@flue/sdk
2.3.1 · 18 Nov 2026Apache-2.0309.1k / weekRepositoryHomepagenpmAtom feed
Upgrading @flue/sdk From 0.3.11 to 0.4.0
1 release in betweenevery note read
- 12
- changes to make
- 0
- advisories fixed
- 21
- new
- 0
- fixes
Before you upgrade
Tick each one off as you go- Breaking0.4.0
- Breaking0.4.0
- Breaking0.4.0
- Breaking0.4.0
- Breaking0.4.0
- Breaking0.4.0
- Breaking0.4.0
- Breaking0.4.0
- Breaking0.4.0
- Breaking0.4.0
Release by release
- Breaking
There are some breaking changes to be aware of, when upgrading from
v0.3tov0.4. - Breaking
commandsanddefineCommandare removed. The original idea — register first-party CLI tools the agent could shell out to — only worked under just-bash and saw little real use. The same surface is better expressed today by passingenvto scope what a connector sees, or by choosing a sandbox connector that gives you the isolation you want. just-bash itself still supports custom commands — you just register them on your bash instance directly instead of through Flue's helper. Removed: thecommands?:option oninit()/prompt()/skill()/task()/shell(); theCommand,CommandDef,CommandOptions,CommandExecutor,CommandExecutorResulttypes; thedefineCommandexport from@flue/sdk/nodeand@flue/sdk/cloudflare; thecommand_start/command_endFlueEventvariants; and theBashLike.registerCommand/SessionEnv.scopeconnector hooks. - Breaking
Default
thinkingLevelchanged from'off'to'medium'. Reasoning-capable models (e.g. gpt-5, claude-opus-4-7) will now reason by default on everyprompt()/skill()/task()call. Non-reasoning models are unaffected (clamped to'off'per the model'sthinkingLevelMap). To restore the old behavior, setthinkingLevel: 'off'explicitly oninit(), your role frontmatter, or the call options. - Breaking
sandbox: 'local'now runs locally. Originally,'local'was a half-isolated layer — ajust-bashsubprocess with aReadWriteFs/MountableFsoverlay mountingprocess.cwd()at/workspace. That made sense when every Flue agent ran on a developer laptop, but increasingly people are deploying the agent itself inside a real sandbox (a container, a microVM, a Cloudflare Sandbox), where wrapping the host in a second virtual filesystem is pure overhead — and actively confusing, because paths get remapped twice.sandbox: 'local'now binds directly to the host:execruns through the user's shell with fullprocess.env, file methods hit the real filesystem, defaultcwdisprocess.cwd(), and there are no path remappings or command restrictions. Agents that hard-coded/workspacepaths must migrate to real host paths. If you want isolation on a developer laptop, reach for a real sandbox connector (Daytona, E2B, Mirage, etc.).
- Breaking
2 other lines we could not sort, mostly version bumps
- 0.4.0
With the new
'local'sandbox, the host shell is exposed directly. The agent'sbashtool can rungh issue view,npm test, etc. with whatever's on$PATHand whatever env you launched flue with. The runner / container / VM is the isolation boundary. - 0.4.0
For non-
'local'sandboxes, install the binaries inside the sandbox image, or wrap the operation as a custom tool withinit({ tools: [...] }). Tools have a structured parameter schema, are visible to the model directly, and recover the "secrets stay on the host" property — the tool readsprocess.env, the agent only sees the tool's params and result.
A whole project to check? Paste its package.json in the dependency check: every deprecated, outdated or vulnerable dependency, with what to upgrade to. Free, no account.