`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
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
Machine-produced by `pnpm run rescope-vendor --apply` plus the regeneration it
prints: `pnpm install` for the lockfile, `pnpm run gen-third-party-notices`,
`verify-translation-pairing --write` for the touched bilingual pairs,
`gen-doc-graphs`, and one typert snapshot whose ids embed character offsets.
`pnpm run rescope-vendor --check` verifies the result.
Renames nine vendored packages (cordis, cosmokit, schemastery and the six
@cordisjs plugins) and every reference that resolves them: manifest names and
dependency keys, module specifiers including declare-module merges, cordis.yml
plugin names, tsconfig paths, every Markdown fence, and `docs/` prose.
Directory names, upstream versions, and dependency ranges are unchanged, so
vendor/README.md still reads as an upstream snapshot; its manifest table gains
an upstream-name column so THIRD_PARTY_NOTICES keeps MIT attribution pointed
at each fork's origin.
The tutorial tier follows the rename end to end: its yaml fences named plugins
the Loader can no longer resolve, its `ts ignore-check` fences disagreed with
the compiled fences beside them, and its prose quoted both. The contracts that
told readers to keep upstream names — the root convention and the vendoring
cookbook's tree comment and manifest invariant — now say to rescope instead.
Two rules read `@deepseek-ai/` as "another workspace plugin": the client bundle
purity gate now names the vendored libraries a browser bundle inlines, and the
files where a bare `cordis` is an agent-preset id keep that product data.
A restriction was compiled against the global tool layer alone: only
global-layer tools were tested against `admits()`, and every chain-layer
tool was overlaid unfiltered afterward. That read the exempt set as "the
global layer" when what it means is "what this scope registers itself" —
two descriptions of the same set only while every model-facing tool sat in
the host composition.
Moving those rows onto the agent plane separated them. A preset's tools are
an ANCESTOR contribution to a joined agent, so a subagent's `toolFilter`
stopped constraining anything it was given; and with the global layer empty
`restrict()` rejected every name it received as unknown, failing the child
outright. With the same tools in the global layer the filter still admits
and applies normally, which is what makes this a regression of the move
rather than a standing limitation.
`view()` now filters everything a scope inherits — the global layer and
every ancestor layer on its chain — and exempts only the layer the scope
owns. That exemption is load-bearing rather than incidental: the delegation
runtime registers a child's `report` and structured-output tools into the
child's own layer, and a filter naming the capabilities the child may use
must not strip the machinery it answers through. Tool order, and with it
prefix-cache reuse, is unchanged: inherited names keep their global-then-
ancestor position and own-layer names still come last.
The diagnostic said "unknown global tool" while listing what is really the
inherited surface; it now names the surface it checks and says why an
own-layer name is not restrictable.
Fixes#2185
The model reads a PNG/JPEG/WebP/GIF file, the bytes commit through the
durable attachment lifecycle, and the tool result carries the real
ImageBlock so the image enters context from the next request onward.
FileSystem gains a bounded readBytes primitive (local + E2B providers);
registration is conditional on the attachment store, and a strict
execution gate refuses routes that do not declare image input, so a
text route's durable history stays free of image blocks. llm-replay
models may declare inputModalities, letting keyless ACP snapshots pin
both the sha256-referenced success and the verbatim refusal.
Supersedes the withdrawn route-scoped design of PR #598; the decision
record is .agents/notes/implemented/feature/2026-08-10-minimal-read-image-tool.md.
modeFor read only the exact scope's layer, so a `code`-preset session
advertised the native catalog: the mode is declared on the preset's STANDING
scope, and the agent only parents to it. Nearest scope wins along the chain —
the mode decides what the model SEES, which is the class of fact the chain
inherits. Caught live (the model politely computed with bash and said
run_code was not in its list); the chain test pins it.
