Metasearch for humans (web UI) and agents (HTTP API, MCP) with a shared TTL cache. v3 is a
Rust rebuild; the authoritative plan is .agents/plans/2026-09-21-v3-rust-core.md and its
per-wave subplans under .agents/plans/v3/. This file is a summary, not a replacement.
legacybranch: v1 (Python, PyPI 0.4.0).v2-legacybranch: v2 (Python FastAPI + React SPA, abandoned 2026-09-21). Reusable as a reference for wire shapes (SearXNG JSON, Exa adapter, MCP tool names) and for the retro in.agents/notes/2026-09-17-velocity-retro.md. Do not lift code from it into v3.main: v3, orphan history started 2026-09-21.
GitHub Project: https://github.com/users/espetro/projects/23/views/1
All work is linked to a refined task there (EPIC per wave, one issue per subplan step) before
implementation starts. A task is refined when it has Iteration, effort (S/M/L/XL), start and
target dates, and a classification label (feature / bug / cosmetic / infra /
documentation). Use ghx to manage it; when ghx isn't installed, fall back to gh project
(e.g. gh project item-list 23 --owner espetro, gh project item-edit) — note that a GitHub
App token (e.g. the devin-ai-integration bot) can read items but cannot add/edit them, so
board mutations that fail with "Resource not accessible by integration" are escalated to the
user.
Full procedure: .agents/plans/v3/README.md. Short form:
- Pick an issue titled
W<n>-<nn> ...whose dependencies areCompleted; move it toWIP. - Read the parent plan sections 4 to 7, the wave file's "Settled inputs", then the step.
- Worktree
~/.worktrees/cauce-<step-id>on branchv3/<step-id>-<slug>frommain. - Implement only the step's "Do"; its "Acceptance" list is the PR's test list; the golden
path stays green; commits signed off (
-s). - PR
Closes #<issue>; on merge, move toCompletedand read the step's "Follow-up".
If a step cannot be done without changing a settled contract, stop and comment on the issue with the proposed amendment to the parent plan. Do not improvise the contract.
crates/cauce-core domain types, pipeline, scheduler, Store + Engine traits (MPL-2.0)
crates/cauce-agent grounded-answer agent loop; provider/tool/observer seam (MPL-2.0)
crates/cauce-store-sqlite rusqlite Store impl (MPL-2.0)
crates/cauce-engines declarative / exec / replay engine runtimes (MPL-2.0)
crates/cauce-server axum, embedded Svelte SPA (`/app`), SSE, MCP, Exa adapter (MPL-2.0)
crates/cauce-cli cauce serve / search / engine test / cache / record (MPL-2.0)
engines/ YAML engine specs + fixtures (Apache-2.0)
sdk/python exec-protocol SDK + ddgs reference engine (Apache-2.0)
tests/e2e golden path integration tests
.agents/ plans/, decisions.md, MEMORY.md, notes/, docs/
One binary: cauce serve (full: UI + API + MCP, < 80 MB idle), cauce serve --headless (API +
MCP, < 50 MB), cauce mcp (stdio only, < 40 MB). Cargo features ui mcp ai archive semantic postgres otlp; defaults are ui mcp ai archive (otlp is non-default). Port 4479, loopback by
default.
Nothing yet. Gates land with wave 0 (.agents/plans/v3/wave-0-skeleton.md) and are tagged
here the same way v2 did: every normative line carries a [gate: <task or path>] marker that a
test resolves.
- Conventional Commits, atomic commits, no AI attribution trailers.
- Plans live in
.agents/plans/; a plan exists on disk before implementation begins. - Agent memory: read
.agents/MEMORY.mdat session start, write dated notes at session end. - Verification policy is the plan's section 7: golden path on the
replayengine first, live canaries gated behindCAUCE_LIVE=1, no fixture-only UI states, every mounted route in the routes table.