Skip to main content

harn_vm/llm/capabilities/
model.rs

1//! Capability DTOs and the wire-dialect model.
2//!
3//! Pure data types: the on-disk [`CapabilitiesFile`] schema, per-provider
4//! [`ProviderDefaults`], the resolved [`Capabilities`] struct callers consume,
5//! and the [`WireDialect`] enum that types a route's message wire format. The
6//! `ProviderRule` matrix row and the resolution engine that turns these DTOs
7//! into a `Capabilities` live in `super::rule`.
8
9use std::collections::BTreeMap;
10
11use serde::Deserialize;
12
13use super::rule::ProviderRule;
14
15/// Parsed on-disk capabilities schema. Public so harn-cli can
16/// construct one directly when wiring harn.toml overrides.
17#[derive(Debug, Clone, Deserialize, Default)]
18pub struct CapabilitiesFile {
19    /// Per-provider ordered rule lists. The first matching rule wins; a
20    /// matching rule with `extends = true` contributes only the fields it
21    /// sets and lets resolution continue to later matching rules (see
22    /// [`ProviderRule::extends`]).
23    #[serde(default)]
24    pub provider: BTreeMap<String, Vec<ProviderRule>>,
25    /// Per-provider defaults applied to every matching row and to
26    /// provider/model pairs that have no model-specific row. This keeps
27    /// transport-shape facts in data without repeating them on every
28    /// generation-specific capability row.
29    #[serde(default)]
30    pub provider_defaults: BTreeMap<String, ProviderDefaults>,
31    /// Sibling → canonical family mapping. Providers with no rule of
32    /// their own fall through to the named family (recursively).
33    #[serde(default)]
34    pub provider_family: BTreeMap<String, String>,
35    /// Per-provider adaptive rate/concurrency governor limits, keyed by
36    /// provider id. Consumed by `crate::llm::rate_governor` when the
37    /// `llm.rate_governor` flag is enabled, so provider limits stay catalog
38    /// data instead of call-site branches.
39    #[serde(default)]
40    pub provider_limits: BTreeMap<String, ProviderLimits>,
41}
42
43/// Adaptive-governor limits for one provider. Every field is optional so a
44/// catalog fragment can pin just the axes it knows; unset axes fall back to the
45/// governor's conservative built-in defaults.
46#[derive(Debug, Clone, Deserialize, Default, PartialEq)]
47pub struct ProviderLimits {
48    /// Ceiling the AIMD concurrency limiter additively climbs toward on
49    /// sustained success.
50    #[serde(default)]
51    pub max_concurrency: Option<u32>,
52    /// Floor the AIMD limiter multiplicatively decreases toward on a throttle
53    /// signal.
54    #[serde(default)]
55    pub min_concurrency: Option<u32>,
56    /// Requests-per-minute token bucket. `None` disables the RPM bucket.
57    #[serde(default)]
58    pub rpm: Option<u32>,
59    /// Tokens-per-minute token bucket, charged by estimated input + output
60    /// tokens. `None` disables the TPM bucket.
61    #[serde(default)]
62    pub tpm: Option<u64>,
63    /// Whether the AIMD adaptive concurrency loop is active. When `false`, the
64    /// concurrency limit is pinned at `max_concurrency`.
65    #[serde(default)]
66    pub adaptive: Option<bool>,
67    /// Circuit-breaker / backoff parameters. Absent means built-in defaults.
68    #[serde(default)]
69    pub backoff: Option<GovernorBackoff>,
70}
71
72/// Exponential-backoff-with-jitter parameters for the governor circuit breaker.
73/// Provider `Retry-After` values always take precedence over the computed
74/// window.
75#[derive(Debug, Clone, Deserialize, PartialEq)]
76pub struct GovernorBackoff {
77    /// First OPEN window, in milliseconds.
78    #[serde(default)]
79    pub base_ms: Option<u64>,
80    /// Ceiling for the OPEN window, in milliseconds.
81    #[serde(default)]
82    pub max_ms: Option<u64>,
83    /// Growth factor applied per consecutive OPEN cycle.
84    #[serde(default)]
85    pub multiplier: Option<f64>,
86    /// Full-jitter toggle.
87    #[serde(default)]
88    pub jitter: Option<bool>,
89}
90
91/// Provider-wide default fields merged into matching rules.
