Clarify the SystemPromptDecisionInput.startsSeries contract (a new request series no longer implies consolidation), record the tool-schema series-start exception consistently in the decision and consequences, and correct the complete-declaration attribution to the unsupported-route projection.
An in-history route carries prompt updates as later system nodes while node 0 stays stale. A surface replacement (compaction, tool-result prune, image offload) starts a new request series, and the projection used to re-baseline node 0 even when the rendering already matched the latest surviving node: it emptied that node and wrote identical text into node 0, so the request differed at its first message and the provider prefix cache missed from the first token (#6973).
Re-baseline node 0 only when the effective text must move: an incapable route, an empty rendering, or a series start whose latest surviving non-empty node differs from the rendering. A capable series start with matching text emits no system event; the loop still logs the request series and rebuilds declarations. Retry compaction still folds the prompt to the head when it shadows the latest node.
ctx.tools.register(definition, { when }) adds a synchronous or asynchronous
condition. AgentLoop resolves conditions once per proposed step with the
agent, the model selection that step preparation captured, and the turn
signal, before prompt assembly and agent/pre-step. The resulting
ResolvedToolSet governs tool declarations, Native dispatch, PTC bindings,
and nested calls for that step and its retries, and the turn disposes it
when the proposal is rejected or fails, when no step opens, and after the
step. A false condition removes the tool everywhere; a thrown condition
ends the turn with TOOL_ACTIVATION_FAILED. Calls whose registration was
replaced or removed during the step return UNKNOWN_TOOL.
system-prompt/assemble now transforms PromptPresentation without a tool
array, so tool declarations stay provider-owned. exit_plan_mode is declared
only for steps whose selected plan state is active, the same state that
renders plan guidance; it still rejects a call before the selection is
logged.
Model selection stays as master defines it; this change adds no model
selection event or Agent field.
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.
Delete @deepseek-ai/dsh-invariants, every package's ./invariant companion
and its tests, the invariant gates and Vitest setup, the scoped-event
resolver generator, and the INVARIANT-coded listener rethrow paths.
Update docs, READMEs, Agent Notes, and add an upgrade guide.
Tool changes previously forced prompt consolidation even when the prepared
route could activate new tools through history. Rewriting the first system
message, or clearing an earlier appended prompt, reset the declaration
series and turned additions into immediate declarations.
Only routes without toolUpdate need that tool-triggered consolidation.
Capable routes append the full changed prompt and keep existing messages.
Explicit series starts, surface replacements, and unsupported prompt
routes still consolidate; retained-name definition changes still rebuild
tool declarations in Session.
Cover both update modes and reset paths, and pin the combined prompt/tool
transition through the shipped headless profile with shared schema goldens.
Log tool changes as developer messages independently of model capability,
fold committed headers and updates into Session.toolHistory(), and let the
LLM runtime construct deferred and retained declarations for routes that
declare a toolUpdate mode. DeepSeek serializes updates as system-role
tool_addition/tool_removal blocks with defer_loading and the tool-changes
beta header. An unchanged restored tool continues the series and is
re-offered through its recorded blocks instead of rewriting declarations.
Replace the obsolete rejection tests with wire assertions and add coverage
for the session fold, loop emission, route projection, config validation,
and the beta header. Chat and Trajectory changes land separately.
The 2026-07-16 explicit-turn-cancellation note still described a frozen,
detached signal reason and a two-variant cause union. The README paragraph
added here points readers at that note as the owner of the signal
lifecycle, so the stale mechanism was load-bearing.
It now records why freezing the caller's cause is not an alternative:
undici assigns `stack` from strict-mode code, so a frozen reason makes
fetch reject with `TypeError: Cannot add property stack` in place of the
abort reason.
Also names `turn/end`'s declared `TurnEndCancelCause` in the README and
takes the reviewer's helper-doc wording.
Node's fetch assigns an enumerable `stack` string onto the abort reason
object it receives, and the loop logged that same object into `turn/end`.
The undici trace became durable session data under a field no reader
declares. A transport adding a property JSON discards instead made
`Session.append` reject the ending, leaving a turn with no close.
`abortedCancelCause()` narrows the untyped `AbortSignal.reason` once and
copies the declared fields, replacing the three `as AgentCancelCause`
assertions at the two read sites and the log write.
Give durable user context its producer-owned source kind instead of a generic plugin wrapper. The V3-to-V4 conversion preserves the frozen rename and collision rules, all own JSON metadata, event coordinates, and historical request-only kinds. Native catalog and JSONL admission retain source validation so unknown attribution can survive without its producer while malformed owned sources are refused.
Use an explicit user-only RequestUserInput for auxiliary LLM prompts that do not own durable message identity. Auto Review and compaction remove their request-only source registrations while retaining prompt content and freezing; provider serializers and projections keep the same wire data, assistant replay isolation, image ordering, and tool correlation. Session writes, Agent delivery, and recorded title requests continue to require identified durable messages.
Bind the saved compatibility policy only to user sources and qualify tmux location attribution without weakening fixed system, model, or tool semantics. Rebuild the current inventory and unpublished V4 acknowledgement on the actual tool-role parent, preserving its rationale and all accepted historical records. The catalog format version records tooling metadata; Session remains V4. Developer-role policy and activation belong to their own feature.
The latest-master Messages adapter, serialization, request-input typing, EpochHeader reservation, and attribution checks pass 103 tests across five files. The retained V3 Messages input-history snapshot passes through the source profile; Host/Client typechecks and all 41 documentation gates pass. The generated comparison records four changed roots and 447 unchanged type fingerprints. Latest-master ARIA changes are retained, with producer source labels applied independently; new Messages fixtures retain their producer names as declared source kinds and use the fixed system-message helper; recorded Session generations are unchanged by this regrouping.
Co-authored-by: Tianyi Cui <[email protected]>
mektpoy