Edit cards showed English words such as "crop · flip · extend" to every user.
edit_image now takes an optional title written by the bot in the user's
language, used for the card and the file name like edit_video. Without one,
the card falls back to the repaint prompt or to language-neutral tokens only:
sizes, ratios, degrees, arrows and format names.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
A Postgres suite starts edit_video and export_video through the real
runtime and local artifact store: outputs become linked versions with
posters on the card, exports store one version per preset, appended clips
and subtitle files of another user are refused at start, and an
out-of-range cut fails the job with a readable error. It skips without
Postgres or ffmpeg.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
- The GPU guard waited for free VRAM at the operation's floor even on a card
whose total memory is below it (an 8 GB card and a 12 GB video floor), so
such jobs deferred until their budget ran out. The requirement is now capped
at 90% of the card's total; a card with less than half the floor fails at
once with "The GPU does not have enough memory for this job."
- Tests pin that media tools reach a run only through the space's
capabilities and within the bot's allowlist now that the image tool
definitions exist.
- Type-only imports are top-level `import type` in the media runtime, tools,
RPCs and tests (AGENTS.md).
- The ComfyUI settings test gives saveMediaSettings a store with nothing saved,
since the save reads the previous settings before validating.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Picks up ComfyUI generation, media tool dispatch and the streaming runner.
The video tools now share the image tools' schema helper, and both sets of
operations join the runtime's operation list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Binary paths, the ffmpeg version, the probe's container name and a
configurable stall timeout had no caller; detection is cached per process,
so every helper looks the binaries up itself.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Media tools and card actions send prompts, images, video and audio to the
deployment owner's ComfyUI and transcription servers. Like the embeddings
server behind built-in memory, those self-hosted endpoints are recipients in
the AI data-sharing disclosure: they appear with the model uses ("ComfyUI
(<origin>)", "Transcription (<origin>)", keyed by the full endpoint so a new
server asks again), and starting a card job asks for model consent on clients
that run the foreground check.
Also adds unit tests for the media settings store: URL normalization and
refusals, encryption, the key-free view, one decryption per saved version, key
retention per server and removing empty settings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Bots and users need to start and follow media jobs, and the deployment owner
needs to point the deployment at their GPU and transcription servers.
- Media tools are named like their operations. A run sees one only when the
space can run that operation (MediaRuntime.capabilities); the per-bot
allowlist still applies. applyTool hands every media tool to one
dispatcher (media-tools.ts) that splits off continue_after, starts the job
under the tool call's effect key (a replay returns the same job) and returns
at once; refusals (invalid input, unknown artifact, unavailable) go back to
the model as tool errors.
- media.jobs.start starts an export, transcription or image edit from a card:
the conversation is resolved like every thread procedure, the card's bot
(or the group's lead) posts the new card, and the runtime refuses source
artifacts outside the actor's space. media.jobs.get and media.jobs.cancel
only see the actor's own jobs.
- media.settings.get/update are deployment-owner only, never return keys and
answer invalid URLs with a 400 that says what is wrong.
- createApp accepts media operations as a test seam.
- Tests: a Postgres suite over the RPCs (owner-only settings, encryption,
key retention per server, URL policy, start/get/cancel isolation between
users, card jobs end to end) and unit tests for tool exposure and dispatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
The image server is the deployment's, not necessarily the user's, and
repainting an edit only works when generation is available.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
edit_video and export_video now run as media jobs: both registrations join
the shared operation list the API and the worker build their runtime from,
name every artifact they read (source, appended clips, subtitle files) so
starting a job refuses artifacts the user cannot read, and their tool
schemas use the runtime's continue_after helper.
Container deployments get ffmpeg, ffprobe and the subtitle font in the
API/worker image. Tests cover the handlers against a fake job context:
cuts, blurred 9:16 reframes with subtitles, appended silent clips, audio
extraction, single frames, batch exports with posters and progress,
rotated phone video, clear input errors and cancellation. They skip where
ffmpeg is not installed; the planner and parser tests always run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Queue, interrupt and prompt requests now go through one JSON helper, so the
cancel path reads like the rest of the client.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Bots can now make images and short clips on the user's own GPU server and
rework images locally, through the media job runtime.
- generate_image: prompt, aspect preset or exact size, up to 4 images, optional
source image for image-to-image. Sizes target what the default models are
tuned for, and sources are fitted to that size before upload, so large
photos cannot exhaust GPU memory.
- generate_video: prompt, aspect, duration up to 10 s, optional start frame.
- edit_image: crop, rotate, flip, resize to a size or format (crop, pad or
blurred background), extend, convert, in that order, with sharp; an optional
prompt repaints the result on ComfyUI. Edits keep the source as parent.
- Registrations declare the input artifacts so jobs refuse ones the user
cannot read, and generation declares a fixed GPU memory floor (6 GB images,
12 GB video) that a custom workflow cannot lower.
- Tool definitions reuse the operation input schemas plus continue_after.
- Saving media settings rejects custom workflows the provider could not fill.
- Tests: sharp edits on generated fixtures, handlers against the ComfyUI
emulator, and a PostgreSQL suite running the operations through the media
runtime (card outputs, artifact versions, foreign artifacts, GPU wait).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Media files can be hundreds of megabytes, so the runner no longer buffers them:
inputs are cloned from the local artifact file when the store has one and
streamed otherwise, outputs go through storeArtifactFromFile (per-kind size cap
and MIME check while streaming; an oversized result fails the job with a short
card error).
A Postgres suite runs the runtime with fake operations: success with outputs,
posters and parent versions; input, availability, artifact and thread checks
before a job exists; tool-call replay; failure with temp-dir cleanup; outputs
outside the temp dir and sources of another user; deferral with backoff and its
budget; the GPU memory guard; cancel of a queued job and of a job running in
another process; one job per lease and one job on the GPU at a time; crash
recovery and its restart limit; reconciliation inside and outside the queue;
the continue_after wake; a cleared conversation; the run time limit; and
capabilities.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Brings in artifact streaming and the media job runtime, settings and tool
helpers that the ComfyUI operations register with.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Brings in artifact streaming and the media job runtime that the video
operations register with.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Bots can now cut, reframe and export videos for every common format without
a hosted service: ffmpeg and ffprobe from PATH do the work on the server.
