ZacharyZhang-NY/Kigi-CLI
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

F5: web search/fetch on the Kimi services (kimi-cli parity)

Browse Source 
web_search now speaks the Kimi search service (kimi-cli tools/web/search.py,
wire-verified against api.kimi.com):
- POST {coding_base}/search with {text_query, limit 1-20 (default 5),
  enable_page_crawling, timeout_seconds: 30}, OAuth bearer +
  X-Msh-Tool-Call-Id; results render in kimi-cli's Title/Date/URL/Summary
  schema with result URLs as citations.
- The old implementation called the xAI Responses API with a search model;
  that client is fully replaced and the entire model-based config surface
  is excised root-and-branch: web_search_model config keys/env/CLI plumbing,
  resolve_web_search_sampling_config, toolset web_search SamplerConfig,
  RemoteSettings.web_search_model, default_web_search_model.
- Enablement is now purely structural: the service exists only on the Kimi
  Code subscription channel, so OAuth sessions get Enabled and API-key-only
  sessions get Disabled (tool absent) — per PRD F5.

web_fetch gains the Kimi fetch service as its primary path (kimi-cli
tools/web/fetch.py): POST {coding_base}/fetch with {url}, Accept:
text/markdown, OAuth bearer + X-Msh-Tool-Call-Id; the 200 body is the
extracted markdown (still overflow-budgeted). Any service failure falls
back to the existing local pipeline (SSRF guards, cache, extraction).
The tool gate defaults ON now (kimi-cli always offers FetchURL) and the
egress User-Agent no longer claims grok-agent/x.ai.

Verified end-to-end against the scripted mock service: a headless session
drove web_search (limit/crawling/call-id observed on the wire) then
web_fetch (Accept + call-id observed) to completion.
This commit is contained in:
ZacharyZhang-NY
2026-07-17 19:33:19 -04:00
parent 78cd94a751
commit 74b210535e
34 changed files with 659 additions and 1209 deletions
Download Patch File Download Diff File Expand all files Collapse all files
crates/codegen/kigi-shell/src/agent/config.rs
+9 -149
View File
@@ -490,7 +490,6 @@ pub struct RuntimeResolutionContext<'a> {
pub is_headless: bool,
/// `Some(true)` = CLI explicitly enabled, `None` = defer to config/env/remote.
pub cli_subagents: Option<bool>,
pub cli_web_search_model: Option<&'a str>,
pub cli_session_summary_model: Option<&'a str>,
/// CLI `--experimental-memory` flag. Enables cross-session memory.
pub cli_experimental_memory: bool,
@@ -846,8 +845,6 @@ pub struct ModelsConfig {
#[serde(skip_serializing_if = "Option::is_none")]
pub default_reasoning_effort: Option<ReasoningEffort>,
#[serde(skip_serializing_if = "Option::is_none")]
pub web_search: Option<String>,
#[serde(skip_serializing_if = "Option::is_none")]
pub session_summary: Option<String>,
/// Vision model used to transcribe user-supplied
/// images via a separate endpoint.
@@ -1310,9 +1307,6 @@ pub struct Config {
/// CLI override for reasoning effort.
#[serde(skip)]
pub reasoning_effort_override: Option<ReasoningEffort>,
/// CLI override for the web search model ID.
#[serde(skip)]
pub web_search_model_override: Option<String>,
/// CLI override for the session summary model ID.
#[serde(skip)]
pub session_summary_model_override: Option<String>,
@@ -1446,9 +1440,6 @@ pub struct Config {
/// Enforced requirement pins from `requirements.toml`.
#[serde(skip)]
pub requirements: Requirements,
/// Model ID for web_search.
#[serde(skip)]
pub web_search_model: String,
/// Session title model. Resolved to the compiled default
/// (`default_session_summary_model`) when unset; see `ModelOverrideConfig::resolve`.
