管住它 · 配到「敢放手」再自动跑Keep It in Check · Configure Until You Dare Let Go.
一个会自己跑命令的 agent,信任是要管的。这章从零配出一个敢放手的 profile:选模型(官方 credits / BYOK / custom endpoint 三条账本)、定权限(四档加 allowlist 与 denylist)、审命令(跑前你看见,跑完你看 diff)—— 配完用一次真实改动验收。An agent that runs commands on its own is a trust you manage. This chapter builds a profile you dare let go from scratch: pick the model (three ledgers — official credits / BYOK / custom endpoint), set permissions (four tiers plus allowlist and denylist), review commands (you see it before it runs, you read the diff after) — then validate the setup on one real change.
一个会自己跑命令、还能接外部工具的 agent,信任是要管的 —— 不能全凭它自觉。这章带你从零配出一个「敢放手」的 profile:建出来,权限定好,allowlist 与 denylist 写死,然后用一次真实改动当场验收 —— 看它该自动的自动、该拦的拦住。配完这三节细讲背后的三本账:选模型、定权限、审命令。An agent that runs commands on its own and reaches external tools is a trust you manage — you can't rely on its good behavior. This chapter builds a profile you dare let go from scratch: create it, set the permissions, pin the allowlist and denylist, then validate it on one real change — watching it auto-run what should auto-run and stop where it should stop. After that, three sections unpack the three ledgers behind it: model, permissions, review.
— I
主线:配一个敢放手的 local-dev profileThe Main Line: A local-dev Profile You Dare Let Go.
01
建:Settings > Agents > Profiles 新建 local-devCreate: a new local-dev profile under Settings > Agents > Profiles
新建一个 profile,命名 local-dev,base model 先选你常用的那档(Auto 也行,模型账第二节再细算)。Profile 的意义是把「这台机器上干什么活、放到什么程度」固定成一个可以复用的组合,而不是每次会话现调。1注 1Note 1Warp Docs · Agent Profiles & Permissions —— 截至 2026-08-06:Profile 可配置 base model(也用于 Planning)、autonomy 与各类权限(Apply code diffs / 读文件 / 做计划 / 执行命令 / Full Terminal Use / 问澄清问题),每项分 Agent Decides / Always ask / Always allow / Never 四档;Apply code diffs 的 Agent decides 目前等同 Always ask,只有 Always allow 跳过 diff 审查;问问题权限三档(Never ask / Ask unless auto-approve / Always ask);allowlist 接正则,denylist 优先于 allowlist 与 Agent Decides;Run until completion(⌘⇧I)连 denylist 都绕过。2026-07-23 起 /fast-forward 可切换单会话 autoexecute(GUI+TUI);2026-07-31 起 execution profiles 全员可从设置文件配置。Warp Docs · Agent Profiles & Permissions — as of 2026-08-06: a Profile configures the base model (also used for Planning), autonomy, and per-type permissions (Apply code diffs / Read files / Create plans / Execute commands / Full Terminal Use / Ask clarifying questions), each at Agent Decides / Always ask / Always allow / Never; for Apply code diffs, Agent decides currently behaves like Always ask — only Always allow skips diff review; the ask-questions permission has three settings (Never ask / Ask unless auto-approve / Always ask); the allowlist takes regex, the denylist takes precedence over the allowlist and Agent Decides; Run until completion (⌘⇧I) bypasses even the denylist. Since 2026-07-23, /fast-forward toggles per-conversation autoexecute (GUI + TUI); since 2026-07-31, execution profiles are configurable from settings files for all users.Create a profile named local-dev; for the base model, pick the tier you already use (Auto is fine — the model ledger gets its own section next). A profile's point is fixing 'what work happens on this machine, and how far it may go' into a reusable combination, instead of re-tuning every conversation.1注 1Note 1Warp Docs · Agent Profiles & Permissions —— 截至 2026-08-06:Profile 可配置 base model(也用于 Planning)、autonomy 与各类权限(Apply code diffs / 读文件 / 做计划 / 执行命令 / Full Terminal Use / 问澄清问题),每项分 Agent Decides / Always ask / Always allow / Never 四档;Apply code diffs 的 Agent decides 目前等同 Always ask,只有 Always allow 跳过 diff 审查;问问题权限三档(Never ask / Ask unless auto-approve / Always ask);allowlist 接正则,denylist 优先于 allowlist 与 Agent Decides;Run until completion(⌘⇧I)连 denylist 都绕过。2026-07-23 起 /fast-forward 可切换单会话 autoexecute(GUI+TUI);2026-07-31 起 execution profiles 全员可从设置文件配置。Warp Docs · Agent Profiles & Permissions — as of 2026-08-06: a Profile configures the base model (also used for Planning), autonomy, and per-type permissions (Apply code diffs / Read files / Create plans / Execute commands / Full Terminal Use / Ask clarifying questions), each at Agent Decides / Always ask / Always allow / Never; for Apply code diffs, Agent decides currently behaves like Always ask — only Always allow skips diff review; the ask-questions permission has three settings (Never ask / Ask unless auto-approve / Always ask); the allowlist takes regex, the denylist takes precedence over the allowlist and Agent Decides; Run until completion (⌘⇧I) bypasses even the denylist. Since 2026-07-23, /fast-forward toggles per-conversation autoexecute (GUI + TUI); since 2026-07-31, execution profiles are configurable from settings files for all users.
