`USER_PRESET_DIR` was exported with no production consumer — only the new test
read it, and a test that imports the implementation constant cannot catch that
constant being wrong. It stays module-internal, and the test spells the segment
it expects.
Overlaps master on the preset e2e, the web scaffold, and the generated
config-catalog triplet; all merged textually, so the catalog is regenerated and
its pairing hashes re-recorded rather than trusted.
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.
CI captured `- text: Loading history…` where the golden holds the grandchild's
first message. The only wait before that capture is the offline banner, which
renders from the descriptor and says nothing about the transcript below it,
and `captureStableAria` calls two identical polls stable — two of which can
both be the loading placeholder on a loaded machine.
Waiting for the message the golden asserts makes the precondition match the
assertion. The race is not observable here (the history is already loaded when
the banner appears, so the wait returns in ~1ms), so this is reasoned from the
CI log and the helper's semantics rather than from a local reproduction; if
the history instead never arrives on CI, the lane now times out on the missing
message rather than committing a loading placeholder to the diff.
Empty `roots` stopped meaning "no roster" when the harness-home root became
a default, so the case that keeps the unjoined-agent diagnostic silent is the
one opting out of every root — which is what it already passes and what its
name now says.
Regenerates docs/config-catalog and re-records its pairing hashes: master's
side of that generated file merged textually, which leaves the recorded blob
hashes describing neither side's content.
The new default's whole claim is that a launcher configuring nothing still
finds a person's presets, and nothing asserted it through a real composition —
`user-root.spec.ts` plugs the service directly, and every other lane pins
`includeUserRoot: false` so the machine's home cannot decide a golden.
This lane boots the real base and web-app bundles with only the shipped root
patched in, points `$DSH_HOME` at a temp home before boot (the derived root is
resolved when the plugin is constructed), and asserts a preset placed there is
listed as `user`, reported healthy, authorable, and mountable into an agent
whose tool catalog it decides.
The service moved its "is a roster composed" reads to the derived root set;
the invariant companion still read `config.roots`. In the shape this change
exists for — an app configures nothing and the roster is the harness home
alone — that made the advisory warning fire while the fail-loud invariant
stayed silent, so an agent could address a model against an empty global
layer unchecked.
Both now read one source: `roots` exposes the resolved set, and the invariant
asks it. That decides the behavior deliberately rather than by omission — a
composition that mounts the roster now fails an unjoined agent whether its
roots were configured or derived, and `includeUserRoot: false` with no
configured roots is how a deployment keeps its agents on the host plane. Both
shapes are pinned; the derived-only case fails against the old predicate.
Three pieces of prose went stale with the first commit: the web-app bundle
comment still called the writable root an assembly fact patched in by
AppCLIEntry (removed in the profile-plugin-bundles refactor — `composeProfile`
owns it now, and only for the shipped root), and the shipped skill and its
Agent Note still called both roots "configuration". The README gains the
resolved-roster reader and the discoverable-but-undeletable preset a second
writable root produces.
`Config.roots` defaulted to `[]`, so a deployment that did not patch both roots
in got a roster with none — `dsh run` once booted exactly that and failed
resolving `standard`. The user root never needed an app: `<dshHome>/.agent-presets`
is the same place in every deployment, resolvable here the way
`dsh-skill-local` resolves `<dshHome>/skills`.
The roster now derives that root itself unless `includeUserRoot` is false, and
`apps/cli` supplies only the SHIPPED root, whose path an installed app alone
can resolve. The derived root is appended after every configured root, so a
shipped id still shadows a home directory claiming it and `writableRoot()`
still prefers an explicitly configured `user` root; the set is resolved once at
construction, because a root set that changed between a `list()` and the
`copy()` acting on its answer would author into a directory the caller never
saw.
Every test that pins an exact roster now says `includeUserRoot: false` — the
machine's real harness home must not decide what a golden or an assertion
counts.