- ffmpeg wrapper: PATH detection, ffprobe JSON probing (rotation, pixel
aspect, frame rate, audio, packet scan when the container has no
duration), runs with -progress parsing, abort and stall kills, an output
size cap, partial-output cleanup and short errors without temp paths.
Inputs go through the file protocol with the demuxer forced from the MIME
type, so a crafted upload cannot be probed as a playlist that reads other
files or URLs. extractAudio (16 kHz mono WAV) and renderPoster for
transcription and media cards.
- A pure planner turns edit steps (trim, keep, concat, reframe, speed,
volume, mute, subtitles, extract_audio, poster) into one filter_complex
pass: leading cuts seek the input, appended clips are normalized to the
frame, fps and 48 kHz stereo (silence where a clip has none), reframes
crop, pad or blur to the presets, speed chains atempo, subtitles burn in
from generated file names with a readable style that stays clear of the
controls on 9:16 frames. H.264/AAC MP4 with yuv420p and faststart;
location metadata is stripped.
- edit_video and export_video operations and tool definitions; the inputs
are shared contracts so the media card can start exports.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Media operations (generation, edits, transcription) need a runner that works
outside the agent turn, keeps a card in the conversation current and never
lets two processes put work on one GPU. Operations land in later commits; this
is the runtime they register with.
- Background job `media.run { jobId }` on a named `media` queue. Jobs carry an
optional queue: the Graphile publisher passes it as queueName and the
in-memory driver runs each queue one job at a time, in order.
- `createMediaRuntime` implements MediaRuntime (capabilities, start, get,
cancel) over a list of `MediaOperationRegistration`s and runs jobs:
start validates the input, creates the job and posts the bot's media card in
one transaction (idempotent per tool call through an operation key), then
enqueues it. The runner claims a fenced lease (an advisory lock keeps one
live lease at a time, so the GPU stays single even when the queue is
bypassed), renews it while polling for cancellation, writes progress and
outputs into the card, defers with backoff when the GPU is short of memory
or the operation asks to wait, bounds deferrals, crash restarts and run
time, always cleans the job's temp directory and always ends the card in a
terminal state. With continue_after the starting bot is woken in the same
transaction that ends the job (new run trigger `media`).
- The reconciler re-enqueues jobs whose enqueue was lost, whose runner died
or whose wait is over; a job stuck behind a dead worker's queue lock runs
outside the queue.
- Media backends are deployment-wide settings stored encrypted as the
`media` IntegrationProviderConfig row: `loadMediaSettings` and
`saveMediaSettings` with the same address policy as OpenAI-compatible
servers (no metadata hosts, public hosts only when allowed, HTTPS when a key
is sent), a saved key never follows the settings to another server, and the
view never returns keys. Contract in contracts/media-settings.ts.
- `createMediaOperations` is the one list both the API and the worker build
their runtime from (empty until the operations land), and media-tools.ts has
the shared `continue_after` schema helper for the tool definitions.
- Migration: lease, deferral, operation key and continueAfter columns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Videos and audio can be hundreds of megabytes. Artifacts only travelled as
base64 inside RPC calls, capped at 10 MiB and held whole in memory, so media
could neither be uploaded, played with seeking nor produced by media jobs.
- ArtifactStore gains stat, createReadStream (inclusive byte ranges),
putStream (size cap enforced while writing, real SHA-256) and an optional
localPath so ffmpeg can read stored files in place. LocalArtifactStore
implements them with the same safety as before (O_EXCL, O_NOFOLLOW, 0600)
and now validates space ids and keys before building paths.
- storeArtifactStream / storeArtifactFromFile store a stream or a job's
output file as an artifact row, with width, height, duration and parent,
and remove the blob again if the row cannot be written.
- GET /api/artifacts/:id serves an artifact with Range/206/416, ETag,
inline disposition for media and attachment otherwise (safe RFC 6266
filenames), nosniff and same-origin CORP. Authorization matches
artifacts.get: the owner's artifact of an active bot thread or group,
with membership of the artifact's space (media elements cannot send the
space header; an explicit header must name the same space).
- POST /api/artifacts takes a raw streamed body with the target and name in
the query, authorized like artifacts.create, with the MIME allowlist and
the per-kind caps checked from Content-Length and again while writing.
- The API CSP allows media from the app origin and blob URLs.
- Web: artifactUrl(artifactId) and uploadArtifactFile for the UI.
Tests cover the streaming store, range parsing, filename sanitizing, the
file helper, and a Postgres suite for both routes: ranges, caching, other
users, shared-space members, other spaces, archived threads, signed-out
requests and upload limits including chunked bodies.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Image and video generation run on the user's own GPU server. ComfyUI is the
common self-hosted backend, so this adds a provider for its HTTP API that the
generation operations can call.
- Core: API-format workflow templates with a small placeholder format
({{prompt}}, {{width}}, {{seed}}, ...) that is validated and filled without
evaluating anything. Defaults: text-to-image and image-to-image on core nodes
for any SD-family checkpoint, and text/image-to-video on ComfyUI's core
LTX-Video nodes; a README lists the placeholders and model files.
- Adapters: createComfyUiProvider submits a workflow, polls its history,
uploads input images and downloads outputs through the address-checked fetch
used for other self-hosted model servers, with timeouts, capped reads and
readable errors. A custom workflow from settings replaces the default it
matches. Cancelling dequeues or interrupts only its own prompt. GPU status
counts memory ComfyUI can free for itself as available.
- Contracts: input schemas for generate_image, generate_video and edit_image.
- Tests: an in-process ComfyUI emulator, a provider-neutral generation
conformance suite and provider tests, all offline.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Groundwork for image and video generation, video editing and transcription.
Nothing calls it yet; the media runner, providers, tools and UI build on it.
- Contracts: video, audio and subtitle MIME types, per-kind size caps for
streamed and generated artifacts, video/audio message blocks and the live
media card block, media operations, job statuses and output presets
(9:16, 16:9, 1:1, 4:5, 4:3, 21:9).