02
定:四档权限 + 两张清单Set: four tiers plus two lists
执行命令设 Agent decides(确信就做、不确定才问);Apply code diffs 注意:Agent decides 目前等同 Always ask,只有 Always allow 才跳过 diff 审查 —— 想看 diff 就不用动它。allowlist 放三条只读正则(git (status|diff|log) 这类),denylist 放 rm、curl、wget、eval 这类红线 —— denylist 优先于一切。1注 1Note 1Warp Docs · Agent Profiles & Permissions —— 截至 2026-08-06:Profile 可配置 base model(也用于 Planning)、autonomy 与各类权限(Apply code diffs / 读文件 / 做计划 / 执行命令 / Full Terminal Use / 问澄清问题),每项分 Agent Decides / Always ask / Always allow / Never 四档;Apply code diffs 的 Agent decides 目前等同 Always ask,只有 Always allow 跳过 diff 审查;问问题权限三档(Never ask / Ask unless auto-approve / Always ask);allowlist 接正则,denylist 优先于 allowlist 与 Agent Decides;Run until completion(⌘⇧I)连 denylist 都绕过。2026-07-23 起 /fast-forward 可切换单会话 autoexecute(GUI+TUI);2026-07-31 起 execution profiles 全员可从设置文件配置。Warp Docs · Agent Profiles & Permissions — as of 2026-08-06: a Profile configures the base model (also used for Planning), autonomy, and per-type permissions (Apply code diffs / Read files / Create plans / Execute commands / Full Terminal Use / Ask clarifying questions), each at Agent Decides / Always ask / Always allow / Never; for Apply code diffs, Agent decides currently behaves like Always ask — only Always allow skips diff review; the ask-questions permission has three settings (Never ask / Ask unless auto-approve / Always ask); the allowlist takes regex, the denylist takes precedence over the allowlist and Agent Decides; Run until completion (⌘⇧I) bypasses even the denylist. Since 2026-07-23, /fast-forward toggles per-conversation autoexecute (GUI + TUI); since 2026-07-31, execution profiles are configurable from settings files for all users.Set Execute commands to Agent decides (act when confident, ask when unsure); note for Apply code diffs: Agent decides currently behaves like Always ask — only Always allow skips diff review, so if you want to see diffs, leave it. Put three read-only regexes on the allowlist (git (status|diff|log) and friends), and red lines like rm, curl, wget, eval on the denylist — the denylist outranks everything.1注 1Note 1Warp Docs · Agent Profiles & Permissions —— 截至 2026-08-06:Profile 可配置 base model(也用于 Planning)、autonomy 与各类权限(Apply code diffs / 读文件 / 做计划 / 执行命令 / Full Terminal Use / 问澄清问题),每项分 Agent Decides / Always ask / Always allow / Never 四档;Apply code diffs 的 Agent decides 目前等同 Always ask,只有 Always allow 跳过 diff 审查;问问题权限三档(Never ask / Ask unless auto-approve / Always ask);allowlist 接正则,denylist 优先于 allowlist 与 Agent Decides;Run until completion(⌘⇧I)连 denylist 都绕过。2026-07-23 起 /fast-forward 可切换单会话 autoexecute(GUI+TUI);2026-07-31 起 execution profiles 全员可从设置文件配置。Warp Docs · Agent Profiles & Permissions — as of 2026-08-06: a Profile configures the base model (also used for Planning), autonomy, and per-type permissions (Apply code diffs / Read files / Create plans / Execute commands / Full Terminal Use / Ask clarifying questions), each at Agent Decides / Always ask / Always allow / Never; for Apply code diffs, Agent decides currently behaves like Always ask — only Always allow skips diff review; the ask-questions permission has three settings (Never ask / Ask unless auto-approve / Always ask); the allowlist takes regex, the denylist takes precedence over the allowlist and Agent Decides; Run until completion (⌘⇧I) bypasses even the denylist. Since 2026-07-23, /fast-forward toggles per-conversation autoexecute (GUI + TUI); since 2026-07-31, execution profiles are configurable from settings files for all users.
