Model configurations

A running ledger of which provider, model and role combinations have been tried, and what happened. models.toml says what is shipped; this page says what is known to work, and why the rest does not.

Update it whenever a role changes model or provider, or a provider surprises you. An entry is only works if a real run finished and its output was used. A failure is recorded with its cause when the cause is known, and with the exact error name when it is not.

Status	Meaning
works	A real run finished and its output was used
fails	It was tried and failed; the cause is in the notes
unconfirmed	Configured, but no run has been recorded yet

Providers

What each endpoint is known to accept. This is the part that carries over to the next model you try.

Alibaba Model Studio (DashScope): Qwen and hosted DeepSeek

Endpoint: https://dashscope-intl.aliyuncs.com/compatible-mode/v1, or the workspace endpoint of the same form shown in the console.

• works: Chat Completions honours thinking_budget exactly. Probed on qwen3.8-max-0902 on 2026-09-28: a budget of 500 gave 500 reasoning tokens, 2000 gave 2000, both finish_reason=stop.
• works: the switches are extra_body fields: enable_thinking, thinking_budget and enable_search.
• works: DeepSeek V4 models served through DashScope do run web search (2026-09-27).
• fails: the Responses API. It ignored token limits (8242 reasoning tokens against a budget of 6500) and heavy calls stalled past 600 s. ats-strike stays on Chat Completions.
• Some models cap max_tokens at 8192; set it per model when the endpoint says so.
• works: max_completion_tokens, which is what LangChain's ChatOpenAI sends for max_output_tokens. Probed on qwen3.8-flash on 2026-10-08: a cap of 300 with thinking_budget = 100 gave 100 reasoning tokens and 303 output tokens in all, finish_reason=length. It counts the thinking too — the old max_tokens (403 tokens out under the same test) capped the reply alone — and it must be larger than thinking_budget, or the call is a 400.
• works: a strict json_schema for TailoredCVData, every key required and every object closed (qwen3.8-flash, 2026-10-08).

DeepSeek (native)

Endpoint: https://api.deepseek.com.

• Model ids: deepseek-flash (V4.1-Flash, current) and deepseek-v4-pro. deepseek-v4-flash is still accepted but is a retired name served by V4.1-Flash (pricing page, checked 2026-10-08).
• fails: web search. The endpoint accepts a web_search tool and runs no search; its docs list every built-in tool as ignored (2026-09-27). Search needs DashScope-hosted DeepSeek instead.
• In thinking mode DeepSeek ignores temperature, so the review role's temperature = 0 has no effect there (per DeepSeek's docs; not yet observed here).
• works: function_calling with thinking = "off", through model_provider = "deepseek" (deepseek-flash, 2026-10-08). With thinking on, DeepSeek refuses the forced tool call function_calling and json_schema both send; use json_mode there.
• works: json_mode with thinking on (deepseek-flash, 2026-10-08). The call succeeds; the reply follows the schema only when the prompt carries it, which json_mode roles now always do.
• works: tool calls that are not forced, with thinking on and parallel_tool_calls — what the validation role sends (deepseek-flash and deepseek-v4-pro, adapticv-ai's bench, 2026-10-19).

OpenAI

Endpoint: https://api.openai.com/v1.

• gpt-6-luna is a valid API id (released 2026-09-22) and takes reasoning_effort from none to max (OpenRouter listing).
• fails (reported, not yet observed here): temperature other than 1 while reasoning is on. GPT-5.6-family models answer 400 Unsupported value: 'temperature' does not support 0.4 with this model (OpenAI forum); LiteLLM's Luna notes say the same for Luna. Fix: ATS_<ROLE>_TEMPERATURE=1, or empty to send none.
• fails (expected): presence_penalty with api = "responses": the Responses API has no such parameter, and the openai SDK raises before sending (TypeError, checked offline 2026-10-08). Leave the client's presence_penalty unset for a role on the Responses API.
• fails (expected, not yet observed here): enable_search. It is a DashScope parameter, shipped in the analysis and letter roles' extra_body. Fix: ATS_<ROLE>_EXTRA_BODY={}.
• Do not leave thinking = "on" on an OpenAI role: it sends enable_thinking, which OpenAI does not take. Use ATS_<ROLE>_THINKING=auto.
• Chat Completions function calling was reported to need reasoning_effort = "none". It did not on gpt-6.1-sol: adapticv-ai's validators made their replace_in_* tool calls with medium (2026-10-19). The validation role sends tools; if a model refuses them while reasoning, set its reasoning_effort to none.

