This release makes subagents in Herdr and Rex panes much more dependable, and lets you open one on your own machine from a Pi running inside Herdr without SSH. Editors and other hosts can now show what a subagent is doing while it runs, for single subagents and for every child of a workflow. A subagent limited to a list of tools now fails right away with a clear error when another extension swaps out those tools, instead of quietly running without them. It also includes a batch of smaller fixes for worktrees, workflow output files, and models that send placeholder values.
Highlights
- From a Pi running inside Herdr,
machine: "local"opens the subagent in a new tab of your own Herdr session, with no saved machine or SSH. - Subagents in Herdr or Rex panes get time to load a heavy set of extensions, report their tool calls, keep large messages, and no longer end as failed after doing their job.
- Editors and other hosts can show a subagent's transcript while it runs.
- A subagent with a
toolslist stops right away with a clear error when an extension such as pi-fabric's code mode replaces its tools. - Models such as GLM-5.3 that fill unused options with
"null"can callsubagent, andmachine: "local"works on NixOS.
Changelog
Highlights
- From a Pi running inside Herdr,
machine: "local"opens the subagent in a new tab of your own Herdr session, with no saved machine or SSH. - Subagents in Herdr or Rex panes are much more dependable: they get time to load a heavy set of extensions, report their tool calls, keep large messages, and no longer end as failed after doing their job.
- Editors and other hosts can now show a subagent's transcript while it runs, for foreground subagents and for every child of a workflow.
- A subagent limited to a list of
toolsnow stops right away with a clear error when another extension, such as pi-fabric's code mode, swaps out its tools, instead of wasting turns on tools it can't call. - Models such as GLM-5.3 that fill unused options with
"null"can callsubagent, andmachine: "local"works on NixOS.
Added
- Workflow progress now names each child's session file. The
workflowChildrenrows of a workflow's progress updates, status and results carrysessionFilefrom the child's launch, and a finished child keeps it while its siblings still run, so an editor or other host can show any child's transcript during the workflow. Before, a host following a workflow could see a child's state but had no way to read what it was doing until the workflow returned. Thanks to @zeishr for #3290. - Foreground progress updates now name the child's session file. Each
details.progress[]row of asubagenttool update carriessessionFilefrom the first update, when the child has one, so an editor or other host can show the child's transcript as it runs. Before, the path appeared only in the final result, once the child had finished. Thanks to @zeishr for #3289. - From a Pi running inside a Herdr pane,
machine: "local"now opens the subagent in a new tab of your own Herdr session, with no saved machine or SSH. Before,localworked only inside Rex, and placing a child in Herdr on the same machine meant a saved machine and SSH to yourself. When Herdr runs inside a Rex tab, the Herdr session wins, and outside both terminalslocalfails before anything starts. Thanks to @mystery4f for #3266.
Fixed
- A child that times out or is stopped in a folder outside Git now gets a one-line note in its recovery summary ("changed tracked files: unavailable (not a Git work tree)"). Before, the summary and the child's output file held git's entire
git diffoption reference. Thanks to @sternnick for #3287. - A child with a
toolsallowlist now fails before its first model request when an extension loaded in the child replaces the tools sent to the model, and the error names the tools that were left out and the ones added in their place. Before, a background child that inherited pi-fabric's code mode started with no usable tool: the model saw onlyfabric_exec, every call to it failed with "Tool fabric_exec not found", and the child spent turns finding that out. Thanks to @VasantLong for #3291. - The
worktreeSetupHookdocs now say that a new worktree already has anode_moduleslink to the main checkout when the hook runs, and that a hook which installs dependencies must remove the link first. Before, the link was undocumented, so a hook runningpnpm installwrote into the main checkout'snode_modulesand could break it. Thanks to @yxlimo for #3288. subagentcalls from models that fill every unset optional parameter with the string"null"andworkflowwith"false", such as GLM-5.3, now work. Before, those placeholders counted as real values, so every call failed with errors such as "Named workflow resource execution must omit action". Only exact"null"values andworkflow: "false"are ignored; a task that merely contains the word null is unchanged. Thanks to @AlexKucera for #3285.machine: "local"in Herdr or Rex now works on hosts whose basic tools are not in/usr/binor/bin, such as NixOS. Before, its setup scripts ran with the fixedPATHused for remote machines instead of Pi's own, sorm,mkdir,mktempandcatwere not found and the child never started. Remote machines keep the fixedPATH. Thanks to @yxlimo for #3284.- The worktree handoff line in a subagent's result now says, when the child changed files, that its work is in the saved patches and was not applied to your checkout, and after a partial cleanup it points at the task entries that say what was left behind. When the handoff file itself can't be written, the error now points at
