Aevatar File Extract Probe
Use this to confirm an attached file actually reaches a workflow document_extract step and its
text comes back. Reading files works on the WORKFLOW path only (the document_extract tool lives
in the workflow engine, not the chat tool surface). So this probe drives a workflow.
You (the agent) report the run output honestly. As of 2026-06-12 there is a known boundary (below); a FAIL here is a real, useful finding — it tells you the read-file chain isn't wired yet, not that you did something wrong. Never invent file content.
Prerequisite (one-time)
The probe workflow probe_document_extract must be registered in this chat's scope (ships in the
repo's workflows/ directory; submit it to your scope if /workflow run reports it missing).
How to run it
Tell the user to send the file and the trigger in ONE message (attach a small PDF / DOCX / .txt / .json / .md / .csv, then type the command, send together):
/workflow run probe_document_extract(alias/run-workflow probe_document_extract). Image OCR is NOT supported — use a text-bearing document.Read the run output. The workflow returns JSON:
{"extracted": true|false, "preview": "<first 200 chars>", "note": "<verbatim error if extraction failed>"}.Verdict:
extracted: true+ a preview that matches the file the user sent →READ-FILE OK— the chat-file → document_extract chain works; file-dependent workflows can run.extracted: falsewith a fileRef / argument error innote→READ-FILE BLOCKED (known gap). The chat file's reference reaches the run on the InputFileRefs channel, but the workflow tool_call layer does not yet hand it todocument_extract(no fileRef injection / noinput_file_refsexpression). Quote thenoteverbatim — that error is the confirmation.unsupported_media_type→ you sent a type document_extract can't read (e.g. an image); resend a text document.未找到 workflow/暂未绑定可运行的 actor→ the probe workflow isn't registered/bound; setup step.
Output
One block: did the run start, what was extracted, the preview or the verbatim note, and the
verdict above. If BLOCKED, state plainly that file-dependent workflows (invoice / resume /
reimbursement) cannot run until the chat-file → document_extract handoff is wired.
Guardrails
- Read-only: this probe only extracts text; it never writes, submits, or forwards anything.
- A
extracted: falseresult is honest and valuable — report it as a finding, do not retry endlessly or pretend the file was read. - Never ask the user for tokens — your NyxID-brokered tools handle all credentials.