request-logger: no Claude Code + Vertex AI support (wizard prints ANTHROPIC_BASE_URL, which Vertex mode ignores)

[mattpocock]

A student joining the course uses Vertex AI (via their company) to access Anthropic models, not the direct Anthropic API. Their ~/.zshrc has:

export CLAUDE_CODE_USE_VERTEX=1
export CLOUD_ML_REGION=global
export ANTHROPIC_VERTEX_PROJECT_ID=vertex-model-runners

request-logger's wizard, run for Claude Code, printed:

ANTHROPIC_BASE_URL=http://localhost:8787/ ENABLE_TOOL_SEARCH=true claude

Running that produces an empty request-logger/logs/ folder — no requests are captured.

Reported in the course Slack (#ai-coding-crash-course):

I just joined AI coding crash course and have trouble configure the request-logger. my company use Vertex AI to access Anthropic models. after request-logger came up w/ "ANTHROPIC_BASE_URL=http://localhost:8787/ ENABLE_TOOL_SEARCH=true claude" but it doesn't log requests.

The catalogue in request-logger/agents.ts has exactly one provider for Claude Code ("Anthropic", api.anthropic.com, sets ANTHROPIC_BASE_URL). There is no Vertex AI provider entry, so this is an unsupported setup today, not just a misconfiguration.

[mattpocock]

Triage: root cause found, and it's not a request-logger bug so much as a real gap — Claude Code on Vertex AI doesn't read ANTHROPIC_BASE_URL at all.

Per Anthropic's own docs (Claude Code on Google Vertex AI), once CLAUDE_CODE_USE_VERTEX=1 is set, Claude Code switches to a completely different backend: it authenticates via Google Cloud (ADC/gcloud) and talks to a *-aiplatform.googleapis.com host, resolved from CLOUD_ML_REGION + ANTHROPIC_VERTEX_PROJECT_ID. ANTHROPIC_BASE_URL is a variable from the direct Anthropic API code path and is silently ignored in Vertex mode — that's why the wizard's printed command produces an empty logs folder, same failure class as items 1–5 in the README's "If your logs folder stays empty" section, just not one we've enumerated yet.

The correct variable in Vertex mode is a different one, documented on the same page:

Optional: Override the Agent Platform endpoint URL for custom endpoints or gateways
export ANTHROPIC_VERTEX_BASE_URL=https://aiplatform.googleapis.com

Workaround to try right now (swap ANTHROPIC_BASE_URL for ANTHROPIC_VERTEX_BASE_URL, keep your existing Vertex vars):

CLAUDE_CODE_USE_VERTEX=1 \
CLOUD_ML_REGION=global \
ANTHROPIC_VERTEX_PROJECT_ID=vertex-model-runners \
ANTHROPIC_VERTEX_BASE_URL=http://localhost:8787/ \
ENABLE_TOOL_SEARCH=true \
claude

(Port must match whatever request-logger printed — 8787 above, adjust if yours differs.)

Two things worth flagging for whoever picks up the permanent fix, neither of which blocks the workaround:

1. Wire shape differs slightly. Vertex's Claude endpoint is .../publishers/anthropic/models/{model}:streamRawPredict, which drops model from the body (it's in the URL) and adds anthropic_version. The anthropic renderer in render.ts may not render this perfectly — worth checking once we have a real capture.
2. Token-counting filter won't match. shouldLogRequest in agents.ts filters housekeeping calls by checking reqPath.includes("count_tokens"). Vertex's equivalent path uses :countTokens (camelCase, no underscore), so it won't be filtered — the logs folder will include housekeeping captures on this route until that check is updated.

Reporter (via Slack, not the GitHub issue author) — can you confirm the workaround above actually produces captures in request-logger/logs/? If it does, this becomes a real "add a Vertex AI provider entry" enhancement (upstream host per region/global, ANTHROPIC_VERTEX_BASE_URL env var, the two notes above) rather than a doc fix, and we can scope it properly.

[mattpocock]

Opened #35 with a first-cut Vertex AI provider entry — adds ANTHROPIC_VERTEX_BASE_URL support, fixes the token-counting filter for Vertex's :countTokens path, and covers it with tests. Still marked as doc-verified rather than tested end to end (see the PR description) — would appreciate confirmation from the reporter that running the printed command against a real Vertex project actually produces captures in request-logger/logs/ before this merges.