## Summary - Add a generic ACP connector with browser pairing, owner-only execution authorization, exact-commit worktrees, and a run-scoped MCP bridge. - Persist fenced investigation runs and bounded structured results in PostgreSQL; render native tables and charts with collaborative selections. - Record immutable evidence-backed decisions and embed live or saved evidence in documents. - Add Planner proposal/read tools, setup documentation, and an investigation authoring skill.

## Scope Universe scope is data-driven investigations rendered by Chopin-owned components. Local agents retain their normal tools and skills. Visual-design specimens, executable browser bundles, and live application previews are deferred.

The nine implementation increments remain separate commits, followed by the Docker packaging fix. Rebased onto `025d98e3`, preserving the upstream diagram support, sidebar decision counts, and both migration histories. The connector uses the MCP SDK version already pinned by main.

## Verification After the latest rebase: - `bun install --frozen-lockfile` - `bun run types` - `bun run ci` - `bun run build` and production preview-exclusion check - Full unit suite: **4,436 passed, 4 skipped** (`bun test --timeout 30000`) - PostgreSQL suite: **77 passed** (30-second test timeout) - Chromium integration: **20 passed** across connector, investigation evidence, saved diagrams, Document activity, and sidebar decision counts - Linux/amd64 Docker image build and production-image investigation import smoke test

Previously verified during implementation: - Real Copilot CLI ACP smoke test

## References - [Design RFC](https://chopin.githubnext.com/documents/githubnext/chopin/rfc-local-agent-experiments-and-interactive-resources) - Setup and usage: `docs/local-experiments.md`

The preview deployment for **chopin** is ready. 🟢

[Open app](http://419-chopin.githubnext.com) | [Open Build Logs](https://next-appserver.tail49836a.ts.net/project/l9buznc9ua1if0vflxrjwajs/environment/lxqsvczv19eopxk29vfb46mj/application/hb85gdd9mwqqlfwcteflvpbj/deployment/ya7iotah5rwsogmnkxlqfses) | [Open Application Logs](https://next-appserver.tail49836a.ts.net/project/l9buznc9ua1if0vflxrjwajs/environment/lxqsvczv19eopxk29vfb46mj/application/hb85gdd9mwqqlfwcteflvpbj/logs)

Last updated at: 2026-10-09 14:05:47 CET

> **Historical pre-fix report.** The fix is now deployed and the fresh full Chat-driven workflow is documented in [the new report](https://github.com/githubnext/chopin/pull/419#issuecomment-6082743617).

## Live preview E2E — corrected: Investigations dialog is invisible

**The earlier blanket visual “passed” claim is withdrawn.** Several screenshots—including those captioned as authorization, running, results, and shared selection—showed the underlying document instead. They did not support their captions. I should have reviewed the rendered images before publishing them.

### Actual UI defect

At tested head `cdd50149bb599c718372dc7ba6d9d7f11ce4ddf6`, `ExperimentsPanel` passes `motion={{ phase: "open", className: "" }}` to `NavigationDialog`. The shared `.motion-modal` style has **opacity 0** until the `is-open` class is present.

The dialog therefore exists in the accessibility tree and automation can operate its controls, while a person cannot see it. Ordinary Playwright visibility checks do not check opacity. This explains the misleading screenshots and invalidates visual acceptance of that workflow.

**Local fix prepared:** supply `is-open` and add an E2E assertion using `element.checkVisibility({ checkOpacity: true })`. The regression failed before the fix and the complete targeted browser test passed afterward; `bun run types` and `bun run ci` also passed. **This fix has not been pushed/deployed to the preview yet.**

### Evidence that remains valid

The real OpenCode ACP run published result `495b71ff-39b2-410e-b573-0e5d649caf74` from sandbox commit `f7aab2fc09f592e231658f91ae4b1c88128563ce`. The persisted document visibly renders a fixed decision card and a live evidence card, shown below. Shared-state, record, reload, and disconnect checks were programmatic functional observations—not proof that the invisible modal was usable.

The 20-image gallery has been replaced by **three distinct captures whose saved pixels were reviewed**. These show current persisted state and the current UI defect, not reconstructed historical states.

- [Corrected full report and three reviewed images](https://github.com/githubnext/chopin/tree/b9d30726b88e813e05dbfa3e4cc6568ee10c7776) - [Live test document](https://419-chopin.githubnext.com/documents/KC-Test-OSS/TestRepo/pr-419-readme-compression-evidence)

### Saved decision: fixed single-workload values 5 / 25 / 13 B


### Live card: repeated-workload values 500 / 30 / 18 B


<details> <summary>UI defect: Investigations was opened, but no dialog or backdrop is painted</summary>


</details>

The separate [Copilot stdio-injection limitation](https://github.com/github/copilot-cli/issues/3889) and connector SIGINT exit-code issue remain documented in the full report.

