メインコンテンツへスキップ

Agent Features

This catalog records what each Grida Agent capability does, the outcome it should preserve, and the limits that affect use. Desktop chat on macOS and Linux and the headless grida-agent host are the supported surfaces; see Supported surfaces. Within Desktop, availability also depends on the selected model connection.

Present marks a capability that works as described on the supported surfaces. Partial identifies narrower availability or behavior. Deferred marks a deliberate scope limit. Open marks a capability whose product behavior still needs a decision. Linked guides describe what to do; linked scenarios and tests record the evidence behind the described behavior.

Saved LAN HTTP model endpoints stay in Desktop settings but cannot run chat; use an approved HTTPS endpoint or a loopback HTTP endpoint.

Feature catalog​

FeatureCurrent Grida implementationProduct expectationBehavior
Task inputPartial. Home and workspace use the same composer controls for files/images, Plan/Goal, model and effort. Home stages the first request until the chosen workspace opens; unsupported settings or unavailable attachments stay visible for correction instead of sending a reduced request. Desktop accepts one attachment by picker, drop or paste. Explicit PNG, JPEG, WebP and GIF images on a compatible model send bounded prepared pixels in the first inference, plus the verified original disk path or truthful retained-snapshot identity. The original and prepared model image remain separate; downsampling and one-frame animation handling are identified. Source images can be up to 20 MiB within decode limits; the prepared PNG is bounded to 384 KiB, 2,048 pixels per edge and 2 megapixels. A text-only model or unsupported host preserves the entire image draft. Ordinary disk files, including small files, and folders are references for inspection on demand under existing permissions. Folders have no total size, traversal, copy or access grant. Sent images appear alone outside the text bubble and preview the retained original; other attachments keep cards, including video stills. Previews cause no inference. PDF, SVG and HTML rendering and video perception are deferred. Attachment-bearing messages cannot be edited later. Attach an image, files, folders.Accept supported material durably once. Keep supplied items identifiable and show any omitted or unusable material before acting.Input
Projects and foldersPartial. Desktop chat works in one workspace folder. On macOS and Linux, a deck opened in its own window has chat scoped to the deck bundle, including a deck inside a project; a nested deck uses the default model and its picker offers effort only. Headless runs name a primary and additional directories through a trusted permission profile. Selected folders, nested deck caseGroup named directories with a primary default. Show each action's destination and check access separately for every root.Projects
Slash commandsPresent. In Desktop, / lists the commands the chat supports: /plan and /goal, plus /side, /stop, /compact and /fork when the host offers them. The list does not change while the agent works. Completion closes on an outside click or Escape without clearing the draft; typing or moving back into the trigger can reopen it. The + menu shares its full-composer placement. /plan and /goal add a composer badge without starting work; during a turn the Plan message, or the goal start, waits in the queue, and /plan turns steering off. /stop stops background processes and says when none are running. /compact and /fork say they are unavailable while the agent works; /compact also needs an open, empty queue, and /fork copies the latest committed message. ACP clients can send /compact as a prompt. Guide, pane, host commandsKeep commands discoverable and preserve their meaning across client surfaces.Commands
Skills and instructionsPartial. On macOS and Linux, AGENTS.md in a readable primary project folder loads automatically. Headless run --skill and ACP clients select skills; a name resolves from .agents/skills, then .claude/skills, then the bundled dotcanvas, slides and svg. Each accepted message keeps the instruction and skill versions it captured, and file changes apply to later messages. A project-skill selection includes the skill's workspace-relative folder, read under existing permissions. Without a selection, a chat with project file tools, including Desktop workspace chats in Act and Plan, lists each skill's name and description; the agent reads a bundled skill with read_skill and a project skill with its file tools when the request calls for one. Guide, source, testsRetain explicit selection and inspectable instruction sources. Make scope, overrides, refresh, and missing sources clear across continuation.Skills
