integration / e2e)
+          3. Flag any cross-squad dependencies that need sequencing
+          4. Post architecture notes to #ralph-overnight
+
+          Output: ARCHITECTURE_READY
+        verification:
+          type: output_contains
+          value: ARCHITECTURE_READY
+
+      - name: create-branch
+        type: deterministic
+        dependsOn: [tech-architecture]
+        command: >
+          git checkout -b "overnight/$(date +%Y%m%d)" 2>/dev/null ||
+          git checkout "overnight/$(date +%Y%m%d)"
+        captureOutput: true
+
+      # ── Wave 3: Squad leads plan and produce builder specs ───────────────
+      # Leads run first, produce self-contained task specs for each builder.
+      # Builders run in wave 4 after the lead completes, reading specs via
+      # {{steps.squad-alpha.output}} — no channel needed for non-interactive agents.
+
+      - name: squad-alpha
+        type: agent
+        agent: squad-alpha-lead
+        dependsOn: [create-branch]
+        task: |
+          Lead squad alpha on #squad-alpha.
+          Architecture: {{steps.tech-architecture.output}}
+          PM plan (alpha): {{steps.pm-planning.output}}
+
+          Split alpha stories into two self-contained tasks. For each, write
+          a complete implementation spec including: files to change, exact
+          changes needed, and the shell command to verify it works.
+
+          Format your output as:
+          BUILDER_1_SPEC: <complete spec for builder 1>
+          BUILDER_2_SPEC: <complete spec for builder 2>
+          Output: ALPHA_SPECS_READY
+        verification:
+          type: output_contains
+          value: ALPHA_SPECS_READY
+
+      - name: squad-beta
+        type: agent
+        agent: squad-beta-lead
+        dependsOn: [create-branch]
+        task: |
+          Lead squad beta on #squad-beta.
+          Architecture: {{steps.tech-architecture.output}}
+          PM plan (beta): {{steps.pm-planning.output}}
+
+          Split beta stories into two self-contained tasks. Write complete
+          specs for each builder: files, changes, verification command.
+
+          Format:
+          BUILDER_1_SPEC: <complete spec for builder 1>
+          BUILDER_2_SPEC: <complete spec for builder 2>
+          Output: BETA_SPECS_READY
+        verification:
+          type: output_contains
+          value: BETA_SPECS_READY
+
+      # ── Wave 4: Builders implement from lead specs (non-interactive) ──────
+      # Each builder reads its spec from {{steps.squad-X.output}} — no channel
+      # needed. They run in parallel across both squads (same wave).
+
+      - name: squad-alpha-b1
+        type: agent
+        agent: squad-alpha-builder-1
+        dependsOn: [squad-alpha]
+        task: |
+          Implement BUILDER_1_SPEC from the squad alpha lead output:
+          {{steps.squad-alpha.output}}
+
+          Follow the spec exactly. Run the verification command to confirm.
+          Output: ALPHA_B1_DONE:<one-line summary>
+        verification:
+          type: output_contains
+          value: ALPHA_B1_DONE
+
+      - name: squad-alpha-b2
+        type: agent
+        agent: squad-alpha-builder-2
+        dependsOn: [squad-alpha]
+        task: |
+          Implement BUILDER_2_SPEC from the squad alpha lead output:
+          {{steps.squad-alpha.output}}
+
+          Follow the spec exactly. Run the verification command to confirm.
+          Output: ALPHA_B2_DONE:<one-line summary>
+        verification:
+          type: output_contains
+          value: ALPHA_B2_DONE
+
+      - name: squad-beta-b1
+        type: agent
+        agent: squad-beta-builder-1
+        dependsOn: [squad-beta]
+        task: |
+          Implement BUILDER_1_SPEC from the squad beta lead output:
+          {{steps.squad-beta.output}}
+
+          Follow the spec exactly. Run the verification command to confirm.
+          Output: BETA_B1_DONE:<one-line summary>
+        verification:
+          type: output_contains
+          value: BETA_B1_DONE
+
+      - name: squad-beta-b2
+        type: agent
+        agent: squad-beta-builder-2
+        dependsOn: [squad-beta]
+        task: |
+          Implement BUILDER_2_SPEC from the squad beta lead output:
+          {{steps.squad-beta.output}}
+
+          Follow the spec exactly. Run the verification command to confirm.
