English documentation · runtime 7.24.4 · SDK 2.6.7. Content is maintained with runtime development; see each guide's scope and review date.

Questions about fixed source files

Status: Current source Scope: AX Trust fixed-context question command Last reviewed: 2026-09-18 Owner: AX Code runtime maintainers

Use ax-code ask to ask a standalone question about explicitly selected code or documentation through an existing AX Trust connection:

ax-code ask --file src/value.py \
  --model defai-01-ax-trust-com/openai/gpt-oss-120b \
  "What does this function return?"

Replace the provider/model with an authorized model on your AX Trust connection. The provider must be an AX Trust connection. The options.axTrust session affinity setting does not enable or disable this command. The command reads the file again on every invocation, supplies its complete contents as fixed context, and adds the gateway’s semantic-cache contract. You do not need to add headers or install embeddings. AX Trust must have the feature enabled for your pool and runtime; otherwise it can return an ordinary answer.

Ask a paraphrase with the same files and options:

ax-code ask --file src/value.py \
  --model defai-01-ax-trust-com/openai/gpt-oss-120b \
  "What does the function return?"

The answer goes to stdout; stderr reports AX Trust cache: MISS, HIT, EXACT_HIT, BYPASS or UNREPORTED. HIT means the gateway reported semantic reuse; EXACT_HIT means its exact namespace matched. UNREPORTED means the server did not provide a recognized cache header. A bypass is not retried to force a hit. Add --json for answer, provider/model identity, context digest, cache status, optional request ID/score and usage when actually reported by upstream. Cached responses need not include usage; absent usage is not a savings estimate.

Repeat --file to include additional context:

ax-code ask --file src/value.py --file docs/value.md \
  --model defai-01-ax-trust-com/openai/gpt-oss-120b \
  --max-tokens 1024 --json \
  "How is the return value determined?"

Select all dependencies needed to answer the question. The command sends no conversation history, discovered repository instructions, tools, search results or test execution. It does not edit files or run commands. A new invocation reads current bytes; changed code or options produce different fixed context. A file may change after it was read, so the answer describes the supplied snapshot. Use normal ax-code run or the TUI for live investigation, edits and tests.

Limits: 1-14 distinct regular UTF-8 files inside the current directory tree, 96 KiB combined file content, and 128 KiB after request encoding. Context is never truncated. Symlinks cannot escape the directory. Read permissions for the configured default agent apply; a deny or an unresolved ask stops before model transport. Select permitted files or use a normal session with approval.

The question must be one line, at most 1024 UTF-8 bytes. Put inline code in files, not in the question. The output limit is 1-4096 tokens, default 512; temperature is fixed at zero. Requests are non-streaming, have a 90-second deadline and do not fall back to a different model. Incomplete or actionable responses are rejected. Similar questions can still have different meanings: opt in only for questions whose answer depends on the selected fixed content.

In the TUI, type /ask or select Ask about fixed files from the command palette. It uses the currently selected model. Enter one relative path per line (paste multiple lines to select several files), then enter your question. The dialog shows the server directory, answer and actual cache status. Press r to ask another question using the same file selection; the server rereads those files. Escape closes the dialog and cancels a pending request. Results remain in the dialog and are not inserted into the coding session or executed as tools. Attached TUI clients require a server with the fixed-context endpoint.