Models and providersPresent. Desktop runs chat on a connected ChatGPT subscription, Grida Gateway, saved OpenRouter or Vercel AI Gateway keys, or an approved compatible endpoint. Each chat durably remembers its connection, wire model and effort, including unsent changes. Opening a chat restores its pair without changing the new-chat default; a new chat snapshots the complete pair from the last valid explicit model or effort change. Fixed-model document contexts retain their model and inherit effort. A saved unavailable choice stays identifiable until resolved. Legacy identity is restored only when verified; otherwise the user chooses explicitly. Accepted work retains its model, account, effort and funding; changing future selection cannot reroute it or stop another chat. A shared worker may block activation until retained work is resolved. Picker reloads preserve selection and offer Retry after failure. Subscription offerings use Grida's compatibility list, not account-discovered model availability. Choose a model.Honor the selected connection and explain sign-in, quota, and connection failures. Use bounded, interruptible retries when the provider verifies it rejected the request before execution.Connections
ConfigurationPartial. The workspace context input limit is the supported setting, kept in the workspace's config.toml. The agent reads and edits that file with its ordinary file tools; every change asks for approval, no earlier approval covers the next one, and commands cannot reach it. An invalid edit is refused with its reason; a saved change applies to later messages. Desktop has no settings control; a manual edit applies when Desktop next starts the workspace's agent. ACP clients can also inspect, change, reset and reload it. Change agent settings, referenceLet people inspect, explain, change, and reset supported settings through the agent or UI, with scope, provenance, validation, and activation timing. Initial fields require individual review.Settings
HooksPresent. Desktop reads ordered global hook definitions only from the user's config.toml. Settings → Hooks inspects definitions and diagnostics, opens/reloads the file and reviews executable trust. Workspace hook JSON and import management are not used. Fresh connections capture trusted definitions; running and paused work keeps its selection. Existing session-start, input, before/after-tool and completion events retain required/observe behavior and scoped execution. Source-built headless supervisors supply explicit permission profiles separately. GuideAdd inspectable lifecycle commands for context, tool checks, and completion verification. Required-check failures block their action. Preserve completed effects and respect Stop, Pause, and task limits.Hooks
Usage and costsPresent. Turn summaries record reported tokens and exact Gateway request charges, including shared subagent evidence. Desktop shows a context-use ring labelled as estimated or reported, and leaves charges to the billing record. A native client starting a Gateway Goal can set a provider-enforced spending ceiling shared with funded children; unknown outcomes keep their holds, and paid or unknown external tool lanes are withheld from capped Goals. Desktop goals have no spending ceiling. Other connections keep step allowances and reported usage. Goal limits, finished turnsLabel estimates clearly, distinguish charges from estimates, and explain exhausted allowances and credit-related failures.Usage
Agent effortPresent. Effort is remembered with each chat's model; explicit changes save the complete pair for new chats while accepted work keeps its captured effort. Desktop offers one effort scale: Low, Medium (default), High, Extra high and Ultra. Low through Extra high set the reasoning level on models whose provider binding declares it; other models keep their own reasoning. Ultra extends the inference-step allowance and permits one independent reviewer through the same captured connection on every supported native model connection, including subscriptions and compatible endpoints. It also selects the highest supported reasoning level when the model accepts one. Step bounds apply independently of optional provider output-token bounds and do not promise quota or monetary ceilings. Goal turns keep the model's default reasoning at every level; Ultra adds only the review allowance. Headless run --effort ultra selects the highest level when the launch declares reasoning support; headless runs have no option for Low through Extra high. Choose effort, headlessOne visible effort choice whose effect is observable, with unavailable levels explained.Agent effort
Persistent goalsPresent. Desktop starts a goal from /goal or + → Goal, also while the agent works (the start waits in the queue), with the product's default allowance of eight work turns and 64 inference steps, keeps its objective, progress and completion confirmation across turns, and offers clear, pause, resume and edit. Ultra goals on every supported native model connection share their inference allowance with one independent review on the same captured connection. Native clients can also set success criteria, allowances, manual continuation, a named reviewer, a Gateway spending ceiling and a check-in through the goal protocol. Retry, restart and objective edits keep the original allowances; a child cannot create a fresh monetary allocation. Guide, native clientsA goal keeps its objective, allowance and progress across turns and reports completion only when the objective is met.Goals
Plan and ActPresent. /plan adds a removable Plan badge; selecting it starts no work. A Plan message can read permitted material and ask questions; commands, file changes (including settings), editor tools, image generation and check-ins are unavailable, even where Act may use them. Each sent or queued message keeps its mode through queue edits and restart. To carry out a plan, remove the badge and send a new message. Headless clients select a launch default or negotiate per-input selection. Plan before actingKeep the operating mode independent of agent effort, including across queued work and recovery.Planning
