X-Config

The X-Config header is a JSON object that provides additional product configuration settings.

This header is optional and can be used in Headers-Only Mode to fine-tune session behavior.

Data Structure

{
  "fsm": {
    "min_num_tools_for_filtering": null,
    "clear_session_meta": null,
    "max_n_turns": null,
    "history_mismatch_policy": null,
    "clear_history_every_n_attempts": null,
    "disable_rllm": null,
    "enable_multistep_planning": null,
    "enabled_internal_tools": null,
    "prune_failed_steps": null,
    "force_to_cache": null,
    "max_pllm_steps": null,
    "max_pllm_failed_steps": null,
    "max_tool_calls_per_step": null,
    "reduced_grammar_for_rllm_review": null,
    "retry_on_policy_violation": null,
    "wrap_tool_result": null,
    "detect_tool_errors": null,
    "detect_tool_error_regex_pattern": null,
    "detect_tool_error_max_result_length": null,
    "strict_tool_result_parsing": null,
    "tool_result_transform": null,
    "n_pllm_plans": null,
    "pllm_candidate_models": null,
    "pllm_context_pruning": null
  },
  "prompt": {
    "pllm": {
      "flavor": null,
      "version": null,
      "debug_info_level": null,
      "clarify_ambiguous_queries": null,
      "context_var_visibility": null,
      "query_inline_roles": null,
      "query_role_name_overrides": null,
      "query_include_tool_calls": null,
      "query_include_tool_args": null,
      "query_include_tool_results": null,
    },
    "rllm": {
      "flavor": null,
      "version": null,
      "debug_info_level": null
    },
    "tllm": {
      "flavor": null,
      "version": null,
      "add_tool_description": null,
      "add_tool_input_schema": null
    }
  },
  "response_format": {
    "strip_response_content": null,
    "stream_thoughts": null,
    "include_program": null,
    "include_policy_check_history": null,
    "include_namespace_snapshot": null
  }
}

Top-Level Fields

The configuration is organized into three sections:

• fsm: FSM (Finite State Machine) overrides for execution behavior
• prompt: Per-LLM prompt configuration overrides
• response_format: Response format configuration for dual-LLM sessions

All fields are optional and have sensible defaults.

---

FSM Overrides (fsm)

Shared Fields (Single-LLM & Dual-LLM)

fsm.min_num_tools_for_filtering

Type	Required	Default	Constraints
integer or null	No	null	>= 2

Minimum number of registered tools to enable tool-filtering LLM step. When not set, the server default is 10.

fsm.clear_session_meta

Type	Required	Default
string or null	No	null

When to clear session meta information. When not set, the server default is "never".

• "never": Never clear
• "every_attempt": Clear at the beginning of each PLLM attempt
• "every_turn": Clear at the beginning of each turn

fsm.max_n_turns

Type	Required	Default	Constraints
integer or null	No	endpoint-dependent	>= 1

Maximum number of turns allowed in the session. If null, unlimited turns are allowed. When not set, the server applies a default based on the endpoint type.

Dual-LLM Only Fields

fsm.history_mismatch_policy

Type	Required	Default
string or null	No	null

Controls behaviour when incoming messages diverge from stored history in stateless mode:

• "reject": Reject the request with an error if a mismatch is detected
• "restart_turn": Truncate stored history to the last consistent point and restart the turn
• "continue": Silently accept the caller's version and continue

fsm.clear_history_every_n_attempts

Type	Required	Default	Constraints
integer or null	No	null	>= 1

Single-step mode only. Clear all failed step history every N attempts to save tokens.

fsm.disable_rllm

Type	Required	Default
boolean or null	No	null

Whether to skip the response LLM (RLLM) review step. When not set, the server default is true.

fsm.enable_multistep_planning

Type	Required	Default
boolean or null	No	null

When false (single-step), each attempt solves independently. When true (multi-step), each step builds on previous. When not set, the server default is false.

fsm.enabled_internal_tools

Type	Required	Default
array[string]	No	endpoint-dependent

List of internal tool IDs available to planning LLM. Valid values: "parse_with_ai", "verify_hypothesis", "set_policy", "complete_turn". When not set, the server applies a default based on the endpoint type.

fsm.prune_failed_steps

Type	Required	Default
boolean or null	No	null

Multi-step mode only. Remove failed steps from history after turn completes. When not set, the server default is true.

fsm.force_to_cache

Type	Required	Default
array[string] or null	No	null

List of tool ID regex patterns to always cache their results regardless of the cache_tool_result setting. When not set, the server default is [].

fsm.max_pllm_steps

Type	Required	Default	Constraints
integer or null	No	null	>= 1

Maximum number of steps allowed per turn.

fsm.max_pllm_failed_steps

Type	Required	Default	Constraints
integer or null	No	null	>= 1

Maximum number of failed steps allowed per turn.

fsm.max_tool_calls_per_step

Type	Required	Default	Constraints
integer or null	No	null	>= 1

Maximum number of tool calls allowed per PLLM attempt. If null, no limit is enforced.

fsm.reduced_grammar_for_rllm_review

Type	Required	Default
"none", "bool", "text", boolean, or null	No	null

Whether to paraphrase RLLM output via reduced grammar before feeding back to planning LLM. "none" denotes passing original review to PLLM; "bool" denotes only keeping boolean flags (success/failure); "text" denotes applying reduced grammar to reviews. true is an alias for "text", false is an alias for "none". When not set, the server default is "text".

