Proposed-lifecycle Agent Note (Chinese draft, English body to follow
after review) recording the storage hub / domain form / workspace
design, the deletion semantics deferred to future work, alternatives
considered, and the session-backend migration reuse audit. The dev plan
under missions/ carries the engineering breakdown.
ctx.workspace registry owns WorkspaceId-branded records: realpath-
normalized unique paths (create rejects collisions; resolveByPath shares
the normalization), ordered sessionIds as the single source of ownership
truth, attachSession gated on the session header cwd matching the
workspace path (double-booking structurally impossible), dead session
ids filtered on projection and pruned on the next mutate, status()
reporting missing directories. No delete surface this phase — deletion
ships together with the session-side primitives as future work.
ctx.storage.domain opens declared domains: zod value schemas parsed at
the durable boundary, one write chain per domain (update(fn) is the only
read-modify-write), domain/changed emitted per record after durability
(new snapshot + operation, no old value, per repo event convention).
Domain-to-backend routing is configuration (default backend + per-domain
overrides); unknown names and missing facets fail loud. Ships the
MemoryStorageBackend test helper and a runtime invariant asserting every
change event matches the in-memory state.
node:sqlite DatabaseSync with the session-persistence-sqlite open
sequence (0o700 dir, exclusive 0o600 create, foreign_keys, configurable
journal mode, user_version stamp-or-reject). STRICT tables throughout:
units/unit_globals meta tables plus one document-per-row table per
declared unit table, keeping per-key durable updates precise.
Atomic whole-file replacement (same-dir temp + fsync + rename + parent
fsync); the in-memory unit state is authoritative and the file is always
the current net state, pretty-printed. Missing files open as empty units
and materialize on first write; foreign or unparsable files reject with
malformed-medium, stored-version drift with version-mismatch.
ctx.storage is a pure registration hub: multiple named backends stay
mounted side by side, data forms (domain first) mount via the
merge-extensible StorageForms map. src/backend.ts is the normative
KV-facet contract; tests/contract.ts is the shared conformance suite
every backend runs. Backends expose data-shape facets (kv now, an
append-log facet reserved for the future session-backend migration).
Review fixes (ds-review-bot on #623):
- Capture-at-delegation: the driver now reads overrideOf(parent.session)
for both knobs synchronously before its first await, and the prompt-submit
listener stamps those captured values — a parent switch racing the child's
asynchronous creation belongs to the parent's future, not the child. The
inheritOverride(parent, child) service method is split into its two halves
(overrideOf / stampOverride) accordingly.
- Veto safety: the one-shot prompt-submit listener registers with
prepend: true, so a veto-capable listener (a denying UserPromptSubmit
hook) cannot close the child's first turn without the durable stamp.
Both regressions are pinned red-first in inheritance.spec.ts: the
delegation-vs-late-switch race (delegate tool flips the caller wider while
the creation transaction is pending) and a blocking prompt-submit listener
(stamp survives a promptless first turn). Service contract tests renamed to
the split API; READMEs and the bilingual Agent Note updated.
Lint (bridge JSDoc params, service-class export shape, async invariant
listener form), regenerated doc catalogs/graphs with role classifications
for httpServer and clientModuleHost, catalog type-link exemptions for the
route/graph contracts, knip alignment (apps/cli composes via cordis.yml so
its yml-named deps are runtime edges knip cannot see; webserver's deleted
test dir), the zh side of the loading-model note brought along with its
pairing records, and coverage exclusions for the new web-transport halves
under the GUI test-lane TODO (real-composition harnesses land with that
lane).
New note records the dsh web composition decision (flat cordis.yml, the
AppCLIEntry/AppWebEntry class pair, one declaration place per config
source, the five-way transport split) with its rejected alternatives; the
client-plugin-loading note's roster endgame and HMR sections are brought
current with what shipped.
apps/cli/cordis.yml holds the whole composition flat — the host runtime
rows, the api-gateway row, the webserver row, and the ten dshClient rows.
