master extracted this layer's inline cold-resume resolver into
@deepseek-ai/dsh-api-remotes, whose `setup` was a fixed AgentSetup. A resumed
session composes the preset ITS header recorded, so the option becomes a
function of that header; the resolver builds the setup before the published
re-checks so those stay adjacent to `resume`.
Conflicts:
docs/cordis-catalog/services.md
docs/module-graph.md
packages/host/apiproxy/package.json
packages/host/apiproxy/src/api-proxy.ts
pnpm-lock.yaml
A default naming a route the Models page has since removed left the
composer saying 选择模型 while the input still accepted a message, which
then failed inside the adapter mid-turn.
`session.prompt` now refuses with `model-unavailable` before opening a
turn. That is the enforcement boundary: the method stays callable no
matter what a client disables. `session.models` reports the same fact as
`routable`, and ui-model pushes a block through the new
`ctx.conversation.blocks` registry so the bar renders the disabled
textarea it already renders without a workspace, carrying the blocker's
own reason. The push direction is forced — ui-model already depends on
ui-conversation, so ui-conversation cannot read it back.
The gate is `routable`, not "matches no advertised group": catalog
membership is advisory, so a route serving a model it stopped advertising
is missing from the groups yet perfectly usable, and `null` before the
first load never blocks so a slow Host cannot lock a working composer.
The scaffold gains a route-only adapter for fixture-less keyless
scenarios. Registering zero providers is a test artifact — every product
composition mounts one — and the goldens that froze the seat's fallback
label now show the model those scenarios actually route to.
A row's stored profile could not tell a hand-declared gateway from a
shipped provider whose models someone narrowed — both look identical from
outside the adapter — so the Models page had no way to mark the routes a
deployment added itself.
The directory entry now carries `declared`, answered by the owning adapter
against its own installed catalog, and the page renders a Custom tag from
it. Absence stays "this adapter draws no such distinction" rather than
"shipped", so a route no adapter claims is labelled neither way.
Also records the default-route work's Agent Note and the e2e evidence for
all three changes: the composer switch writing the section, and the Models
page declaring a route with its own reasoning effort.
The route a new session starts from was frozen into the gateway's
composition entry, so switching models in a conversation reached only that
conversation and every later session went back to the shipped default.
The gateway now owns an `api-gateway` settings section: the entry is the
base layer and the user document layers over it, so `session.selectModel`
records an accepted switch as the default for the next session. The write
is wholesale rather than a merge — switching to a model with no reasoning
effort has to clear a stored one — and a storage failure is reported
without undoing the switch, which already applies to its own session.
`targetFor` now resolves its tiers on every read instead of seeding once:
an explicit selection, else the session's own logged request header, else
the live default. That is what keeps a session that has run a turn deriving
its route from its log forever after, while a session still blank — New
Session reuses one rather than minting another — starts from a default
saved after it was created.
- subagents.history computes its projections best-effort on both arms
(a hostile unit's fold rejection serves the page without the block,
matching the session-list precedent) with dual-arm coverage
- the lifecycle-witness spec probes every field of the seven-key
witness, protecting the key list itself
- list-children's own module docs catch up with the seq-gate contract,
and the design note records the rung-two later-event window as an
accepted, self-healing deviation of the corruption class
None of these change behavior; each said something that was not true.
`SessionCwdConflict`'s doc block had been left stranded above the
`AgentPresetConflict` inserted under it, so one class carried a comment
about the other and the second carried none.
The roster comment named a `.system` directory that does not exist; the
shipped root is `config/agent-presets/`, and `system` is the trust its
entries carry.
The real-composition test attributed the disabled `api-gateway` row to
"side effects outside this process" alongside the port and the exporter.
It is disabled for a different reason — the api-proxy cannot mount in
this layer at all — and hiding that behind the same phrase would leave a
later layer unable to tell whether the line can come out.
One test claimed to refuse an adoption while asserting only that the
header records the preset; it now says what it checks.