The client half of the label fix: the create echo and the session-added
frame's agentPreset now reach the session list (newest wins in the upsert —
every producer of the field reports the CURRENT composition), and a confirmed
blank-session switch publishes through the new ISessions.noteAgentPreset, so
the header label moves with the composition instead of waiting for a reload.
interrupt_agent(agent_id) passes the calling agent as the ancestor
authority for ctx.subagents.interrupt(); the service verifies live
registry identity and recorded lineage, so a direct child or deeper
descendant stops with the same generic parameter while send_message keeps
its exact-direct-parent authority.
Discovery: list_agents gains an optional scope. descendants walks the new
SubagentService.listDescendants() — one lineage trace flattened in stable
pre-order across ordinary and one-shot intermediates, each entry carrying
its verified parentId and depth — and every status now comes from the
live Agent registry (running/idle/complete).
Refs #1535
Code Mode was a deployment-wide field on the host `tools` row: a
deployment ran every session that way or none. The obvious product
shape — 代码模式 beside 标准/极简/创造 in the preset picker — had
nothing to hang on.
The registry itself cannot move into a preset; the agent loop's
scheduler, the api-proxy's presenters, and every tool plugin are its
consumers. So split the registry from its projection: `presentAs(mode)`
writes one cell on the calling agent's scope layer, exactly as
`restrict()` does, and the three reads that decided presentation take
that scope's mode instead of the service's. The config `mode` becomes
the default agents shadow rather than a process-wide fact.
Two consequences are load-bearing. `run_code` now enters a view only
for scopes whose own mode presents it — a native agent must not find it
dispatchable because another agent in the process does — and the
reserved name holds whatever the configured mode, since any agent may
select a code mode later.
`dsh-agent-tool-mode` is the row a preset carries to declare this. A
code mode waits for the host's `codeRuntime` rather than assuming it,
so a runtime-less deployment fails the preset at mount, naming the
row, instead of at the session's first request.
The shipped `code` preset is `standard` plus that row, ordered second.
"Two emitted positions" was followed by an enumeration counting
occurrences — the field twice, the class statement once — so the two
halves of the sentence disagreed. The sentence now states what needs the
tables without a count.
The file's convention is a \uXXXX escape for a character with no visible
width (\u0301, \u1100, \u1161, \ud800 are all written that way) and a
literal only for a visible one; the new case wrote nine raw ZWNJ.
The comment also named only the field as needing 15.1 tables. Two emitted
code positions do: the bare field, once in each class, and the
Tool\u200CbArgs class name. The subscript comment is not one.
The docstring names ZWNJ/ZWJ as a skew instance but nothing checked the
predicate's tail-position accept branch; this is its first test, and it
also covers camelCase's Tool-prefix branch for a head that is
XID_Continue but not XID_Start.
The equivalence sentence pinned its evidence to Node 22.23.1 against
CPython 3.9.6 without saying the samples sit inside those two versions'
shared tables, next to five named characters where that same pair
diverges.
`keeps a non-ASCII field name…` asserted that Code Mode omits the native
schemas so nothing else carries a dropped field's name, requiredness, and
type. That holds under `mode: 'code'` only; under `both` the native schemas
ship alongside the SDK, as the module header says. Earlier rounds swept
`py-types.ts` for this family and qualified five sites there; the spec was
never in scope, so this is the family's last unqualified member rather than
residue from those fixes.
The previous commit's `UNPRINTABLE` sentence said the raw-reach point for all
three characters is `pyScalar`'s `JSON.stringify`, and the test comment said
that route is the only one. Both are exclusive claims and both are false: the
subscript tool-name comment calls `JSON.stringify` itself, and a tool name
carrying NEL, LS, or PS always lands there, none of the three being
`XID_Continue`. `pyScalar`'s own docstring already recorded that inheritance,
so the file contradicted itself. Both sentences now name the two call sites.
The note's evaluation axis was introduced as "the `typing` names the block
spells", which excludes one of its own members (`A | B` is operator syntax)
and omitted PEP 585 builtin generics — `dict[str, Any]` and `list[…]` appear
in nearly every render and need 3.9. The axis is now "the names and syntax the
block would evaluate at definition time", enumerated 3.8 through 3.11.
