ship 0.2.0: single-agent MVP with LangChain quickstart under 10 minutes

[cchinchilla-dev]

Description

Master tracking issue for the 0.2.0 release. Scope is the single-agent MVP: from a user's perspective, pip install agentanvil + a contract YAML + a LangChain agent = a passing test run in under 10 minutes, with a reproducible report.

Current version: 0.1.x (foundation patches) → Target: 0.2.0.

The version jump is justified by the scope:

• Static consistency analyzer on the contract model
• Analyzer, Generator, DockerRunner, Reporter — the full orchestration loop lands.
• Evaluator layers 1 + 2 (single judge) — covers all objective metrics of Layer 1 and a single LLM-as-judge for Layer 2 (ensemble deferred to 0.3.0).
• AgentLoomBackend guarded by agentloom[contracts,observability]>=0.5.0 optional dependency.
• Record/replay envelope complete for single-agent flows with determinism CI job.
• Quickstart LangChain example ≤ 10 min verified in CI.
• Portability invariant CI job certifies pip install agentanvil (no extras) runs end-to-end via DirectBackend.
• mkdocs-material docs site scaffolded.

Out of scope for 0.2.0 (deferred to 0.3.0 and later) is listed at the bottom.

How to use this issue

• Each item is a coherent body of work with shared dependencies.
• Items within a phase are parallelizable unless explicitly noted.

Phase A — Contract model completion

The static consistency analyzer is the deliverable of 0.2.0.

• add static consistency analyzer to core.contracts #4 add static consistency analyzer to core.contracts
• add analyzer module with framework detection and AgentProfile #5 add analyzer/ module with framework detection and AgentProfile
• add generator module with happy_path, edge_case, and policy scenario categories #6 add generator/ module with happy_path, edge_case, policy scenario categories

Parallelization: #4 is independent of #5 and #6. #5 and #6 are independent once contract types are stable.

Phase B — Runners and backends

• add DockerRunner with container digest pinning #7 add DockerRunner with container digest pinning
• flesh out DirectBackend with streaming, bounded retries, rate-limit handling, pricing tables #8 flesh out DirectBackend with streaming, bounded retries, rate-limit handling, pricing tables
• add AgentLoomBackend guarded by optional agentloom[contracts,observability] dependency #9 add AgentLoomBackend guarded by optional agentloom dependency

Parallelization: all three in parallel. #7 depends on Runner ABC from 0.1.x; #8 depends on DirectBackend skeleton from 0.1.x; #9 depends on AgentLoom 0.5.0 availability.

Phase C — Evaluator and reporter

• add evaluator Layer 1 (objective) #10 add evaluator Layer 1 (objective) — latency, cost, tool correctness, format compliance, policy_violations_rule, constraint_compliance
• add evaluator Layer 2 (single LLM-as-judge) with structured output #11 add evaluator Layer 2 (single LLM-as-judge) — structured output, configurable model
• add reporter with JSON, HTML, Markdown outputs #12 add reporter/ with JSON, HTML, Markdown outputs

Parallelization: #10 and #11 in parallel after contract types are stable. #12 depends on evaluator output shape.

Phase D — CLI, record/replay, quickstart, CI

• add CLI #13 add CLI: validate, run, replay, version commands
• add quickstart LangChain example with end-to-end CI and 10-minute budget #14 add quickstart LangChain example with end-to-end CI and ≤ 10 min budget
• add portability invariant and determinism CI jobs #15 add portability invariant and determinism CI jobs
• scaffold mkdocs-material docs site #16 scaffold mkdocs-material docs site
• add first Tier A single-agent case study under examples/case-studies/ #17 add first Tier A single-agent case study under examples/case-studies/

Parallelization: #13 depends on all earlier modules. #14 depends on #13. #15 can start once #14 lands. #16 independent.

Cross-phase dependency map

Phase A (Contract + Analyzer + Generator)
  |
  +--> Phase C (Evaluator + Reporter)
  |     |
  |     +--> Phase D (CLI, Quickstart, CI, Docs)
Phase B (Runners + Backends)
  |
  +--> Phase C

What is deliberately not in 0.2.0

The deferred items are tracked but not gated by 0.2.0:

Multi-agent and A2A (defer to 0.3.0):

• Multi-agent topologies, ConversationGraph, pathology detection.
• A2A spec pinning, conformance suite, differential testing.

Corpus and statistical analysis (defer to 0.3.0):

• Dataset ingestion adapters.
• Ensemble Agent-as-Judge.
• Active sampling.
• Five inferential contrasts.

Transversal (defer to 0.4.0):

• Regression detection.
• Security coverage (prompt injection, PII, secrets).
• CI/CD integration templates.
• K8sRunner.
• ≥ 15 case studies execution.

0.5.0 release (defer to 0.5.0):

• Replication package generator.
• SBOM + Sigstore.
• DOI Zenodo.

What 0.2.0 unlocks

After 0.2.0 ships:

• First Tier A single-agent case studies are runnable: contract + scenario + recording + report.
• Portability invariant is CI-enforced: pip install agentanvil (no extras) works end-to-end via DirectBackend.
• Determinism is CI-enforced: record → replay × 100 yields identical reports.
• Docs site is live: getting started, contract reference, backends, evaluators.
• the formal-contract reference can be drafted from the static analyzer implementation.
• Quickstart LangChain < 10 min is verifiable on a fresh venv.

Issue inventory (14 total)

Phase	Count	Issues
A — Contract completion	3	#4, #5, #6
B — Runners + backends	3	#7, #8, #9
C — Evaluator + reporter	3	#10, #11, #12
D — CLI, quickstart, CI, docs, case study	5	#13, #14, #15, #16, #17

Total in 0.2.0: 14 issues (12 enhancement + 1 chore + 1 docs).

Notes

• This is a release-tracking issue, not an implementation issue. Each child issue retains its own discussion and scope.
• Dependencies on AgentLoom 0.5.0 concern only add AgentLoomBackend guarded by optional agentloom[contracts,observability] dependency #9 (AgentLoomBackend). All other items ship without AgentLoom.
• The portability invariant added by add portability invariant and determinism CI jobs #15 is the first CI-level guarantee that AgentAnvil works standalone; it holds across all subsequent phases.