refactor(agents): migrate analyst + browser to hub (#1102)

[kovtcharov-amd]

Why this matters

AnalystAgent (data) and BrowserAgent (web) were the last two registry builtins backing gaia analyze / gaia browse. They now ship as standalone gaia-agent-analyst / gaia-agent-browser wheels under hub/agents/python/, discovered via the gaia.agent entry point — so the core framework wheel no longer hardcodes them and they version independently. Their full+lite model tiers are preserved exactly; the shared tier builder is promoted to registry.build_model_tiers so hub packages reuse the identical ~4B preset rather than re-deriving it.

Continues the chat-family core wave for #1102 (after connectors-demo #1442). gaia browse/gaia analyze now resolve their agent through the registry and fail loudly with an install hint if the wheel is absent.

Test plan

• python util/lint.py --agents — clean (analyst/browser optional)
• pytest tests/unit/agents/test_registry.py tests/unit/test_agents_split.py tests/unit/cli/test_cli_smoke.py — green (framework-only tolerates the now-installed agents)
• pip install -e hub/agents/python/analyst hub/agents/python/browser && pytest hub/agents/python/{analyst,browser}/tests/ — 8 passed (incl. discovery + tier shape)
• black / isort / pylint -E — clean
• CI: new Analyst Agent Tests + Browser Agent Tests workflows install the wheels and run their tests

[github-actions]

Summary

Clean, well-scoped refactor — Approve with suggestions. This moves the last two registry builtins (data/AnalystAgent, web/BrowserAgent) out to standalone gaia-agent-analyst / gaia-agent-browser wheels under hub/agents/python/, discovered via the gaia.agent entry point, and faithfully mirrors the connectors-demo migration (#1442). The model tiers are preserved exactly, the shared tier builder is promoted to registry.build_model_tiers so hub packages reuse one source of truth instead of re-deriving the lite preset, and gaia browse/gaia analyze now resolve through the registry and fail loudly with an install hint when the wheel is absent. No functional regression: the CLI's config kwargs are all valid *AgentConfig fields, so routing them through the factory's field filter preserves behavior.

The one thing worth acting on before merge: the analyst package lost test parity with browser — it has no discovery/registration test, so an entry-point or build_registration() regression in the analyst wheel would ship unnoticed.

Issues

🟢 Minor — Analyst package missing discovery + registration tests (hub/agents/python/analyst/tests/test_analyst_agent.py)

The browser package's fresh test file covers the three things that actually validate the packaging change — build_registration() shape, agent import, and registry discovery when installed. The analyst tests were carried over from the old tests/unit/agents/ location and only cover the agent itself (import / config / tool set), so the new wheel's entry-point wiring is untested. Add the two parity tests so a broken gaia.agent entry point or malformed build_registration() fails CI:

def test_build_registration_shape():
    import gaia_agent_analyst as m

    reg = m.build_registration()
    assert reg.id == "data"
    assert reg.namespaced_agent_id == "installed:data"
    assert reg.source == "installed"
    assert [t.name for t in reg.model_tiers] == ["full", "lite"]


def test_discovered_when_installed():
    from gaia.agents.registry import AgentRegistry

    reg = AgentRegistry()
    reg.discover()
    assert "data" in {a.id for a in reg.list()}

🟢 Minor — CLAUDE.md agent table now stale (CLAUDE.md:500-501)

The agent table still points to agents/analyst/agent.py and agents/browser/agent.py, and the project-structure tree lists analyst/ and browser/ under src/gaia/agents/. Prior migrations updated CLAUDE.md to note the wheel move (e.g. the gaia-code / gaia-emr lines). Worth a one-line update to the two rows so the dev-facing map stays accurate — optional, non-blocking.

Strengths

• Faithful pattern reuse. The hub package layout, lazy __getattr__ re-exports, build_registration(), CI workflow, and setup.py extras all match the connectors-demo migration exactly — nothing novel to second-guess.
• One source of truth for tiers. Promoting build_model_tiers (and LITE_MIN_MEMORY_GB / lite_models()) to module level lets the hub wheels declare the identical full+lite preset without copy-pasting the platform-conditional ~4B list — the right call over duplicating it per package.
• Fail-loud CLI. registry.get(agent_id) is None → RuntimeError with a concrete pip install hint (src/gaia/cli.py) follows the "No Silent Fallbacks" rule and names what's wrong + how to fix it.
• Test isolation handled correctly. Framework-only suites guard the moved imports with pytest.importorskip, and test_registry.py narrows _BASE_AGENTS to the resident chat profiles — so the core wheel's CI stays green without the hub wheels installed, while the new per-wheel workflows cover them.

Verdict

Approve with suggestions. No blocking issues — both findings are minor. The test-parity gap is the only one I'd recommend closing before merge; the CLAUDE.md tidy-up can follow.