Skip to content

todo.md

Latest commit

 

History

History
13 lines (12 loc) · 2.77 KB

todo.md

File metadata and controls

13 lines (12 loc) · 2.77 KB
Raw

Status - PLANNED (NOT IMPLEMENTED)

  • Test Rule 3 (Citation-Strength Gating, ide_agent_rules.md) on VS Code/Copilot, pi.dev, and Claude Desktop — currently only confirmed working in Antigravity (Gemini 3.1 Pro).
  • Investigate whether the embedding-docstring skill's 4-line/400-char prose body limit needs a complexity exception (e.g., orchestration/coordinator functions like run_council, run_chairman) without becoming project-specific — must stay generic to the skill (no hardcoded function names or thresholds tied to one codebase). Triggered by 3 size-limit violations found auditing a non-zerikai_memory project where trimming may force real information loss rather than just tightening prose.
  • Run the "stale-but-confident document" test case from memory-synthesis-reliability-spec.md Section 6 — closest analog to the original import-extraction incident, not yet tested (existing tests only cover weak/missing citations, not a wrong-but-well-cited source).
  • Feature: Add show_raw=True parameter to query_memory — when enabled, surface the verbatim stored docstring from ChromaDB alongside the LLM synthesis, allowing the caller to compare ground-truth entity description vs. synthesized interpretation. Addresses the gap where synthesis paraphrases the docstring rather than returning it verbatim (observed in _build_system_message main.py:828). Relevant to trust/auditability and the reviewer's source-of-truth concern. Add under MCP Tools Reference in README_FULL.md once implemented.
  • Track whether absence-framed queries ("does X extract imports", "is X missing") are more prone to synthesis fabrication than presence-framed queries ("how does X work"), per memory-synthesis-reliability-spec Section 7 Q2. Current sample too small to confirm (one fabrication observed: import extraction). Revisit once more real-world Rule 3 cases accumulate across agents/sessions.
  • Add fetch_cap to the config.py and the .env file, and use it in the codebase to limit the number of documents fetched from ChromaDB for reranking.
  • Diff preview in scan_status ("3 files changed since last scan") and a simple collection backup before scan are genuinely good ideas and easy to implement.
  • Persist last_scanned_at timestamp per workspace in SQLite registry; expose via scan_status output
  • Test temperature 0 + explicit "return no answer if uncertain" prompt vs current behavior; measure hallucination rate on poorly documented codebases
  • Full hybrid search, BM25 index + reciprocal rank fusion to replace keyword-overlap reranking.
  • Constant Upgrade: tree-sitter indexes UPPER_CASE constants (Python/JS/TS) with preceding comments as pseudo-docstrings, and markdown checkboxes as searchable entities — config files and todo lists now visible to query_memory.