#[serde(skip)]
@@ -1662,7 +1653,6 @@ impl Default for Config {
storage_mode: StorageMode::resolve(None, None),
default_model_override: None,
reasoning_effort_override: None,
web_search_model_override: None,
session_summary_model_override: None,
default_yolo_mode: false,
default_auto_mode: false,
@@ -1693,7 +1683,6 @@ impl Default for Config {
auto_wake_enabled: true,
compat_resolved: CompatConfig::default(),
requirements: Requirements::default(),
web_search_model: crate::models::default_web_search_model().to_owned(),
session_summary_model: None,
image_description_model: None,
prompt_suggest_model_pin: crate::config::PromptSuggestModelPin::Unpinned,
@@ -1780,9 +1769,7 @@ impl Config {
if config.client_version.is_none() {
config.client_version = Self::default().client_version;
}
let model_overrides =
crate::config::ModelOverrideConfig::resolve(None, None, raw_config, None);
config.web_search_model = model_overrides.web_search;
let model_overrides = crate::config::ModelOverrideConfig::resolve(None, raw_config, None);
config.session_summary_model = model_overrides.session_summary;
config.image_description_model = model_overrides.image_description;
config.prompt_suggest_model_pin = model_overrides.prompt_suggestion;
@@ -1813,7 +1800,7 @@ impl Config {
/// - respect_gitignore via `ToolsConfig::resolve`
/// - disable_zdr_incompatible_tools via `ToolsConfig::resolve`
/// - managed_mcps_enabled via `ManagedMcpsConfig::resolve`
/// - web_search_model / session_summary_model / image_description_model /
/// - session_summary_model / image_description_model /
/// prompt_suggest_model_pin via `ModelOverrideConfig::resolve`
/// - memory_config via `MemoryConfig::resolve`
/// - disable_web_search (CLI flag ORed with config.toml)
@@ -1824,7 +1811,6 @@ impl Config {
/// `resolve_worktree_type` since it's an agent-level field, not a Config field.
pub fn resolve_runtime_fields(&mut self, ctx: &RuntimeResolutionContext<'_>) {
self.cli_subagents = ctx.cli_subagents;
self.web_search_model_override = ctx.cli_web_search_model.map(|s| s.to_owned());
self.session_summary_model_override = ctx.cli_session_summary_model.map(|s| s.to_owned());
let cli_flag = ctx.cli_subagents.unwrap_or(false);
self.resolve_subagents(cli_flag, ctx.raw_config, ctx.cwd);
@@ -1843,12 +1829,10 @@ impl Config {
self.managed_mcps_enabled = mcps.enabled;
self.managed_mcp_gateway_tools_enabled = mcps.gateway_tools_enabled;
let models = crate::config::ModelOverrideConfig::resolve(
ctx.cli_web_search_model,
ctx.cli_session_summary_model,
ctx.raw_config,
ctx.remote_settings,
);
self.web_search_model = models.web_search;
self.session_summary_model = models.session_summary;
self.image_description_model = models.image_description;
self.prompt_suggest_model_pin = models.prompt_suggestion;
@@ -1889,7 +1873,6 @@ impl Config {
cwd: Option<&std::path::Path>,
) {
let remote_settings = self.remote_settings.clone();
let cli_web_search_model = self.web_search_model_override.clone();
let cli_session_summary_model = self.session_summary_model_override.clone();
let laziness_debug_log = self.laziness_debug_log.clone();
let ctx = RuntimeResolutionContext {
@@ -1898,7 +1881,6 @@ impl Config {
cwd,
is_headless: self.mode == AgentMode::Headless,
cli_subagents: self.cli_subagents,
cli_web_search_model: cli_web_search_model.as_deref(),
cli_session_summary_model: cli_session_summary_model.as_deref(),
cli_experimental_memory: self.cli_experimental_memory,
cli_no_memory: self.cli_no_memory,
@@ -2021,6 +2003,9 @@ impl Config {
.feature_flag(ff)
.resolve()
}
/// `web_fetch` tool gate; default ON (kimi-cli parity: `FetchURL` is
/// always offered). `KIGI_WEB_FETCH=0` / `[features] web_fetch = false`
/// are the kill switches.