PermissionsPresent. Desktop workspace chats start with the Default preset: project reads, writes and sandboxed commands without asking. Desktop does not ask for access beyond the project: file tools stay inside it, commands have no network, and no additional folders can be granted, which diverges from the policy's Ask default for outside paths. Trusted headless profiles name the folders, operations and actions that need review. Full access is an explicit, in-app-confirmed per-chat choice on compatible native builds: commands use ordinary OS-user files and networking without confinement, file tools gain a filesystem root, and credentials/state readable by that user lose their isolation. The selection survives reopening; new Desktop chats remain Default. Changes require an idle chat, an empty queue and no running processes. Accepted work and workers keep captured authority. Older hosts retain Default. No automatic reviewer is required. Guide, policy, source, testsNew Desktop chats ask before granting additional access. Bind approval to the exact action and scope, with host limits effective across tools and permission changes.Permissions
Automatic reviewPartial. A trusted headless launch can select a reviewer for one exact file read. Ask remains the default. GuideReview eligible actions within existing host limits. Allow manual answers when review cannot decide. Broader eligibility and a Desktop selector remain open.Review
ProgressPartial. Streaming replies, pending questions, expandable tool progress and Ultra review activity are shown, and finished turns group their work. A provider retry wait keeps the turn running without a separate status in Desktop. Detailed model reasoning stays hidden. A tool call whose arguments the model wrote malformed gets a failed result explaining that the arguments could not be parsed, and the model can send it again within the same turn. Read a finished turn, provider retryShow durable task, interaction, child-work, retry and waiting status, including clear exhausted or blocked outcomes. Let the model correct a malformed tool call within the turn.Progress
SteeringPresent. In Desktop, Steer on a waiting message sends it into the running turn at its next decision. Messages queued with Plan, Ultra or a goal, and non-text messages, wait for their own turn. A message leaves the queue before it is offered; if the turn refuses it, the text returns to the composer as an unsent draft. ACP clients use the steer operation. GuideAdd acknowledged updates to the current task at recorded decision points.Steering
QueueingPartial. Messages sent while a turn runs or waits join a visible queue. Hovering or focusing a waiting message shows its full multiline text without editing it; image attachments have leading thumbnails when an authorized preview is available. Waiting messages can be dragged to reorder or removed. Supported runtimes can edit text and attachment captions in the main composer, including clearing a caption while keeping its image or file; a captured Plan or Ultra choice is shown and kept on edit. Every row retains its queue icon beside any thumbnail. Save replaces the queued text; Cancel or Escape restores the ordinary draft and its attachment/modes. Failed saves retain the edit for retry, and steering waits until editing ends. Stop or an interruption holds the queue until Resume. The captured model is not shown, editing cannot change captured settings, and goals cannot be edited. Attachment caption editing requires the negotiated native operation; older runtimes retain plain-text editing only. GuideShow the model and settings captured with each waiting request and allow explicit setting changes on edit. Preserve unsent drafts and give each request one durable disposition.Queue
QuestionsPresent. Desktop supports choices, text and Skip. Answers can reach the current turn while independent work continues; follow-up messages queue separately. Pending questions survive restart, and answering keeps an explicit Pause. A source-built headless run --questions returns a retained input_required question; ACP clients enable questions by negotiating both question operations. Source, testsAn answer given after Stop or a history change keeps its question and receives an explicit disposition. Questions explain what the answer affects.Questions
Pause and resumePresent. In Desktop, the composer's main button pauses the running turn at once; ▶ resumes the same turn. A reply cut off by the pause stays visible, the model is asked to continue from where it stopped, and completed tool work is kept. Sending a message instead ends the paused turn as stopped and runs the message next; Desktop does this in two steps, so if the send fails the paused turn cannot be resumed and the message may wait in a held queue until Resume. Paused turns survive restart with their retained input and captured model connection, including after an earlier between-turn model change. ACP clients use the pause and resume operations. Guide, testsWhile pausing, show work that must still settle. Keep accepted external jobs visible.Pause