03
验:一次真实改动当场验收Validate: one real change, reviewed on the spot
派一个真任务(第四章那类修 bug 就行),盯两件事:allowlist 里的 git status 它不问就跑 —— 这是省事;denylist 里的命令它必须停下来问你 —— 这是保命。两个都发生,profile 才算配成;拦了不该拦的、放了不该放的,回去改正则。Dispatch one real task (a chapter-4-style bug fix works), and watch two things: an allowlisted git status runs without asking — that's the convenience; a denylisted command stops and asks — that's the insurance. Both happening means the profile works; if it blocks what it shouldn't or releases what it shouldn't, go back and fix the regexes.
04
调:信任按类别逐步松手Tune: loosen trust by command class
同一类命令审过几次都没问题,把它的正则挪进 allowlist;没底的那类一直留在「每次问」。别一上来全放,也别永远每步都拦 —— 敢放手不是一次决定,是一条曲线。单个会话想临时放开,有 /fast-forward(GUI 和 TUI 都有);长期组合则从 2026-07-31 起全员可以从设置文件配置,不用在 UI 里逐台机器点。Once a class of command has passed review a few times, move its regex to the allowlist; the classes you're unsure of stay on ask-every-time. Don't open everything at once, and don't block every step forever — daring to let go isn't one decision, it's a curve. To temporarily open up a single conversation there's /fast-forward (GUI and TUI); and since 2026-07-31, long-term combinations are configurable from settings files for everyone — no per-machine UI clicking.
— II
选模型:官方 credits、BYOK、custom endpoint 三本账Pick the Model: Three Ledgers — Official Credits, BYOK, Custom Endpoint.
模型这件事有三条路:用 Warp 官方模型和 credits,BYOK 接自己的 provider key,或者 custom inference endpoint 接自己的网关。官方 Model Choice 里有四档 Auto 和 OpenAI、Anthropic、Google、xAI、Fireworks-hosted 开源模型(完整名单和价格见第十一章);BYOK 支持 OpenAI、Anthropic、Google,Free 就能用,选中带 key 图标的具体模型时推理走你的 provider 账单、不耗 Warp credits;4注 4Note 4Warp Docs · Agent model choice —— 截至 2026-08-06:Auto 模式四档(Responsive / Cost-efficient / Genius / Open-weights),模型名单覆盖 OpenAI、Anthropic、Google、xAI 与 Fireworks-hosted 开源模型(完整名单与档位价格见第十一章);存在 model fallback chain;每个 Profile 可设独立 base model;Grok 可走连接的 SuperGrok 订阅(xAI 账号额度,不耗 Warp credits);ZDR 可用性按模型变化(Fable 5 要求 Anthropic 数据保留,无 ZDR)。Warp Docs · Agent model choice — as of 2026-08-06: four Auto modes (Responsive / Cost-efficient / Genius / Open-weights), a roster across OpenAI, Anthropic, Google, xAI, and Fireworks-hosted open-weight models (full roster and tier pricing in chapter 11); a model fallback chain exists; each Profile can set its own base model; Grok can run through a connected SuperGrok subscription (xAI account quota, no Warp credits); ZDR availability varies by model (Fable 5 requires Anthropic data retention, no ZDR).2注 2Note 2Warp Docs · Bring Your Own API Key —— 截至 2026-08-06:BYOK 支持 OpenAI、Anthropic、Google;Free 与符合条件的付费档可用(个人及 10 人以内组织,大组织需 Business / Enterprise);key 存本地 OS keychain、请求经 Warp backend in-flight、不存服务端;选中具体 BYOK 模型时优先于 credits;失败 / 限流默认不 fallback 到 Warp credits,有 opt-in 的 credit fallback;ChatGPT / Claude 消费级订阅不能接入(SuperGrok 是唯一消费订阅路径);BYOK 不适用于 Cloud Agents;中心化 admin-managed BYOK 尚未上线,Enterprise 的托管推理走 BYOLLM(当前 AWS Bedrock,Azure Foundry 与 Google Vertex coming soon)。配置入口:Settings 搜 API keys。Warp Docs · Bring Your Own API Key — as of 2026-08-06: BYOK supports OpenAI, Anthropic, and Google; available on Free and eligible paid plans for individuals / orgs up to 10 employees, larger orgs need Business / Enterprise; keys are stored in the local OS keychain, requests transit Warp's backend in-flight, nothing is stored server-side; a selected BYOK model takes priority over credits; failure / rate-limit does not fall back to Warp credits by default — an opt-in credit fallback exists; ChatGPT / Claude consumer subscriptions cannot be connected (SuperGrok is the only consumer-subscription path); BYOK does not apply to Cloud Agents; centrally admin-managed BYOK is not yet available — Enterprise managed inference goes through BYOLLM (AWS Bedrock today, Azure Foundry and Google Vertex coming soon). Configure via Settings search: API keys.custom endpoint 接任何 OpenAI-compatible 目标,2026-07-31 起还能选 schema(OpenAI Chat Completions / OpenAI Responses / Anthropic Messages),localhost 和私网地址不行。