Skip to main content

firstpass_core/
config.rs

1//! Routing configuration (SPEC §8.4) — declarative, agent-first.
2//!
3//! Config is the policy an operator (or agent) hands Firstpass: an ordered list of routes,
4//! each matching a slice of traffic to a mode, a model ladder, and gates; plus budget caps
5//! and escalation rules. The first matching route wins, so specific routes go first and a
6//! bare `match = {}` catch-all goes last.
7
8use crate::error::{Error, Result};
9use crate::features::{Features, TaskKind};
10use serde::{Deserialize, Serialize};
11use std::time::Duration;
12
13/// Serving mode for a route (SPEC glossary).
14#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
15#[serde(rename_all = "lowercase")]
16pub enum Mode {
17    /// Gate before serving; escalate on fail. The output is proven before the caller sees it.
18    Enforce,
19    /// Serve immediately, gate asynchronously — verdicts feed learning only. Zero added latency,
20    /// zero risk; the default for onboarding unfamiliar traffic.
21    Observe,
22}
23
24/// Named routing-mode presets: coherent (accuracy/cost/latency) constraint bundles layered
25/// on top of the existing escalation knobs. A mode is a **resolver** over existing config,
26/// not a new engine — it overrides only the knobs it has an opinion on; everything else
27/// comes from the route/global config unchanged.
28///
29/// `Balanced` (the default) is a strict no-op: with no mode set anywhere, behaviour is
30/// byte-identical to existing behaviour. All other modes are opt-in overlays.
31///
32/// Set per-request via the `x-firstpass-mode` header, per-route via `routing_mode = "..."`,
33/// or globally via the `FIRSTPASS_MODE_PROFILE` env var.
34#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
35#[serde(rename_all = "lowercase")]
36pub enum RoutingMode {
37    /// Forward without routing or gating; verdicts are asynchronous learning signals only.
38    /// Maps to the existing observe passthrough path. Zero added latency, zero quality proof.
39    Observe,
40    /// Prefer cheapest start; no speculative prefetch.
41    /// Tradeoff: lowest token spend; can serve lower quality on hard or heavy traffic.
42    Cost,
43    /// Today's default — bandit start, configured thresholds, configured speculation.
44    /// **This preset is a strict no-op: byte-identical to existing behaviour when applied.**
45    #[default]
46    Balanced,
47    /// One extra escalation rung allowed; serial (no speculative prefetch waste).
48    /// Tradeoff: higher quality ceiling at higher cost; still bounded by gate verification.
49    Quality,
50    /// `speculation = 1` always: always prefetch one rung ahead to cut p95 latency.
51    /// Tradeoff: pays 1 wasted speculative call when the cheap rung passes.
52    Latency,
53    /// Start at the top (most expensive) ladder rung; verification as insurance.
54    /// Tradeoff: highest quality, highest cost per call; bandit is bypassed; savings are minimal.
55    Max,
56}
57
58impl RoutingMode {
59    /// Lowercase string name, for use in headers and trace fields.
60    #[must_use]
61    pub fn as_str(self) -> &'static str {
62        match self {
63            Self::Observe => "observe",
64            Self::Cost => "cost",
65            Self::Balanced => "balanced",
66            Self::Quality => "quality",
67            Self::Latency => "latency",
68            Self::Max => "max",
69        }
70    }
71
72    /// All named modes in display order, for discovery surfaces (capabilities, MCP, CLI).
73    pub const ALL: &'static [Self] = &[
74        Self::Observe,
75        Self::Cost,
76        Self::Balanced,
77        Self::Quality,
78        Self::Latency,
79        Self::Max,
80    ];
81
82    /// The knob overrides this mode applies on top of the route/global escalation config.
83    /// `None` fields are left unchanged — modes only override where they have an opinion.
84    ///
85    /// **`Balanced` returns all `None`/`false`: it is a strict no-op by invariant.**
86    #[must_use]
87    pub fn preset(self) -> ModePreset {
88        match self {
89            // Observe: handled as a passthrough redirect before EnforceCtx; preset is unused.
90            Self::Observe => ModePreset {
91                speculation: None,
92                max_rungs_delta: None,
93                start_at_top: false,
94                description: "shadow mode: forward without gating; verdicts are async learning signals only",
95                tradeoff: "zero added latency, zero quality proof; use to observe before enforcing",
96            },
97            // Cost: kill speculative spend; let the bandit start cheap.
98            Self::Cost => ModePreset {
99                speculation: Some(0),
100                max_rungs_delta: None,
101                start_at_top: false,
102                description: "no speculative prefetch; bandit start on cheapest rung",
103                tradeoff: "lowest token spend; can serve lower quality on hard or heavy traffic",
104            },
105            // Balanced: all None — strict no-op so existing behaviour is preserved.
106            Self::Balanced => ModePreset {
107                speculation: None,
108                max_rungs_delta: None,
109                start_at_top: false,
110                description: "bandit start, configured thresholds and speculation (default — no behavioral change)",
111                tradeoff: "today's tuned balance; add a mode only when you need a different point",
112            },
113            // Quality: one more rung; no speculative waste.
114            Self::Quality => ModePreset {
115                speculation: Some(0),
116                max_rungs_delta: Some(1),
117                start_at_top: false,
118                description: "one extra escalation rung; serial (no speculative waste)",
119                tradeoff: "higher quality ceiling at higher cost; still bounded by gate verification",
120            },
121            // Latency: speculation=1 always; pay 1 wasted call to cut p95.
122            Self::Latency => ModePreset {
123                speculation: Some(1),
124                max_rungs_delta: None,
125                start_at_top: false,
126                description: "speculation=1: always prefetch one rung ahead to cut p95 latency",
127                tradeoff: "pays 1 wasted speculative call when cheap rung passes; speculation_band is overridden",
128            },
129            // Max: jump to top rung immediately; verification as insurance.
130            Self::Max => ModePreset {
131                speculation: Some(0),
132                max_rungs_delta: None,
133                start_at_top: true,
134                description: "start at the top (most expensive) ladder rung; verification as insurance",
135                tradeoff: "highest quality, highest cost per call; bandit bypassed; savings minimal",
136            },
137        }
138    }
139}
140
141/// Knob overrides a [`RoutingMode`] applies on top of the route/global escalation config.
142/// Only non-`None` fields override; absent fields leave the config value unchanged.
143///
144/// `Balanced` returns all `None`/`false` — it is a strict no-op.
145#[derive(Debug, Clone, Copy)]
146pub struct ModePreset {
147    /// Override for `escalation.speculation`. `None` = no override (config value unchanged).
148    pub speculation: Option<u32>,
149    /// Additive delta applied to `max_rungs_per_request`. `None` = no change.
150    /// Applied as `(current + delta).max(1)`.
151    pub max_rungs_delta: Option<i32>,
152    /// When `true`, start at the last ladder rung (top quality), bypassing the bandit.
153    pub start_at_top: bool,
154    /// One-line description of what this mode targets.
155    pub description: &'static str,
156    /// Honest tradeoff: what you gain and what you give up.
157    pub tradeoff: &'static str,
158}
159
160/// Top-level configuration document.
161#[derive(Debug, Clone, Default, Deserialize)]
162#[serde(deny_unknown_fields)]
163pub struct Config {
164    /// Ordered routes; first match wins.
165    #[serde(rename = "route", default)]
166    pub routes: Vec<Route>,
167    /// Spend caps.
168    #[serde(default)]
169    pub budget: Budget,
170    /// Escalation limits and promotion rules.
171    #[serde(default)]
172    pub escalation: Escalation,
173    /// Served-failure guardrail (ADR 0009 D3). Absent (the default) means nothing watches the
174    /// served-failure rate and behaviour is byte-identical to before.
175    #[serde(default)]
176    pub guardrail: Option<crate::guardrail::Guardrail>,
177    /// Seconds a demoted route stays demoted before the guardrail will consider it again.
178    /// Deliberately long: re-promoting on one good window makes routing flap, and every flap is
179    /// a visible behaviour change for real users.
180    #[serde(default = "default_guardrail_cooldown")]
181    pub guardrail_cooldown_secs: i64,
182    /// User-defined subprocess gates (SPEC §8.1), referenced by `id` from a route's `gates` /
183    /// `deferred_gates`. Declared as `[[gate]]` sections in TOML.
184    #[serde(rename = "gate", default)]
185    pub gate_defs: Vec<GateDef>,
186    /// Per-deployment price overrides, replacing the built-in defaults for the named models.
187    /// List prices drift and enterprise contracts differ — savings math is only honest when the
188    /// operator can pin THEIR prices. Declared as `[[price]]` sections in TOML.
