Looper v2.13.143

Looper v2.13.143

Release Decision

Who should upgrade: Operators who run the Manager on an api Profile (the api runtime). Everyone else can stay on v2.13.142: no other runtime changes behavior.

What this release solves: #687 lets the user's stop end an api Manager's reply by cancelling the request, which the api runtime refused until now. It marks usage unknown instead of zero when a request ends without it, and adds per-request details to the run record and one structured log line per request.

Breaking changes: No. The api runtime and its session now declare can_interrupt, so stop is accepted instead of refused; nothing is removed.

How to verify: Check the installed version.

Contributors: Maintainer @foolzzz authored #687; no community contribution was included in this patch release.

looper --version   # looper 2.13.143

Behavior changes

• Stop (#687): ApiRuntime.can_interrupt (looper/agents/runtimes/api/adapter.py) and ApiSession.can_interrupt (looper/agents/runtimes/api/session.py) become true. ApiSession.run_turn keeps the turn's TurnCancel and a done event; interrupt() requests the cancel, waits up to 5 seconds for the turn to end (a tool call that is still finishing is never cancelled mid-write) and returns StopMethod.CANCEL. With no turn running it returns nothing. A stop requested before the turn starts, after the platform tool specs load, or a cancelled turn that made no request ends as a user-stop result with empty usage and StopMethod.CANCEL. A cancelled turn that made requests also reports StopMethod.CANCEL. The description also says the stored conversation stays valid after a stop; the changed production code contains no separate transcript handling for it, so that claim rests on the new tests.
• Usage unknown, not zero (#687): token totals are summed only over requests that reported a count, and stay unset when none did (before, a missing count added as 0). In ClaudeCli run finishing (looper/agents/claude_cli.py), a result with usage_complete false clears cost_usd.
• Run record (#687): a new RuntimeAdapter.result_notes hook (looper/agents/runtimes/__init__.py, default empty) is merged into the finished run's notes; the api adapter returns the result's api block. That block now also carries incomplete (the number of requests with incomplete usage), and its upstream_ids are passed through log_redaction.redact.
• Log line (#687): each request of a turn writes one api request record at info level on the looper.agents.runtimes.api logger with agent, protocol, model, base-URL host name, status, duration, first-token time, retries, input, cache read, cache write and output tokens, completeness and the redacted upstream id.
• Docs (#687): the status lines of the API runtime RFC and the long-horizon context plan (both zh-CN) now say stop and usage records are implemented on the branch.
• Tests (#687): new tests/test_api_stop_db.py and tests/test_api_usage_records_db.py; tests/test_api_session.py, tests/test_api_runtime_kind.py, tests/test_api_manager_turns_db.py and tests/test_per_turn_interrupt.py are updated.
• Model-visible changes: none. No prompt, tool declaration or tool result changes; the model sees the same requests. A stopped turn now ends earlier than before; the host-side log line and record notes are not sent to the model.
• Line counts: the description reports tests 510 and production 67 added lines (fixtures excluded).

Upgrade and rollback

Re-run install.sh once v2.13.143 is published. No schema migration. Rolling back to v2.13.142 makes stop on an api Manager reply refused again and drops the per-request notes and log line from new runs; records already written keep their notes.

Optional Capability Activation & Use

No new optional capability activation is introduced in this release.

Validation

Retained implementation evidence: the description of #687 lists the tests that cover the change (tests/test_api_stop_db.py, tests/test_api_usage_records_db.py, tests/test_api_session.py, tests/test_api_runtime_kind.py, tests/test_api_manager_turns_db.py, tests/test_per_turn_interrupt.py, tests/test_provider_pricing.py, tests/test_run_usage.py, tests/test_llm_records_db.py, tests/test_api_loop.py, tests/test_api_responses.py), "automated reviews" and "CI before the merge". It gives no counts or verdicts, and no local full-suite result is recorded. This draft did not re-run any of it.

CI: CI runs after the release is drafted; its result is not recorded here.

Not tested for this draft: Local test runs, a local full suite, a live service, a real provider or chat channel, the built wheel. The release documentation gate checks the body only.