`PERSONA_SECTION`/`PERSONA_ORDER` existed twice, once in the registry
that declares the slot and once restated in the row that replaces it —
a drift that would land a preset's persona beside the deployment's
instead of shadowing it. The registry exports them now.
The preset conflict message read "already runs agent preset undefined"
for a session that records none, which is the shape a deployment with no
roster produces; it names that case instead, with the regression that
reaches it through the gateway.
Finally, `PresetTree.write()` drops the `loader/config-update` the
inherited method emits — recorded on the override, since a future
edit-while-running flow needs its own persistence path.
A preset publishes its services behind `isolate` realms, which is what
makes them per session — and what makes them invisible to every host
context. The api-proxy kept reading the root realm, so requests that are
ABOUT a session but arrive from outside it answered for a singleton that
no longer exists: `goal.pause`/`clear` and `skill.list` returned "this
deployment does not mount @deepseek-ai/dsh-goal / dsh-skill" for sessions
whose composition mounts exactly that. Verified against a running host
before and after.
`agentPresets.serviceFor(agent, name)` addresses the instance instead,
reading the same subtree-ownership relation `leakedServices` already
uses, inverted. It is read addressing for a caller holding the agent: a
host row that `inject`s a service cannot use it, because injection
resolves before any session exists — which is why `tools` and
`subagents` stay host-plane and this is not a way around that.
Tool presenters had the same shape and the same cure: `viewFor` looked
definitions up without a scope while the global layer is empty by
design, so every card degraded to the generic renderer. It now takes the
owning agent.
Cold resume through `agentFor()` mounted no preset at all, so every
generic entry point — prompt, models, commands — rebuilt a restarted
session on host tools and the deployment persona. It composes the
recorded preset now, as the other resume path already did.
Two consequences of moving the agent plane behind presets, both invisible
until the host plane stopped carrying model-facing rows.
`sessions.fork` built its child with a bare `installTarget` and a `meta`
without `agentPreset`. That was harmless while every tool sat in the host
plane — the child inherited them for free. It now comes up with an EMPTY
tool set. The child composes the parent's preset instead, for the same
reason a resumed session keeps its own: the seeded history was produced
under those tools.
`bashEnv` lives in its own `dsh-bash-env` row rather than inside
`tool-bash`, so a preset that isolates the realm must compose the provider
beside its consumer; the host row is disabled here like every other
model-facing one. Nothing outside the agent plane injects `bashEnv`, so it
stays per-session.
The Web overlay disables base's 32 agent-plane rows and mounts the preset
roster instead, so each session composes its own tools and prompt rather than
sharing one process-wide set. The TUI keeps base unchanged: it is single-session
and composing its agent process-wide is correct there.
`roots` is patched in by AppCLIEntry, like `distIndex`: the shipped presets sit
beside the composition that names them and the user's live under the Harness
home, neither of which a config author chooses.
A session's preset is fixed at creation. Naming a different one for an existing
identity is `agent-preset-conflict` rather than a switch, because that
session's history was produced under the first preset's tools. The guard sits
after `await creation`, beside the cwd check, so it covers every path that
yields a live agent — freshly created, adopted live, resumed, or recovered by
the concurrent-creation catch. A request naming no preset adopts the session as
it is, keeping reconnect and retry ordinary.
Two bugs the real-composition test caught, both invisible to unit tests:
`PresetTree` now refuses to write. The Loader persists a tree whose plugin
self-disposed, and tearing an agent down disposes its whole subtree — inherited,
that rewrote the shipped composition, truncating a 241-line preset to `[]` the
first time a session ended.
`dsh-tool-skill` compared against a lookup of its own name in the global layer,
so it threw inside any preset: `register()` files into the calling context's
scope. It now compares against the definition it registered, which is what the
identity check meant all along.
