Skip to content

armgit / ODS

Public Imported from GitHub
0

ODS V3 Pre-Release: Public testing and refinement ahead of the official V3 launch. Turn your PC, Mac, or Linux box into a private AI server.

15.4 KB
Raw Blame

Portal task interface and goal execution

The empty conversation uses a silver prompt bar. Once a message exists, the normal compact composer returns. The existing theme controls still determine the panel, text and diff colors. Motion honors prefers-reduced-motion and controls remain usable by keyboard and in narrow browser windows.

Using goals

Choose Goal in Prompt commands or enter /goal followed by the desired outcome. No agent count is required. Portal saves the goal, public plan, actual responses and clarification answers. A successful partial turn can start a new, individually identified continuation in the same worker session.

The durable controller shares the existing agent-team execution lane: only one managed worker uses the shared model at a time. Closing or reloading the browser does not submit another run. Stop records the cancellation request before contacting the runtime. An interrupted controller requires an explicit stop; transport failures and uncertain effects are never automatically replayed.

Completion requires a terminal transport receipt, a nonempty delivered answer, an acceptable host verification outcome, and a valid completed public plan. The plan is reported by the model, not independent proof of arbitrary work. Existing file/publication verification and tool permission checks still apply. The controller pauses after four turns without completed-step progress or twelve total work turns. Clarifications pause for the owner, with at most four answer rounds and 4,000 retained answer characters. Completed work, preferences and the original objective are carried forward; the UI offers explicit continuation after a stopped or incomplete goal.

What the activity and response show

  • The compact expandable activity timeline replaces the boxed Live activity UI. Its last 24 real tool calls mix public updates, plan steps, searches/sources, tool operations and clarification traces in chronological order. Failed, blocked and unconfirmed operations remain visible; a recovered attempt does not label a completed run as failed. These are execution observations and deliberately public reports, never private model reasoning.
  • The optional pixel_ods_activity tool sends a bounded public progress message. The header shows that message while work runs and measured elapsed time. It never plays a canned reasoning sequence. Running logs follow new events until the owner scrolls upward; completed logs collapse and can be reopened.
  • Clicking command/file rows opens bounded details. Successful writes/edits can show syntax-highlighted excerpts, addition/deletion counts and Copy. Current edits[] and older replacement schemas are supported. A write has no known previous content, so it does not claim a created file or a full diff. Truncated excerpts do not claim complete line counts. Failed edits do not display proposed changes as successful effects.
  • Display metadata uses schema v3 with closed fields and limits: 160-character public labels, 400-character details, three safe HTTP(S) source links, eight plan steps and 1,000 characters per change side. Known credential files and common secret-bearing lines are omitted from excerpts. This is bounded display filtering, not a guarantee that arbitrary authored text is secret-free. All transport/UI readers retain compatibility with schemas v1 and v2.
  • The context ring uses the last assistant call's input/output/cache token measurements and its actual context budget. It never substitutes cumulative billing usage or an estimated zero. Without measurements, only the configured capacity is shown. Hover, keyboard focus or touch opens details.
  • Newly delivered response text fades in without a synthetic typing queue. Earlier chunks retain their DOM nodes; code and Markdown remain intact. Reduced-motion settings disable the effect. The runtime may retain text until its existing verification finishes; CSS does not bypass that.
  • Inline source badges preserve HTTP(S) links supplied in the answer. They do not invent citations, fetch link previews or claim that a link was verified.
  • Published artifact changes continue to use the existing real snapshot diffs. Activity excerpts are separate tool observations, not snapshot verification.

Component provenance

The following UImaxxing registry layouts were adapted to the existing React, Tailwind and Lucide setup; attribution and source provenance remain in the files:

Their supplied notices allow use, modification and shipping in products, including commercial products, and prohibit republication as a component library or use as machine-learning training data. No registry installer or new UI dependency was required. The reasoning/activity, streaming, plan, question and inline-citation equivalents are Portal implementations; no paid Kobra component, license token or protected source was used.