92#[derive(Debug, Clone, Deserialize, Default)]
93pub struct ProviderDefaults {
94    /// Message/request/response wire format used by shared helpers.
95    /// Known values are `openai`, `anthropic`, `gemini`, and `ollama`.
96    #[serde(default)]
97    pub message_wire_format: Option<String>,
98    /// Which synchronous endpoint family the dialect's routes dispatch to.
99    /// See [`LiveEndpointFamily`].
100    #[serde(default)]
101    pub live_endpoint_family: Option<LiveEndpointFamily>,
102    /// Native tool definition wire shape. Known values are `openai`
103    /// and `anthropic`.
104    #[serde(default)]
105    pub native_tool_wire_format: Option<String>,
106    /// Whether image content blocks may reference remote URLs.
107    #[serde(default)]
108    pub image_url_input_supported: Option<bool>,
109    /// File-upload transport used by `std/files.upload`. Known values
110    /// are `anthropic` and `gemini`.
111    #[serde(default)]
112    pub file_upload_wire_format: Option<String>,
113    /// Provider-specific reasoning request shape for OpenAI-compatible
114    /// transports. Known values are `openrouter` and `enabled`.
115    #[serde(default)]
116    pub reasoning_wire_format: Option<String>,
117    #[serde(default)]
118    pub files_api_supported: Option<bool>,
119    #[serde(default)]
120    pub batch_api: Option<bool>,
121    #[serde(default)]
122    pub batch_wire_format: Option<String>,
123    #[serde(default)]
124    pub batch_input_mode: Option<String>,
125    #[serde(default)]
126    pub batch_discount_percent: Option<u32>,
127    #[serde(default)]
128    pub batch_turnaround_hours: Option<u32>,
129    #[serde(default)]
130    pub batch_max_requests: Option<u64>,
131    #[serde(default)]
132    pub batch_max_input_bytes: Option<u64>,
133    #[serde(default)]
134    pub batch_result_retention_days: Option<u32>,
135    #[serde(default)]
136    pub batch_result_ordering: Option<String>,
137    #[serde(default)]
138    pub batch_partial_failure: Option<String>,
139    #[serde(default)]
140    pub batch_cancellation: Option<String>,
141    #[serde(default)]
142    pub batch_security_notes: Option<Vec<String>>,
143    #[serde(default)]
144    pub batch_operational_notes: Option<Vec<String>>,
145    /// Regions where the provider/model batch route is explicitly supported.
146    /// Empty means the capability is not region-scoped.
147    #[serde(default)]
148    pub batch_regions: Option<Vec<String>>,
149    /// Explicit prompt-cache TTL values this provider can honor on request.
150    /// Empty means the route may cache, but Harn has no explicit TTL knob for
151    /// it. Known values today: `5m`, `1h`.
152    #[serde(default)]
153    pub prompt_cache_ttls: Option<Vec<String>>,
154    /// Shortest prefix this provider will actually cache, in tokens. See
155    /// [`crate::llm::capabilities::rule::ProviderRule::prompt_cache_min_prefix_tokens`].
156    #[serde(default)]
157    pub prompt_cache_min_prefix_tokens: Option<u32>,
158    #[serde(default)]
159    pub seed_supported: Option<bool>,
160    #[serde(default)]
161    pub top_k_supported: Option<bool>,
162    #[serde(default)]
163    pub temperature_supported: Option<bool>,
164    #[serde(default)]
165    pub top_p_supported: Option<bool>,
166    #[serde(default)]
167    pub frequency_penalty_supported: Option<bool>,
168    #[serde(default)]
169    pub presence_penalty_supported: Option<bool>,
170    #[serde(default)]
171    pub stop_supported: Option<bool>,
172}
173
174/// Copies `src` into `dst` when `src` is set (last-writer-wins overlay).
175pub(super) fn overlay_opt<T: Clone>(dst: &mut Option<T>, src: &Option<T>) {
176    if src.is_some() {
177        dst.clone_from(src);
178    }
179}
180
181/// Copies `src` into `dst` only when `dst` is still unset (fill-the-gaps).
182pub(super) fn fill_opt<T: Clone>(dst: &mut Option<T>, src: &Option<T>) {
183    if dst.is_none() {
184        dst.clone_from(src);
185    }
186}
187
188/// Visits every `ProviderDefaults` field once, applying `$op` (`overlay_opt`
189/// or `fill_opt`) to each `(dst, src)` pair. The field roster lives here only;
190/// `overlay`/`fill_missing_from` differ solely in the merge rule they pass.