- Core: attachment kinds include video and audio, artifacts become video or
audio blocks, and agent history keeps media artifact ids so later turns
can refer to them.
- Schema: artifacts record width, height, duration and the artifact they
were made from; media_jobs tracks generations, edits and transcriptions.
- Adapter kit: provider-neutral image generation, video generation and
transcription interfaces, the media job context operations run in, and
the MediaRuntime that tools use to start jobs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Chat markdown hard-coded a near-white text color, so on light surfaces
(ask cards, delegated task briefs, peer and file previews) the text was
white on white. The desktop conversation panel and the bubble and glass
surfaces were hard-coded dark while the text on them follows the theme,
so in the light theme the user's own messages rendered as empty pills.
- Chat markdown uses the foreground token; heading, code and quote colors
come from the shared @rakazo/chat-ui markdown styles instead of
white/emerald overrides. Quotes read in the body color, as on mobile,
so they stay legible in the dimmed transcript.
- User bubbles use the chat-user tokens; bubble and glass card surfaces use
card and border tokens.
- The conversation panel uses the background token on desktop too, as it
already did on mobile. Bot reply footers, summary cards, quick prompt
chips, the awareness strip and the chart expand button drop hard-coded
white/zinc/emerald classes that would vanish on that panel.
- New web E2E measures the contrast of the user's messages, a markdown bot
reply, a delegated brief and an ask card in a group transcript in both
themes and captures screenshots.
Claude-Session: https://claude.ai/code/session_01WnqLpAn8253EY1eCLKVJb2
Co-authored-by: Claude <noreply@anthropic.com>
* feat(db): main agent, main conversation, user model and delegated task foundation
The product needs one main agent that leads a main conversation, knows the
user across spaces, and hands work to visible specialists. Today none of
that exists in the data model: "Chief" is a client-side naming convention,
the "Hauptthread" is a group the web client creates by name, and nothing
stores what the agents know about the user.
This adds the shared foundation the Phase 1 features build on:
- Bot.kind ("main" | "specialist", one main bot per user) plus capability
fields (tags, builtin tool allowlist, model tier).
- ChatGroup.kind ("group" | "main", one main conversation per user) and
ChatGroup.leadBotId.
- UserProfile and UserFact with category, confidence, source, consent
status and per-space visibility; embeddings are stored as float arrays
because none of the Postgres images ship pgvector.
- DelegatedTask for work the main bot hands to a specialist.
- SpaceModelTier to map a model tier to a connected model per space.
The migration backfills each user's onboarding bot in the home space as the
main bot and turns an existing "Hauptthread" group containing it into the
main conversation. CHECK and foreign key constraints on existing tables are
added NOT VALID and validated in a separate migration, as elsewhere in this
repo.
New db helpers ensure the main bot and the main conversation (idempotent and
race-safe, repairing a lost lead), add same-space specialists to the main
conversation and announce it with a group.updated event, and load the user
context a run may see. The main conversation is exempt from the two-member
floor and cannot have its members replaced through a group update. The
contracts expose the new fields and the user model and delegated task
schemas.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* feat(adapters): every run knows who it is and who the user is, in the user's time zone
Runs dropped the bot's name, title and description whenever custom
instructions existed, always stated the time in UTC, and knew the user only
through free-form durable memory. Groups with a lead, like the main
conversation, still got the peer-group lines that forbid handing work back.
And every group follow-up run failed during setup: the inbound handoff lookup
selected a `bot` relation that Message does not have, which Prisma rejects.
- Instructions now open with the bot's identity, then its own instructions,
a main-agent role paragraph for the main bot, the current time in the
user's time zone with its offset (UTC when unknown), and a `<user>` block
built from loadUserContext: profile fields and the active facts visible in
the run's space, escaped, clipped and framed as data rather than
instructions. Runs that reply in a shared messaging channel get only the
name, zone, locale, language and style, matching how those runs already
leave out durable memory.
- run_subagent helpers receive the identity, time and user block through a
new optional AgentRunRequest.subagentContext instead of nothing.
- Groups with an active lead get lead and specialist lines and matching
inbound handoff instructions. Members may hand a stage back to the lead and
the lead may delegate the next stage; the bounce rules still apply between
members. Ordinary groups keep today's lines.
- The teammate directory is rendered only when the run has message_bot, which
group runs never have, and tool guidance names a tool only when the run has
it.
- The auto-review judge, history compaction and the team chat judge read the
time in the user's zone, and schedule_create defaults to it.
- The inbound handoff sender is read from the handoff block and looked up
among the user's bots; a Postgres test covers the follow-up run, the prompt
it receives, and handing the result back to the lead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* feat(memory): learn the user profile from conversations; add built-in semantic memory
Why: the user profile only knew what onboarding wrote, so people had to repeat themselves,
and semantic recall needed a hosted or separately run memory service. The profile now grows
from everyday conversations, and recall can run on the user's own server with a local
embeddings endpoint (for example llama.cpp). Both stay optional and are tested offline.
What:
- adapter-kit: EmbeddingProvider contract (describe, embed) and the profile.consolidate job
(payload schema and a replace-keyed job factory, one pending job per user).
- adapters: OpenAI-compatible embeddings (POST {baseUrl}/embeddings through the guarded
OpenAI-compatible fetch, batching, bounded responses, clear errors), a deterministic
emulator, an in-process /v1/embeddings server emulator, and a shared conformance suite.
- adapters: "builtin" semantic memory stored in Postgres (semantic_memory_items with Float[]
embeddings, cosine ranking in process, bot/shared/history scopes matching the existing
providers), registered in the memory provider factory and configured through
memory.connectProvider with an embeddings URL, model and optional API key. Private
endpoints require the deployment owner, like the other local memory providers.
- adapters: profile consolidation reads the user's settled app conversations since
UserProfile.consolidatedAt (never messaging channels), asks the main bot's model for strict
JSON facts and an updated summary, drops restatements (embedding cosine when built-in
memory is configured, normalized text otherwise), stores sensitive categories as pending,
and advances the cursor with a compare-and-set. A reconciler hook enqueues it at most daily
per active user; it skips under the scripted runtime and without a usable model.