## Fresh deployed E2E — Chat → investigation → evidence: passed

The visibility fix is deployed and this is a **new end-to-end run** on `9785424c6ad2beb4b7ffce05e1002facd84c3c08`. The trusted preview-ready timestamp advanced to **2026-10-09 14:05:48 UTC**. [All CI jobs passed](https://github.com/githubnext/chopin/actions/runs/37940125533).

- Preview: https://419-chopin.githubnext.com - [Fresh test document](https://419-chopin.githubnext.com/documents/KC-Test-OSS/TestRepo/pr-419-chat-driven-investigation) - Investigation: `2723d2d5-6717-4e54-a877-55a7a720bd90` - Sandbox source: `KC-Test-OSS/TestRepo` at `f7aab2fc09f592e231658f91ae4b1c88128563ce` - Local agent: **OpenCode 2.0.24**, via the existing generic `opencode acp` connector path

### Workflow verified

| Step | Result | Reviewed screenshot | | --- | --- | --- | | Pair the real connector with the fresh document | Passed | [Connection](https://raw.githubusercontent.com/githubnext/chopin/be5acdfe7c2d72d7b2e17e52856bf521d2c2e88b/screenshots/01-connector-paired.png) | | Send `@chopin` request through Chat | Planner called `propose_investigation`; no manual proposal-form substitute | [Chat request and tool action](https://raw.githubusercontent.com/githubnext/chopin/be5acdfe7c2d72d7b2e17e52856bf521d2c2e88b/screenshots/02-chat-request.png) | | Authorize execution in Investigations | Request stayed idle until the owner's Run click | [Visible authorization dialog](https://raw.githubusercontent.com/githubnext/chopin/be5acdfe7c2d72d7b2e17e52856bf521d2c2e88b/screenshots/03-owner-authorization.png) | | Run locally and receive results | Exact-commit worktree; real ACP result published; chart/table visible | [Results inside Investigations](https://raw.githubusercontent.com/githubnext/chopin/be5acdfe7c2d72d7b2e17e52856bf521d2c2e88b/screenshots/04-investigation-results.png) | | Share selections and embed the live result | Filters/Raw selection synchronized across two tabs; Insert view created the live document card | [Live embed](https://raw.githubusercontent.com/githubnext/chopin/be5acdfe7c2d72d7b2e17e52856bf521d2c2e88b/screenshots/05-live-document-embed.png) | | Ask Chat to interpret the evidence | Planner used `list_investigations` and `read_investigation`, reporting the captured numbers correctly | [Evidence readback](https://raw.githubusercontent.com/githubnext/chopin/be5acdfe7c2d72d7b2e17e52856bf521d2c2e88b/screenshots/06-chat-evidence-readback.png) | | Save a decision and insert fixed evidence | Recorded conclusion/rationale and the acknowledged single/Raw snapshot; record appeared in Decisions | [Saved decision embed](https://raw.githubusercontent.com/githubnext/chopin/be5acdfe7c2d72d7b2e17e52856bf521d2c2e88b/screenshots/07-fixed-decision.png) | | Change live selection, reload, and disconnect | Live view changed to repeated/Zlib; fixed snapshot stayed single/Raw. Both persisted and live filters still synchronized without the connector | [Final live evidence](https://raw.githubusercontent.com/githubnext/chopin/be5acdfe7c2d72d7b2e17e52856bf521d2c2e88b/screenshots/08-live-after-disconnect.png) |

The six values matched an independent calculation: **single Raw/Gzip/Zlib = 5/25/13 bytes; repeated = 500/30/18 bytes**. The saved conclusion is **“Keep the single README uncompressed.”** This is a tiny-file size experiment, not a general performance benchmark.

**[Full step-by-step report with eight distinct, pixel-reviewed screenshots](https://github.com/githubnext/chopin/tree/be5acdfe7c2d72d7b2e17e52856bf521d2c2e88b)**

The images were inspected from their saved PNGs before publication. The modal and backdrop are actually painted in this deployment. Later persistence captures are explicitly labeled as final state, not as earlier transient actions.

### Scope and remaining observations

- Two tabs used the same test account; cross-owner permissions were not tested with another identity. Reload and connector-disconnect persistence were tested, not a server restart. - The connector is stopped and both sandbox worktrees are clean. SIGINT still exits with code 1 despite successful disconnect/lock cleanup (`MCP error -32001: AbortError`). - The narrow dialog clips long investigation labels; its chart has a horizontal scroller. The numeric table and full-width document embeds are readable. - The previously diagnosed [Copilot ACP stdio limitation](https://github.com/github/copilot-cli/issues/3889) remains. This successful workflow used OpenCode.

<details> <summary>Key visual evidence: Chat, visible result dialog, fixed decision</summary>




</details>