v0.2.0

Release Notes v0.2.0

Date: 2026-08-23
Package: planner-critic v0.2.0 — PyPI

What's New

Deterministic Loop Efficiency (M2)

• Topological auto-repair (#130) — re-orders tasks to fix ordering-only violations without LLM revision cost
• Deterministic precondition closer (#131) — auto-injects template-matched precondition steps, eliminating unverified_precondition blockers without LLM revision
• Oscillation detection + auto-converge (#152) — detects structural cycling and auto-converges non-oscillating tasks

Domain Pack Framework & Scaffolding (M3–M4)

• Domain Pack protocol (#139) — DomainPack protocol for domain-specific gates, manifest loader, engine integration
• 4 domain packs (#140–143) — SecOps, Supply Chain, FinOps, Data Engineering — each with domain-specific deterministic gates
• Policy-as-Code engine (#129, #156) — OPA/Rego + CEL policy evaluation; RegoGate/CelGate integrate with the deterministic gate pipeline
• plancritic init --template (#155) — scaffold new domain packs from templates
• Inverse-rollback synthesizer (#160) — auto-generates rollback steps for planned actions
• pytest-planner-critic plugin (#156) — use gate assertions in your test suite

Security & Trust Oracle (M5)

• SWE-bench security corpus (#123) — 7 instances across 7 CWE buckets as ground truth
• Security oracle harness (#124) — validate critic against human-labeled ground truth
• Adversarial injection harness (#125) — 21 injection trap patterns, per-layer attribution
• Deterministic-gate security regression (#126) — 35 flawed variants, 100% blocked
• Standing-rule promotion (#127) — label-migration harness for security-policy promotion

Enterprise-Scale Safety (M6)

• Dynamic posture resolver (#149) — resolve risk_tolerance per-run from policy context
• Run budgets (#150) — hard limits on LLM spend per planning session
• State locking (#151) — StateLock with WAIT/ESCALATE strategies for concurrent planning
• Precondition ledger (#158) — audit trail of preconditions established across plan versions
• Blast-radius quotas (#132) — restrict concurrent task scope by action type and cluster
• Secret/PII redaction (#159) — regex-based redaction with audit trail, wired into every external-output surface
• Gate-rationale metadata (#174) — every gate result includes explanation of why it passed or blocked
• Plan-signature persistence (#176) — cryptographic plan signatures for integrity verification

Developer & Interactive Surfaces (M7)

• plancritic check (#137) — quality-check a plan against all gates
• plancritic diagnose (#175) — diagnose why a plan failed gates
• plancritic domains (#162) — list available domain packs and their gates
• plancritic policy (#178) — evaluate Rego/CEL policies against a plan
• plancritic templates (#179) — list and scaffold from domain-pack templates
• @guardrail decorator (#137) — Python decorator for inline plan guardrails
• Seed Rego library (#178) — 4 starter policies (blast-radius, naming, least-privilege, quota)
• Drift observability (#181) — plancritic findings CLI + drift alert z-score

Enterprise Integration & Adoption (M8)

• GitHub Action (#128) — community action for CI pipeline integration
• GitLab CI template (#128) — ready-to-use .gitlab-ci.yml template
• AutoGen adapter (#134) — integrate PlannerCritic as an AutoGen agent
• Webhook notifier (#161) — Slack/Teams/Generic webhook delivery with HMAC + JWT verification

Scale Validation & Enhancement (M9)

• 5 new enterprise field-test corpora — IDP, MAO, SRE, SCP, FNG — 14 goal files, 16 new reason codes
• Auto-repair benchmark (#177) — ≥30% revision reduction on ordering-violation corpus
• Rollback credibility field test (#182) — 21 goals across 8 domains, 3 credibility patterns
• Family-histogram stasis benchmark (#183) — ≥20% revision reduction from family-based convergence

Release-Gate Field Test (M10)

• 170 goals across 40 domains — 73 approved (balanced), 97 escalated (strict), 8 escalated (adversarial), 0 true failures
• 90 deterministic subsystem tests — covering all v0.2.0 features
• 3 benchmarks completed — auto-repair, rollback credibility, family-histogram stasis
• Security oracle: 7/7 correct plans pass, 35/35 flawed variants blocked, 21 injection traps generated
• Release gate: PASS on all 8 blocking criteria

Code Quality

• 31 bugs found and fixed by pre-release code review (#184–#214) — including 6 critical (redaction offset corruption, RegoGate --data→--input, dead @re_gate, MCP critic cache, diagnose KeyError, oracle double-count)
• Coverage ≥95%
• Lint clean (ruff + mypy strict)

Breaking Changes

• Configuration format — The plancritic-fieldtest.toml v0.1.0 config file is superseded by the new provider registry. Use plancritic providers add or write a plancritic.toml with [providers.*] sections.
• Domain-pack config — pack_config is now part of the DomainPack Protocol, not a module-level global. Existing domain packs must migrate to the protocol.
• API key handling — Hardcoded api_key values in config files were removed in favor of ${ENV_VAR} references. Set OPENROUTER_API_KEY or equivalent in your environment.

Upgrade Path

1. Config migration: Replace v0.1.0 config files (plancritic-fieldtest.toml, plancritic-asymmetric.toml, etc.) with the new format. Use plancritic migrate --path .plancritic/plans.db to upgrade the plan store schema if needed.
2. Domain packs: If you authored custom domain packs in v0.1.0, update them to implement the DomainPack protocol (see docs/design/domain-pack-design.md).
3. API references: Replace api_key = "sk-..." with api_key = "${OPENROUTER_API_KEY}" in config files.

Field Test Results

Metric	Result
Balanced goals approved	73/73 (100%)
Strict goals escalated	97/97 (100%)
Adversarial goals escalated	8/8 (100%)
True failures	0
Deterministic gate passes	170/170 (100%)
Security oracle	7/7 correct, 35/35 flawed, 21 traps
Scorecard A (pre-amended)	PASS
Scorecard B (pass* semantics)	100%
Deterministic tests	90/90 pass
Benchmarks	3/3 complete

Known Issues

• Planner capability gap — strict tolerance + LLM critic = structurally unable to approve non-trivial plans. Usable with balanced tolerance in production; strict is for adversarial/security testing.
• Local model support — DeepSeek-R1-8B cannot produce valid structured JSON; Qwen3-4B critic is too weak (false-positive approvals on strict goals). Use openai/gpt-4o-mini via OpenRouter.
• TUI / studio / IDE surfaces — deferred to v0.3.0.
• Backstage plugin / Slack dashboard — deferred to v0.3.0.
• Adaptive revision cap — not yet implemented.

Documentation

Doc	Location
Field Test Results	docs/field-test/v0.2.0/field-test-results-0.2.0.md
Field Test Plan	docs/field-test/README.md
Architecture	docs/architecture/architecture-v0.1.0.md
API Reference	docs/reference/api.md
Design Decisions	docs/design/design-decisions.md
Domain Pack Design	docs/design/domain-pack-design.md
Policy Engine Design	docs/design/policy-engine-design.md
Enterprise Safety Design	docs/design/enterprise-safety-design.md
Developer Surfaces Design	docs/design/developer-surfaces-design.md
Integration Surfaces Design	docs/design/integration-surfaces-design.md
Security	SECURITY.md
Changelog	CHANGELOG.md

What's Changed

• docs(wbs): add v0.2.0 WBS — 10 milestones across 9 part files by @deghosal-2026 in #170
• M5: Security & Trust Oracle (SWE-bench Verified) by @deghosal-2026 in #180

Full Changelog: v0.1.0...v0.2.0