examples/README.md

Schema-Checked Loop Contracts

Twenty JSON examples turn the operational patterns into reviewable machine-readable contracts. Every file validates against schemas/loop-contract.schema.json in CI.

A pattern explains how a class of loop should operate. Its contract pins down what one implementation may read, change, verify, spend, and escalate. A runtime starter then wires that contract to a session, schedule, CI job, or durable worker.

What A Contract Fixes

A Loop Contract is a reviewable operating specification for one recurring agent job. It is not a prompt or a runtime. In a supervised session, a person can correct scope, request evidence, remember a failed attempt, and decide when to stop. A recurring agent needs those decisions written down before a trigger starts the next run.

Decision	Contract fields	What becomes explicit
Authorize	Objective, trigger, intake, workspace	Which outcome matters, what starts a run, which work qualifies, and where action is allowed.
Operate	Context, agents, verification, state	What must be loaded, who acts and checks, what proves progress, and which receipts survive.
Govern	Budget, escalation, exit	How autonomy is capped, who owns exceptions, and what success or a blocked stop means.

Without these boundaries, a recurring agent can select the wrong work, expand its own permissions, approve itself, repeat failed attempts, or spend indefinitely. The JSON schema makes the same policy readable by people and enforceable by tooling.

Contract Catalog

Build And Maintain

Pattern and contract	Use when	Trigger	Deterministic gate	Durable receipt
PR babysitter · JSON	A PR stalls on comments, CI, or conflicts	PR event + schedule	Required checks pass and review threads are resolved	Last SHA, check URLs, thread status, blockers
CI repair · JSON	A required check fails	CI failure event	Original failing command passes	Failure excerpt, patch, commands, passing output
Docs drift · JSON	Docs may disagree with code	Weekly schedule	Examples and links pass against current behavior	Verified drift items and false positives
Dependency triage · JSON	Update PRs accumulate	Weekly + bot PR events	Tests, build, typecheck, and audit pass	Update group, risk class, commands, deferral reason
Bug hunting · JSON	A team needs recurring bug discovery	Weekly schedule	Reproducible steps or a failing test	Checked surface, trace, minimal reproduction
Release notes · JSON	A release needs evidence-backed notes	Tag or release branch	Every entry links to a merged source	Processed PRs/issues and draft changelog

Operate And Observe

Pattern and contract	Use when	Trigger	Deterministic gate	Durable receipt
Deploy verifier · JSON	A rollout needs monitoring	Deployment event + polling	Synthetic checks and thresholds hold	Rollout phase, anomalies, metric links, decision
Incident response · JSON	An alert needs evidence and ownership	Pager or SLO event	Severity and cause cite concrete impact signals	Timeline, hypotheses, owner, handoff
Data quality · JSON	A dataset refresh may drift	Refresh event or nightly	Hard schema and quality rules pass	Profile metrics, exceptions, quarantine decision
Cost control · JSON	Agent spend rises	Schedule + spend threshold	Quality holds on a comparable sample	Baseline spend, traces, proposed savings
Model routing · JSON	Model selection is inconsistent	Schedule + model/price change	Replay quality stays within tolerance	Task-class metrics and routing decision
Performance regression · JSON	Latency, throughput, or memory worsens	Metric or benchmark regression	Correctness passes and repeated measurements recover	Environment, profile, samples, candidate decision

Learn And Optimize

Pattern and contract	Use when	Trigger	Deterministic gate	Durable receipt
Feedback clusterer · JSON	Feedback is noisy	Daily or weekly schedule	Themes cite sources and preserve frequency/severity	Processed IDs, clusters, outliers
Evaluation regression · JSON	Agent evals fall below baseline	Nightly + score drop	Targeted tasks recover without scorer changes	Run IDs, traces, hypotheses, rerun scores
Benchmark optimization · JSON	A stable system should improve	Bounded experiment window	Repeated gain with correctness and protected metrics intact	Hypothesis, diff, raw scores, cost, decision
Knowledge freshness · JSON	An agent retrieval corpus is stale	Daily refresh + source change	Provenance, freshness, retrieval, and leakage gates pass	Manifest diff, checksums, evals, promotion pointer