Background notificationsPartial. Desktop workspace chats can notify when a turn finishes, fails or ends with an uncertain result, or when it waits for access approval, while the chat is out of focus; a cancelled turn is silent. A pending question raises no notification, so reopen the chat to answer it. A click opens the chat on its original model connection, waiting for a starting chat to connect. ControlsOpen the exact chat on click. Show no private chat content in system notifications.Notifications
StopPresent. Stop ends the active turn, labels it You stopped, leaves an interruption note for the next turn, and holds queued inputs until Resume queue. In Desktop chat, sending a message while a turn is paused ends it as stopped. Headless and ACP clients use the stop operation, and a signal stops a headless run. Guide, interruption note, queue holdEnd the active turn with a partial account and persistently pause queue advancement.Stop
Finished turnsPresent. Each finished Desktop turn collapses its commentary and tool activity beneath a Worked for … header, with the final answer visible; a turn that only answered shows no header. Expanding shows the recorded work in order, with consecutive actions summarized on one line. Headers persist after reopening and on earlier pages. Stopped, failed and uncertain turns are labelled; review partial work before retrying an uncertain one. Images appear below the latest finished turn's answer only and leave the chat once another turn starts; generated images stay in Tools → Recents. Read a finished turnShow retained results with every finished turn and link more result types from the retained account. Activity wording follows tool classification only; no extra model request describes the work.Finished turns
ReconnectPresent while the host runs. Closing the pane or switching chats leaves work running. Reopening loads retained history, then live output, and starts no model request; late output from a previous chat stays out of the newly selected chat. A refused send keeps the draft, including its attachment and Plan badge. Guide, host test, controller testsRestore the same work, preserve rejected input, and keep late output in its original chat. Viewing status grants no execution authority.Reconnect
Chat managementPresent. Chat-specific model and effort survive switching and restart; restoring them never changes the new-chat default. Desktop lists active chats in the chat picker; Settings → Archived chats lists local workspace archives and restores them without opening a chat or starting model work. Chat options retains rename/archive/delete. Deletion requires an idle chat with an empty queue, a finished or cleared Goal and no unresolved or uncertain work. Deleting removes the chat's private data and named subagents; project files and independent forks stay. Archiving keeps work running. GuideKeep deletion eligibility clear for outstanding work. Archiving keeps work running.Saved chats
Restart recoveryPartial. Questions, approvals and paused turns survive a restart, and a paused turn resumes with its retained input. Reopening only loads retained work: waiting messages start after an explicit Resume. Work whose provider outcome is unknown is shown as result uncertain and is not sent again. A changed provider credential blocks resuming a retained task. Working files have separate lifetimes. Guide, master tests, process testsReconcile uncertain work before retrying. Preserve pending decisions and make continuation, interruption, and unavailable results explicit.Recovery
Long conversationsPresent. The host can compact older context before eligible queued Plan or Act text, including after a provider reports a context limit; the failed turn remains recorded and is not replayed. Manual compaction starts from an idle chat and can summarize a bounded prefix while keeping later turns, even when the full context exceeds the projection limit. The original transcript remains available. Desktop opens the latest messages and pages backward without rerunning work. History guide, contextRetain original history and continuation evidence. Failed or stale compaction preserves usable context and new input. Explain limits after bounded recovery.Context
BranchingPresent. Desktop /fork creates a separate chat from the latest committed message; ACP clients can fork from a selected committed message. The source keeps running. Both chats work on the same workspace files under the host's current permissions; approvals are not copied. Copied context carries portable text and reports omitted media, tool material and provider continuation. Guide, ACP, testPreserve the original chat, identify the branch point, and distinguish copied context from shared files and permissions.History