中文摘要

升级决策

**谁需要升级：**在 api Profile（api 运行时）上运行 Manager 的运维者。其他人可留在 v2.13.142：其他运行时的行为不变。

解决了什么：#687 让用户的停止操作可以通过取消请求来结束 api Manager 的回复（此前 api 运行时拒绝停止）；请求结束时没有用量则标记为未知而不是 0；并为运行记录增加每个请求的细节，每个请求写一条结构化日志。

**是否有破坏性变更：**否。api 运行时及其会话现在声明 can_interrupt，停止由拒绝变为接受；没有移除任何内容。

**如何验证：**运行 looper --version 核对版本为 2.13.143。

**贡献者：**维护者 @foolzzz 提交 #687；本次 patch release 无社区贡献。

looper --version   # looper 2.13.143

行为变化与入口

• 停止（#687）：ApiRuntime.can_interrupt（looper/agents/runtimes/api/adapter.py）和 ApiSession.can_interrupt（looper/agents/runtimes/api/session.py）变为 true。ApiSession.run_turn 保存该轮的 TurnCancel 和完成事件；interrupt() 发出取消请求，最多等待 5 秒让该轮结束（仍在收尾的 tool 调用不会在写入中途被取消），返回 StopMethod.CANCEL。没有进行中的轮次时返回空。在轮次开始前、平台工具 spec 加载后请求停止，或被取消的轮次没有发出任何请求，都以用户停止结果、空用量和 StopMethod.CANCEL 结束；已发出请求的被取消轮次同样报告 StopMethod.CANCEL。描述还称停止后已存对话保持有效；改动的生产代码中没有单独的对话记录处理，这一点只由新增测试支撑。
• **用量未知而非零（#687）：**token 总数只对报告了数值的请求求和，没有任何请求报告时保持未设置（此前缺失的数值按 0 计入）。looper/agents/claude_cli.py 的运行收尾中，usage_complete 为 false 的结果会清除 cost_usd。
• **运行记录（#687）：**新增 RuntimeAdapter.result_notes 钩子（looper/agents/runtimes/__init__.py，默认为空），其结果合并进已完成运行的 notes；api 适配器返回结果中的 api 块。该块现在还带 incomplete（用量不完整的请求数），upstream_ids 经 log_redaction.redact 处理。
• **日志（#687）：**一轮中的每个请求在 looper.agents.runtimes.api logger 上以 info 级别写一条 api request 记录，含 agent、协议、模型、base URL 主机名、状态、耗时、首 token 时间、重试次数、输入、缓存读、缓存写和输出 token、完整性以及脱敏后的上游 id。
• **文档（#687）：**API 运行时 RFC 和长程上下文计划（均为 zh-CN）的状态行改为说明停止与用量记录已在分支实现。
• **测试（#687）：**新增 tests/test_api_stop_db.py 和 tests/test_api_usage_records_db.py；更新 tests/test_api_session.py、tests/test_api_runtime_kind.py、tests/test_api_manager_turns_db.py 和 tests/test_per_turn_interrupt.py。
• **面向模型的变化：**无。提示词、tool 声明和工具结果均不变，模型看到的请求相同。被停止的轮次现在会更早结束；主机侧的日志行和记录 notes 不会发给模型。
• **行数：**描述报告新增测试 510 行、生产代码 67 行（不含 fixtures）。

升级与回退

v2.13.143 发布后重跑 install.sh。无表结构迁移。回退到 v2.13.142 会让 api Manager 回复的停止重新被拒绝，新运行不再有每请求的 notes 和日志行；已写入的记录保留其 notes。

可选能力启用与使用

本版本未新增可选能力启用入口。

验证

保留的实现证据：#687 的描述列出覆盖该改动的测试（tests/test_api_stop_db.py、tests/test_api_usage_records_db.py、tests/test_api_session.py、tests/test_api_runtime_kind.py、tests/test_api_manager_turns_db.py、tests/test_per_turn_interrupt.py、tests/test_provider_pricing.py、tests/test_run_usage.py、tests/test_llm_records_db.py、tests/test_api_loop.py、tests/test_api_responses.py）、“自动评审”和“合并前的 CI”，没有给出数量或结论，也没有记录本地完整套件结果。本草稿未重新运行。

**CI：**CI 在发布草稿后运行，其结果未在此记录。

**本次草稿未测试：**本地测试、本地完整套件、运行中的服务、真实 provider 或聊天渠道、构建出的 wheel。发布文档 gate 只检查正文。