- db: additive migration 20261002210400_ws4_semantic_memory.
- web: a "Built-in" option in Memory settings; AI data sharing names the embeddings server.
Tests: new embedding conformance, OpenAI-compatible wire, built-in provider and consolidation
unit tests; new Postgres suites for built-in memory (including a real loopback embeddings
server through the memory provider resolver) and for consolidation and its scheduling query,
both added to the integration harness. The settings E2E spec now opens the new form.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* feat(models): map model tiers per space and let bots and helpers ask for one
A space can now map the tiers fast, standard and code to connected models,
including several models served by one OpenAI-compatible router endpoint
(one credential, many model ids). A bot can ask for a tier, and a single
run_subagent call can ask for one. A tier with no model mapped falls back
to the usual model choice, so nothing breaks before tiers are set up.
- db: SpaceModelTier repo (list, find, set or clear); tier mappings are
copied into a new space like model and voice preferences; createBot
carries a tier, so duplicates keep it and spawned bots inherit it.
- contracts/api: models.tiers.list and models.tiers.set;
models.probeOpenAiCompatible can list a connected server's models by
credential id with its stored address and key; bots.update accepts
modelTier and keeps a pinned model and a tier mutually exclusive, so a
model picked on a surface that knows no tiers (mobile) is never silently
overridden. Thinking-level validation follows the tier's model. AI
consent status discloses tier models; desktop local settings allow the
tier procedures.
- adapters: selectConfiguredModel runs a mapped tier ahead of the pinned
model and the space default; the executor resolves bot tiers for runs
and model lookups, and helper tiers through resolveModelTier.
run_subagent takes model_tier; an explicit model wins over it.
- web: a collapsed Tiers section in model settings (shown once a model is
connected) and mapped tiers as choices in the bot Model select.
- docs: tiers in self-host model setup.
- tests: resolution order and fallbacks (selection, executor, Pi helper),
repo, API and contract units, authorization and offline Pi Postgres
suites, and a model-tiers web E2E spec.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* fix(adapters): state the consolidation time in the user's zone; name the memory migration after its content
Profile consolidation now passes the profile's time zone to the shared time
instruction, which takes the zone since every run states the time in the
user's zone. The semantic memory migration is renamed after what it creates.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* fix(adapters): keep the user's summary and facts out of replies to other owners' agents
An agent connection wakes a bot with another owner's message, and the reply
goes back to that owner. Such runs now get only the basics of the user block
(name, time zone, locale, language, style), like runs in shared messaging
channels, so confirmed facts such as health never reach another person's
agent.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* feat(profile): readable, writable and forgettable user model
The main agent is meant to know the user, but nothing could read or change
the user model the foundation added. Onboarding pasted the user's answers
into the first bot's instructions as a "User Context" block, and the user
had no way to see or remove what their bots knew about them.
This makes the user model a user-controlled resource:
- db: a profile repository scoped by user. One atomic update writes the
account name, profile fields, new facts (deduplicated by normalized
text, capped per user), fact edits including per-space visibility, and
confirmations. Health and values facts wait for confirmation unless the
user wrote them. Forgetting hard-deletes facts by id, category or all,
clears the summary that may restate them and marks earlier chats as
consolidated so they cannot teach them again; forgetting everything also
clears the profile fields. Writes take a per-user advisory lock.
- api/contracts: profile.get, profile.update and profile.forget. Unknown
fact ids are a 404 on update and ignored on forget; spaces outside the
user's memberships are rejected.
- adapters: an update_user_profile tool for the main bot only, hidden from
specialists and shared messaging channels and refused if called anyway.
Facts carry source "conversation" and the bot, thread and message they
came from. The scripted runtime records "learn about me: [category:]
statement" for tests.
- web: onboarding writes the name, time zone (browser default, now
prefilled and editable), locale, chosen language, communication style
and the hobby and work-style answers to the profile instead of
overwriting the bot's instructions, and the harvested memory section no
longer repeats them. A new "About you" settings section lists profile
fields and facts by category with their origin, confirms pending facts,
and forgets one fact or everything.
Tests cover the contracts, the repository (unit and PostgreSQL), tool
selection and refusal, the scripted trigger, the onboarding mapping and
view helpers, an API suite that runs the tool through the executor with a
real main bot, authorization for the new procedures, and an E2E spec.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* feat(delegation): main bot hands work to specialists as tracked tasks
The main agent had no way to give a specialist a piece of work and keep
track of it. message_bot is a fire-and-forget peer message whose work stays
hidden from the user, and handoff_to_bot ends the main bot's turn and only
works between members of one group. Delegation now runs the specialist's
work where the user follows the main agent and brings the result back so
the main agent can close the task.
- delegate_task({bot, title, brief, deadline?}) is offered to the main bot
only and refused unless the caller leads the main conversation. It
resolves the specialist by id or name across the user's spaces, creates a
DelegatedTask and posts a task card in the main conversation.
- A specialist from the main conversation's space joins it and works there
in a run with the new trigger "delegation", which the peer filters do not
hide, so its progress, asks and approvals appear in the main conversation
with the specialist as sender. A specialist from another space works in
its own chat; its written result is posted into the main conversation as
its message.
- The executor mirrors the run on its task (running, waiting, done, failed,
cancelled) after the run starts and when each attempt ends, and updates
the card in place through thread.message.updated. Done and failed tasks
wake the main bot in the main conversation with the result. The job
reconciler settles tasks whose run ended without the executor (stop,
clear, archive, deletion, crash).
- tasks.list({threadId?, status?}) returns the caller's tasks for the
current space.
- The task card shows who, what and status with Lingui labels; the brief
opens on demand. The message_bot card no longer shows the recipient's
unrelated overall status or hard-coded German copy.
- The scripted runtime understands "delegate to the bot named X: ..." and
the task wake prompt; activity rows name delegation runs by task title.