AppCLIEntry is the pre-cordis glue: layered env (ambient > cwd .env >
$DSH_HOME/.env, fixing DSH_HOME=... dsh web not finding its key), patch
composition from the three non-yml sources (profile json through the static
PROFILE_MAPPINGS table, CLI flags, the resolved frontend distIndex), the
Loader include boot (--dev appends the hmr row before the settle), and the
fail-loud triple (assertEntriesLoaded + installFailLoud + an all-ACTIVE
sweep for PENDING fibers). web.ts shrinks to argv parsing + the URL line.
The four free functions in boot.tsx become one kernel class holding what
must exist before cordis: the parsed BootManifest, the ClientModuleSystem
instance, and the loading-page handles. Context/Loader setup runs in
parallel with the immediately-tier prefetch, but entry creation awaits the
prefetch: materialization is tree.import's synchronous require, so
cross-package require edges (i18n -> runtime/client) need every
immediately-tier factory registered first — unbarriered creation raced
10-25% of boots. The kernel adopts the modules entry (writes the
__DSH_MODULES__ slot pre-cordis, creates the entry first, skips its graph
row), and provide('modules') now lives in the adoption apply. apps/web
drops its host-package edges (composition is apps/cli's job).
connection binds the web transport: it injects httpServer + apiProxy and
registers toFetchHandler(ctx.apiProxy) under the /api prefix (the node:http
to fetch bridge moves in from the webserver, keeping the res-close disconnect
detection and drain/close backpressure waits). hmr owns dev reload: a
stat-poll watch per graph row driven by clientModuleHost.onGraphChanged,
rebuilt(id) on content change, and the /plugins/events SSE route (GET/HEAD
guarded); frame types are single-sourced in events.ts shared by both halves.
The node half is ClientModuleHostService (ctx.clientModuleHost): it composes
the __DSH_BOOT__ graph by scanning loader entries for dshClient packages,
serves /plugins/<id>/client.js, taps the index render, and exposes
rebuilt/onRebuilt/onGraphChanged. Scanning is incremental per package — no
full-rescan path exists: internal/plugin marks the fiber's entry name dirty,
a flush reconciles each name against live entries, package metadata
(including negative verdicts) caches forever, and re-hashing is reachable
only through rebuilt(id). The browser half moves wholesale to the standard
./client export (ClientModuleSystem, parseBootManifest with the dual-view
BootManifest, and the adoption plugin face that reads the
window.__DSH_MODULES__ slot and provides ctx.modules).
HttpServerService provides ctx.httpServer: register(route) -> disposer
(duplicate patterns throw), tapIndex transforms in registration order, and
the bound port; matching is exact > longest prefix > static dist fallback
(403/405/SPA semantics preserved). The server listens on activation, answers
per-request failures with 400 + a log line instead of exiting the process,
and knows no harness concepts — the boot graph, bundle routes, SSE channel,
and /api prefix all moved to their owning plugins.
createApiProxy moves from dsh-host-runtime into dsh-host-apiproxy (the
dependency direction already pointed this way); the package now
default-exports ApiProxyService (config {provider, model}, provides
ctx.apiProxy) while staying transport-agnostic — it registers no routes.
Runtime keeps bootHost/startHost for the headless path with its import
re-anchored, and drops the mountWebPlugins roster mounting helper.
Per-session policy overrides (sandbox/mode, approval/policy) never crossed
the delegation boundary: a spawn child of a read-only-switched parent ran
under the wider deployment default, and a fork child missed any switch made
after its seed boundary — delegation was a bypass channel for a user's
tightening.
The in-process driver now snapshots the delegating parent's override chain
and stamps it onto the child through the canonical write paths
(SandboxPolicyService.inheritOverride / ApprovalService.inheritOverride),
anchored inside the child's first turn via a one-shot agent/prompt-submit
listener: turn-enclosed (durable), ahead of the first request (an inherited
'never' reaches the child's first system prompt), and positioned after any
stale fork-seed switch so the ordinary last-event-wins fold resolves it.
Only overrides are copied — an unswitched parent stamps nothing and the
child follows the live deployment default; both services are consumed
opportunistically, so compositions without them delegate unchanged. Nesting
composes by construction (each stamp folds the already-stamped parent log).
