feat: add agent catalog/auth API and safer orchestrator switching

[nikhilachale]

Summary
This PR adds a daemon-backed agent catalog, exposes installed/authorized agent state to the frontend, and uses that data in project settings so users can choose worker/orchestrator agents more safely.

It also adds orchestrator replacement handling: when the saved orchestrator agent changes, AO starts the replacement first and only retires the previous orchestrator after the new one is up, so a failed replacement does not cause downtime.

A key caveat is that agent-auth/login flows can interfere with replacement startup. If switching agents triggers the agent’s own bootstrap path, the replacement may come up outside AO’s normal orchestrator initialization path and miss the AO orchestrator system prompt.

What Changed
Backend
Added agent inventory service for:

supported agents
installed agents
authorized agents
counts for each
Added optional AgentAuthChecker capability on adapters.

Added shared CLI auth probing helper for adapters with cheap local auth checks.

Added GET /api/v1/agents.

Extended registry inventory entries to carry adapter manifest metadata for user-facing labels.

Added orchestrator replacement flow in the session service:

spawn replacement first
retire previous orchestrator only after successful replacement
preserve previous orchestrator when replacement startup fails
Added backend tests for agent catalog, controller responses, session replacement behavior, and related project/service wiring.

Frontend
Regenerated API types for the new agents endpoint/DTOs.

Updated ProjectSettingsForm to:

load agent catalog from the daemon
show authorized agent options
handle installed-but-not-authorized states
surface orchestrator replacement pending state
allow retry once replacement is safe to perform
Added/updated tests for the new settings behavior.

aoagents/ReverbCode#276. extended version of this pr

closes #2310

[nikhilachale]

@neversettle17-101 @illegalcall please review this

[nikhilachale]

https://htmlpreview.github.io/?https://gist.githubusercontent.com/nikhilachale/e5e76b1a77f89a2ffdc89706311493f7/raw/agent-installation-auth-orchestrator-spawn.html

[nikhilachale]

Agent adapters gained binary-resolution behavior across Agy, Aider, Amp, Auggie, AutoHand, Claude Code, Cline, Codex, Continue, Copilot, Crush, Cursor, Devin, Droid, Goose, Grok, Kilocode, Kimi, Kiro, OpenCode, Pi, Qwen, and Vibe.
The agent service and ports were extended to support refreshed catalog status and availability checks.
HTTP agent controllers and DTOs were updated, with OpenAPI generation kept in sync.
Added refresh support for agent catalog data, exposed refresh in the API contract, disabled unavailable agents in selection flows.
@illegalcall

[illegalcall]

I found a few issues that should be fixed before merging:

1. P1: Preserve the no-resurrection marker cleanup

session_worktrees is used as the shutdown-restore marker. The intended flow is: shutdown writes the marker, next boot restores the session, then the marker is consumed/deleted.

In this PR, Kill marks a session terminated but no longer clears that stale marker. Example: AO shutdown saves proj-1, next boot restores it, the user later explicitly kills proj-1, and the next daemon boot sees the old marker and relaunches the killed session. Explicit kill intent must win.

Please restore marker cleanup on Kill and after RestoreAll consumes a marker, and keep the no-resurrection tests.

2. P2: Return arrays, not null, for the initial agent cache

On a fresh daemon before POST /agents/refresh, Installed and Authorized are nil slices. JSON encodes those as null, but the OpenAPI/generated TS contract says these fields are arrays.

Example response today can be:

{ "supported": [...], "installed": null, "authorized": null }

Clients should receive:

{ "supported": [...], "installed": [], "authorized": [] }

Please initialize/clone nil slices as empty slices so first-load clients always get arrays.

3. P2: Make OpenCode binary resolution fail when absent

/agents/refresh treats a successful ResolveBinary as installed. OpenCode currently falls back to the bare string "opencode", so a machine without OpenCode can still report it as installed. Later, spawn does the real binary check and fails.

For catalog/install detection, ResolveBinary should only succeed when a real executable path is found; otherwise return ports.ErrAgentBinaryNotFound.

4. P2: Avoid sending a real Continue prompt during auth refresh

Refreshing the catalog should be local/status-only. The Continue auth fallback runs cn -p hi, which is a prompt execution path, not an auth status check.

Example: opening project settings and refreshing agents could submit an actual model prompt, create local session/history, consume quota, or hang on provider/login state. If Continue has no safe local auth signal, return unknown rather than running a live prompt.