189    #[serde(rename = "price", default)]
190    pub price_defs: Vec<PriceDef>,
191    /// Extra model providers a ladder can route to, beyond the built-in `anthropic` / `openai`.
192    /// Any OpenAI-compatible endpoint (Groq, Together, Fireworks, DeepSeek, Mistral, xAI,
193    /// OpenRouter, Ollama, vLLM, Azure, …) is one `[[provider]]` entry — no rebuild. Declared as
194    /// `[[provider]]` sections; a ladder rung is then `<id>/<model>`.
195    #[serde(rename = "provider", default)]
196    pub providers: Vec<ProviderDef>,
197}
198
199/// The wire API a provider speaks. `anthropic` = Messages API; `openai` = Chat Completions API
200/// (the de-facto standard that nearly every hosted and open-source model host implements);
201/// `gemini` = Google's Generative Language API (`generateContent`).
202#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize)]
203#[serde(rename_all = "lowercase")]
204pub enum Dialect {
205    /// Anthropic Messages API (`POST /v1/messages`).
206    Anthropic,
207    /// OpenAI Chat Completions API (`POST /v1/chat/completions`).
208    Openai,
209    /// Google Gemini Generative Language API (`POST /v1beta/models/<model>:generateContent`),
210    /// authenticated with an API key in the `x-goog-api-key` header.
211    Gemini,
212}
213
214/// How a provider call is credentialed — orthogonal to [`Dialect`] (ADR 0006): dialect shapes the
215/// request/response body, auth scheme shapes how the request is signed/credentialed. Default
216/// (`api_key_env` / BYOK header) is unchanged for every existing `[[provider]]` entry.
217#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Deserialize)]
218#[serde(rename_all = "snake_case")]
219pub enum AuthScheme {
220    /// API key in a header (`x-api-key`, `Authorization: Bearer`, or `x-goog-api-key`) — today's
221    /// behavior for `anthropic` / `openai` / `gemini`.
222    #[default]
223    ApiKey,
224    /// AWS SigV4 request signing (Bedrock) — credentials from `AWS_ACCESS_KEY_ID` /
225    /// `AWS_SECRET_ACCESS_KEY` / `AWS_SESSION_TOKEN`, region-scoped.
226    AwsSigv4,
227    /// GCP OAuth2 bearer token (Vertex AI) — minted and cached by `gcp_auth` from
228    /// `GOOGLE_APPLICATION_CREDENTIALS` or the ambient environment.
229    GcpOauth,
230}
231
232/// A model provider a ladder can route to. Declared as `[[provider]]` in TOML; referenced from a
233/// ladder as `<id>/<model>` (e.g. `groq/llama-3.3-70b-versatile`).
234#[derive(Debug, Clone, Deserialize)]
235#[serde(deny_unknown_fields)]
236pub struct ProviderDef {
237    /// Ladder prefix for this provider, e.g. `"groq"`. Overrides a built-in of the same id.
238    pub id: String,
239    /// Which wire API it speaks.
240    pub dialect: Dialect,
241    /// Base URL, e.g. `"https://api.groq.com/openai"` or `"http://localhost:11434"` for Ollama.
242    /// Unused for `aws_sigv4/gcp_oauth` auth (those construct the URL from `region`/`project`).
243    #[serde(default)]
244    pub base_url: String,
245    /// Env var the API key is read from at call time, e.g. `"GROQ_API_KEY"`. Omit for a keyless
246    /// endpoint (local Ollama / vLLM). Per-request BYOK headers still apply to the built-in
247    /// `anthropic` / `openai` providers.
248    #[serde(default)]
249    pub api_key_env: Option<String>,
250    /// How this provider's requests are credentialed. `api_key` (default) is byte-identical to
251    /// today; `aws_sigv4` / `gcp_oauth` are the Bedrock/Vertex auth schemes (ADR 0006).
252    #[serde(default)]
253    pub auth: AuthScheme,
254    /// Cloud region, e.g. `"us-east-1"` — required for `aws_sigv4` (Bedrock) and `gcp_oauth`
255    /// (Vertex).
256    #[serde(default)]
257    pub region: Option<String>,
258    /// GCP project id — required for `gcp_oauth` (Vertex).
259    #[serde(default)]
260    pub project: Option<String>,
261}
262
263/// A per-deployment price override for one model (`provider/model`), USD per 1M tokens.
264#[derive(Debug, Clone, Deserialize)]
265#[serde(deny_unknown_fields)]
266pub struct PriceDef {
267    /// The ladder id this price applies to, e.g. `"anthropic/claude-haiku-4-5"`.
268    pub model: String,
269    /// USD per 1M input (prompt) tokens.
270    pub input_per_mtok: f64,
271    /// USD per 1M output (completion) tokens.
272    pub output_per_mtok: f64,
273}
274
275/// A user-defined gate (SPEC §8.1). Exactly one kind per definition:
276/// - **subprocess** (`cmd`): any executable that reads the candidate as JSON on **stdin** (never
277///   argv — injection-resistant) and emits `{"verdict":"pass|fail|abstain", ...}` on stdout.
278/// - **judge** (`judge`): a native LLM-judge gate that grades the candidate against a rubric.
279/// - **consistency** (`consistency`): a self-consistency uncertainty gate that scores the
280///   candidate by agreement with k fresh samples of the same model (Wang et al. 2022).
281/// - **schema** (`schema`): validates the candidate (parsed as JSON) against a JSON-Schema
282///   subset (top-level `type` / `required` / per-property `type`).
283#[derive(Debug, Clone, Deserialize)]
284#[serde(deny_unknown_fields)]
285pub struct GateDef {
286    /// The id a route references this gate by (must be unique and not shadow a built-in gate id).
287    pub id: String,
288    /// Subprocess command: program first, then its args — e.g. `["pytest", "-q"]`. Set this **or**
289    /// `judge` / `consistency` / `schema`, not both.
290    #[serde(default)]
291    pub cmd: Vec<String>,
292    /// Hard timeout in milliseconds for a subprocess gate; it abstains (`timeout`) if the process
293    /// runs longer.
294    #[serde(default = "default_gate_timeout_ms")]
295    pub timeout_ms: u64,
296    /// LLM-judge configuration. Set this **or** `cmd` / `consistency` / `schema`, not both.
297    #[serde(default)]
298    pub judge: Option<JudgeDef>,
299    /// Self-consistency configuration. Set this **or** `cmd` / `judge` / `schema`, not both.
300    #[serde(default)]
301    pub consistency: Option<ConsistencyDef>,
302    /// JSON-Schema (subset) the candidate must satisfy. Set this **or** `cmd` / `judge` /
303    /// `consistency`, not both.
304    #[serde(default)]
305    pub schema: Option<serde_json::Value>,
306    /// What an **abstain** from this gate means for serving (§7.2). `fail_open` (default): an
307    /// abstaining gate never blocks serving — availability over strictness, today's behavior.
308    /// `fail_closed`: an abstain blocks serving exactly like a `Fail` — strictness over
309    /// availability, for gates whose silence must never be mistaken for approval.
310    #[serde(default)]
311    pub on_abstain: AbstainPolicy,
312}
313
314/// Per-gate abstain policy (§7.2): what happens to serving when the gate can't produce a verdict.
315#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Deserialize)]
316#[serde(rename_all = "snake_case")]
317pub enum AbstainPolicy {
318    /// An abstain never blocks serving (the historical behavior).
319    #[default]
320    FailOpen,
321    /// An abstain blocks serving exactly like a `Fail`.
322    FailClosed,
323}
324
325/// Configuration for a native LLM-judge gate (SPEC §8.3): a separate model grades the candidate
326/// against a rubric. The runner enforces maker ≠ checker (a model never grades its own output) and
327/// treats the candidate as data, not instructions.
328#[derive(Debug, Clone, Deserialize)]
329#[serde(deny_unknown_fields)]
330pub struct JudgeDef {
331    /// The judge model, as `provider/model` (must differ from the candidate's model at runtime).
332    pub model: String,
333    /// Pass iff the judge's score ≥ this threshold, in `[0, 1]`.
334    #[serde(default = "default_judge_threshold")]
335    pub threshold: f64,
336    /// What "good" means for this route — handed to the judge as the grading rubric.
337    #[serde(default)]
338    pub rubric: Option<String>,
339}
340
341/// Configuration for a self-consistency uncertainty gate: resample the original request `k` times
342/// on the same model and score the candidate by agreement with the fresh samples.
343///
344/// Research basis: Wang et al. 2022 (self-consistency) and Farquhar et al., *Nature* 2024
345/// (semantic entropy). Answers a model reliably reproduces are far likelier correct; disagreement
346/// flags hallucination. This produces a continuous confidence score the conformal threshold
347/// machinery can calibrate.
