[jeffdupont]

What

• Dependency audit in CI: new make audit target that runs pip-audit (pinned 2.9.0) over uv export --frozen of the committed uv.lock, and a Dependency audit CI job that runs it. It audits the lock as committed (no uv lock first), so a vulnerable pin fails the job.
• Dependabot: .github/dependabot.yml with uv daily at the workspace root and github-actions weekly, both with a 7-day cooldown — same shape as go-ai-sdk (go-ai-sdk#3).
• Lockfile upgrades so the new job passes. No pyproject.toml ranges changed:

Package	Was	Now
anyio	4.14.1	4.15.1
langgraph-sdk	0.4.2	0.4.7
mcp	1.28.1	1.30.0
pyjwt	2.13.0	2.15.1
urllib3	2.7.0	2.8.0
virtualenv	21.5.1	21.14.7

These cover all 25 open Dependabot alerts on main (2 critical, 12 high, 11 moderate), plus one mcp advisory (GHSA-84m7-p3x7-pcfv) that pip-audit reports and GitHub doesn't.

Verified locally

• make audit → No known vulnerabilities found
• Control: pip-audit on pyjwt==2.13.0 exits 1 with 14 advisories, so the job fails on a vulnerable pin
• uv lock --check, ruff check, ruff format --check, mypy packages/*/src clean; pytest 2536 passed, 14 skipped

Not verified: the CI run itself (this PR's checks will show it).

🤖 Generated with Claude Code

---

Note

Overview
Adds automated dependency security: a make audit target runs pinned pip-audit over a frozen uv export of the committed lockfile, and CI runs it in a new Dependency audit job (non-blocking on PRs, failing on push/schedule). CI also gets a weekly Monday cron so advisories against an unchanged uv.lock still surface.

Introduces Dependabot for daily uv lock updates at the repo root and weekly GitHub Actions bumps (7-day cooldown on both).

Lockfile refresh (no pyproject range edits) bumps packages including claude-agent-sdk (0.2.121 → 0.2.173), OpenTelemetry 1.43 → 1.45.x, brotlicffi, and virtualenv, so audit passes.

Follow-on code/test tweaks: Claude Agents handler documents and excludes new ClaudeAgentOptions keys from config forwarding (verbatim_prompts, session resume fields, forward_subagent_text) to stay aligned with the upgraded SDK. The lifecycle test for failed telemetry setup now patches _setup_telemetry to raise instead of using a malformed OTEL_EXPORTER_OTLP_TIMEOUT, because OTel 1.45 no longer fails on that input.

^{Reviewed by Cursor Bugbot for commit ce4c397. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit bcc3de1. Configure here.}

[jeffdupont]

Pushed three commits:

• a52a220 merges origin/main. uv.lock was taken from main and regenerated, not hand-merged. The Makefile .PHONY line keeps both api-report targets and audit.
• 2210f7b runs uv lock -P for brotlicffi, claude-agent-sdk, virtualenv and the OpenTelemetry packages. make audit now reports no known vulnerabilities. Two fixes the upgrade needed:
  • claude-agent-sdk 0.2.173 adds four ClaudeAgentOptions fields. All four go on the excluded list: resume_session_at and resume_drops_turn (session state), forward_subagent_text (stream shape), verbatim_prompts (whether prompt text can make Claude Code read host files).
  • OpenTelemetry 1.45 no longer raises on a malformed OTEL_EXPORTER_OTLP_TIMEOUT, so test_a_failed_telemetry_setup_leaves_no_client_behind now injects the failure itself.
• ce4c397 makes the audit job continue-on-error on pull_request only, and adds a weekly schedule (Mondays 13:00 UTC).

Needs a human reviewer.

🤖 Generated with Claude Code