OpenUdon

OpenUdon is the public UWS workflow authoring, review, package, and executor-handoff tool. It can run directly or under optional Symphony-managed orchestration, and it hands approved packages to a trusted executor boundary such as the udon runtime.

It owns project templates, optional Symphony workflow policy, example artifacts, deterministic validation, review handoff evidence, package digests, credential policy, and trusted-runner glue. Public workflow semantics belong in github.com/OpenUdon/uws; OpenAPI search/discovery/import/indexing belongs in github.com/OpenUdon/apitools.

Quick Start

Install the main CLI:

go install github.com/OpenUdon/openudon/cmd/openudon@latest

Optional companion tools:

go install github.com/OpenUdon/openudon/cmd/icot@latest
go install github.com/OpenUdon/openudon/cmd/udon-runner@latest

Useful checks:

go test ./...
go vet ./...
go run ./cmd/openudon check
go run ./cmd/openudon check-apitools-boundary
go run ./cmd/openudon validate ./examples/uws-validation
make check
make release-check
git diff --check

Execute through openudon run and the portable run-config handoff. Configure the final executor with OPENUDON_EXECUTOR as either an absolute binary path or docker://<image>.

Layout

• cmd/openudon: local CLI for checks, synthesis, assessment, eval, readiness, approval templates, and trusted execution.
• cmd/icot: guided authoring CLI for project.md and workflows/intent.hcl.
• internal/: reusable OpenUdon implementation.
• examples/: committed examples and eval corpus.
• templates/project.md: starter project brief.
• docs/: detailed architecture, safety, operator, XRD, and release notes.

Execution Boundary

The intended lifecycle is:

natural-language project brief
  -> Symphony-managed implementation issue or local authoring session
  -> generated OpenAPI/UWS artifacts
  -> deterministic validation and review
  -> approved handoff package
  -> trusted executor handoff

openudon synthesize, openudon build, openudon promote, openudon assess, cmd/icot, and eval commands generate, compile, validate, and report on artifacts. They do not execute production workflows.

openudon run is separate. It validates the handoff manifest, stored and current quality, approval JSON, package digest, and tier before writing a non-secret openudon.executor-run.v1 run config and invoking the Go trusted executor runner. The runner stages the reviewed UWS/OpenAPI files into the run workdir before calling the configured executor. The runner is also available directly as go run ./cmd/udon-runner --config <run-config.json>. OPENUDON_EXECUTOR accepts either an absolute path to an executable file or docker://<image>. The outer OPENUDON_UDON_RUNNER override must be an absolute path to an executable file.

Authoring With iCoT

iCoT turns a project idea into reviewed authoring artifacts. It writes project.md and workflows/intent.hcl; it does not synthesize compiled artifacts or execute workflows.

go run ./cmd/icot --example ./examples/<name>

Common modes:

# Print rendered project.md and intent.hcl without writing files.
go run ./cmd/icot --example ./examples/<name> --print

# Use the fixed manual flow without LLM extraction.
go run ./cmd/icot --example ./examples/<name> --no-llm

# Seed from an existing example.
go run ./cmd/icot --from-example ./examples/eval/runtime-only-render --example ./examples/<name>

# Use YAML or JSON session/legacy answers.
go run ./cmd/icot --answers ./answers.yaml --example ./examples/<name>

# Rebuild project.md from workflows/intent.hcl.
go run ./cmd/icot reconcile --example ./examples/<name>

# Check authoring quality, intent parseability, and advisory drift.
go run ./cmd/icot lint --example ./examples/<name>

# Replay eval references through iCoT and save ignored transcripts.
go run ./cmd/icot replay-eval --root ./examples/eval --provider copilot-api --model gpt-5.4-mini

iCoT autosaves incomplete local sessions under <example>/.icot/session.yaml and resumes by default. Successful saves delete the autosave. Transcripts are written under <example>/.icot/transcript.json unless --no-transcript is used. These local files are ignored by git.

Side-effect scope in iCoT:

• read-only: generate and validate artifacts only.
• sandbox-only: sandbox proof runs require approved_for_sandbox, approved bindings, and a trusted runner.
• after-approval: sandbox and production execution require the full OpenUdon/Symphony approval path.

Synthesize And Assess

Generate all reviewed artifacts for an example:

export COPILOT_API_BASE_URL=http://localhost:4141
export OPENUDON_LLM_PROVIDER=copilot-api
export OPENUDON_LLM_MODEL=gpt-5.4-mini

go run ./cmd/openudon synthesize \
  --example ./examples/support-email \
  --provider "$OPENUDON_LLM_PROVIDER" \
  --model "$OPENUDON_LLM_MODEL" \
  --max-attempts 5

The command reads project.md, discovers or imports OpenAPI documents under openapi/, writes workflows/intent.hcl when needed, and generates equivalent public UWS HCL/YAML workflow artifacts:

expected/plan.json
expected/plan.md
expected/discovery.json
expected/refinement.json
expected/refinement.md
expected/review.md
expected/symphony-handoff.json
expected/quality.json
expected/quality.md

Use narrower stages after editing artifacts:

# intent.hcl -> workflow/UWS/plan/review/quality
go run ./cmd/openudon build --example ./examples/support-email --max-attempts 5

# workflow.hcl -> UWS/review/quality
go run ./cmd/openudon promote --example ./examples/support-email

# quality reports only
go run ./cmd/openudon assess --example ./examples/support-email

The bounded refinement loop records retried stages, failed checks, and stop reason in expected/refinement.json.

