v0.7.2

Changes

• Add .kin metadata for Kindex integration
• Add Constrain artifacts (constraints.yaml, component_map.yaml, trust_policy.yaml, schema_hints.yaml, prompt.md)

v0.7.1

Changes

• Wire orphaned generate_emission_compliance_test() into decomposition pipeline — emission tests now auto-generated alongside contract and Goodhart tests
• Add save_emission_test() and emission_test_path() to ProjectManager
• Create PACT_KEY_STANDARD.md as canonical format specification for the stack
• Fix README phase count (was "Eleven phases", now accurate "Nine phases")

1766 tests passing.

v0.7.0 — Stack Evolution

What's New

Sentinel Extraction

The monitoring subsystem (Sentinel, signals, incidents, remediator, triage) has been extracted to a standalone package. PACT log key generation is retained in Pact for production traceability.

Audit Repo Separation

Two-repo separation-of-privilege model. The coding agent and auditing agent operate in different repositories — the agent being evaluated cannot modify the tests that judge its work.

• pact audit-init — initialize audit repo
• pact sync — sync visible tests (never Goodhart) to code repo
• pact certify — tamper-evident certification with SHA-256 artifact hashes

Contract Schema Evolution

ComponentContract gains two new required sections:

• data_access: reads/writes classification tiers, rationale, side_effects with field-level detail
• authority: data domain ownership declarations

Anti-cliche enforcement rejects vague rationale strings. Emission compliance tests are auto-generated from the contract interface (no LLM needed).

Stack Integration

Pact is now the contract-first build system in a larger stack:

• Constrain (--constrain-dir): seed decomposition with policies, component maps, trust policies
• Arbiter (phase 8.5): POST access_graph.json for blast radius analysis and trust scoring
• Ledger (--ledger-dir): load field-level audit assertions into contract test suites
• Sentinel (pact sentinel status|push-contract|list-keys): accept tightened contracts from production monitoring

Structured Event Emission

All generated implementations accept optional event_handler and log_handler. Every public method emits structured events with PACT keys, classification metadata, and side effects.

New Pipeline Phase: 8.5 Arbiter Gate

After Integrate, before Polish. Generates access_graph.json, registers with Arbiter if configured, handles HUMAN_GATE responses. Skippable with --skip-arbiter.

Breaking Changes

• Sentinel CLI commands (pact watch, pact report, pact incidents, pact incident) removed — use the standalone Sentinel package
• Monitoring config fields (monitoring_enabled, monitoring_auto_remediate, monitoring_budget) removed from GlobalConfig
• Per-project monitoring fields (monitoring_log_files, monitoring_process_patterns, etc.) removed from ProjectConfig
• Phase order now includes arbiter between integrate and polish

Stats

• 1,766 tests passing
• 46 files changed, 1,728 insertions, 4,039 deletions

v0.6.2

Changes

Interview: acceptance criteria -- The interview phase now asks "what does done look like?" Extracts specific, testable acceptance criteria before decomposition. Criteria flow into decomposition context and north-star validation warns when contracts don't cover them.

Resume state fix -- pact resume no longer loses component task progress. Previously, re-entering the decompose phase overwrote all component_tasks with fresh empty objects. Now preserves existing completed/failed status. State also flushes to disk immediately after task setup, surviving crashes and validation failures.

v0.6.1 — MCP Server

What's new

Complete MCP server for Claude Code and other MCP clients.

• 7 tools: pact_status, pact_contracts, pact_contract, pact_budget, pact_retrospective, pact_validate, pact_resume
• 5 resources: pact://status, pact://contracts, pact://contract/{id}, pact://budget, pact://retrospective
• Stdio transport via FastMCP
• Auto-detects project from PACT_PROJECT_DIR env var or walks up from CWD

Install

pip install pact-agents[mcp]

Run

pact-mcp                              # standalone entry point
pact mcp-server --project-dir ./proj   # via CLI

Claude Code config

{
  "mcpServers": {
    "pact": {
      "command": "pact-mcp",
      "env": { "PACT_PROJECT_DIR": "/path/to/project" }
    }
  }
}

v0.5.1: Smoke Test Generation

What's New

Mechanical Smoke Tests (pact adopt)

pact adopt now generates smoke tests from AST-extracted function signatures — no LLM required, zero cost. Each public module-level function gets:

• An import verification test (module loads without error)
• A callable check test (function exists and is callable at module level)

Filters out class methods, private functions (_-prefixed), and nested closures (detected via source line indentation).

Output: tests/smoke/ — conventional location, committable, runnable with standard pytest.

Self-Adoption Results

Codebase	Components	Functions	Smoke Tests
Pact	57	682	248 (57 files)
Baton	35	401	103 (35 files)

Test Suite

1,886 total tests (1,638 hand-written + 248 smoke) — all passing.