Evidence: inheritance.spec.ts drives scripted-model children into the real
dsh-fs-sandbox fence through the real write tool (disk-state + denial-marker
assertions; spawn, stale-seed fork, grandchild, escalation fail-closed, and
no-stamp guards), inheritOverride contract tests in both service suites, and
the recorded subagent-sandbox-inheritance ACP snapshot (read-only preset →
delegate → child denied, replayed keylessly).
See .agents/notes/implemented/feature/2026-07-25-subagent-policy-inheritance.md.
Second review round, both pinned red-first:
- Overlapping turn-boundary flush hints now JOIN the outstanding flush
promise (Promise.all) instead of displacing it: the SDK's
concurrent-flush guard resolves an overlapping forceFlush()
immediately, so retaining only the latest promise let shutdown()
proceed while the first export was still in flight — the same silent
drop the single-flush fix closed.
- Adoption replay contains failures per event, matching the firehose:
one rejected record is withheld fail-closed while the rest of the
historical log still hands off. Wrapping the whole loop let a single
failure silently skip the remainder on an already-adopted session.
Three review findings, each pinned by a red test first:
- The OTel backend retains the latest turn-boundary flush promise and
shutdown() awaits it before provider.shutdown(): the SDK's
concurrent-flush guard makes the shutdown-internal flush return early
while one is in flight, silently dropping everything enqueued after
the flush snapshot (including the coordinator's dispose-time shutdown
markers).
- The coordinator retires sessions from the adopted set on
session/disposed (mirroring the persistence coordinator): a long-lived
backend no longer retains closed sessions and their event logs, and
final unload no longer stamps shutdown markers for dead sessions.
- The exporter config passes through whole to OTLPLogExporter and its
type/JSDoc now advertise the full OTLPExporterNodeConfigBase shape:
rebuilding url/headers only silently dropped documented SDK options
(timeoutMillis, compression, keepAlive, ...).
Master's i18n batches added Chinese counterparts to ~50 docs this PR
edits in English. Bring each zh side along with the minimal edits
covering the en diff (recorded-hash diffs, not re-translations),
reunite the stream-workflow-progress pair under rejected/ with its
manifest entry, re-record all pairing hashes, and regenerate the
event/persistence/tool catalogs and doc graphs over the merged tree.
tests/otel.e2e.ts self-skipped on $DSH_OTLP_E2E_ENDPOINT, which nothing in
the repo sets — it never ran; the mock-collector wire spec and the keyless
Loader-composition e2e already cover the pipeline both ways.
The severityOf compact/end probe parsed another package's merged event
shape by string comparison — an untyped cross-package contract that breaks
silently — and its only consumer was the test's own stand-in declaration.
Unknown event types now uniformly fall through as info; outcome semantics
stay with the owning package.
The seam keeps the telemetry/redact scrubbing interface but ships no rules
of its own: the innermost next() passes records through unchanged, and
deployments mount their rules as waterfall listeners. As an SDK we cannot
know which patterns are secrets in a given deployment; a shipped list
invites false confidence while catching only known shapes, and false
positives would corrupt exported bodies. Mechanism stays with the seam,
policy moves to the deployment; both READMEs and the Agent Note state the
raw-export default plainly.
The loader-composition e2e now mounts a deployment-style rule fixture and
pins the same wire behavior: secret absent, placeholder present, canonical
log untouched.
Revive the reviewed session-telemetry packages from the closed
session-telemetry-otlp-rfc branch (PR #222/#231) on current master, renamed
to @deepseek-ai/dsh-session-telemetry{,-otel} (the SDK component-telemetry
package holds the dsh-telemetry name).
Delta over the branch version: every record now passes a telemetry/redact
waterfall between projection and emit() — the innermost next() applies a
non-configurable conservative credential-shape rule set, listeners stack
stricter rules, a throwing rule withholds the record fail-closed, and the
canonical log is never rewritten. This answers the export-side concern that
closed PR #222; the boundary axiom (our aspect ends at emit(); delivery is
the reporting SDK's) is unchanged, and the runtime-telemetry RFC's outbox /
readCommitted lane is recorded as deferred in the Agent Note.