The current activity and streaming implementations adapt the visual patterns of:

They use Portal components and real runtime events instead of the reference demos' hardcoded progress or timed text. No additional UI dependency was added.

Validation and portability

Regression coverage includes durable continuation, idempotency, cancellation, question/answer recovery, malformed plans, stalled models, owner isolation, schema compatibility, actual token telemetry and preservation of complete text. Receipt availability is retried with the same read-only run ID; the model request is not resubmitted. Frontend tests cover the command menu, inline questions, history reload, accessible tooltips, citation links and diff behavior.

The controller tests ran on Windows Python and Linux containers; native ingress and installer checks ran in Linux/WSL. The interface was inspected in a compact browser viewport. A real local Qwen 3.5 4B run continued over multiple turns, completed both public plan steps and delivered a checked ten-word welcome sentence; the ring displayed 8,456 / 65,536 tokens. This is a smoke test, not a claim that every task or small model will complete correctly. No macOS machine was available for a native execution test. Product code contains no developer machine paths or dependency on this workstation's Docker overlay configuration.

A second live run calculated 17 + 25, checked the subtraction, and completed both plan steps after durable continuation, with 5,768 / 65,536 tokens measured. Earlier live runs exposed intermittent upstream transport failures. Receipt reads now have bounded retries; ambiguous work is still never resubmitted. The edge records only the exception category for diagnosis. The latest run did not reproduce the failure, which does not establish that every possible transport failure has been eliminated.

The activity smoke test exposed a preview-intent bug: an ordinary filename under portal-.../notes.txt was interpreted as a website request. The intent projection now excludes nonvisual file operands, with regression coverage for Windows and POSIX paths. Explicit website/HTML publication still requires its existing verification. The local 4B model can still select the wrong tool or misinterpret an edit; rendering accurate activity does not eliminate model errors.

The latest real-model checks displayed a completed edit with +1/-1, expanded its before/after excerpt, copied the displayed diff, then read the file and delivered its contents. A separate web search returned real source links and a final answer linking to the official React useState documentation. Source titles remove the runtime's technical evidence envelope for display only; model-facing evidence remains unchanged. The first unconstrained file test still exposed hallucinated tool names; the guard stopped that run and kept its actual failed/completed steps. These checks establish the activity integration, not universal model reliability.

Tabbed workspace and published-file review

Publication cards in the conversation now open Preview, Review or the selected changed file in a compact workspace. Review displays one verified diff with a searchable folder tree on the right. Opening its source creates a closable, deduplicated file tab. Verified Markdown is rendered as a document with a source toggle; relative file links resolve within that same publication.

The preview iframe stays mounted across tab switches and panel collapse, so reading a file does not reset an interactive preview. The workspace uses the latest verified publication of the selected project, without a version selector. Review lists every current manifest file plus explicitly deleted paths, and opens unchanged files in the verified source viewer. Small panels use a file drawer; keyboard tab navigation, focus restoration and transparent, theme-aware scrollbars cover the chat and workspace. No external editor, undo or live repository state is implied: this view reads the existing verified publication snapshots.

Focused regressions cover opening a workspace before publication, publication switches during manifest loading, failed-manifest retries, obsolete pending file requests, exact file selection and draft-only publication requests. Existing snapshot integrity, clipboard, source and conversation recovery tests are retained.

The code/file-list divider supports pointer dragging, keyboard resizing and container-size changes. Folder trees and breadcrumbs show the publication's real relative directory; paths sent to the snapshot API remain relative to that publication. Existing projects are not relabeled as Playground.

On wallpaper themes, chat publication cards use a transparent blurred surface that picks up the wallpaper color. The jump-to-latest action and response copy icon have no box. The known successful publication footer is hidden only when the same verified publication already has a card. Stored text, model history, failure messages and useful answer content are preserved. A comparison fetch failure still leaves the verified web preview accessible.