[nikhilachale]

-session_manager
- Kill now deletes stale session_worktrees restore markers so explicitly killed sessions do not resurrect on the next boot.
- RestoreAll now deletes the marker after a successful restore.
- Added tests for both cleanup paths.

• service/agent

  • Initial agent inventory now returns empty arrays for installed and authorized, not null.
  • cloneInventory now preserves empty slices with make + copy.
  • Added test assertions for non-nil empty slices.

• opencode

  • ResolveOpenCodeBinary no longer falls back to "opencode" when absent.
  • It now returns ports.ErrAgentBinaryNotFound if no real executable path is found.
  • Added tests for missing binary behavior and canceled context.

• continueagent

  • Removed refresh-time auth probing.
  • Continue no longer implements ports.AgentAuthChecker, so catalog refresh will mark it installed if cn exists but leave authStatus as unknown.
  • Removed the cn -p hi probe and related tests.
  • Added a regression test ensuring Continue does not implement AgentAuthChecker.

is these changes are good or something else is needed ? @illegalcall

[illegalcall]

Remaining issues on latest head abbd29c:

1. P2: unknown auth agents are still unselectable

Catalog auth is advisory, and Continue now intentionally returns authStatus: "unknown" because there is no safe local auth probe. But RequiredAgentField still disables every agent not present in authorized (CreateProjectAgentSheet.tsx builds authorizedIds and sets disabled: !isAuthorized).

Example: Continue is installed, /agents/refresh returns it in installed with authStatus: "unknown", but not in authorized; the UI shows “Needs auth” and disables it. The user can never select Continue from the guided create/settings/task flows even though spawn is supposed to be the authoritative validation point. Installed + unknown should be selectable, probably with a warning/status label; spawn can surface the real auth failure.

Refs: backend/internal/service/agent/service.go, frontend/src/renderer/components/CreateProjectAgentSheet.tsx

2. P2: Fresh daemon still starts with an empty usable catalog

NewWithAgents initializes installed and authorized as empty arrays, and _shell preloads only GET /agents through ensureQueryData(agentsQueryOptions). Since GET /agents does not run probes, a fresh daemon presents installed=[] and authorized=[] until the user manually clicks Refresh.

Example: user chooses a project folder, the agent sheet opens, both dropdowns are required, but every supported agent is disabled because the cached catalog has no authorized entries yet. First load should trigger refresh, or the UI should distinguish “not refreshed yet” from “known unavailable.”

Refs: backend/internal/service/agent/service.go, frontend/src/renderer/routes/_shell.tsx, frontend/src/renderer/components/CreateProjectAgentSheet.tsx

3. P2: Auth text parser can mark explicit false status as authorized

StatusFromText only has false patterns for loggedin. Outputs like { "authenticated": false } or { "authorized": false } miss the negative block, then match the positive substrings authenticated / authorized and return authorized.

That can put an unauthenticated agent into authorized, making the UI show it as ready until spawn fails later. Please handle structured/common false keys before broad positive substring matches, or parse JSON-ish status output more carefully.

Ref: backend/internal/adapters/agent/authprobe/authprobe.go

4. P3: Docs links are broken

The PR deletes docs/agent/README.md, but README.md and docs/README.md still link to it. docs/README.md also links agent/switching.md, but no docs/agent/* files exist in this branch.

Refs: README.md, docs/README.md

[nikhilachale]

• Unknown installed agents: fixed in frontend/src/renderer/components/CreateProjectAgentSheet.tsx
  - Installed agents with authStatus: "unknown" are selectable.
  - They show Auth unknown with a warning icon.
  - Explicit unauthorized still stays disabled as Needs auth.

  • Fresh daemon empty catalog: fixed in frontend/src/renderer/routes/_shell.tsx

    • First ready daemon load now uses refreshAgents under the existing agents query key, so /agents/refresh runs
      instead of only cached GET /agents.

  • False auth parser: fixed in backend/internal/adapters/agent/authprobe/authprobe.go

    • Explicit false keys like "authenticated": false, "authorized": false, logged_in=false, etc. are checked before
      broad positive substring matches.

    • Covered by new tests in authprobe_test.go.

  • Broken docs links: fixed in README.md and docs/README.md

    • Removed stale links to missing docs/agent/README.md and docs/agent/switching.md.

@illegalcall