4.3 KiB
Global Agent continuation: bounded QA recipe
This is a separate increment after the plugin platform foundation. The transport
review checkpoint is #4012 at ea619555; its isolated correction is 63f0572e0.
The runtime continuation carries the correction through an ordinary feature-branch
merge. It is not merged to main or accepted for release.
Reproduce the installable sample
From the exact runtime head reported in the handoff:
npm ci
npm run build:plugin-runtime-deps
node examples/plugins/global-agent/build.mjs
node packages/plugin-cli/dist/cli.js pack examples/plugins/global-agent/.package --out /tmp/netcatty-global-agent.ncpkg
sha256sum /tmp/netcatty-global-agent.ncpkg
The reproducible sample SHA-256 is
0c13d89cb80965db978a42f1d6e84e9cf23d7289bfa432a14f04b7f3a2651b70.
Two independent package builds matched. Build scripts and tests are excluded
from the six-file installed package.
Native fixture acceptance (parent-scheduled GUI only)
- Use an isolated synthetic profile, no real Notes, hosts, histories or keys.
Start
node scripts/plugin-global-agent-provider-fixture.mjs --port 9218. It performs no inference or outbound request and logs no bodies/headers. - Through normal host settings, configure an OpenAI-compatible Chat Completions
provider at
http://127.0.0.1:9218/v1, modelglobal-agent-fixture, with the fixture-only literalFIXTURE_ONLY_API_MARKER. Do not use real credentials, paid services, allowlist edits, debugger injection or the builtin AI sidebar. - Install the
.ncpkgin Settings → Plugins and open its contributed Global Agent Chat view. Verify light/dark/accent updates using existing theme tokens. - Choose Catty. Choose no targets, or only owned synthetic terminal fixtures,
then Open conversation. The fixture's default tool is
get_environment; it never requests a terminal command. Record Send → streamed result/usage andGET /fixture/statuscounters. This proves a deterministic pipeline, not real model inference. - Send
fixture:wait, wait for output, press Stop. Verify aborted count rises once and no late text appears. Capture screenshot/video with fixture labels. - Exercise Actions Ask using the existing in-app policy dialog: approve and deny the scoped environment tool separately. Actions Deny permits model-only chat with no tools. Agent-use Deny prevents dispatch. A policy change revokes existing handles; reselect/reopen explicitly before another Send.
- Close/reopen and disable/re-enable only this synthetic plugin. Reselect its targets and reopen its owned conversation. History should survive, while the private per-open actor ID is fresh. At five-minute expiry, the sample asks for explicit reselection and keeps the draft, without resending or renewing scope.
- Revoke/disable or lock while streaming and verify cleanup/no late events. Disconnect/re-own a selected synthetic worker during selection and a turn; stale selections must fail instead of binding a replacement connection.
Do not activate a configured real Claude backend in this fixture recipe. Claude MCP's code/handler boundary is covered with synthetic tests; authenticated service and genuine CLI acceptance remain unverified and require separately approved access. Other backends are explicitly unsupported/unavailable. No background service, blanket OS/Vault access, merge or release is implied.
Evidence and limits
- Installed Catty path uses actual
AgentRuntime, driver, main provider admission, source policy and worker dispatcher, with JSDOM/in-process plugin transport and synthetic workers. It proves selected B dispatch, stream, Stop, revocation and history after runtime restart. This is not a native BrowserWindow sandbox test. - Main/core tests cover source/session/resource substitution, scope expiry, approval waits, worker lifecycle, cancellation, late completion, stable source budgets, source-isolated history and no key serialization. Filesystem API reads cannot access the core history root. Explicitly granted terminal commands still have that terminal's native OS authority.
- No GUI ran in this executor. Native screenshots/video and independent runtime review remain parent-owned gates. The exact-head CI result must be reported separately from focused local evidence.