fix(opencode-go): send x-opencode-session header for session affinity

[chelsealong]

What does this PR do?

OpenCode Go's chat_completions relay returns HTTP 400 Model is unavailable for
some backends (confirmed for deepseek-v4-flash) unless the request carries a
stable, opaque x-opencode-session header. Hermes' opencode-go provider profile
never sent that header, so the very first request to an affected model fails.

This adds x-opencode-session to OpenCodeGoProfile.build_api_kwargs_extras,
resolved with the same pattern already used for OpenRouter's x-grok-conv-id
sticky-routing header: the ambient conversation context first, falling back to
the explicit session_id, normalized through the existing
_cache_scope_from_session_id helper. That keeps the value:

• stable across turns of the same conversation,
• correctly rotated after session rotation/compression (it's derived from the
  session-lineage root, same as the OpenRouter/xAI header),
• scoped to this provider only — it's added inside the opencode-go profile's
  own hook, so it can't leak into other providers' requests.

This does not send any official-OpenCode-CLI impersonation headers
(User-Agent: opencode-cli, x-opencode-client: cli) — only the honest,
non-impersonating session-affinity header, matching the approach the issue
explicitly asks for.

Related Issue

Fixes #81584
Fixes #81832

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• plugins/model-providers/opencode-zen/__init__.py: OpenCodeGoProfile.build_api_kwargs_extras
  now attaches extra_headers["x-opencode-session"] whenever a conversation/session
  id is available, on top of the existing reasoning-control logic (moved unchanged
  into a private _reasoning_api_kwargs_extras helper).
• tests/plugins/model_providers/test_opencode_go_profile.py: added
  TestOpenCodeGoSessionAffinityHeader covering header presence with a
  session_id, cron-timestamp normalization/stability (mirrors the existing
  OpenRouter grok-header test), header omission with no session context, and
  coexistence with the existing reasoning top_level kwargs.

How to Test

1. python -m pytest tests/plugins/model_providers/test_opencode_go_profile.py -q
2. Reproduce the failure without the fix: git checkout HEAD~1 -- plugins/model-providers/opencode-zen/__init__.py,
   rerun the new tests (they fail with KeyError: 'extra_headers'), then
   git checkout HEAD -- plugins/model-providers/opencode-zen/__init__.py to restore the fix.

Actual output

Without the fix (source reverted, tests kept):

FAILED ...::TestOpenCodeGoSessionAffinityHeader::test_session_id_sets_affinity_header - KeyError: 'extra_headers'
FAILED ...::TestOpenCodeGoSessionAffinityHeader::test_header_normalizes_cron_timestamp - KeyError: 'extra_headers'
FAILED ...::TestOpenCodeGoSessionAffinityHeader::test_header_coexists_with_reasoning_top_level_kwargs - KeyError: 'extra_headers'
3 failed, 1 passed, 20 deselected in 1.02s

With the fix:

........................
24 passed in 2.07s

Also ran the broader related suites with no regressions:

tests/plugins/model_providers/test_opencode_go_profile.py
tests/providers/test_provider_profiles.py
tests/hermes_cli/test_opencode_go_validation_fallback.py
tests/hermes_cli/test_opencode_zen_model_limit.py
tests/hermes_cli/test_model_switch_opencode_anthropic.py
tests/hermes_cli/test_opencode_go_flat_namespace.py
tests/hermes_cli/test_opencode_go_in_model_list.py
tests/agent/transports/test_chat_completions.py
tests/agent/test_auxiliary_client.py
=> 51 + 45 + 172 passed, 0 failed

ruff check on both changed files: all checks passed.

Checklist

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(opencode-go):)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix
• I've run pytest tests/ -q (relevant subset — see above) and all tests pass
• I've added tests for my changes
• Tested on Ubuntu 24.04

Documentation & Housekeeping

• N/A — no user-facing config/docs changes

AI assistance disclosure

This PR was prepared with AI assistance (Claude, via an autonomous coding agent),
reviewed and tested before submission.

[Enough1122]

Thanks for this fix! I applied it locally (cherry-picked both files into the provider profile) and can confirm it resolves the issue end-to-end: deepseek-v4-flash via opencode-go now works reliably, where previously every chat/completions request returned HTTP 400 "Model is unavailable" without the x-opencode-session header. Verified with the new session-affinity tests plus the broader opencode-go suites (24 + 51 passed, no regressions), and in live desktop usage. Looking forward to seeing this merged into main!

[spfcraze]

This was generated by AI during triage.

Summary:
Fixes #81584 will auto-close the issue on merge, but the header is attached only where build_api_kwargs_extras runs — the anthropic_messages and codex_responses transports never call it, and the issue's regression coverage item 1 names all three routes.

Problems:

• build_api_kwargs_extras is invoked only from agent/transports/chat_completions.py:670 and agent/auxiliary_client.py:8031; neither agent/transports/anthropic.py nor agent/transports/codex.py calls it, so opencode-go turns on the anthropic_messages or codex_responses routes still send no x-opencode-session header after this change.
• The profile's per-model routing sends GPT models to codex_responses and MiniMax/Qwen to anthropic_messages (docstring in plugins/model-providers/opencode-zen/__init__.py), so opencode-go serves those routes.
• Issue [Bug]: OpenCode Go deepseek-v4-flash fails with 400 "Model is unavailable" unless x-opencode-session is sent #81584's regression coverage item 1 asks for the header on chat_completions, anthropic_messages, and responses routes.

Solution:
Extend the header to the anthropic_messages and codex_responses transport paths as well.

---

Checked against 8fed6e0 — the tip of fix/opencode-go-session-affinity-header when this was written — and 2389564, main at the same moment.

[chelsealong]

Good catch — pushed 26654cc: build_api_kwargs in agent/chat_completion_helpers.py now merges the x-opencode-session header onto the anthropic_messages route (MiniMax/Qwen on OpenCode Go) and the codex_responses route, the same way Nous Portal's tags/session_id are merged onto the Messages wire — build_api_kwargs_extras is only consulted by the chat_completions transport otherwise. Added TestOpenCodeGoSessionHeaderOtherTransports covering both routes; confirmed the new tests fail (KeyError: 'extra_headers') without the fix and pass with it. Ran the full previously-listed suite plus these + codex/anthropic transport tests (368 passed) and ruff check on both changed files.

[DavidMetcalfe]

This PR also fixes the failure reported in #81832 (same root cause: opencode-go + deepseek-v4-flash, HTTP 400 without x-opencode-session; different upstream error message text but the same missing-header trigger). Consider adding "Fixes #81832" to the PR body so both issues auto-close on merge.

[chelsealong]

Added "Fixes #81832" to the PR body per your note — same root cause, so it'll auto-close alongside #81584 on merge. No code changes needed.