Session format v2 embeds each attempt's compact stream in
assistant/message and assistant/attempt, but Host and client consumers
still expanded it into per-member TimedStreamChunk arrays and did
per-member work; expandAssistantStream materializes the full array before
find/toReversed/break can answer. Session Stats (the projection phase of
every Session open), the token meter's usage and provider-assembly folds,
the subagent output fold, and the Session Controller image lookup still
paid O(members) allocation and time per settlement.
The Chat and Trajectory definitions were already settled from
message.content on master; the remaining per-member folds stay.
dsh-llm now exports record-level readers (first token, visible content,
visible text, last raw chunk of a type, raw chunks of a type, joined
text, run-aware assembly, per-run first-token/first-visible times) that
scan the compact records once with early exit. Session Stats reads
assistantStreamFirstTokenTime, the token meter reads
lastAssistantStreamChunk(stream, 'usage') and assembles through
assembleAssistantStream, the subagent output fold appends
joinAssistantStreamText, and the Session Controller scans
assistantStreamChunks(stream, 'block-end').
expandAssistantStream is deliberately not memoized: retaining expansions
costs roughly ten times the compact stream for the Session's lifetime.
It remains the validating path at durable boundaries.
Synthetic 200-turn v0 migration benchmark, median of five: first-open
projection 28.0 ms -> 5.4 ms, first-open total 76.9 -> 50.0 ms, peak RSS
137.2 -> 94.9 MB; reopen projection 17.8 -> 5.6 ms; all phase budgets and
the 128 MB heap constraint keep passing.
Run the dsh-archive-agent-notes audit over every active Agent Note on
current master, judging each record by whether its rationale still guides
work rather than by size or age.
- Archive 453 implemented bilingual triplets (417,882 English words):
completed UI chrome, narrow adapters, closed bug fixes, implementation
walkthroughs whose package READMEs, docs pages, generators, or successor
notes now carry the useful behavior, and 51 records fully superseded by
a later active note. Keep 201 implemented notes whose ownership rules,
negative guarantees, durable or wire semantics, security rules,
reintroduction conditions, or still-tempting rejected alternatives
remain useful.
- Reject 7 proposals whose premise is gone or whose work shipped in
amended form under other records; delete 2 rejected notes that no
longer prevent a plausible mistake.
- Retarget every remaining inbound link to the archived path, and repair
active prose that named an archived record as the owner of a live fact:
parenthetical citations drop, ownership sentences redirect to the
README, docs page, or active note that states the fact, and history
citations say so. Chinese files link the English archived path because
the pairing gate treats the frozen tree as outside the bilingual corpus.
- Seal 1,359 new frozen artifacts; existing seals are unchanged and
outbound links from archived notes are neither inspected nor repaired.
- Regenerate docs/config-catalog.md after the hook-bridge comment edits
shifted two source line numbers.
Three review follow-ups on the consolidated jobs seam.
The spill file a pull source keeps is job metadata now: the pump reports it
to the registry on every read (present or withdrawn), the projection lists
it under output.spillPaths, and the model's dropped-output notice names it
from there. Attached to a gap chunk it was lost whenever the ring itself
evicted (a pump that kept up never produced a gap) or whenever the gap chunk
was evicted later, and it could not follow a source that withdrew its file.
A job_output wait's claim on the settlement now carries a live flag that
the caller's abort clears synchronously from the abort event. The registry
rejects an aborted wait on a later microtask, so a settlement announced in
between used to find the claim counted and skip the notice the tool result
was not going to carry either.
The job-controller client entry and the tool-jobs entry stop exporting
implementation classes and render helpers nothing outside the packages
consumes; the client entry keeps IJobs, which also carries the ctx.jobs
Context augmentation to consumers.
The caller-bound operation set now says what it is: the jobs one caller
may operate on. The caller argument is required (SessionId | undefined),
so an anonymous view has to be requested explicitly instead of falling
out of a forgotten argument.
Write handles now hold a durable lease (session.lock.json beside the
log): a random owner token, diagnostic pid, and an expiry. Acquisition
wins by exclusive create; a second process's create or write open rejects
while the record is unrenewed for less than leaseTtlMs (default 5 min),
and takes over after that — a crashed holder is waited out, never
reclaimed by pid. The holder renews every leaseRenewIntervalMs (default
4 min); a renewal that finds a foreign, vanished, or expired record — or
fails outright — marks the lease lost permanently, so every later
append/flush rejects with SessionOwnershipLostError while reads continue.
Close releases the record; read handles never touch it. Takeover of an
expired record is eventually exclusive: a replaced holder stops within
one renewal interval.
Refs #3245
The settled-retention JSDoc, the required observed streams, and the session-controller type shifts changed what the config, persistence, Cordis, and event catalogs embed; CI's static and coverage lanes compared the stale copies. Regenerated, with the zh mirrors updated in step.