Side chatsPartial. In Desktop, /side with a question opens a new temporary side chat in its own workbench tab and sends the question there; /side alone shows the side chat used last, or opens one. A chat can keep up to eight side chats open, each in its own tab (in a file window, the latest shows under the chat); closing one leaves the others. Hovering a side-chat tab shows its latest visible user and Grida messages without selecting the tab or starting work. The main conversation is unchanged. A side chat starts empty: the copied conversation is model context only, and its transcript, history and edits cover its own messages. After its first message it is titled like a chat, and its tab shows the title instead of Side chat. The tab shows the same chat view and message box as any chat, without chat switching or the controls a side chat refuses; model and effort start as the parent's. A side chat is an ordinary chat forked from the parent's committed conversation: same model, tools and permission rules, approvals asked the same way, live output, queueing and stopping. A side-chat policy and a history boundary keep it to questions and non-mutating exploration unless the person asks for a change; the policy steers the model and does not narrow its permissions. A long copied conversation is summarized automatically before the first side question and stays background. Goals, scheduled follow-ups, subagents, delegated effort, forking, manual compaction, nested side chats and catalog changes are refused and listed in its snapshot. It and its title never appear in the chat list, and it is discarded with its data and processes on close, parent deletion or switch, disconnect, or the next start after a crash; Desktop keeps retrying a close or an open whose outcome it could not confirm until the host settles it. Only the Desktop window that opened a side chat can show or drive it. ACP clients forward answers themselves. Guide, ACP guideAnswer side questions in temporary forked chats kept apart from the parent, with explicit context, the parent's permissions and bounded retention; an answer reaches the main task only when the person brings it there.Side chats
Earlier-message editingPartial. Desktop edits a retained text message in its original bubble, with multiline input and Cancel/Send actions that preserve the unsent composer draft. Editing requires an idle chat with an empty queue and a resolved Goal; a message that started a Goal, or one with an attachment, cannot be edited. A pending edit locks its text and Cancel action; an unconfirmed reply offers a retry of the same request. The host retains the previous view, which ACP history can read and Desktop does not yet show; effects of earlier work stay. GuideMake prior views directly inspectable in Desktop.Editing
Artifact presentationPresent. Images support zoom, copy and download. Desktop chat can reopen and save a retained PNG crop. Generated images appear below the latest finished turn only and stay in Tools → Recents, which can retry saving a copy. Supported workbenches can place a retained generated image in an existing folder after confirmation, with overwrite protection. Results guide, previewShow previews during work, respect navigation, and continue when a viewer closes. Distinguish generation, saving, availability, and verified display. Recover completed results after save failure.Results
Visual toolsPartial. Desktop Act on supported saved OpenRouter or Vercel connections can generate an image and keep it in Tools → Recents. On an OpenRouter model that takes images, generation can use one Library choice, attachment or workspace PNG; local PNGs need explicit transfer confirmation. The workbench agent can ask for a Library reference that you refine, choose or decline. The agent can view PNG, JPEG, WebP and GIF paths on a connection that returns image-bearing tool results, including supported subscription Responses, Responses API, compatible/OpenRouter and Gateway profiles. A user-image-capable direct Gemini profile does not yet offer that tool. Viewing uses bounded prepared pixels and normal read permissions; it does not broaden generation-reference transfer. Deferred: the agent cannot render SVG, HTML or documents it writes to an image, so its visual claims about its own output are not pixel-checked. Generate an image, choose a reference, view an imageKeep tools capability-scoped. Preserve requested properties such as transparency and references, and identify the visual checks actually performed.Visual work
Editor context and editsPartial. Desktop Act reads and opens workbench files, and can edit selected SVG text in a workbench, standalone SVG or slide through the editor's Undo and Save. Under the Default preset, Act can also write workbench files directly on disk; those writes are not on the editor's Undo stack. Open documents are expected to reload from disk, with unsaved edits deferring the refresh, but that has not been verified for agent writes. SVG text, refresh caseUse document-owned operations and representations. Preserve the addressed target and report the document tool's edit and save outcomes.Live material