The `standard` catalog is asserted exactly, not spot-checked: a row that
registers into the wrong layer mounts cleanly and simply contributes nothing, so
an omission is this design's quietest failure. It matches the shipped TUI
catalog plus `glob`/`grep`, the pair that composition documents as
ripgrep-dependent.
Re-records `cordis-inspect-jsdoc`, whose rendered `SessionHeader` gains the
`agentPreset` field. `fs-glob-sampling` fails identically on pristine master
and is untouched here.
The browser e2e scaffold gains the roster fact AppCLIEntry supplies. `roots` is
resolved and patched in by the CLI entry, like `distIndex` on the webserver row,
and this lane boots the shipped tree without that entry — so it has to supply
the same fact or the roster resolves nothing and every session in the lane
composes an agent with no tools, no persona, and no token meter. Only the
shipped root: a developer's own `~/.dsh/.agent-presets` must not decide a golden. The
`cordis:group` builtin comes with it, exactly as `boot()` registers it, because
a preset resolving package names from its own directory cannot reach
`@cordisjs/plugin-group` by name.
The lane stays red through this layer and the next four for the reason stated
above — the api-proxy injects `subagents`, `workspace`, and `tools`, so
`api-gateway` cannot activate and the browser has no `/api` at all. It goes
green again in the layer that returns those registries to the host plane; this
change is what makes that layer's fix sufficient rather than partial.
`session.create` takes an optional `agentPreset`, and the host resolves it,
mounts it during pre-publication setup, and records the resolved id on the
session header so a later resume rebuilds the same agent.
Resolution happens BEFORE the session exists, not inside setup: the session
boundary snapshots `meta` before asynchronous setup begins, so an id discovered
during setup could never reach the header. Mounting still happens in setup,
where a failure rolls the whole creation back rather than publishing a session
whose capabilities are half-installed.
Resume ignores whatever the request names and rebuilds from the stored id. A
resumed session's history was produced under that composition; restoring a
different one would replay tool calls the model can no longer make.
`dsh-agent-presets` now throws `UnknownPresetError` / `PresetMountError` so
the host can tell a bad request from a broken preset — they become
`agent-preset-not-found` and `agent-preset-invalid`.
Ships the two built-in compositions (`standard`, `core-web`) and the persona
row that lets them differ in identity. Nothing mounts them yet: no roster is
configured, so `composeAgent` finds no service and every session keeps the host
composition. Wiring the roster and moving base's agent-plane rows behind it is
the next commit, so the switch happens atomically with a real-composition test.
subagents.history drops its session-query read: a live child serves its
in-memory events with the registry watermark snapshot, a cold child goes
through inspectServable with a detached fold, and the retired
SESSION_QUERY_* arms disappear while every wire face stays put. The
design note records the shared read source.
- the subagent projection's empty state becomes a serializable null
sentinel (undefined fields vanish in JSON push frames, leaving stale
identities in clients); consumers treat null and a missing key alike,
and cached sentinels defer to the authoritative refold
- the tool catalog generator mounts the projection registry for
list_agents; the shared unavailable-error copy goes catalog-neutral
- the design note's verification section records the new
subagent-diagnostic snapshot honestly, the 2026-07-22 note's test
inventory is rewritten to the current mechanism, and the session-store
error code is documented
- listChildren contains per-child projection faults on both ladder rungs
(any registered unit's fold/schema rejection maps to that child's corrupt
diagnostic) and pins the whole-enumeration listing-failure rethrow
- the base bundle mounts session-projection (web-app's own insert retired
to avoid the double mount); stale composition comment updated
- the shared projections-unavailable wire face is pinned across
list/history/prompt; retired session-query arms removed from the catalog
paths
- the design note records the unknown-parent semantics shift and the
fold-fault isolation rule
All agent/* and agent-loop/config-start-failed events take one payload
object carrying the agent subject; waterfall/serial payloads require a
signal and keep next as the final argument. PreStepContext and
RequestFailureContext are unfolded into payloads and retired.