pub(crate) fn resolve_web_fetch(&self) -> Resolved<bool> {
let ff = self
.remote_settings
@@ -2030,6 +2015,7 @@ impl Config {
.requirement(self.requirements.web_fetch.pinned())
.config(self.features.web_fetch)
.feature_flag(ff)
.default(true)
.resolve()
}
/// `ask_user_question` tool gate; default ON. remote settings
@@ -4387,85 +4373,6 @@ pub fn resolve_model_to_sampling_config(
alpha_test_key,
))
}
fn resolve_hidden_default_web_search_sampling_config(
model_id: &str,
session_key: Option<&str>,
alpha_test_key: Option<String>,
endpoints: &EndpointsConfig,
) -> SamplerConfig {
let entry = ModelEntry {
info: ModelInfo {
id: None,
model: model_id.to_owned(),
base_url: endpoints.resolve_inference_base_url(),
name: None,
description: None,
max_completion_tokens: None,
temperature: None,
top_p: None,
api_backend: ApiBackend::Responses,
auth_scheme: Default::default(),
extra_headers: IndexMap::new(),
context_window: NonZeroU64::new(200_000).unwrap(),
auto_compact_threshold_percent: None,
system_prompt_label: None,
use_concise: false,
agent_type: default_agent_type(),
inference_idle_timeout_secs: None,
max_retries: None,
hidden: true,
user_selectable: true,
supported_in_api: true,
reasoning_effort: None,
supports_reasoning_effort: false,
reasoning_efforts: Vec::new(),
capabilities: Vec::new(),
supports_backend_search: false,
compactions_remaining: None,
compaction_at_tokens: None,
show_model_fingerprint: false,
stream_tool_calls: None,
laziness_detector: LazinessDetectorPerModelConfig::default(),
},
api_key: None,
env_key: None,
api_base_url: None,
};
let credentials = resolve_credentials(&entry, session_key);
sampling_config_for_model(&entry, credentials, alpha_test_key)
}
pub fn resolve_web_search_sampling_config(
model_id: &str,
models: &IndexMap<String, ModelEntry>,
session_key: Option<&str>,
alpha_test_key: Option<String>,
endpoints: &EndpointsConfig,
) -> Option<SamplerConfig> {
let resolved = if let Some(entry) = find_model_by_id(models, model_id).cloned() {
let credentials = resolve_credentials(&entry, session_key);
Some(sampling_config_for_model(
&entry,
credentials,
alpha_test_key,
))
} else if model_id == crate::models::default_web_search_model() {
Some(resolve_hidden_default_web_search_sampling_config(
model_id,
session_key,
alpha_test_key,
endpoints,
))
} else {
None
};
if resolved.is_none() {
tracing::warn!(
web_search_model = % model_id,
"configured web_search model not found; disabling web search"
);
}
resolved.map(crate::tools::config::web_search_sampling_config)
}
pub fn to_acp_model_info(
models: &IndexMap<String, ModelEntry>,
) -> IndexMap<acp::ModelId, acp::ModelInfo> {
@@ -4803,7 +4710,6 @@ reasoning_effort = "low"
cwd: None,
is_headless: false,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -4832,7 +4738,6 @@ reasoning_effort = "low"
cwd: None,
is_headless: true,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -4855,14 +4760,9 @@ reasoning_effort = "low"
assert!(cfg.disable_web_search);
}
#[test]
fn new_from_toml_cfg_restores_web_search_and_session_summary_models() {
fn new_from_toml_cfg_restores_session_summary_and_image_description_models() {
let empty: toml::Value = toml::Value::Table(toml::map::Map::new());
let cfg = Config::new_from_toml_cfg(&empty).expect("empty config should parse");
assert_eq!(
cfg.web_search_model,
crate::models::default_web_search_model(),
"empty config should produce the compiled-in default web_search model"
);
assert_eq!(
cfg.session_summary_model,
