A picker showed directory names, so the settings page could only ever list
`standard` / `core-web` / `cordis` and hope the reader knew what they meant.
A preset may now publish display text in an optional `preset.yml` beside
its composition, and the section renders cards — name, description, and the
one in use — instead of rows.
The file carries display text ONLY. `id` is the directory name and `trust`
comes from the root a preset was discovered under, so neither is writable
there: otherwise a locally authored preset could name itself into the
shipped set. It is a separate file because a composition is a top-level list
of plugin rows — YAML cannot carry sibling keys beside it, and a fake
metadata row would hand the Loader something to load.
Every read failure degrades to no metadata; absent, malformed, wrongly
typed, and blank all mean the same thing and the picker falls back to the
id. Presentation is not capability: a preset whose name is broken still
mounts.
The editor gained name and description fields above the YAML, and clearing
both removes the file rather than storing a blank name.
A composition is a file, but "edit it on the filesystem" is not a browser
affordance. The roster gains `read`/`write`/`remove` beside `select`, and
the browser gains a settings section over them: the presets as rows, one
composition open in a YAML editor at a time, and per-row default, duplicate,
and delete.
All four authoring methods are loopback-pinned. A composition names the
plugins a session runs, so reading one is reconnaissance, writing one is
arbitrary capability, and selecting one can move a session onto a preset
that edits the live runtime. `agentPreset.list` deliberately stays ordinary
and now reports `authorable`, so a surface knows whether creating is
possible at all rather than offering a button whose save always fails.
Authoring starts by duplicating: a shipped preset opens read-only because
the deployment's copy is what a broken local one is compared against. Ids
are contained before they become directory names, and the text is parsed
with the loader's own schema, so a save cannot leave a file no session
could load.
Fixes a defect the real-composition test found: a preset written under the
user's home could never mount, because the loader resolves a row against the
composition's own directory and Node's `node_modules` walk from there never
reaches the installed harness. The mount now records the host base and sends
bare specifiers there, leaving relative paths resolving from the preset.
Also closes the coverage the earlier surfaces in this stack shipped without —
the General row, the composer seat, and the plugin halves now have tests.
The creation header names the preset a session STARTED with and is frozen,
which is correct — it is a creation fact. Switching is legal only while a
session is blank, and that looked like enough: no history exists yet.
It is not, because the switch's effect outlives the blank window. The user
switches, then sends the first message; every turn from there runs under the
new composition while the header still names the old one. The session is
then locked around a misrecorded preset, and resume reads the header to
rebuild it — composing one preset's tools over a history another produced,
which is exactly the replay the blank-only lock exists to prevent, reached
by another route. A picker showed `standard` for a session running
`core-web`.
A switch is now an `agent-preset/selected` event appended after the swap
commits, and `resolveSessionPreset()` (last selection, else the header) is
what every reconstruction reads: the summary, resume, the conflict guard,
and the fork introduced one layer down.
`agentPreset.select` recomposes one session's agent from a different preset.
It is allowed only while the session is blank — once a turn has run, that
history was produced under the preset's tools and swapping them would strand
logged tool calls, so the attempt answers `agent-preset-locked`.
The agent and the session survive; only the preset subtree is swapped. That
was forced by what the host actually owns: api-proxy discards the `AgentHandle`
it creates, and there is no delete RPC, so neither disposing nor recreating the
session was available. Swapping the subtree is also the better answer — the
session id, its workspace attachment, and its projections all stay put.
`recompose` is unmount-then-mount because two compositions cannot coexist: both
would register the same tool names into one layer. So it resolves the new
preset BEFORE tearing anything down (an unknown id is a no-op) and restores
the previous composition when the new one fails to mount, rather than leaving
the agent with no tools at all. Both paths are pinned by test.
Also restores the English half of the `agentPreset.list` README paragraph,
which was lost before the previous commit — and `verify-translation-pairing
--write` recorded the pair as consistent anyway, because it records whatever
state it finds rather than checking the two sides say the same thing.
`config.default` becomes the composition base of an `agent-presets` settings
namespace, so the user document layers over the deployment's engineering
default and a person can change which preset new sessions get without a
restart.
The value is read per resolution rather than snapshotted: a hot-reloaded
document takes effect on the next session created, and every running session
stays on the preset it was composed from — which is the same rule the
session-header guard enforces from the other side.