Tests: unit tests for the prompts, status mapping, card, reducer, tool
selection and scripted runtime; a Postgres suite for same-space, group,
other-space, ask and answer, failure, stop and replay; tasks.list in the
authorization suite; and an end-to-end spec for the whole flow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* fix(adapters): show the main agent the specialists it can delegate to
The main conversation is a group run, and group runs no longer list the
teammate directory, so the main agent could not see the names of the
specialists delegate_task expects. Runs that have delegate_task now list
the user's specialists across the spaces they belong to, those next to the
conversation first and others marked with their space, and that list
replaces the teammate directory so no bot appears twice. It also says when
to prefer delegate_task over handoff_to_bot.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* feat(inbox): show what specialists need from the user in the main conversation
Specialists ask questions, request approvals and the computer screen in
their own threads, and the user only finds out by opening each thread. With
the main conversation as the single entry point, those requests have to
reach the user there, with the specialist as the visible sender.
- inbox.list returns, for the current space and the user's own bots, pending
asks and approvals (bot messages with a pending ask block whose run waits
for input), takeover requests, and unread specialist threads. Asks already
in the main conversation are left out; takeovers stay because their
controls live in the bot's own thread. Each item carries the bot, the
thread target, run and message ids, and the ask block for asks and
approvals.
- The main conversation shows the items above the composer. Asks and
approvals are answered in place through threads.answer with the source
thread target; takeovers and unread threads open the bot's thread. The
list refreshes when the bots' status, unread state or newest message
changes, every 15 seconds, and after an answer.
- The inbox shrinks and scrolls so it never pushes the composer out of view.
The transcript now clips its scroller: with no room left in the column,
the scroller's padding used to spill over whatever sat below it.
- The ask block schema is extracted from MessageBlock so the inbox contract
reuses it, and AskCard accepts a className.
Tests: unit tests for the item builder and the component, a Postgres suite
for the query (added to the harness list), inbox cases in the authorization
suite, and an E2E spec in which a specialist asks in its own thread and the
user answers from the main conversation.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* fix(adapters): write learned facts under the profile's lock and limit
Profile consolidation wrote facts beside the user profile repository: without
its per-user lock, its duplicate check at write time, or its limit of 500
facts. A long-running profile could grow past the limit and then refuse the
user's own additions. Consolidation now takes the same lock, drops facts the
profile gained since the run read it, and keeps only what still fits.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* fix(adapters): keep profile writes out of replies to other owners' agents
A run woken over an agent connection hears another owner's bot, not the
user, so update_user_profile is withheld there: nothing such a conversation
says can become a fact about the user. Detection moves ahead of tool
selection; the reduced user block uses the same flag.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* feat(capabilities): least-privilege builtin tools per bot and schedules on behalf
Every bot saw every builtin tool, and nothing outside the runtime's own
declaration kept a run to the tools it was given: the executor ran any tool a
runtime asked for, and the Pi runtime exposed all builtins whenever a caller
passed an empty tool list, which the safety checker, the history summarizer
and the team chat judge do on purpose. A specialist could not be kept to the
tools of its lane, bots had no capability tags, and the main bot could only
schedule routines for itself.
- A bot can carry a builtin tool allowlist. Its runs see only the allowlisted
builtins, and an empty allowlist means none. MAIN_ONLY_BUILTIN_TOOLS (empty
until the user profile and delegation tools land) never reach other bots.
applyTool refuses any tool the run did not expose, and an empty tool list
now means no tools for the Pi runtime and its subagents.
- A bot spawned by a restricted bot inherits the restriction and a duplicate
keeps it, so neither path widens a bot's tools.
- bots.create and bots.update accept tags, restrictBuiltinTools and
allowedBuiltinTools; the router rejects names that are not builtin tools.
bots.builtinTools lists the names a bot's allowlist can use.
- Bot settings get a Tags field and a switch with a tool checklist under
Advanced.
- The main bot's schedule_create takes an optional bot: one of the user's
bots in any space the user still belongs to. The routine belongs to that
bot in its space and wakes in the main conversation when the bot is a
member there, otherwise in its 1:1 thread. Routines record the bot that
scheduled them (additive migration), so schedule_list and schedule_cancel
also cover exactly the routines a bot scheduled for others. Cancelling by
a name that matches several routines is refused.
Tested with unit tests for the per-bot filter, the applyTool guard, the empty
tool list, and on-behalf resolution, creation, listing and cancelling; with
Postgres journeys for on-behalf routines waking across spaces and in the main
conversation and for allowlist enforcement through the scripted runtime; and
with a web E2E for the settings section.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* test(adapters): give the consolidation unit tests the profile lock
The write transaction now takes the per-user profile lock, so the mocked
client answers the lock query, and the first case checks the lock is taken.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* feat(adapters): run profile consolidation on the fast tier when one is mapped
Fact extraction is short structured work, so consolidation asks the model
resolver for the fast tier. The resolver takes a per-call tier and, when the
space has not mapped it, keeps the bot's own tier or model as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* feat: land in the main conversation, led by the main bot
After onboarding the user ended up in a 1:1 chat with "Chief", and the
"Hauptthread" was a group the web client created and found by name. The
main bot had no conversation it actually led, and addressing another bot
there did nothing unless the user had added it as a member first.
Now the main conversation is a real place:
- bots.ensureMain creates or returns the main bot. Bootstrap and space
navigation return `main` ({ botId, groupId, spaceId }) once the main
bot exists; reading navigation never creates a bot.
- The main bot leads the main conversation and answers when nobody is
addressed: message routing, follow-ups, reactions and clearing the
thread fall back to the group's lead instead of its first member.
- Addressing a same-space bot that is not a member yet, by @Name or a
mention chip, makes it join before the message is routed, so it
answers in the conversation. Specialists from other spaces do not join.
- The main bot cannot be archived or deleted, the main conversation
cannot be archived or deleted or have its members edited, and child
bots never dissolve it. Renaming the main bot renames the conversation.
- Web: onboarding uses bots.ensureMain and finishes in the main
conversation. The main conversation is the first sidebar entry in every
space, shown with the main bot's avatar and name; from another space it
switches back to the home space. The main bot is not listed again among
the bots, its 1:1 thread stays reachable by URL. The client no longer
creates or detects the "Hauptthread" by name, refetches a group when it
changes, and shows the speaker of any of the user's bots in a group.