348///
349/// **maker == checker is intentional** — unlike a judge gate, self-consistency is definitionally
350/// self-referential. This is the mechanism, not a bug.
351#[derive(Debug, Clone, Deserialize)]
352#[serde(deny_unknown_fields)]
353pub struct ConsistencyDef {
354    /// The model used for resampling, as `provider/model`. May equal the candidate's model —
355    /// self-consistency is self-referential by design.
356    pub model: String,
357    /// Number of resample calls. Must be in `[2, 8]`; defaults to `3`.
358    #[serde(default = "default_consistency_k")]
359    pub k: u32,
360    /// Pass iff the mean agreement score ≥ this threshold, in `[0, 1]`.
361    pub threshold: f64,
362}
363
364/// Default subprocess-gate timeout: 30s. Long enough for a test suite, short enough to bound the
365/// enforce-path tail.
366fn default_gate_timeout_ms() -> u64 {
367    30_000
368}
369
370/// Default judge pass threshold.
371fn default_judge_threshold() -> f64 {
372    0.7
373}
374
375/// Default self-consistency k (resample count).
376fn default_consistency_k() -> u32 {
377    3
378}
379
380/// One routing rule.
381#[derive(Debug, Clone, Deserialize)]
382#[serde(deny_unknown_fields)]
383pub struct Route {
384    /// The traffic this route claims. An empty match (`{}`) matches everything.
385    #[serde(rename = "match", default)]
386    pub match_: Match,
387    /// Serving mode.
388    pub mode: Mode,
389    /// Model ladder, cheapest first, as `provider/model` strings.
390    #[serde(default)]
391    pub ladder: Vec<String>,
392    /// Inline gates run before serving (enforce mode).
393    #[serde(default)]
394    pub gates: Vec<String>,
395    /// Gates run asynchronously after serving; their verdicts attach to the trace as a
396    /// learning signal and never block the response.
397    #[serde(default)]
398    pub deferred_gates: Vec<String>,
399    /// Per-route routing-mode preset. Overrides `FIRSTPASS_MODE_PROFILE`; overridden by the
400    /// per-request `x-firstpass-mode` header. Absent (default `None`) falls through to the
401    /// global default, then `Balanced` — byte-identical to existing behaviour.
402    #[serde(default)]
403    pub routing_mode: Option<RoutingMode>,
404    /// Percentage rollout (ADR 0009 D1): enforce only a slice of this route's matched traffic,
405    /// held constant per session so a conversation never flips arm mid-thread. Absent (the
406    /// default) enforces all matched traffic — byte-identical to existing behaviour.
407    #[serde(default)]
408    pub rollout: Option<crate::rollout::Rollout>,
409    /// Shadow scoring (ADR 0009 D2): on a sample of *observed* requests, run the ladder off the
410    /// hot path and record what would have been served and what it would have cost. Absent (the
411    /// default) leaves observe exactly as it was. Makes real model calls when enabled, which is
412    /// why `max_usd_per_day` is required rather than defaulted.
413    #[serde(default)]
414    pub shadow: Option<crate::rollout::Shadow>,
415}
416
417/// Predicate over a request's [`Features`]. Every present field is an AND-constraint; absent
418/// fields are wildcards.
419#[derive(Debug, Clone, Default, Deserialize)]
420#[serde(deny_unknown_fields)]
421pub struct Match {
422    /// Require a specific calling agent.
423    #[serde(default)]
424    pub agent: Option<String>,
425    /// Require the subagent to be one of these (or this one, if a bare string).
426    #[serde(default)]
427    pub subagent: Option<StringOrVec>,
428    /// Require a specific task kind.
429    #[serde(default)]
430    pub task_kind: Option<TaskKind>,
431    /// Require the language to be one of these (or this one, if a bare string).
432    #[serde(default)]
433    pub language: Option<StringOrVec>,
434}
435
436impl Match {
437    /// Whether `f` satisfies every present constraint.
438    #[must_use]
439    pub fn matches(&self, f: &Features) -> bool {
440        if let Some(agent) = &self.agent
441            && f.agent.as_deref() != Some(agent.as_str())
442        {
443            return false;
444        }
445        if let Some(subs) = &self.subagent {
446            match &f.subagent {
447                Some(s) if subs.contains(s) => {}
448                _ => return false,
449            }
450        }
451        if let Some(tk) = self.task_kind
452            && f.task_kind != tk
453        {
454            return false;
455        }
456        if let Some(langs) = &self.language {
457            match &f.language {
458                Some(l) if langs.contains(l) => {}
459                _ => return false,
460            }
461        }
462        true
463    }
464}
465
466/// A field that accepts either a single string or a list of strings in TOML/JSON.
467#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
468#[serde(untagged)]
469pub enum StringOrVec {
470    /// A single value.
471    One(String),
472    /// A set of values.
473    Many(Vec<String>),
474}
475
476impl StringOrVec {
477    /// Whether `needle` is contained in this value.
478    #[must_use]
479    pub fn contains(&self, needle: &str) -> bool {
480        match self {
481            StringOrVec::One(s) => s == needle,
482            StringOrVec::Many(v) => v.iter().any(|s| s == needle),
483        }
484    }
485}
486
487/// What to do when a spend cap is hit (SPEC §8.4 — never brick the customer).
488#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
489#[serde(rename_all = "snake_case")]
490pub enum OnExhausted {
491    /// Serve the best attempt seen so far (default).
492    #[default]
493    ServeBestAttempt,
494    /// Return a structured error instead of serving.
495    Error,
496}
497
498/// Spend caps. `None` means "no cap".
499#[derive(Debug, Clone, Default, Deserialize)]
500#[serde(deny_unknown_fields)]
501pub struct Budget {
502    /// Max USD per single request (across all rungs + gates).
503    #[serde(default)]
504    pub per_request_usd: Option<f64>,
505    /// Max USD per session.
506    #[serde(default)]
507    pub per_session_usd: Option<f64>,
508    /// Max USD per day.
509    #[serde(default)]
510    pub per_day_usd: Option<f64>,
511    /// Behaviour when a cap is reached.
512    #[serde(default)]
513    pub on_exhausted: OnExhausted,
514}
515
516/// Escalation limits and session-level promotion.
517#[derive(Debug, Clone, Deserialize)]
518#[serde(deny_unknown_fields)]
519pub struct Escalation {
520    /// Hard ceiling on rungs climbed within one request.
521    #[serde(default = "default_max_rungs")]
522    pub max_rungs_per_request: u32,
523    /// Optionally start higher for the rest of a session that keeps failing.
524    #[serde(default)]
525    pub session_promotion: Option<SessionPromotion>,
526    /// Prefetch depth: fire this many rungs ahead concurrently to trade wasted spend for latency.
527    /// `0` (default) is serial — one call at a time. The served result is identical either way.
528    #[serde(default)]
529    pub speculation: u32,
530    /// Calibrated conformal serve threshold: a rung is served iff its aggregate gate score is
531    /// `>=` this value (SPEC §10.1). `None` (default) keeps the original rule — serve iff the
532    /// aggregate gate verdict is `Pass` — byte-identical to today.
533    #[serde(default)]
534    pub serve_threshold: Option<f64>,
535    /// Online/adaptive conformal (Gibbs-Candès ACI): when set, the serve threshold is tracked
536    /// **live** from deferred feedback instead of held fixed, so served-failure stays at target under
537    /// distribution shift. `None` (default) uses the fixed `serve_threshold` above — byte-identical.
538    #[serde(default)]
539    pub adaptive: Option<AdaptiveConfig>,
540    /// Route tool-calling / multimodal requests through enforce (ADR 0005). **Default `true`**:
541    /// agent traffic — the target workload — is gated out of the box. A per-request fidelity
542    /// guard still applies: structured content only routes when every ladder rung's provider
543    /// carries it verbatim (Anthropic-dialect today); otherwise the request falls back to
544    /// transparent observe passthrough rather than risking corruption. Set `false` to restore
545    /// the pre-ADR-0005 behavior (structured requests always pass through un-gated).
546    #[serde(default = "default_enforce_structured")]
547    pub enforce_structured: bool,
548    /// UCB1 start-rung bandit: learn which rung to START the ladder on per request context, to
549    /// cut expected cost by skipping rungs that almost always fail for this context. `None`
550    /// (default) starts every request at rung 0 — byte-identical to today. Prediction may only
551    /// choose where the ladder STARTS; gating, escalation, and serving are untouched.
552    #[serde(default)]
553    pub bandit: Option<BanditConfig>,
554    /// Last-resort context condensing. `None` (default) = a conversation that overflows every
555    /// rung's context window fails, which is today's behaviour.
