Releases: azank1/cdv
Releases · azank1/cdv
Release list
v1.0.1 — dependency guardrails
Fixed
- Cap
mcpdependency at<2— the unpinnedmcp>=1.0.0let pip resolve the new mcp 2.0 line, which the CDV server isn't validated against yet - Pin ruff/mypy in dev extras so CI and local runs use identical lint gates
No functional changes.
v1.0.0 — CDV: the judge for AI coding agents
CDV (formerly PromptLoop/loopllm) — full rebrand and 1.0. pip install "cdv[mcp]"
Breaking changes
- PyPI package
loopllm→cdv; CLI commandloopllm→cdv - Python imports
loopllm.*→cdv.*(public API symbols unchanged) - All 36 MCP tools
loopllm_*→cdv_* - Env vars
LOOPLLM_*→CDV_*; state dir~/.loopllm→~/.cdv; audit artifact.loopllm/audit.json→.cdv/audit.json - VS Code extension:
loopllm-prompt-gauge→cdv-judge("CDV — Judge for AI Coding Agents") - Repo:
azank1/loop-llm→azank1/cdv(old URLs redirect) cdv migrate-legacystill imports your old~/.loopllm/store.db— learned priors and episodes carry over
What CDV is
Your agent said it works. CDV is the receipt. Every step an agent claims is judged through two channels — a deterministic floor plus an LLM critic, stricter score wins — with Bayesian stop/continue logic, a git-stamped audit trail, and a per-project memory that learns how your tasks actually converge. Local-first, model-agnostic, no API key.
309 tests, strict mypy, ruff — all green. Full details: CHANGELOG.md
v0.11.0 — CI audit-gate + install-mcp --rules
Added
- CI verification gate:
loopllm audit --export <path>writes the audit trail as a portable JSON artifact you commit alongside the code it verifies;loopllm audit-gate --since <ref> [--min-score X] [--require-verified]checks every commit in a PR against it (report-only by default). - Reusable composite action
.github/actions/loopllm-gate+ workflow wiring the gate into this repo's own PRs. loopllm install-mcp --rulesdrops project-scoped agent-instruction files (Cursor.mdc, VS Code instructions,CLAUDE.md) so the IDE agent actually consults loopllm instead of merely having the tools registered.
Fixed
- structlog was never configured and could write to stdout, corrupting the MCP stdio JSON-RPC stream; new
logging_config.configure_logging()routes all output to stderr with level filtering.
Full details: CHANGELOG.md
v0.7.0 — Conservative Dual-Verify for agent loops
Conservative Dual-Verify (CDV)
Agents submit step_output artifacts; the MCP server scores each step through two independent channels:
- Channel A: deterministic evaluators (regex, JSON, completeness, composite)
- Channel B: separate critic via MCP sampling (verifier hat)
- Final score:
min(channel_a, channel_b)— the stricter channel wins
Added
step_scorer.py,guards.py,evaluator_factory.py- Composable guard stack: timeout, token budget, output-repeat, plateau, threshold, Bayesian ROI
- Async artifact-primary
loopllm_loop_step; legacyscorefallback with deprecation warning - Verdict JSON exposes
channel_a_score,channel_b_score,score_source,deficiencies
Changed
AgentLoopControllerusesGuardStack- README restructured around three-layer architecture
- Internal launch docs pruned;
docs/launch/show-hn.mdupdated for CDV
Full changelog: CHANGELOG.md