Some(crate::models::default_session_summary_model().to_owned()),
@@ -4876,14 +4776,12 @@ reasoning_effort = "low"
let with_overrides: toml::Value = toml::from_str(
r#"
[models]
web_search = "custom-ws-model"
session_summary = "custom-ss-model"
image_description = "custom-id-model"
"#,
)
.unwrap();
let cfg2 = Config::new_from_toml_cfg(&with_overrides).expect("config should parse");
assert_eq!(cfg2.web_search_model, "custom-ws-model");
assert_eq!(
cfg2.session_summary_model,
Some("custom-ss-model".to_owned())
@@ -4894,26 +4792,6 @@ reasoning_effort = "low"
);
}
#[test]
fn hidden_default_web_search_resolution_is_explicit_and_responses_only() {
let endpoints = EndpointsConfig::default();
let resolved = resolve_web_search_sampling_config(
crate::models::default_web_search_model(),
&IndexMap::new(),
Some("session-token"),
None,
&endpoints,
)
.expect("hidden default web search model should resolve");
assert_eq!(resolved.model, crate::models::default_web_search_model());
assert_eq!(resolved.base_url, endpoints.proxy_url());
assert_eq!(resolved.api_backend, ApiBackend::Responses);
assert_eq!(
resolved.api_key.as_deref(),
Some("session-token"),
"hidden default should still use normal credential resolution"
);
}
#[test]
fn finalize_image_describe_sampler_none_uses_active_session_model_not_forced_helper() {
let active = SamplerConfig {
model: "composer-session-model".into(),
@@ -6671,14 +6549,12 @@ reasoning_effort = "low"
r#"
[models]
default = "my-enterprise-model"
web_search = "enterprise-search"
session_summary = "title-model"
"#,
)
.unwrap();
let cfg = Config::new_from_toml_cfg(&raw).expect("config should parse");
assert_eq!(cfg.models.default.as_deref(), Some("my-enterprise-model"));
assert_eq!(cfg.models.web_search.as_deref(), Some("enterprise-search"));
assert_eq!(cfg.models.session_summary.as_deref(), Some("title-model"));
}
#[test]
@@ -8903,7 +8779,6 @@ agent_type = "cursor"
unsafe {
std::env::remove_var("KIGI_SUBAGENTS");
std::env::remove_var("KIGI_RESPECT_GITIGNORE");
std::env::remove_var("KIGI_WEB_SEARCH_MODEL");
std::env::remove_var("KIGI_SESSION_SUMMARY_MODEL");
std::env::remove_var("KIGI_CURSOR_SKILLS_ENABLED");
std::env::remove_var("KIGI_CURSOR_RULES_ENABLED");
@@ -9189,7 +9064,6 @@ hooks = true
cwd: None,
is_headless: false,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9215,7 +9089,6 @@ hooks = true
cwd: None,
is_headless: false,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9228,10 +9101,6 @@ hooks = true
assert!(!cfg.respect_gitignore);
assert!(cfg.managed_mcps_enabled);
assert!(!cfg.managed_mcp_gateway_tools_enabled);
assert_eq!(
cfg.web_search_model,
crate::models::default_web_search_model()
);
assert_eq!(
cfg.session_summary_model,
Some(crate::models::default_session_summary_model().to_owned())
@@ -9251,7 +9120,6 @@ hooks = true
cwd: None,
is_headless: true,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9283,7 +9151,6 @@ hooks = true
cwd: None,
is_headless: false,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9306,7 +9173,6 @@ hooks = true
cwd: None,
is_headless: false,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9329,7 +9195,6 @@ hooks = true
cwd: None,
is_headless: false,
cli_subagents: Some(true),
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9353,7 +9218,6 @@ hooks = true
cwd: None,
is_headless: false,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9377,7 +9241,6 @@ hooks = true
cwd: None,
is_headless: false,
cli_subagents: None,
