Skip to content

v1.7.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 26 Apr 04:57
· 9 commits to main since this release
v1.7.0
61574ef

Changelog

1.7.0

Two new capabilities: true parallel step execution with live streaming dashboard, and autonomous workflow self-improvement.

Parallel execution

  • max_parallel is now readable from manifest["graph"] (not just the policy pack). Set "graph": {"max_parallel": 4} to run up to 4 independent steps concurrently. The ThreadPoolExecutor-based DAG scheduler was already in place — this unlocks it from workflow definitions.

Live streaming dashboard (scripts/live_dashboard.py)

  • New --live [PORT] flag on run_workflow.py starts a local HTTP server (default port 7474) in a background thread alongside the workflow run.
  • Serves an SSE stream from the current run's events.jsonl (which already captured step_started, step_completed, step_failed events).
  • Browser page shows a real-time step grid: pending → running (pulsing) → complete/failed/skipped — no refresh needed. Zero external dependencies (SSE via stdlib http.server).
  • Also standalone: python3 scripts/live_dashboard.py <workflow-dir> [--port N]

Autonomous improvement loop

  • New scripts/mutation_classifier.py — risk-scores mutation proposals as low / medium / high based on mutation type and which step fields are being changed. remove_step and script changes are high; timeout_seconds/retry_limit changes are low.
  • run_improvement_cycle() in run_workflow.py — auto-approves pending mutations that meet the configured risk threshold and prints a summary. Calls analyze_run_history() to surface unhealthy steps (>20% failure rate).
  • New --improve flag — runs the improvement cycle on demand. --improve-max-risk low|medium|high overrides the threshold.
  • Workflows opt in to post-run auto-improvement with "auto_improve": {"enabled": true, "max_risk": "low"} in workflow.json.

21 new tests; 99/99 pass.

1.6.0

Five security fixes from adversarial code review:

  • MCP policy bypass — enforce_security_policy() now fully evaluates MCP steps: validates the server name against an optional allowed_mcp_servers allowlist in the policy pack, and blocks all MCP steps when network_mode: offline.
  • Webhook security — Generated webhook_server.py now binds to 127.0.0.1 only (not 0.0.0.0), validates the request path against a configurable EXPECTED_PATH, and enforces constant-time token auth via hmac.compare_digest. install_webhook_trigger() accepts path and secret from the trigger dict and forwards them to the template.
  • Dependency field — Importer wrote step["needs"] but the scheduler reads step.get("depends_on", []). Fixed to write depends_on; existing test updated accordingly.
  • Branch/switch contracts — type: "branch" steps now include condition, on_true, and on_false derived from n8n connection output indices (output[0]=true, output[1]=false). type: "switch" steps include a cases dict keyed by outputKey, each mapping to downstream step IDs.
  • Mutation envelope — Mutation proposal files are now written and read as {"mutations": [...]} to match the envelope format that run_workflow.py expects; backward-compatible reader accepts both formats.
  • 12 new tests; 78/78 pass.

1.5.0

  • type: "browser" — Chrome MCP-backed browser automation step. Claude navigates, clicks, fills forms, reads page content, inspects network, and captures results — all driven by a natural language instruction field with full artifact template expansion. No API required.
  • type: "computer-use" — Desktop automation step via computer-use MCP. Claude controls any native app via screenshots, mouse, keyboard, and scroll. Automates workflows that have no API and no browser interface.
  • Both step types write their output as artifacts, support output_artifact naming, full retry/timeout/auto-heal, and integrate with the approval gate and audit system.
  • 5 new tests; 66/66 pass.

1.4.0

  • --import-n8n <file> — Convert any n8n workflow export JSON to a workflow.json. Maps 30+ n8n node types: executeCommand → shell, httpRequest → http, if → branch, switch → switch, merge → merge, wait → wait, code → shell, langchain/AI nodes → claude, service nodes (Slack, Gmail, GitHub, Notion, Jira, Linear, S3, Postgres…) → http with placeholder URLs. Extracts schedule and webhook triggers. Topologically sorts steps and injects needs dependencies.
  • Auto-improvement proposals — Imported workflows automatically get pending mutation proposals for service nodes (upgrade to type:mcp), inline code stubs (port to script), placeholder URLs, and auto-generated branch conditions. Reviewable with --list-mutations.
  • scripts/import_n8n.py — Standalone converter; also callable via run_workflow.py --import-n8n.
  • 7 new tests; 61/61 pass.

1.3.0

  • type: "skill" — Run any Codex or Claude Code skill as a workflow step. Loads SKILL.md, builds a combined prompt with artifact context, and calls the claude CLI. Skill name resolves by exact match or case-insensitive prefix.
  • scripts/discover_skills.py — Auto-discovers all installed skills from ~/.codex/skills/, ~/.claude/plugins/cache, ~/.claude/skills/, and ~/.claude/plugins/. Deduplicates by name (more-specific paths win). Returns name, path, source, description, and whether a SKILL.md exists.
  • --discover-skills flag — Prints a table of all discovered skills with their source and description.
  • --generate skill-awareness — Available skills are injected into the system prompt so generated workflows can include type: "skill" steps where appropriate.
  • 5 new tests; 54/54 pass.

