Changelog

1.1.0

• Reject empty or blank acceptance criteria and required checks, including legacy empty completion plans.
• Require passing verification to reference a successful Sentinel command or hashed observation files; text alone cannot pass.
• Recheck source inputs and evidence hashes at completion, and block any recorded failure.
• Compare artifacts before and after builds. Unchanged artifacts pass only with a prior successful record for the same command, inputs, and artifact hash.
• Label reported observations explicitly: file integrity does not prove the observation is true.
• Add regression tests for old artifacts, cached builds, tampering, stale verification, empty plans, and invented command evidence.
• Integrate semantic freshness and protocol fixes from 1.0.2; explicitly superseded tasks preserve their failed/unfinished evidence in history.

1.0.2

• Purge stale semantic rows when indexed files are deleted or become unreadable
• Include non-ignored untracked source files in semantic refreshes
• Prevent a new task from silently replacing an active task
• Negotiate only explicitly supported handshake-era MCP versions and counter-offer 2025-11-25 for unknown/modern initialize requests
• Clarify that the command allowlist is a process governor, not an OS sandbox
• Add regression tests for deletion, untracked indexing, active-task replacement, and protocol negotiation

1.0.1

• Fixed stdio transport to use newline-delimited UTF-8 JSON-RPC as required by MCP instead of Content-Length framing
• Replaced the self-confirming framing test with newline-delimited transport coverage
• Added an end-to-end test through the official MCP Python client SDK
• Added best-effort redaction for common credentials in returned source snippets and governed command output
• Kept runtime dependencies standard-library-only; the MCP SDK is test-only

1.0.0

• Added adaptive per-task context budgeting
• Added exact duplicate-context suppression for unchanged worktree generations
• Added ranked context_bundle orientation
• Added explicit context-budget status and justified extension workflow
• Added compact final task_report
• Added artifact SHA-256 reporting
• Enforced hard stop of governed context and deployment polling after completion
• Added cross-platform GitHub Actions tests

0.3.0

• Added persistent local SQLite semantic repository memory
• Added incremental indexing of definitions, imports and references
• Added semantic_find, dependency_context, change_impact and stale-index detection

0.2.0

• Rebuilt the original proof of concept as a governed MCP bridge
• Added task state, required verification, bounded commands, artifact-aware build stopping and deployment polling limits

0.1.0

• Initial bounded repository orientation proof of concept