fsm.retry_on_policy_violation

Type	Required	Default
boolean or null	No	null

When true, allow planning LLM to retry after policy violation. When not set, the server default is false.

fsm.wrap_tool_result

Type	Required	Default
boolean or null	No	null

Whether to wrap tool results in Ok/Err types.

fsm.detect_tool_errors

Type	Required	Default
string or null	No	null

Whether and how to detect errors in tool results:

• "none": Do not detect tool result errors
• "regex": Use regex patterns to detect common error messages in tool results
• "llm": Use an LLM to analyze tool results and detect potential errors

fsm.detect_tool_error_regex_pattern

Type	Required	Default
string or null	No	null

The regex pattern to use for detecting error messages in tool results when detect_tool_errors is set to "regex".

fsm.detect_tool_error_max_result_length

Type	Required	Default
integer or null	No	null

The maximum length of tool result to consider for error detection. Longer results will be truncated. If null, no limit is enforced.

fsm.strict_tool_result_parsing

Type	Required	Default
boolean or null	No	null

If true, only parse external tool results as JSON when the tool declares an output_schema. When false, always attempt json.loads on tool results.

fsm.tool_result_transform

Type	Required	Default
string or null	No	null

Transform applied to tool results before processing:

• "none": No transformation
• "codex": Apply codex-style transformation to tool results

fsm.n_pllm_plans

Type	Required	Default	Constraints
integer or null	No	null	>= 1

Number of PLLM plans to generate per step. When > 1, multiple plans are generated concurrently and the best one is selected by a selection LLM. Models are assigned round-robin from pllm_candidate_models; when the model list is shorter than n_pllm_plans, extra candidates reuse the same models with different sampling seeds for diversity. When not set, the server default is 1.

fsm.pllm_candidate_models

Type	Required	Default
array[string] or null	No	null

Distinct model names to round-robin across when generating multiple PLLM plans. When empty or null, the request model is used for all candidates.

fsm.pllm_context_pruning

Type	Required	Default
boolean or null	No	null

Whether to hide context variable values from all PLLM steps except the current one. Context vars are cumulative, so older steps contain redundant data. When not set, the server default is true.

---

Prompt Overrides (prompt)

Per-LLM prompt configuration. Each sub-object corresponds to a different LLM in the pipeline.

prompt.pllm

Planning LLM prompt overrides:

Field	Type	Default	Description
flavor	string	null	Prompt template variant to use (e.g., "universal").
version	string	null	Prompt template version. Combined with flavor to load template.
debug_info_level	string	null	Level of detail for debug/execution information in planning LLM prompt: "minimal", "normal", "extra". When not set, the server default is "normal".
clarify_ambiguous_queries	boolean	null	Whether planning LLM is allowed to ask for clarification on ambiguous queries.
context_var_visibility	string	null	The visibility level of context variables in the PLLM prompts: "none", "basic-notext", "basic-executable", "all-executable", "all".
query_inline_roles	array[string]	null	List of roles whose messages will be inlined into the user query: "assistant", "tool", "developer", "system".
query_role_name_overrides	object	null	Overrides for message role names in the inlined user query. For example, {"assistant": "developer"} will change the role of assistant messages to developer.
query_include_tool_calls	boolean	null	Whether to include upstream tool calls in inlined query.
query_include_tool_args	boolean	null	Whether to include arguments of upstream tool calls.
query_include_tool_results	boolean	null	Whether to include results of upstream tool calls.

prompt.rllm

Review LLM prompt overrides:

Field	Type	Default	Description
flavor	string	null	Prompt template variant to use (e.g., "universal").
version	string	null	Prompt template version. Combined with flavor to load template.
debug_info_level	string	null	Level of detail for debug/execution information in RLLM prompt: "minimal", "normal", "extra".

prompt.tllm

Tool-formulating LLM prompt overrides:

Field	Type	Default	Description
flavor	string	null	Prompt template variant to use (e.g., "universal").
version	string	null	Prompt template version. Combined with flavor to load template.
add_tool_description	boolean	null	Whether to include tool descriptions in tool-filtering prompt.
add_tool_input_schema	boolean	null	Whether to include tool input JSON schemas in tool-filtering prompt.

Other LLM Prompts

prompt.grllm, prompt.qllm, prompt.tagllm, prompt.policy_llm, prompt.error_detector_llm all support:

Field	Type	Default	Description
flavor	string	null	Prompt template variant to use (e.g., "universal").
version	string	null	Prompt template version. Combined with flavor to load template.

---

Response Format Overrides (response_format)

response_format.strip_response_content

Type	Required	Default
boolean	No	null

When true, returns only essential result value as plain text, stripping all metadata.

response_format.stream_thoughts

Type	Required	Default
boolean or null	No	null

Whether to stream the model's thinking process in the response.

response_format.include_program

Type	Required	Default
boolean	No	null

Whether to include the generated program in the response.

response_format.include_policy_check_history

Type	Required	Default
boolean	No	null

Whether to include policy check results even when there are no violations.

response_format.include_namespace_snapshot

Type	Required	Default
boolean	No	null

Whether to include snapshot of all variables after program execution.