The rewritten test tolerated the interleaving where a fast command
settles before the running turn's next pre-step claim. The notice is
then folded into a step whose scripted reply is final, the turn closes
with an empty next-step inbox, and the collection entries are never
reached — a real timeout, not a tolerated ordering.
The command now blocks on a sentinel the test creates only after the
agent has gone idle, so the wake is the only path that can deliver the
notice, and the test asserts exactly two turns.
Also apply the review's smaller points: key the wake budget by Agent
rather than object, register the budget-refill listener only under
wakeup delivery, pin the schema default and rejection like
reportDelivery does, record the retirement-window stranding as a Known
Limitation, and cross-link the partial supersession both ways.
A throwing producer cancel jumped to the force-fail branch before
`reported` was set, so `settle()` announced an unreported completion and
the default wakeup delivery started a model turn on an owner the host
was already destroying — the exact failure mode marking the record
reported exists to prevent.
Teardown claims the report before calling the producer, because that
decision does not depend on whether the producer's cancel succeeds.
Also reject a `maxConsecutiveWakes` that cannot bound anything: the
field exists to cap a runaway chain, and `Infinity` removed the cap
while a fraction never named a turn.
Correct the module JSDoc and the background-task runtime note, both of
which still promised that notices never wake an idle agent.
The previous pass over-corrected: presenting both roots as unknowable made the
guidance vaguer without making it truer. Every `dsh` launcher puts authored
presets under `$DSH_HOME/.agent-presets`, so the skill and persona state that
as the answer to "where do my presets live" and reserve the read-back for the
path an agent actually opens or edits, which `copy()` already reports.
Neither root is a property of the preset layer. `Config.roots` defaults to `[]`;
the two the shipped agent sees are patched in by the dsh CLI, whose user root is
`dshHomePath('.agent-presets')` and whose system root ships beside the install.
`writableRoot()` then takes the first `user` root from whatever configuration
supplied it.
The skill and the persona stated `${DSH_HOME:-$HOME/.dsh}/.agent-presets/<id>/`
as where authored presets live, which is only the default install's answer, and
no call reports either root: `authorable` says whether a writable one exists,
and `list()` cannot reveal a user root that is still empty. Both now keep that
path as what to tell a user asking where to look, and direct every path an agent
acts on to `list()` or `resolve()` — which is also what `copy()` already relies
on, since it chooses the writable root itself and reports the file it created.
#2201 landed three moves this branch sits on top of. `bindSettingsScope` became
`ctx.settingsScope.bind` on the Settings surface, so the section injects that
service and the Remote face instead of importing a free function; the
`SettingsScope` types moved into the runtime contract module, which is where
this branch's `base`, `user`, and `unset` now live; and the client
`credentials/changed` event became the forwarded owner event
`credentials/updated`, which the web-search card subscribes to through
`ctx.remote.$on`.
The generated module graph crossed mermaid's 1000-edge render guard with this
branch's 14 new edges. The guard doubles again, exactly as it did at 500 — the
graph counts real package edges, and trimming them to fit a renderer default
would make the document lie about the dependency set.
Review found two errors in the first pass.
The skill quoted `service "<name>" has been registered at <Owner>` as the
rejection for a service published into the root realm. That message only
appears when the loose name collides with one the host already supplies. A
preset publishing its OWN new service name reaches the root realm successfully
and is rejected by the mount audit instead, with `row(s) published
process-global service(s) [...]` — which is the shape a forgotten realm
actually takes. Both messages are now listed, verified by mounting one preset
of each kind.
The Agent Note claimed every harness package publishes exactly three file
patterns. 139 do; 28 also publish `lib/client.js`, 22 publish
`lib/types/**/*.js`, and `dsh-message-feedback` publishes `src`. The note now
states only the fact it needs: no package's `files` includes its README.
Also from review: the copy step no longer repeats the display name it already
set, the id rule spells out `[a-z0-9][a-z0-9-]*`, and the sandbox step says
reads need no escalation.