556    ///
557    /// Deliberately not a general trimming knob: condensing routinely would mean the gate verifies
558    /// an answer produced from a prompt the client never sent, and the receipt would attest to a
559    /// decision about a different question than the one asked. This fires only once the ladder is
560    /// exhausted, where the choice is no longer "faithful vs degraded" but "degraded vs failed".
561    #[serde(default)]
562    pub condense: Option<CondenseConfig>,
563    /// Insert Anthropic prompt-cache breakpoints on the stable prefix (system + tools).
564    /// `false` (default) = today's behaviour: whatever the caller sent, unchanged.
565    ///
566    /// Opt-in because it is a bet on your traffic, not a free optimisation. A cache write costs
567    /// 1.25x base input and a read 0.1x, so roughly one reuse of the prefix covers the premium and
568    /// everything after it saves heavily. A long agent session with a fixed system prompt and tool
569    /// set reuses constantly; single-shot traffic never does, and pays a 25% surcharge on the
570    /// prefix for nothing.
571    #[serde(default)]
572    pub prompt_cache: bool,
573    /// Replay a previously **verified** answer for an identical prompt, skipping the ladder
574    /// entirely. `None` (default) = every request runs the ladder, byte-identical to today.
575    ///
576    /// Unlike an ordinary response cache, an entry is written only when the receipt shows the
577    /// request was served on a passing verdict — so a hit cannot serve something that was never
578    /// proven. The receipt records `served_from: cache` and `cache_source`, naming the decision
579    /// whose gate passed.
580    #[serde(default)]
581    pub verified_cache: Option<VerifiedCacheConfig>,
582    /// Speculative-deferral band: when set (and the bandit has a warm gate-pass estimate for
583    /// the chosen start rung), speculative prefetch fires **only** when that estimate falls
584    /// inside `[low, high]` — the marginal zone where the next rung is *probably but not
585    /// certainly* needed, which is where parallel spend buys the most latency (speculative
586    /// cascades). Outside the band (confident pass or confident fail-through) requests run
587    /// serial and the speculative tokens are saved. `None` (default) = `speculation` applies
588    /// unconditionally, byte-identical to today. No bandit / cold context ⇒ band inapplicable
589    /// ⇒ configured `speculation` applies.
590    #[serde(default)]
591    pub speculation_band: Option<[f64; 2]>,
592    /// Epsilon-greedy start-rung exploration: randomise a fraction of start-rung choices and
593    /// record propensities so IPS/SNIPS off-policy estimates are valid. `None` (default) →
594    /// deterministic policy — byte-identical to today. See [`ExplorationConfig`].
595    #[serde(default)]
596    pub exploration: Option<ExplorationConfig>,
597    /// Shadow probe (ADR 0008 Phase 1): measure the k-sample gate-pass-count signal on a
598    /// sampled fraction of enforce requests without changing serving or the served cost.
599    /// `None` (default) = off = zero extra provider calls = byte-identical to today.
600    #[serde(default)]
601    pub probe: Option<ProbeConfig>,
602    /// Per-query gate-pass predictor (ADR 0008 Phase 2): a learned `P(gate-pass | rung,
603    /// features)` model, trained online from receipts and recorded on the receipt in shadow.
604    /// `None` (default) = off = byte-identical to today (no predictor, no `predicted_pass`).
605    #[serde(default)]
606    pub predictor: Option<PredictorConfig>,
607    /// Elastic verification (ADR 0008 Phase 3): skip the named expensive gates on a serve when the
608    /// cheap **visible** gate score clears the conformally-calibrated threshold λ. `None` (default)
609    /// runs every gate on every serve — byte-identical to today's uniform verification. See
610    /// [`ElasticConfig`].
611    #[serde(default)]
612    pub elastic: Option<ElasticConfig>,
613}
614
615/// Per-query gate-pass predictor configuration (ADR 0008 Phase 2). The model is an online
616/// logistic regression (`firstpass_core::predictor::PassPredictor`) trained from the trace
617/// store. Its prediction is recorded on the receipt in **shadow** — it does not change routing
618/// in this phase; whether it is good enough to act on is decided offline (`firstpass
619/// predictor-eval`).
620#[derive(Debug, Clone, Copy, Deserialize)]
621#[serde(deny_unknown_fields)]
622pub struct PredictorConfig {
623    /// SGD learning rate, in `(0, 1]`. Validated by [`Config::parse`].
624    #[serde(default = "default_predictor_lr")]
625    pub lr: f64,
626    /// L2 shrinkage on the weights (bias excluded), `>= 0`. Validated by [`Config::parse`].
627    #[serde(default = "default_predictor_l2")]
628    pub l2: f64,
629}
630
631fn default_predictor_lr() -> f64 {
632    0.05
633}
634
635fn default_predictor_l2() -> f64 {
636    1e-4
637}
638
639/// Config for online/adaptive conformal serving ([`crate::conformal::AdaptiveConformal`]).
640#[derive(Debug, Clone, Deserialize)]
641#[serde(deny_unknown_fields)]
642pub struct AdaptiveConfig {
643    /// Target served-failure rate the online loop holds to.
644    pub alpha: f64,
645    /// Step size (default 0.02) — larger tracks shift faster but noisier.
646    #[serde(default = "default_adaptive_gamma")]
647    pub gamma: f64,
648}
649
650fn default_adaptive_gamma() -> f64 {
651    0.02
652}
653
654/// Epsilon-greedy start-rung exploration: a fraction `epsilon` of requests start at a
655/// uniformly-random rung so that propensities are logged and IPS/SNIPS off-policy estimates
656/// are valid (Horvitz-Thompson 1952; SNIPS: Swaminathan & Joachims 2015). The bandit still
657/// observes all gate verdicts — including from exploration draws — so learning is uninterrupted.
658///
659/// Every request under this policy records `policy.propensity` in the trace: the probability the
660/// logging policy had of choosing the start rung it actually chose. Old traces (before this field
661/// was added) serialize byte-identically (propensity is omitted when `None`).
662///
663/// Absent (default) → deterministic policy — no exploration, no propensity recorded,
664/// byte-identical to today.
665#[derive(Debug, Clone, Deserialize)]
666#[serde(deny_unknown_fields)]
667pub struct ExplorationConfig {
668    /// Fraction of requests routed to a uniformly-random start rung instead of the greedy
669    /// choice. Must be finite and in `(0, 0.5]`; validated by [`Config::parse`].
670    /// A small `epsilon` (0.05–0.1) is usually enough: it keeps learning alive and makes
671    /// IPS/SNIPS estimates valid at low cost in exploration waste.
672    pub epsilon: f64,
673}
674
675/// Shadow-probe configuration (ADR 0008 Phase 1): measure the validated k-sample
676/// gate-pass-count signal on a sampled fraction of requests WITHOUT changing serving.
677///
678/// **Cost caveat**: `sample_rate × k` extra model calls (at the `start_rung` model) per
679/// sampled request — measurement only, nothing served changes. Default absent = off = zero
680/// extra provider calls, byte-identical to today. Enable only when you want to collect the
681/// regime signal for offline analysis.
682#[derive(Debug, Clone, Copy, Deserialize)]
683#[serde(deny_unknown_fields)]
684pub struct ProbeConfig {
685    /// Number of parallel probe samples per request. Must be in `[2, 8]`; validated by
686    /// [`Config::parse`]. Each sample is one model call at the `start_rung` model.
687    pub k: u32,
688    /// Fraction of requests that trigger the probe, in `[0, 1]`. `0.0` = never; `1.0` = always.
689    /// Must be finite; validated by [`Config::parse`].
690    pub sample_rate: f64,
691}
692
693/// Elastic-verification serving config (ADR 0008 Phase 3). When present, the gates named in
694/// `expensive_gates` are treated as *skippable*: a rung serves without running them when the
695/// **visible** (cheap, always-run) gate score is `>= lambda` — the elastic three-regime rule,
696/// validated offline (`firstpass-bench --elastic`, artifact `docs/benchmarks/elastic-validation.txt`).
697///
698/// The load-bearing invariant (ADR 0008): the un-verified serves carry the *same* distribution-free
699/// served-failure bound as the verified ones **only when `lambda` was conformally calibrated on
700/// representative data and the workload has not drifted**. `lambda` is therefore operator-supplied
701/// from a calibration run — obtain it from `firstpass calibrate` / the offline `--elastic` harness,
702/// never hand-tuned. `alpha`/`delta`/`calibration_id` are recorded verbatim on every skip receipt so
703/// an auditor can check the bound held against realized deferred outcomes.