191macro_rules! merge_provider_defaults {
192    ($dst:expr, $src:expr, $op:path) => {{
193        $op(&mut $dst.message_wire_format, &$src.message_wire_format);
194        $op(&mut $dst.live_endpoint_family, &$src.live_endpoint_family);
195        $op(
196            &mut $dst.native_tool_wire_format,
197            &$src.native_tool_wire_format,
198        );
199        $op(
200            &mut $dst.image_url_input_supported,
201            &$src.image_url_input_supported,
202        );
203        $op(
204            &mut $dst.file_upload_wire_format,
205            &$src.file_upload_wire_format,
206        );
207        $op(&mut $dst.reasoning_wire_format, &$src.reasoning_wire_format);
208        $op(&mut $dst.files_api_supported, &$src.files_api_supported);
209        $op(&mut $dst.batch_api, &$src.batch_api);
210        $op(&mut $dst.batch_wire_format, &$src.batch_wire_format);
211        $op(&mut $dst.batch_input_mode, &$src.batch_input_mode);
212        $op(
213            &mut $dst.batch_discount_percent,
214            &$src.batch_discount_percent,
215        );
216        $op(
217            &mut $dst.batch_turnaround_hours,
218            &$src.batch_turnaround_hours,
219        );
220        $op(&mut $dst.batch_max_requests, &$src.batch_max_requests);
221        $op(&mut $dst.batch_max_input_bytes, &$src.batch_max_input_bytes);
222        $op(
223            &mut $dst.batch_result_retention_days,
224            &$src.batch_result_retention_days,
225        );
226        $op(&mut $dst.batch_result_ordering, &$src.batch_result_ordering);
227        $op(&mut $dst.batch_partial_failure, &$src.batch_partial_failure);
228        $op(&mut $dst.batch_cancellation, &$src.batch_cancellation);
229        $op(&mut $dst.batch_security_notes, &$src.batch_security_notes);
230        $op(
231            &mut $dst.batch_operational_notes,
232            &$src.batch_operational_notes,
233        );
234        $op(&mut $dst.batch_regions, &$src.batch_regions);
235        $op(&mut $dst.prompt_cache_ttls, &$src.prompt_cache_ttls);
236        $op(
237            &mut $dst.prompt_cache_min_prefix_tokens,
238            &$src.prompt_cache_min_prefix_tokens,
239        );
240        $op(&mut $dst.seed_supported, &$src.seed_supported);
241        $op(&mut $dst.top_k_supported, &$src.top_k_supported);
242        $op(&mut $dst.temperature_supported, &$src.temperature_supported);
243        $op(&mut $dst.top_p_supported, &$src.top_p_supported);
244        $op(
245            &mut $dst.frequency_penalty_supported,
246            &$src.frequency_penalty_supported,
247        );
248        $op(
249            &mut $dst.presence_penalty_supported,
250            &$src.presence_penalty_supported,
251        );
252        $op(&mut $dst.stop_supported, &$src.stop_supported);
253    }};
254}
255
256impl ProviderDefaults {
257    pub(super) fn overlay(&mut self, other: &ProviderDefaults) {
258        merge_provider_defaults!(self, other, overlay_opt);
259    }
260
261    pub(super) fn fill_missing_from(&mut self, other: &ProviderDefaults) {
262        merge_provider_defaults!(self, other, fill_opt);
263    }
264
265    pub(super) fn has_any_field(&self) -> bool {
266        self.message_wire_format.is_some()
267            || self.live_endpoint_family.is_some()
268            || self.native_tool_wire_format.is_some()
269            || self.image_url_input_supported.is_some()
270            || self.file_upload_wire_format.is_some()
271            || self.reasoning_wire_format.is_some()
272            || self.files_api_supported.is_some()
273            || self.batch_api.is_some()
274            || self.batch_wire_format.is_some()
275            || self.batch_input_mode.is_some()
276            || self.batch_discount_percent.is_some()
277            || self.batch_turnaround_hours.is_some()
278            || self.batch_max_requests.is_some()
279            || self.batch_max_input_bytes.is_some()
280            || self.batch_result_retention_days.is_some()
281            || self.batch_result_ordering.is_some()
282            || self.batch_partial_failure.is_some()
283            || self.batch_cancellation.is_some()
284            || self.batch_security_notes.is_some()
285            || self.batch_operational_notes.is_some()
286            || self.batch_regions.is_some()
287            || self.prompt_cache_ttls.is_some()
288            || self.prompt_cache_min_prefix_tokens.is_some()
289            || self.seed_supported.is_some()
290            || self.top_k_supported.is_some()
291            || self.temperature_supported.is_some()
292            || self.top_p_supported.is_some()
293            || self.frequency_penalty_supported.is_some()
294            || self.presence_penalty_supported.is_some()
295            || self.stop_supported.is_some()
296    }
297}
298
299/// The message/request/response wire dialect a route speaks.
