Repository navigation
fix(llm): default to claude-opus-5-5 from one owner; keep claude-opus-5 on the ladder - #1941
Merged
Merged
Conversation
…-5 on the ladder The default Claude model was the previous-generation claude-opus-5, and it was spelled out again in the quantization runner, the report client (three times), the proposal-scorer defaults and the --proposal-scorer-models help text, so a default change had to be found and repeated by hand. DEFAULT_CLAUDE_MODEL is now claude-opus-5-5, the current Opus, and every code-side default derives from it (the help text renders %(default)s). claude-opus-5 stays in the allowlist as the second rung, so a gateway that serves Opus 5 but not 5.5 falls back one generation, not two, and an explicit --claude-model claude-opus-5 is still accepted under the strict allowlist. Docs and the setup skill name the new default; the AMD proxy template keeps its pinned ids, which an operator chooses for their gateway. Refs #1936 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HktBcjPwjE7QgkA8MgKH6q
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HktBcjPwjE7QgkA8MgKH6q
Code metrics gate: PASSEDJudged: the 23 files this change touches. Files it does not touch never fail it; anything found there is listed under Outside this change for information.
Module length warning: over 800 lines, within the 1200-line failure limit (not failing): 1
Already over at the base (backlog, not this change's): 1
96 findings in files this change does not touchOutside this change (informational): 96
...and 71 more; the job log lists every row.
Thresholds and sources: |
chaojhou
approved these changes
Oct 16, 2026
chaojhou
left a comment
chaojhou
left a comment
Collaborator
There was a problem hiding this comment.
PR #1941 -- fix(llm): default to claude-opus-5-5 from one owner; keep claude-opus-5 on the ladder
What it does: moves DEFAULT_CLAUDE_MODEL (common/llm_config.py) from claude-opus-5 to claude-opus-5-5 and makes the copies that hard-coded the old id (quantization runner, the three report-client defaults, the proposal-scorer list and its --proposal-scorer-models help) derive from it; claude-opus-5 stays second on the strict allowlist ladder, so a gateway without 5.5 falls back to it and an explicit claude-opus-5 remains valid. Docs, the setup skill and the GEAK shell fallback name the new default.
Blocking issues: none
Checked: llm_config.py default and resolve path; credentials._CLAUDE_ALLOWED_MODELS; the --proposal-scorer-models default and help at parser.py:924-937; every remaining claude-opus-5 literal at head; model-id keyed logic (KernelForge adaptive-thinking parser, provider-reported cost in tracker/usage.py); docs naming the default | Ran: git grep over the head tree, gh api for checks and comments | Base: 7ff5758 | Head: bead7c2
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
DEFAULT_CLAUDE_MODEL(common/llm_config.py) moves fromclaude-opus-5toclaude-opus-5-5, the current Opus (per the Claude model table: Opus 5 is still served but is the previous generation; 5.5 is the successor, $4/$20 vs $5/$25 per MTok). The literal was also repeated inagents/quantization/driver/runner.py,breakdown/reporters/llm_client.py(×3),orchestrator/scoring/proposal_scorer.py, and the--proposal-scorer-modelshelp string; those now derive from the constant (the help renders%(default)s)._CLAUDE_ALLOWED_MODELSkeepsclaude-opus-5as the second rung, so the catalog fallback ladder is5-5 → 5 → 4-8 → 4-7 → 4-6, and an explicit--claude-model claude-opus-5stays valid underINFERENCE_OPTIMIZER_ALLOW_CUSTOM_ORCH_MODEL=0. Docs (install, slurm, authentication, SKILL, troubleshooting), the setup skill's default option andinstall_kernel_tools.sh's shell fallback name the new default.Deliberately unchanged:
assets/slurm/proxy.env.templatekeeps its pinnedclaude-opus-5values (AMD gateway operators pin what their gateway serves; GEAK has no catalog fallback), only its ladder comment is updated;.env.template's commented "pin an older id" example.Opus 5.5 request-shape check: no forced
tool_choice(any/tool) anywhere;temperatureis only sent when a caller sets it, and the one caller that does is the OpenAI-protocol report client; KernelForge's_supports_adaptive_thinkingparsesclaude-opus-5-5as major 5 → adaptive thinking, neverbudget_tokens(now pinned by a test case).test_agent_roles_and_policyladder,test_proposal_scorer,test_reporters_llm_client, KernelForgetest_resolve_and_cli,test_agent_env_contract,test_claude_resume,test_provider_registry); addedclaude-opus-5-5to the adaptive-thinking cases. Tests that passclaude-opus-5as an explicit input are unchanged. 93 test files touching the default / allowlist / scorer / report client / runner / parser / docs: 3616 passed; 13 fail identically on cleanmainin this environment (all intest_install_rocprof_compute.py, host profiler-install probes).ruffclean;bash -non the installer;code-metricsPASSED.Live: an 8 h GLM-5.2 MXFP4 session ran its orchestration, critic and GEAK on
claude-opus-5-5(subscription auth, official endpoint).DEFAULT_CLAUDE_MODEL) remains, plus the shell fallback ininstall_kernel_tools.sh, which cannot import it.CLAUDE_MODEL/--claude-model, runs (orchestration, critic, report narrative, quantization agent, proposal scoring, GEAK fallback) now useclaude-opus-5-5; setup offersUse default (claude-opus-5-5). Before, they used the previous-generationclaude-opus-5. On a gateway without 5.5, preflight now printsWARNING — 'claude-opus-5-5' not in gateway catalog; falling back to 'claude-opus-5'.CLAUDE_MODEL=claude-opus-5keeps it, including under the strict allowlist.🤖 Generated with Claude Code