A rejected background spawn produces no process output, so the observed stderr reader the job registry pumps was empty and the model's job_output after such a failure showed only killed before exit, where master's consuming read rendered [stderr] followed by spawn failed: … with the cause. The local providers now serve the note as the whole observed stderr stream once the spawn rejected, and the consuming read still folds it in exactly once, so the model's first read after the failure is byte-identical to master's. ShellProcess.observed and both of its readers are required: every background process exposes offset readers, and the shell tools' pull sources no longer carry a no-readers branch.
The consuming shell read on master told the model where a truncated stream's complete text lived: [some output was dropped from memory; full output: <paths>]. The ring-backed read rendered a notice only for the ring's own eviction and ignored the gapBefore mark the pump puts on a chunk whose source lost bytes between two polls, so a fast producer's tail read as continuous output and no path survived. A lossy source read now carries the source's spillPath onto its gap chunk (JobSourceRead, JobAppendOptions, JobChunk), the pump forwards it, the ring retains it, and the model renderer emits master's notice for a producer-side gap or a ring eviction, listing the distinct spill files in offset order. The bash and pwsh sources already returned the subprocess reader's spillPath; the type now admits it.
Settlement trimmed the ring to settledRetainBytes before releasing waiters, so a job that finished before its first job_output handed the model only the last 16 KiB plus a dropped-output notice where master's consuming read kept the executor's whole tail. The ring now keeps every byte the model cursor has not consumed on top of the settled cap, and the first terminal model read trims to the cap.
The registered event was emitted after startPump, whose first drain runs synchronously and may announce output, so a source readable at once produced an output event for an id no subscriber had seen registered. The announcement now precedes the pump; the events JSDoc states that a job's first event is registered.
JobView.kind was typed as the merge-extensible JobKind, so the job-controller program, which compiles only the merges it can see, generated a strict codec of bash | subagent for job.rows and job.observe. The first pwsh, pty-send, or workflow job failed frame validation on both ends and the browser dropped the whole roster. The projection now carries kind as a string; JobSpec.kind keeps the Host-side JobKind. A built-artifact smoke parses a pty-send row through the generated schemas, and the seam spec pins the open type.
The `start` frame and the reconnect baseline attempt carried a wall-clock
`startedTime` that no Host or Client consumer read: the Host accumulator
and the Client reconciler copied it into their attempt maps and nothing
looked at it again. Presentation frames now carry only the coordinates
consumers use (attempt, revision, turn, step, chunk index, and v1 seq
provenance); the type-equivalence block, event graph, READMEs, and the
live-stream Agent Note follow the type.
- docs/subsystems/jobs.md (+zh) records JobSpec, JobHandle, JobHooks,
JobOutcome, the pull-source types, JobChunk, JobOutputRead, JobView,
JobRead, VisibleJobs, and the event types as type-equivalence blocks; the
manifest and the cordis catalog link map follow
- the jobs, jobs-local, and tool-jobs READMEs (+zh) describe the ring, the
caller-bound view, the event stream, and the notice ledger; the
architecture table and the adding-a-tool cookbook drop record: true and
readOutput
- Agent Note 2026-09-03-jobs-seam-consolidation owns the decision; the
2026-09-01 record note and the 2026-08-08 web display note carry
Superseded pointers and the display note links the rows spec
- regenerated catalogs (cordis, api, inspect, client, tool, config,
persistence, doc graphs, module graph) with their Chinese mirrors; the
cordis catalog drops the jobOutput exemption and the tool catalog mounts
the agent registry for tool-jobs; the client model imports JobView from
its leaf so the typert analyzer resolves it
- `JobStart` is a union discriminated by `record`: `RecordingJobStart.run`
receives `RecordingJob` (with `append`), `PlainJobStart.run` receives the
plain `RunningJob`, and the registry builds the plain face without the
method, so the former warn-and-drop path is unrepresentable.
- Remove the duplicated JSDoc opener in the registry seam and correct the
job-list slot comment (order 20 sits before the subagent catalog's 30).
- Restore the `cordis-tool-round` session fixture whose only change was a
re-recorded `createdAt`.
- Update the 2026-08-08 job-display note's stale service-disposal and slot
ordering facts; archive the superseded 2026-08-25 unified-task-list note
and point its remaining inbound links at the archived snapshot.
Review asked for the observation wire and its client service to leave
session-controller: a job should reach the Host only through its own
Remote namespace, and the session plugin's apply had to materialize
{ $stream, session } for a service that is not session state.
@deepseek-ai/dsh-api-job-controller now owns the Host JobController
(job.observe over ctx.jobs, requiring the agent and job registries) and
the client ClientJobOutput (ctx.jobOutput) over ClientJobOutputModel,
resolving both Remote faces in its own apply. The remotes assembly
mounts the job namespace; ui-jobs injects the new package; the wire
types drop their Session prefix and spell the status union out so the
client face pulls no Host declaration merges into browser programs.
session-controller keeps the jobsBySession roster and loses observe-job,
job-output, and the observe cadence config.
ui-chat's slots and snapshot-builder import CommandNode and friends from
the declaring package: the typert analyzer's cross-package rule trips on
the snapshot re-export as soon as another client face joins the root
Client program (the new controller does), which is why this branch
carried the change before. JobPumpSource documents why it is an offset
snapshot reader rather than a Node.js Readable.