Completion notices went through agent.inject(), which never reserves a
driver, so a task settling after its turn closed left the notice parked
until unrelated input woke the agent — while the same prompt told the
model not to poll for it.
An unreported completion now picks its lane from the owner's state: a
busy owner is injected as before, an idle owner is woken with
followup(). This adopts the delivery rule the subagent continuation
manager already ships. maxConsecutiveWakes bounds the self-exciting
chain and is reset by user-authored input; completionDelivery: quiet
restores the old lane for deterministic transcripts.
Teardown cancellation now claims the terminal report the way kill()
already does, so an owner being destroyed is never woken, and settle()
announces completion last so a reporter that opens a turn synchronously
sees a committed record.
Master regenerated this dialog golden on its own side, where the section does
not exist; the refreshed file differs from it by exactly the one nav row.
Three seams: the tsconfig path map gained a mapping on each side and keeps
both; the event-producer matrix is generated, so it was regenerated rather
than hand-merged row by row.
The card read the credential only when its settings scope published, and a
credential is not part of any settings section: a key written from the Models
page — which addresses the same reference — left this badge reporting a state
the Host had already replaced. It now re-reads on credentials/changed for the
reference it watches, and ignores the event for any other reference.
The `cordis` preset's `editing-cordis-compositions` skill is the only guidance
an agent has when it authors a preset, and four of its statements were false.
`tool-bash` was named as the worked example of a row that hides a service; it
provides nothing and injects `bashEnv` from the host's own `bash-env` row, so
following that advice strands the row behind its realm and the preset fails to
mount. The `isolate` example composed `tasks-local` with `tool-tasks`, which
the shipped compositions' own comments say breaks `run_in_background`. A string
realm label was described as pooling one instance; labels join realms and
`provide()` still throws on the second registration. Rows were to be checked
against a package README, which no harness package publishes.
Verification is now the agent's own: `standingKeyFor(id)` runs the same mount a
session start performs and rejects an unresolvable package, an invalid config, a
service in the root realm, and a row that never activated. The skill states that
`list()`'s `broken` field is a shape check that every one of those passes, ships
the `cordis_mount` plugin that reaches the roster service, and names `copy()` as
the authoring write. The prohibition on touching the shipped install is promoted
to its own section and extended to the host composition.
Fixes#2266
Only `docs/module-graph.*` conflicted; generated, so taken from master and
regenerated. `api-proxy.ts` auto-merged — this branch's history read
(`historySourceFor` / `presenterScopeFor` / `historyCutOf`, in that order)
is intact.
Hook bypassed as before: the staged-pairing check hands an archived note path
to `verify-translation-pairing`. The full-corpus gate passes.
The bubble clamped horizontally and did nothing vertically, so a long
label under an anchor low on the page ran off the bottom. Preset cards
made that the common case: custom presets sit at the end of the roster
and carry the longest descriptions.
A `top` or `bottom` bubble now flips to the other side, and only into a
side that genuinely fits — an anchor with room on neither keeps the
requested placement rather than oscillating. Sliding the bubble
vertically instead would cover the text being read.
Only `docs/module-graph.*` conflicted, from master's new
`@deepseek-ai/dsh-message-feedback` package; generated, so taken from master
and regenerated.
Hook bypassed as before: the staged-pairing check hands an archived note path
to `verify-translation-pairing`. The full-corpus gate passes.
A preset publishes its own description, of any length, and `.cards` sizes
rows with `grid-auto-rows: 1fr` — which makes every implicit row the same
height, not just the row holding the tall card. One long description
therefore set the height of the whole roster.
The description now clamps to four lines and offers the rest through the
shared Tooltip, attached only while the element actually overflows. Card
height stays derived: with the description bounded, `grid-auto-rows: 1fr`
already equalizes, and a card carrying the broken-preset reason or a
revealed path still sizes itself.
Tooltip gains an optional `maxWidth`; its default half-viewport cap
renders a description wider than the settings dialog it belongs to.