300///
301/// This is the single typed representation of what used to be encoded two
302/// different, drift-prone ways: the stringly `Capabilities.message_wire_format`
303/// field (compared against `"anthropic"`/`"gemini"`/`"ollama"` literals at a
304/// dozen call sites) and the `(is_anthropic_style, is_ollama)` boolean pair
305/// threaded independently through the transport/response layers. A closed enum
306/// makes an unhandled or mistyped dialect a compile error and removes the
307/// boolean-blindness where two `bool`s could silently disagree.
308#[derive(Debug, Clone, Copy, PartialEq, Eq)]
309pub enum WireDialect {
310    /// Anthropic native Messages API (`/v1/messages`). The only dialect that
311    /// surfaces Claude's extended-thinking stream. `message_wire_format =
312    /// "anthropic"`.
313    Anthropic,
314    /// OpenAI-compatible Chat Completions (`/v1/chat/completions`). The default
315    /// for hosted/openai-shape routes. `message_wire_format = "openai"`.
316    OpenAiCompat,
317    /// Ollama native `/api/chat`. `message_wire_format = "ollama"`.
318    Ollama,
319    /// Google Gemini `generateContent`. `message_wire_format = "gemini"`.
320    Gemini,
321}
322
323impl WireDialect {
324    /// Parse the catalog's `message_wire_format` string. Unrecognized values
325    /// (including the explicit `"openai"`) resolve to [`WireDialect::OpenAiCompat`],
326    /// exactly matching the pre-cutover behavior where every
327    /// `== "anthropic"/"gemini"/"ollama"` check fell through to the
328    /// OpenAI-compatible path.
329    pub fn from_message_wire_format(value: &str) -> WireDialect {
330        match value {
331            "anthropic" => WireDialect::Anthropic,
332            "ollama" => WireDialect::Ollama,
333            "gemini" => WireDialect::Gemini,
334            _ => WireDialect::OpenAiCompat,
335        }
336    }
337
338    /// The canonical `message_wire_format` string for display and round-trip.
339    pub fn as_str(self) -> &'static str {
340        match self {
341            WireDialect::Anthropic => "anthropic",
342            WireDialect::OpenAiCompat => "openai",
343            WireDialect::Ollama => "ollama",
344            WireDialect::Gemini => "gemini",
345        }
346    }
347
348    /// Whether this route speaks Anthropic's native Messages shape.
349    pub fn is_anthropic(self) -> bool {
350        matches!(self, WireDialect::Anthropic)
351    }
352
353    /// Whether this route speaks Ollama's native `/api/chat` shape.
354    pub fn is_ollama(self) -> bool {
355        matches!(self, WireDialect::Ollama)
356    }
357
358    /// Whether this route speaks Google Gemini's `generateContent` shape.
359    pub fn is_gemini(self) -> bool {
360        matches!(self, WireDialect::Gemini)
361    }
362}
363
364/// Which synchronous ("live") endpoint family a route dispatches to.
365///
366/// [`WireDialect`] answers "what does the JSON look like"; most dialects have
367/// exactly one live endpoint, so the two questions collapse. Gemini is the
368/// exception: Google serves the same models over two incompatible synchronous
369/// shapes, so a second axis is needed to say *which* one a route uses.
370///
371/// This axis is deliberately independent of `batch_wire_format`. Gemini Batch
372/// only accepts `generateContent`-shaped bodies, so a route can select
373/// [`LiveEndpointFamily::GeminiInteractions`] for its live traffic while its
374/// batch submissions stay `batch_wire_format = "gemini"`.
