fix: use max_completion_tokens for Azure OpenAI

[zerone0x]

Summary

Fixes #461

• send max_completion_tokens instead of max_tokens for Azure OpenAI chat completions
• add a unit test covering Azure request args

Testing

• not run (unit test added)

[evalstate]

This was made with orbit?

[evalstate]

Triage result

Human attention: ⚠️ Required
Human decision needed: ready for human landing decision
Recommendation: 🏁 escalate to a human

Quick read

This PR appears ready for human landing review. It fixes the Azure OpenAI request-shaping bug at the right seam, the bug was locally reproduced and shown fixed, a small test-strengthening follow-up was added, Codex review is clear, CI is green, and the branch is conflict-free against origin/main.

Intent

Make Azure OpenAI requests work by sending the token limit under Azure’s expected field name and add a test so this provider-specific behavior is covered.

Why

fast-agent was sending max_tokens for Azure OpenAI chat completions, but some Azure models require max_completion_tokens instead. That causes 400 unsupported-parameter errors and blocks those Azure deployments.

Codex review

• GitHub Codex review findings: none collected for the current head.
• Local Codex review: successful and usable.
• Local Codex conclusion: no concrete regression or blocking issue identified in the diff.
• Minor follow-through was applied in this lane: the new Azure unit test was tightened so it explicitly exercises the non-reasoning Azure branch that originally emitted max_tokens, rather than being able to pass through a reasoning-model path.

CI/CD

• Validation path: bug reproduction and fix confirmation.
• Local validation established the bug and the fix:
  • pre-change behavior for a non-reasoning Azure model produced max_tokens
  • PR behavior for the same request produced max_completion_tokens
• Active CI state: green/unrelated.
• Reported GitHub Actions check: update_release_draft passing.
• Initial conflict check: clean against origin/main.
• Final conflict check: clean against origin/main.

Recommendation

🏁 Escalate to a human for the final landing decision. The change is small, targeted, validated, reviewed, and conflict-free, so this looks ready for normal human merge judgment.

[evalstate]

Triage result

Human attention: ⚠️ Required
Human decision needed: Please confirm whether this Azure-specific parameter translation is the desired fix and approve the change, or request a design adjustment if Azure/OpenAI handling should be generalized differently.
Recommendation: 🏁 escalate to a human

Quick read

This PR is addressing a real Azure OpenAI compatibility issue: some Azure deployments reject max_tokens and require max_completion_tokens instead.

Intent

Make fast-agent work with Azure OpenAI by sending the token limit under the parameter name Azure expects, and add a test to lock that behavior in.

Why

The issue report describes 400 errors from Azure OpenAI when max_tokens is sent. The patch changes the OpenAI request builder so Azure uses max_completion_tokens while non-Azure behavior stays the same.

Codex review

The code change is right-shaped and includes a unit test that asserts the Azure request payload uses max_completion_tokens and does not include max_tokens.

CI/CD

Validation could not be completed in this handoff because I could not execute the focused repro or test run in this environment. No conflict issues were detected in the merge state.

Recommendation

🏁 escalate to a human