`resolve()` read `config.default` directly, which would have made the whole
setting inert; it now goes through `defaultId` like every other caller.
The write-protection test is rewritten against a temp profile root. It was
passing vacuously: the un-overridden Loader REWRITES the composition it read —
stamping `disabled: true` onto the self-disposing row — so the committed
fixture had been mutated by the very run that proved the bug, and every later
run compared against the damaged file and passed. Building the preset in a
temp directory makes the assertion immune to its own failure mode, and it now
fails with a visible `+ disabled: true` when the override is removed.
Review follow-ups on this layer. The exported schema is
`AgentPresetSettingsSchema`, symmetric with the `AgentPresetSettings`
interface it resolves and self-describing at an import site. The `session.create`
JSDoc promised "the deployment's default preset" for an omitted `agentPreset`,
which this layer makes false — it now names the effective default. The
constructor records why it does not use `installSettingsSection`: that helper
re-judges what a consumer DERIVED across attach and detach, and nothing here is
derived. The provider-unload test disposes the fiber `ctx.plugin()` handed back
instead of reaching into `ctx.reflect.store`, and the write-protection wait says
why slack is the right shape for an absence assertion.
The real composition covers the layering too. `apps/cli` boots the shipped
`cordis.yml`, stores `agent-presets.default`, and asserts an unnamed session
composes from it — the package suite proves the layering against a hand-built
context, this proves the roster and the settings provider are wired to each
other. That test also pins the settings row at a temp file: it defaulted to
`$DSH_HOME/settings.yaml`, so a developer's own stored default decided the
outcome of a file whose whole point is that only the shipped root does.
The Agent Note records the per-resolution read and its correspondence with the
session header, and the vacuous-test finding above.
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.
`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.
Records are pruned by observation rather than by a disposal hook, for the
reason the module already states: three different owners can tear a
subtree down, and a cleared `uid` is what they share. That leaves the
pruning to whoever reads — and the only production reader is the
invariant companion, whose package is a development composition a
shipped host never loads.
So a live host pruned nothing: every session ever composed left a record
retaining its whole disposed subtree, since the fiber holds its config
and that config is the key its EntryTree is stored under.
Prune on the mount path too. Every session takes it, which bounds the
set at one generation of dead records instead of one per session.
A preset is a directory holding one `agent.cordis.yml`. Mounting it under an
agent's scope context during `setup(agentCtx)` gives that one session its own
tools and prompt sections while every other live session keeps its own.
No registry gains a tier. `dsh-tools` and `dsh-system-prompt` already file
registrations into the calling context's scope layer, and entry contexts chain
to the context a subtree was plugged into, so a composition mounted under
`agent.ctx` is that agent's alone and unwinds with it.
The mount audits itself because a directly-plugged subtree is absent from
`ctx.loader.entries()` and no boot audit covers it. It rejects an unscoped
target, a row that never became usable, and a row that published a service into
the root service realm — that last one is process-global rather than
per-session, and its collision with the next session surfaces as an unhandled
rejection `setup` never observes, leaving a half-composed agent that looks
healthy. The package invariant re-checks that rule on every service
notification, since a row publishing from a timer would escape a one-shot audit.
Raises the `packages/README.md` word ceiling from 920 to 980: the group table
must enumerate every group, and the new `preset/` row is necessary content.
Design: .agents/notes/implemented/architecture/2026-08-03-per-session-agent-presets.md
The webserver's built-in static dist serving becomes a single-owner fallback
seat (registerFallback/applyIndexTaps); the SPA server moves to the new
@deepseek-ai/dsh-frontend-static plugin so the composing application owns its
dist as composition, not carrier config. distIndex leaves the webserver
schema; unclaimed fallback answers 404.
The contextBreakdown and contextPressure units carried the full priced
surface, so each session's persisted projection checkpoint grew without
bound. A surface replacement is now priced by the shadow-price event
logged directly before it — compact/summary for compaction, the new
compact/prune from tool-result pruning (priced through the injected
token meter) — and the unit states shrink to a fixed handful of numbers.
Regenerate the persistence/cordis/module/config catalogs.