- Onboarding no longer resets the step when its bootstrap resolves after
the user has already moved on, and new spaces open in the app instead
of walking onboarding again.
Tests: group mention routing with a default responder, lead follow-ups,
child-bot guards, a journey for the main conversation (lead answers,
addressed bot joins, guards, rename), authorization for bots.ensureMain,
and an E2E spec that lands in the main conversation, gets an answer from
the lead, sees an addressed bot join, and reaches the conversation from
another space. Existing E2E specs start in the main bot's thread through
a helper or use the main conversation where that is what they test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
---------
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* ci: publish amd64 images under a lowercase name and deploy only after green CI
Image publishing never succeeded: GHCR rejects the capitalized repository
slug, and the arm64 jobs failed on the macOS runners, which have no Docker
buildx. Resolve the namespace in lowercase and build amd64 only until an
arm64 runner with buildx exists; the manifest check expects one digest.
The Evo deploy ran on every push to main regardless of the result of
Product CI, so red commits went live. Trigger it from the completed ci
workflow instead and deploy the head it verified; manual dispatch stays.
The Playwright report publish failed every nightly run because the S3
settings are not configured. The script now skips the gallery with a
notice when they are missing instead of failing the job.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* ci: give each runner on a shared host its own E2E ports
Two Linux runners live on the same host. Every Web E2E job first frees
"stale" listeners on the fixed ports 3110, 3111, 5180 and 4321 with
fuser -k, so a job starting on the second runner killed the Vite server
of the job already running on the first one. The victim died with
SIGKILL (exit 137), and its remaining tests failed on reloads that could
no longer reach a server. This is what made the Web E2E run on this PR
red, and it explains the recurring reload failures on main whenever two
runs overlapped.
The workflow now derives API_PORT, WEB_PORT and WWW_PORT from the runner
name (suffix -N means an offset of N times 100), and the port-freeing
script follows those settings instead of a hard-coded list. The harness,
Vite, the API and the marketing site already read these variables.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* test(testkit): cover the env-driven E2E ports instead of a fixed list
The ports test parsed a hard-coded array that no longer exists. It now
runs the script on a fake host whose fuser finds nothing, and checks the
default ports, the per-runner override, the rejection of a non-numeric
setting, and that the workflow picks the ports before freeing them.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* test(updater): expect the amd64-only publish and the lowercase namespace
The workflow test still described the multi-arch publish. It now checks
that only amd64 builds natively, that arm64 is documented as paused, that
one digest feeds the manifest, and that every image name comes from the
lowercase namespace step.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* ci: keep every runner off the default E2E ports
Branches that still free the fixed ports 3110, 3111, 5180 and 4321 keep
running on the shared host (main, the nightly run, other open PRs) until
they carry the per-runner ports. A job of ours on the runner without a
numeric suffix used exactly those defaults, so such a neighbour could
still kill it, and its own port freeing could kill the neighbour. The
unsuffixed runner now starts at offset 100, so no job from this workflow
touches the defaults in either direction.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* fix(web): recover lazy routes and overlays from a failed module fetch
When the browser aborts a module request while a page loads, the lazy
import rejects and the route throws, leaving a blank page. Chromium does
this whenever a network interface appears or disappears mid-request
(net::ERR_NETWORK_CHANGED), which happens on a laptop that changes
networks and on the CI host every time a container starts or stops. A
Web E2E run hit exactly this: two modules of the onboarding page failed
to load and the suite waited on an empty document.
Lazy routes and overlays now load through a helper that retries the
import once and, when the browser keeps rejecting it from its module
map, reloads the page a single time before surfacing the error.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
(cherry picked from commit d5dacb12f14380071ef14ba8845d18c9d39e3fac)
* fix(web): reload once when the module graph is aborted while booting
The lazy-route recovery cannot help when the browser aborts a module of
the static import graph: nothing of the app runs, the root stays empty,
and the only trace is a console error. A Web E2E run hit this right
after a page reload, when four modules of the shell failed with
net::ERR_NETWORK_CHANGED and the shell never rendered.
An inline script in the HTML entry now listens for the error event of
the module script and reloads the page once, guarded by a session flag
that the entry module clears after a successful boot. The existing
preload recovery stays in place; it only covers Vite's chunk preloads of
production builds.
A new E2E spec aborts one module request during boot and one lazy route
request, and expects the sign-in page to render in both cases.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
(cherry picked from commit e5b7e036e1a60e7e270592c4b9382ee23ed45ccb)
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(adapters): route evOo self-modification requests to code tools
Oos lacked product self-knowledge in their system instructions and were
steered toward browser/computer tools for UI text changes. Add explicit
guidance that evOo changes belong in source code (file/shell tools,
project/ when linked, cloud agent when configured), never via claude.ai
login or browser CSS injection.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(adapters): avoid space query that broke executor test mocks
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(e2e): wait for shell-ready after RPC bot create
createNamedBot now waits for data-ready and the Message composer after
navigating to a freshly created bot, instead of asserting immediately on
a cold page load where bootstrap had not finished.
Also fix onboarding scan advancing to bot creation (onComplete was a no-op)
and use a fast simulated scan when AGENT_RUNTIME=scripted so CI E2E does
not walk the real filesystem during onboarding.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* style: apply biome format to touched onboarding files
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* style: clear biome lint in onboarding scan files touched by the fix
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* style: format onboarding-scan after rebase onto main
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* style: fix biome lint on main files blocking Product CI
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(onboarding): allow scan before first bot exists
The scan RPC required a non-empty botId, but onboarding runs scan before
Chief is created and the UI sent an empty string. That rejected the RPC,
left the flow on /onboarding, and every completeOnboarding helper timed out.