The failed-shape-check badge reads "Failed to load" rather than "Broken":
discovery reports a composition that is missing, unreadable, or malformed,
which overstates as damage.
Fixes#2238
Generated docs (`module-graph`, `event-producer-consumer`) taken from master
and regenerated. The `cordis-inspect-jsdoc` golden likewise: master's copy is
the base, and this branch's `presentAs` per-scope rewording is re-applied on
top, since `cordis_inspect` renders that JSDoc into model-visible output.
Hook bypassed as before: the staged-pairing check hands an archived note path
to `verify-translation-pairing`. The full-corpus gate passes.
Master moved the browser-half declaration from a top-level `dshClient` to
`dsh.client`, and the new package still carried the old key. The Loader row
resolved, the host half applied, and the bundle built — the browser simply
never mounted the entry, so the session header rendered no task control at all
while `ctx.tasks` held a running task. Nothing reported it; the web e2e is
what caught it.
Also carried the merged tree onto master's current package rules: the release
manifest shape (version, repository, workspace-protocol peers, no `private`)
and `@deepseek-ai/cordis` in place of the old `cordis` specifier, plus the
connection handle the locale plugin's settings scope now reads in tests.
Master gave the agent-preset nav its own glyph; this branch's plugins glyph
sits beside it, and both keep their own icon. Master's stricter
no-base-to-string rule also reaches this branch's new provider test, whose
fetch-call assertion now names the types it reads instead of stringifying an
unknown body.
Three findings from review survived against the staged-save head:
The search provider read its options thunk per property, so a settings write
landing inside credential resolution sent the key resolved from the old
section to the endpoint named by the new one. Each operation now snapshots
once at its entry and threads that snapshot into credential resolution; a
regression test drives a commit into the middle of a search and pins that the
endpoint, model, and key all come from the section the search started on.
The /client entry exported components, controllers, and namespace constants
with no consumer, which the client export discipline allows only with sign-off.
Only types remain. The duplicate per-card Injected/Face interface pairs are
one declaration each now, so a member added to one side cannot silently miss
the other.
The credential state carries the reference it describes and its writability: a
reference change no longer projects the old answer onto the new name, an
out-of-order response for a stale reference is dropped, and a key that a
deployment sources from the process environment disables the control instead
of inviting a write the Host must refuse.
Also corrected three prose claims against the code they describe: the card's
fields do not differ by platform (the served schema does), the section's empty
line counts registered rather than visible cards and is read once, and the
search README overstated what a configuration surface learns about a key.
Master moved every workspace edge to workspace:^ and added release-member
fields; this branch's manifests follow, keeping only the dependency edges it
contributes. The client-runtime README keeps this branch's paragraph: master
did not touch it, and the base/user layers and `unset` it describes are what
this branch added.
Master moved peerDependencies to `workspace:^`; this branch's two added peers
(`dsh-agent`, `dsh-system-prompt`) take the same form. `docs/module-graph.*`
is generated; regenerated after the merge.
Hook bypassed for the same reason as the previous merge: the staged-pairing
check hands an archived note path to `verify-translation-pairing`. The
full-corpus gate passes.
Review found two defects in the previous commit's ordering fix.
@pku-xht: `historyStateFor` copied the attached session's events, the handler
then awaited `presenterScopeFor`, and only then read the projection baseline
off the still-live Session. An append during that await served events cut at N
beside a baseline folded to N+1 — one response describing two moments. The
same restructure had also moved the baseline read outside the `try`, so a
failing snapshot escaped the structured `internal` error.
Both awaits now happen before the cut: `historySourceFor` resolves which
session serves the read, `presenterScopeFor` ensures the recorded composition,
and `historyCutOf` then reads events and baseline adjacently with nothing
between them. The whole sequence is back inside the try.