3注 3Note 3Warp Docs · Custom inference endpoint + changelog 2026-07-31 —— endpoint 接 OpenAI-compatible 目标(OpenRouter、LiteLLM、z.ai、公开 HTTPS 的内部网关);localhost / 私网被拒(本机模型经 ngrok 类隧道暴露);key 存本地;不适用于 Cloud Agents;Business / Enterprise 本地经 endpoint 仍消耗 platform credits;2026-07-31 起 endpoint 可选 OpenAI Chat Completions / OpenAI Responses / Anthropic Messages 三种 schema(文档页仍只写 Chat Completions,以 changelog 为准)。router 现在可以与 BYOK 组合(解析出模型后套用你的 BYOK key),但永远不能用 custom endpoint。Warp Docs · Custom inference endpoint + the 2026-07-31 changelog — endpoints take OpenAI-compatible targets (OpenRouter, LiteLLM, z.ai, a public-HTTPS internal gateway); localhost / private URLs are rejected (local models via an ngrok-style tunnel); the key is stored locally; endpoints do not apply to Cloud Agents; Business / Enterprise local runs via an endpoint still consume platform credits; since 2026-07-31 an endpoint can select one of three schemas — OpenAI Chat Completions / OpenAI Responses / Anthropic Messages (the docs page still describes Chat Completions only; the changelog is the source of truth). Routers now compose with BYOK (your BYOK key applies after the router resolves a model), but never with custom endpoints.The model question has three paths: Warp-provided models and credits, BYOK with your own provider key, or a custom inference endpoint through your own gateway. Official Model Choice has four Auto tiers plus OpenAI, Anthropic, Google, xAI, and Fireworks-hosted open-weight models (full roster and pricing in chapter 11); BYOK supports OpenAI, Anthropic, and Google, works on Free, and a selected key-icon model bills inference to your provider account, not Warp credits;4注 4Note 4Warp Docs · Agent model choice —— 截至 2026-08-06:Auto 模式四档(Responsive / Cost-efficient / Genius / Open-weights),模型名单覆盖 OpenAI、Anthropic、Google、xAI 与 Fireworks-hosted 开源模型(完整名单与档位价格见第十一章);存在 model fallback chain;每个 Profile 可设独立 base model;Grok 可走连接的 SuperGrok 订阅(xAI 账号额度,不耗 Warp credits);ZDR 可用性按模型变化(Fable 5 要求 Anthropic 数据保留,无 ZDR)。Warp Docs · Agent model choice — as of 2026-08-06: four Auto modes (Responsive / Cost-efficient / Genius / Open-weights), a roster across OpenAI, Anthropic, Google, xAI, and Fireworks-hosted open-weight models (full roster and tier pricing in chapter 11); a model fallback chain exists; each Profile can set its own base model; Grok can run through a connected SuperGrok subscription (xAI account quota, no Warp credits); ZDR availability varies by model (Fable 5 requires Anthropic data retention, no ZDR).2注 2Note 2Warp Docs · Bring Your Own API Key —— 截至 2026-08-06:BYOK 支持 OpenAI、Anthropic、Google;Free 与符合条件的付费档可用(个人及 10 人以内组织,大组织需 Business / Enterprise);key 存本地 OS keychain、请求经 Warp backend in-flight、不存服务端;选中具体 BYOK 模型时优先于 credits;失败 / 限流默认不 fallback 到 Warp credits,有 opt-in 的 credit fallback;ChatGPT / Claude 消费级订阅不能接入(SuperGrok 是唯一消费订阅路径);BYOK 不适用于 Cloud Agents;中心化 admin-managed BYOK 尚未上线,Enterprise 的托管推理走 BYOLLM(当前 AWS Bedrock,Azure Foundry 与 Google Vertex coming soon)。配置入口:Settings 搜 API keys。Warp Docs · Bring Your Own API Key — as of 2026-08-06: BYOK supports OpenAI, Anthropic, and Google; available on Free and eligible paid plans for individuals / orgs up to 10 employees, larger orgs need Business / Enterprise; keys are stored in the local OS keychain, requests transit Warp's backend in-flight, nothing is stored server-side; a selected BYOK model takes priority over credits; failure / rate-limit does not fall back to Warp credits by default — an opt-in credit fallback exists; ChatGPT / Claude consumer subscriptions cannot be connected (SuperGrok is the only consumer-subscription path); BYOK does not apply to Cloud Agents; centrally admin-managed BYOK is not yet available — Enterprise managed inference goes through BYOLLM (AWS Bedrock today, Azure Foundry and Google Vertex coming soon). Configure via Settings search: API keys. a custom endpoint connects any OpenAI-compatible target — since 2026-07-31 with a selectable schema (OpenAI Chat Completions / OpenAI Responses / Anthropic Messages) — but localhost and private-network addresses don't work.3注 3Note 3Warp Docs · Custom inference endpoint + changelog 2026-07-31 —— endpoint 接 OpenAI-compatible 目标(OpenRouter、LiteLLM、z.ai、公开 HTTPS 的内部网关);localhost / 私网被拒(本机模型经 ngrok 类隧道暴露);key 存本地;不适用于 Cloud Agents;Business / Enterprise 本地经 endpoint 仍消耗 platform credits;2026-07-31 起 endpoint 可选 OpenAI Chat Completions / OpenAI Responses / Anthropic Messages 三种 schema(文档页仍只写 Chat Completions,以 changelog 为准)。router 现在可以与 BYOK 组合(解析出模型后套用你的 BYOK key),但永远不能用 custom endpoint。Warp Docs · Custom inference endpoint + the 2026-07-31 changelog — endpoints take OpenAI-compatible targets (OpenRouter, LiteLLM, z.ai, a public-HTTPS internal gateway); localhost / private URLs are rejected (local models via an ngrok-style tunnel); the key is stored locally; endpoints do not apply to Cloud Agents; Business / Enterprise local runs via an endpoint still consume platform credits; since 2026-07-31 an endpoint can select one of three schemas — OpenAI Chat Completions / OpenAI Responses / Anthropic Messages (the docs page still describes Chat Completions only; the changelog is the source of truth). Routers now compose with BYOK (your BYOK key applies after the router resolves a model), but never with custom endpoints.模型旋钮上还有一层:custom model router。它不是第四本账,是把「每次手挑模型」变成写好一次的策略:Settings > AI > Custom Routers 里建,或写 YAML 放进 ~/.warp/custom_model_routers/,按任务复杂度分档或自然语言规则匹配 —— 「改数据库迁移」走便宜模型,「排查生产事故」走 frontier 模型。它在模型选择器里像模型一样被选,每次请求解析成一个具体模型,你始终能看到实际跑的是哪个。两处更新:router 现在能和 BYOK 组合(解析出模型后走你的 key);但 router 永远不能用 custom endpoint。5注 5Note 5Warp Docs · Custom routers —— 截至 2026-08-06:在 Settings > AI > Custom Routers 建,或写 YAML 放 ~/.warp/custom_model_routers/;按任务复杂度(easy / medium / hard 加必填 default)或自然语言规则(自上而下首条命中)把每次请求解析到一个具体模型;出现在模型选择器里;可与 BYOK 组合;Team-synced routers 是 Enterprise 功能。Warp Docs · Custom routers — as of 2026-08-06: create in Settings > AI > Custom Routers or author YAML in ~/.warp/custom_model_routers/; resolve each request to one concrete model by task complexity (easy / medium / hard plus a required default) or by natural-language rules matched top to bottom; appears in the model picker; composes with BYOK; team-synced routers are an Enterprise feature.The model knob has one more layer: custom model routers. Not a fourth ledger — it turns picking a model by hand each time into a policy you write once: create it in Settings > AI > Custom Routers, or drop YAML into ~/.warp/custom_model_routers/, routing by task complexity or natural-language rules — database migrations go to a cheap model, production incidents to a frontier one. It appears in the model picker like any model, each request resolves to one concrete model, and you can always see which actually ran. Two updates: routers now compose with BYOK (your key applies after resolution); but a router can never use a custom endpoint.5注 5Note 5Warp Docs · Custom routers —— 截至 2026-08-06:在 Settings > AI > Custom Routers 建,或写 YAML 放 ~/.warp/custom_model_routers/;按任务复杂度(easy / medium / hard 加必填 default)或自然语言规则(自上而下首条命中)把每次请求解析到一个具体模型;出现在模型选择器里;可与 BYOK 组合;Team-synced routers 是 Enterprise 功能。Warp Docs · Custom routers — as of 2026-08-06: create in Settings > AI > Custom Routers or author YAML in ~/.warp/custom_model_routers/; resolve each request to one concrete model by task complexity (easy / medium / hard plus a required default) or by natural-language rules matched top to bottom; appears in the model picker; composes with BYOK; team-synced routers are an Enterprise feature.
