When TheAuditor Cannot Answer, Your Agent Can Still Read
TheAuditor steers coding agents to a query instead of a whole-file read. That only helps while TheAuditor can answer. Now the steering steps aside when it cannot, files you just changed stay readable, and no lookup waits on a re-index.
Ten days ago we described a gate that declines an agent’s whole-file read of source code and names the query to use instead, in Every Answer Says What It Left Out. A gate like that earns its place on a good day. Whether it can be trusted depends on a bad one: no analysis yet, a refresh that failed, a project in the middle of an edit. A gate that declines the file and then cannot answer the question leaves the agent with neither, and an agent with neither stops working.
A pre-release audit of our own server found exactly that edge. It is closed, and the rule now is simple: the gate only stands where an answer stands.
The gate steps in only while TheAuditor is answering
The gate engages only while TheAuditor is running and answering for this project. If nothing is answering, nothing is declined, and the agent reads the way it would without us.
When TheAuditor cannot answer a lookup, the refusal says why and tells the agent so in plain words: “Until TheAuditor can answer again, reading files directly is allowed.” The agent is never told to use a query that cannot run.
Files you just changed stay readable
An index describes the code as it was when it was built. A file edited since then is one the index has not seen, so it stays readable while the gate is engaged, and so does any file modified after TheAuditor last looked. For every other file the answers still hold, so the gate keeps steering the agent to the query for those.
No lookup waits on a re-index
No call waits for a refresh. While one runs, or while files are changed against the index, the answer comes from the last finished analysis and says so, naming how many files changed since. The index refreshes once your edits settle, rather than in the middle of your agent’s own test run.
If the last analysis did not complete, the answer says that as well, so a partial index never passes for a whole one.
A failure is said once, not paid for on every call
A refresh that fails is not retried on every lookup. The answer comes from the last analysis with a notice, TheAuditor tries once more after a pause, and after that it waits for your next edit. The agent hears about the failure instead of waiting through it again and again.
Every lookup has a limit
A query that runs past its time limit is stopped, and the answer says so and suggests narrowing it. The next call works. The server, not the agent’s client, says why a call ended, so a stopped query never reads as an empty result.
Honest scope note
This keeps an agent working when TheAuditor cannot answer. It does not make every question answerable: where the index holds no fact for something, the answer still says so rather than guessing. While reading is open, the agent works from the file itself, exactly as it would have without TheAuditor installed. Nothing is hidden from it at any point.
Where this sits
A tool an agent calls in the background has to fail in a way the agent can work around. That is the ground-truth discipline behind Code Reality Labs: say what you know, say what you cannot answer, and never stand between the agent and the code when you have nothing better to offer. TheAuditor is in final commercial release preparation and ships when its hardening checks pass. Subscribe on the main site for launch news.
Was this useful?