Almost four hours after I launched it, intake finished and told me the interview was still open. It had read the whole clone, graded it B-, and produced roughly 32,000 words of repository research. There were still decisions it needed me to make.

I cloned AgentWatch into a separate agentwatch-cyoa directory. The source checkout stayed untouched. Both the original and the clone started at commit 768a5fc9fd1ab2e170b8e02e10c4295dc5fd0baf, with a clean main branch.

In Projects I opened Import Existing, selected only agentwatch-cyoa, kept the action on Ingest, and launched it. That created the gitignored .aidd metadata skeleton and started project-intake. The first visible step was Analyze codebase, a nested codebase-analysis recipe launched by intake itself.

The actual chain

The outer recipe has nine steps: analysis, onboarding interview, project profile, artifact check, feature coverage, testing scenarios, audits, parking, and the intake report.

The interview is another recipe rather than one opaque prompt. Its own nine-step chain generates questions, answers them, and reviews those answers before anything downstream is allowed to read them. Nothing reaches the remediation step until the answers exist and have been checked.

Most of the three hours forty-four was real work rather than one stuck prompt. The nested interview alone took two hours and twenty minutes, over half of that spent answering its 12 sections. Every outer step completed, and both interview steps finished before anything downstream read their output. Only after they succeeded did doc2feature receive .aidd/responses and create its candidate remediations.

What intake found

Three findings came back High: incomplete MCP pagination, silent partial scans, and credentials which could cross cleartext HTTP before the tool warned about them. That last one caught my attention. The analysis took thirteen minutes, and I wanted to follow up on the credential handling before spending time on the smaller findings.

The analysis also flagged terminal-control handling and unbounded notification bodies as worth following up.

The interview generated 50 prompts, most of them Critical or High. The review called the technical assessment strong, but left the interview open because product authority, assurance policy, credential-bearing HTTP, release ownership, and several other decisions still need an owner.

The final inventory holds 21 feature records. Eight describe capabilities already present in the source and stay completed with passes: true. Thirteen are remediations parked at waiting_approval: eight out of the interview and five that survived the audit review. Those cover the placeholder architecture map, branch-heavy untrusted-source matchers, a missing dead-code gate, a deploy command pointing at an absent guide, and two production modules close to the 300-line ceiling. Intake also wrote 10 end-to-end testing scenarios and left 43 reports in the evidence directory, mostly clean, duplicate, or not applicable. Only the reviewed findings became feature records.

The early analysis couldn’t run several quality checks because the clean clone had no installed dependencies. Later intake work established the dependency environment and recorded a passing smoke:qc plus 130 passing tests. Those are two dated observations from different points in the same run.

The source boundary held throughout. The clone stayed at 768a5fc9fd1ab2e170b8e02e10c4295dc5fd0baf with a clean tracked tree, matching the untouched original checkout. Everything intake learned or generated stayed under ignored .aidd metadata.

The stop

Near the end, intake changed all 13 open records from backlog to waiting_approval, validated the inventory, and only then wrote its report. No open record was left sitting where an ordinary coding run could select it quietly.

The final report also preserved some awkward metadata facts. spec.md and screen-map.md are absent, the old project-structure file describes a backend/frontend template which AgentWatch is not, and the artifact snapshot was taken before the new testing-scenarios file existed. Those gaps still need work.

I still need to supply the ownership and policy answers. The source gave intake a lot to work with, but those decisions weren’t in it.

What do you do next?