375///
376/// It is also deliberately independent of the advisory `recommended_endpoint`
377/// string: this is the value the runtime actually switches on, so it is typed
378/// and an unknown value in a capability source fails the load instead of
379/// silently falling back to the legacy transport.
380#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
381#[serde(rename_all = "snake_case")]
382pub enum LiveEndpointFamily {
383    /// Google Gemini `POST /v1beta/models/{model}:generateContent`. The legacy
384    /// path, still the only shape Gemini Batch accepts and the only one Vertex
385    /// delegates to.
386    GeminiGenerateContent,
387    /// Google Gemini `POST /v1beta/interactions` (GA June 2026). Adds
388    /// server-side conversation state (`previous_interaction_id`), observable
389    /// execution steps, streaming, and background execution.
390    GeminiInteractions,
391}
392
393impl LiveEndpointFamily {
394    /// The canonical capability-source string for display and round-trip.
395    pub const fn as_str(self) -> &'static str {
396        match self {
397            Self::GeminiGenerateContent => "gemini_generate_content",
398            Self::GeminiInteractions => "gemini_interactions",
399        }
400    }
401
402    /// Whether this endpoint family keeps conversation state on the provider,
403    /// making the `previous_response_id` / `store` / `background` options
404    /// meaningful on the route.
405    pub const fn is_stateful(self) -> bool {
406        matches!(self, Self::GeminiInteractions)
407    }
408
409    /// Whether this route dispatches over Gemini's Interactions transport.
410    pub const fn is_gemini_interactions(self) -> bool {
411        matches!(self, Self::GeminiInteractions)
412    }
413}
414
415/// How the neutral `computer` tool projects onto a route's native computer-use
416/// surface (the `computer_use_style` capability). A typed enum rather than a raw
417/// string so an unknown value in a capability source is a load-time
418/// deserialize error instead of a silently-disabled computer tool.
419#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
420#[serde(rename_all = "snake_case")]
421pub enum ComputerUseStyle {
422    /// Anthropic `computer_20251124` native tool.
423    NativeAnthropic,
424    /// OpenAI Responses `computer` native tool.
425    NativeOpenai,
426    /// Accessibility / set-of-marks grounding over the universal function tool.
427    Grounded,
428    /// The plain function-schema `computer` tool (the universal default).
429    Function,
430}
431
432/// Screenshot downscaling policy applied before an image reaches the model (the
433/// `screenshot_scaling` capability). Typed for the same reason as
434/// [`ComputerUseStyle`] — an unknown value fails the capability load loudly.
435#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
436#[serde(rename_all = "snake_case")]
437pub enum ScreenshotScaling {
438    /// Fit within Anthropic's XGA (1024x768), preserving aspect ratio.
439    Xga,
440    /// Send the capture at its native resolution (OpenAI et al.).
441    Original,
442}
443
444/// How a route carries a `system`/`developer`-role message that appears at a
445/// non-leading position in the conversation — the operator-instruction channel
446/// (OpenAI developer messages, Anthropic mid-conversation system messages).
447/// Leading system content is always the system prompt; this governs only the
448/// *interleaved* case. Typed so an unknown value in a capability
449/// source fails the load loudly. `None` on [`Capabilities`] derives a safe
450/// default from the wire dialect — see
451/// [`resolve_system_message_placement`](super::resolve_system_message_placement).
452#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
453#[serde(rename_all = "snake_case")]
454pub enum SystemMessagePlacement {
455    /// The wire API accepts a `system`/`developer` message verbatim at any
456    /// position (OpenAI Chat Completions & Responses, Ollama). Pass through.
457    Inline,
458    /// The API accepts an interleaved `system`-role directive but enforces
459    /// placement rules (Anthropic Opus 4.8: one message must follow a `user`
460    /// turn — or an assistant turn ending in `server_tool_use` — and be the
461    /// last message or be followed by an `assistant` turn; never
462    /// `messages[0]`). A valid consecutive section is merged and kept native;
463    /// anything else folds. `developer` collapses to `system` (Anthropic has no
464    /// developer role).
465    NativeDirective,
466    /// The API has no positional system channel (Anthropic pre-4.8, Gemini
467    /// `systemInstruction`, Bedrock Converse `system[]`). A leading run folds
468    /// into the top-level system prompt; an interleaved directive folds into
469    /// the adjacent user turn as a `<system-reminder>` block so its position
470    /// and operator intent survive. Never 400s, never silently repositioned to
471    /// the global system prompt.
