nsaspy/quasar-ui

Fork 0

Bootstrap the browser coding-agent interface #110

Open

lost-rob0t wants to merge 80 commits from agent-bootstraping into main

lost-rob0t commented

2026-07-27 07:14:37 +00:00

(Migrated from github.com)

Goal

Turn Quasar’s floating agent entrypoint into a persistent browser coding-agent interface without replacing the existing supervisor, providers, actor runtime, graph operations, MCP support, budgets, recovery, or standalone management console.

Implemented

Full conversation timeline inside the floating modal; opening the bubble no longer redirects to settings or requires the standalone Agents page.
Conversations, drafts, active selection, and partial provider streams share one versioned PouchDB state record; legacy local browser state migrates once and is removed after a confirmed write.
Conversation creation, searchable recency-sorted switching, deletion, active-conversation restoration, per-conversation draft persistence, edit/resubmit, retry, copy controls, run controls, contextual route/dataset/graph state, draggable desktop position, persisted mouse/touch/keyboard resizing, and full-screen/mobile layouts.
Inline assistant, user, system, permission, error, streaming, and tool execution rendering.
Supervisor phases map consistently to idle, thinking, running-tool, waiting-for-permission, paused, failed, cancelled, completed, and budget-exhausted UI states.
Tool inputs, outputs, failures, affected records, and run-history entries remain inspectable.

Unified prompts and commands

Normal prompts and slash commands use the same composer and agent execution entrypoint.
Registry-derived command palette with aliases, fuzzy search, persisted recent-command ranking, keyboard/touch selection, generated signatures, schema-aware argument hints, missing-argument diagnostics, generated help, quoted/escaped values, named and positional arguments, arrays, boolean flags, and trailing natural-language instructions.
Dynamic actor and configured MCP commands are discovered without hard-coding them into the chat component.

Runtime permissions

Adds explicit capability permissions for web search, URL fetch, JavaScript, database, graph, documents, actors, filesystem, shell, and MCP.
Splits legacy sources.external so it authorizes web search only and never implicitly authorizes URL fetching.
Supports action, chat, session, permanent allow, action deny, and permanent deny scopes.
Permission checks run immediately before model-selected and command-selected tool execution.
Permission requests open the active chat, render inline, redact secret-shaped arguments, and resume or reject the exact blocked in-memory tool call.
Stop and pause cancel pending permission requests.

Explicit unavailable capabilities

Adds typed /file-read, /file-write, and /shell commands to capability discovery.
Clearly marks filesystem and shell commands unavailable when no trusted adapter is configured.
Blocks unavailable commands before permission prompts or agent execution and explains the missing adapter inline.

Real provider streaming

Replaces the adapter streaming stub with real SSE parsing for OpenAI-compatible and Anthropic providers.
Streams text and tool-call arguments incrementally, records finish state and usage, supports AbortSignal, normalizes errors, and falls back to JSON for compatible endpoints that do not stream.
Partial responses are shown in the modal and restored after refresh.

Browser-local attachments

Ingests supported text, source, JSON, JSONL, CSV, YAML, and markup files directly in the modal.
Enforces 8-file, 1 MB per-file, and 4 MB total limits before persistence.
Persists pending attachments with the active conversation and restores them after reload.
Renders removable attachment chips and records consumed files on the submitted user message.
Appends file contents to provider prompts behind an explicit data-not-instructions boundary.
Rejects unsupported binary files with a visible error instead of silently dropping content.

Document capabilities

Registers model-selectable document_read, document_create, document_patch, and document_delete tools.
Document writes run through StarIntel validation, Quasar operation history, persistence confirmation, affected-ID reporting, and undo.
Browser tests cover create, read, patch, delete, and undo restoration.
Dedicated /doc-read, /doc-create, /doc-patch, and /doc-delete composer commands expose typed schemas, aliases, permission checks, and destructive-risk labeling.

Sandboxed JavaScript

Registers javascript_execute for natural-language agent selection and /js command use.
Runs code in a disposable Worker, never in the page, service worker, console, extension context, or unrestricted Node runtime.
Blocks direct DOM, storage, network, cache, service-worker, and Quasar-global access.
Exposes only bounded console, input, result, and typed tools.call primitives.
Nested calls return through the normal registry and permission runtime, persist as parent-linked child tool-call records, and render as independently inspectable chat cards.
Enforces wall-clock, input, output, console, message, nested-call, and depth limits; stop/pause terminates active Workers.
Results include status, returned value, console output, duration, usage, nested calls, sanitized errors, and termination reason.