Full Changelog

• generate_smoke_tests() in adopt.py — mechanical test generation
• _is_module_level() — source line indentation check for nested function filtering
• Smoke test output moved from .pact/test-gen/smoke/ to tests/smoke/
• 8 new adopt unit tests covering smoke generation
• Updated docs, README, and landing page

v0.5.0: Goodhart Tests — Hidden Acceptance Criteria

What's New

Goodhart Tests — hidden adversarial acceptance criteria that implementation agents never see. Counters Goodhart's Law: when agents can see all tests, they optimize for those specific inputs rather than truly satisfying the contract.

Features

• Goodhart test author: Single-LLM-call adversarial test generation during the Test phase. Probes for hardcoded returns, boundary-adjacent inputs, invariant generalization, and postcondition universality.
• Isolated storage: Hidden tests live in .pact/contracts/<cid>/goodhart/, completely isolated from visible test loading and handoff briefs.
• Polish-phase evaluation: After all visible tests pass, Goodhart tests run as a final acceptance gate.
• Graduated-disclosure remediation: Failing components get behavioral hints (never actual test code):
  • Level 1: Vague behavioral hint from test description
  • Level 2: Specific contract invariant/postcondition
  • Max 2 attempts (configurable via max_goodhart_attempts)
• Wizard: New pact wizard command for guided project setup
• Lifecycle improvements: Polish phase integration, integrator enhancements

Cost Impact

~$0.07/component for Goodhart test generation (1 LLM call, no research/plan). For a 5-component project: ~$0.35 total overhead.

Stats

• 1,573 tests, all passing
• 23 files changed, +2,131 lines

Full changelog: v0.4.0...v0.5.0

v0.4.0: Reliability & Performance

What's New

Reliability (R2)

• Contract quality gates: audit_contract_specificity() now runs during contract authoring and validation, flagging vague language patterns
• Structured side effects: Contracts declare structured_side_effects (reads_file, writes_file, network_call, etc.) — wired into prompts and test rendering
• Performance budgets: Functions can declare performance_budget (p95 latency, max memory, Big-O complexity) — wired into prompts and test rendering
• Answer provenance: AuditedAnswer tracks source, confidence, and timestamp for every interview answer
• Dependency normalization: Case-insensitive + underscore-transposition matching in the main validation gate
• MCP resume tool: pact_resume added to MCP server for resuming failed/paused runs

Performance (R3)

• Optimized handoff briefs: include_test_code=False in implementer — agents read test files directly from disk
• Fixed Opus 4.6 pricing: Corrected from $15/$75 to $5/$25 per million tokens

New Features

• pact pricing: View current model pricing table, --export to override at ~/.config/pact/model_pricing.json
• Dynamic pricing: External pricing file loaded at config init, overridable per-project
• BudgetTracker.summary(): Structured dict with all budget metrics including cache stats

Documentation

• README: updated pricing, test count (1501), added 6 missing CLI commands
• Landing page: updated test count

Stats

• 20 files changed, +491/-35 lines
• 1501 tests passing

v0.3.1: Documentation & Landing Page Updates

What's Changed

Landing Page

• Added feature cards for wavefront scheduling, prompt caching, drift detection, retrospective learning, contract quality gates, and resume/error classification
• Updated test count from 1,260 to 1,482 across all copy

Documentation

• Updated CLAUDE.md test count (1,260 -> 1,482) and fixed working directory path
• Updated README.md test count (1,458 -> 1,482)

Stats

• 1,482 tests passing
• 22,488 lines of production code
• 18,816 lines of test code

Install

pip install pact-agents

v0.3.0: Health Monitoring, Test Generation, Codebase Adoption

What's New

Health Monitoring (Dysmemic Pressure Detection)

• Pipeline monitors its own coordination health across phases
• Tracks output-to-planning token ratio, cascade failures, budget velocity
• Automated remedies for informational findings; user-controlled remedies via FIFO for structural changes
• pact health <project> CLI command with detailed metrics, per-phase token table, and remedy suggestions
• Configurable thresholds via health_thresholds in pact.yaml
• skip_cascaded is proposed (not auto-applied) — respects C5 Brittleness Trap

Test Generation

• pact test-gen <dir> — generate tests and security audit for any codebase
• Static analysis identifies untested functions, complexity hotspots, and security-sensitive patterns

Codebase Adoption

• pact adopt <dir> — bring existing codebases under pact governance
• Analyzes structure, generates contracts and decomposition tree

Spec-Compliance Audit

• pact audit <project> — verify every requirement in task.md is covered
• Gap report showing covered, partial, and missing requirements

Other

• Python, TypeScript, and JavaScript support
• 1482 tests passing
• Multi-provider support (Anthropic, OpenAI, Gemini)

Install

pip install pact-agents