[KarthikAvinashFI]

1. Why

The harness test suite (tests/test_harness.py and tests/harness/) runs in no CI job, so it drifted: a merge glued an old test body into another test, fakes fell behind the code they stand in for, and a few checks still describe a tool surface the harness no longer has. This PR makes the suite green again and adds a CI job that keeps it that way.

2. What changed

File	Change
.github/workflows/harness-tests.yml (new)	On every PR and on pushes to dev and main: uv sync --extra livekit --extra harness-stores --extra harness-ui --extra hosted-tracing, then uv run pytest -q -p no:cacheprovider tests/test_harness.py tests/harness
tests/harness/test_grade.py	Splits the fan-out test a merge had glued into test_the_cap_holds_for_the_session_that_saves_too (the NameError on st); then drops that fan-out test: it drove the generate_suite tool, which sub-agent writers replaced, and write_in_parallel has no caller left
tests/harness/test_dispatch_ack.py	The duck-typed report and failure fakes carry the fields the call runner now logs (run_id, test_case_id, failure.details, failure.stage) and the runner context it reads (usage_reporter, evidence_seam)
tests/test_harness.py	test_a_skill_only_names_tools_its_stage_actually_has derives every field of the contract's nested models, the values Literal fields and the sub-goal contract offer, and the sub-agent name, instead of a hand list; it keeps the hand list only for the example tools the scenario skill uses to show a check going wrong. test_hosted_claude_session_uses_durable_resume_and_restricted_tools runs on a Gemini model (the harness refuses Claude models) and checks the current tool surface: allowed_tools is exactly the spec's builtins, no MCP server, strict config. test_the_system_prompt_stays_within_budget caps the prompts at today's size plus room

3. Behavior callouts

• No product code changes; tests and CI only.
• The prompt-size cap moves from 90,000 / 130,000 characters to 185,000 / 270,000: the write-scenarios skill alone is 124,000 characters today. The cap still fails a skill that keeps growing; shrinking the skill is its own work.
• The session test no longer asserts options.tools == builtins or that AskUserQuestion stays out of allowed_tools: since the scenario-writers change, the Claude backend grants builtins through allowed_tools and restricts the rest with the gate hooks and permission callback.

4. Tests

Run	Result
uv run pytest -q tests/test_harness.py tests/harness with the CI extras, on this head	5163 passed, 20 skipped
Same, before this PR, without the extras	18 failed, 36 errors; 46 of them were missing livekit / psycopg, which the CI job installs

Mutation check for the skill test: a skill naming a tool that does not exist (seed_the_wrold) fails it.

Not in this job: tests/test_cli_examples.py (11) and tests/test_config_and_facades.py (37) fail on dev too, back to at least 8 Oct (a0e2ef23); they exercise the SDK examples and optimizer scoring, not the harness.

5. Commands

uv sync --extra livekit --extra harness-stores --extra harness-ui --extra hosted-tracing
uv run pytest -q -p no:cacheprovider tests/test_harness.py tests/harness

Linear

Fix TH-8768