v0.7.3

CLI ergonomics + a full worked example. Two clean, separate inputs for proof run, a richer AgentContext.from_dir(), and a copy-me examples/credit_agent/.

Install / upgrade:

pip install -U proofagent-harness

Added

• proof run --context-dir DIR — load the full agent context via AgentContext.from_dir() (system prompt + tool schemas + memory + an optional agent.yaml manifest). Passing it lifts the limited-context ceilings on instruction_following / safety.
• proof run --domain-knowledge-dir DIR — supply the domain-knowledge corpus as a SEPARATE input from the agent context (grounding for hallucination scoring).
• proof run --seed — deterministic multi-turn runs from the CLI (was Python-API only).
• Manifest support in AgentContext.from_dir() — reads role / goal / business_case from an optional agent.yaml (or agent.json / manifest.*); explicit CLI flags override it. Also accepts system_prompt.txt / system.md, and a domain_knowledge/ folder as an alias of knowledge/.
• examples/credit_agent/ — a complete worked example showing how to feed your files (agent context + domain knowledge) for both multi-turn and artifact modes.
• Run-configuration table — before each evaluation, proof run / proof artifact print a summary (mode, Harness LLM + fallback, turns, consensus, metrics, context / domain-knowledge dirs, and — when uploading — the dashboard agent name, version, and gate profile). Suppress with --quiet.

Changed

• proof run --turns default is now 15 (was 8).
• proof artifact domain-knowledge flag is now --domain-knowledge-dir (with --knowledge-dir / -k kept as a back-compat alias); accepts .md / .txt / .json / .yaml.
• --api-key help now points to where to get a key (https://app.proofagent.ai → Settings → API Keys).
• --agent help clarifies it is the name shown on the governance dashboard.

Fixed

• Import ordering and lint cleanups so CI is green.

Full CLI reference: see the README and the docs.