The test title covered two of the three characters it asserts; NEL is NEXT
LINE, neither a line nor a paragraph separator.
Four non-blocking review suggestions, all prose plus one assertion.
`UNPRINTABLE`'s new sentence named three characters but only two raw-reach
points, leaving "and NEL?" open; it now says all three reach text through
`pyScalar`, and how the description path handles each. `pyScalar`'s
raw-pass-through list already covered NEL under "the C1 controls", and the
test now pins it alongside LS and PS, so the docstring's claim has a
mechanical check for every character it names.
The test title said "paragraph separators" for a pair whose first member is
LINE SEPARATOR. Two docstring paragraphs are reflowed to the file's ~80
columns after the earlier inserts left short lines.
The note's CPython-floor obligation gains a second axis: the `typing` names
the block spells (`TypedDict` 3.8, `NotRequired` 3.11, `A | B` annotations
3.10) are definition-time evaluation floors, not parse floors, so the floor PR
does not read "parseable on the supported range" as "executable on it".
Review read `JSON.stringify`'s raw pass-through of U+0085/U+2028/U+2029 as a
parse hazard: an LS in a `Literal[...]` value or in a `# tools["..."]` comment
would end the physical line and take the SDK block down. Measured on CPython
3.9.6 (Unicode 13.0) and 3.12.13 (15.0): all three are accepted in both a
string literal and a `#` comment, value round-tripping, and only LF and CR
terminate either. The set is the tokenizer's, not `str.splitlines()`'.
Both existing claims were accurate, so nothing changes behaviorally. Name the
distinction where it was assumed: `UNPRINTABLE`'s terminator sentence now says
which set it means, and `pyScalar`'s raw-pass-through list, previously "DEL and
the C1 controls", now also names LS/PS, which are neither. A test pins the raw
form for `const` and `enum` so escaping them later cannot land as a silent
divergence from the TypeScript flavor.
The `mode: 'code'` qualification landed on the module header and
`isBareIdentifier` but not on the other three sites the same reviews
enumerated. `UNPRINTABLE`'s "only declaration of the tools", the open-object
comment's "only signal that extra keys are accepted", and the docstring
comment's "only description of what a tool does" are each false under
`mode: 'both'`, where the native schemas ship alongside the SDK.
Widen the note's predicate-path sentence past head and last position: a
character added to `XID_Continue` passes `IDENTIFIER`'s trailing quantifier
anywhere after the head, the middle of a name included.
Record the ƛ test's table provenance. U+A7DC and the U+019B mapping to it
both arrive in Unicode 16.0, and the engines floor sits exactly there:
Node 22.19.0 reports Unicode 16.0 (ICU 77.1) and produces the mapping.
`camelCase`'s `toUpperCase()` is a fourth reader of the engine's Unicode
tables, on a table distinct from XID membership and with a wider window: a
tool named U+019B passes `isBareIdentifier` and compiles as `async def` on
CPython 3.9.6, but Node maps the head to U+A7DC and the declared
`class Args` fails there with `invalid non-printable character`. Record it
alongside the three XID read points in the renderer docs and in the note's
CPython-floor obligation, and pin the derivation with a test.
Correct three over-quantified sentences: a `camelCase`-derived class name is
evaluated for every tool but only reaches emitted text when some object shape
in the schema declares a `TypedDict`. Attribute the `str.isidentifier()`
equivalence to `IDENTIFIER` rather than to the predicate, which is
deliberately stricter, and restore the antecedent the mode qualification
dropped.
