[Backlog Discovery] feat(backlog): llm-token-budget-guardrails

[bestony]

[Backlog Discovery]

• Requirement title: 为 LLM 调用增加 Token 预算与超限处理
• Priority: P1
• Requirement file: backlog/20260221073930-llm-token-budget-guardrails.md
• Dedupe key: llm-token-budget-guardrails
• Source run: https://github.com/bestony/self/actions/runs/22252887858

---

[Backlog Discovery]

Update Record 2026-02-21 15:42:32 +08:00

Update summary:

• Added an Internal Evidence Needed section requesting a 30-day analysis of token limit failures, impacted workflows, logs, and scope.

• Expanded the proposed solution with input-priority trimming rules and a regression evaluation baseline for quality, failure rate, and cost impact.

• Updated acceptance criteria to require the 30-day evidence, deterministic trimming priority rules, and regression evaluation metrics.

• Renumbered sections to accommodate the new evidence requirements.

• Status: committed

• Commit: 5daa2871c111e0acbbc6b2988b6512db6a2ca21a

• Trigger: workflow_run.completed

• Comment: https://github.com/bestony/self/actions/runs/22252941516

• Actor: @github-actions[bot]

• Updated At (Asia/Shanghai): 2026-02-21 15:42:32 +08:00

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[github-actions]

[Reviewer Workflow]
Reviewer: Product Manager

需求价值评估

• 是否有价值: 有价值
• 优先级: P1
• Reviewer 视角结论: 该需求直指长上下文导致的稳定性与成本问题，但需补齐内部失败数据以校准收益与投入。

价值点

• 通过 token 预算与超限策略降低 LLM 请求失败/隐式截断，提升 self 的定时/自驱流程稳定性。
• 统一裁剪与摘要规则可提升输出一致性与可复现性，降低维护者排查成本。
• 预算可观测与 fail-fast 能减少无效调用与成本浪费，改善运营效率。

风险与建议

• 内部证据缺失（最近 30 天失败/截断统计）会影响价值量化与优先级判定，建议先补齐基线数据。
• 摘要/裁剪策略可能牺牲关键信息，建议明确输入优先级并用固定用例回归验证。

[github-actions]

[Backlog Discovery]
Reviewer: Product Manager

• Scope boundary is vague: “all LLM calls” isn’t enumerated by workflow (e.g., reviewer/product/backlog/scheduler). Suggest listing in-scope workflows and any explicit out-of-scope areas so completion is verifiable.
• Acceptance criteria around regression metrics lacks measurable targets (e.g., acceptable failure-rate/cost-change thresholds). Suggest defining baseline sources and success thresholds to make go/no-go decisions testable.