v0.162.0

[0.162.0] - 2026-09-09

Added

• orgChart now shows the work that actually happened, not only the work that could (#3386).
  The chart has always drawn the capability graph — a node per delegate, an edge wherever
  one agent can delegate to another. It now overlays the delegation ledger on top, so each
  node carries how many dispatches it actually received, how they ended, and when it was
  last used. Two things become visible that a wiring diagram cannot show: a delegate that
  is configured but has never once been used, sitting beside one used constantly; and a
  delegate failing every dispatch while its health dot stays green, because a probe answers
  "can I reach it?" and a dispatch answers "did the work go through?".

• Subagents appear on the chart for the first time (#3386).
  They are drawn from the ledger, never from config: a subagent definition is a job title
  rather than a worker, so a node per definition would fill the chart with roles nobody has
  ever delegated to. For an agent whose delegation is all in-process this is the entire
  picture — its chart was previously empty of its own work. A subagent node reports no
  liveness, because it runs inside its owner and there is nothing to probe.

• A detached delegation is now closed out when it lands (#3386).
  A background job's edge was recorded at dispatch and left reading ok with a zero
  duration forever — a claim about an outcome nobody observed. The edge is now settled with
  its real outcome and duration when the job completes, hooked at mark_complete because
  that is the one funnel every settle path goes through (the background manager's own
  completion and cancel paths, and the A2A terminal hook). A redundant settle cannot
  restate an outcome already recorded, and a cancellation is still not a failure.

• A read API for the ledger (#3386).
  GET /api/ledger returns recent edges, newest first, optionally scoped to one
  originating chat session; GET /api/ledger/edges returns them aggregated per target —
  the shape a graph draws. The aggregate reports priced alongside cost_usd, so a
  partial sum over the rows that had a measurable cost is never mistaken for a total.

• Plugin setup gaps can carry a bounded, declarative remediation action (#3389). registry.report_setup_gap() (and the underlying setup_gaps.report()) now accept an optional action — a single hint or a list — that a future console can map to a "fix this" affordance. Actions are closed, server-validated data, never behavior: a fixed kind vocabulary (plugin_config, which is scoped to the reporting plugin, plus a reserved global_settings target) with bounded plain-text label/fields. Any unknown kind, callback, arbitrary URL, HTML, or oversized payload is dropped rather than stored, and a malformed action never raises into plugin loading. Legacy report_setup_gap(key, message) calls are untouched — identical storage, warnings() output, and clear behavior — and the runtime HTTP status response is unchanged for now.

Fixed

• Interjecting mid-turn no longer renders the agent's answer twice (#3387).
  Sending a second message while a turn streams makes the console split its live
  assistant bubble, so the interjection lands where the agent actually consumed it:
  what had been said freezes into its own bubble, and an emptied continuation keeps
  streaming below. The A2A terminal frame then re-sends the whole turn's canonical
  text — which landed on that continuation in full, drawing the frozen prose a second
  time. The two halves carry different ids by design, so the existing id-dedupe never
  saw it and the duplicate persisted to local history for good.
  A turn's canonical text is now distributed across the bubbles that render it,
  exactly once (apps/web/src/chat/turnText.ts), on every path that carries it: the
  live terminal frame, reattach after an agent switch or reload, the post-stream task
  reconcile, and boot hydration. Transcripts already holding a duplicate are repaired
  when they load. Delegation exchanges, which split a turn the same way, were exposed
  to the same bug and are covered by the same fix.

• Expanding a tool chip mid-turn no longer collapses the moment the turn finishes (#3390).
  ToolCalls renders its summary chips from several branches that hold different numbers of
  children — a live turn has a spotlight slot, a settled one doesn't — and the chips carried no
  React key. React matches unkeyed children by position, so on settle a chip shifted index,
  landed in a slot that had held something else, and was rebuilt from scratch; the DS
  ToolCardSummary keeps its open state internally, so the rebuild silently threw away whatever
  the operator had expanded. Both the "N background jobs" chip and the "N tools" fold chip now
  keep their identity across the transition, so an expansion made while the agent is still
  working survives it.
  This was also reddening unrelated PRs: the console e2e suite failed roughly one random spec per
  parallel run, and this chip was one of two causes (the other, an SSE delivery race in the
  server-turn spec, is fixed alongside it).

---

Full Changelog: v0.161.0...v0.162.0