fix(llm): omit temperature for major-only Opus ids (claude-opus-5)

[husamemadH]

Summary

_anthropic_rejects_temperature() decides whether to send temperature on the native Anthropic path — Opus 4.7+ returns HTTP 400 if the field is present at all. Its version pattern required a minor component (opus-(\d+)[-_.](\d{1,2})), so a major-only id like claude-opus-5 never matched, the guard reported "this model accepts temperature", and the field was sent. Every such call fails with `temperature` is deprecated for this model, aborting the stream in ~0.4s with zero tokens — surfaced in the UI as "The model returned an empty response."

This makes the minor optional and treats a missing minor as .0, so claude-opus-5 parses as (5, 0) and is correctly gated as >= 4.7.

The major is also capped at 1–2 digits with a no-trailing-digit lookahead, mirroring the minor. This is required once the minor becomes optional: with a greedy (\d+) major, claude-3-opus-20240229 (legacy Claude 3 Opus, date directly after opus-) would parse as version 20240229 and silently lose its temperature — a model that accepts it. The old pattern avoided that only as a side effect of requiring a minor.

Interactive chat never hit this: it leaves temperature as None so nothing is sent. The scheduled-task path calls stream_agent_loop() without a temperature and inherits its temperature: float = 0.3 default, which is sent — so background tasks broke while chat looked healthy.

Target branch

• This PR targets dev, not main.

Linked Issue

Fixes #5753

Type of Change

• Bug fix (non-breaking — fixes a confirmed issue)

Checklist

• I searched open issues and open PRs — this is not a duplicate.
• This PR targets dev
• My changes are limited to the scope described above — no unrelated refactors or whitespace changes mixed in.
• I actually ran the app and verified the change works end-to-end. Verified at the payload level instead — see Verification below.

How to Test

1. Reproduce on dev (the guard answers False for a model that is newer than 4.7):

   docker compose run --rm --no-deps -T --entrypoint python odysseus -c '
   from src.llm_core import _anthropic_rejects_temperature, _build_anthropic_payload
   print("rejects:", _anthropic_rejects_temperature("claude-opus-5"))
   p = _build_anthropic_payload("claude-opus-5", [{"role":"user","content":"hi"}], 0.3, 100)
   print("temperature in payload:", "temperature" in p)'

   Before: rejects: False / temperature in payload: True (the 0.3 is the scheduled-task default).
   After: rejects: True / temperature in payload: False.

2. Full predicate table from the issue, plus the legacy Claude 3 Opus case:

   model id	before	after	expected
   claude-opus-5	False	True	True
   claude-opus-5-20260101	False	True	True
   claude-opus-4-8	True	True	True
   claude-opus-4-7	True	True	True
   claude-opus-4-6	False	False	False
   claude-opus-4-20250514	False	False	False
   octopus-4-8	False	False	False
   claude-3-opus-20240229	False	False	False

3. Tests:

   python -m pytest tests/test_llm_core_anthropic_temp_omit.py tests/test_llm_core_anthropic_temp_clamp.py -q

   42 passed. Also ran python -m pytest tests/ -k "llm_core or temperature or anthropic" — 235 passed.

Note on the fix suggested in the issue

The issue proposes making the minor optional while leaving the major as (\d+). That regresses claude-3-opus-20240229 to True (the greedy major swallows the date as version 20240229), which fails the existing case at tests/test_llm_core_anthropic_temp_omit.py:51. Capping the major at 1–2 digits handles both; the issue's table didn't include a legacy Claude 3 Opus id, which is why it looked clean.

Verification

Verified at the payload level, which is where this bug lives — the field is either in the request body or it isn't:

• Ran _anthropic_rejects_temperature() and _build_anthropic_payload() against the real module in a container built from dev, before and after the change. Before: claude-opus-5 keeps temperature. After: it is omitted.
• Confirmed the full predicate table above, including the two dated-id edge cases and octopus-4-8.
• tests/test_llm_core_anthropic_temp_omit.py + tests/test_llm_core_anthropic_temp_clamp.py: 42 passed.
• pytest tests/ -k "llm_core or temperature or anthropic": 235 passed, no regressions.

The HTTP 400 itself is as described in #5753; this change removes the field that triggers it.

Disclosure

The patch was drafted with Claude Code. I reproduced the bug myself, reviewed the regex change line by line, and ran the tests before opening this. Flagging it per CONTRIBUTING.md's note on agent-generated PRs — the issue (#5753) was filed independently and this is a single targeted fix, not a bulk submission.

Files changed

• src/llm_core.py — version pattern in _anthropic_rejects_temperature(), plus comments explaining both digit caps.
• tests/test_llm_core_anthropic_temp_omit.py — major-only ids added to the "rejects" set (claude-opus-5, dated, provider-prefixed, future claude-opus-6); payload test for the scheduled-task path; explicit payload guard for legacy Claude 3 Opus.

No UI/rendering changes.