PaperClipAI/tests/runner-e2e/product-journeys/AGENT-SETTINGS.md

Saved agent instructions

agent-settings-persistence verifies an ordinary instruction edit on an owned native agent and the result of its next task. It has one provider work run and requires the agent-settings-write target capability. The target provider grants that capability only for a fresh disposable agent dedicated to this case. Its exact name must match the prepared context and contain the current campaign ID. It must have no previous work runs or QA settings proof block. The name is an additional fixture binding; the target's capability grant establishes ownership. Target configuration rejects campaigns that combine this case with another paid journey before acquiring a target. Browser-only journeys may share the campaign. The journey does not edit the shared source agent, provider, model, accounts, or feature flags.

1. Read the original managed instruction entry, exact bytes, content hash, and content revision token. Require an empty work-run history. Empty or whitespace-only files use the ordinary editor value because the read-only page shows an empty-state message. Preserve those exact bytes too.
2. Open Instructions, select Edit for the entry, change the Instruction file editor, and use Save changes. The new procedure contains a fresh synthetic reference that does not occur in the task request. Select the configured entry through the file tree on every open or reload, expanding its parent folders. An existing AGENTS.md must not divert the edit from a different configured entry. Confirm persisted bytes and the new save audit before navigating away. The same confirmation applies when restoring the original instructions.
3. Reload Instructions. Require the exact saved bytes in both the browser and the normal read API. Verify a new agent.files_updated audit event with the signed-in user, exact owned agent, entry path, and content hash. Activity reads use both entity filters before the API limit. Unrelated company history cannot consume that page. A full owned page or foreign entity fails the coverage check rather than being silently filtered.
4. Create one ordinary task through the shared composer. The agent must use the saved procedure to produce the settings-proof task document. Require the exact reference, one committed native run on the configured provider, model, and environment, and current-run document authorship. Pending questions, checkout, or recovery cannot pass. The dedicated nativePhase column must be committed. The grader does not require resultJson.finalizationPhase, which normal bounded run reads can omit.
5. Reload the task and read the saved document. Verify the entry file is still unchanged after the agent's task.
6. Restore the original bytes through the same Instructions controls. Verify the restored browser/API content, its original content token, a distinct board save audit, and no second task run. Return to the task result for the ordinary authenticated screenshot.

Current instruction revisions are content tokens. Identical restored bytes have the original token; the journey does not invent a database parent/child revision chain. The independently retained board audit establishes each actual save.

The independent grader rejects wrong or missing settings, leaked task inputs, foreign agents/tasks/runs/documents, a changed entry, unknown runtime identity, uncommitted work, extra runs, and an unconfirmed restoration. A failed work run remains failed. The public flow does not edit instructions during failed or active work; the target's normal cleanup stops and terminates that owned agent.

Private evidence consists of agent-settings-observations.json and observation-read-failures.json under this case's private-stories directory. The first contains closed IDs, hashes, the synthetic reference, checks, and both matched board-save receipts. Each receipt includes its audit ID, actor, agent entity, entry path, content hash, and save source. It does not contain raw instruction files, account bindings, or provider payloads. Discovery and local controls do not qualify a staging run. The daily plan may admit this case only after a genuine staging pass and reviewed source merge.