Validation

Latest CI run: 30391940864 — green.

Passed:

npm run format:check
npm run lint
npm run typecheck
architecture boundary checks
static service-worker checks
unit and integration tests
production build
application artifact validation
Playwright browser tests

Browser coverage now includes durable text attachment ingestion, reload restoration, removal, unsupported binary rejection, durable dependency-aware task orchestration, task lifecycle persistence, header control pointer isolation, modal restoration, acknowledged PouchDB draft and partial-stream persistence, legacy-state migration, persisted desktop resizing, searchable conversation switching, recent-command ranking, complete execution-state mapping, unavailable filesystem/shell gating, command discovery, typed document commands, argument validation, draft restoration, partial-stream restoration, mobile viewport fit, document mutations plus undo, JavaScript DOM/storage/network isolation, nested capability bridging plus running/completed child-call lifecycle events, timeout termination, and capability-bridge startup readiness.

Acceptance audit: still incomplete

This remains a draft because the following parts of the full specification are not honestly complete yet:

Pending permission cards survive reload, but an in-memory suspended tool promise cannot resume across a full browser restart; the restored run pauses instead.
Markdown/code rendering is intentionally lightweight and does not yet provide a full parser and syntax-highlighting pipeline.
The Worker sandbox enforces wall-clock and serialization limits but cannot honestly guarantee hard memory or instruction limits like a QuickJS/WASM isolate.

Branch state

Head: agent-bootstraping
Base: main
GitHub reports the PR mergeable.
GitHub reports the branch mergeable; CI run 30391940864 passed on the current PR merge candidate.

