Add @deepseek-ai/dsh-agent-model. It provides ctx.agentModel
(resolveForNextStep, select, applyConfiguredRoute) and owns the
modelSelection projection, which the Agent registry no longer registers.
dsh-agent-loop now injects agentModel; the base and SDK-minimal bundles
and the Python SDK runtime include it.
The provider keeps the projection registered until consumers that are
still unloading finish their cleanup. The model-switch notice becomes an
opt-in per-Agent plugin at @deepseek-ai/dsh-agent-model/notice, and
agent/request now also rejects reasoning-effort changes after step
preparation.
Mount the provider in every composition that loads dsh-agent-loop, and
update consumers, the switch-notice snapshot driver, docs, and the
upgrade guide.
Address the ds-review-bot round on #6272: the README summary, completion
notice, token effect, and Known Limitations lines no longer call the
payload the final answer; the empty-result error reads "without an
answer"; the ordering spec also interleaves commentary; the partial_answer
branch cites its upstream source.
Codex app-server 0.162 adds `partial_answer` to the agentMessage phases
(stable answer text that later assistant output or tool calls may follow).
The wire treated every phase other than final_answer, commentary, and null
as fatal, so one such message failed the whole delegation. Record the
latest partial answer and select final, then partial, then unphased text;
unknown phases still fail the run. The pinned @openai/codex 0.153.4 payload
never emits the phase, so behavior with the shipped binary is unchanged.
Route fields and list_subagent_models are now registered only by the
tool-subagent instance configured with modelSelectionSettings, which keeps
one fixed definition per composition for Web presets while headless and SDK
profiles return to their previous schemas. Discovery registers with that
instance instead of a separate list-models plugin entry, so the delegation
description never references an absent tool. modelSelectionSettings on a
provider that inherits the parent conversation now fails the mount.
The subagent guidance now tells the model to end its turn when only waiting
for delegated results remains, since each result starts a new turn. The read
guidance names sed, head, and tail alongside cat and covers line ranges.
The compaction-output-reserve and compaction-summary-headroom context windows
grow by 40 tokens to keep their triggers between the history and its
condensed replacement.