SubagentsPartial. On supported native connections, ordinary Act turns can create up to eight general workers independently of effort, with captured model/reasoning/connection, delegated native permissions and skills, and bounded child allowances. The parent collects exact assignment results and resource-retirement evidence before successful completion. Information sent to an active worker joins its current assignment for the next model request; the receipt confirms recording rather than consumption, and sending never wakes paused or finished work. Workers receive one parent-turn-scoped assignment; the native host automatically retires their remaining processes when it ends. Workers cannot delegate again or inherit live editor handles; additional approval is unavailable inside a worker. Plan, persistent Goals, side chats and existing child roles retain narrower contracts. Desktop's read-only, parent-scoped Subagents view separates Active and Done history and opens actual live/retained child conversation with captured metadata. On compatible hosts, start/finish events interleave with parent commentary and individual tool activity using retained publication order, including after reopen; missing historical order has a labeled fallback. In a workspace, transcript indicators and lifecycle links open or reveal the same main content tab, while the parent chat and composer remain usable. Closing its tab only hides the view; passive updates never focus it. Standalone file windows retain a chat-local viewer. It offers no manual creation or execution controls. Observation is independently negotiated; older connections retain read-only bounded inspection, with incomplete counts/history explicit. Private ACP named-client assignments and Ultra's bounded independent reviewer remain available. A child cannot be closed individually; deleting the parent deletes its children. Guide, interaction caseDelegate independent work through the main agent and inspect each child's progress and conversation without leaving the parent chat.Subagents
External agentsPartial. The native CLI can run one text turn with a reviewed local ACP peer using read-only workspace access and no network, and retains the text and outcome. The Linux Insiders Local agent preview runs a provisioned peer read-only with Stop and reconnect. Normal backend selection and other distributions remain open.Advertise each backend's actual tools, controls, execution location, and recovery guarantees.Agent choices
MCP tool integrationsPartial. Headless Act can invoke an allowlisted set of tools from one trusted, authenticated numeric-loopback MCP server per launch. Development and Insiders Desktop builds can bind one such server to a selected workspace on Gateway through launch options; Desktop shows no server list or discovery status. Only Act uses these tools; Plan and goals omit them, and a side chat has its parent's. Retained work keeps its original connection, and a lost connection needs a host restart. Native clientsShow each server's tools and availability, make discovery failure visible, and extend scoped configuration to ordinary Desktop use.Integrations
Grida tools for other agentsPartial. Development and Insiders Desktop builds export a read-only list of the active and open workbench files. The standalone MCP host and headless Act offer surface_open, which reports that no editor is attached. No document edit operations are exported yet. Native clients, guideExpose reviewed Grida capabilities through MCP independently of the built-in agent, with scoped access, shared operation definitions, and document-owned results and errors.External access
Command executionPresent on macOS and Linux. Desktop workspace chats run commands in the project under the Default preset: login-shell environment, no network, writes limited to the project and the system temporary folder, and existing .git and agent metadata read-only, so the agent can start a repository but not commit to an existing one under Default. Explicit Full access instead uses the unconfined native backend for ordinary-user files/networking, retaining process ownership and honest cleanup evidence; escaped descendants may survive on either platform. A command still running after its short wait, such as a development server, stays with the chat as a numbered process: later turns read its output, type into it or interrupt it, and ending or stopping a turn leaves it running. Stopping it, deleting the chat or quitting ends it; a chat keeps at most 64. Private ACP clients list and stop a chat's processes and receive the list as it changes; Desktop reuses that list and its Stop controls in the bottom Workspace utilities panel's Processes tab for the selected parent chat, including host-retained ended entries when none run; the sibling Terminal tab keeps the separate interactive shell; its app interaction is untested in the app. Trusted headless hosts select the same capability through a launch profile. Plan withholds the command tools. Windows is deferred. Permissions, stop running processes, confined commandsRun programs, including long-running and interactive ones, under scoped native macOS/Linux confinement, with the chat able to read, interrupt and stop what it started. Native Windows is deferred.Execution
Scheduled wake-upsPartial. Private ACP clients can allow one check-in on an idle Act message or on the first turn of a non-Ultra Act goal. The chat shows its purpose and due time; the check-in can be cancelled, Pause holds delivery, and after a restart its schedule is restored explicitly. Desktop chat does not offer check-ins. Client guideLet the task resume when due under its existing permissions. Pause holds delivery. Cancelling a check-in leaves the task available to continue explicitly.Wake-ups
Reusable personal memoryOpen. No personal-memory store; conversation history is the only retained context.Define retention, inspection, correction, and removal before committing the feature.Memory

