Session-log ZIP entries always used DEFLATE level 6 even though compression level is a deployment tradeoff: CPU-constrained hosts may prefer low latency while bandwidth-constrained hosts may prefer smaller archives. A hardcoded level also violated the repository rule that deployment-varying plugin choices live in validated Config.
Add sessionExportCompressionLevel to ApiProxyService.Config as an integer 0-9 with default 6, resolve the same default once for direct createApiProxy callers, and pass the required level into the streaming module. Tests prove schema defaulting and rejection as well as a level-0 versus level-9 archive-size difference with identical extracted content. The generated config catalog, bilingual gateway README, and feature note document the knob and its tradeoff.
The ZIP loop checked desiredSize only after a push and responded to an overfull queue with setTimeout(0). A timer turn does not mean the consumer drained anything, so a slow or disconnected client still allowed the producer to enqueue the complete compressed archive while later artifact and attachment reads ran eagerly.
Give the ReadableStream a 64 KiB byte queuing strategy and block the single producer on a pull-released capacity gate whenever desiredSize is non-positive. Cancellation wakes that gate through the existing producer signal; synchronous fflate output is therefore bounded to the queue high-water mark plus one input push. A regression test exhausts timer turns without consuming and proves the next media entry remains unread until response pulling begins, and the bilingual contracts now describe the real bound.
Only the root raw-artifact read received the request signal. Lineage discovery and descendant reads could continue after disconnect, response-body cancellation did not stop the producer, and the root error boundary converted an abort rejection into an ordinary HTTP 500.
Combine request and response-consumer cancellation into the ZIP producer signal, forward it through every cancellable read, check it around the attachment seam, and terminate fflate exactly once when production stops. The pre-stream boundary now rethrows the original abort instead of translating it. Regression tests cover signal propagation, exact cancellation identity at the HTTP boundary, and a reader cancellation interrupting an in-flight descendant read; the bilingual host contract records these lifecycle semantics.
The exporter read persistence artifacts directly even when the requested root or a descendant was still live. Buffered session events could therefore be omitted from a successful download, so the advertised verbatim-artifact guarantee described storage accurately but captured an arbitrarily stale durability boundary.
Resolve each id against SessionStore and cross its authoritative flush barrier immediately before readRaw. Cold sessions remain a no-op, while live roots and descendants are made durable independently; this intentionally yields a per-session read-boundary snapshot rather than claiming an atomic lineage snapshot. A host-path regression test proves both artifacts change from stale to durable only through flush, and the bilingual host contract and Agent Note document the boundary.
SessionPersistence.readRaw previously used undefined for two unrelated states: a supported backend could not find the requested session, or the backend had no per-session artifact concept at all. The export endpoint consequently reported an existing SQLite-backed session as HTTP 404, which falsely diagnosed storage capability as session absence.
Make raw-artifact support an explicit backend capability. Unsupported backends now fail their inherited readRaw path loudly and the host answers 501 before reading, while undefined retains the single meaning of an absent artifact on a supporting backend. First-party backends, test providers, generated API catalogs, bilingual persistence docs, and export error contracts now state that distinction; focused tests cover both the 501 and the inherited rejection.
The export endpoint already streams a ZIP response, but the web client immediately converted that response into a Blob. That forced the complete archive through JavaScript memory before a download could start and coupled transport, buffering, object-URL lifetime, and filename handling to the trajectory view.
Navigate a temporary download anchor directly to the export endpoint instead. The browser now owns streaming and HTTP failure presentation, while a standalone delivery module owns URL construction and filename sanitization. Focused tests cover the handoff, rejection behavior, and the assembled session view; the package README and feature note record the new ownership boundary.
Only `docs/module-graph.*` conflicted; generated, so taken from master and
regenerated. `api-proxy.ts` auto-merged — this branch's history read
(`historySourceFor` / `presenterScopeFor` / `historyCutOf`, in that order)
is intact.
Hook bypassed as before: the staged-pairing check hands an archived note path
to `verify-translation-pairing`. The full-corpus gate passes.
Only `docs/module-graph.*` conflicted, from master's new
`@deepseek-ai/dsh-message-feedback` package; generated, so taken from master
and regenerated.
Hook bypassed as before: the staged-pairing check hands an archived note path
to `verify-translation-pairing`. The full-corpus gate passes.
Generated docs (`module-graph`, `event-producer-consumer`) taken from master
and regenerated. The `cordis-inspect-jsdoc` golden likewise: master's copy is
the base, and this branch's `presentAs` per-scope rewording is re-applied on
top, since `cordis_inspect` renders that JSDoc into model-visible output.
Hook bypassed as before: the staged-pairing check hands an archived note path
to `verify-translation-pairing`. The full-corpus gate passes.
One assistant message carrying several subagent calls now overlaps them
under the rolling pool (maxParallelToolCalls) instead of serializing each
foreground delegation behind an exclusive barrier; results still commit
in model order.
tool-subagent declares isConcurrencySafe: () => true for every call form:
children work in their own sessions, a run never mutates the parent
session, and sibling workspace coordination already belongs to the model
for background, continuable, and workflow children. The former
serial-execution pin flips to a parallel pin plus a gated genuine-overlap
test, and the authored subagent-parallel snapshot pins the assembled
transcript (tool/call, tool/call, tool/result, tool/result) with
interchangeable twin children, which replay race-free under the existing
first-call binding and createdAt harvest ordering.
The provider seam now requires concurrent starts and continuable
preparations for distinct children to isolate operation-local state,
cancellation, settlement, and cleanup; a continuable gate test proves a
cancelled preparation leaves no Agent or durable Session while its
sibling persists independently.
Closesdeepseek-harness/deepseek-harness#1489
list_agents called a storage-only continuable child 'complete', inviting
parents to poll it for a result that will never appear there. Rename the
status to 'ready' — resumable, not terminal — and rewrite the tool
description to say a parent is told when a child finishes, so listing is
for recall, not polling.
A continuable child that stopped without reporting — an error, a token
ceiling, cancellation, teardown — left its parent nothing to act on.
The continuation manager now delivers an unconditional settlement
notice to the durable direct parent before releasing ownership, folding
consumed work (foldConsumedWork supersedes findLastMessageTurnEnd) so a
claimed-but-unrun prompt reads as aborted rather than completed, waking
an idle parent, steering a busy one, and never waking a closing tree.
A continuable child could finish without giving its parent a usable
result: the report tool was described as optional and its default
delivery stayed quiet. Install the child-scoped report obligation (tool
description plus the tool:report prompt section), default reportDelivery
to wakeup, bind shipped subagent_fork delegations to one-shot so fork
keeps its prefix reuse, and extend the ACP snapshot harness to pin child
system prompts alongside child tool schemas.
Retarget PR #2187 from master 8a763b34645fc9371c0c21595ac65438d603d9bf to c757901957abdfd87f1cd7b11ecab16f3050e65d so the reconnect changes are evaluated against the current release and package metadata.
Resolve the mcp-client package.json conflict by preserving master's publishable-package metadata and workspace peer ranges while retaining the PR's direct @deepseek-ai/dsh-timeout peer and development dependency. The reconnect implementation therefore keeps its explicit runtime contract without discarding current-master release configuration.
