b4dbd262 · zzmq.dev@gmail.com/zz
Commit
b4dbd2627c897a5a705be21513efa262bf867246
Implement the zz-git/zz-ladder-lanes tool surface and zz work In-process tools (app/agent/tools.ts) wrapping existing engine functions only -- no new business logic. zz-git mirrors engine/ ingest.ts's fetchers; zz-ladder-lanes exposes ladder_read/ ladder_write_rung, lane_propose/lane_get, and ask_post/ ask_query_claimable/ask_claim/ask_answer (the claim/work/answer steps split apart, since a general agent's "work" spans many tool calls across turns, not one synchronous callback like claimAndAnswer assumes). app/agent/loop.ts is a minimal tool-calling loop against Ollama's /api/chat, confirmed directly to support real OpenAI-style function calling for qwen2.5:14b-instruct before building this -- no BYOK, no Claude Agent SDK. zz work <repo> [--task] posts a "work"-kind Ask (a new, generic AskKind for a prose payload.description with no dedicated handler) and runs the loop with the system prompt and skill from site/content/mcp/index.md as its operational copy. Verified against thoward/bsharptree with a bounded task (one not-yet-onboarded commit, not the whole repo): the agent correctly claimed the ask, called git_diff/git_show_file, and wrote a real changeset+filechange rung on a clean run. Along the way, found and fixed real gaps the first attempts exposed: ladder_write_rung crashed with opaque internal errors on a rung missing children/sources instead of a clear message (now normalizes and validates with actionable errors: missing commit source to anchor to, invalid top-level kind, empty summary); a non-string source field rendered as "[object Object]" instead of being coerced; and the agent answered an Ask "complete" despite ladder_write_rung never having succeeded, so the system prompt now explicitly requires an honest "deny" over a false "complete". A later run surfaced a real remaining limit, not a code bug: the model occasionally doesn't retry a third time after two validation errors -- left as a known reliability gap, not papered over. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Troy Howard <thoward37@gmail.com> -- 2026-10-08T07:36:23.000Z