Finish incomplete imported-agent setup with Use here

[loganj]

🤖

Summary

• An older, incomplete import cannot start until the user selects Use here in the destination community. The action completes setup with the same agent key and leaves the agent stopped.
• Keep Stop available for recovery. Do not silently move an agent that is already configured in another community.

Details

• The local signing service signs the agent public key and destination community. The desktop app checks that signature against the source owner recorded with the import. This is not administrator approval or channel membership.
• The Use here button, signed request, desktop command, and setup checks arrive together in this draft. Setup does not reread the old installation or acquire a new key.
• Start, automatic restore, credential access, and mention selection all exclude incomplete imports. A separate Start is needed after setup.
• The newer Start with buzz-app preference cannot bypass this. Restore skips an incomplete import even when that preference is on.
• The desktop app permission for the Use here command is granted to the main window only. The production permission test covers it, including rejection for guest and remote pages.
• Setup turns off any start-on-app-launch setting left over from the incomplete import. The agent starts only after an explicit Start.
• Use here in a new community reuses the stored key of the source setup. Delete keeps that key while another setup of the same identity still uses it.
• Based on Show separate identity cards and prevent duplicate imports #225. Native metadata discovery and the combined inventory page are no longer part of this draft; they arrive together later in the stack.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Changes requested: two blockers in the setup delivery contract, detailed inline. Merge criteria: authorize the new IPC command through the production capability and preserve stopped/manual-start state when completing an incomplete import, with regression coverage for both.

Reviewed head c89f21c18b963b764027f43d58046b913c5f08c6 against base 0d2b9223c128880cb8aedd8c7cecf6e41578513d. Existing CI is green. Validation: complete diff and cross-layer source review; isolated controller reproduction of all three retained-startup-intent cases. No broad CI rerun or live desktop/Keychain/ACP acceptance.

Non-blocking UX follow-up: cross-community setup leaves the source card’s Use here prompt unchanged while adding a configured sibling. Directing focus/status to the completed destination would make success clearer.

[wesbillman]

[P1] Grant the Use here command through the production app ACL

Registering this handler does not make it callable. agent_control_use_here is absent from src-tauri/build.rs’s AppManifest::commands and src-tauri/capabilities/default.json. With this app’s ACL manifest, pinned Tauri 2.11.5 rejects a command without a resolved grant before dispatching the handler (webview/mod.rs:1794–1852). Consequently Use here obtains the signature, then fails at native invocation instead of completing setup.

Add the command manifest entry and main-webview permission, and include it in browser_permissions_tests.rs’s production-context ACL test (including guest/remote rejection). The new test mocking invoke cannot catch this failure.

[loganj]

Larry (agent), replying via Logan's account.

Fixed in 76d00bc. agent_control_use_here is now in src-tauri/build.rs AppManifest::commands, and allow-agent-control-use-here is in capabilities/default.json. browser_permissions_tests.rs checks it in the production ACL test, including guest and remote rejection. I also found the same gap for the Clone commands later in the stack (agent_control_clone_settings, agent_control_local_clone_settings) and fixed it there too.

[wesbillman]

[P2] Clear inherited automatic-start intent when completing setup

For a retained incomplete import with startOnAppLaunch: true, this transition changes configured but preserves the startup preference. The new-target branch also clones that preference, despite clearing enabled. Controller::launch_ids() therefore admits the completed target on the next app launch, and native restore starts it without the separate Start promised by this PR. In-place setup additionally preserves legacy enabled: true, which becomes the startup preference when the explicit field is absent.

An isolated run against this head reproduced all three cases: same destination with explicit startup true, different destination with explicit startup true, and same destination with legacy enabled true. Each had zero launch candidates before Use here and one afterward. No actual credentials/processes were used.

When transitioning an incomplete target to configured (or creating its destination record), save enabled = false and start_on_app_launch = Some(false) and test post-setup/reopen restore eligibility. Keep already-configured idempotent retries from resetting a legitimately running agent.

[loganj]

Larry (agent), replying via Logan's account.

Fixed in 76d00bc. When Use here moves a target from incomplete to configured, it now saves enabled = false and start_on_app_launch = Some(false). This covers the in-place record and the new destination record. A retry on an already-configured target changes nothing. The regression test use_here_clears_retained_startup_intent_until_an_explicit_start covers all three cases you reproduced: same destination with startup true, a new destination with startup true, and same destination with legacy enabled: true. It checks that restore has no launch candidates after setup.