cli_web_search_model: Some("custom-ws"),
cli_session_summary_model: Some("custom-ss"),
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9386,7 +9249,6 @@ hooks = true
laziness_debug_log: None,
storage_mode: None,
});
assert_eq!(cfg.web_search_model, "custom-ws");
assert_eq!(cfg.session_summary_model, Some("custom-ss".to_owned()));
}
#[test]
@@ -9405,7 +9267,6 @@ hooks = true
cwd: None,
is_headless: false,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9428,7 +9289,6 @@ hooks = true
cwd: None,
is_headless: false,
cli_subagents: None,
cli_web_search_model: None,
cli_session_summary_model: None,
cli_experimental_memory: false,
cli_no_memory: false,
@@ -9441,12 +9301,12 @@ hooks = true
let first_subagents = cfg.subagents_enabled;
let first_gitignore = cfg.respect_gitignore;
let first_mcps = cfg.managed_mcps_enabled;
let first_ws = cfg.web_search_model.clone();
let first_ss = cfg.session_summary_model.clone();
cfg.resolve_runtime_fields(&ctx);
assert_eq!(cfg.subagents_enabled, first_subagents);
assert_eq!(cfg.respect_gitignore, first_gitignore);
assert_eq!(cfg.managed_mcps_enabled, first_mcps);
assert_eq!(cfg.web_search_model, first_ws);
assert_eq!(cfg.session_summary_model, first_ss);
}
#[test]
fn version_overrides_apply_into_typed_config() {
crates/codegen/kigi-shell/src/agent/mvp_agent/agent_ops.rs
+44 -18
View File
@@ -834,20 +834,29 @@ impl MvpAgent {
tier_restricted: false,
}
}
pub(super) fn prepare_web_search_sampling_config(&self) -> Option<SamplingConfig> {
let model_id = self.cfg.borrow().web_search_model.clone();
let models = self.models_manager.models();
let session = self.current_or_buffered_auth();
let alpha_test_key = self.cfg.borrow().endpoints.alpha_test_key.clone();
let client_version = self.cfg.borrow().client_version.clone();
let mut cfg = config::resolve_web_search_sampling_config(
&model_id,
&models,
session.as_ref().map(|a| a.key.as_str()),
alpha_test_key.clone(),
&self.cfg.borrow().endpoints,
)?;
Some(cfg)
/// Web search config (PRD F5). The Kimi search service exists only on
/// the Kimi Code subscription channel (`POST {coding_base}/search`,
/// kimi-cli `auth/platforms.py`), so this is `Enabled` only for OAuth
/// sessions — API-key-only sessions get `Disabled` and the tool is
/// absent. The live token is refreshed per request via the session
/// api-key provider; the config key is just the bootstrap value.
pub(super) fn prepare_web_search_config(
&self,
) -> kigi_tools::implementations::WebSearchConfig {
use kigi_tools::implementations::WebSearchConfig;
if self.cfg.borrow().disable_web_search {
return WebSearchConfig::Disabled;
}
let Some(auth) = self.current_or_buffered_auth().filter(|a| a.is_session_auth()) else {
tracing::info!("web_search disabled: no Kimi Code OAuth session");
return WebSearchConfig::Disabled;
};
let base = self.cfg.borrow().endpoints.proxy_url();
WebSearchConfig::Enabled {
search_url: format!("{}/search", base.trim_end_matches('/')),
api_key: auth.key,
extra_headers: indexmap::IndexMap::new(),
}
}
/// Returns `Err` with a user-facing message on invalid config; the caller at
/// the process boundary prints it and exits.
@@ -867,11 +876,14 @@ impl MvpAgent {
/// Prepare the web fetch configuration based on feature flags.
///
/// Enabled gate: `disable_web_search` kill-switch > `KIGI_WEB_FETCH` env >
/// remote settings `web_fetch_enabled` > default (false).
/// remote settings `web_fetch_enabled` > default ON (kimi-cli parity:
/// `FetchURL` is always offered).
