From the Inside Claude Code - The Claude Code Ecosystem architecture map
The triage and dedupe agents are boxed in identity, tools and network
strength · medium · verified
Both Claude jobs authenticate with a short-lived token from workload identity federation instead of a stored key, hold only issues: write, run in auto permission mode with Write, Edit, WebFetch and WebSearch disallowed, call GitHub only through the wrapper scripts, and run on a runner whose egress is limited to an explicit host list.
Impact: A compromised run can at most mislabel or comment on the one issue that triggered it.
This is a strong template for any further agent in this repo: per-job token, script-mediated writes, explicit egress list and per-script call caps.
The trail
Open on the mapMore from this board
- Optional SessionStore methods raise instead of being absent, so a change in Types reaches 13 modules
- The Lowerer is one struct spread over 38 files and 29,700 lines with no unit tests
- Day 1: start with marketplace.json, the contract everything else serves
- Every SDK module imports its exceptions from one file, _errors.py
- Partner MCP plugins launch unpinned packages on every user's machine
- PR-controlled agent config is restored from the base branch before Claude starts
- Control-protocol failures escape the SDK exception hierarchy
- Both public entry points share one control-protocol engine
- Label decisions by the triage agent leave no reason a maintainer can see
- Spend and run length are bounded at the transport
- The subprocess transport carries five responsibilities in one file
- Command-line construction is hardened against argument injection
- Every agent-driven closure has a human override before it happens
- Skip for now: config-only integrations, CI plumbing and single-file plugins
- External-PR workflows never execute contributor code
- Channel servers keep bot tokens owner-only and refuse injected pairing requests