The configuration contract records the reviewed input-limit field and shared inspect/update operations.

Additional media HTTP/SDK operations are documented by the retained host package. Their presence does not make every operation a tool available to the agent.

Supported surfaces​

SurfaceScope
Desktop chatmacOS and Linux. The web-rendered chat runs its agent through the native host. Windows keeps file and media workflows; chat is unavailable there.
Headless grida-agentmacOS and Linux, built from source. Scripted run, ACP over stdio or a local WebSocket, and the standalone MCP tool host. Client guide
grida CLIGateway chat under your Grida account and local external ACP peers. Use your Grida login
Retired surfaceWhat it providedDisposition
Web SVG assistantHosted chat and tool execution for the browser SVG editor.Hosted execution, chat panel and assistant controls removed. SVG editing and storage retained.
Web canvas assistantHosted chat and design-tool execution for the browser canvas editor.Hosted execution and the Agent tab removed. Canvas editing and inspector retained.
Earlier Desktop chatChat the hosted editor served to Desktop builds without the native host.Removed. Those builds show no chat; update Desktop to use chat.

Desktop pages open only in Desktop; use the headless host for headless development.

Checking coverage​

Review each capability from accepted input through authority, execution, observation, interruption, recovery, and eventual retention or deletion. Check its interaction with concurrent work and unavailable dependencies. A missing answer belongs in the owning contract before another feature label is added.

The experience scenarios describe observable outcomes, including failures and interrupted work. Existing test coverage and remaining proof live in the harness assessment and regression obligations.

Supporting systems​

These are required working mechanisms behind the features above.

SystemCurrent implementationRequired contract
Scratch working storagePresent. Explicit raster input retains the exact source separately from bounded prepared model pixels and any cropped working result. Queueing, retry and restart use accepted immutable bytes. Headless selected general-file inspection keeps its existing bounded retained-original and working-copy workflow. Missing required bytes stop dependent work; preview reads do not authorize tools or inference. Attachments, file inspection, scratch.Provide isolated working copies, clear retention, safe cleanup, and recovery of required bytes. Scratch
Headless execution hostPartial. One native grida-agent executable serves Desktop over private stdio ACP, scripted run, serve over ACP stdio or WebSocket, and serve-mcp, on macOS and Linux. Windows is deferred, and live provider and operating-system qualification is limited. Entrypoint, tests, ACPProvide one native engine/server executable for Desktop, automation, and headless behavioral testing. ACP carries shared interactions, with negotiated richer controls and custom client presentation. Host design
Filesystem and workspace bindingsPresent. Native file tools read, list, search, write and edit files under granted named roots with a captured primary; version-control and agent metadata folders stay read-only under Default. Explicit Full access adds the filesystem root and permits metadata writes while retaining bounded file-format and symlink checks. Desktop's Default preset allows writes in the project. Folder references grant no access: protected roots, containing ancestors and aliases are refused, outside paths retain their existing permissions, and Plan cannot inspect them with commands. Selected folders, filesystem servicesResolve named roots and scratch scopes with explicit grants. Concrete file and document tools own their content and edit contracts. Workspaces

Document ownership​

DocumentOwns
This catalogFeature identity, current facts, intended promise, consequential caveats, and supported scope.
ExperienceDetailed behavior, edge cases, and acceptance scenarios for those promises.
MasterProduct strategy and composition that fulfill the promises.
ComponentsReusable execution and durability contracts.

Keeping this compact and current​

  • Update a feature's row in the same change that introduces, changes, or retires its user-visible behavior. Elaborate the behavior in the linked document.
  • Describe the situation and observable result in product terms. Preserve the reason for a surprising rule and the limits that affect use. Keep internal functions, storage details, and test procedures in their engineering owners.
  • Update current coverage from the real entrypoint and its evidence. Mark a feature Available in release documentation only with a named host and release plus acceptance evidence. Keep target contracts separate from that claim.
  • Give each user-visible capability one row. Implementation choices and internal components belong in the engineering documents.
  • Keep FAQ answers and guides derived from this catalog and their linked behavior contracts. Resolve disagreements in those owners first.
  • Automating or retiring a manual test changes its verification method. The feature's promise remains in its canonical description until the product behavior itself changes.