An Agent Hub row reported a subagent as having run on a model that never
spoke. All 97 of its requests, 421K tokens and $6.19 of cost were served
by the primary; a transient stall then armed a fallback, that fallback
errored on its first request with an exhausted quota, and the run died.
Attribution followed the routing switch rather than the output.
Three surfaces lied independently, each re-deriving "the current model"
and calling it the run's model: the executor's progress snapshot, the
session's fallback selector that the hub row reads first, and the
transcript walk behind a settled row.
Sessions now own attribution. `AgentSession.servingModel` names the model
that produced this session's output, holding the last model that actually
served while a candidate is armed but unproven. A switch is a routing
decision, not evidence the target can produce anything, so the answer only
moves once a turn on the target settles.
Consumers read it instead of reconstructing it. The executor's observer
dropped its own event bookkeeping: that bus also carries advisor turns
running on a different model, and it was reading `retry_fallback_applied`
as proof of service. The hub row reads the same getter, so the main
session — which has no executor progress and no persisted history — stops
rendering an unproven candidate as its plain configured model. A fallback
armed before anything has served is still shown, marked as a fallback,
because there is no earlier work to miscredit there.
One predicate decides "this turn produced output", shared by the live
session and the offline replay so they cannot disagree. `error` and
`aborted` are both failures — a stalled stream is finalized as `aborted`
with its partial block still attached, so a stop reason alone proves
nothing — and a turn needs actionable content, which a `length` stop
burning its budget on unsigned thinking does not have. It tolerates
malformed content blocks: transcripts outlive the shapes that wrote them,
and one bad line previously blanked a whole row's history.
Ordering matters at two swap sites. Both the chain advance and the
cooldown-expiry restore move the model and fan `model_changed` out to
subscribers synchronously, so each now updates fallback state before the
swap rather than after; otherwise an observer reading attribution inside
that window sees the incoming candidate carrying the outgoing one's proof.
A startup-selected fallback owns the run from its first request only on a
fresh session. A resumed transcript already holds turns another model
produced, so there the candidate stays unproven until it answers.
`retryFallbackModel` is removed: every consumer reads `servingModel`, and
keeping a parallel derived getter alive for tests is the duplicate surface
this change set exists to remove.
A single-model subagent is pinned to a `subagent:<id>` role whose
`retry.fallbackChains` entry shadows every configured role chain, so the
chain it inherits decides where the child retries. Inheritance resolved
the role by re-deriving it from the child's `modelPatterns` — but every
spawn path expands the role alias into `modelOverride` before calling
`runSubprocess` (`modelPatterns = normalizeModelPatterns(modelOverride ??
agent.model)`), so `@task` never reached the derivation and it returned
`undefined` every time. Every task subagent inherited `chains.default`.
With `modelRoles.task: anthropic/claude-sonnet-5`, `task` chained to
sonnet alone, and `default` chained to sonnet plus a second provider, a
transient stall on sonnet routed the child onto the default chain's
second model — one the operator had deliberately kept out of the `task`
chain — and a quota error there killed a 28-minute run.
#7694 fixed only the shape where an unexpanded alias reaches the
executor, which no production caller produces; its tests supplied a bare
`agent.model: ["@smol"]` with no `modelOverride`. The incident above
happened on v17.2.10, which contains that fix.
Route inheritance off the role identity the spawn path already computes
and passes as `modelRole`. Since that leaves the pattern-derived operand
unreachable, drop it and the parameter it was the only user of.
The vibe worker path had the same defect independently: `#resolveWorker`
expanded `@task`/`@smol` for the bundled `task`/`sonic` workers and kept
no role, so vibe children inherited `default` no matter what the
executor did. It now carries `modelRole` on `ResolvedVibeWorker` and
`VibeRecord` through both the spawn and rehydrate sites.
To stop the two halves drifting apart again — the mistake that caused
this bug — `resolveAgentModelSelection` returns the expanded `patterns`
and the pre-expansion `role` from one call, and both spawn paths take
both from it. `resolveAgentModelSource` is removed: its only use was
being fed to `resolveExplicitModelRole`, and keeping it invites the same
split derivation. `resolveAgentModelPatterns` stays for the UI callers
that legitimately want patterns alone.
Tests cover the producible shapes: the incident's chain layout (role
chain equal to the primary, default chain a superset), role identity
surviving expansion for every alias-routed bundled agent, and the
patterns/role pairing itself. #7694's two tests are re-anchored to a
shape a real caller produces.
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.
- Update hashline block resolution formatting to correctly incorporate anchor lines within operation labels.
- Fix and update test assertions and mock contexts across coding agent tests.
Reverted branch-side edits to spawn-policy prompts/tests, settings tab
groups, mermaid cache typing, prewalk todo gating, and packages/ai test
churn back to merge-base content; trimmed their changelog entries. These
repaired stale CI against an older main and are stale or conflicting
against current main.
Forward unresolved explicit subagent model selectors into child session startup so modelRoles.task cannot disappear during executor preflight and fall through to an unrelated provider default.\n\nFixes #4421
Kept OpenRouter and Vercel upstream routing suffixes in subagent retry fallback selectors so same-base routed candidates stay distinct.
Resolved retry fallback candidates from the raw selector before model switching so routed fallback models keep their requested upstream route.
Placed subagent-scoped fallback chains ahead of inherited retry chains so overlapping primary selectors prefer the explicit subagent model order.
Expanded the regression to pin fallback chain insertion order when a global chain shares the same primary selector.
Installed subagent-scoped retry fallback chains from ordered task model candidates so provider failures can advance to the next configured worker model.
Updated task result tracking to surface the fallback-applied final model and added focused regression coverage for the executor wiring.
Fixes#2750
can1357