472    Fold,
473}
474
475/// Provider field that must carry Harn's private reasoning from an assistant
476/// turn into that provider's next request.
477///
478/// This is deliberately an enum, rather than an arbitrary catalog string:
479/// replaying private reasoning changes the provider-visible transcript, so an
480/// unknown field must fail capability loading rather than silently leak or
481/// discard it. The durable Harn transcript keeps this as `reasoning`; the
482/// OpenAI-compatible request boundary projects it only for the selected mode.
483#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
484#[serde(rename_all = "snake_case")]
485pub enum ReasoningHistoryWireField {
486    /// Moonshot's OpenAI-compatible API requires this on the prior assistant
487    /// message before matching tool results.
488    ReasoningContent,
489}
490
491impl ReasoningHistoryWireField {
492    pub const fn as_str(self) -> &'static str {
493        match self {
494            Self::ReasoningContent => "reasoning_content",
495        }
496    }
497}
498
499/// Resolved capabilities for a `(provider, model)` pair. Unset rule
500/// fields resolve to `false` / empty / `None` so callers never have to
501/// unwrap an `Option<bool>` for what are really boolean gates.
502#[derive(Debug, Clone, PartialEq, Eq)]
503pub struct Capabilities {
504    pub native_tools: bool,
505    pub message_wire_format: WireDialect,
506    pub native_tool_wire_format: String,
507    pub defer_loading: bool,
508    pub tool_search: Vec<String>,
509    pub responses_api: bool,
510    pub hosted_tools: Vec<String>,
511    pub remote_mcp: bool,
512    pub conversation_state: bool,
513    pub compaction: bool,
514    pub background_mode: bool,
515    pub batch_api: bool,
516    pub batch_wire_format: Option<String>,
517    pub batch_input_mode: Option<String>,
518    pub batch_discount_percent: Option<u32>,
519    pub batch_turnaround_hours: Option<u32>,
520    pub batch_max_requests: Option<u64>,
521    pub batch_max_input_bytes: Option<u64>,
522    pub batch_result_retention_days: Option<u32>,
523    pub batch_result_ordering: Option<String>,
524    pub batch_partial_failure: Option<String>,
525    pub batch_cancellation: Option<String>,
526    pub batch_security_notes: Vec<String>,
527    pub batch_operational_notes: Vec<String>,
528    pub batch_regions: Vec<String>,
529    pub tool_approval_policy: Option<String>,
530    pub max_tools: Option<u32>,
531    pub prompt_caching: bool,
532    pub prompt_cache_ttls: Vec<String>,
533    /// Shortest prefix this route will actually cache, in tokens. `None` means
534    /// the route has no model-specific floor and the wire-dialect default in
535    /// [`crate::llm::cache_conformance::CacheControlProfile`] applies.
536    pub prompt_cache_min_prefix_tokens: Option<u32>,
537    pub cache_breakpoint_style: String,
538    pub vision: bool,
539    pub audio: bool,
540    pub pdf: bool,
541    pub video: bool,
542    pub files_api_supported: bool,
543    pub file_upload_wire_format: Option<String>,
544    pub structured_output: Option<String>,
545    /// Legacy mirror for CLI display and older callers.
546    pub json_schema: Option<String>,
547    pub prefers_xml_scaffolding: bool,
548    /// See [`ProviderRule::reserved_tool_call_token`].
549    pub reserved_tool_call_token: bool,
550    pub prefers_markdown_scaffolding: bool,
551    pub structured_output_mode: String,
552    pub supports_assistant_prefill: bool,
553    pub prefers_role_developer: bool,
554    pub prefers_xml_tools: bool,
555    pub thinking_block_style: String,
556    /// Which synchronous ("live") endpoint family this route dispatches to,
557    /// when its wire dialect serves more than one. `None` means the dialect has
558    /// exactly one live endpoint, so there is nothing to select. See
559    /// [`LiveEndpointFamily`].
560    pub live_endpoint_family: Option<LiveEndpointFamily>,
561    /// Whether this route emits its reasoning INLINE in the text channel as
562    /// `<think>...</think>` blocks (local Ollama/llama.cpp reasoning models,
563    /// Qwen3 via vLLM, Kimi) rather than in a separate provider reasoning
564    /// field. When true, the `llm_call` envelope builder splits those blocks
565    /// out of `text`/`prose`/`visible_text` and folds them into the reasoning
566    /// channel, mirroring how hosted providers surface a dedicated thinking
567    /// field. Derived from `thinking_block_style == "inline"` — the same
568    /// population that represents reasoning as inline `<think>` in prompt
569    /// context is the one that emits it that way in responses.
