fix: org chart resume button + name collision

[aterrylu]

Summary

Two bug fixes for the org chart sidebar pane, plus a parallel client-side fix caught during polish.

Bug 1: Resume button on stopped agent cards

Stopped/exited agents in the org chart only had a trash (remove) button. Now the hover overlay shows a green debug-start (Resume) button beside the trash, wired to the existing resumeSession store action (isAutonomosAgent: true), which calls POST /api/sessions/:id/resume (from #110) to restore the full spawn config from persisted state.

Only shown when !isRunning — running agents still get just the kill/remove button.

Bug 2: Name collision — both agents show as "online"

Killing an agent and spawning another with the same name made both appear running in the org chart. Two root causes in buildOrgChart():

1. Node Map was keyed by lowercased name — the second session silently overwrote the first.
2. The wire-up loop looked up the parent by name and pushed the same node object into parent.children once for each duplicate, so a child could appear multiple times under one parent.

Fix: Nodes are now keyed by claudeSessionId (unique). Sessions are first grouped by name, then chooseCanonical() picks one canonical session per name:

• Prefer running sessions over exited ones
• Break ties by newest persistedAt (most recently intended instance)
• Falls back to newest exited if all duplicates are exited
• Backward-compat: treats missing status field as "running" (pre-feat: persist exited sessions for resume #110 sessions)

OrgNode interface gains status + claudeSessionId so the frontend can distinguish live/exited directly from the chart data, no name lookup needed.

Bug 2.5 (caught by polish review): Client-side collision hazard

useAgentStatusByName in HierarchyPanel.tsx was keyed by lowercased name too — reintroducing the exact collision class the server fix eliminated, for activity state (currentTool, working/needs_input, etc.). Renamed to useAgentStatusById and keyed by claudeSessionId. AgentCard now looks up activity state by node.claudeSessionId.

Approach — canonical selection (design call)

For duplicate names, we pick one canonical session and hide the duplicates from the org chart entirely (Option A from the design discussion with @aterrylu):

• Running wins over exited
• Newest persistedAt wins among the preferred pool
• Dropped duplicates are reachable via the sidebar "exited agents" toggle

A follow-up issue tracks adding a matching hide-exited toggle to the org chart itself — it has gnarly orphaned-children cases when an exited parent has running descendants, and was deemed out of scope for this PR.

Diagram

flowchart TB
    subgraph Before["Before (buggy)"]
        B1[sessions.json] --> B2["Map&lt;name, Node&gt;<br/>collision loses data"]
        B2 --> B3["parent.children.push(node)<br/>duplicates under parent"]
    end
    subgraph After["After"]
        A1[sessions.json] --> A2["Group by name<br/>Map&lt;name, Session[]&gt;"]
        A2 --> A3["chooseCanonical()<br/>running wins, newest wins"]
        A3 --> A4["Map&lt;claudeSessionId, Node&gt;<br/>unique identity"]
        A4 --> A5["Tree wired by<br/>canonical sessions only"]
    end

Loading

Test plan

• 5 new regression tests in orgChart.test.ts covering: running-wins, newest-wins tiebreak, all-exited fallback, no-duplicate-under-parent, schema shape. Includes a startup sweep that removes stale orgchart-test-* entries to survive crashed test runs.
• make check — 68/68 tests pass, Biome + TSC clean
• Visual QA deferred to reviewer — dev server on this branch would share ~/.autonomos/sessions.json with prod (tracked as a separate follow-up), making local make dev QA unsafe while prod is up. Please eyeball on the live dashboard after pulling the branch:
  • Resume button appears on stopped agent cards in the org chart, hover works, click resumes the agent, card updates to running
  • Kill an agent, spawn a new one with the same name — only the running one shows as online in the org chart
  • Existing org chart features (drag to open pane, set_manager, etc.) still work

Follow-ups (separate PRs, not this one)

1. Config dir isolation — ~/.autonomos/ is shared between prod :3100 and dev :3101; dev's resume sweep nearly cloned the running fleet during my QA attempt. Will ship as a stacked PR immediately after this one.
2. Hide-exited toggle for org chart — needs careful handling of exited parents with running children. Recommend spawning a worker.
3. handleRemove silent failure — confirm dialog stays open with no visible feedback if the backend rejects the remove. Pre-existing in the sidebar path too, out of scope here.
4. Resume button loading feedback — currently just an opacity toggle. Swap for a spinner/pulse.
5. isProduction detection via existsSync(dashboardDist) — breaks when tsc --build writes to dist/ without index.html. Needs an explicit flag or NODE_ENV check.
6. Makefile tsx --env-file=X watch src/index.ts argument ordering — tsx parses watch as the script name. Works in main by accident. Reorder to tsx watch --env-file=X src/index.ts.

Files changed

• packages/server/src/orgChart.ts — buildOrgChart() rewrite, chooseCanonical(), OrgNode schema
• packages/server/src/__tests__/orgChart.test.ts — new, 5 regression tests
• packages/dashboard/src/components/HierarchyPanel.tsx — Resume button, OrgNode type, useAgentStatusById rename + ID key, targetSession lookup
• packages/dashboard/src/components/Codicon.tsx — debug-start icon added

🤖 Generated with Claude Code

[nox-0x]

Solid fix. Name collision by keying nodes on claudeSessionId (unique) instead of lowercased name (collision-prone) is exactly right — the chooseCanonical() tiebreak logic (running > exited, newest persistedAt wins) is clean and well-tested. Resume button wiring is correct. Tests are thorough including the startup sweep for crashed-run cleanup. No blocking issues.