@workspace/output-contract-gate

Prototype runtime for contract-enforced agent output.

Current scope:

• parse list-form EDN response contracts
• normalize them into a usable IR
• parse Markdown into an AST
• validate the five-section ημ response shape deterministically
• compile bounded repair prompts from machine failures

This package is the first implementation slice for:

• specs/drafts/contract-enforced-agent-output-pipeline-2026-03-23.md

Commands

pnpm --filter @workspace/output-contract-gate build
pnpm --filter @workspace/output-contract-gate test
pnpm --filter @workspace/output-contract-gate validate -- --contract ../../specs/drafts/contract-enforced-agent-output-pipeline.example.edn --response ./sample.md

CLI

Validate a Markdown response against an EDN contract file:

pnpm --filter @workspace/output-contract-gate build

node devel/packages/output-contract-gate/dist/cli.js \
  --contract devel/specs/drafts/contract-enforced-agent-output-pipeline.example.edn \
  --response /tmp/candidate.md \
  --artifacts-root devel/artifacts/output-contract-gate

The CLI prints JSON.

• exit 0 = structure passed
• exit 1 = structure failed; JSON includes repairPrompt
• exit 2 = CLI/IO/contract loading error

By default the CLI writes a run bundle under:

./artifacts/output-contract-gate/<run-id>/

Use --artifacts-root <dir> to override or --no-artifacts to suppress writing.

Current artifact bundle:

• input.json
• contract.edn
• contract-ir.json
• candidate.md
• candidate.ast.json
• validation-report.json
• final-decision.json
• repair-prompt.txt when structure fails

Generate mode

Generate a candidate, then pipe it through the structure gate and, on success, the review stub:

node devel/packages/output-contract-gate/dist/cli.js generate \
  --contract devel/specs/drafts/contract-enforced-agent-output-pipeline.example.edn \
  --task-text "Turn this request into the required five-section response." \
  --generator fixture-valid \
  --artifacts-root devel/artifacts/output-contract-gate

Supported generators:

• fixture-valid
• fixture-invalid
• openai-chat

openai-chat uses an OpenAI-compatible POST /chat/completions transport.

Default model for openai-chat remains gpt-5.4. qwen3.5 was verified successfully through local proxx, but must be selected explicitly when desired.

Useful flags:

• --task-file <path>
• --task-text <text>
• --generator <mode>
• --base-url <url>
• --model <id>
• --api-key <token>
• --temperature <n>

Generate mode adds:

• task.txt
• generation-report.json

and, when structure passes, also writes:

• review-report.json

Review stub

Once a structurally valid bundle exists, emit a machine-shaped stub review report:

node devel/packages/output-contract-gate/dist/cli.js review-stub \
  --bundle devel/artifacts/output-contract-gate/<run-id>

This writes:

• review-report.json

and augments:

• final-decision.json

The stub reviewer is deterministic and heuristic-only. It is a placeholder for the later GPT-family semantic reviewer.

Status

Prototype only. The first reference contract is the five-section response shape:

• Signal
• Evidence
• Frames
• Countermoves
• Next