A business package imported the Remote assembly for `ctx.remote` and the
Connection plugin for the wire types it passes around. The assembly now
re-exports the carrier's Client-facing types, so a business package names one
package. The re-export is type-only: the carrier's runtime values keep their
own module edge, since inlining them here would duplicate the carrier inside
the assembly bundle.
Four surfaces that had no Remote assembly dependency declare one now.
The Gateway and the carrier each compiled both halves under one tsconfig, so
the Host aggregate built their browser faces — including the face that owns
`ctx.remote`, the most likely future consumer of a generated `/remote`
contribution. Both packages now expose a host and a client face, and each
aggregate references only its own; three modules the halves share appear in
both file lists, as api/remotes already does.
The two apps/web specs in the Host aggregate restate the conversation engine's
Context key format instead of importing the Client runtime for it. A drift
makes the key miss its rendered node, so the assertion fails loudly.
The Host aggregate now reaches one Client project, the carrier's host face,
which the Gateway's own dispatch face needs; no generated contribution is
reachable from it.
The browse and native backends were dual-face packages: a Node backend plus a
browser surface under one tsconfig that referenced Client packages. That put
Client projects — and through them the Client runtime — inside the Host
compiler aggregate, which builds before the generated Remote contributions
exist. Each browser half moves to its own Client package, and both backends
become Node-only.
The interaction is still one choice: the adaptive chooser mounts the backend
and its surface as a pair of Loader entries and tears both down in reverse, so
a resolved kind still swaps both faces. Compositions that pin an interaction
directly now pin the pair, and the chooser's runtime-string package list keeps
naming everything a composing app must resolve.
`CommandService.list` and `execute` carry the wire contract directly through
`@Remote`, and the Client assembly mounts the generated commands
contribution. The legacy API Proxy route, its schemas, the map rows, the
generated client methods and the fixture's command domain are removed, so the
catalog and the admission call have one owner again.
`Session.command()` keeps a result-shaped public face for parity with the
prompt, cancel and attachment neighbours it sits beside, and reads the
generated namespace through one `SessionRemotes` parameter. The Session
cluster declares that face against the owning business package rather than the
generated contribution: the Host compiler aggregate builds this package, and
it runs before any contribution is emitted.
Migrated calls lose the `title-invalid` class of protocol-only error codes and
report `internal`; no production caller branched on them.
Every generated Remote method now resolves to `RemoteResult<T>`: the Client
face folds a carrier failure, a transport throw and a rejected result payload
into one error branch, so no consumer wraps a call to recover them. Only
assembly faults still reject — a wrong argument count, an unmounted method, a
missing Context binder, an absent Connection.
`RemoteFailure.code` stays an open string because the closed RPC code union
lives in the carrier package, which already depends on type-meta; naming it
here would invert that edge.
The goal surface drops its own try/catch plus the structural probe that
guessed whether a thrown cause was an RPC failure, and reads the result
instead.
Absence crosses the wire as a missing field: an omitted argument and a void
or undefined result both arrive as an absent JSON member, and the wide RPC
result slot accepts a success response without a value. Parameters declared
optional stay optional in the generated consumer declaration, so a business
signature is never widened to `T | undefined` to suit the wire. The weak SRC
descriptor reads parameter names from a JavaScript signature and cannot see
optionality, so a source-launched Host accepts an absent field and the strict
LIB pass owns rejecting a genuinely missing required parameter.
Adding COLLAPSE_SECTION_ORDER and CODE_ONLY_INSTRUCTION shifted every
later line in packages/core/tools/src/index.ts by 14, and three generated
artifacts cite those lines: the subsystems cordis-surface region, the
event producer/consumer matrix, and the Agent Note pair record left
inconsistent by the cherry-pick resolution.
`collapses()` read `defaultMode`, so the collapse only applied when the
DEPLOYMENT was `code`. An agent handed `code` by an agent preset under a
native default announced `[run_code]` on the wire and still executed a
model-direct native call -- the bypass this collapse exists to close,
reopened for exactly the composition `dsh-agent-tool-mode` produces.
`modeFor(scope)` is the same resolution `wireSchemas` and the SDK section
already use, so presentation and execution cannot disagree, and a mode
inherited from a standing preset scope collapses like a declared one.
The per-agent and preset tests asserted only the wire, which is why the
regression passed them. They now assert through the executor: the body
never runs, the call resolves UNKNOWN_TOOL, and the native sibling beside
it still executes.