Govern And Protect

Pattern and contract	Use when	Trigger	Deterministic gate	Durable receipt
Security review · JSON	A sensitive diff needs review	Sensitive change + weekly sweep	Findings cite reproducible evidence	Reviewed SHA, findings, false positives
Enterprise approval · JSON	A change needs formal sign-off	Gate or approver event	Every required gate has a human decision	Approval ledger, status, SLA, owner
Accessibility regression · JSON	A UI accessibility check fails	PR or preview failure	Exact rule plus required human criteria pass	Before/after scan, interaction trace, reviewer decision
Adversarial red team · JSON	An agent needs active security testing	Pre-release campaign	Independent reproduction and policy citation	Seed, full/minimized trace, severity, disclosure status

Read And Validate A Contract

Preview a contract as a compact human-readable checklist:

python3 scripts/preview_loop_contract.py examples/ci-repair-loop.json

Validate the complete contract library with the dependency-free CI validator:

python3 scripts/check_loop_contract_examples.py

Four Worked Paths

1. Repair A Failing Check

Symptom: a pull request's pytest -x job is red.

1. Choose the CI repair pattern.
2. Adapt ci-repair-loop.json: set the event trigger, protected files, three-retry budget, and escalation owner.
3. Run the test-repair starter in an isolated worktree:

CHECK_CMD="pytest -x" AGENT_CMD="codex exec" ./examples/runnable/test-repair-loop.sh

Done means: the original command passes. The agent's claim is not the gate.

2. Refresh A Support Knowledge Base

Symptom: answers cite policies and release notes that may be stale.

1. Choose the knowledge freshness pattern.
2. Adapt knowledge-freshness-loop.json: name approved sources, freshness windows, retrieval evals, leakage checks, and atomic promotion.
3. Schedule it with the Codex automation, desktop task, or your durable worker.

Done means: a versioned candidate passes provenance, freshness, retrieval, and sensitive-data gates before promotion. A successful fetch alone is not enough.

3. Process A Bounded Work Queue

Symptom: a set of narrow, independently verifiable tasks needs unattended handling.

Create tasks.jsonl:

{"id":"docs-101","objective":"Update the CLI install example","context":["README.md","src/cli.ts"],"allowed_paths":["README.md"],"verification":"python3 scripts/check_docs.py"}
{"id":"schema-204","objective":"Add the missing status enum test","context":["schemas/status.json","tests/test_status.py"],"allowed_paths":["tests/test_status.py"],"verification":"pytest tests/test_status.py -q"}

Validate the queue first, then run a bounded batch:

python3 examples/runnable/queue-worker-loop.py --queue tasks.jsonl --dry-run
python3 examples/runnable/queue-worker-loop.py \
  --queue tasks.jsonl \
  --agent-command "codex exec" \
  --verify-command "python3 verify_item.py {id}" \
  --max-items 2 \
  --max-retries 2 \
  --command-timeout 900

Done means: the external verifier exits successfully and the item ID is persisted in QUEUE_PROGRESS.jsonl; reruns skip completed work.

4. Watch A Threshold Without Auto-Remediation

Symptom: a rollout, latency metric, pass rate, or spend metric needs bounded monitoring and evidence-backed escalation.

PROBE_CMD="./scripts/p95_latency_ms.sh" \
THRESHOLD=250 \
DIRECTION=max \
MAX_SAMPLES=12 \
INTERVAL_SECONDS=300 \
AGENT_CMD="codex exec" \
./examples/runnable/threshold-monitor-loop.sh

Done means: all samples stay within policy, or the first breach produces a durable receipt and read-only diagnosis. The starter never changes production.

Adaptation Checklist

Before running any contract unattended:

• replace every generic source, command, file, and destination with a real one;
• keep the acting agent away from tests, scorers, policies, or thresholds that judge its work;
• narrow allowed paths, tools, credentials, and network access;
• make receipts reviewable and safe to retain;
• set time, retry, token, cost, and concurrency budgets;
• test the escalation path before trusting the success path;
• run first in a branch, worktree, sandbox, synthetic target, or read-only mode.