feat(client): manage imported agents from Settings

[bryantderosier]

Users can import portable JSON or YAML agents from an individual file or a folder into the selected environment, then edit, enable, disable, or remove the imported copies in Settings. Duplicate imports ask which agents to replace, preserving skipped copies and existing on/off settings.

The Import menu sits beside Scoped agents. Compact rows show availability beside the name, with edit and destructive trash controls; action feedback uses toasts. The editor changes only the imported copy. Model menus show only signed-in, enabled providers with available models; hovering opens their models to the side. Model and reasoning pickers sit side by side. Reasoning stays separate and offers every reasoning option advertised by the selected model; unrelated edits preserve existing values. Assigned tasks retain their saved persona and fixed model route in both composers.

Web and desktop share the Settings UI; mobile provides native file/folder pickers, replacement toggles, and editing. The rebase preserves upstream's new composer behavior and semantic mobile theme classes, and keeps editor actions visible while fields scroll.

Validation

• 487 focused tests across 21 files passed after the rebase (including the corrected Claude runtime-prompt assertion).
• Server, web, mobile, shared client runtime, and contract typechecks passed. Changed TypeScript files passed lint with existing upstream warnings.
• Real web-client pass: YAML file and folder imports, imported-copy editing, silent off/on toggles, selective replacement, preserved disabled state, trash removal, and toast feedback. Source YAML stayed unchanged.
• Model menu update: 20 focused helper tests, web/mobile/client-runtime typechecks, and targeted lint passed. Browser verification confirmed the provider-only menu, side submenus, compact picker layout, all advertised reasoning options, and persisted primary/fallback selections. Tests cover signed-out, unknown-auth, disabled, and uninstalled providers.
• Native mobile, packaged desktop, and relay/tunnel interaction were not exercised.

UI evidence

Before: published original stack UI. After: rebased imported-agent Settings UI. Additional images show the editor and per-agent replacement choices.

Current editor: signed-in provider submenus and side-by-side model/reasoning controls. The earlier editor image above shows the previous layout.

Reasoning menu before restoring all options:

Reasoning menu after restoring all options:

Merge order

Part of GitHub stack 120. Merge from bottom to top:

#75 → #76 → #78 → #79 → #80 → #81 → #83 → #82 → #85 → #118

Stack rebuilt 2026-09-09 (stack pingdotgg#120): #84 and #86 were dropped, and fork-boundary refactors moved the persona RPCs, guards, and client wiring into J5-owned files. Review threads were preserved.

Generated with GPT-6 using the Codex desktop app.

2026-09-10 additions: a Create agent button beside Import authors a personal agent (name, stable ID, description, markdown instructions, runtime policy, primary/fallback models) stored as an imported definition, on web and mobile; and removed example or folder agents now stay listed as Removed with a Restore action (Tyler's review finding). Both ride the existing J5 RPC group with no new upstream files.

Later on 2026-09-10: the edit dialog and sheet now load and save instructions; every row gains a more menu with Duplicate as personal agent (prefilled create dialog with a fresh ID, the way to build on a bundled example) and Export YAML; and acceptedInput, inputArtifacts, and outputArtifact are optional in the definition schema so personal agents carry no synthetic artifact. One new read RPC in the J5 group serves all three.

Also on 2026-09-10: every catalog entry carries its content digest, and the assignment lock controls on web and mobile show a Changed badge when a thread's launch snapshot no longer matches the current definition (roadmap item 4).

Later on 2026-09-10 (roadmap items 1, 2, 4, 5, then 7, 8, 9): editing instructions in the app through a readAgentPersona RPC, Duplicate as personal agent and Export YAML, a Changed drift badge on locked threads, and optional handoff-artifact fields for personal agents. The final pass adds a per-agent usage line (tasks, runs, completed/failed, mean duration, provider-reported tokens, last launch, resolved routes) aggregated on request from the existing thread, run, and provider-turn projections; an origin badge plus a Library sources section that lists, adds, and removes folders in agent-personas.json with read-only git nudges (uncommitted changes, remote ahead) and open-in-editor on web; and Copy to environment in the row menu. Three more methods on the J5 RPC group, no new upstream touches. Verified with the server J5 suite (88 tests), client-runtime (39), web (22), and mobile (4), plus typecheck and lint on every touched package.

2026-09-11: Add folder in Library sources opens a folder picker that browses the environment's filesystem through the existing filesystem browse RPC (web dialog, mobile sheet) instead of a typed path field; the resolved absolute path is what gets saved.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Advanced

Run ID: fbbbfbd0-c9e3-49d3-bb53-33ebeca11ba7

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[bryantderosier]

Closing: the two subagent label helpers existed only to feed the Agents panel restyle in #86, which is dropped from this stack to keep upstream rebases clean. The branch is kept as a pass-through pointing at #84's tip so #118 keeps a correct base; GitHub's stack feature refuses base changes through the API.

[tyler-barton-horizon]

Found one functional issue in the stack review:

Settings exposes the destructive remove action for every persona row, including bundled examples, and the server-side complete removal records the id in removed-source-agent-personas.json before deleting any imported override. That makes the agent disappear from future catalogs. The docs say users can restore by using Import -> Agent file with the original definition, but imports only accept .yaml/.yml while the bundled examples in this stack are shipped as JSON (apps/server/src/j5/agents/examples/*.json).

So removing a bundled agent like scout or builder is effectively irreversible from Settings unless the user manually converts/recreates YAML or edits environment state. That feels like a one-way door in a normal Settings workflow.

Suggested fixes: add an explicit Restore/Reset action for removed source/bundled agents, hide/remove-disable the trash action for bundled examples until restore exists, or ship supported YAML copies of the bundled definitions and point the docs at that supported restore path.

I checked the existing Bryant-stack comments before posting this. The earlier #75 review themes appear addressed in the combined stack, and I did not see this specific issue already raised.

[bryantderosier]

Thanks for the catch, Tyler. Agreed it was a one-way door once imports went YAML-only. Fixed in 789a499cf with an explicit Restore action:

• The catalog now lists removed source and bundled definitions as removed entries (availability unavailable/removed), so they never disappear from the library view and never appear in the @agent picker.
• A new restoreSourceAgentPersona RPC in the J5 group clears the exclusion; it requires orchestration-operate like the other mutations.
• Web and mobile Settings show a Restore button in place of the trash for those rows. Imported copies still have no restore, since their content lived only in the deleted import, and the docs now say so instead of pointing at Import.
• Library tests cover remove, list, restore, the not-removed error, and that an imported override hides the removed source; the RPC scope test covers the new method.

Zero new upstream-file lines: the RPC rides the J5 group that is already merged into WsRpcGroup.

requested a followup review after pushing a new commit.