704///
705/// Default-off: absent config = today's uniform cascade, byte-identical. Applies to the serial
706/// engine only; if `speculation > 0` is also configured the expensive gates still run (conservative
707/// — running more gates never weakens the bound).
708#[derive(Debug, Clone, Deserialize, PartialEq)]
709#[serde(deny_unknown_fields)]
710pub struct ElasticConfig {
711    /// Gate ids that may be skipped when the visible signal clears `lambda`. Ids **not** listed
712    /// here are "visible" (cheap, e.g. `non-empty`/`json-valid`/`schema`) and always run — they are
713    /// the probe signal. Must be non-empty; validated by [`Config::parse`].
714    pub expensive_gates: Vec<String>,
715    /// Calibrated skip threshold λ: serve without the expensive gates iff the visible gate score is
716    /// `>= lambda`. Must be finite and in `[0, 1]`; validated by [`Config::parse`].
717    pub lambda: f64,
718    /// Target served-failure rate α the λ was calibrated at. Recorded on every skip receipt for
719    /// audit; not enforced per-request (the calibration already baked it into λ).
720    #[serde(default)]
721    pub alpha: Option<f64>,
722    /// Confidence level (1−δ) the λ was calibrated at. Recorded on every skip receipt for audit.
723    #[serde(default)]
724    pub delta: Option<f64>,
725    /// Provenance id of the calibration run that produced λ. Recorded on every skip receipt so the
726    /// skip traces back to the exact conformal calibration that authorized it.
727    #[serde(default)]
728    pub calibration_id: Option<String>,
729}
730
731/// Last-resort context condensing (§8.4). See [`Escalation::condense`].
732#[derive(Debug, Clone, Deserialize)]
733#[serde(deny_unknown_fields)]
734pub struct CondenseConfig {
735    /// Opening turns to keep — usually where the task definition lives.
736    #[serde(default = "default_condense_keep_head")]
737    pub keep_head: usize,
738    /// Closing turns to keep — the live thread.
739    #[serde(default = "default_condense_keep_tail")]
740    pub keep_tail: usize,
741}
742
743const fn default_condense_keep_head() -> usize {
744    2
745}
746
747const fn default_condense_keep_tail() -> usize {
748    6
749}
750
751/// Verified-response cache (§8.4). See [`Escalation::verified_cache`].
752#[derive(Debug, Clone, Deserialize)]
753#[serde(deny_unknown_fields)]
754pub struct VerifiedCacheConfig {
755    /// How long a proven answer stays replayable.
756    ///
757    /// A proof has a shelf life: the model, the gate, and the world all move. Short by default
758    /// because replaying a stale answer confidently is worse than paying to re-derive it.
759    #[serde(default = "default_cache_ttl_secs")]
760    pub ttl_secs: u64,
761    /// Hard cap on entries; oldest evicted first. In-process store only — a Redis-backed cache is
762    /// bounded by its TTL and the server's own eviction policy.
763    #[serde(default = "default_cache_max_entries")]
764    pub max_entries: usize,
765    /// Redis URL for a **shared** cache across replicas. Absent (default) = in-process only.
766    ///
767    /// Without it the cache is per-replica: behind N instances the same answer is cached N times
768    /// over, the hit rate drops roughly by N, and — the part that matters — a retraction only
769    /// reaches the replica that received the feedback, leaving the other N-1 serving an answer
770    /// that has been disproven.
771    ///
772    /// Requires the `redis-cache` build feature. Configuring a URL without it is rejected at
773    /// startup rather than silently ignored: a cache that quietly stays per-replica looks
774    /// identical to one that is working.
775    #[serde(default)]
776    pub redis_url: Option<String>,
777}
778
779const fn default_cache_ttl_secs() -> u64 {
780    15 * 60
781}
782
783const fn default_cache_max_entries() -> usize {
784    10_000
785}
786
787/// Config for the UCB1 start-rung bandit (`firstpass_proxy::bandit::StartRungBandit`).
788///
789/// Absent (`None`) → start every request at rung 0 (byte-identical to today).
790#[derive(Debug, Clone, Deserialize)]
791#[serde(deny_unknown_fields)]
792pub struct BanditConfig {
793    /// Minimum total gate-verdict observations in a context bucket before the bandit's
794    /// prediction kicks in. Below this, every request starts at rung 0 (cold-start safety).
795    #[serde(default = "default_bandit_min_observations")]
796    pub min_observations: usize,
797    /// UCB1 exploration constant `c` (Auer et al. 2002). Higher values explore more; 1.0 is the
798    /// theoretical default. Must be finite and `>= 0`; validated by [`Config::parse`].
799    #[serde(default = "default_bandit_exploration")]
800    pub exploration: f64,
801    /// Selection algorithm. `"ucb1"` (default) is deterministic and auditable; `"thompson"`
802    /// samples Beta posteriors — stochastic by nature, so every decision logs a non-degenerate
803    /// propensity (clean IPS/SNIPS/DR off-policy estimates without the epsilon overlay) and
804    /// pairs with `discount` for non-stationary traffic.
805    #[serde(default)]
806    pub algorithm: BanditAlgorithm,
807    /// Per-observation multiplicative decay of a context's counts, in `(0, 1]`. `1.0`
808    /// (default) = no forgetting. `0.99` adapts to model churn / workload drift at the cost
809    /// of a slightly noisier estimate. Validated by [`Config::parse`].
810    #[serde(default = "default_bandit_discount")]
811    pub discount: f64,
812}
813
814/// Start-rung bandit selection algorithm.
815#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Deserialize)]
816#[serde(rename_all = "lowercase")]
817pub enum BanditAlgorithm {
818    /// Deterministic UCB1 (Auer et al. 2002).
819    #[default]
820    Ucb1,
821    /// Thompson sampling with Beta posteriors (Chapelle & Li 2011).
822    Thompson,
823}
824
825fn default_bandit_discount() -> f64 {
826    1.0
827}
828
829fn default_bandit_min_observations() -> usize {
830    50
831}
832
833fn default_bandit_exploration() -> f64 {
834    1.0
835}
836
837const fn default_max_rungs() -> u32 {
838    3
839}
840
841/// Structured (tools/images) enforce is on by default — the fidelity guard in the proxy keeps
842/// it safe (verbatim-carry rungs only).
843fn default_enforce_structured() -> bool {
844    true
845}
846
847impl Default for Escalation {
848    fn default() -> Self {
849        Self {
850            max_rungs_per_request: default_max_rungs(),
851            session_promotion: None,
852            speculation: 0,
853            serve_threshold: None,
854            adaptive: None,
855            enforce_structured: default_enforce_structured(),
856            speculation_band: None,
857            bandit: None,
858            condense: None,
859            prompt_cache: false,
860            verified_cache: None,
861            exploration: None,
862            probe: None,
863            predictor: None,
864            elastic: None,
865        }
866    }
867}
868
869/// "After N failures within a window, promote this session's starting rung."
870#[derive(Debug, Clone, Deserialize)]
871#[serde(deny_unknown_fields)]
872pub struct SessionPromotion {
873    /// Failure count that triggers promotion.
874    pub after_failures: u32,
875    /// Sliding window, e.g. `"30m"`. Also the idle TTL: a session untouched for this long forgets
876    /// its promotion, so a reused session id never inherits a stale rung.
877    pub window: String,
878    /// Re-probe one rung below the promotion every `probe_every` turns.
879    ///
880    /// Without this, promotion is a **ratchet**: a session that escalates once stays expensive for
881    /// its whole life, so every later turn — including the trivial ones — pays the higher rung.
882    /// The periodic probe is the only way a promotion comes back down. `1` re-probes every turn
883    /// (promotion effectively off); large values approach a permanent ratchet.
884    #[serde(default = "default_promotion_probe_every")]
885    pub probe_every: u32,
886    /// Hard cap on tracked sessions; oldest evicted first. A routing optimisation must never be
887    /// able to exhaust memory on a proxy under load.
888    #[serde(default = "default_promotion_max_sessions")]
889    pub max_sessions: usize,
890}
891
892const fn default_promotion_probe_every() -> u32 {
893    // Every 5th turn probes one rung down: quick enough to recover from an over-promotion, rare
894    // enough that the escalation tax is still mostly avoided.
895    5
896}
897
898const fn default_promotion_max_sessions() -> usize {
899    10_000
900}
901
902impl SessionPromotion {
903    /// Parse [`SessionPromotion::window`] into a [`Duration`].
904    ///
905    /// # Errors
906    /// Returns [`Error::BadDuration`] if the window is not `<int><unit>` with unit in `s`/`m`/`h`/`d`.
