fix(cli): point integration_test users at the platform flow

[ayushnangia]

Summary

A user following the official 01-coding-template README's authoring loop runs:

hud eval <env-name> integration_test --task-ids my-task -y

Against a local source (or before deploying), the CLI rejects the agent with a bare Invalid agent: integration_test. Must be one of: claude, openai, gemini, openai_compatible — leaving the documented flow with no explanation that integration_test is the platform's authoring agent (it pre-stages Task.validation, runs the graders, and expects Reward 1.0).

Changes

• hud/cli/eval.py: the agent_type validator recognizes integration_test and raises a precise, actionable error — deploy the environment and run it against the platform (hud eval <env-name> integration_test --task-ids <slug> -y); not available with --runtime local. Unknown agents keep the existing generic message.

Validation

• New test in hud/cli/tests/test_eval_config.py (test_parse_agent_type_points_integration_test_to_the_platform).
• pytest hud/cli/tests/test_eval_config.py → 38 passed; full local suite → 136 passed.

Scope note

Deliberately an error-message fix, not a local implementation of the integration_test agent — that belongs on the platform side and would be a feature PR.

---

Note

Low Risk
Changes are user-facing error text and Harbor schema guards on adapt/export; no auth, runtime execution, or agent behavior changes beyond clearer validation.

Overview
CLI: Choosing agent integration_test no longer surfaces a generic “invalid agent” message. EvalConfig raises a specific error explaining that integration_test is the platform authoring agent (pre-stages golden Task.validation, expects Reward 1.0) and must be run against a deployed environment on the platform—not with --runtime local.

Harbor integration: Introduces HARBOR_SCHEMA_VERSION = "1.0". harbor.adapt rejects task.toml files that declare a different schema_version; missing schema_version still adapts for legacy exports. export writes schema_version into generated task.toml for round-trip consistency. Contract tests cover unknown, supported, and absent schema versions.

^{Reviewed by Cursor Bugbot for commit 612ccb7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ca13eb0. Configure here.}

[ayushnangia]

Update: the real thing now exists locally. #586 implements integration_test as a local agent — it replays Task.validation through the task's own MCP capabilities, lets the scenario graders run, and the CLI enforces the Reward-1.0 authoring gate (exits non-zero below 1.0, naming the failing raw grade). If #586 lands, this PR's error path is superseded: the name parses, and the only remaining refusal is for the gateway shortcut (integration_test makes no LLM calls). Happy to close this one in #586's favor at that point — or keep it if reviewers want the hint to remain for the platform-only flags.

[ayushnangia]

Thanks Bugbot — the guidance is intentionally the interim behavior: this PR only adds the helpful error, and #586 (same author) implements the real local agent, which makes the name parse and replaces this error path. Cross-comment is on #586's description; happy to close this one in #586's favor if reviewers prefer.