Covered by seam/redact/OTel-wire unit tiers (100% per-file) and a keyless
Loader-composition e2e that boots the examples fixture against a mock OTLP
collector and pins redaction on the wire plus the untouched canonical log.
The dev bundle watch missed rebuilds that landed while the registry was
constructing: fs.watchFile captures its comparison baseline with an
ASYNCHRONOUS first stat, so a write racing that window is absorbed into
the baseline and never reported. Standalone repro missed 24/400 same-tick
rewrites; the CI flake in web-plugins.spec.ts ('watch mode: a bundle
content change re-hashes the row...') was exactly this — the spec writes
immediately after createHostWebPluginRegistry returns.
The watch now polls from one registry-owned setInterval against a stat
baseline the scan itself captures synchronously, stat-before-read: a
write landing between stat and read leaves the hash newer than the
baseline (next tick re-hashes to the same rev, no spurious notify); a
write landing after the read leaves the baseline older (next tick
detects and notifies). No blind window. The poll iterates the live
table, so rescans retarget the watch for free and dispose clears one
timer. Stress: real-registry same-tick rewrite 0/600 missed (was 1/300);
spec watch tests 0/50.
New regression test pins the same-tick-as-construction write. Its
rewrite deliberately differs in size from the seed: a same-millisecond
same-size rewrite is invisible to any mtime+size poll (coarse fs
timestamps) — a stat-polling limit, not this regression. Loading-model
Agent Note updated in both languages (pair re-recorded).
The pool selector is defense-in-depth only — pull_request executes the
PR's own workflow definition, so YAML cannot enforce runner trust. Make
the actual enforcement boundary explicit in the decision record:
org-side disabled forking (the public release is an isolated read-only
mirror under a separate org), with migration to a repo-restricted
org-level runner group with base-branch workflow pinning as a hard
gate before forking could ever be enabled.
The hosted 32-core runner is exclusive to one job, but the vm-backup
pool shares one 64-core VM across four runner instances; concurrent
PRs could stack 4×24 = 96 Vitest workers and re-trigger the documented
aggregate-contention failures in the timing-sensitive process suites.
Bound the self-hosted leg at 12 workers per job (48 host-wide fully
loaded) and keep 24 on the hosted leg, selected by the same expression
as the pool.
The row rendered a status only for error, so a call cancelled before
tool/result read as a completed plan update even though no todo/write
occurred. Non-ok states now ride the generic row's StateDot semantics
(ongoing dot while running, warning dot + 已中断 marker when interrupted);
the ok badge stays for settled successful updates.
Keep inherited child prompt markers bounded by the normal silence fallback. Stage web-plugin rescans atomically and retain missing watch state until a successful rebuild.
additionalProperties: false on the todo_write item schema is model-visible
(tool schemas ride the request header and the code-mode prompt types), so
the pinned ACP/headless expected outputs re-record. Keyless refresh; the
two locally-failing scenarios are this machine's known environment issues
(HOME-symlink cwd normalization, SQLite ExperimentalWarning), not the diff.
Sweep all remaining sources that still described coverage as an
enterprise 32-core job: the ci.yml jobs preamble, the three-job
paragraph of the larger-hosted-runners note, and the required-pool
sentence of the portable-recovery note — English and Chinese sides of
both notes, with their i18n pairing records re-recorded.
Keep the cache restore for the ephemeral hosted (untrusted-PR) leg where
it is a genuine speedup, gated by the same expression as the runs-on
pool selector; the self-hosted leg skips it and installs from the
persistent local store.
master gave 2026-06-29-todo-write-tool a Chinese counterpart; the English
side's web-consumer sentences now translate across (in-body links keep
their .md targets per the pairing contract) and the pair is re-recorded.
The todo-panel fake gains the time/callTime fields ToolResultNode now
requires.
imccyu
kingwl
Turtle
Yichen Jiang
Tianyi Cui
Chinesezjc