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.
pi-ai resolves a request's apiKey override only through a provider that
declares an api-key method: resolveProviderAuth short-circuits to that
method when the override is present, and otherwise falls through to the
credential store and then to ambient discovery. A provider with no
api-key method at all therefore resolves to nothing, and the request
fails with "Provider is not configured" before any network I/O.
Two routes hit that. openai-codex ships OAuth alone, so moving off the
/compat dispatch broke a profile that names a key for it — the old path
handed the token straight to the provider. And a catalog route naming an
api was being rebuilt with the harness's own auth, so `openai: {api:
openai-completions}` stopped reading OPENAI_API_KEY, contradicting the
documented promise that omitting a credential keeps provider-native
discovery.
Auth is now one decision for both constructions. A catalog route keeps
its installed provider's auth, through an api override too: which
environment a provider reads belongs to the provider, not to the wire
format its models speak. A catalog provider with no api-key method gets
the harness method beside its own, but only when the profile names a
credential — a keyless codex profile keeps the honest refusal, since
this adapter holds no OAuth store to resolve through.
Materialization now spreads the installed entry instead of enumerating
the result, so a Model field this package does not model survives a
pi-ai upgrade; headers went missing from an nvidia route exactly that
way once already. providerInfo reports the configured displayName, which
also joins the registration facts so a rename re-registers rather than
leaving the old label in every selector. A refused registration swap
gets its own diagnostic naming the route, matching the directory swap
beside it.
The README documented endpoint interrogation this layer does not
implement, and still described unknown providers as kept-last-good after
they became legal declarations refused at the write point. The Agent
Note claimed per-model reasoning configurability the schema never had,
required capacities the route now defaults, and stated an apiKey
override that short-circuits unconditionally.
Three defects surfaced while driving the Models page.
A hand-declared model needed an explicit contextWindow and maxTokens,
but a provider listing usually returns ids and nothing else — so the
page happily wrote a profile the adapter then rejected, which took the
whole namespace down silently. Capacities now fall back to the route's
`defaultContextWindow` (262,144) and `defaultMaxTokens` (32,768). Both
are guesses by construction, which is why they are route fields a
deployment corrects once rather than constants buried in the adapter;
the fallback sizes the model and never becomes a per-request cap.
That silent failure was the second defect. A schema-valid profile the
adapter could not serve was stored and only rejected later, disabling
every route in the namespace with nothing said. `dsh-settings` gains an
optional `validate` on registration — a check for what a schema cannot
express — and `llm-pi-ai` refuses an unserviceable section at the write
that produced it. A stored section that fails keeps the namespace's last
good value, as a schema failure already did, so an externally edited
document still cannot strand the owner. The plugin's own last-good
fallback goes with it: nothing reaching it can fail any more.
Third, a model with no reasoning metadata advertised the single level
`off`, which pi-ai translates to *omitting* the reasoning option — the
same request naming no effort produces. Selecting it disabled nothing,
so a provider whose default is to think kept thinking with `off` shown
as selected. Such a model now reports no reasoning capability at all,
which is the seam's way of saying the control is unavailable, and the
per-model `reasoning` flag is gone: without a thinkingLevelMap to spell
levels it could only invent them.
The protocol table narrows to the three a hand-declared route reaches
today, most-reached first so a surface offering a choice defaults to the
one gateways actually speak.
Review found four defects in the declared-provider work.
`PiAiAdapter` reused one `Models` collection and mutated it whenever the
configuration changed. `Models.streamSimple()` resolves its provider
lazily — when the stream is first consumed, which is after the adapter
awaits the route's credential — so a configuration change landing in
that window let an in-flight request finish under a configuration it
never resolved against, or fail on a provider that no longer existed.
Each resolution now produces an immutable snapshot and every operation
captures one before its first await, which is what makes the seam's
per-step freeze (`llm.prepareCall()`) hold end to end: switching models
mid-reply takes effect on the next step, never inside the one in flight.
`defaultMaxTokens` was materialized from the catalog's `Model.maxTokens`.
The two answer different questions: pi-ai requires that field as the
model's output capability, while the seam's is a cap the deployment
chose to send on requests naming none, so every request had started
carrying a number nobody picked. Only an explicitly configured cap
reaches the seam now.
The configurable-provider directory was refreshed by disposing its
registration and making a new one. A candidate set the registry refuses
— a profile keyed `deepseek-official`, which llm-deepseek declares —
left the whole directory withdrawn and the Models page empty, silently,
because the settings callback contains the failure. The seam's
registration handle now carries `replace()` with the same
validate-first atomicity `registerAdapter` has.
