Add AGY, Pi, and Goose agent support

[laulpogan]

Summary

• add built-in AGY, Pi, and Goose agents across creation menus, settings, keybindings, process recognition, and custom-agent launch handling
• use each CLI's interactive launch contract: agy --prompt-interactive, positional Pi prompts, and Goose session / run --interactive --text
• enable linked context sharing for all three agents in local and SSH sessions
• install the get-linked-context skill into Claude-compatible, Agent Skills, and Antigravity skill directories
• preserve Goose launch arguments when restarting linked sessions and select the remote context shim for SSH nodes

Validation

• 234 focused tests pass
• npm run typecheck passes
• npm run build passes
• production Electron build launched real AGY 1.1.24, Pi 0.84.4, and Goose 1.45.0 TUIs
• Pi loaded get-linked-context; Goose discovered it through goose skills list
• authenticated context shim listed linked AGY and Goose nodes and read their live terminal output
• live run recorded zero console errors and zero failed requests

Full local suite: 10,172 passed, 25 failed, 37 skipped. All 11 failing test files are unchanged by this branch; failures are existing macOS/environment-sensitive real-shell, socket-path, localStorage, account-environment, and platform-copy expectations.

Scope

Context sharing uses the terminal-output fallback for AGY, Pi, and Goose. Structured transcript parsing for these CLIs is not included.

[eneskirca]

Reviewed on a clean worktree off origin/main. npm run typecheck clean; the touched suites (agents, agent-launch, context-link core + server, noteLink, keybindings, workspace, remote-hooks, plus all of src/renderer/lib and canvas-control-core) are green — 1915 tests, no regressions against main. The boundary/guard tests (no-electron both shells, fs-atomic.guard, config.capabilities) pass.

This is careful work. The discipline I most wanted to see is here: the three agents join exactly one capability list, and config.capabilities.test.ts pins the negative side (hasHooks/canResume/hasUsage/canChat/canReadTitle/… all false) rather than leaving it to inference. resolveLinkTranscript has no locator for them, so it returns '' and the reader says "no transcript yet" — fails closed, and the PR description says so instead of implying structured transcripts work. assembleLaunchCommand/assembleResumeCommand stay byte-identical for every existing agent (empty launchArgs collapses through filter(Boolean)), and both launch paths — the shell assembler and core/agent-launch.ts — got the new fields, which is the parity that usually gets missed. The contextLinkShimPath change also fixes a real pre-existing bug: SSH endpoints were being handed the local shim path.

Three things to address.

1. verify gates on canContextLink, not hasHooks — and this PR is the first to make those two lists differ.

Until now every CONTEXT_LINK_CAPABLE agent was also in AGENT_HOOK_TARGETS, so verify's gate was accidentally sufficient. With agy/pi/goose in the list, verify --node <agy node> and verify --agent goose both pass the two canContextLink checks in Canvas.tsx (~9811 / ~9819), then arm the reviewers and the judge through armAfter. depSatisfied (renderer/lib/pendingLaunch.ts) requires state === 'done', and these agents emit no hook events at all, so the state stays unknown forever — deliberately not treated as satisfied. Result: the whole panel sits QUEUED with only the manual ▶ escape, and nothing tells the user why.

Note the --after flag path already refuses exactly this, with a named error (resolveAfter → hasHooks: "not an agent session that reports when it is done"). verify should refuse on the same predicate — for the target and --agent — rather than building a panel that cannot fire.

2. Unconditional skill installs into two more directories, local and remote.

installSkill() now creates ~/.agents/skills/get-linked-context/ and ~/.gemini/antigravity-cli/skills/get-linked-context/ on every boot, on every machine, whether or not Pi/Goose/AGY are installed — and RemoteHooks.installContextLink now does three sequential remote writes per SSH connect instead of one. Two asks: (a) is creating third-party tool directories that don't exist intended, or should the write be conditional on the parent existing? (b) ~/.agents/skills is a shared directory — a skill placed there is discoverable by tools that are not ours and not in CONTEXT_LINK_CAPABLE; worth a line in the comment saying that's understood.

3. Custom-agent launch parity between the two paths.

custom-agent.ts inherits launchArgs/promptLaunchArgs/promptFlag from the base harness, so the shell path gives a baseAgent: 'goose' custom agent goose-wrap run --interactive --text '…' (nicely covered by your new test). But resolveTrustedAgentConfig in core/agent-launch.ts returns none of those three for the custom branch, so the session-host path emits the bare command with a defaulted --prompt. Either thread them through or say in a comment why the trusted path deliberately doesn't.

Minor, no change requested: the one-shot context-link note is idle-gated on status[selfId]?.state === 'working', which is inert for an agent with no hook state — for these three the note can land mid-turn. It's user-initiated on edge draw so the blast radius is small, but a comment noting the gate is a no-op here would be honest.

What I could not verify: none of agy, pi, goose are installed on my machine, so every launch grammar (agy --prompt-interactive, positional Pi, goose session / goose run --interactive --text) and both discovery-path claims (~/.agents/skills, ~/.gemini/antigravity-cli/skills) are author-reported here, not reproduced. Your validation notes name versions and say the TUIs were launched and goose skills list found the skill, which is the right level of evidence — I'm recording it as device-verified by you, unverified by me.

Verdict: changes requested, item 1 being the one that matters. The rest is close to merge-ready.

[laulpogan]

Addressed all three requested changes in 073d660a and rebased onto current origin/main.

• verify now requires completion hooks for both the target and reviewer. AGY, Pi, Goose, and hookless custom agents fail with a named error instead of remaining queued forever.
• Optional AGY/Pi/Goose skill roots are written only when their parent directories already exist, locally and over SSH. The shared ~/.agents/skills visibility is documented.
• Session Host custom agents now inherit their base agent’s launchArgs, promptLaunchArgs, prompt mode, promptFlag, and separator. A Goose wrapper test pins session and run --interactive --text.
• The context-link comment now states that hookless agents can receive the user-triggered note mid-turn.

Verification: 289 focused tests passed (7 platform skips), TypeScript clean, production build passed. The broader shortcut/keybinding set also passed 87 tests after expanding its synthetic key pool for the three new commands.