LumabyteCo/clarifyprompt-mcp

Summary

Four targeted additions, one per engine pillar (memory / agentic / models / context). 3 new MCP tools (20 → 23). 6 new eval fixtures (23 → 29). Fully back-compat with 1.5.x.

Memory — explicit fact CRUD

Three new tools complement the existing reflection-on-save_outcome path:

• memory_remember(subject, predicate, object, scope?, confidence?) — explicitly insert a fact. Source tagged user:explicit, auto-embedded for semantic retrieval.
• memory_forget(id) — bi-temporal soft-delete by id. Idempotent.
• memory_list_facts(scope?, predicate?, limit?) — list live facts.

Closes the obvious UX gap where the engine could only learn from outcomes — users can now say "remember I prefer X" directly.

Agentic — compose_prompt revise loop

New max_iterations field (default 1, hard max 5). With auto_revise: true AND post_critique: true, each iteration's improvedPrompt feeds back through optimize+critique until verdict=accept, no improvedPrompt is available, or the cap is reached.

{ "prompt": "...", "post_critique": true, "auto_revise": true, "max_iterations": 3 }

pre_clarify only fires once. Response includes new iterations field.

Models — per-stage routing in compose_prompt

Each stage can target a different model:

{
  "prompt": "...",
  "clarify_model": "qwen2.5-coder:7b-instruct-q4_K_M",
  "optimize_model": "claude-sonnet-4-20250514",
  "critique_model": "gpt-4o-mini"
}

optimization.metadata.model + critique.judgeModel in the response reflect the actual model that ran each stage.

Context — git-state + environment signals

Two new ContextBundle fields:

• bundle.git — branch, short SHA, dirty flag, last 5 commit titles. Fails soft when cwd isn't a repo.
• bundle.environment — nowIso / weekday / timezone. Pure JS, never fails.

Both feed the Context Curator as low-utility candidates — they won't dominate budget but surface as grounding sources when relevant.

Eval coverage

29 fixtures total (was 23). Local baseline on qwen2.5-coder:7b: 25 passed / 1 failed / 3 skipped / 97% avg. The lone failure remains the persistent analyzer-creative-media model-class signal.

CI eval gate against gpt-4o-mini will fire automatically on this commit.

Compatibility

• No env-var surface changes. All new fields are additive.
• 23 MCP tools, 1 resource. No tools removed; no result-shape changes.
• Existing memory.db files are fully forward-compat (1.5.2's dimension fix still applies).

Install

npx clarifyprompt-mcp

Full diff: https://github.com/LumabyteCo/clarifyprompt-mcp/compare/v1.5.2...v1.6.0