907    pub fn window_duration(&self) -> Result<Duration> {
908        parse_window(&self.window)
909    }
910}
911
912/// Parse a compact duration like `30m`, `2h`, `90s`, `1d`.
913fn parse_window(s: &str) -> Result<Duration> {
914    let s = s.trim();
915    let split = s
916        .find(|c: char| c.is_ascii_alphabetic())
917        .ok_or_else(|| Error::BadDuration(s.to_owned()))?;
918    let (num, unit) = s.split_at(split);
919    let n: u64 = num
920        .trim()
921        .parse()
922        .map_err(|_| Error::BadDuration(s.to_owned()))?;
923    let secs = match unit {
924        "s" => n,
925        "m" => n.saturating_mul(60),
926        "h" => n.saturating_mul(3600),
927        "d" => n.saturating_mul(86_400),
928        _ => return Err(Error::BadDuration(s.to_owned())),
929    };
930    Ok(Duration::from_secs(secs))
931}
932
933/// A parsed `provider/model` reference.
934#[derive(Debug, Clone, PartialEq, Eq)]
935pub struct ModelRef {
936    /// Provider segment, e.g. `anthropic`.
937    pub provider: String,
938    /// Model segment, e.g. `claude-haiku-4-5`.
939    pub model: String,
940}
941
942impl ModelRef {
943    /// Parse a `provider/model` string.
944    ///
945    /// # Errors
946    /// Returns [`Error::BadModelRef`] if the input is not exactly one `provider/model` pair.
947    pub fn parse(s: &str) -> Result<Self> {
948        match s.split_once('/') {
949            Some((p, m)) if !p.is_empty() && !m.is_empty() && !m.contains('/') => Ok(Self {
950                provider: p.to_owned(),
951                model: m.to_owned(),
952            }),
953            _ => Err(Error::BadModelRef(s.to_owned())),
954        }
955    }
956}
957
958const fn default_guardrail_cooldown() -> i64 {
959    3_600
960}
961
962impl Config {
963    /// Parse a TOML configuration document.
964    ///
965    /// # Errors
966    /// Returns [`Error::Config`] on invalid TOML or unknown fields, or [`Error::InvalidConfig`] on
967    /// an invalid gate definition (not exactly one of `cmd`/`judge`, a judge threshold outside
968    /// `[0,1]`, or a duplicate/blank `id`).
969    pub fn parse(toml_str: &str) -> Result<Self> {
970        let config: Self = toml::from_str(toml_str)?;
971        // A nonsensical rollout percent must fail loudly at parse: silently clamping it would
972        // ramp a different amount of traffic than the operator asked for, and they would only
973        // notice via a savings figure that never moved.
974        for (i, route) in config.routes.iter().enumerate() {
975            if let Some(r) = &route.rollout {
976                r.validate()
977                    .map_err(|e| Error::InvalidConfig(format!("route[{i}]: {e}")))?;
978            }
979            if let Some(sh) = &route.shadow {
980                sh.validate()
981                    .map_err(|e| Error::InvalidConfig(format!("route[{i}]: {e}")))?;
982            }
983        }
984        if let Some(g) = &config.guardrail {
985            g.validate().map_err(Error::InvalidConfig)?;
986        }
987        if let Some(p) = &config.escalation.session_promotion {
988            // `window` is also the idle TTL, so an unparseable one would silently disable expiry.
989            p.window_duration()?;
990            // probe_every = 0 would make promotion one-way: a session that escalated once stays on
991            // the expensive rung for its whole life, including the trivial turns at the end. That
992            // is a cost regression an operator would only find in a bill, so it is rejected here
993            // rather than quietly coerced to something they did not ask for.
994            if p.probe_every == 0 {
995                return Err(Error::InvalidConfig(
996                    "[escalation.session_promotion] probe_every must be >= 1. It is how a promoted \
997                     session ever returns to a cheaper rung; 0 would pin it permanently."
998                        .to_owned(),
999                ));
1000            }
1001            if p.max_sessions == 0 {
1002                return Err(Error::InvalidConfig(
1003                    "[escalation.session_promotion] max_sessions must be >= 1".to_owned(),
1004                ));
1005            }
1006        }
1007        let mut seen = std::collections::HashSet::new();
1008        for def in &config.gate_defs {
1009            if def.id.trim().is_empty() {
1010                return Err(Error::InvalidConfig("gate id must not be empty".to_owned()));
1011            }
1012            // Exactly one kind: `cmd`, `judge`, `consistency`, or `schema` — never more, never none.
1013            let kinds_set = [
1014                !def.cmd.is_empty(),
1015                def.judge.is_some(),
1016                def.consistency.is_some(),
1017                def.schema.is_some(),
1018            ]
1019            .iter()
1020            .filter(|&&b| b)
1021            .count();
1022            if kinds_set != 1 {
1023                return Err(Error::InvalidConfig(format!(
1024                    "gate {:?} must set exactly one of `cmd`, `judge`, `consistency`, or `schema`",
1025                    def.id
1026                )));
1027            }
1028            if let Some(judge) = &def.judge
1029                && !(0.0..=1.0).contains(&judge.threshold)
1030            {
1031                return Err(Error::InvalidConfig(format!(
1032                    "gate {:?} judge threshold {} is outside [0, 1]",
1033                    def.id, judge.threshold
1034                )));
1035            }
1036            if let Some(c) = &def.consistency {
1037                if !(2..=8).contains(&c.k) {
1038                    return Err(Error::InvalidConfig(format!(
1039                        "gate {:?} consistency k {} is outside [2, 8]",
1040                        def.id, c.k
1041                    )));
1042                }
1043                if !(0.0..=1.0).contains(&c.threshold) {
1044                    return Err(Error::InvalidConfig(format!(
1045                        "gate {:?} consistency threshold {} is outside [0, 1]",
1046                        def.id, c.threshold
1047                    )));
1048                }
1049            }
1050            if !seen.insert(def.id.as_str()) {
1051                return Err(Error::InvalidConfig(format!(
1052                    "duplicate gate id {:?}",
1053                    def.id
1054                )));
1055            }
1056        }
1057        for price in &config.price_defs {
1058            if price.model.trim().is_empty() {
1059                return Err(Error::InvalidConfig(
1060                    "price model must not be empty".to_owned(),
1061                ));
1062            }
1063            let ok = |v: f64| v.is_finite() && v >= 0.0;
1064            if !ok(price.input_per_mtok) || !ok(price.output_per_mtok) {
1065                return Err(Error::InvalidConfig(format!(
1066                    "price for {:?} must be finite and >= 0",
1067                    price.model
1068                )));
1069            }
1070        }
1071        config.validate_ladders_are_priced()?;
1072        if let Some(b) = &config.escalation.bandit {
1073            if !b.exploration.is_finite() || b.exploration < 0.0 {
1074                return Err(Error::InvalidConfig(format!(
1075                    "bandit.exploration must be finite and >= 0, got {}",
1076                    b.exploration
1077                )));
1078            }
1079            if !b.discount.is_finite() || b.discount <= 0.0 || b.discount > 1.0 {
1080                return Err(Error::InvalidConfig(format!(
1081                    "bandit.discount must be in (0, 1], got {}",
1082                    b.discount
1083                )));
1084            }
1085        }
1086        if let Some([lo, hi]) = config.escalation.speculation_band {
1087            let ok = |v: f64| v.is_finite() && (0.0..=1.0).contains(&v);
1088            if !ok(lo) || !ok(hi) || lo > hi {
1089                return Err(Error::InvalidConfig(format!(
1090                    "escalation.speculation_band must satisfy 0 <= low <= high <= 1, got [{lo}, {hi}]"
1091                )));
1092            }
1093        }
1094        if let Some(exp) = &config.escalation.exploration
1095            && (!exp.epsilon.is_finite() || exp.epsilon <= 0.0 || exp.epsilon > 0.5)
1096        {
1097            return Err(Error::InvalidConfig(format!(
1098                "escalation.exploration.epsilon must be finite and in (0, 0.5], got {}",
1099                exp.epsilon
1100            )));
1101        }
1102        if let Some(probe) = &config.escalation.probe {
1103            if !(2..=8).contains(&probe.k) {
1104                return Err(Error::InvalidConfig(format!(
1105                    "escalation.probe.k must be in [2, 8], got {}",
1106                    probe.k
1107                )));
1108            }
1109            if !probe.sample_rate.is_finite() || !(0.0..=1.0).contains(&probe.sample_rate) {
1110                return Err(Error::InvalidConfig(format!(
1111                    "escalation.probe.sample_rate must be finite and in [0, 1], got {}",
1112                    probe.sample_rate
1113                )));
1114            }
1115        }
1116        if let Some(pred) = &config.escalation.predictor {
1117            if !pred.lr.is_finite() || !(0.0..=1.0).contains(&pred.lr) || pred.lr == 0.0 {
1118                return Err(Error::InvalidConfig(format!(
1119                    "escalation.predictor.lr must be finite and in (0, 1], got {}",
1120                    pred.lr
1121                )));
1122            }
1123            if !pred.l2.is_finite() || pred.l2 < 0.0 {
1124                return Err(Error::InvalidConfig(format!(
1125                    "escalation.predictor.l2 must be finite and >= 0, got {}",
1126                    pred.l2
1127                )));
1128            }
1129        }
1130        if let Some(elastic) = &config.escalation.elastic {
1131            if elastic.expensive_gates.is_empty() {
1132                return Err(Error::InvalidConfig(
1133                    "escalation.elastic.expensive_gates must name at least one gate to skip".into(),
1134                ));
1135            }
1136            if !elastic.lambda.is_finite() || !(0.0..=1.0).contains(&elastic.lambda) {
1137                return Err(Error::InvalidConfig(format!(
1138                    "escalation.elastic.lambda must be finite and in [0, 1], got {}",
1139                    elastic.lambda
1140                )));
1141            }
1142        }
1143        Ok(config)
1144    }
1145
1146    /// The first route whose match claims `f`, or `None` if no route matches.