Quality And Repair Loop

The pipeline is validation-first:

1. Run synthesize for a new or substantially changed project.md.
2. If it fails, read expected/refinement.json and expected/quality.json.
3. Repair the earliest failing stage.
4. For openapi.*, add a valid local OpenAPI file or explicit OpenAPI URL.
5. For intent.*, edit project.md or workflows/intent.hcl, then rerun build.
6. For workflow.*, prefer improving intent and rerunning build; use promote and assess for narrow workflow repairs.
7. For uws.*, review.*, symphony_handoff.*, or artifacts.*, repair the generated artifact or evidence, then run promote or assess.
8. Stop after the configured attempt limit and report blocking checks if quality still fails.

Eval And Release Evidence

Use deterministic checks for routine development:

go test ./...
go vet ./...
make check
git diff --check

Use the eval harness when changing prompts, synthesis/refinement behavior, model defaults, or quality gates that could affect generated artifacts:

go run ./cmd/openudon eval --root ./examples/eval --provider copilot-api --model gpt-5.4-mini

Eval reports are written under ignored eval/runs/. They include pass/fail summaries, provider/model/mode/prompt-version breakdowns, approximate prompt-token totals, generated workspace paths, provider drift watch data, and comparison against a previous report when available.

Use release gates only for candidate release evidence:

make release-eval

make release-eval uses OPENUDON_LLM_PROVIDER and OPENUDON_LLM_MODEL, defaulting to copilot-api and gpt-5.4-mini, and requires the current eval corpus size as the minimum brief count.

Readiness

Local readiness reports record optional sibling checkout presence, deterministic gate results, git state, ignored local artifacts, provider credential environment presence as booleans only, and current maintainer automation policy.

go run ./cmd/openudon readiness --out eval/readiness/local.json
go run ./cmd/openudon readiness --run-gates --out eval/readiness/local.json

GitHub Actions runs public-module vet/test gates without local sibling checkouts. Real-provider release evidence remains local/manual.

Trusted Execution

After artifacts pass review, generate approval JSON with the current package digest:

mkdir -p approvals
go run ./cmd/openudon approval-template \
  --example ./examples/support-email \
  --state approved_for_sandbox \
  --reviewer "Reviewer Name" \
  > approvals/support-email-sandbox.json

Validate approval, quality, handoff policy, package digest, and tier compatibility before trusted executor handoff:

go run ./cmd/openudon run \
  --example ./examples/support-email \
  --tier sandbox \
  --approval approvals/support-email-sandbox.json

Use --dry-run to validate all gates and write run config without invoking the executor.

Approval JSON shape:

{
  "version": "openudon.approval.v1",
  "scope": "examples/support-email",
  "state": "approved_for_sandbox",
  "reviewer": "Reviewer Name",
  "approved_at": "2026-04-29T12:00:00Z",
  "expires_at": "2026-05-06T12:00:00Z",
  "package_sha256": "<current handoff package digest>",
  "notes": "optional"
}

Tier rules:

• sandbox accepts approved_for_sandbox or approved_for_production.
• production accepts only approved_for_production.
• Expired approvals fail.
• Scope mismatch fails.
• Package digest mismatch fails.
• Stored or current quality failures fail.
• Malformed handoff manifests fail.
• Credential-value artifacts and direct production execution remain prohibited.
• Approval JSON and saved run configs from before the OpenUdon package rename should be regenerated so scope, version, and package digest fields match the current artifact set.

Symphony Agent Workflow

OpenUdon issues may be run through Symphony-managed Codex sessions. Agents should follow this policy:

• Use UWS as the workflow interchange format.
• Use OpenAPI for HTTP method, path, schema, server, and security details.
• Use extension-owned UWS operations for non-HTTP runtimes such as SMTP, command execution, SSH, SQL, or LLM calls.
• Use ../uws for public schema/model validation.
• Use openudon run to hand approved UWS/OpenAPI packages to a trusted executor such as udon.
• Use ../symphony only as the work orchestration service.
• Do not execute production side effects directly from an agent session.
• If execution is requested, produce or update the approved artifact and document the trusted runner command.

Expected artifact locations:

examples/<name>/project.md
examples/<name>/openapi/
examples/<name>/workflows/intent.hcl
examples/<name>/workflows/workflow.hcl
examples/<name>/workflows/workflow.uws.yaml
examples/<name>/expected/plan.json
examples/<name>/expected/plan.md
examples/<name>/expected/discovery.json
examples/<name>/expected/refinement.json
examples/<name>/expected/refinement.md
examples/<name>/expected/review.md
examples/<name>/expected/symphony-handoff.json
examples/<name>/expected/quality.json
examples/<name>/expected/quality.md

Before handoff:

go test ./...
go vet ./...
make check
git diff --check
go run ./cmd/openudon validate examples/uws-validation
go run ./cmd/openudon assess --example examples/<name>

If side-effectful execution is explicitly requested, use openudon run with approval JSON. Do not run production effects from synthesis, build, promote, assess, iCoT, or eval.

Model And Credential Guidance

Use the local copilot-api proxy with gpt-5.4-mini as the default model for synthesis. OpenUdon reliability comes mostly from prompt preprocessing, structured output when available, deterministic quality gates, and bounded repair attempts. Escalate to a larger model only after the default model fails deterministic checks.

LLM credentials must come from provider environment variables such as COPILOT_API_BASE_URL, COPILOT_API_KEY, GEMINI_API_KEY, OPENAI_API_KEY, or ANTHROPIC_API_KEY. Do not place tokens in prompts, commands, examples, or workflow artifacts.

Use OPENUDON_LLM_PROVIDER and OPENUDON_LLM_MODEL when you want shell-level defaults for local LLM-assisted commands; explicit --provider and --model flags still take precedence.

More Documentation

• Project authoring
• Intent contract
• Data flow
• Safety
• Eval gallery
• Release note template
• Contributing
• License