模型路径Model path
适合Best for
你承担什么What you own
官方模型 / official creditsOfficial models / official credits
省心、Auto、custom routers、Cloud Agents、团队默认使用Convenience, Auto, custom routers, Cloud Agents, team defaults
credits、模型可用性、计划限制credits, model availability, plan limits
BYOK
已有 OpenAI / Anthropic / Google 账单,希望控制 provider 成本You already have OpenAI / Anthropic / Google billing and want provider-level cost control
API key 安全、provider 成本、rate limit、provider 数据策略API key security, provider cost, rate limits, provider data policy
OpenAI-compatible router / gateway / self-hosted public endpointOpenAI-compatible router / gateway / public self-hosted endpoint
兼容性、公开可达性、延迟、稳定性、日志与隐私compatibility, public reachability, latency, stability, logs and privacy
local modellocal model
仅当通过公开 HTTPS 的 OpenAI-compatible gateway 暴露给 WarpOnly if exposed to Warp through a public-HTTPS OpenAI-compatible gateway
它不是本机 localhost 直连It is not direct localhost access
提示词PromptBYOK 设置前 checklistBYOK setup checklist
准备在 Warp 里配置 BYOK / custom endpoint。
先确认:
- 我用的是具体 provider model,不是 Auto;
- 这个 key 的权限最小、可撤销、有预算上限;
- provider 侧 retention / training 设置符合项目要求;
- 我知道请求会经 Warp backend in-flight;
- 这不会覆盖 Cloud Agents;
- 失败默认不 fallback,要不要开 opt-in credit fallback。Preparing to configure BYOK / custom endpoint in Warp.
Confirm first:
- I will select a concrete provider model, not Auto;
- the key has minimal permissions, can be revoked, and has a budget limit;
- provider-side retention/training settings match the project;
- I know requests transit Warp's backend in-flight;
- this will not cover Cloud Agents;
- failure doesn't fall back by default — whether to enable the opt-in credit fallback.
准备在 Warp 里配置 Anthropic BYOK。
先确认:
- 我选的是 Claude Sonnet 这个具体模型,不是 Auto;
- 这个 Anthropic key 只开了 inference 权限、能在控制台一键撤销、设了每月 50 美元上限;
- Anthropic 侧 retention 设为 30 天、已关掉 training,符合项目要求;
- 我知道请求会经 Warp backend in-flight;
- 这不会覆盖 Cloud Agents;
- 开 opt-in credit fallback 兜底,免得限流时断活。Preparing to configure Anthropic BYOK in Warp.
Confirm first:
- I will select the concrete model Claude Sonnet, not Auto;
- this Anthropic key has inference-only permission, can be revoked from the console in one click, and has a $50/month budget limit;
- Anthropic-side retention is set to 30 days with training off, which matches the project;
- I know requests transit Warp's backend in-flight;
- this will not cover Cloud Agents;
- enable the opt-in credit fallback so rate limits don't stall work.
— III
定权限:四档、两清单、一个绝不能开错的开关Set Permissions: Four Tiers, Two Lists, One Switch You Can't Get Wrong.