The executor collapse landed without telling the model it exists. Every
tool contributes its own guidance section naming its tool, none of them
qualify how that tool is reached, and they all render before the SDK
(orders 100-199 against SDK_SECTION_ORDER 150), so the prompt said "Use
the read tool" eleven times and never said only run_code is callable.
A real session shows the consequence: the model emitted a native call,
read `unknown tool "read"` for a tool the same prompt declares, and
concluded the deployment was inconsistent rather than routing through
run_code.
The registry now contributes `tools:code-only` at order 99 -- ahead of
the guidance band -- stating the rule, registered wherever `tools:sdk`
is and rendering empty outside an effective `code`. `both` renders it
empty because its native calls do execute, which is also why
both-mode-turn no longer shares code-mode-turn's expected prompt. The
denial itself now names the route back, since a bare UNKNOWN_TOOL for a
declared tool is what misled the model.
wireSchemas() already advertised only run_code under mode: 'code', but the
executor resolved every call through get(), which returns the full visible
map plus the reserved transport. A model could name a native tool directly
and bypass run_code entirely. Route the execution-path lookups through a
new private resolveExecution() that applies the mode collapse at the
operation boundary: model-direct calls under 'code' may only name run_code
(UNKNOWN_TOOL otherwise), while SDK sub-dispatches (parent token set) keep
every visible tool. get()/schemas() public semantics are unchanged.
The denial happens at createExecution, before the extensible policy
pipeline — pre-execute listeners, approval ask, and guards never observe
a call that is deterministically denied. A collapsed call honors the
pre-dispatch cancellation contract, routes aborted results through the
visible tool's finalizeContent, and captures the finalizer before
argument materialization.
Regenerated docs, catalogs, graphs, scoped events, re-recorded
translation pairs, and updated test assertions.
Fixes#1815
LegGasai noted that filtering prompt sections by tool:* prefix is a poor
heuristic: it conflates section naming convention with presentation
semantics and would incorrectly drop tool:structured_output. The executor
collapse already enforces the boundary — a model-direct native call is
rejected as UNKNOWN_TOOL regardless of what the prompt says, so filtering
the prompt adds no security and only risks losing useful guidance.
The tool:read/tool:bash/etc sections describe capability usage patterns
that apply to both native and code presentations, and keeping them does
not reopen the native direct-call path because the executor blocks it.
wireSchemas() already advertised only run_code under mode: 'code', but the
executor resolved every call through get(), which returns the full visible
map plus the reserved transport. A model could name a native tool directly
and bypass run_code entirely. Route the execution-path lookups through a
new private resolveExecution() that applies the mode collapse at the
operation boundary: model-direct calls under 'code' may only name run_code
(UNKNOWN_TOOL otherwise), while SDK sub-dispatches (parent token set) keep
every visible tool. get()/schemas() public semantics are unchanged.
The denial happens at createExecution, before the extensible policy
pipeline — pre-execute listeners, approval ask, and guards never observe
a call that is deterministically denied. A collapsed call honors the
pre-dispatch cancellation contract, routes aborted results through the
visible tool's finalizeContent, and captures the finalizer before
argument materialization.
Under code mode, a system-prompt/assemble listener filters out tool:*
guidance sections that told the model to call native tools directly.
The tools:sdk section and SDK types remain so programs can still use
all tools through run_code.
Regenerated docs, catalogs, graphs, scoped events, and re-recorded
translation pairs.
Fixes#1815
wireSchemas() already advertised only run_code under mode: 'code', but the
executor resolved every call through get(), which returns the full visible
map plus the reserved transport. A model could name a native tool directly
and bypass run_code entirely. Route the execution-path lookups through a
new private resolveExecution() that applies the mode collapse at the
operation boundary: model-direct calls under 'code' may only name run_code
(UNKNOWN_TOOL otherwise), while SDK sub-dispatches (parent token set) keep
every visible tool. get()/schemas() public semantics are unchanged.
The denial happens at createExecution, before the extensible policy
pipeline — pre-execute listeners, approval ask, and guards never observe
a call that is deterministically denied. A collapsed call honors the
pre-dispatch cancellation contract, routes aborted results through the
visible tool's finalizeContent, and captures the finalizer before
argument materialization.
Under code mode, a system-prompt/assemble listener filters out tool:*
guidance sections that told the model to call native tools directly.
The tools:sdk section and SDK types remain so programs can still use
all tools through run_code.
Fixes#1815
imccyu
Yichen Jiang
Chinesezjc
Tianyi Cui
Dudu-0223
Ziya
ZiyaZhang
Wenlu Wang