The predicate is not the only reader of the engine's XID tables. camelCase
reads them through its split set and its head test, and the class name it
derives is emitted for EVERY tool -- including one the predicate rejected,
whose TypedDict is still declared and named. A tool named `zz-` + U+1E4D0
never reaches the skew in the predicate, since the `-` rejects it outright,
yet still emits `class Zz<U+1E4D0>xArgs`, which CPython 3.9.6 refuses the
same way. A backend PR executing "pin the predicate against tables for the
floor" literally would leave that path open, so the note and the docstring
now name all three read points.
Two corrections in the same paragraph. The failing direction is a character
added to XID_Start OR XID_Continue -- one added only to the latter passes
the trailing `\p{XID_Continue}*` in a tail position and fails identically.
And the safe direction routes a name to the subscript/`dict[str, Any]`
path: a rejected FIELD name degrades its whole enclosing object rather than
just itself, which the predicate's opening paragraph already said.
Also qualify the module header's "ONLY source" claim, which holds under
`mode: 'code'` but not `both`, where wireSchemas ships every native schema
alongside the SDK section; record the measured str.isidentifier()
equivalence (21 samples, zero divergence, Node 22.23.1 vs CPython 3.9.6)
where the versions it is relative to already live; and attribute the
`FInd` spelling in the ligature test to full case mapping rather than to
the NFKC step, which is the identity there.
Two tests. The fold-collision half of the childClassName fix: sibling joins
that are byte-distinct before NFKC and equal after, so `usedClassNames`
dedupes by raw bytes and the counter only sees the collision because the
join is normalized. And the argument-side oneOf-of-objects branch naming,
which reaches the same childClassName path the output side already pins.
camelCase normalized `joined` and then prefixed, so the seam the `Tool`
prefix creates was never covered: `Tool` ends in `l`, a combining-mark
head composes with it, and a name headed by U+0301 was emitted as
`Tool` + U+0301 while CPython compiles `Too` + U+013A. childClassName
has the same shape -- both sides separately NFKC-stable, their join not:
a base ending in a Hangul L jamo or LV syllable composes with a V or T
jamo head. Beyond the declared-name/compiled-symbol mismatch, two
byte-distinct names can fold onto one, and usedClassNames dedupes by raw
bytes, so the collision counter never sees it. Normalize after the
prefix decision and at the join, before the cap. The remaining joins
need nothing: `Args`/`Output` and the digit suffix cannot compose
backwards.
Also record the Unicode-table skew. The predicate reads the engine's
tables (Node 22.23.1: 17.0) and the interpreter reads its own (CPython
3.9.6: 13.0.0), so an interpreter older than the engine takes a bare
name its tokenizer refuses -- U+1C89, U+10570, U+1E290 and U+1E4D0 are
accepted here and rejected there. The other direction only degrades a
legal name to subscript. Closing it needs the CPython floor, which the
backend PR owns; state the asymmetry in the docstring and make the
decision an explicit obligation in the note.
The identifier test was ASCII-only, so an object with a `路径` field
degraded to dict[str, Any] -- dropping every sibling field's name,
requiredness and type, with no native schema behind it in Code Mode to
carry them. Python identifiers are `xid_start xid_continue*`, so match
that instead, and widen camelCase's split and head check to the same
sets (naming `_` explicitly in the split, since it is XID_Continue).
NFKC stability is a second and separate condition. CPython normalizes
identifiers at compile time while a JSON key is compared as written, so
a U+FB01 ligature key would be declared and reachable under its ASCII
expansion, a key the tool never accepts, and two keys that normalize
together would collapse
into one declaration. Those names take the subscript path. Generated
class names are normalized instead of rejected -- they are never matched
against a key. Astral characters can now reach the class-name cap, whose
slice counts UTF-16 code units, so drop a split surrogate half.
