v0.2.2 — Cohere Chat v2 adapter

Added

• New cohere-chat adapter: imports Cohere Chat API v2 message lists (tool_plan / tool_calls assistant messages, document-block tool results, message-level citations) into the canonical trace format, with auto-detection.
• Cohere shares the tool_calls wire shape with OpenAI, so detection scans for Cohere-only markers (tool_plan, assistant message-level citations, document tool-result blocks) before the OpenAI rules.
• An assistant tool_plan maps to a model_response step marked metadata.source_field: "tool_plan" so plan text stays visible to safety checks without being mistaken for the final answer.

Full changelog: https://github.com/felmonon/agent-reliability-harness/blob/main/CHANGELOG.md

agent-reliability-harness v0.2.1

Fixed

• Modernized packaging license metadata to the PEP 639/SPDX form
  (license = "MIT", license-files = ["LICENSE"]) and removed the
  deprecated license classifier; builds no longer emit setuptools license
  deprecation warnings.
• Converted README documentation links to absolute GitHub URLs so they work
  from the PyPI long-description renderer.
• Corrected stale compatibility language that incorrectly claimed every
  v0.1 score remained identical despite the documented nested-safety-scan
  fix (70.0 -> 70.83 for one sample; verdict unchanged).

agent-reliability-harness v0.2.0

Added

• Trajectory rules (sequence): required tools, forbidden tools,
  partial-order call_order, per-tool min_calls/max_calls
  (ARH-SEQ-001..005).
• Flow rules (error_handling, side_effect): ignored tool errors,
  retry storms, duplicate side-effect protection (ARH-FLW-001..003).
• Completion rules (completion): required final response, max step
  count (ARH-CMP-001..002).
• Argument value constraints: enum, full-match pattern, numeric
  min/max on tool arguments (ARH-SCH-007..009), alongside the legacy
  bare type-name form.
• Token budgets (budgets.max_total_tokens, ARH-BUD-004) with an
  explicit warning when the trace carries no token data (ARH-BUD-005)
  instead of a silent pass.
• Citation validity (grounding.require_valid_citation_urls,
  ARH-GRD-003) and step-level status/error, input_tokens/
  output_tokens, and metadata trace fields.
• Stable rule IDs (ARH-*) on every finding, plus expected/observed/
  remediation evidence fields and a generated rule reference
  (docs/rules.md).
• Regression engine: arh compare diffs a candidate run against a
  baseline report (new/resolved findings, pass/fail transitions,
  score/latency/cost deltas) with gate modes regressions (default),
  failures, never, and --max-score-drop.
• Adapters: --format openai-chat (Chat Completions tool_calls
  transcripts) and --format anthropic-messages (tool_use/tool_result
  conversations), with deterministic auto-detection and lossless handling of
  unmappable input via metadata.adapter.notes.
• Reports: JUnit XML (--junit-out) and SARIF 2.1.0 (--sarif-out)
  renderers; rule IDs shown in console and Markdown output; JSON reports
  carry schema_version.
• Schema versioning: traces/policies accept schema_version (major 1);
  unknown major versions are rejected with a precise error.
• JSON Schemas for trace, policy, and report formats (schemas/).
• Benchmark suite: 34 seeded failure cases with a generator, a runner
  enforcing thresholds (precision/recall 1.0, byte-identical determinism,
  <50 ms/trace, metamorphic invariances, adapter equivalence, regression
  scenarios), plus BENCHMARK-METHODOLOGY.md and generated
  BENCHMARK-RESULTS.md.
• GitHub Action (action.yml): validate + optional baseline compare,
  step-summary reports, SARIF/JUnit outputs, fork-safe, no secrets.
• CI: 3-OS × Python 3.11-3.13 matrix, ruff, strict mypy, benchmark
  thresholds, packaging build with clean-environment install check.
• Documentation set (docs/), specs (SPEC.md, TRACE-SPEC.md,
  POLICY-SPEC.md, ARCHITECTURE.md, and more), and community files.

Fixed

• (review) Regression loose-matching could silently cancel two findings
  with different rule IDs on the same step whenever any legacy finding was
  present, hiding real regressions from the CI gate. Loose matching now
  applies only when at least one side of a pair lacks a rule_id.

• (review) Safety scanning now recurses into nested lists/dicts and the
  step error string; unsafe content one level deep no longer evades
  detection.

• (review) Warning-only traces can no longer fail: unverifiable budget
  warnings (ARH-BUD-005, new ARH-BUD-006 latency, new ARH-BUD-007 cost) are
  score-neutral.

• (review) Malformed telemetry (negative/non-finite/mistyped latency,
  cost, tokens; non-string text/tool_name/error) is rejected at parse time.

• (review) ArgSpec patterns are compiled at policy load (precise error
  instead of a traceback), min > max and non-finite bounds are rejected,
  NaN values can no longer bypass range constraints, and unknown declared
  types no longer disable enum/pattern/range checks.

• (review) JUnit output strips XML-1.0-illegal control characters so
  strict CI consumers can always parse it; Markdown reports escape
  trace-derived fields in table cells.

• (review) The GitHub Action passes all inputs via env: bindings
  instead of interpolating ${{ inputs.* }} into the script, closing a
  shell-injection vector.

• (review) CLI: --fail-under is bounds-checked (usage error, exit 2);
  pathologically nested JSON fails with a clean message instead of a
  RecursionError traceback; the module docstring documents real exit
  codes.

• v0.1.x reported type errors for required arguments twice (the argument
  was checked in two loops). Each argument is now checked exactly once; the
  duplicate finding is gone. Scores and verdicts are unchanged.

Compatibility

• v0.1.x trace and policy files produce identical scores, verdicts, and
  finding messages (pinned by golden tests). New report fields are additive.
  arh compare accepts v0.1.x baselines via a rule-id-agnostic fallback
  match. Exit codes and console markers are unchanged. See COMPATIBILITY.md.