自主度不是一个开关,是一条从「每步都问」到「全放行」的滑杆 —— 你该停在中间。四档(Agent Decides / Always ask / Always allow / Never)分别管 diff 落地、读文件、做计划、执行命令、Full Terminal Use、问澄清问题;常用的中间档是「确信就做、不确定才问」。两侧用清单兜底:allowlist 在 Settings > Agents > Profiles,接正则,把信得过的只读命令放进去自动跑;denylist 放绝不能自动跑的,优先于 allowlist 和「Agent decides」,命中必须人工批准。1注 1Note 1Warp Docs · Agent Profiles & Permissions —— 截至 2026-08-06:Profile 可配置 base model(也用于 Planning)、autonomy 与各类权限(Apply code diffs / 读文件 / 做计划 / 执行命令 / Full Terminal Use / 问澄清问题),每项分 Agent Decides / Always ask / Always allow / Never 四档;Apply code diffs 的 Agent decides 目前等同 Always ask,只有 Always allow 跳过 diff 审查;问问题权限三档(Never ask / Ask unless auto-approve / Always ask);allowlist 接正则,denylist 优先于 allowlist 与 Agent Decides;Run until completion(⌘⇧I)连 denylist 都绕过。2026-07-23 起 /fast-forward 可切换单会话 autoexecute(GUI+TUI);2026-07-31 起 execution profiles 全员可从设置文件配置。Warp Docs · Agent Profiles & Permissions — as of 2026-08-06: a Profile configures the base model (also used for Planning), autonomy, and per-type permissions (Apply code diffs / Read files / Create plans / Execute commands / Full Terminal Use / Ask clarifying questions), each at Agent Decides / Always ask / Always allow / Never; for Apply code diffs, Agent decides currently behaves like Always ask — only Always allow skips diff review; the ask-questions permission has three settings (Never ask / Ask unless auto-approve / Always ask); the allowlist takes regex, the denylist takes precedence over the allowlist and Agent Decides; Run until completion (⌘⇧I) bypasses even the denylist. Since 2026-07-23, /fast-forward toggles per-conversation autoexecute (GUI + TUI); since 2026-07-31, execution profiles are configurable from settings files for all users.Autonomy isn't a switch but a slider from ask-every-step to full-allow — and you should stop in the middle. Four tiers (Agent Decides / Always ask / Always allow / Never) govern applying diffs, reading files, planning, executing commands, Full Terminal Use, and asking clarifying questions; the everyday middle is act-when-confident-ask-when-unsure. Backstop both ends with lists: the allowlist lives under Settings > Agents > Profiles, takes regex, and auto-runs the read-only commands you trust; the denylist holds what must never auto-run, takes precedence over both the allowlist and 'Agent decides,' and a match always needs manual approval.1注 1Note 1Warp Docs · Agent Profiles & Permissions —— 截至 2026-08-06:Profile 可配置 base model(也用于 Planning)、autonomy 与各类权限(Apply code diffs / 读文件 / 做计划 / 执行命令 / Full Terminal Use / 问澄清问题),每项分 Agent Decides / Always ask / Always allow / Never 四档;Apply code diffs 的 Agent decides 目前等同 Always ask,只有 Always allow 跳过 diff 审查;问问题权限三档(Never ask / Ask unless auto-approve / Always ask);allowlist 接正则,denylist 优先于 allowlist 与 Agent Decides;Run until completion(⌘⇧I)连 denylist 都绕过。2026-07-23 起 /fast-forward 可切换单会话 autoexecute(GUI+TUI);2026-07-31 起 execution profiles 全员可从设置文件配置。Warp Docs · Agent Profiles & Permissions — as of 2026-08-06: a Profile configures the base model (also used for Planning), autonomy, and per-type permissions (Apply code diffs / Read files / Create plans / Execute commands / Full Terminal Use / Ask clarifying questions), each at Agent Decides / Always ask / Always allow / Never; for Apply code diffs, Agent decides currently behaves like Always ask — only Always allow skips diff review; the ask-questions permission has three settings (Never ask / Ask unless auto-approve / Always ask); the allowlist takes regex, the denylist takes precedence over the allowlist and Agent Decides; Run until completion (⌘⇧I) bypasses even the denylist. Since 2026-07-23, /fast-forward toggles per-conversation autoexecute (GUI + TUI); since 2026-07-31, execution profiles are configurable from settings files for all users.
