…-agent#200)

Phase-1 follow-up to open-multi-agent#205. Owner explicitly carved this out as separate
work in their RFC open-multi-agent#200 reply: "OpenAI reasoning_effort is worth adding,
but keep it separate from the phase-1 protocol fix."

The IR field `AgentConfig.thinking.effort` was added in open-multi-agent#205 but only
wired up for Anthropic / Gemini. This PR adds the OpenAI side.

Changes:
- openai.ts chat() and stream(): one-line `reasoning_effort: options.thinking?.effort`
  added before `...options.extraBody` so extraBody can override (consistent
  with how the other sampling params are ordered).
- types.ts AgentConfig.thinking JSDoc: refreshed to document per-provider
  mapping now that OpenAI is wired up.

Design notes:
- The `'minimal'` effort value (gpt-5) isn't yet in the SDK 4.104 type
  union (`'low' | 'medium' | 'high' | null`), but the existing `as
  ChatCompletionCreateParamsNonStreaming` cast at the bottom of the params
  object already covers this, consistent with how `top_k` / `min_p` are
  forwarded. Test verifies pass-through.
- `thinking.enabled` and `thinking.budgetTokens` are intentionally ignored
  by the OpenAI adapter — OpenAI's reasoning surface is qualitative
  (`effort`) only, and the explicit-effort design was the owner's
  open-question resolution in open-multi-agent#200.
- Scope deliberately tight: openai.ts only. Other OpenAI-compatible
  adapters (copilot, azure-openai, deepseek, grok, minimax, qiniu) are
  unchanged — they each route to providers with their own reasoning
  semantics, so a separate PR per adapter is the right unit.

+6 tests covering: chat forwarding, stream forwarding, gpt-5 'minimal'
pass-through, omission when thinking absent, omission when effort unset
(with budgetTokens set, to verify enabled/budgetTokens are correctly
ignored), extraBody override of effort.

770/770 passing.

…ch openai

The Copilot and Azure-OpenAI adapters were quietly dropping seven
AgentConfig fields that openai.ts has always forwarded:

  frequency_penalty, presence_penalty, top_p, top_k, min_p,
  parallel_tool_calls, extraBody (spread)

A user setting `topP: 0.9` or `extraBody: { logit_bias: {...} }` on a
Copilot/Azure agent would see no effect on the wire, with no warning —
the AgentConfig accepted the field, the adapter silently ignored it.

Mirror the openai.ts param object exactly: same field set, same ordering
(sampling first, ...extraBody, then structural fields so extraBody can
override sampling defaults but cannot clobber model/messages/tools/stream).

This was discovered while tracing the reasoning_effort change — it's not
strictly part of open-multi-agent#200, but the Copilot/Azure adapters were already in
this PR's surface area for the reasoning_effort fix and the silent-drop
behaviour is the same bug class as the budget-vs-maxTokens issue you
caught on open-multi-agent#205. Folding the parity fix in while the diff is open.

+6 tests (3 per adapter):
- forwards full sampling-param set + extraBody
- extraBody overrides sampling params (field-ordering contract)
- extraBody cannot override structural fields (model/stream)

784/784 passing.

…ted by these endpoints

Mid-PR correction. The previous "mirror openai.ts exactly" reasoning was
sloppy: openai.ts forwards `top_k` and `min_p` deliberately for the
local-server case (`provider: 'openai', baseURL: 'http://localhost:8000'`
pointing at vLLM), but Copilot and Azure OpenAI both have hard-coded cloud
endpoints with no equivalent escape hatch.

Verified against real docs:
- Azure OpenAI Chat Completions reference lists `frequency_penalty` /
  `presence_penalty` / `top_p` / `parallel_tool_calls` / `logit_bias`,
  but explicitly does NOT include `top_k` or `min_p`.
- vLLM docs confirm `top_k` / `min_p` are vLLM-only sampling extensions
  not part of the OpenAI Chat Completions spec.
- Copilot uses `api.githubcopilot.com` — also a fixed cloud proxy that
  normalises to standard OpenAI Chat Completions.

So forwarding these as top-level fields on Copilot / Azure would be dead
weight at best and a 400 at worst. Users who genuinely need them on
either adapter can still pass via `extraBody` (e.g. for a local proxy
that re-routes to vLLM).

Removed from copilot.ts and azure-openai.ts param objects in both chat()
and stream(); updated cast comments to drop the `top_k` / `min_p`
mention; replaced the "forwards full set" tests with two narrower tests
each: (a) forwards the OpenAI-cloud-compatible subset, (b) does NOT
forward `top_k` / `min_p` (regression guard).

786/786 passing.

…ous")

After @MyPrototypeWhat's audit ("是否必须的?"), trimmed three categories:

1. Copilot chat() narrow-surface comment: 9 lines → 3 lines.
   The explicit list of forwarded/not-forwarded fields was redundant with
   the source code immediately below it. Kept the load-bearing WHY
   ("Copilot's API isn't publicly documented").

2. Three test rationale comments removed entirely (4-6 lines each).
   The test names already convey both WHAT and WHY:
   - "forwards the OpenAI-cloud-compatible sampling params and extraBody"
   - "does NOT forward vLLM-only top_k / min_p (Azure runs MS-hosted OpenAI models)"
   - "does NOT forward sampling params or extraBody that lack public Copilot spec coverage"
   The verification process and bug history live in PR open-multi-agent#209's description
   and the per-commit messages, not in long-lived test source.

Net -21 lines. 784/784 still passing. Lint clean.

…JackChen-me)

Per owner's call on open-multi-agent#209: option 2. Narrow `ThinkingConfig.effort` to the
SDK 4.104-declared values (`'low' | 'medium' | 'high'`) and document
that newer API values (gpt-5 `'minimal'`, GPT-5.5 `'none'`) travel via
`extraBody: { reasoning_effort: '<value>' }` — the same escape hatch
already used for vLLM-only `top_k` / `min_p`.

Cleanup:
- src/types.ts: ThinkingConfig.effort union narrowed; JSDoc points users
  to extraBody for newer SDK-untyped values.
- src/llm/openai.ts: object-level cast comment trimmed (no longer needs
  to mention 'minimal'; cast still required for top_k/min_p/extraBody).
- src/llm/copilot.ts + src/llm/azure-openai.ts: field-level cast on
  `reasoning_effort` removed entirely. The IR type now satisfies the SDK
  type union, so no boundary cast is needed.
- tests/openai-adapter.test.ts + tests/azure-openai-adapter.test.ts:
  rewrote the 'minimal' tests to verify the extraBody route (the now-
  supported path) instead of the old IR-direct route.
- tests/copilot-adapter.test.ts: deleted the 'minimal' test outright.
  Copilot doesn't forward extraBody (narrow surface, no public spec),
  so neither route exists for Copilot users wanting 'minimal'. They
  use 'low' instead, or wait for SDK bump.

783/783 passing. Lint clean. Net -7 lines.

No behaviour change for users on `'low' | 'medium' | 'high'` (the entire
SDK-typed surface).