Project directories and composer controls

New projects use real Playground/<descriptive-name> directories below the configured agent workspace. The runtime reserves unused names, persists the conversation binding, and routes file operations, patch headers, exec workdirs and publication to the same directory. Explicit owner paths and legacy projects stay where they are. Shell source is never rewritten. A new project must begin with a file write before running project commands, including when a small model tries to start with an exec or patch. Native Windows and Linux tests cover name collisions, continuation, restart, unsafe paths and routing provenance.

A local Qwen 3.5 4B run created Playground/cafeteria-aurora/index.html, then published it. Its actual workspace bytes matched the HTTP preview's verified entry hash. This verifies creation and publication, not every website interaction.

The composer has a compact model selector in place of the character counter. It loads the installed catalog on first use and uses the existing model-switch API and readiness checks. The input length limit and context usage ring remain. The task list reflects validated goal or public plan events, with no simulated completion. Streaming text progressively reveals new answers even when all network deltas arrive in one batch; saved history, reduced-motion preferences, stops and errors render immediately. Display animation does not delay persistence or tool execution. These presentations adapt the task-list and streaming-response references to Portal's existing React components rather than installing a second UI stack.

Subagent workspace

Agent icons and conversation links open a closable Subagents tab in the same workspace as Preview, Review and source files. The list groups the current conversation's agents by active, completed and attention states. Selecting an agent shows its recorded conversation, live activity, questions and actual retry/stop actions; the back button returns to the list. Question drafts survive tab switches and reopening the panel. A new publication does not displace the active agent conversation or reload the existing web frame.

Workspace requests are consumed once and scoped to the conversation so closing the tab or changing chats cannot reopen a stale selection. File requests leave the agent view before loading, keeping retry controls visible if a manifest cannot be fetched. Review omits the first-publication explanatory sentence.

Workspace, Review, source files and Subagents share the Settings panel's glass surface on wallpaper themes. Inner canvases stay transparent; file drawers retain their own readable surface. Reduced transparency and browsers without backdrop blur use an opaque fallback. Published iframe content is not filtered.

Answer links and tool sources show site favicons, including inside Subagents. The browser requests icons lazily from a fixed Google favicon-cache endpoint; only the public hostname is sent, without article paths, query strings or referrer. Local addresses and authenticated URLs use a globe. Missing/offline icons also fall back without removing or changing the link. The CSP permits only the icon endpoint, not arbitrary remote images. An icon is decoration, not source verification.

Document downloads without a website

Portal can attach an existing workspace document with the specialist tool pixel_ods_workspace_artifact in an ordinary owner-interactive Portal turn. Team workers, subagents and background goal rounds cannot attach documents through this surface yet; their publication attempts are rejected before contacting the broker. The tool takes one relativePath, for example {"relativePath":"Playground/report.pdf"}. A misspelled argument such as path is rejected before publication with exact schema guidance; it is not reported as a file-permission failure. The tool does not accept content, host paths, URLs or a destination. A website and index.html are not required. Markdown, TXT, CSV, TSV, JSON, PDF, ZIP, RAR, DOCX, XLSX and PPTX are supported as opaque files, up to 4 MiB each and four publication attempts per response. Macro document extensions are not accepted.

The owner-scoped broker captures the exact file with no-link descriptor reads, rejects unsafe ownership, permissions, hard links and changed source bytes, then creates an immutable snapshot and verifies its HTTP readback. The tool's policy admission and receipt acceptance are bound to the active run and session. Stop or a superseding run prevents a late receipt from being attached. Raw MEDIA: text and model-authored links never grant download authority.

A successful receipt adds a file card to the answer without opening the website panel. Clicking Download retrieves the existing authenticated snapshot route; the browser checks the byte count and SHA-256 before saving. Validated receipts survive completed-response history and replay. A hash, network or missing-file failure leaves an explicit retry state. Files are not fetched before that click.

This verifies delivery and byte identity, not PDF layout, archive integrity, Office compatibility or factual correctness. The agent must verify those aspects separately and report any limits. Existing website publication keeps its own entry-point, manifest and preview behavior. No document is executed, unpacked or rendered by this delivery tool.