1147    #[must_use]
1148    pub fn route_for(&self, f: &Features) -> Option<&Route> {
1149        self.routes.iter().find(|r| r.match_.matches(f))
1150    }
1151
1152    /// The effective price table: built-in defaults with this config's `[[price]]` blocks applied.
1153    #[must_use]
1154    pub fn price_table(&self) -> crate::cost::PriceTable {
1155        self.price_defs
1156            .iter()
1157            .fold(crate::cost::PriceTable::defaults(), |table, p| {
1158                table.with_override(
1159                    p.model.clone(),
1160                    crate::cost::ModelPrice {
1161                        input_per_mtok: p.input_per_mtok,
1162                        output_per_mtok: p.output_per_mtok,
1163                    },
1164                )
1165            })
1166    }
1167
1168    /// Every ladder rung must resolve to a price.
1169    ///
1170    /// An unpriced rung is not a missing number, it is a **fabricated one**: the router falls
1171    /// back to `$0.00` for that call, so the receipt reports a spend nobody measured and a
1172    /// `[budget]` cap fed by that total can never trip. The built-in table only knows the
1173    /// first-party models, so this is precisely the bring-your-own-provider path.
1174    ///
1175    /// A genuinely free rung (a local model you host) is still legal — declare it as an
1176    /// explicit `[[price]]` of `0`, so the receipt's `$0.00` is a stated fact rather than a
1177    /// silent fallback.
1178    fn validate_ladders_are_priced(&self) -> Result<()> {
1179        let prices = self.price_table();
1180        for (i, route) in self.routes.iter().enumerate() {
1181            for rung in &route.ladder {
1182                if prices.get(rung).is_none() {
1183                    return Err(Error::InvalidConfig(format!(
1184                        "route[{i}] ladder rung {rung:?} has no price, so its cost would be \
1185                         recorded as $0.00 and any [budget] cap would never trip. Declare it:\n\n\
1186                         [[price]]\nmodel = {rung:?}\ninput_per_mtok = 0.0\noutput_per_mtok = 0.0\n\n\
1187                         (use the real per-1M-token prices; 0.0 is correct only for a model you \
1188                         host yourself)"
1189                    )));
1190                }
1191            }
1192        }
1193        Ok(())
1194    }
1195}
1196
1197#[cfg(test)]
1198mod tests {
1199    use super::*;
1200    use crate::features::Features;
1201
1202    const SPEC_CONFIG: &str = r#"
1203[[route]]
1204match = { agent = "claude-code", subagent = ["test-runner", "explore"] }
1205mode  = "enforce"
1206ladder = ["anthropic/claude-haiku-4-5", "anthropic/claude-sonnet-5"]
1207gates  = ["schema", "judge-diff"]
1208
1209[[route]]
1210match = { task_kind = "code_edit" }
1211mode  = "enforce"
1212ladder = ["anthropic/claude-sonnet-5", "anthropic/claude-opus-4-8"]
1213gates  = ["patch-applies", "lint-diff", "judge-diff"]
1214deferred_gates = ["compiles", "tests"]
1215
1216[[route]]
1217match = {}
1218mode  = "observe"
1219ladder = ["anthropic/claude-opus-4-8"]
1220
1221[budget]
1222per_request_usd = 0.50
1223per_session_usd = 10.00
1224per_day_usd     = 250.00
1225on_exhausted    = "serve_best_attempt"
1226
1227[escalation]
1228max_rungs_per_request = 3
1229session_promotion = { after_failures = 3, window = "30m" }
1230"#;
1231
1232    #[test]
1233    fn parses_the_spec_example() {
1234        let c = Config::parse(SPEC_CONFIG).unwrap();
1235        assert_eq!(c.routes.len(), 3);
1236        assert_eq!(c.routes[0].mode, Mode::Enforce);
1237        assert_eq!(
1238            c.routes[0].ladder,
1239            ["anthropic/claude-haiku-4-5", "anthropic/claude-sonnet-5"]
1240        );
1241        assert_eq!(c.routes[1].deferred_gates, ["compiles", "tests"]);
1242        assert_eq!(c.budget.per_request_usd, Some(0.50));
1243        assert_eq!(c.budget.on_exhausted, OnExhausted::ServeBestAttempt);
1244        assert_eq!(c.escalation.max_rungs_per_request, 3);
1245        let sp = c.escalation.session_promotion.as_ref().unwrap();
1246        assert_eq!(sp.after_failures, 3);
1247        assert_eq!(sp.window_duration().unwrap(), Duration::from_secs(1800));
1248    }
1249
1250    #[test]
1251    fn first_matching_route_wins() {
1252        let c = Config::parse(SPEC_CONFIG).unwrap();
1253
1254        // claude-code / test-runner -> route 0
1255        let mut f = Features::new(TaskKind::Explore);
1256        f.agent = Some("claude-code".into());
1257        f.subagent = Some("test-runner".into());
1258        let r = c.route_for(&f).unwrap();
1259        assert_eq!(r.ladder[0], "anthropic/claude-haiku-4-5");
1260
1261        // a code_edit with no agent -> route 1 (route 0's agent constraint fails)
1262        let f2 = Features::new(TaskKind::CodeEdit);
1263        let r2 = c.route_for(&f2).unwrap();
1264        assert_eq!(r2.gates, ["patch-applies", "lint-diff", "judge-diff"]);
1265
1266        // anything else -> catch-all observe route
1267        let f3 = Features::new(TaskKind::Chat);
1268        let r3 = c.route_for(&f3).unwrap();
1269        assert_eq!(r3.mode, Mode::Observe);
1270    }
1271
1272    #[test]
1273    fn shipped_example_config_parses() {
1274        // The repo ships `firstpass.example.toml` for users to copy. If it drifts
1275        // from the schema, this fails in CI rather than at a user's first run.