Make botId optional, omit it from the client when unknown, advance to bot
creation when scan completes, and keep scripted/fake runs on the fast scan path.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* ci: add Product CI merge gate and Autopilot auto-merge path
Aggregate lint, typecheck, builds, unit, Postgres journeys, and Web E2E
into a single Product CI check for branch protection. Soft-fail artifact
uploads and skip headless macOS screenshots so infra noise does not block
merge. Add autopilot/auto-merge workflow and MERGE_GATES.md.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): keep autopilot auto-merge inert without label
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): run autopilot auto-merge on self-hosted runner
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): free stale E2E ports before Playwright bind on Linux
Self-hosted evoo-ci runners can retain listeners on harness ports
(3110 API, 5180 web, 4321 marketing, 3111 device control) after a
failed run, causing EADDRINUSE before the next job. Add a guarded
free-e2e-ports.sh that uses fuser on Linux only and never touches
live dev ports 3100/5173. Wire it into the Playwright workflow and
the E2E harness with clear logging.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): stabilize message More menu Escape in hover E2E
The desktop hover spec assumed Escape always closed the first page-level
menu within one keypress. On self-hosted Chromium the Copy menu could stay
open while the rail remained visible, failing getByRole('menu') count.
Wait for the Copy item to prove the menu opened, then retry Escape until
that item hides while asserting the More control stays on the hover rail.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): retry sign-up navigation in Playwright helpers
Late-suite specs can miss the sign-up heading when Vite is still
warming on self-hosted runners. Retry goto until the form renders
instead of failing once at the default timeout.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): restore browser preview module and onboarding e2e
The web bundle imported a browser-preview store that was never added, so production builds and every Playwright signup failed before the shell rendered. Mobile check failed because Expo packages were a patch behind the SDK.
Restore the store, type the optional desktop bridge, and walk whichever onboarding step scripted CI actually opens.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): stop the blank web client and the send 500
The core barrel pulled git checkout into the browser, so node:child_process crashed the page before sign-up. Serve that module from a node export instead.
Add the missing run tool-call budget columns that threads/send writes. Guard an indexed shell flag for typecheck, and expect the files key the desktop preload already exposes.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): isolate pnpm setup and the takeover membership
The Team Computer takeover journey inserted a space member the
default-space trigger had already created, so shared Postgres rejected
the unique pair. Self-hosted jobs now keep a per-job pnpm home and
store, and artifact uploads no longer fail a job that already passed.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): clear typecheck, lint, and unit failures
Inbound channel rosters can mark a participant list as authoritative.
Answered channel asks stay off the group, and clearing a thread waits
until delegated returns land. Lint, unit, and web e2e contracts follow
the current shell, composer, and settings.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): align e2e with the shipped shell and onboarding
Web E2E still expected older focus buttons, nested project rows, and
settings that never opened MCP. Persist the signed-in space, list
non-default spaces under Projekte, and keep the quote control on screen.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): open group reply from a revealed hover rail
The first user row covered its Reply control until the rail accepted
pointer events. The group transcript now reveals that rail, then replies.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): stop the shell render loop and settle computer and mention UI
A working bot rewrote preview state on every render, the computer menu never stayed still, and the mention picker did not expose its expanded state or option name.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): stop idle probes hanging and accept a follow-up send
Background probes walked every process fd and timed out on the shared runner. Linux now tracks the marker with flock and /proc/locks. A new composer draft can send while an earlier send is still refreshing, and group members stay mentionable.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(adapters): detect background work when flock ids are padded
The idle probe compared /proc/locks with an unpadded device key, so majors
or minors under 16 never matched. It deleted the live marker and reported
idle while the command still held the lock. Probe by taking the lock, and
print the cancel key the way the kernel does.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(adapters): stabilize background-work probe tests on self-hosted CI
Clear stale marker files before each probe scenario and allow longer
polling so launch can acquire flock on busy persistent runners.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): do not block onboarding on spaces list
Render the first onboarding step from me and integration setup alone.
Resolve the optional skip-to-app redirect after that so Playwright
signups are not stuck on the loading shell when spaces.list is slow.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): restore green gates after second main rebase
Re-apply biome formatting on Shell/SortableGroup and strict TypeScript
assertions in computer-idle tests dropped when skipping the duplicate
conflict-resolution commit.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): wait for onboarding bootstrap before E2E proceeds
Expose data-onboarding-step on the onboarding shell and have Playwright
wait until loading finishes. Ignore ad-hoc screenshot scripts on
self-hosted runners so they do not fail lint.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): fall back when onboarding bootstrap stalls
Race me/integration setup against a timeout and promote loading to the
emoji step after 20s so Playwright is not blocked on a hung RPC.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): render onboarding immediately for E2E signups
Start on the emoji step instead of a loading shell that waits on RPC.
Bootstrap can still redirect once me and integration setup resolve.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(ci): stabilize web e2e on self-hosted runners
Per-job pnpm stores broke node_modules symlinks when setup-node restored
cached installs, so Vite could not resolve react during dependency
optimization. Keep a runner-scoped store, clear stale Vite caches, and pin
Playwright to one worker in CI to reduce onboarding RPC contention.
Add Vite resolve aliases for react/react-dom, wait for an interactive
onboarding step in E2E helpers, and include the React hook smoke spec from
main.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): resolve React aliases through pnpm package layout
Root node_modules/react does not exist under pnpm, so Vite aliases pointed
at missing paths and broke production builds and the Playwright webServer.