goal/changed follows the same shape so agentEvents keeps its listener
error containment. ReactLoopAgent builds its scope carrier once in the
constructor. Regenerates scope resolvers, tool-cordis api catalog, and
docs catalogs; updates all affected listeners, tests, and the
core-data-structures docs (en + zh).
llm.discoverModels was reachable from any declared trusted host. The
method takes a caller-supplied baseURL and makes the host issue a GET to
it, then reports the status or the parsed body — so on a LAN deployment
an anonymous caller had a probe for whatever the host can reach and the
browser cannot, plus a path that carries a draft credential. The
PRIVILEGED_METHODS doc already states the rule this broke: trustedHosts
is a DNS-rebinding fence, not authentication, so the configuration plane
stays loopback-same-origin. It is in that set now, asserted both against
the hand-built fence and over real HTTP beside the catalog reads that
deliberately stay reachable.
supportsDiscovery and listModelDiscoveryNamespaces are gone. The field
was required on the wire and read by nobody: its own contract said a
surface should offer the action "instead of naming an adapter family it
would have to hardcode", while the surface hardcodes llm-pi-ai in two
places and gates the button on whether there is anything to probe. Its
shape did not fit the second caller either — the create card has no row
to read a per-row field from. Keeping a required field alive for a
consumer that may never arrive costs every producer and fixture a value
nobody consults, which is exactly how the fixtures drifted. The registry
that fed it had no other production consumer, so registration and
disposal are now observed through the offer itself.
The Agent Note claimed the key is never logged, which the wire schema
beside it already contradicts, and predated both the provider field and
the catalog-answer path. The two new public types pointed at core.md
without a type-equiv block or manifest entry, so the generated service
catalog named documentation that did not exist.
Clicking "fetch available models" on a built-in provider went to the
network. That is the wrong source: pi-ai's registry is the authoritative
list for its own providers, and it carries the context windows and output
caps a `GET /models` listing does not disclose. Asking api.deepseek.com
what DeepSeek serves is both slower and worse, and against an endpoint
that answers a different shape it failed outright.
Interrogation is still keyed by settings namespace — the provider being
added has no route — but the request may now name the route it is
editing. An adapter that already describes that route answers from what
it knows, needs no endpoint at all, and never touches the network; only a
route the catalog does not describe reaches the wire, and one naming no
endpoint is told to set one or enter its models by hand.
`ConfigurableProviderView` gained `supportsDiscovery` so a surface offers
the action where a namespace can answer instead of hardcoding an adapter
family.
Three narrower corrections ride along. Discovery no longer claims Azure
or Codex: Azure authenticates with an `api-key` header and an
`api-version` query despite its OpenAI lineage, and Codex uses OAuth, so
both reported an authentication failure as a provider with no models.
Cancellation during the body read escaped as the raw abort reason rather
than a coded ABORTED. And the schema comment claiming the probe key is
never logged overstated it: the host neither stores nor returns it, but
it rides the client's outgoing envelope like every other secret-bearing
payload, and redacting that tap is a configuration-plane-wide change.
Once a pi-ai route became a declaration rather than a catalog lookup,
adding an OpenAI-compatible gateway meant knowing its model ids up
front. Most such endpoints publish that list at `GET /models`, but no
seam operation could ask: every one is keyed by a registered provider
route, and the provider being added has no route, no stored profile,
and no stored credential — the endpoint and key are values in a form.
Interrogation is therefore keyed by settings namespace, which a
configuration surface already holds from the configurable-provider
directory. `registerModelDiscovery` offers it per namespace,
`discoverModels` asks, and the request carries the draft itself. The
reply is candidates, not a catalog: every field but the id is optional
because most listings disclose nothing else, and adopting one is a
settings write like any other. Nothing here reads or writes settings or
credentials, so `settings.yaml` still decides what a route serves.
`llm.discoverModels` carries the same draft over the wire. Its apiKey is
the third and last payload a secret may ride, and it is never stored,
logged, or echoed; every refusal folds into `model-discovery-failed`,
naming the endpoint asked but never the credential offered.
