Claim Boundary Harness v1.2.0

Highlights

• Adds a generic task-local external retrieval planner that preserves original queries and exact identifiers.
• Routes official entities, DOI, RFC, arXiv, CVE, standards, PyPI, npm, Hugging Face, GitHub, and unknown namespaces through source-native or capability-discovery paths.
• Tracks evidence coverage per target and keeps provider/index misses below verified absence.
• Connects the planner receipt to the PowerShell research gate, intake router, bounded action consumer, deployment profiles, and Windows regression checks.
• Retains the nonblocking behavior-correction lifecycle and keeps task execution, tool use, semantic judgment, and final claims with the host model agent.

The external retrieval planner creates a plan and reduces supplied attempt receipts; it does not browse, block execution, or write durable memory.

Claim Boundary Harness v1.0.0

CBH v1.0.0 establishes the model-agent-first public release line.

Highlights:

• keeps the host LLM agent as the owner of planning, semantic judgment, tool use, recovery, and the final answer
• adds route-declared memory source hints and explicit memory/research action bindings
• adds bounded, provenance-preserving context selection for compound natural-language retrieval
• keeps exact indexed anchors selected while handing weak candidates to the model for semantic reranking
• extends the WorkBuddy model Agent Loop contract and deployment profiles
• keeps the safe default WorkBuddy hook chain to UserPromptSubmit plus PreToolUse; Stop remains opt-in after compatibility validation
• synchronizes VERSION, README, CHANGELOG, CITATION, adapter manifest, and contract tests to v1.0.0

Validation:

• Windows reference smoke checks passed
• Ubuntu Bash path/router smoke checks passed
• 117 public pytest contract tests passed
• 91 WorkBuddy adapter tests passed

See CHANGELOG.md for the full release boundary.