The invariant judged any scoped assembly with a chain of one as an unjoined
agent, which rejects a legitimate assembly in a standing preset key (that key
has no parent of its own). It now gates on `context.agent` — a scope-only read
is not an agent and is out of range by construction rather than by a premise
about who else calls `assemble` — and asks the roster's own
`composedPreset()` instead of introspecting chain length. The advisory warning
uses the same relation.
Also from review: the `2026-08-05-per-agent-tool-presentation` note still
described `presentAs` as per-agent, which standing mounts made false and this
branch's own rewording contradicts; the duplicated "process-wide unit table"
argument collapses to the Agent Note with pointers from the five copies; a
dead `.sort()` before `arrayContaining`; and change-history narration in the
tool-cordis README.
The plugin-config section added a fourth id to the nav-glyph branch, and no
test rendered one, so CI's per-file gate caught the uncovered path. The nav
now asserts what it is for: each named id draws its own glyph, and a section
this package never heard of still renders the gear.
The pre-commit staged-pairing hook was bypassed for this merge: master's
`docs(notes): archive superseded dsh run decision` (7ee9e16001) arrives as a
rename into `.agents/notes/archived/`, and the hook hands that path to
`verify-translation-pairing`, which correctly refuses an archived note as
out-of-corpus. The full-corpus gate passes (851 pairs).
The web-app bundle gained dsh-app-boot on master beside this branch's
dependency list; both rows stand, and verify-cordis-config confirms every
patch row still resolves from that manifest.
Merging master's rename of the client manifest field left this package on the
old `dshClient` name. The row still composed and its empty node half still
activated, but the browser roster scan never matched it, so the whole settings
section vanished with no error anywhere.
verify-cordis-config now requires a packages/client package's "./client"
export and its dsh.client declaration to agree in both directions; the
composition file cannot tell a surface plugin from a Host plugin, so the
manifests are where this is checkable. The check is scoped to that group
because a Host package's "./client" export is the typed wire face its browser
consumers import, not a plugin the roster serves.
`minimal` was rewritten upstream into a fixed-prompt PTY surface that now
carries its own compaction group, so this branch's token-meter move applies to
it too: the row leaves the group and `tokenMeter` leaves the realm, exactly as
in the other three presets. Master's header prose is kept whole — the note this
branch had added there described a composition that no longer exists.
`docs/event-producer-consumer.*` is generated; regenerated after the merge.
`session.history` built its projection baseline in `historyStateFor` and only
then awaited `presenterScopeFor`, which is what ensures the recorded preset's
standing mount. The unit table is process-wide while the units themselves are
registered by preset rows, so on a first cold read the baseline was serialized
before `todos` (and every other preset-owned key) existed, and every later read
served a complete one. Until this branch, a web lane happened to mount a preset
in `beforeAll` for unrelated reasons, which hid the ordering.
The baseline is now a thunk the handler reads after the scope resolves. Also
covers the roster-less early return of the unjoined-agent warning, which the
per-file coverage gate reported uncovered.
The display name read only the user layer while its placeholder claimed a
cleared field falls back to the route id. A `cordis.yml` can pin a name
for a route the catalog does not ship, and clearing then restores that
name, not the id. The placeholder now reads the composition layer — the
same layer `inheritedModels` reads — and names the route id only when
nothing pins one.
The saved notice echoed the target captured when the card opened, which
never lied while the name could not change. It now names the provider as
the refreshed directory reports it, and is announced only once that
refresh is in the snapshot it reads from.
The protocol select's blank option is named, because a screen reader
announces it either way and an unnamed one is a choice with no identity.
`protocolChoices` no longer rehydrates the section schema for the two
layouts that have no per-route protocol to find.
Documentation caught up with the same broom: the module JSDoc and the
e2e header both still described one field where the fold now carries
two, the English user guide carried a Chinese UI label among English
ones, and the README's DeepSeek row inventory omitted `maxTokens`.
`select.input` caps the control at 240px, and a `<select>` carrying only
`.input` keeps the OS arrow, which paints flush inside that shrunk right
edge — visibly tighter than every other control on the page. Both
protocol dropdowns took `.input` alone; the provider select beside them
already added `.selectInput`, which removes the OS arrow, reserves the
right pad, and paints the shared 12px chevron.
