HVE Cockpit: a host-agnostic web GUI for steering agentic coding sessions#2271
Draft
MichaelDanCurtis wants to merge 185 commits into
Draft
HVE Cockpit: a host-agnostic web GUI for steering agentic coding sessions#2271MichaelDanCurtis wants to merge 185 commits into
MichaelDanCurtis wants to merge 185 commits into
Conversation
Host-agnostic web cockpit that makes hve-core RPI agent sessions legible and interactive via an MCP bridge, without owning orchestration. Includes the design spec and a themeable Fluent mockup (Fluent / VS Code / Mica).
…ide frames; add startup error handling
- server: bind HTTP listener to 127.0.0.1 explicitly (was binding all interfaces); EADDRINUSE/ephemeral fallback preserved. - handlers: present_options now passes a finite fallback timeout to bridge.presentOptions so a stuck decision can't block the agent forever. Configurable via RPI_COCKPIT_DECISION_TIMEOUT_MS (default 1800000 ms). - package.json: add prepare/postbuild scripts and engines node>=20. - gitignore: ignore rpi-cockpit/dist build output. - tests: cover the finite-timeout fallback for present_options.
…into host surfaces Add `rpi-cockpit init [--host claude|codex|vscode|all] [--codex-global]`: - Prepend shebang to src/index.ts so tsc emits it as dist/index.js line 1. - Dispatch on process.argv[2]==='init' before starting the MCP server; parse --host/--codex-global, derive defaults from import.meta.url, run runInit(), print summary, exit 0. - src/init.ts: idempotent read-merge writers for .mcp.json (CLAUDE_PROJECT_DIR), .vscode/mcp.json (workspaceFolder + cwd), hand-rolled .codex/config.toml ([mcp_servers.rpi-cockpit] with absolute entryPath/cwd, startup_timeout_sec=20), and marker-delimited narration blocks inlined from the contract into CLAUDE.md, AGENTS.md, and .github/copilot-instructions.md (preserving existing content). - Ignore /.codex/ (machine-specific absolute path; never commit). - TDD: tests/init.test.ts covers config shapes, host filters, preservation, and idempotency.
Run `rpi-cockpit init --host all` to wire the cockpit into every host
surface and inline the narration contract:
- .mcp.json (Claude): mcpServers.rpi-cockpit, type stdio, ${CLAUDE_PROJECT_DIR}
- .vscode/mcp.json (VS Code): servers.rpi-cockpit, type stdio, ${workspaceFolder}
- CLAUDE.md / AGENTS.md / .github/copilot-instructions.md: narration block
- .gitignore: allowlist .vscode/mcp.json so the generated config is committable
- README: per-host setup, config locations, 7-tool/UI verification, Codex TOML
.codex/config.toml is machine-specific and stays git-ignored (not committed).
Remove .github/instructions/hve-core/rpi-cockpit-narration.instructions.md: its applyTo glob targeted agent/prompt definition files, so VS Code Copilot never auto-applied it during real work. Narration now lives in each host always-on instruction file (CLAUDE.md, AGENTS.md, .github/copilot-instructions.md) written by init. Repoint the five RPI .agent.md references to the canonical rpi-cockpit/agents/cockpit-instructions.md.
The old "Register with Claude Code" step told users to cp .mcp.json.example over
.mcp.json, which clobbers the correct init-generated config (type stdio +
${CLAUDE_PROJECT_DIR} + env). Remove that step (the per-host section already
covers Claude Code) and update .mcp.json.example to match the generated form.
The RPI Agent handoffs used send: true and targeted the RPI Agent itself
(/rpi continue=all, etc.). VS Code Copilot honors send: true as auto-submit
("the prompt automatically submits to start the next workflow step"), so after
every response a handoff auto-submitted back into the same agent -> infinite
self-invocation on any request. Switch the handoffs to send: false so they are
user-click buttons (their intended use). Native frontmatter issue, exposed by
running the agent in VS Code Copilot; not caused by the cockpit instrumentation.
VS Code Copilot auto-submits send: true handoffs ("the prompt automatically
submits to start the next workflow step"), which auto-advances the whole RPI
pipeline (planner -> implementor -> reviewer ...) with no pause for the user.
Flip the remaining seven hve-core agents handoffs to send: false so each step
waits for an explicit click, matching the RPI Agent fix in d046b33.
The Phase 5 cockpit narration told the agent to call present_options for the Suggested Next Work list "then act on the returned id" — which, with the finite present_options timeout falling back to the recommended option, could auto-start a new RPI cycle with no human input. Reword so Discover surfaces the options and acts only on an explicit user selection, otherwise stops and yields.