## Goal Turn Quasar’s floating agent entrypoint into a persistent browser coding-agent interface without replacing the existing supervisor, providers, actor runtime, graph operations, MCP support, budgets, recovery, or standalone management console. ## Implemented ### Persistent modal chat - Full conversation timeline inside the floating modal; opening the bubble no longer redirects to settings or requires the standalone Agents page. - Conversations, drafts, active selection, and partial provider streams share one versioned PouchDB state record; legacy local browser state migrates once and is removed after a confirmed write. - Conversation creation, searchable recency-sorted switching, deletion, active-conversation restoration, per-conversation draft persistence, edit/resubmit, retry, copy controls, run controls, contextual route/dataset/graph state, draggable desktop position, persisted mouse/touch/keyboard resizing, and full-screen/mobile layouts. - Inline assistant, user, system, permission, error, streaming, and tool execution rendering. - Supervisor phases map consistently to idle, thinking, running-tool, waiting-for-permission, paused, failed, cancelled, completed, and budget-exhausted UI states. - Tool inputs, outputs, failures, affected records, and run-history entries remain inspectable. ### Unified prompts and commands - Normal prompts and slash commands use the same composer and agent execution entrypoint. - Registry-derived command palette with aliases, fuzzy search, persisted recent-command ranking, keyboard/touch selection, generated signatures, schema-aware argument hints, missing-argument diagnostics, generated help, quoted/escaped values, named and positional arguments, arrays, boolean flags, and trailing natural-language instructions. - Dynamic actor and configured MCP commands are discovered without hard-coding them into the chat component. ### Runtime permissions - Adds explicit capability permissions for web search, URL fetch, JavaScript, database, graph, documents, actors, filesystem, shell, and MCP. - Splits legacy `sources.external` so it authorizes web search only and never implicitly authorizes URL fetching. - Supports action, chat, session, permanent allow, action deny, and permanent deny scopes. - Permission checks run immediately before model-selected and command-selected tool execution. - Permission requests open the active chat, render inline, redact secret-shaped arguments, and resume or reject the exact blocked in-memory tool call. - Stop and pause cancel pending permission requests. ### Explicit unavailable capabilities - Adds typed `/file-read`, `/file-write`, and `/shell` commands to capability discovery. - Clearly marks filesystem and shell commands unavailable when no trusted adapter is configured. - Blocks unavailable commands before permission prompts or agent execution and explains the missing adapter inline. ### Real provider streaming - Replaces the adapter streaming stub with real SSE parsing for OpenAI-compatible and Anthropic providers. - Streams text and tool-call arguments incrementally, records finish state and usage, supports `AbortSignal`, normalizes errors, and falls back to JSON for compatible endpoints that do not stream. - Partial responses are shown in the modal and restored after refresh. ### Browser-local attachments - Ingests supported text, source, JSON, JSONL, CSV, YAML, and markup files directly in the modal. - Enforces 8-file, 1 MB per-file, and 4 MB total limits before persistence. - Persists pending attachments with the active conversation and restores them after reload. - Renders removable attachment chips and records consumed files on the submitted user message. - Appends file contents to provider prompts behind an explicit data-not-instructions boundary. - Rejects unsupported binary files with a visible error instead of silently dropping content. ### Document capabilities - Registers model-selectable `document_read`, `document_create`, `document_patch`, and `document_delete` tools. - Document writes run through StarIntel validation, Quasar operation history, persistence confirmation, affected-ID reporting, and undo. - Browser tests cover create, read, patch, delete, and undo restoration. - Dedicated `/doc-read`, `/doc-create`, `/doc-patch`, and `/doc-delete` composer commands expose typed schemas, aliases, permission checks, and destructive-risk labeling. ### Sandboxed JavaScript - Registers `javascript_execute` for natural-language agent selection and `/js` command use. - Runs code in a disposable Worker, never in the page, service worker, console, extension context, or unrestricted Node runtime. - Blocks direct DOM, storage, network, cache, service-worker, and Quasar-global access. - Exposes only bounded console, input, result, and typed `tools.call` primitives. - Nested calls return through the normal registry and permission runtime, persist as parent-linked child tool-call records, and render as independently inspectable chat cards. - Enforces wall-clock, input, output, console, message, nested-call, and depth limits; stop/pause terminates active Workers. - Results include status, returned value, console output, duration, usage, nested calls, sanitized errors, and termination reason. ## Validation Latest CI run: `30391940864` — **green**. Passed: - `npm run format:check` - `npm run lint` - `npm run typecheck` - architecture boundary checks - static service-worker checks - unit and integration tests - production build - application artifact validation - Playwright browser tests Browser coverage now includes durable text attachment ingestion, reload restoration, removal, unsupported binary rejection, durable dependency-aware task orchestration, task lifecycle persistence, header control pointer isolation, modal restoration, acknowledged PouchDB draft and partial-stream persistence, legacy-state migration, persisted desktop resizing, searchable conversation switching, recent-command ranking, complete execution-state mapping, unavailable filesystem/shell gating, command discovery, typed document commands, argument validation, draft restoration, partial-stream restoration, mobile viewport fit, document mutations plus undo, JavaScript DOM/storage/network isolation, nested capability bridging plus running/completed child-call lifecycle events, timeout termination, and capability-bridge startup readiness. ## Acceptance audit: still incomplete This remains a draft because the following parts of the full specification are not honestly complete yet: - Pending permission cards survive reload, but an in-memory suspended tool promise cannot resume across a full browser restart; the restored run pauses instead. - Markdown/code rendering is intentionally lightweight and does not yet provide a full parser and syntax-highlighting pipeline. - The Worker sandbox enforces wall-clock and serialization limits but cannot honestly guarantee hard memory or instruction limits like a QuickJS/WASM isolate. ## Branch state - Head: `agent-bootstraping` - Base: `main` - GitHub reports the PR mergeable. - GitHub reports the branch mergeable; CI run `30391940864` passed on the current PR merge candidate.

nsaspy changed title from ~~WIP: Bootstrap the browser coding-agent interface~~ to Bootstrap the browser coding-agent interface

2026-08-27 19:57:32 +00:00

This pull request has changes conflicting with the target branch.

src/App.jsx
src/app/main.tsx
src/lib/agent-supervisor.js
src/lib/provider-adapters.js
src/lib/provider-adapters.test.js

View command line instructions

Manual merge helper

Use this merge commit message when completing the merge manually.

Merge commit title

Merge pull request 'Bootstrap the browser coding-agent interface' (#110) from agent-bootstraping into main

Merge commit body

Reviewed-on: https://git.starintel.actor/nsaspy/quasar-ui/pulls/110

Merge pull request 'Bootstrap the browser coding-agent interface' (#110) from agent-bootstraping into main

Reviewed-on: https://git.starintel.actor/nsaspy/quasar-ui/pulls/110

Checkout

From your project repository, check out a new branch and test the changes.

git fetch -u origin agent-bootstraping:agent-bootstraping

git switch agent-bootstraping

Merge

Merge the changes and update on Forgejo.

Warning: The "Autodetect manual merge" setting is not enabled for this repository, you will have to mark this pull request as manually merged afterwards.

git switch main

git merge --no-ff agent-bootstraping

git switch agent-bootstraping

git rebase main

git switch main

git merge --ff-only agent-bootstraping

git switch agent-bootstraping

git rebase main

git switch main

git merge --no-ff agent-bootstraping

git switch main

git merge --squash agent-bootstraping