git worktree list. Before, the line gave only counts, so a parent reading "1 changed patches" could take the change as already landed. Thanks to @yxlimo for #3283. - When
worktree: truerefuses to start because the checkout has uncommitted changes, the error now names that checkout and lists the changed files (up to 10). Before, it said only "requires a clean git working tree" and named neither the checkout nor the files. Thanks to @samclams for #3276. - A workflow given an
outputfile no longer writes its error into that file when it fails, stops, or pauses. The workflow now removes any earlier file at that path when it starts and writes the file only when it completes; the error stays in the tool result, the completion notice, andstatus.json. Before, the file appeared even when the workflow failed within milliseconds of starting, so a script or parent that waited for the file treated the failure as a delivered report. Thanks to @samclams for #3275. - When a subagent in a managed worktree commits its work, cleanup now keeps its branch instead of deleting it. Single, chain and parallel runs name the kept branch and its last commit in their result or completion notice; for a workflow child it is recorded in the worktree handoff file listed in the child's
artifactPaths. Before, cleanup force-deleted the branch, so the commits' messages and authorship survived only as unreachable objects you had to find withgit fsck. A branch whose commits are already on another branch, a remote branch, or a tag is still deleted. Thanks to @samclams for #3274. - A subagent placed in a Herdr or Rex pane no longer loses a large message or crashes when the parent reconnects. Before, a message too big to send to the parent in one piece, such as a screenshot it read or a very large file it wrote, never reached the parent, and the next time the parent reconnected to that pane the subagent's Pi crashed. Now the parent gets the message shortened: images and very long text are replaced by a note saying what was left out, and very long tool arguments are cut. Two large messages sent back to back also no longer break the connection.
- A subagent with
outputSchemaplaced in a Herdr or Rex pane is now refused before any pane opens, with an error namingoutputSchemaamong the options a placed Pi can't use yet. Before, it started, could never callstructured_output, and failed at the end with an error telling it to call a tool it didn't have. - A subagent placed in a Herdr or Rex pane no longer fails to start when another extension in that pane opens its own Pi session at startup. Before, the pane was closed with "did not start the pi-subagents bridge" or the launch timed out, because loading pi-subagents a second time in the same Pi cut the connection to the first copy. Thanks to @mystery4f for finding and fixing this in #3263.
- A subagent placed in a Herdr or Rex pane now gets 30 seconds to start. Before, the parent gave up after 100 checks, about 6 seconds for a local Rex pane, so a Pi with many extensions was closed as "did not start the pi-subagents bridge" while it was still loading. Thanks to @mystery4f for finding and fixing this in #3263.
- When a subagent in a Herdr pane fails after its Pi started, for example on a version mismatch or a missing required tool, the error now says why. Before, it could say only "Could not remove the exact owned remote runtime directory", because the closing Pi had already removed that folder. Thanks to @mystery4f for finding and fixing this in #3263.
- Subagents placed in a Herdr or Rex pane now report their tool calls. Status, the fleet view and results show their tool count and current tool, per-tool timeouts apply to them, and time spent waiting on your answer no longer counts against their
timeoutMs. Before, the parent saw none of a placed subagent's tool calls, so it showed 0 tools, a stuck tool was never stopped, and a wait on your answer used up the subagent's time. Thanks to @mystery4f for the report in #3263. - A subagent placed in a Herdr or Rex pane (
machine) no longer ends as failed when it did its job. It now gets the same instructions for reporting its result as other subagents. Before, it never saw them, so it could not write the acceptance report and every run under the default acceptance was rejected. Thanks to @mystery4f for the report in #3263.