Agents with a write workspace (developer/product_owner/head_marketing/documenter)
run with cwd = their git workspace clone. Claude Code launches each MCP server
(flow/do/git-readonly/optimal/docs/search) and the SDK server as
`uv run python -m ...` from that cwd. When the clone's uv.lock drifts from the
baked image, `uv run` re-resolves and re-syncs /app/.venv against the clone's
lock — a multi-minute stall on a cold wheel cache — so the servers never reach
"connected": they sit at status="pending" and the agent gets ZERO gateway
verbs. It then can't claim/commit/idle (all MCP verbs), its Stop is rejected,
and it respawns in a loop redoing work it can't submit.
UV_PROJECT_ENVIRONMENT pins the venv location but does NOT stop the cwd-relative
resolve/sync (confirmed empirically on uv 0.11.1); `--no-sync` does, so the
servers reuse the baked /app/.venv as-is and start instantly. The /app-cwd roles
(qa/cell_pm/main_pm/auditor) were unaffected because their env already matches.
- orchestrator.py: --no-sync on all 6 generated MCP servers
- docker/scripts/sdk-startup-hook.sh: --no-sync on the agent_sdk.server launch
- test_spawn_strict_mcp.py: assert every server's args start with run,--no-sync
Every agent MCP server is launched as `uv run python -m roboco.mcp.<server>`
(via the orchestrator-generated mcp-config.json) and the SDK server via
`uv run python -m roboco.agent_sdk.server` (sdk-startup-hook.sh) — both with
cwd = the agent's WORKSPACE, not /app. `uv run` then resolves a cwd-relative
`.venv` (≠ the image's baked /app/.venv), ignores VIRTUAL_ENV with a warning,
and RE-SYNCS the full dependency set (torch/lancedb/pyarrow/scipy, ~350MB)
into a fresh venv on every spawn.
A warm host uv wheel cache masks this (fast re-resolve from cached wheels —
earlier runs this session opened PR #26/#28/#29 fine). On a COLD cache (first
spawn after an image rebuild — exactly when deploying new fixes) the download
takes minutes, the MCP servers never register, and the agent burns its whole
budget with "No such tool available: mcp__roboco-*" before reaping. Observed
this session: be-dev-1 never claimed; /tmp/sdk-server.log showed the live
torch/lancedb download + the `VIRTUAL_ENV ... will be ignored` warning.
Fix: set UV_PROJECT_ENVIRONMENT=/app/.venv in (1) every MCP server's env in
the generated mcp-config.json (one place — shared mcp_env dict) and (2) the
SDK startup hook. uv then reuses the pre-baked image venv instantly,
regardless of cwd or cache state. Not a regression from this session's code
(none of #172b/#175/#176/#177/#178 touched the launch/venv path — verified);
a pre-existing launch-cwd fragility that rebuilding to deploy exposed.
Test: _generate_mcp_config asserts every server env pins
UV_PROJECT_ENVIRONMENT=/app/.venv. make quality green.
Smoke run 3 showed agents loading builtin Anthropic connectors
(mcp__claude_ai_Gmail__authenticate, Google Calendar, Notion, Drive)
alongside our 5 roboco MCP servers. The connectors bloat the tool
surface and give the LLM 'discover' targets it shouldn't have.
The Claude Code CLI's --strict-mcp-config flag tells it to load ONLY
the servers from --mcp-config, ignoring all builtin defaults. Added
to _append_image_and_claude_args next to --mcp-config.
Note: the existing --tools allowlist (Read,Write,Edit,Bash,Grep,Glob,
Task,TodoWrite) only filters builtin tools, not MCP-prefixed ones —
that's why the connectors slipped through.
Spec ref: docs/superpowers/specs/2026-05-12-post-smoke-3-fixes-design.md
section E3.
Renn F