Fix open issues: security, routing, and reliability (#17-#27)

[OnlyTerp]

Summary

Resolves the open issues in the tracker. Each is a focused, stdlib-only fix; no new dependencies. The offline self-test, compileall, doctor --no-test --ci, auto_router_demo, and the new ASCII guard all pass locally, and test_proxy.py now exercises every fix below.

#14 (long-context openai_compat hygiene: tool-only content=null + context-window hint) is already resolved in main and is covered by the existing openai_compat long-context hygiene test — no code change needed here.

Security

• #19 config leak / unauthenticated routing mutation — /healthz exposed the full backend config and /uc/select let any local web page flip routing. When bound to loopback (the default), every endpoint now requires a loopback Host header (_guard_local() → 403 otherwise). A browser can't forge Host, so this blocks DNS-rebinding. Opt out by binding a non-loopback UC_LISTEN_HOST.

• #22 credential leak / SSRF to custom upstreams — Claude Code's own inbound Anthropic credential was forwarded to any route upstream. Now inbound creds (authorization, x-api-key) are stripped before dispatch to a non-default upstream unless the route explicitly opts in with auth: "passthrough".

  • Found and fixed a real bug while writing the test: _routes_to_slots dropped the "passthrough" intent entirely (if auth and auth != "passthrough"), so the opt-in could never re-enable forwarding. The intent is now preserved as slot["auth_passthrough"] and honored at dispatch:

    # _routes_to_slots
    if auth == "passthrough": slot["auth_passthrough"] = True
    elif auth:               slot["auth"] = _expand_env(auth)

• #23 prompt injection in cursor_agent — the <CLAUDE_TOOL_CALL>{...}</CLAUDE_TOOL_CALL> bridge marker is our private control channel, but it was parsed out of cursor-agent output after untrusted transcript content (user text, tool results) could smuggle one in. _flatten_messages now defangs the tag in all rendered transcript content, so only our own trailing instructions can emit a real call.

• #25 settings TOCTOU / world-readable ANTHROPIC_BASE_URL — launchers now create the state dir private (chmod 700 / icacls lock to the current user) and write session settings via a umask 077 temp file + chmod 600 + atomic mv, so another local account can't read or symlink-race the ANTHROPIC_BASE_URL.

• #21 supply-chain integrity — CI actions are pinned to commit SHAs (actions/checkout v4.3.1, actions/setup-python v5.6.0) and a new scripts/check_ascii_ps1.py CI step fails the build on non-ASCII bytes in any .ps1 (regression guard for #17).

• #27 unverified JWT — clarified that the Codex JWT decode is best-effort/unverified and used only for a cosmetic account-id/exp routing hint (never an auth decision; authority is enforced upstream). Also switched REFRESH_CMD.split() → shlex.split(..., posix=os.name != "nt") so a refresh command with a quoted path isn't split on internal spaces.

Reliability / correctness

• #24 resource exhaustion — added a request body cap (UC_MAX_BODY_BYTES, default 64 MiB) that returns 413 before allocating, a BoundedSemaphore(UC_MAX_CONNECTIONS) so a connection flood applies back-pressure instead of spawning unbounded threads, and a per-socket timeout (UC_SOCKET_TIMEOUT) to bound idle/slowloris connections.

• #20 non-deterministic router cache — the cache key used Python's builtin hash(), which is salted per process (PYTHONHASHSEED), so the same task minted a different key every restart and silently re-ran the classifier. Replaced with a stable, tier-scoped SHA256 digest: "%s|%s" % (tier, sha256(task)[:32]).

• #17 PowerShell 5.1 parse failure — Start-UltraCode.ps1 had an em-dash (U+2014) on line 233; PS 5.1 decodes a BOM-less .ps1 as CP1252, where 0x94 is a closing quote and broke string parsing. Replaced with ASCII --.

• #18 workflows fall back to stock Claude — the orchestrator/worker pick lived only in memory, so a proxy restart (or a workflow spawning a fresh proxy) forgot it and sub-agents reverted to stock Claude. The pick now persists to UC_SELECTION_CACHE (set by both launchers) and is restored on startup; a one-time diagnostic surfaces when selection is on but empty so a lost pick isn't mistaken for the workflow ignoring the model.

Tests (#26)

test_proxy.py previously covered only the mock-based core. Added offline coverage for the loopback guard (403 on non-local Host), the oversize-body 413, credential stripping vs. auth:passthrough opt-in, deterministic router cache keys, selection persist/restore across a simulated restart, the cursor_agent injection defang, and best-effort JWT decoding. The provider modules' built-in self-tests gained an injection-guard assertion as well.

Link to Devin session: https://app.devin.ai/sessions/b788d729b84747d094cdff871cb52d4b
Requested by: @OnlyTerp

---

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.