Roles

One row per configuration tried, newest first within a role. Provider is the endpoint the call went to: use_alternate_provider empty means the [provider] default.

analysis (JD analysis)

Until 2026-10-08 this role was summary, and also wrote the cover letter; the rows below are its JD-analysis runs. Settings named in the old keys: web_search = true is now extra_body = { enable_search = true }, json_object is now json_mode.

Provider	Model	Settings	Status	Date	Notes
OpenAI (use_alternate_provider = 3)	gpt-6-luna	api = responses, structured_output = json_schema, reasoning_effort = xhigh, thinking = auto, temperature = 1, extra_body = {}	fails	2026-10-09	provider_error: … OpenAIAuthenticationError (HTTP 401) on the first call, from the deployed server. Routing checked: the role resolves to provider 3 and https://api.openai.com/v1, and the model, API and settings in the server log match. So OpenAI rejected the provider-3 key for /v1/responses. Which reason (invalid key, or a restricted key without the Responses scope) is not yet known.
OpenAI (use_alternate_provider = 3)	gpt-6-luna	reasoning_effort = xhigh, thinking = auto, temperature = 1, web_search = false, json_object	unconfirmed	2026-10-08	Set after the two failures below.
OpenAI, misrouted	gpt-6-luna	use_alternate_provider = 1, thinking = on, temperature = 0.4, web_search = true	fails	2026-10-08	1 means the default provider, so the call went to the DashScope endpoint with the DashScope key. Then 400 BadRequestError once routed; the request carried the temperature and thinking switch OpenAI does not take. The error text itself was not read.
DashScope	qwen3.8-flash (shipped)	thinking = on, reasoning_effort = max, temperature = 0.4, extra_body = { enable_search = true }, json_mode	unconfirmed	2026-10-08	Current models.toml.
DashScope	deepseek-v4.1-flash	reasoning_effort = max, temperature = 0.4, json_object	fails	2026-10-06	Both attempts replied with a search-call shape ({'query': …, 'top_k': 5}) instead of the analysis: 12 JDAnalysis validation errors.
DashScope	deepseek-v4.1-flash	reasoning_effort = high, temperature = 0.4	fails	2026-10-06	AuthenticationError on two postings. Cause not recorded.
DashScope	qwen3.7-flash-2026-07-15	reasoning_effort = high, temperature = 0.4	fails	2026-09-30	PermissionDeniedError on all 64 requests. Cause not recorded.
DashScope	qwen3.7-flash-2026-07-15	reasoning_effort = low, temperature = 0.4	fails	2026-09-29	Finished (stop) but the reply failed JDAnalysis validation, twice.
DashScope	qwen3.7-flash-2026-07-15	reasoning_effort = low, temperature = 0.1	fails	2026-09-28	APIConnectionError on four requests.
DashScope	qwen3.8-flash	reasoning_effort = low, temperature = 0.1, max_tokens = 3000 or unset	fails	2026-09-28	Finished (stop) but the reply failed JDAnalysis validation, twice, on four postings. The analysis schema and prompt were reworked afterwards ("solve errors in analysis", 2026-10-06).

letter (cover letter)

Provider	Model	Settings	Status	Date	Notes
DashScope	qwen3.8-flash (shipped)	thinking = on, reasoning_effort = max, temperature = 0.4, extra_body = { enable_search = true }	unconfirmed	2026-10-08	Current models.toml; split out of summary.

The six CV steps (jd_evaluation, profile_scoring, format_selection, cv, validation, highlight)

Since 2026-10-20 a CV is written in adapticv-ai's six steps, one role each. The rows here are adapticv-ai's own bench of 2026-10-19 on the ten quick JDs of tests/cv_quality/ — the same prompts, called through ChatOpenAI from that project, not through ats-strike — and ats-strike's shipped setting. Each row is the whole pipeline: every step on the same model, reasoning_effort high for profile_scoring and medium for the other five, no temperature.