Also fix two comment claims. The note said one projection reads the
runtime twice per tool; the language-aware getters are installed on
run_code's own definition, so it is twice, both for that schema. And the
182-bracket site's reachability is an array reached from the root
through oneOf arms alone -- a union spine of any depth, not just one
root union; an object ancestor restarts the chain at the 181 site.
pyScalar's docstring attributed the rejection of a String-spelled
beyond-safe-range integer to "the Python runtime". No published backend
makes that call on this base. The fact that does not depend on one: the
padded digits name an integer no double holds, and passing it back would
have to cross the argument boundary as a JSON number. Say that, and say
why String rounds at all -- Number::toString is shortest round-trip, so
2 ** 60 emits the 16 digits that re-read to the same double and pads.
Mirror both in the test comment.
The note's Decision sentence said a test covers the flavor guard through
ctx.tools.schemas(). The test reads the definition's getter directly,
under a language absent from both tables; schemas() reaches the same
getter but has no assertion. Name what is read, and record that a
renderer-without-flavor language is drift this guards against rather
than an existing input -- the two key sets are identical today.
"Reachable only through a raw register() whose parameters is
array-rooted" was too narrow. A root oneOf reaches the same 182: the
union arm propagates listDepth unchanged because `A | B` opens no
bracket, so an array branch starts its chain at 0 exactly as an array
root does. Say "root opens an array chain -- rooted at the array, or at
an array branch of a root oneOf", in the JSDoc and the test comment,
and assert the union shape alongside the array-rooted one.
The note's Decision paragraph said the flavor guard is reached under
"a language that has a renderer but no flavor entry, and a test covers
it". The test uses ruby, absent from both tables, and the mechanism is
that schemas() reaches run_code's getters without passing
requireCodeRuntime -- so any language absent from the flavor table hits
it. State that instead.
The Consequences paragraph recorded the language-binding obligation as
two reads, assembly and execution. Within one projection there are more:
run_code's description and parameters getters each call
resolveFlavor(peekRuntime()) and schemaOf destructures both, so a reload
between them yields one schema whose halves name different languages.
"defineTool compiles an object root, so the annotation is a bare
TypedDict class name that opens nothing" is a false universal:
parameterSchemaSpecToJsonSchema compiles an OPEN object root, so an
empty parameter table and one with unrepresentable field names both
degrade to dict[str, Any], which opens one bracket. The conclusion the
sentence carries is unaffected -- 1 or 2 against a 182 cap -- so say
"a bare TypedDict class name or dict[str, Any], neither of which
carries a chain", in the JSDoc and the test comment that copied it.
pyScalar's docstring said the subscript tool-name comment quotes
"through the same call". It quotes through its own JSON.stringify call
site in renderToolsSdkPy and never reaches pyScalar, which only takes
const/enum scalars. Same function, different call site.
The mode-'both' test attributed assembly.tools to the public schemas().
That projection is wireSchemas, wired at ctx.systemPrompt.tools.
"the two code points CPython refuses" counted classes: NUL is one code
point, unpaired surrogates are the whole 2,048-wide D800-DFFF block. Say
kinds, in both the docstring and the test comment that mirrors it, and
restore the "odd" qualifier the test comment dropped -- an even trailing
backslash run does not eat the closing quote.
The soft-keyword test title still said "only special in statement
position", which the previous commit's own three-way split contradicts
for `case`: `case_block` is a clause head inside a `match` statement, not
a statement.
Add the case the subscript tool-name path lacked. A lone surrogate is
reachable in a name through JSON.parse of MCP wire JSON, and that path
has no UNPRINTABLE / LONE_SURROGATE fallback -- only the same ES2019
well-formed stringification the Literal path leans on.
The 182 the cap is chosen against had no direct case: the existing tests
cover the root chain and the TypedDict field, both of which start one
bracket lower. An array-rooted parameters schema reaches it from a plain
ToolSdkSchema literal, no raw register() needed. Exactly 180 arrays over
a const scalar is the worst case itself -- the root frame starts at
listDepth 0, so every list[ still emits and the innermost Literal[ is
reached rather than degraded; one deeper is where the item degrades.
Name the subscript tool-name comment in pyScalar's docstring: it quotes
through the same JSON.stringify call and inherits the same escapes and
the same pass-throughs.