///
/// Params resolution (TOML > env > remote settings > default):
/// - `proxy_endpoint`: `[toolset.web_fetch] proxy_endpoint` > `KIGI_WEB_FETCH_PROXY` > remote settings > None
/// - `allowed_domains`: `[toolset.web_fetch] allowed_domains` > remote settings > built-in defaults
/// - `service_url`: TOML/env dev override > `{coding_base}/fetch` on
/// OAuth sessions (PRD F5) > None (local pipeline only)
pub(super) fn prepare_web_fetch_config(
&self,
) -> kigi_tools::implementations::grok_build::web_fetch::WebFetchConfig {
@@ -886,7 +898,7 @@ impl MvpAgent {
return WebFetchConfig::Disabled;
}
let context_window = Some(self.sampling_config.borrow().context_window);
let params = cfg
let mut params = cfg
.toolset
.web_fetch
.resolve_params(
@@ -898,6 +910,20 @@ impl MvpAgent {
tracing::info!("web_fetch disabled: allowed_domains is explicitly empty");
return WebFetchConfig::Disabled;
}
// PRD F5: the Kimi fetch service exists only on the OAuth channel.
// TOML/env may pin their own service_url (dev override); otherwise
// OAuth sessions get `{coding_base}/fetch` and API-key sessions
// stay local-only.
if params.service_url.is_none()
&& self
.current_or_buffered_auth()
.is_some_and(|a| a.is_session_auth())
{
params.service_url = Some(format!(
"{}/fetch",
cfg.endpoints.proxy_url().trim_end_matches('/')
));
}
WebFetchConfig::Enabled { params }
}
/// Construct from pre-built components. Use when the caller needs the
@@ -2231,7 +2257,7 @@ impl MvpAgent {
.find(|entry| entry.info.model == sampling_config.model)
.and_then(|entry| entry.info.max_retries);
let origin_client = self.origin_client_info_from_meta(init.meta.as_ref());
let web_search_sampling_config = self.prepare_web_search_sampling_config();
let web_search_config = self.prepare_web_search_config();
let image_gen_config = self.prepare_image_gen_config();
let video_gen_config = self.prepare_video_gen_config();
let app_builder_deployer_config = self.prepare_app_builder_deployer_config();
@@ -2454,7 +2480,7 @@ impl MvpAgent {
origin_client.as_ref().map(|o| o.product.clone()),
inference_idle_timeout_secs,
model_max_retries,
web_search_sampling_config,
web_search_config,
web_fetch_config,
image_gen_config,
video_gen_config,
crates/codegen/kigi-shell/src/agent/mvp_agent/subagent_coordinator.rs
+1 -1
View File
@@ -427,7 +427,7 @@ impl MvpAgent {
terminal,
session_env,
memory_config: self.memory_config.clone(),
web_search_sampling_config: self.prepare_web_search_sampling_config(),
web_search_config: self.prepare_web_search_config(),
web_fetch_config: self.prepare_web_fetch_config(),
image_gen_config: self.prepare_image_gen_config(),
video_gen_config: self.prepare_video_gen_config(),
crates/codegen/kigi-shell/src/agent/subagent/handle_request.rs
+1 -1
View File
@@ -1067,7 +1067,7 @@ pub(crate) async fn handle_subagent_request(
None,
ctx.inference_idle_timeout_secs,
None,
ctx.web_search_sampling_config.clone(),
ctx.web_search_config.clone(),
ctx.web_fetch_config.clone(),
ctx.image_gen_config.clone(),
ctx.video_gen_config.clone(),
crates/codegen/kigi-shell/src/agent/subagent/mod.rs
+1 -1
View File
@@ -206,7 +206,7 @@ pub(crate) struct SubagentSpawnContext {
/// cross-session memory store.
pub memory_config: Option<crate::config::MemoryConfig>,
/// Resolved sampling config for web_search.
pub web_search_sampling_config: Option<kigi_sampler::SamplerConfig>,
pub web_search_config: kigi_tools::implementations::WebSearchConfig,
/// Resolved config for web fetch.
pub web_fetch_config: kigi_tools::implementations::grok_build::web_fetch::WebFetchConfig,
/// Image generation config (parent-inherited).