The protocol table offered every pi-ai streaming API, including four
whose authentication a profile cannot express: Bedrock signs with SigV4
over AWS credentials and a region, Vertex needs a project, a location,
and ADC, Azure needs provider environment plus an api-version, and Codex
uses OAuth. Offering them handed back routes that cannot authenticate.
Catalog routes still reach them through their own provider.
- `relay` resolves its sender in `contextBody` like every other form. It was
the one shape whose marker could claim a form the body did not render: an
unreadable sender fell back inside the body while the row still said relay,
contradicting the contract this PR's own note states.
- `recall` requires the retained, omitted, and truncated fields. Completeness
is what the card exists to report, so a reference that cannot state it is
not a readable recall — showing the label alone presents a confident card
over unknown loss.
- The snapshot body states the supersession its producer framing line carries.
That line is the one part of the model-facing text no section contains, and
unlike an instruction context's `<system-reminder>` it states the form's own
semantics rather than wrapping content.
- `GoalMessageSource` is a discriminated pair, so `{ form: 'notice' }` without
its account no longer compiles. The guarantee this PR claims now holds at
that seam too, not only through `ContextFormed` on plugin sources.
- Goal and tool-goal summaries are bounded by a shared `boundContextSummary`,
which tool-tasks now uses as well. A goal objective is unbounded caller text
in exactly the way a task label is.
- The runtime snapshot interpolates once per request: agent-loop renders the
sections and joins them through `joinContextSections`.
- Every form's fallback branch is pinned, not only the notice one.
Four values complete the vocabulary, so the opaque body is reached only by
producers that genuinely promise no shape.
`snapshot` — current state a later snapshot supersedes. system-prompt now
exposes `renderContextSections()`, the named contributions
`renderContextSnapshot()` already joins for the model, so the body attributes
each part to the subsystem that produced it instead of re-splitting joined
prose. The runtime snapshot, time-context, and tmux-context declare it.
`notice` — a one-off account of what just happened, declared by tool-tasks,
goal state changes, tool-goal wrap-up, plan-mode switches, and
repeat-tool-guard. Its `summary` rides the COLLAPSED row: these five are the
majority of shipped producers and none of them needs expanding to be read.
The task summary bounds itself because its inputs are unbounded caller text.
`relay` — a message another agent addressed to this one; both subagent
sources declare it and the body names the sender above what it said.
`recall` — material lifted from another session's log. session-reference
needed no new field: its references already record retained and omitted
counts and the truncation flag, which the body shows first, because recalled
context is bounded on the way in.
`ContextFormed` is now discriminated by `form`, so a producer cannot declare
a shape without the facts that shape is presented from — a notice without its
summary, or a snapshot without its sections, fails to compile.
Only the two hook bridges stay opaque, by design: their content is whatever
an external program printed, so no shape can be promised for it. Unknown
kinds and unreadable records land there too.
`contextBreakdown.messageTokens` and `measure().surfaceTokens` answer the same
question in the same heuristic vocabulary, and the panel's composition rows are
only honest while they agree. Each owner carried its own copy of the positional
fold — same pricing, same `{seq, tokens}` node list, same replace-range lookup
and guard, differing only in mutable versus immutable application — so an edit
to either one would have moved the panel away from `measure()` with both sides
still green. The duplication gate caught the shared 62 tokens.
`src/surface-fold.ts` now owns `foldSurfaceTokens`: total, allocation-fresh,
returning the event's price, the next surface, and the signed total delta. The
service assigns that result where it used to prepare a commit closure, which
keeps its validate-before-mutate replay transaction intact — the fold throws
before any state is touched, so a malformed event still fails identically on
every retry. `_prepareSurfaceMutation` and `_estimateSurfaceEvent` go away with
it, and the projection's apply drops to one call.
Covers the identity with a session that appends and then compacts, asserting
the projection figure equals the service surface at each boundary; the test
fails when either side of the fold is perturbed.
Assistant footers and the stats line gain TTFT/tok-per-second readings
folded from step timings; context occupancy moves off the stats line onto
a composer ring whose panel shows a heuristic system/tools/messages
breakdown from the new token-meter contextBreakdown session projection.