570    pub emits_inline_reasoning: bool,
571    pub thinking_modes: Vec<String>,
572    pub interleaved_thinking_supported: bool,
573    pub anthropic_beta_features: Vec<String>,
574    pub vision_supported: bool,
575    pub image_url_input_supported: bool,
576    pub preserve_thinking: bool,
577    /// Provider-specific wire field used to replay Harn's private reasoning
578    /// on the preceding assistant turn. `None` preserves the privacy default:
579    /// reasoning is never sent back to an OpenAI-compatible provider.
580    pub reasoning_history_wire_field: Option<ReasoningHistoryWireField>,
581    pub server_parser: String,
582    pub honors_chat_template_kwargs: bool,
583    pub chat_template_options_field: Option<String>,
584    pub requires_completion_tokens: bool,
585    /// True when the route is served ONLY by the provider Responses API and
586    /// rejects `/v1/chat/completions` (OpenAI `*-codex` models). Harn routes
587    /// such calls through the Responses provider automatically.
588    pub chat_completions_unsupported: bool,
589    /// See [`ProviderRule::reasoning_tools_require_responses`].
590    pub reasoning_tools_require_responses: bool,
591    pub requires_streaming: bool,
592    pub reasoning_effort_supported: bool,
593    pub reasoning_effort_levels: Vec<String>,
594    pub reasoning_none_supported: bool,
595    /// See [`ProviderRule::max_thinking_budget`]. `None` means the model uses
596    /// the provider's own default ceiling.
597    pub max_thinking_budget: Option<i64>,
598    pub reasoning_disable_supported: bool,
599    /// See [`ProviderRule::reasoning_required_for_tools`].
600    pub reasoning_required_for_tools: bool,
601    pub reasoning_text_promotable: bool,
602    pub reasoning_wire_format: Option<String>,
603    pub seed_supported: bool,
604    pub top_k_supported: bool,
605    pub temperature_supported: bool,
606    pub top_p_supported: bool,
607    pub frequency_penalty_supported: bool,
608    pub presence_penalty_supported: bool,
609    pub stop_supported: bool,
610    pub allowed_tool_choice_modes: Vec<String>,
611    pub requires_tool_result_adjacency: bool,
612    pub supports_parallel_tool_calls: bool,
613    pub tools_exclude_response_format: bool,
614    pub recommended_endpoint: Option<String>,
615    pub text_tool_wire_format_supported: bool,
616    pub preferred_tool_format: Option<String>,
617    pub tool_mode_parity: Option<String>,
618    pub tool_mode_parity_notes: Option<String>,
619    pub thinking_disable_directive: Option<String>,
620    /// Per-task auto-policy reasoning-level overrides for this route.
621    /// See [`ProviderRule::auto_reasoning_overrides`].
622    pub auto_reasoning_overrides: BTreeMap<String, String>,
623    /// OpenRouter upstream provider names to exclude from routing for this
624    /// row. See [`ProviderRule::provider_route_denylist`]. Empty means "no
625    /// route restriction".
626    pub provider_route_denylist: Vec<String>,
627    /// OpenRouter upstream provider names this row is PINNED to (allowlist), in
628    /// preference order. See [`ProviderRule::openrouter_provider_order`]. Empty
629    /// means "no pin" (free OpenRouter routing).
630    pub openrouter_provider_order: Vec<String>,
631    /// Serving-quality / precision trust verdict for this route. See
632    /// [`ProviderRule::serving_precision`]. `"unverified"` when unset.
633    pub serving_precision: String,
634    /// How the neutral `computer` tool projects onto this route's native
635    /// computer-use surface. `None` means the route exposes no computer-use
636    /// surface. See [`ComputerUseStyle`].
637    pub computer_use_style: Option<ComputerUseStyle>,
638    /// Screenshot downscaling policy applied before the image reaches the
639    /// model. `None` means unset. See [`ScreenshotScaling`].