Also adds exhaustiveness arms in state.ts applyBeat/summarize (passthrough + label), required by strict tsc when extending the Beat union. Task 2 refines the applyBeat arm to set steerMenu.
…rill-in
Addresses the whole-branch review's one Major: the spec required the camera
to frame the graph bounding box on a new node set, but it was hard-pinned at
{x:40,y:40,z:1} and wide pipelines clipped at the right edge. gwFitToView
centers + scales the bbox into the live canvas (zoom clamped to [0.3,2]),
gated by a scope+node-set signature so live-run status updates keep the
user's pan/zoom while a fresh pipeline or a drill-in re-frames.
Author
|
@microsoft-github-policy-service agree |
Author
|
@jkim323 things were done, UX was made ... take a look at the workflow UX ..its pretty cool |
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #2271 +/- ##
===========================================
- Coverage 81.24% 61.84% -19.41%
===========================================
Files 127 10 -117
Lines 18829 76 -18753
Branches 12 12
===========================================
- Hits 15298 47 -15251
+ Misses 3528 26 -3502
Partials 3 3
Flags with carried forward coverage won't be shown. Click here to find out more. 🚀 New features to boost your workflow:
|
Addresses GitHub Advanced Security dependency-pinning-analyzer findings on PR microsoft#2271: replace caret ranges with the exact versions already resolved in package-lock.json (no install change). tsc + 357 tests green.
A real Claude Code slash command (.claude/commands/gallery.md) launches the 65-agent gallery producer and shows it in the preview pane; the cockpit narration contract (cockpit-instructions.md + CLAUDE.md) documents /gallery as a convention mapping to gallery_open, alongside the existing /Nav.
…low DAG The agent-gallery producer predated the flow canvas, so agent microsoft#58's tile narrated the old gh-aw screen card. Add a flow() helper and rewire microsoft#58 to narrate the orchestration pipeline (flow_open + workflow nodes + label/event edges) so its tile shows the n8n-style DAG, matching the new flow surface.
…t#60/microsoft#61/microsoft#62) gallery tiles as their own surfaces The producer predated the memory and promptlab surfaces, so these four meta-utility agents narrated context badges / a findings panel. Add memory() and promptlab() producer helpers and rewire microsoft#57 to the memory view and microsoft#60/microsoft#61/microsoft#62 to the promptlab workbench, so every meta-utility tile reflects the shipped surface.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this is
The HVE Cockpit (
rpi-cockpit/) is a host-agnostic web GUI that makes anagentic coding session legible and steerable. An agent narrates its work
over an MCP bridge; the cockpit renders one view per kind of work (an RPI build
loop, a reviewer's findings panel, a backlog kanban, a team board, a codebase
map, and more), and the user steers through a directive queue the agent drains.
The charter boundary is strict: the cockpit captures intent (steer,
decisions, interventions) and the agent performs; the cockpit never launches or
controls agents itself.
Architecture
A one-way data spine: a typed beat (zod union) -> a pure reducer ->
a pure view-model projection -> WebSocket broadcast -> an unbundled
vanilla-JS browser client. MCP tools emit beats via a bridge. No frontend
framework, no graph library. Tests are Vitest + happy-dom.
What's in it
kanban, team board, 3D codebase map, data profile, gallery, prompt workbench,
memory, and an n8n-style flow canvas for the gh-aw agentic-workflow
pipeline), plus a representation map of shared primitives (timeline, decision
flow, list, question, screen, context badges, app frame).
producer MCP server writes a state snapshot the consumer pane renders).
tscclean.Scope of this PR
rpi-cockpit/— the cockpit package (source, the browser client, tests, andthe design/plan docs under
rpi-cockpit/docs/).initstep: registering the MCP server (.mcp.json,.vscode/mcp.json), amarker-delimited narration-contract block in
CLAUDE.md/AGENTS.md/.github/copilot-instructions.md, narration pointers in the.github/agents/hve-core/*agent files,.gitignoreand a previewlaunch.json, plus the original design docs underdocs/.The root-level changes are the natural discussion point for maintainers: they
are how the cockpit wires into the repo, and are easy to pare back to a
cockpit-only contribution if preferred.
Status
Developed iteratively (brainstorm -> spec -> plan -> subagent-driven build with
per-task and whole-branch reviews), each surface live-verified in the preview
pane. Opened here as a draft for upstream feedback.