pi-subagents Files

This release makes run listings and workflow launches harder to break, and makes the bundled skill much cheaper to use. A single run with a damaged status file can no longer hide all your other runs from /subagents-fleet and status listings. The skill that teaches agents how to delegate now costs about 3k tokens per use instead of 10.6k, and agents following it block on child runs less often. On Opus 5.5, Sonnet 5.5 and Fable 5.1, a workflow launch whose script gets lost now recovers on the next try instead of failing again.

Highlights

• A run with a damaged status file no longer hides your other runs from /subagents-fleet and status listings.
• The bundled skill costs about 3k tokens per use instead of 10.6k, and agents following it block on a child less often.
• On Opus 5.5, Sonnet 5.5 and Fable 5.1, a workflow launch that loses its script recovers on the next try instead of failing again.

Highlights

• A run with a damaged status file no longer hides your other runs from /subagents-fleet and status listings.
• The bundled skill costs about 3k tokens per use instead of 10.6k, and agents following it block on a child less often.
• On Opus 5.5, Sonnet 5.5 and Fable 5.1, a workflow launch that loses its script recovers on the next try instead of failing again.

Changed

• The bundled pi-subagents skill is about a quarter of its old size: about 3k tokens per use instead of 10.6k. It keeps the common calls, the workflow and resume examples, and the rules for when and how to delegate, and points to the guide sections for everything else instead of copying the reference docs. In testing, agents using it blocked on a child with async: false 2 times in 39 tasks instead of 10. /council no longer loads an extra 11k-token reference.
• On Claude 5.x models (Opus 5.5, Sonnet 5.5, Fable 5.1), text written after a tool call reaches Pi only as a short summary without code blocks, so a ```js workflow block written at that point is lost and subagent({ workflow: true }) can't find it. The error now explains this and tells the agent to pass the script as a file for the rest of the turn, and the bundled skill tells these models to use a file from the start once the turn has made a tool call.

Fixed

• A recent background run whose status.json no longer parses, for example a file filled with NUL bytes, no longer stops /subagents-fleet and status listings from showing your other runs. The run is skipped with one warning, shown as a notice while /subagents-fleet is open, which repeats only if the file changes. Its files are left as they are, and asking for that run by id still reports the parse error. Thanks to @ozand for [#3040] and [#3041].