An agent that plans coding tasks, forks candidate implementations, tests each in isolation, and keeps the winner — built on a vendored pydantic-deepagents harness, fronted by a CopilotKit React UI speaking the AG-UI protocol to a FastAPI backend. Everything traced end-to-end with Logfire.
Start here: PDR.md (architecture → files) and
docs/adr/ (16 decision records, 0001–0016).
| Path | What |
|---|---|
vendor/pydantic-deepagents/ |
pristine upstream + vendor/patches/ (one patch: public branch_outcomes()) |
projects/agent-web/ |
the web service: FastAPI + AGUIAdapter + MCP registry + CopilotKit frontend |
projects/eval-optimizer/ |
C5 fork-based plan-viability pipeline (headless) |
docs/adr/, docs/PLAN-agent-web-startup.md |
decisions & the startup/always-on plan |
HANDOFF.md |
what was verified where (sandbox vs Windows-native) |
Obsolete/ |
parked-instead-of-deleted files (never agent-deleted; git-ignored) |
cd projects\agent-web
uv sync # harness installs FROM vendor\, never PyPI
powershell -File scripts\build-frontend.ps1 # once, and after UI changes
powershell -File scripts\Start-AgentWeb.ps1 # one server: http://localhost:8801Secrets are Windows USER environment variables (see .env.example files);
.env holds only non-secret flags. Auto-start: scripts\Register-StartupTask.ps1.
- Every feature is env-gated (
TEAMS,EXECUTE,TOOL_SEARCH, …); withTOOL_SEARCH=1the model must callsearch_toolsto see MCP tools — theskills/external-servicesskill teaches it to. execute(shell) is approval-gated through AG-UI interrupts; browser automation is not — leaveBROWSER_AUTOMATION=0unless a task needs it.- Telemetry is the impartial witness: every claim in the ADRs' resolution logs is backed by a Logfire trace.