Add token passthrough feature

[bettercallsaulj]

No description provided.

[dIvYaNshhh]

The ctypes marshalling in here is careful work — _AgentOptionsStorage gets the buffer-lifetime problem right in a way that's easy to get wrong, _bind_optional_agent_options_symbol is the per-symbol try/except that #10's big shared block should have been, and _missing_options_symbol_message() is exactly the actionable error I was asking for on that PR. Good stuff.

Three things I think block merge:

1. The 12.9 MB of committed binaries. create_by_url_gateway (8.76 MB) and create_by_url_server (4.14 MB) are checked in as executables. I pulled the magic bytes — they're Mach-O 64-bit x86_64, so they won't run on Linux CI and need Rosetta on Apple Silicon, which makes create_by_url_run.sh unrunnable for most of the team. There's also no source or build recipe in the repo, so there's no way to review what they actually do — that's a supply-chain question I can't sign off on.

Could these be built from third_party/gopher-orch in the script, or pulled as release assets? Once they're in history they're there permanently.

2. The native-library search reorder (inline below). It ships to every PyPI user, not just developers, and turns "process happens to run from a writable directory" into arbitrary native code execution.

3. Empty access_token produces Authorization: Bearer , and the shipped example hits that path by default.

Plus a question I couldn't answer from the diff: whether the native side deep-copies the options struct at creation or retains the pointer. If it retains it, the storage is freed too early. Details inline.

Smaller stuff: Optional[object] defeats mypy on every new signature, runtime options get normalised at two different layers depending on which factory you use, and ffi/library.py now imports the public config layer. All noted inline.

Base is iml_new_apis, so this can't land until #10 does — and it shouldn't land alone regardless, because of the --remote submodule change (#13 reverts it).

[dIvYaNshhh]

Re-reviewed at a46cfc58. Eight of nine addressed, and a few went past what I asked. Resolving those threads.

Confirmed fixed:

• Search order (2c21fa03) — installed platform package back at position 1, cwd dev paths after it, and the comment now points at GOPHER_MCP_PYTHON_LIBRARY_PATH. Hijack vector closed.

• Empty access_token (fc5ad5b0) — fixed in both places: access_token == "" normalises to None, and _normalize_headers gates on if access_token:. Covered by a test.

• Options struct lifetime (229ad3ff) — you went and got the answer to the question I couldn't resolve from the diff, and put it where the next reader will find it:

  Native BuildAgentOptions deep-copies this struct into C++ strings/maps during creation, so this storage only needs call-duration lifetime.

  That, plus the comment explaining why _bytes is pinned, is exactly right. Thanks for chasing the C side rather than just asserting it.

• Normalisation at two layers (c012d442) — done once at the agent boundary now, so both factory paths fail at the same point.

• Layering (28764f27) — the new gopher_mcp_python/runtime_options.py actually breaks the ffi → config dependency instead of papering over it.

• Typing — RuntimeOptionsInput replaces Optional[object]; zero occurrences left in agent.py, so mypy can see through it.

• Submodule pinning (6f88270e) — GOPHER_ORCH_TRACK_REMOTE=1 opt-in, pinned by default.

• Temp dir (c75d5e8c).

Suite at the tip of the stack: 151 passed, 116 skipped, 3 failed — all three failures reproduce identically on origin/main (they need ./build.sh), so nothing regressed.

One item left, and it's the one I'd hold merge on: the 12.9 MB of committed binaries — no commit touches them. Detail inline.

Separately: the description is still empty. With a new public module, a new FFI struct, and a build-script behaviour change in here, a few lines of scope would help the next reviewer.

[dIvYaNshhh]

de69016c closes the last item — resolving the thread. Nothing outstanding from me on this PR now.

You took option 1 and did it properly:

• Both binaries deleted, and create_by_url_run.sh now builds header_access_token_create_by_url_{server,gateway} from third_party/gopher-orch via cmake --build.
• build_header_binaries() only runs when the binaries are absent, so repeat runs don't pay for a rebuild.
• It fails with an actionable message when the submodule isn't initialized, and auto-inits the nested gopher-mcp when that's what's missing.
• Core count falls back sysctl -n hw.ncpu -> nproc -> 4, so it works on macOS and Linux.

Net effect: the script now runs on Linux and Apple Silicon, where before it couldn't run at all.

Suite at the tip of the stack: 153 passed, 116 skipped, 3 failed — the three failures reproduce identically on origin/main (they need ./build.sh), so nothing regressed.

---

One thing for whoever clicks merge — not a code change. Deleting the files fixes the tree at HEAD, but the blobs are still reachable in this branch's history:

$ git rev-list --objects feature_token_protect --not origin/main | ...
  4.0 MB  examples/header/create_by_url_server
  8.4 MB  examples/header/create_by_url_gateway

added in e0c4bc47, removed in de69016c. The repo currently has allow_merge_commit, allow_squash_merge and allow_rebase_merge all enabled — and with a merge commit or rebase-merge those 12.4 MB land in main's history permanently, even though no commit on main would have them in its tree.

Squash-merging this PR avoids it entirely (the squash commit carries only the final tree). Worth agreeing before merge — cheap now, needs a history rewrite later. Same applies to #13 and #14, which sit on this branch.