Why: establish the authorized task workspace and its initial baseline.
Evidence: preserved precheck receipt in .task/feature-agent-model-ab.md; R0=43f8e27b74e5e2213e2099ba04e5a69fe398586f.
Impact: task metadata only; the feature workspace starts at 43f8e27b74.
Refs: .task/feature-agent-model-ab.md
Why: Ruling 5 authorized the strong-model comparison with uncapped budget.
The A/B infrastructure is proven end to end (qa7 on remote main, seeded
claude-sonnet-4-6 via the user's existing dmxapi aggregator key, eval/live
ran the full golden set), but the account holds CNY 0.49 against a CNY 0.62
per-request pre-deduction - all six turns failed with insufficient quota,
captured honestly in the report.
Evidence: SELF - .task/model-ab/ab-sonnet-quota-blocked.json + evidence row;
smoke test of the dmxapi streaming endpoint succeeded before the run.
Impact: Task blocked on user input only (top up dmxapi or provide another
provider key). No code changes; worktree retained for immediate resumption.
Why: A small max_tokens probe passed (small pre-deduction), suggesting the
account might have been topped up; redeploying with chat=claude-sonnet-4-6
and lightweight=deepseek-v4-flash showed full-turn pre-deduction (CNY 0.62)
still exceeds the CNY 0.48 balance - 6/6 turns quota-failed again.
Evidence: SELF - .task/model-ab/ab-sonnet-quota-blocked-r2.json + evidence
row; settings confirmed chat=claude-sonnet-4-6, lightweight=deepseek-v4-flash
on the instance before the run.
Impact: Still blocked on user-only payment input. Estimated CNY 50-100 covers
multiple full comparison runs on dmxapi; alternatively any OpenAI-compatible
provider key via ~/.secrets/abmodel or .env.dev SVC_MODEL_*.
Why: Ruling 5's final item. With the user-provided lmuai endpoint
(claude-opus-5-5) and the P5 build unchanged, the same golden task set ran
both sides to completion (6/6 turns each).
Evidence: SELF - B report .task/model-ab/ab-opus.json (2793s) vs A baseline
.task/phase05-eval/after-p5.json; smoke through the live instance confirmed
the provider switch (after fixing the active-secret shadowing: overlayKey
prefers the secret table, so provider changes need /api/secrets rotation).
Impact: Selection report in the card - opus-5-5 eliminates invented numbers
(7 -> 0, the program's core metric), restores independent verification via
lightweight=deepseek-v4-flash; MiniMax stays as the fast/fallback tier.
Recommended default documented for the owner.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
摘要
模型 A/B（裁定 5 收口，整个质量改进计划的最后一项）。任务卡 feature-agent-model-ab；主记录 §4 模型 A/B。
同 golden 任务集、同 P5 构建双模型对照：A=MiniMax-M2.7-highspeed；B=claude-opus-5-5（用户提供的 lmuai 端点）。
建议：默认 chat=claude-opus-5-5 + lightweight=deepseek-v4-flash（独立验证恢复）；MiniMax 留作快速/回退档。
本 PR 仅 .task 变更（选型报告 + 三份 A/B 报告 + evidence 行），无代码改动。
Evidence
make test PASS（evidence 行 SELF）；B 侧全量 evidence 行（2793s）；两轮 quota-blocked 报告如实留档。
运维注意
换供应商时设置页密钥被密钥表活动密钥遮蔽（overlayKey 优先）——经 /api/secrets 新增并激活新密钥（实测踩中并修复）。