640    pub screenshot_scaling: Option<ScreenshotScaling>,
641    /// Whether this route requires echoing acknowledged safety checks on the
642    /// computer-use follow-up turn (OpenAI Responses `pending_safety_checks`
643    /// → `acknowledged_safety_checks`). See [`ProviderRule::safety_ack_flow`].
644    pub safety_ack_flow: bool,
645    /// How this route carries an interleaved `system`/`developer` message.
646    /// `None` derives a safe default from `message_wire_format` (OpenAI/Ollama
647    /// → `Inline`, everything else → `Fold`). See [`SystemMessagePlacement`]
648    /// and [`resolve_system_message_placement`](super::resolve_system_message_placement).
649    pub system_message_placement: Option<SystemMessagePlacement>,
650}
651
652impl Default for Capabilities {
653    fn default() -> Self {
654        Self {
655            native_tools: false,
656            message_wire_format: WireDialect::OpenAiCompat,
657            native_tool_wire_format: "openai".to_string(),
658            defer_loading: false,
659            tool_search: Vec::new(),
660            responses_api: false,
661            hosted_tools: Vec::new(),
662            remote_mcp: false,
663            conversation_state: false,
664            compaction: false,
665            background_mode: false,
666            batch_api: false,
667            batch_wire_format: None,
668            batch_input_mode: None,
669            batch_discount_percent: None,
670            batch_turnaround_hours: None,
671            batch_max_requests: None,
672            batch_max_input_bytes: None,
673            batch_result_retention_days: None,
674            batch_result_ordering: None,
675            batch_partial_failure: None,
676            batch_cancellation: None,
677            batch_security_notes: Vec::new(),
678            batch_operational_notes: Vec::new(),
679            batch_regions: Vec::new(),
680            tool_approval_policy: None,
681            max_tools: None,
682            prompt_caching: false,
683            prompt_cache_ttls: Vec::new(),
684            prompt_cache_min_prefix_tokens: None,
685            cache_breakpoint_style: "none".to_string(),
686            vision: false,
687            audio: false,
688            pdf: false,
689            video: false,
690            files_api_supported: false,
691            file_upload_wire_format: None,
692            structured_output: None,
693            json_schema: None,
694            prefers_xml_scaffolding: false,
695            reserved_tool_call_token: false,
696            prefers_markdown_scaffolding: false,
697            structured_output_mode: "none".to_string(),
698            supports_assistant_prefill: false,
699            prefers_role_developer: false,
700            prefers_xml_tools: false,
701            thinking_block_style: "none".to_string(),
702            live_endpoint_family: None,
703            emits_inline_reasoning: false,
704            thinking_modes: Vec::new(),
705            interleaved_thinking_supported: false,
706            anthropic_beta_features: Vec::new(),
707            vision_supported: false,
708            image_url_input_supported: true,
709            preserve_thinking: false,
710            reasoning_history_wire_field: None,
711            server_parser: "none".to_string(),
712            honors_chat_template_kwargs: false,
713            chat_template_options_field: None,
714            requires_completion_tokens: false,
715            chat_completions_unsupported: false,
716            reasoning_tools_require_responses: false,
717            requires_streaming: false,
718            reasoning_effort_supported: false,
719            reasoning_effort_levels: Vec::new(),
720            reasoning_none_supported: false,
721            max_thinking_budget: None,
722            reasoning_disable_supported: true,
723            reasoning_required_for_tools: false,
724            reasoning_text_promotable: false,
725            reasoning_wire_format: None,
726            seed_supported: true,
727            top_k_supported: true,
728            temperature_supported: true,
729            top_p_supported: true,
730            frequency_penalty_supported: true,
731            presence_penalty_supported: true,
732            stop_supported: true,
733            allowed_tool_choice_modes: Vec::new(),
734            requires_tool_result_adjacency: false,
735            supports_parallel_tool_calls: true,
736            tools_exclude_response_format: false,
737            recommended_endpoint: None,
738            text_tool_wire_format_supported: true,
739            preferred_tool_format: None,
740            tool_mode_parity: None,
741            tool_mode_parity_notes: None,
742            thinking_disable_directive: None,
743            auto_reasoning_overrides: BTreeMap::new(),
744            provider_route_denylist: Vec::new(),
745            openrouter_provider_order: Vec::new(),
746            serving_precision: "unverified".to_string(),
747            computer_use_style: None,
748            screenshot_scaling: None,
749            safety_ack_flow: false,
750            system_message_placement: None,
751        }
752    }
753}