Profile 名称:<safe-ops / local-dev / sandbox-yolo>
Base model:<model>
默认自主度:<Agent decides / Always ask / Always allow / Never>
允许自动读:<目录/文件类型>
允许自动执行:<只读命令正则>
必须询问:写入、删除、网络下载、sudo、MCP 写工具
禁止:生产数据库写入、deploy、secret 输出、home 目录批量改动
验收:每次结束必须给命令列表、变更 diff、验证命令和结果。Profile name: <safe-ops / local-dev / sandbox-yolo>
Base model: <model>
Default autonomy: <Agent decides / Always ask / Always allow / Never>
Auto-read allowed: <directories/file types>
Auto-execute allowed: <read-only command regex>
Must ask: writes, deletes, network downloads, sudo, MCP write tools
Forbidden: production DB writes, deploys, secret output, bulk home-directory changes
Acceptance: every run ends with command list, change diff, verification command and result.
Profile 名称:local-dev
Base model:Claude Sonnet
默认自主度:Agent decides
允许自动读:./src 下的 .ts / .tsx 和 package.json
允许自动执行:^(ls|cat|git (status|diff|log)|npm run lint)\b
必须询问:写入、删除、网络下载、sudo、MCP 写工具
禁止:生产数据库写入、deploy、secret 输出、home 目录批量改动
验收:每次结束必须给命令列表、变更 diff、验证命令和结果。Profile name: local-dev
Base model: Claude Sonnet
Default autonomy: Agent decides
Auto-read allowed: .ts / .tsx and package.json under ./src
Auto-execute allowed: ^(ls|cat|git (status|diff|log)|npm run lint)\b
Must ask: writes, deletes, network downloads, sudo, MCP write tools
Forbidden: production DB writes, deploys, secret output, bulk home-directory changes
Acceptance: every run ends with command list, change diff, verification command and result.
— IV
审命令:它要跑什么、改了什么Review Commands: What It'll Run, What It Changed.
放手不等于不看 —— 它跑前给你命令,跑后给你 diff,这两眼你得花。默认让它「先给命令、确认再跑」,把要执行的那行摆到你眼前;改动则进 Code Review 面板,diff 框起来看,而不是淹在终端输出里。Apply code diffs 那档别贪快设成 Always allow —— 那是把 review 这一眼也省掉。Letting go isn't not looking — it shows you the command before, the diff after, and those two glances are yours to spend. By default have it 'show the command, confirm, then run,' putting the exact line in front of you; changes go to the Code Review pane, where the diff is framed rather than buried in terminal output. Don't get greedy and set Apply code diffs to Always allow — that's skipping the review glance too.
命令审查:能不能先 dry-run、小样本、只读 SELECT。Command review: can it dry-run, sample, or read-only SELECT first.
成本控制:Auto 消耗 Warp credits;router 可以走 BYOK,但 endpoint 不行;provider 账单也要设预算。Cost control: Auto consumes Warp credits; routers can ride BYOK but endpoints can't; provider billing still needs budgets.
成本控制:Cloud Agents 不走 BYOK;上云前先看 credit 余额和验收方式。Cost control: Cloud Agents do not use BYOK; check credit balance and review path before going cloud.
验收这一章 —— 你的 local-dev profile 要当场配完、当场验过:Accept the chapter — your local-dev profile gets configured and validated on the spot:
建好 local-dev profile,选好了模型路径:官方 credits、BYOK,还是 custom inference endpoint。The local-dev profile exists, with a chosen model path: official credits, BYOK, or custom inference endpoint.
allowlist 里有几条只读命令正则,denylist 里有 rm、curl、wget、eval 和指向生产的命令。The allowlist has read-only regexes; the denylist has rm, curl, wget, eval, and anything pointing at production.
用一次真实改动验收过:该自动的自动跑了,该拦的拦住问我了。Validated on one real change: what should auto-run did, what should block asked me.
自主度停在「确信才做、不确定就问」,没误开 Run until completion;临时放开用 /fast-forward。Autonomy sits at act-when-confident-ask-when-unsure; Run until completion is off; temporary release uses /fast-forward.
记住:BYOK / custom endpoint 不覆盖 Cloud Agents;BYOK 失败默认不 fallback;消费级订阅接不进来。Remember: BYOK / custom endpoint don't cover Cloud Agents; BYOK failure doesn't fall back by default; consumer subscriptions can't connect.
你能放手的程度,
取决于你设界的程度.
How far you can let go
is how far you set the bounds.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return