Resolve react and react-dom from the web package and map jsx-runtime to the
actual entry files.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(core): type the lazy child_process spawn cache
Main's lazy require left spawn typed as null after assignment, which broke
Typecheck on the rebased branch.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(db): mock git-checkout export path in unit tests
Vitest resolves vi.mock against package exports; the .js suffix is not
declared for @rakazo/core/git-checkout.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): stabilize E2E after main sidebar and onboarding changes
Use first() for duplicate transcript lines, wait for interactive onboarding
before model-required assertions, and allow project sidebar to settle after
create. Extend the Playwright job timeout so the single-worker suite can finish.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): restore Projekte sidebar and integrations MCP entry
Use the default space for Unterhaltungen/Projekte separation instead of the
active space header, persist the bootstrapped space in localStorage, wire
SettingsOverlay to MCP management, and compare hover tones on inactive rows.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): satisfy biome format on homeSpaceId line
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): align sidebar E2E with nested Projekte shell layout
Scope Unterhaltungen assertions to conversation rows only, target project-space
Chief actions under .sidebar-projekte, and stabilize hover-tone polling.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix: satisfy lint after main merge (format and unused var)
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): show project-space 1:1 bots in Projekte sidebar
Include lone bots in projectSidebarGroups, render bot rows in Projekte,
and stabilize E2E for nested sidebar layout and duplicate group entries.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): disambiguate duplicate group sidebar buttons in E2E
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): align spaces and group-chats E2E with Projekte sidebar IA
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): dedupe Projekte sidebar rows and stabilize E2E specs
- Deduplicate project space and group-chat entries in projectSidebarGroups
- Scope group-chats/spaces E2E clicks to first Projekte match after IA split
- Wait for shell ready before mention-picker keyboard test; restore main locators
- Fix PrivacySettingsOverlay lint and align biome schema with CLI 2.5.13
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): restrict project drag handle and stabilize space E2E
- Move SortableGroup drag listeners to ProjectDragGrip only so sidebar clicks work
- Scope space delete actions to project-space rows; wait for menu in E2E
- Switch group settings E2E via direct navigation after unsaved draft check
- Merge main (DataTrustPanel) and fix Shell import order for CI lint
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): stabilize message hover Escape assertion in E2E
Assert the More menu closes and the hover rail stays visible instead of
requiring focus restoration, which is flaky across Chromium builds.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* chore: format apps/api/src/app.ts after main merge
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(web): stabilize team-computer E2E send after bot switch
Wait for shell ready and use the Send button instead of Enter after
switching bots and closing the computer overlay. The prior run timed
out on threads/send with only an apps/api format commit in between.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
- Move spawn import in git-checkout.ts to lazy require() to prevent
node:child_process from being included in browser bundles
- Add ./git-checkout subpath export in package.json
- Update project-workspace.ts to use direct subpath import
- Update test mock to match new import path
Fixes: Vite crash from child_process externalization error,
API crashloop from missing readGitCheckout export
- Add Vite resolve.alias for react/react-dom to root node_modules
- Removes duplicate React instances from pnpm transitive deps
- Create E2E smoke test verifying no React hook errors
- Fix: pnpm resolves @dnd-kit deps to react@19.2.8, conflicting with
app react@19.2.3 causing 'Invalid hook call' at runtime
Fixes: dark screen / Invalid hook call on /app
python -m, stdin programs, and deno/bun eval or kill-named scripts could still tear down a managed graphical desktop after the shell guard tokenized the command.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
The desktop-protection shell guard inspected shell syntax and nested
bash -c, but not script files passed to python, node, perl, ruby, or
php. Read small workspace scripts and reject kill-family contents, and
fall back to a kill/pkill/killall/xkill filename when the path cannot
be read.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
The Live Trading control stored its mode in localStorage, so the shell could
claim live money while the bot was in dry-run. Show the Tradingbot status
when TRADING_BOT_URL is set, and an unknown state when it is not.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
The desktop-protection shell guard only walked nested shell -c programs,
so interpreter inline code could still invoke kill/pkill. Inspect python/
node/perl/ruby/php -c/-e/-r payloads for the same lifecycle kills.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* Voice UX: simple speak toggle, onboarding pick, chat voice change
Replace Kokoro-branded settings with a Kann sprechen / Spricht nicht toggle,
onboarding voice personalization, and local chat commands for voice changes.
German voice presets (Eva, Bernd) surface as friendly names; Kokoro stays
under the hood via KOKORO_BASE_URL.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* Fix import order in voice onboarding and settings
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(adapters): stop desktop shell from inheriting host secrets
The this-mac shell spawned with the server process environment and ignored
the command environment, so Autopilot could read deployment credentials that
are not on the run redaction list. Drop credential-shaped host variables,
then apply the request environment.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* style(adapters): biome format desktop-sandbox after env scrub
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
Extend the existing NewSpaceDialog project flow so an optional host
folder becomes the space workspace root. When .git metadata is present,
remote URL, default branch, and head commit are stored on the space.
Team Computer file APIs expose the linked folder under project/ without
a second VCS stack. Desktop adds a native folder picker; web falls back
to an absolute path field.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
When a run worker loses its lease while the runtime yields no events
(long tool or model work), lease loss was only detected on the next
heartbeat renewal (up to 60s). A reclaimed worker could execute in
parallel during that window.
Add a dedicated lease poll that aborts stale workers within a few
seconds and call runtime.abort so pi-runtime stops promptly.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
The executor calls finalizeRun with toolCallBudgetCount and
toolCallBudgetExceeded in the updateMany data clause, but these
fields were missing from the Prisma schema, causing an Invalid
updateMany invocation error that blocked all runs from completing.
Also updated CLAUDE-RESUME-NOW with full infrastructure status.
The executor loaded every agent secret in a Space and injected decrypted
values into shell environment variables for any member's bot run. Restore
agentSecretScopeForRun scoping accidentally dropped in 0a8a831.
Add an executor test that fails when findMany is scoped only by spaceId.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
When a user records a taught skill on one Team Computer bot, manual
takeover on another bot in the same space cleared the teaching lease
without stopping the session. Teaching input then failed silently and
skills/stop could not release the replacement takeover lease.
Reject computer/takeover while another bot on the same computer is
recording, matching the existing Stop teaching first messaging guard.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
Team mode stores one disk per Space. computer/readFile and computer/files
only verified the caller owns the entry bot; an explicit bots/<otherBotId>/…
path let one member read another member's workspace. Wire existing path guards
into the API router and executor file/shell tools, and add a shared-Space
authorization regression test.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
claimSteering matched steering rows by runId even after they were
claimed, and still returned them when updateMany updated zero rows.
That could inject duplicate user context into an active run.
Filter findMany to claimedAt: null and return nothing when the claim
update does not land.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* fix(api): scope routine remove to acting user in shared Spaces
routines/remove only filtered by spaceId, so any member of a shared Space
could delete another member's routine by ID. Match update/testRun by also
requiring userId in the lookup.
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* test: drop unused ownerActor in routine remove authz case
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