Related fix: main added Delete (#241). A new-destination Use here setup shares the source credential reference, so deleting either card removed the shared Keychain entry. Delete now keeps the credential while another setup of the same identity still uses it. Test: delete_keeps_a_key_shared_by_another_setup_of_the_same_identity.

[baxen]

LGTM, checked the rust side in more detail than the TS

[wesbillman]

Star Lord’s automated source review (via Wes’s account)

No actionable findings in this follow-up. The two prior findings are addressed in the reviewed source:

• Use here IPC permission: agent_control_use_here is now included in the generated application-command manifest and the capability restricted to the main webview. The existing permission regression also includes this command for local-main admission and guest/remote rejection (src-tauri/build.rs, src-tauri/capabilities/default.json, src-tauri/src/browser_permissions_tests.rs).
• Stopped setup / startup intent: both incomplete-target paths clear enabled and start_on_app_launch. The added regression source covers same- and other-community setup, legacy enabled fallback, persistence across reopen, and preserving an already-configured setup’s deliberate startup preference on retry (crates/agent-controller/src/store.rs:186-210, crates/agent-controller/src/runtime/tests.rs:1504-1565).

I also inspected the shared-key deletion change against the serialized native caller: deleting one setup retains custody while another saved setup uses the same credential reference and public key; final deletion still removes custody. The frontend/native caller chain, destination-scoped unmount guard, incomplete-import launch/selection exclusions, and Stop recovery were inspected for integration regressions. The earlier optional cross-community card/focus UX observation is not a new blocker.

Pinned scope: head 76d00bcc0a2fef7c815d636f91e02258e1cc02d9; base/merge-base 657ea06871cf38c6fbdf393085cb541fec413459 (stacked on #225). This is a bounded follow-up to the feedback on c89f21c18b963b764027f43d58046b913c5f08c6; actual branch changes were distinguished from incoming base changes. It does not clear findings on ancestor PRs.

Validation limits: source-only review of immutable Git objects, with extracted source bytes/blob IDs/SHA256 verified and no dirty source inputs. Regression tests were read, not executed. No builds, app launches, credential operations, or live setup/delete workflows were run; CI was not assessed. Native/runtime behavior and human acceptance remain unverified. This COMMENT is not an approval or merge authorization; the earlier submitted review was not dismissed.

[wesbillman]

Star Lord’s automated source review (via Wes’s account)

No new actionable findings in this follow-up. Direct tree comparison with the last reviewed revision, 76d00bcc0a2fef7c815d636f91e02258e1cc02d9, shows changes in only four files; the previously reviewed Use here implementation and its two blocker fixes remain unchanged.

• The import preview now excludes a public key already managed in any community. The caller supplies the complete native agent snapshot, and native import rejects duplicate keys both during preparation and commit (src/bundled/agents/AgentImport.tsx:70–78, AgentControlPanel.tsx:181–190, crates/agent-controller/src/import.rs:153–165,252–259). The added component regression covers a different-community record, case normalization, and repeated loading.
• Inventory definition aliases now require an exact local definition ID. A reference to an omitted, slug-colliding definition remains unmatched instead of being grouped under another profile (src/features/agents/inventory.ts:14–22,43–52). I inspected the regression and the definition/avatar/grouping consumers.
• The production Use here command manifest/main-webview capability and the stopped-setup fix remain present. Both incomplete-target paths clear enabled and start_on_app_launch; already-configured retries preserve their deliberate preferences. These source files and the earlier regression coverage are byte-identical to the last reviewed revision.

Pinned scope: head 3e749fdb8a8e4ec60f9246091017f8850a427423; base cc4b36f735fcfd7ce44111a4115456d66addbfc7 (stacked on #225). This is a bounded follow-up, including the incoming-base changes; it does not clear findings on ancestor PRs or reopen the earlier optional card/focus observation.

Validation limits: source-only review of immutable GitHub blobs, verified against pinned trees; no dirty working-tree inputs. Regression tests were read, not run. No builds, app launches, credential operations, or live setup/import/delete workflows were executed. CI was not assessed; native/runtime behavior and human acceptance remain unverified. This COMMENT is not an approval or merge authorization, and no earlier review was dismissed.