The pi-ai side is a plain GET for OpenAI-compatible protocols only —
their listing shape is the one gateways, self-hosted servers, and the
official endpoints agree on. Others say so, sending the user to
hand-entry rather than reporting a guessed shape as an empty provider.
The reply is read under a four-megabyte ceiling held on the bytes
actually received, because the endpoint is a URL the user typed.
next-step items now split by origin: user-origin messages keep the
'steering' placement, while agent.inject context (approval notices, task
completion notices, attached snapshots) carries a new 'context' placement
that no surface renders until it is claimed as a durable user/message
context card. Widen the placement unions on the wire and in the client
runtime, add a projection test case, and document the split in the
apiproxy/ui-conversation READMEs and the web-steer chrome agent note.
Steer, inject, and followup now land as durable user/message events on the
session surface; the steering/message event type and its ConversationNode
kind are removed from the client projection. Update tests, docs, generated
catalogs, and agent notes to match, and align the steering e2e fixture and
prompt inventory assertions with the durable user/message landing.
The live fast-path fence and the raced-collision catch duplicated the same
subagent-ownership classification, tripping the duplication gate. Extract
`fencedLiveAgent` so both paths resolve one live identity through the
fence identically.
The raced-collision catch mirrored only the subagent-owned half of
ensureSession's `.catch`: a concurrent plain-agent publish winning the
identity still fell through to `internal`, where ensureSession returns
the winner. Mirror in full — classify a subagent-owned winner as
`agent-busy`, return a clean plain-agent winner directly.
The commands entry's inline comment described the old routing shape
("clients only send a sessionId for a published session") without the
ownership fence that agentFor now applies on every path — the fence's
contract home is the api/commands.ts module JSDoc, so trim the duplicate
and point at the routing shape only, keeping one home per fact.
Explicit-id adoption of a cold session-backed subagent under a *different*
cwd answered `session-conflict` because the cwd check ran before the
persistence inspection classified the identity. The api/commands.ts
contract states explicit-id `session.create` adoption rejects
session-backed subagents with `agent-busy` — ownership is an identity
property, so it must win regardless of the requested workspace.
Reorder the stored-session branch to inspect and classify ownership
first, then enforce the cwd match, making the response match the
documented contract.
When a generic `agentFor` cold resume loses the identity to a parent's
concurrent `enter()` — the collision rejection arrives from
`ctx.agents.resume` publication after the pre-resume re-check — the error
fell through to the `internal` mapping. Clients retrying then see a
transient-looking internal failure instead of the stable ownership error
that `ensureSession`'s `.catch` already produces for the exact same
published-winner case.
Mirror that re-classification in `agentFor`'s resume error path: after the
typed errors, re-check the registry and attached store and answer
`agent-busy` when the raced winner is subagent-owned. Adds a regression
test whose resume mock publishes the subagent winner before throwing the
ID-collision error.
`agentFor` fenced subagent ownership through the attached session store
(`ctx.sessions.get`) and only then returned a live registered agent. A
registered agent whose session is ever absent from the attached store —
an invariant nothing in this package guarantees — would therefore be
handed out through generic Host routing unfenced, bypassing subagent
delivery entirely.
Fence `live.session` directly whenever a live agent exists, and keep the
attached-store check only for the not-live durable classification.
`ensureSession`'s race `.catch` already fences `live.session`; this makes
the fast path the same check instead of an asymmetric weaker one.
`hasSubagentDescriptor` sliced the whole own-suffix events array on every
Agent-bound RPC — including each `session.prompt` and `sessions.models`
call on long transcripts — and `ensureSession` rescans the same suffix
after creation. Replace the slice-then-some with an indexed loop from the
seed boundary, so the classification is a plain O(suffix) read with no
allocation.
Yichen Jiang
imccyu
_Kerman
imccyu
Tianyi Cui
kingwl