Provider	Model	Settings	Status	Date	Notes
DeepSeek (use_alternate_provider = 2)	deepseek-flash (shipped)	model_provider = "deepseek", thinking = auto, reasoning efforts as above	unconfirmed	2026-10-20	Current models.toml; not yet run through ats-strike.
DeepSeek, native (adapticv-ai)	deepseek-flash	thinking by default, reasoning efforts as above	works	2026-10-19	21–27 calls a CV, 3.5–5.9 min, 205k–258k tokens, most of them thinking (9.4k–14.9k on the longest scoring calls). Validation and tool calls (parallel_tool_calls) accepted.
DeepSeek, native (adapticv-ai)	deepseek-v4-pro	the same	works	2026-10-19	20–25 calls, 9.7–13.4 min, 282k–354k tokens. The validators made replacements through the tool calls.
OpenAI (provider 3, adapticv-ai)	gpt-6.1-sol	Chat Completions, reasoning efforts as above	works	2026-10-19	20–22 calls, 2.2–3.1 min, 71k–81k tokens. Tool calls with reasoning_effort = medium on Chat Completions: the validators made replacements.

cv (CV writing)

Until 2026-10-20 this role wrote the whole CV in one call, to the TailoredCVData schema, in the generate → review → condense loop. It now writes the body and each experience in the six-step pipeline above; the rows below are the old loop's.

Provider	Model	Settings	Status	Date	Notes
OpenAI (use_alternate_provider = 3)	gpt-6-sol	Chat Completions, json_schema, reasoning_effort = xhigh, thinking = auto, extra_body = {}, max_output_tokens = 24500	unconfirmed	2026-10-16	Set after the failure below.
OpenAI (use_alternate_provider = 3)	gpt-6-sol	Chat Completions, json_schema, thinking = auto, extra_body = { reasoning = { mode = "pro", effort = "xhigh" } }, max_output_tokens = 24500	fails	2026-10-16	400 Unknown parameter: 'reasoning'. The reasoning object is the Responses API's; Chat Completions takes only a top-level reasoning_effort, and pro mode (reasoning.mode = "pro") is documented for Responses only (OpenAI reasoning guide).
DashScope	qwen3.8-max-0902 (shipped)	json_schema (strict), thinking = on, extra_body = { thinking_budget = 7500 }, max_output_tokens = 24500	unconfirmed	2026-10-08	Same room for the reply as the max_tokens = 17000 before it: the new cap counts the thinking.
DashScope	qwen3.8-max-0902	json_schema_strict, thinking = on, thinking_budget = 7500, max_tokens = 17000	unconfirmed	2026-10-08	Current models.toml. The budget behaviour is confirmed under Providers. CVs were produced on 2026-10-06, but the client logs of the time did not record the model id; the server's log did (a CV's log.log now records it per call).
DashScope	qwen3.8-max	json_schema_strict, thinking_budget = 6000, max_tokens = 16000, temperature = 0.9	works	2026-09-23	Generate finished in 133 s with 5610 reasoning tokens, under the budget. The review pass on the same model finished in 42 s.

review (CV against the master profile) — retired

Retired on 2026-10-20 with the generate → review loop; the six-step pipeline's validation role replaces it.

Provider	Model	Settings	Status	Date	Notes
DeepSeek (use_alternate_provider = 2)	deepseek-flash	reasoning_effort = high, no thinking_budget, temperature = 0, plain text	unconfirmed	2026-10-08	Thinking mode makes DeepSeek ignore the temperature (see Providers).
DashScope	deepseek-v4-pro-0813 (shipped)	extra_body = { thinking_budget = 9000 }, max_output_tokens = 24000, temperature = 0, plain text	unconfirmed	2026-10-08	Same room for the reply as max_tokens = 15000 before.
DashScope	deepseek-v4-pro-0813	thinking_budget = 9000, max_tokens = 15000, temperature = 0, plain text	unconfirmed	2026-10-08	Current models.toml. The review loop ran on 2026-10-06 (violations listed, CV regenerated); the model id is in the server log.

detect (is it a job posting?)

Provider	Model	Settings	Status	Date	Notes
DeepSeek (use_alternate_provider = 2)	deepseek-flash	temperature = 0, max_tokens = 1000, thinking = off, plain text	unconfirmed	2026-10-08	thinking = off sends enable_thinking: false, which DeepSeek's native API does not document.
DashScope	qwen3.8-flash (shipped)	same settings	unconfirmed	2026-10-08	Current models.toml.

highlight (the **bold** pass)

Since 2026-10-20 the highlight is the last of the six CV steps, with adapticv-ai's prompt; the row below is the old loop's.

Provider	Model	Settings	Status	Date	Notes
DashScope	qwen3.8-flash (shipped)	thinking = off, temperature = 0.1, json_object	unconfirmed	2026-10-08	Current models.toml.

Recording a result

Add a row under the role, with the provider, the exact model id, the settings that differ from models.toml (including any ATS_<ROLE>_* override), the date and the outcome. For a failure, the error name from the job log (provider_error: … PermissionDeniedError) or the server's own line for it (a CV's log.log carries every one) belongs in the notes; the cause belongs there once it is known. Never record keys, workspace URLs or request payloads.