Merge into Agents API

[cm2435-hcomp]

What

Route HoloDesktop CLI runtime and session ownership through the shared Python Agent API SDK. Retain CLI/MCP/ACP/A2A entry points, inference/source settings, authenticated attachment and emergency Stop; close SDK-owned device bridges before closing HTTP transport.

Why

HoloDesktop already uses the local Agents API pattern. Moving lifecycle and sessions to the shared SDK removes duplicate ownership and gives products the same placement interface.

How

Delegate launcher, client and session lifecycle to the SDK; preserve request options and protocol entry points, and close device bridges before their HTTP transport.

Validation and dependencies

P6, depends on hcompai/hai-agents-python#222 and the reviewed runtime/driver/contract releases. MUST NOT merge or publish with the current permissive hai-agents>=1.0.12 floor: update to the actual compatible released SDK and lock before landing. No future version is fabricated here. Legacy CLI artifact ownership remains until this migration ships.

Validation: 453 non-E2E CLI tests pass, 16 skipped, against review SDK sources. Prior isolated CLI/MCP/ACP/A2A live probes exercised files, follow-ups and owned-command Stop. Full installed-product visual/packaged parity remains unverified. Local-agent/local-environment and local-agent/cloud-environment QA covers the new runtime path; candidate hosted combinations were not demonstrated. HoloWork migration is deferred.

Stack navigation: P1 contract → P2 runtime; P3 generation + P4 SDK; P5 checks: runtime, generation, SDK pins; P6 CLI. P7 duplicate pin cleanup follows only after P6 ships. HoloWork integration is subsequent work.

Review guide: intent, architecture, service tradeoffs and merge order.

CI dependency constraint: released SDK/contract packages lack local-runtime modules and Desktop.host, so current installer/unit/type checks require the compatible dependency releases. Ruff lint and format steps pass on the current CI head; mypy and message/session tests pass with candidate SDK/contract sources.

[cm2435-hcomp]

bugbot run

[cm2435-hcomp]

bugbot run

[cm2435-hcomp]

cursor review

[abonneth]

Pushed directly (as agreed), aligning with the SDK changes on hai-agents-python#222:

• 415ee0c: launch and attach go through the SDK's LocalRuntime and AsyncClient.local(runtime=...), sharing ~/.hai/agent-runtime. Removed the CLI's own launcher and its unproven health probe. Why: the CLI and the SDK used different state dirs on the same port and broke each other, and an unproven probe hands the bearer to whatever holds the port.
• fc1bb0f: holo agent-api starts on the shared recipe, so other clients can attach.
• ba8bda7, 1f742f8: README/SECURITY updated for the new state paths, identity proofs and the stop --force pid check.

Tests: 370 passed (e2e 84). Smoke against a real runtime:

• run --fake spawns and answers;
• doctor verifies the token;
• a squatter on the port is rejected and never receives the bearer;
• stop --force works;
• 3 concurrent attaches all succeed.

Before merge:

• Released runtimes don't send proofs yet, so this fails closed against them. Needs a runtime release plus a pin bump.
• Daemons started by released CLIs keep their token in ~/.holo. The new CLI can't attach to them, but holo stop --force still finds them.
• The CLI pins runtime 0.1.12 and the SDK pins 0.1.8, which triggers a version-skew warning. Unify them at release time.

[abonneth]

Pushed 42fc682 + b593a60: Windows ARM64 installer.

Why: hai-agents[desktop] pulls pyautogui + 5 deps that ship only sdists. --no-build made ARM64 installs unresolvable.

Change: --no-build → --no-build-package <name> for each manifest wheel (cryptography). Hosted wheels stay binary-only; pure-Python sdists build without a compiler. CI/release smokes drop UV_NO_BUILD so they test the installer's real policy.

Status: ARM64 install now resolves (89 packages). Remaining failures (ARM64 bootstrap No module named hai_agents_local.runtime, calculator did not prove…) are release ordering: they clear once the runtime and SDK with local runtime ship and the floor bumps.

[abonneth]

Simplification pass (no behavior change), edeb018:

• Removed RuntimeSpawnSettings.api_token: no readers left (the SDK reads the env var).
• Removed AUTH_TOKEN_ENV/PORT_ENV re-exports from the launcher; callers import from settings.
• run_turn catches ApiError only: the SDK client raises it for every HTTP error; the httpx.HTTPStatusError branch was only hit by the test fake.
• Test stub imports the identity header names from the SDK.
• Dropped "removed text stays removed" asserts in the static workflow tests.

Tests: 454 passed, 16 skipped; ruff clean.

[abonneth]

Commit	Fix
df91e36	SDK source tracks the current local-runtime head
51793dc	Shared recipe: missing macOS grants point at the app running holo (clean error, no futile runtime restart); doctor checks this process directly. --fast walkthrough unchanged
80bc951	holo serve fails the A2A task with the same grant guidance

Tests: 455 passed, 16 skipped. ruff + mypy clean. Live holo doctor: green.