- Bounded rewind report recovery to messages created after the active checkpoint.
- Added resumed-context regression coverage for late stale rewind results.
Fixes#7739
- Implemented in-house, zero-dependency utility modules in `pi-utils` covering DOM manipulation, markdown parsing, templating, browser automation helpers, and terminal buffers.
- Migrated packages across the repository to consume the new internal utilities and `omptype` schema validators instead of external dependencies.
- Removed multiple external runtime and development dependencies including Zod, Marked, LRU cache, Turndown, and Puppeteer browser packages.
Repeated /mcp reauth commands could remain blocked by a prior
unfinished login because each command receives a fresh controller.
Coordinate flows through session-shared state so a replacement cancels
and cleans up the old flow before proceeding.
A single-model subagent has no fallbacks of its own, so it inherits one
and is pinned to a `subagent:<id>` role. That pin is inserted first in
`retry.fallbackChains` so no other role can capture its routing — which
also means it shadows every configured role chain at runtime.
Inheritance was hardcoded to `chains.default`, so a subagent spawned
through a role alias (the bundled scout's `model: "@smol"`) retried on
the default role's chain instead of its own. With `smol` chained to
composer/grok/luna and `default` chained to gpt-5.6-sol, every scout
fell back onto sol.
Resolve the inherited chain from the role identity still present in the
raw pattern (`@smol` -> `smol`), falling back to `default` when that
role configures no chain. An explicitly empty role chain still means
"no fallbacks", mirroring `expandDefaultRetryFallbackChains`. Explicit
model selectors keep inheriting `default`: they carry no role identity,
and a role assigned the same model must not capture the child's routing.
Closed worker and cmux run signals before yielding for floating-rejection drainage. Stale promise continuations can no longer begin page navigation after evaluated code returns.
Classified only marked browser failures and evaluated-run stack frames as run-owned rejections. Unrelated tab-worker failures now remain on the worker guard's fatal path.
Used a run-scoped Promise subclass instead of mutating native combinator methods. Evaluated code can now freeze its Promise constructor without breaking cleanup or later browser runs.
Observed Promise.all and Promise.race results derived from browser calls during each evaluated run. User catch continuations that rethrow browser failures now fail the owning run without changing native await behavior.
Logged late user continuation failures in cmux runs and delayed worker rejection folding until request-interception cleanup completed. This closes both windows where missing awaits could be silently dropped.
Logged user continuation rejections that settle after their browser run has ended. This preserves the completed result while making missing awaits visible instead of silently dropping them.
Tracked whether user continuation callbacks create each descendant rejection. Browser errors that user code rethrows now fail the owning run instead of being contained as propagated helper failures.
Scoped browser-error markers to each run and contained only propagated browser failures. Routed floated user continuations into failed runs and added worker coverage for native await plus every continuation method.
Registered live session resume commands with postmortem handling so a
fatal rejection or exception identifies every recoverable agent before
cleanup. Escaped terminal control characters in recovery output.
Observed every browser facade continuation so fire-and-forget helper
timeouts cannot wedge or kill a tab worker. Preserved native Promise
identity for callers and test matchers.
Scanned the Windows host USERPROFILE .agents directory when running under WSL so globally installed Agent Skills are available alongside Linux-home skills.
Fixes#3779
(cherry picked from commit c3468dae4f9b91646bf50f2cf4ded9075572290e)
can1357