The stylesheet spec now reads the package's own sources and fails any
`<select>` that takes `.input` without `.selectInput`, so the next one
cannot be added bare.
Resolves the vendored-package rescope (cordis -> @deepseek-ai/cordis,
schemastery -> @deepseek-ai/schemastery) against this branch's settings
namespaces: the bash capability namespace, its two executors' section
installs, and the new plugin-config client package all move to the scoped
specifiers.
The vendored Cordis rescope (`cordis` -> `@deepseek-ai/cordis`,
`@cordisjs/plugin-*` -> `@deepseek-ai/cordis-plugin-*`) collided with this
branch's two added edges. Resolved by taking master's package names and
keeping this branch's additions: `@deepseek-ai/dsh-agent` and
`@deepseek-ai/dsh-system-prompt` peers, the `scopeChainOf` import, and the
`agent/created` type-only edge.
`docs/event-producer-consumer.*` is generated; regenerated after the merge
rather than hand-resolved.
Controls committed on blur, which turned leaving a field into a durable,
revision-fenced document write the user could neither preview nor undo, and
silently discarded a draft the field did not accept.
A card's form now owns the staged text every control renders, and Save is the
only point where drafts become writes. Reset stages the composed default the
same way; an invalid draft blocks the save with its reason instead of being
dropped; Discard drops the drafts; a collapsed card marks that it holds some.
The Host stays the only authority on whether a value was accepted, so the save
reads the section back and keeps the drafts of a save that did not land.
The Models editor card curated its pi-ai fields by what every route has,
so `displayName` and `api` — the two a hand-declared route names for
itself — were asked for at creation and then reachable only through
settings.yaml. The editor now renders both for a route the directory
reports as declared, from the same namespace schema the create card
reads. A catalog route gets neither: it defaults its name from its
catalog entry, and its models each carry their own protocol, so a
route-level one could only override all of them.
Clearing the name unsets it and the route falls back to its id, which is
what the field's placeholder shows; storing the empty string would be
refused by the adapter. A declared profile naming no protocol selects
nothing rather than reading as if it had picked the first choice.
The Provider ID stays fixed: it is the settings dict key, it is
referenced from `agent-default-model` and every logged request header,
and it is the stem of a credential reference the page can never read
back to move.
Fixes#2204
Moving the agent plane behind presets left two readers on the wrong side of
the host/agent line.
`dsh-token-meter` was disabled on the host and mounted inside each preset's
`compaction` realm, but its three projection units register into the
process-wide `sessionProjections` table. A unit registered from one preset
answers for every session, so whether a `minimal` session showed a context
meter depended on whether some other session had mounted `standard` since
boot, and a process that only ever ran `minimal` showed none. The meter takes
no configuration, keys every fold by Session, and registers no tool or prompt
section, so it returns to the host composition and leaves the presets'
`isolate` map; the realm and `compact-basic` stay, because what a preset
chooses is whether its agent compacts, not whether its tokens are counted.
Nothing named an agent that joined no preset. The join is a scope-parent link,
and without it the tools, prompt-section, and skill views resolve the empty
global layer: the agent publishes, the turn runs, and the model receives
nothing. `AgentPresets` now logs one warning per such agent while a roster is
configured, and the invariant companion fails outright — at
`system-prompt/assemble` rather than at publication, because an unjoined agent
is legal until it addresses a model and `recompose` binds exactly such an
agent. The warning stays advisory: a synchronous `agent/created` throw vetoes
publication, and the ACP bridge, SDK server, and headless bundle all create an
unjoined agent today.
Three limits are recorded rather than fixed: projection key presence is not a
per-session capability signal, a superseded standing generation is never
reclaimed, and a `cordis_mount` temporary plugin belongs to the composition
rather than the session that mounted it.
Fixes#2203
Yichen Jiang