fe: proposal preview, per-agent edit, diff-vs-previous, accept-to-roster #1215

New issue

Closed

opened 2026-05-15 17:03:20 +00:00 by architect · 1 comment

architect commented

2026-05-15 17:03:20 +00:00

Collaborator

Context

Part of the Agents Roster AI Assistant feature. The sidebar (#1213) ships the chat and renders raw JSON. This ticket replaces that debug view with the real structured preview, makes every field editable in place, shows a diff against the previous proposal on refinements, and wires the Accept handlers that write the chosen agents into the roster.

Scope

Per-agent card

For each ProposedAgent in the envelope, render a card with editable fields:

Name (text input)
Role / purpose (text input)
System prompt (textarea, monospace)
Model tier (select: haiku / sonnet / opus) with the model's model_tier_reason shown beneath
Skills (multi-select, options drawn from the catalog snapshot)
Tool permissions (allow/deny toggles per tool, with per-grant rationale shown inline; sensitive tools flagged visually)
MCP servers (chips; ones with setup_required: true show a "setup needed" badge linking to docs)
Rationale (read-only, one line)

Editing a card mutates a local draft copy — the original proposal stays intact in case the operator wants to refine again.

Fleet-level summary

Render ProposalEnvelope.summary above the cards as a short narrative of how the agents work together.

Selection

Checkbox per card — default checked
"Select all / none" toggle
Accept button is disabled when nothing is selected

Diff vs previous proposal

When parent_proposal_id is set, fetch the previous envelope from local session state and compute a card-by-card diff
Visual treatment:
- New agent — green border + "new" badge
- Removed agent — show as a ghost card at the bottom with strike-through + "removed" badge
- Changed agent — yellow border, with field-level highlighting on changed values
A "show changes only" filter collapses unchanged cards

Accept handler

For each selected (and edited) agent, write to the existing roster store via the same write path the manual editor uses
On success: close the assistant session, dismiss the sidebar, scroll the roster list to surface the newly created agents
On partial failure: keep successful writes, surface the failures with a retry per-agent

Catalog drift guard

Compare envelope.catalog_hash to current catalog (cheap refetch on accept)
If drifted, show a non-blocking warning: "your catalog changed since this proposal was generated — review skill/MCP picks before accepting"

Deliverables

Per-agent editable card component
Fleet summary header
Selection state + select-all/none
Diff-vs-previous renderer with new/removed/changed states + filter
Accept handler using existing roster write path; partial-failure UX
Catalog-drift warning
Component tests for: edit doesn't mutate original, diff highlights changed fields, accept writes only selected agents, partial-failure surfaces retries
E2E test of the full path: describe → propose → edit one agent → accept → see in roster

Notes for implementer

Reuse the existing manual-agent-editor field components wherever possible. Two divergent editors for the same data shape will rot fast — make the assistant preview a wrapper around the same primitives.

Depends on: #1210 (types), #1213 (sidebar host)
Depended on by: nothing — leaf on the FE side

## Context Part of the **Agents Roster AI Assistant** feature. The sidebar (#1213) ships the chat and renders raw JSON. This ticket replaces that debug view with the real structured preview, makes every field editable in place, shows a diff against the previous proposal on refinements, and wires the **Accept** handlers that write the chosen agents into the roster. ## Scope ### Per-agent card For each `ProposedAgent` in the envelope, render a card with editable fields: - **Name** (text input) - **Role / purpose** (text input) - **System prompt** (textarea, monospace) - **Model tier** (select: haiku / sonnet / opus) with the model's `model_tier_reason` shown beneath - **Skills** (multi-select, options drawn from the catalog snapshot) - **Tool permissions** (allow/deny toggles per tool, with per-grant rationale shown inline; sensitive tools flagged visually) - **MCP servers** (chips; ones with `setup_required: true` show a "setup needed" badge linking to docs) - **Rationale** (read-only, one line) Editing a card mutates a local draft copy — the original proposal stays intact in case the operator wants to refine again. ### Fleet-level summary Render `ProposalEnvelope.summary` above the cards as a short narrative of how the agents work together. ### Selection - Checkbox per card — default checked - "Select all / none" toggle - Accept button is disabled when nothing is selected ### Diff vs previous proposal - When `parent_proposal_id` is set, fetch the previous envelope from local session state and compute a card-by-card diff - Visual treatment: - **New** agent — green border + "new" badge - **Removed** agent — show as a ghost card at the bottom with strike-through + "removed" badge - **Changed** agent — yellow border, with field-level highlighting on changed values - A "show changes only" filter collapses unchanged cards ### Accept handler - For each selected (and edited) agent, write to the existing roster store via the same write path the manual editor uses - On success: close the assistant session, dismiss the sidebar, scroll the roster list to surface the newly created agents - On partial failure: keep successful writes, surface the failures with a retry per-agent ### Catalog drift guard - Compare `envelope.catalog_hash` to current catalog (cheap refetch on accept) - If drifted, show a non-blocking warning: "your catalog changed since this proposal was generated — review skill/MCP picks before accepting" ## Deliverables - [ ] Per-agent editable card component - [ ] Fleet summary header - [ ] Selection state + select-all/none - [ ] Diff-vs-previous renderer with new/removed/changed states + filter - [ ] Accept handler using existing roster write path; partial-failure UX - [ ] Catalog-drift warning - [ ] Component tests for: edit doesn't mutate original, diff highlights changed fields, accept writes only selected agents, partial-failure surfaces retries - [ ] E2E test of the full path: describe → propose → edit one agent → accept → see in roster ## Notes for implementer Reuse the existing manual-agent-editor field components wherever possible. Two divergent editors for the same data shape will rot fast — make the assistant preview a wrapper around the same primitives. --- **Depends on**: #1210 (types), #1213 (sidebar host) **Depended on by**: nothing — leaf on the FE side

architect added the

area:agents

area:dashboard

labels

2026-05-15 17:03:27 +00:00

architect referenced this issue

2026-05-15 17:03:39 +00:00

meta: Agents Roster AI Assistant — tracking issue #1216