+          Output: BETA_B2_DONE:<one-line summary>
+        verification:
+          type: output_contains
+          value: BETA_B2_DONE
+
+      # ── Wave 4: Full test suite gate ─────────────────────────────────────
+
+      - name: ci-gate
+        type: deterministic
+        dependsOn: [squad-alpha-b1, squad-alpha-b2, squad-beta-b1, squad-beta-b2]
+        command: >
+          npm run lint 2>&1 | tail -5 &&
+          npx tsc --noEmit 2>&1 | tail -5 &&
+          npm test 2>&1 | tail -40 &&
+          echo "CI_GREEN" || echo "CI_RED"
+        captureOutput: true
+        failOnError: false
+
+      # ── Wave 5: QA writes integration tests + security audit (parallel) ──
+
+      - name: qa-tests
+        type: agent
+        agent: qa-engineer
+        dependsOn: [ci-gate]
+        task: |
+          CI results: {{steps.ci-gate.output}}
+          Alpha built: {{steps.squad-alpha.output}}
+          Beta built: {{steps.squad-beta.output}}
+
+          Write integration tests for the newly implemented features.
+          Focus on the happy path and the two most likely failure modes per feature.
+          Run the full test suite including your new tests.
+
+          Output: QA_DONE:<test count added> new tests, all green
+        verification:
+          type: output_contains
+          value: QA_DONE
+
+      - name: security-audit
+        type: agent
+        agent: security-auditor
+        dependsOn: [ci-gate]
+        task: |
+          Security audit of all changes in this overnight session.
+
+          Run: git diff main 2>/dev/null | head -1000
+
+          Check for:
+          - SQL/command injection vectors
+          - Authentication bypass risks
+          - Secrets or tokens in code
+          - Unvalidated user input reaching sensitive operations
+          - Insecure direct object references
+
+          Output: SECURITY:CLEAN or SECURITY:ISSUES:<list>
+        verification:
+          type: output_contains
+          value: 'SECURITY:'
+
+      # ── Wave 6: Tech lead reviews and gates the PR ────────────────────────
+
+      - name: tech-review
+        type: agent
+        agent: tech-lead
+        dependsOn: [qa-tests, security-audit]
+        task: |
+          QA: {{steps.qa-tests.output}}
+          Security: {{steps.security-audit.output}}
+          CI: {{steps.ci-gate.output}}
+          Alpha: {{steps.squad-alpha.output}}
+          Beta: {{steps.squad-beta.output}}
+
+          If security issues found: fix them first, then output TECH_REVIEW:PASS
+          If CI is red: identify root cause, apply fix, re-run test, output TECH_REVIEW:PASS
+          If all green: output TECH_REVIEW:PASS
+
+          Do NOT output TECH_REVIEW:PASS with outstanding security issues.
+        verification:
+          type: output_contains
+          value: TECH_REVIEW:PASS
+        maxIterations: 2
+
+      # ── Wave 7: Commit, push, create PR ───────────────────────────────────
+
+      - name: push-and-pr
+        type: deterministic
+        dependsOn: [tech-review]
+        command: >
+          git add -A &&
+          git commit -m "feat: overnight session $(date +%Y-%m-%d) — autonomous batch" &&
+          git push -u origin HEAD 2>&1 &&
+          gh pr create --title "Overnight session $(date +%Y-%m-%d)" \
+            --body "Autonomous dev session. See session-log.txt for details." \
+            --draft 2>&1 ||
+          echo "PR_CREATED_OR_UPDATED"
+        captureOutput: true
+        failOnError: false
+
+      # ── Wave 8: PM writes morning summary ────────────────────────────────
+
+      - name: morning-summary
+        type: agent
+        agent: product-manager
+        dependsOn: [push-and-pr]
+        task: |
+          Write the morning summary to ${SESSION_LOG:-session-log.txt}.
+
+          Results: alpha={{steps.squad-alpha.output}} | beta={{steps.squad-beta.output}}
+          CI={{steps.ci-gate.output}} | QA={{steps.qa-tests.output}}
+          Security={{steps.security-audit.output}} | PR={{steps.push-and-pr.output}}
+
+          Include: completed stories, CI/quality status, learnings, next priorities.
+          Mark completed stories done in ${BACKLOG_PATH:-backlog.json}.
+
+          Output: SESSION_COMPLETE
+        verification:
+          type: output_contains
+          value: SESSION_COMPLETE
+
+coordination:
+  barriers:
+    - name: squads-complete
+      waitFor: [squad-alpha-b1, squad-alpha-b2, squad-beta-b1, squad-beta-b2]
+      timeoutMs: 7200000
+    - name: quality-gates-complete
+      waitFor: [qa-tests, security-audit]
+      timeoutMs: 3600000
+
+state:
+  backend: memory
+  ttlMs: 86400000
+
+errorHandling:
+  strategy: continue # one story failing shouldn't abort the whole night
+  maxRetries: 1
+  retryDelayMs: 10000
+  notifyChannel: ralph-overnight
+
+trajectories:
+  enabled: true
+  reflectOnBarriers: true
+  reflectOnConverge: true
+  autoDecisions: true