Changelog

This page is for an agent checking what changed in the Zaun API and documentation before it writes or updates an integration. Entries are newest first and list only changes a customer can see. If you are the person, this is the page to watch for changes that need action; those entries say what to do.

2026-10-13: Documentation for agents#

• The docs are published for agents: /llms.txt lists every page with its Markdown URL, /llms-full.txt holds every page in one file, and each page is served as Markdown at /docs/<page>.md, including the API reference at /docs/openapi.md.
• New pages: Set up with your agent, Try it, Choose a capture path, one page each for Claude Code, Claude Cowork, Codex, Cursor and OpenTelemetry for custom agents, and Hex v1.3 known limits.
• Limits and Responses state the POST /v1/responses time budget: a non-streaming response must finish within 180 seconds, and a stream must release its first text within 180 seconds, or the request returns 504 deadline_exceeded. Action: send stream: true for long generations and high reasoning effort.

2026-10-13: Scans that could not be assessed are not billed#

• A scan Hex v1.3 could not complete, which returns as an alert for review, is no longer charged: its usage reports zero tokens. Images read for such a scan are still charged.

2026-10-13: Session calls are billed once, with cached tokens#

• A scan with a session_uuid is billed for the tokens it assesses for the first time at the scan rate ($0.20 per million) and for the session's earlier tokens, read back as context, at $0.02 per million. Before, every session call was billed for the session's whole count at the scan rate.
• usage.input_tokens now reports only the new tokens, and the new usage.cached_input_tokens reports the earlier tokens. A resend with nothing new bills zero new tokens. Requests without a session and document scans are unchanged. Action: if you total usage yourself, add cached_input_tokens at its own rate.

2026-10-13: GPT-6 reasoning effort#

• The GPT-6 routes on POST /v1/responses accept every reasoning.effort level; gpt-6-sol and gpt-6-luna default to medium, and gpt-6-astra has no documented default, so its zaun.effort.default is null. On gpt-6-sol and gpt-6-luna, send function tools or an effort, not both. See Responses.

2026-10-13: Detection samples and the test string#

• GET /v1/detection-samples lists Zaun-written sample requests, one or more per category plus benign ones, each with the outcome expected under the default controls of a new self-serve project. No key is needed. The console Playground offers the same samples as one-click runs, up to 30 runs a minute per organization.
• The test string ZAUN-TEST-VIOLATION-7f3a9c is always reported as a blocked input check named zaun_test, in prompts and inside tool arguments. No project setting can lower it. Use it to prove a key, a hook, a gateway or an alert; see Try it.

2026-10-13: One assessment path for every project#

• Project configurations no longer accept classifier.mode: reference. Projects that selected it receive a new configuration version that selects cascade_v1, the default. Projects on primary are unchanged. Action: if you publish configurations through the administration API, stop sending reference.
• Results keep stages.embed and stages.knn, now always skipped, so existing clients keep parsing them. New results never carry signals.knn.
• The error codes embedder_unavailable, vector_search_unavailable and embedding_invalid are no longer returned.
• A new organization's keys can scan as soon as the organization is enabled, and the console Playground no longer waits for setup.

2026-10-05: Completion assessment usage and interrupted completions#

• Input and tool-call assessments on POST /v1/responses report the tokens they are charged in zaun.scan.usage, zaun.output_scan.usage and GET /v1/result, instead of zero. What you pay is unchanged: they were always billed at the scan rate. Assessments recorded before this change keep reporting zero until they expire.
• A completion that stops before the model call finishes now settles its credit hold. Usage the model already reported is charged when the input assessment blocks or fails, a streaming client stops reading, or held text exceeds its cap. A request that never reached the model, or that was refused before generating, releases its hold without a charge.

2026-09-30: /v1/docs redirects to the console#

• GET /v1/docs returns a 308 redirect to https://console.zaun.ai/docs, and the API reference documents it.

2026-09-26: GPT-6 models on POST /v1/responses#

• gpt-6-astra, gpt-6-sol and gpt-6-luna are available through POST /v1/responses and listed by GET /v1/models. See models for limits and prices.
• Requests large enough to reach the provider's higher long-context price are rejected before any charge.

2026-09-22: Generation moves to the Responses API#

• POST /v1/responses is compatible with the OpenAI Responses API and its official SDKs. It takes input as a string or as message, function_call and function_call_output items, plus instructions, function tools, tool_choice, max_output_tokens and stream.
• POST /v1/chat/completions and POST /v1/completions are removed and return 404. Action: move callers to Responses.
• Responses are stateless: store: true and previous_response_id return 422. Send the full conversation as input items on every call.
• Streams send typed Responses events with sequence_number and end with response.completed, response.incomplete or error; there is no [DONE] line.
• A blocked output ends as status: incomplete with incomplete_details.reason: content_filter; a token cutoff reports max_output_tokens.
• usage uses the Responses spelling: input_tokens, output_tokens, input_tokens_details.cached_tokens and cache_write_tokens.

2026-09-22: Results no longer name models or vendors#

• GET /v1/result, GET /v1/sessions/{session_uuid} and completion assessments no longer return signals.classifier, signals.path, usage[].model, revisions.model or revisions.corpus. revisions.pipeline is an opaque fingerprint that changes whenever the assessment changes.
• Only latency_ms.total is reported; per-stage timings are gone.
• Document results drop extraction_evidence.provider and extraction_evidence.model_version and report coverage.extraction_source: ocr for recognized text. OCR regions and confidence remain.
• usage on results is the billed {input_tokens, output_tokens}, matching the scan response.
• GET /v1/usage no longer accepts group_by=model or group_by=stage. Day grouping is unchanged. Action: remove those parameters from usage queries.

2026-09-22: Faster, fail-soft document scans#

• A PDF whose own text layer covers every page is read from that layer (extraction_source: pdf_text_layer) in milliseconds. Scanned or partly scanned PDFs are read with OCR, page by page.
• TXT, Markdown, CSV and HTML are scored exactly as uploaded. DOC, DOCX, XLS and XLSX go through a native parser first.
• A document with no text anywhere is refused.
• A transcription that is refused, times out or comes back empty is reported inside an HTTP 200 body with status: failed, error.stage, review_recommended and review_reason. HTTP 4xx is reserved for requests rejected before any processing. Idempotent replays return the same body. See document scans.