fix(backends): send the CLI/API contracts these backends actually have

[Yifan Yang (Yif-Yang)]

Follow-up to #170 and #190. Both PRs moved in the right direction, and their test scaffolding is kept — but each shipped a parameter its target does not accept, so as merged neither backend does what it says. This corrects both against the real contracts.

Copilot — skillopt_sleep/backend.py

--allowed-tools is not a GitHub Copilot CLI option:

$ copilot --allowed-tools=Bash -p hi
error: unknown option '--allowed-tools=Bash'

Both Copilot call paths (_call and the task-replay path) sent it, so both aborted before doing any work.

Two independent axes were also conflated:

• --allow-all-tools waives the interactive approval prompt. The CLI's own help says it is "required for non-interactive mode" — dropping it breaks headless runs. This is a permission flag, not a scoping flag.
• --available-tools is the visibility axis: "Only these tools will be available to the model." This is where scoping belongs.

So: keep --allow-all-tools, scope with --available-tools (default bash, override via COPILOT_AVAILABLE_TOOLS).

The selector is case-sensitive, which is easy to get wrong. Verified live against Copilot CLI 1.0.80:

argv	result
--available-tools bash	"toolName":"bash" — tool runs
--available-tools Bash	tool never appears — silently blocked
--available-tools write	bash blocked — scoping demonstrably works

That last row is the control: the flag really does narrow the surface, rather than being accepted and ignored.

MiniMax — skillopt/model/minimax_backend.py

chat_template_kwargs.enable_thinking is a Qwen/HuggingFace-serving convention. It appears nowhere in MiniMax's OpenAI-compatible reference, so the endpoint ignores it and thinking stays at the server default no matter what the flag says.

This predates #190 — the field was already on main; #190 built model-aware logic on top of it, so the routing worked but the value never reached the model.

The documented control is the top-level thinking object:

{"thinking": {"type": "adaptive"}}
{"thinking": {"type": "disabled"}}

Also per the docs, the M2.x family accepts {"type": "disabled"} but keeps thinking on regardless. Sending disabled there would record a request that misrepresents what the model actually did, so M2.x is sent adaptive. This preserves #190's intent (M2.7 keeps always-on thinking) without a value that lies. Unknown deployments default to adaptive, matching the documented API default of thinking-on-when-omitted.

Reference: https://platform.minimax.io/docs/api-reference/text-openai-api

Tests

Both suites now assert the exact wire payload / argv rather than the shape the code happens to produce, plus explicit regression guards (test_nonexistent_allowed_tools_flag_is_never_sent, test_unsupported_chat_template_kwargs_is_never_sent).

• Full suite: 1389 passed, 10 skipped, 269 subtests
• Negative control: reverting each fix fails exactly the 9 tests that assert it, and nothing else — the tests exercise the new behavior rather than passing incidentally.