The session-title snapshot exposed a real activation race: ui-trajectory
and ui-question register into conversation-declared slots but only
injected 'slots', so nothing ordered their applies after ui-conversation's
— register() into the undeclared slot threw and the entry FAILED. Both now
inject 'conversation' as an ordering edge (documented as such; specs stub
the service where the bench declares the slot itself).
Review-bot findings, all three applied: the module loader's load sink
cross-checks the handoff id against the arriving row (a mis-stamped bundle
can no longer register under another entry's identity); the default
execute seam removes the inline script node right after its synchronous
execution (repeated HMR rebuilds no longer accumulate dead nodes); a
throwing onRebuilt subscriber is contained per-listener and routed to
onError instead of escaping the fs.watchFile callback.
ui-question landed on master as a full dshClient plugin package (composer
question flow); it enters the flat roster, apps/cli deps/refs, and the
smoke graph. Restore the workspace-context and user-interaction host deps
the conflict resolution had dropped.
client-hmr is a normal plugin package composed into dev graphs only. It
listens on /plugins/events and reloads one plugin per rebuilt frame,
serialized: invalidate, prefetch (fresh factory registers while the old
fiber still serves), registry.delete before touching the fiber, drain
disposers, drop owned style tags, entry.refresh(), fiber.await() loud.
Dependency cascade costs zero client code — fiber activation epochs
re-load dependents through cordis itself. Reload is coarse by design;
no rollback in v1; self-reload works with a frame gap the next rebuild
heals.
scripts/dev-web.ts (pnpm run dev:web) is the convenience watch-build:
it discovers its package list by scanning packages/*/*/package.json for
dshClient platform "web" at startup — no hardcoded roster — and talks
no protocol to the host. Gate bookkeeping rides along: knip entries for
the new packages, README model-experience allowlist rows.
The registry scans mounted Loader entries' dshClient declarations and
composes __DSH_BOOT__ {rev, entries} — inject edges and the immediately
mark come from manifests, never hand-copied; malformed fields fail loud
at load. The composing app owns one flat roster plus the --dev switch
(hmr row and bundle watching are dev-graph decisions).
The rebuild signal is the webserver's own observation: in dev mode the
registry stat-polls each scanned bundle (fs.watchFile; polling because
network mounts deliver no inotify), re-hashes on change, and broadcasts
a rebuilt frame on the /plugins/events SSE channel only when the rev
actually changed. Watch membership follows the table across rescans;
dispose drops all watches; a torn read self-heals on the next tick.
The POST /plugins/rebuilt endpoint is gone — builders and the host
share zero protocol. dsh web --dev logs the watched bundle list and
each rebuilt id with its rev transition.
Every client plugin package carries dshClient ({platform, inject,
immediately?}) and emits lib/client.js through the shared clientBundle
preset; exports["./client"] points at the bundle. The infrastructure
tier (connection, runtime, ui-theme, i18n, hmr) declares immediately: true
in its manifest — absent means lazy. The bundle purity gate covers all
nine packages: platform modules stay external, INLINE_SAFE wire layers
inline, any other cross-plugin value import is a build error. Migrations
that rule forced: scopeOf became a SessionsService method and
transportError moved into dsh-host-apiproxy's wire layer; the store
engine stays in runtime under a documented temporary exemption
(TODO(webload/store-rehome)).
The module system moves out of dsh-client-runtime (./loader retired) into
its own package: a lazy CJS table where executing a bundle only registers
its factory and materialization happens at first require, memoized, with
recursive requires self-ordering. ClientModuleSystem is a class; index.ts
keeps the types and a thin factory. Boot is two-phase: phase one prefetches
the immediately tier in parallel (registration only, failures deferred to
phase two's loud import); phase two mounts the vendored Loader with the
module system as internal, creates one entry per graph row plus the
app-shell pseudo-row the kernel appends itself, and settles on an
all-ACTIVE sweep. The shell kernel is self-sufficient: hand-rolled
loader-status stores, no plugin value imports, platform seed list single-
sourced in platform.ts.
- The real-browser collapse smoke tracked the old chrome: visible HARNESS
text (the wordmark svg is aria-hidden now), an 'Expand sidebar' label
(renamed 'Open sidebar'), a 300px settle (default is 280), and an
immediate focus assert (rail search defers focus past the slide). The
case now tracks the brand span, polls the deferred focus, and uses the
current labels and width.
- Tooltip treated hover and focus as one trigger: leaving with the mouse
dropped the bubble of a still-focused anchor (and vice versa). The two
triggers are tracked independently; the bubble hides only after both
clear. Spec pins both orders.
- The ui-sidebar README and the bilingual collapse note still described
the retired geometry morph; both now state the slide + crossfade
contract, the fixed-width (never-conceding) sidebar, and the rail's
whale-mark/tooltip chrome.
- tui.spec: exercise a goal-sourced injected context card (labels by
source kind, not plugin name), closing the last uncovered branch in
tui/src/index.ts that the CI coverage gate caught.
- time-context.e2e / goal.e2e: filter injected context by source now
that it is a user/message (plugin/goal source), and count goal
continuation rounds by round>0 rather than event type.
Address a second fresh-eye review of the review fixes:
- MAJOR: late steering that lands after runTurn returns terminally
stopped (e.g. during the post-turn flush) was drained by runLoop and
dropped without a discard, leaving a dangling outstanding id the
negative-only invariant can't catch. Emit agent/inbox/discard for it,
symmetric with the in-turn terminal-stop drop.
- remove the dead cancel() idle-settle branch: whenIdle's fast path
already resolves for a lone quiet item, so no waiter is ever left for
it to settle. Document why.
- gen-cordis-api classShape now drops private/protected/#private members
and strips getter/setter bodies, so Session no longer leaks private
fields and getter bodies into the model catalog.
- document that AgentMessage intentionally omits meta (durable-only).
Adds a regression test for the late-steering discard.
Resolve six review findings on the unified-send change:
- quiet (wakeup:false) queued items no longer un-park the driver; the
inbox distinguishes hasWakingQueued (drives the loop, idle/quiescence)
from hasQueued (anything to dequeue), so a lone quiet item parks at idle
and rides the next waking send. whenIdle/cancel settle off the waking
signal, so cancelling a parked quiet item no longer hangs whenIdle.
- SendOptions.meta on queued/steering sends now reaches the durable
user/message and steering/message (was dropped except on injection).
- a terminal agent/turn-stop that drops pending steering emits
agent/inbox/discard so the enqueue-dequeue-or-discard ledger balances.
- the loop-authored continuation reason is snapshotted and frozen like a
public send.
- gen-cordis-api collects exported classes (body-stripped) so the now-
abstract-class Agent and its transitive shapes reappear in the API
catalog.
Adds regression tests for each and re-records the affected snapshot.
Add a branded AgentMessageId assigned to each accepted send message and
returned from send/followup/steer/inject (was void). Rename the inbox
event payload InboxItemInfo to AgentMessage, carrying that id so a caller
can correlate a queued item with its enqueue/dequeue/discard events.
Hypatia May
imccyu
Tianyi Cui
Turtle
fz
_Kerman