Original test gates

New task runs retain the eligible tests present in their workspace before the first model call. Git-ignored inputs are not included. The runner records the files' bytes, modes and original suite command in a content-addressed ZIP archive. The archive keeps internal copies out of ordinary test discovery (including npm test with Vitest). The runner verifies each member against the saved inventory before using it and extracts originals only into temporary storage outside the project for independent replay. Existing repository-internal file symlinks retain their relative link text and entire file-link chain, including the terminal file even when its name is not a test. Every target must be present in the source snapshot. Directory symlinks, absolute or external targets, cycles, dangling links and ignored targets are refused. Replay preserves the original link topology and never writes through a candidate's replacement link. Stage handoffs include the binding identity, relevant assigned tests and the full inventory's location, rather than repeating every test's contents. A run started with --package PATH retains only the eligible tests inside that directory. The command is the package's detected suite, which runs from the package (cd PATH && SUITE), unless --test-command gives one: that command is retained exactly as given and runs from the repository root, so write it as cd PATH && … to run it in the package. The replay of a command in that form is read inside the package as the same runner's command is read at the root: a unittest, pytest, node:test or go test replay that runs no test, or only skipped tests, is refused. See Monorepo packages.

When an existing test is edited, deleted, renamed or replaced, a passing independent validation must also pass the retained original suite against the current implementation. Both candidate and original executions happen in separate source copies. A weakened candidate test cannot replace a failing original assertion. New coverage remains allowed and is executed alongside the candidate suite. If no suite command was available when the inventory was bound, changed tests remain unverified until an explicit user revision supplies one. Unchanged original files use the ordinary independent verification gates.

--status exposes the binding under view.evidence.protected_tests and the replay result under view.evidence.check_replay.protected_tests. A failed replay has a unique receipt and log directory; a restart preserves it. Successful replay receipts are pinned and rechecked before completion.

To explicitly revise a gate, first stop at a reconciled pause before the Validator or combined checkpoint. Supply --revise-protected-tests FILE with --resume-paused. FILE must be JSON with exactly these fields:

• previous_hash: the exact current protected binding hash shown by status.
• files: every currently eligible test (inside the package, for a run started with --package) and every target in its file-link chain. Regular files map to sha256, size and Unix permission mode (an integer). Links map to symlink, the exact relative link text. All entries must match the actual workspace, including terminal targets not named as tests.
• command: the nonempty suite command authorized for the revised inventory.
• reason: the user's rationale for the revision.

For example, test_layout.py pointing to shared/oracle.py requires both entries: "test_layout.py": {"symlink": "shared/oracle.py"} and "shared/oracle.py": {"sha256": "…", "size": 123, "mode": 420}. Replace the illustrative hash and size with the actual target's values. The binding shown by status points to its complete saved inventory for comparison.

The runner records the revision as a user CLI event, preserves the original bundle, and requires fresh independent validation. A Builder cannot supply this authority in its report, and --test-command alone does not replace the original gate. A gate revision does not approve, edit or relax the product contract; changed requirements still need the normal contract revision and approval process.

Regular-file inventories retain their version-one format; inventories with file links use version two. Those logical inventories and their binding hashes stay unchanged when storage is compacted. On an idle saved-run invocation, AutoCode converts only that run's own current and historical retained directories into verified archive companions. It preserves the saved settings, inventory and revision history, then removes only the copied files named in that inventory. Interrupted compaction resumes without recapturing the current implementation. Corrupt retained inputs fail verification; extra files and other runs' artifacts are left untouched.

An older run that is still active keeps its existing storage until an idle invocation. Other old runs in the same project may still have discoverable backups until they are individually resumed at an idle boundary. New runs never sweep another run's .autocode/ contents. The saved root is a storage locator: new bindings point to a ZIP file; legacy directory locators can resolve a verified ZIP companion after compaction.

Existing saved runs without a binding keep their prior behavior. Their original tests cannot be inferred safely from an already modified checkout; use a new run for this guarantee. The test inventory follows the runtime's existing test path classification. Treat this as protection against accidental gate weakening, not isolation against deliberately hostile test runners or application code.