1.2.0

  • type: "http" — Full HTTP request step: method, URL, headers, body, bearer/basic auth, response stored as JSON artifact. Fails on 4xx/5xx by default (fail_on_error: false to override). Template expansion supported in url/headers/body.
  • type: "switch" — Multi-way branch: evaluate an expression against named cases, mark non-matching steps as skipped. Equivalent to n8n's Switch node.
  • type: "loop" — Iterate a script over every item in a JSON array or newline-separated artifact. Each iteration receives LOOP_ITEM, LOOP_INDEX, LOOP_TOTAL env vars. Results written as JSON array artifact.
  • type: "wait" — Pause execution for seconds or poll a condition script until it exits 0 (until field). Configurable timeout and poll interval.
  • type: "merge" — Combine multiple artifact inputs into one: concat (arrays → flat list, objects → merged dict), zip, or first. Closes the merge-path gap vs n8n.
  • type: "workflow" — Run another workflow.json as a sub-workflow. Pass artifacts in, collect artifacts out. Enables composable workflow libraries.
  • 7 new tests; 49/49 pass. No remaining gaps vs n8n's core action nodes.

1.1.0

  • Feature A — type: "claude" step: run a Claude prompt as a workflow step with {{artifact:id}} / {{env:VAR}} template expansion, optional output_schema JSON validation, and automatic artifact capture. Calls claude CLI; falls back to anthropic SDK.
  • Feature B — --generate "description": generate a full workflow.json + scaffolded directory from a natural language description using Claude.
  • Feature C — Auto-heal on step failure: when "auto_heal": true is set on a workflow or step, Claude automatically proposes a mutation on failure. Proposal stored as a pending mutation for human review.
  • Feature D — type: "branch" step: conditional DAG branching via a condition script (exit 0 = true path). Unchosen-branch steps are marked skipped and treated as complete for dependency resolution.
  • Feature E — Trigger system: new triggers array in workflow.json supports schedule (cron) and webhook (HTTP) triggers. --install-triggers installs launchd plists on macOS or crontab entries on Linux. New scripts/schedule_workflow.py and auto-generated webhook_server.py.
  • Feature F — Run history dashboard: --dashboard generates a self-contained dashboard.html showing all past runs (status, duration, step breakdown) with links to per-run visualizations. New scripts/dashboard.py.
  • 11 new tests; 42/42 pass.

1.0.3

  • Feature 2 — MCP steps: new type: "mcp" step executes a tool call via the MCP protocol using the .mcp.json server registry. Params support {{artifact:id}} and {{env:VAR}} template expansion. Added assets/mcp-servers.json.example.
  • Feature 3 — Sidecar mutation proposals: sidecar scripts can emit structured JSON proposals (add_step / modify_step / remove_step) via a ---PROPOSE_MUTATION--- sentinel. Runner captures and stores proposals in state/proposed-mutations.json. New CLI flags: --list-mutations, --approve-mutation ID, --reject-mutation ID.
  • apply_mutation() fills in required schema defaults so proposals without full step specs can be approved without manual editing.
  • Visualization updated: nodes with pending mutations show an amber pulsing ring; inspector panel lists mutation details.
  • UX fixes: workflow execution now prints per-step progress (→ running, ✓ complete, ✗ failed); approval gates print a clear pause message with instructions; run_workflow.py accepts workflow dir as a positional argument; init scaffold no longer generates success_gate: TODO.

1.0.2

  • Added scripts/visualize_workflow.py: n8n-style interactive HTML DAG viewer auto-generated after every run.
  • Added --visualize flag to run_workflow.py to generate the graph without executing steps.
  • Visualization features: live status polling, color-coded step types, bezier edges, GATE badges, sidecar nodes, inspector panel, search/filter, minimap, Export SVG, keyboard shortcuts (F/Esc//).
  • Fixed inspector close: canvas click and Esc now correctly dismiss the inspector panel.
  • Fixed package_skill.py: runtime directories (artifacts, logs, state, audit, runs) are now explicitly excluded from release archives to prevent data leaks.
  • Fixed run_workflow.py --step: removed duplicate setup_run_audit() call that created orphan audit runs without finalizing them.

1.0.1

  • Added project governance files: LICENSE, CONTRIBUTING.md, SECURITY.md, and CODEOWNERS.
  • Added repo maintenance tooling with ruff, pre-commit, and a packaging script for release archives.
  • Added CI smoke coverage for packaged install artifacts and a tag-driven GitHub release workflow.
  • Clarified Python support policy and release process in the documentation.

1.0.0

  • Promoted the workflow manifest to schema v4.
  • Added runtime contract enforcement for produced and consumed artifacts.
  • Added structured approvals, structured event logs, metrics, doctor/repair flows, and rollback support.
  • Added prompt-asset pinning, security policy enforcement, and migration tooling.
  • Added native step executors and parallel DAG execution support.
Assets 3
Loading