1276        let toml = include_str!("../../../firstpass.example.toml");
1277        let c = Config::parse(toml).expect("firstpass.example.toml must parse");
1278        assert_eq!(c.routes.len(), 3);
1279        assert_eq!(c.routes[0].mode, Mode::Enforce);
1280    }
1281
1282    #[test]
1283    fn parses_gate_definitions() {
1284        let toml = r#"
1285[[route]]
1286match = {}
1287mode  = "enforce"
1288ladder = ["anthropic/claude-haiku-4-5"]
1289gates  = ["my-tests"]
1290
1291[[gate]]
1292id  = "my-tests"
1293cmd = ["pytest", "-q"]
1294
1295[[gate]]
1296id         = "judge"
1297cmd        = ["bash", "-c", "./judge.sh"]
1298timeout_ms = 60000
1299"#;
1300        let c = Config::parse(toml).unwrap();
1301        assert_eq!(c.gate_defs.len(), 2);
1302        assert_eq!(c.gate_defs[0].id, "my-tests");
1303        assert_eq!(c.gate_defs[0].cmd, ["pytest", "-q"]);
1304        assert_eq!(c.gate_defs[0].timeout_ms, 30_000, "default timeout applies");
1305        assert_eq!(c.gate_defs[1].timeout_ms, 60_000);
1306    }
1307
1308    #[test]
1309    fn rejects_invalid_gate_definitions() {
1310        let empty_cmd = "[[gate]]\nid = \"g\"\ncmd = []\n";
1311        assert!(matches!(
1312            Config::parse(empty_cmd),
1313            Err(Error::InvalidConfig(_))
1314        ));
1315
1316        let dup = "[[gate]]\nid = \"g\"\ncmd = [\"a\"]\n[[gate]]\nid = \"g\"\ncmd = [\"b\"]\n";
1317        assert!(matches!(Config::parse(dup), Err(Error::InvalidConfig(_))));
1318    }
1319
1320    #[test]
1321    fn parses_consistency_gate_definition() {
1322        let toml = r#"
1323[[gate]]
1324id          = "uncertainty"
1325consistency = { model = "anthropic/claude-haiku-4-5", k = 5, threshold = 0.6 }
1326"#;
1327        let c = Config::parse(toml).unwrap();
1328        assert_eq!(c.gate_defs.len(), 1);
1329        let cons = c.gate_defs[0].consistency.as_ref().unwrap();
1330        assert_eq!(cons.model, "anthropic/claude-haiku-4-5");
1331        assert_eq!(cons.k, 5);
1332        assert!((cons.threshold - 0.6).abs() < 1e-9);
1333    }
1334
1335    #[test]
1336    fn consistency_k_defaults_to_3() {
1337        let toml = "[[gate]]\nid = \"u\"\nconsistency = { model = \"anthropic/claude-haiku-4-5\", threshold = 0.7 }\n";
1338        let c = Config::parse(toml).unwrap();
1339        assert_eq!(c.gate_defs[0].consistency.as_ref().unwrap().k, 3);
1340    }
1341
1342    #[test]
1343    fn rejects_exactly_one_of_violations_for_consistency() {
1344        // cmd + consistency — two kinds set
1345        let both = "[[gate]]\nid = \"g\"\ncmd = [\"x\"]\nconsistency = { model = \"a/b\", threshold = 0.5 }\n";
1346        assert!(matches!(Config::parse(both), Err(Error::InvalidConfig(_))));
1347
1348        // consistency k out of bounds
1349        let bad_k =
1350            "[[gate]]\nid = \"g\"\nconsistency = { model = \"a/b\", k = 1, threshold = 0.5 }\n";
1351        assert!(matches!(Config::parse(bad_k), Err(Error::InvalidConfig(_))));
1352
1353        let bad_k2 =
1354            "[[gate]]\nid = \"g\"\nconsistency = { model = \"a/b\", k = 9, threshold = 0.5 }\n";
1355        assert!(matches!(
1356            Config::parse(bad_k2),
1357            Err(Error::InvalidConfig(_))
1358        ));
1359
1360        // consistency threshold out of bounds
1361        let bad_thresh =
1362            "[[gate]]\nid = \"g\"\nconsistency = { model = \"a/b\", threshold = 1.1 }\n";
1363        assert!(matches!(
1364            Config::parse(bad_thresh),
1365            Err(Error::InvalidConfig(_))
1366        ));
1367    }
1368
1369    #[test]
1370    fn parses_adaptive_conformal_config() {
1371        let c = Config::parse("[escalation.adaptive]\nalpha = 0.1\n").unwrap();
1372        let a = c
1373            .escalation
1374            .adaptive
1375            .expect("[escalation.adaptive] should parse");
1376        assert!((a.alpha - 0.1).abs() < 1e-9);
1377        assert!((a.gamma - 0.02).abs() < 1e-9, "gamma defaults to 0.02");
1378        // Absent => None (fixed-threshold serving, default byte-identical behavior).
1379        assert!(Config::parse("").unwrap().escalation.adaptive.is_none());
1380    }
1381
1382    #[test]
1383    fn parses_provider_entries_and_ladders_can_reference_them() {
1384        let c = Config::parse(
1385            r#"
1386[[provider]]
1387id = "groq"
1388dialect = "openai"
1389base_url = "https://api.groq.com/openai"
1390api_key_env = "GROQ_API_KEY"
1391
1392[[provider]]
1393id = "ollama"
1394dialect = "openai"
1395base_url = "http://localhost:11434"
1396
1397# A bring-your-own provider has no built-in price, so the ladder must declare one or the
1398# receipt would report $0.00 for every Groq call.
1399[[price]]
1400model = "groq/llama-3.3-70b-versatile"
1401input_per_mtok = 0.59
1402output_per_mtok = 0.79
1403
1404[[route]]
1405match = {}
1406mode = "enforce"
1407ladder = ["groq/llama-3.3-70b-versatile", "anthropic/claude-sonnet-5"]
1408"#,
1409        )
1410        .unwrap();
1411        assert_eq!(c.providers.len(), 2);
1412        let groq = &c.providers[0];
1413        assert_eq!(groq.id, "groq");
1414        assert_eq!(groq.dialect, Dialect::Openai);
1415        assert_eq!(groq.base_url, "https://api.groq.com/openai");
1416        assert_eq!(groq.api_key_env.as_deref(), Some("GROQ_API_KEY"));
1417        // A keyless local endpoint (Ollama) parses with no api_key_env.
1418        assert_eq!(c.providers[1].id, "ollama");
1419        assert!(c.providers[1].api_key_env.is_none());
1420        // Absent => no extra providers (built-ins only).
1421        assert!(Config::parse("").unwrap().providers.is_empty());
1422    }
1423
1424    #[test]
1425    fn parses_aws_sigv4_provider_and_defaults_auth_to_api_key() {
1426        let c = Config::parse(
1427            r#"
1428[[provider]]
1429id = "bedrock"
1430dialect = "anthropic"
1431auth = "aws_sigv4"
1432region = "us-east-1"
1433"#,
1434        )
1435        .unwrap();
1436        let bedrock = &c.providers[0];
1437        assert_eq!(bedrock.auth, AuthScheme::AwsSigv4);
1438        assert_eq!(bedrock.region.as_deref(), Some("us-east-1"));
1439        assert!(bedrock.project.is_none());
1440
1441        // Omitting `auth` on an existing provider entry defaults to ApiKey (I1: no behavior
1442        // change for today's `[[provider]]` entries).
1443        let c2 = Config::parse(
1444            r#"
1445[[provider]]
1446id = "groq"
1447dialect = "openai"
1448base_url = "https://api.groq.com/openai"
1449"#,
1450        )
1451        .unwrap();
1452        assert_eq!(c2.providers[0].auth, AuthScheme::ApiKey);
1453    }
1454
1455    #[test]
1456    fn all_documented_provider_shapes_parse() {
1457        // The exact `[[provider]]` shapes shown in the README / usage page / firstpass.example.toml.
1458        // This is the guard that the documented provider instructions stay valid — one entry per
1459        // dialect/auth combination we tell users to write.
1460        let c = Config::parse(
1461            r#"
1462[[provider]]
1463id = "groq"
1464dialect = "openai"
1465base_url = "https://api.groq.com/openai"
1466api_key_env = "GROQ_API_KEY"
1467
1468[[provider]]
1469id = "ollama"
1470dialect = "openai"
1471base_url = "http://localhost:11434"
1472
1473[[provider]]
1474id = "gemini"
1475dialect = "gemini"
1476base_url = "https://generativelanguage.googleapis.com"
1477api_key_env = "GEMINI_API_KEY"
1478
1479[[provider]]
1480id = "bedrock"
1481dialect = "anthropic"
1482auth = "aws_sigv4"
1483region = "us-east-1"
1484
1485[[provider]]
1486id = "vertex"
1487dialect = "anthropic"
1488auth = "gcp_oauth"
1489region = "us-east5"
1490project = "my-gcp-project"
1491"#,
1492        )
1493        .expect("every documented provider shape must parse");
1494        assert_eq!(c.providers.len(), 5);
1495
1496        let gemini = c.providers.iter().find(|p| p.id == "gemini").unwrap();
1497        assert_eq!(gemini.dialect, Dialect::Gemini);
1498        assert_eq!(gemini.auth, AuthScheme::ApiKey);
1499
1500        let vertex = c.providers.iter().find(|p| p.id == "vertex").unwrap();
1501        assert_eq!(vertex.auth, AuthScheme::GcpOauth);
1502        assert_eq!(vertex.region.as_deref(), Some("us-east5"));
1503        assert_eq!(vertex.project.as_deref(), Some("my-gcp-project"));
1504    }
1505
1506    #[test]
1507    fn empty_match_is_wildcard() {
1508        let m = Match::default();
1509        assert!(m.matches(&Features::new(TaskKind::Other)));
1510    }
1511
1512    #[test]
1513    fn subagent_list_membership() {
1514        let c = Config::parse(SPEC_CONFIG).unwrap();
1515        let route0 = &c.routes[0];
1516        let mut f = Features::new(TaskKind::Other);
1517        f.agent = Some("claude-code".into());
1518        f.subagent = Some("docs-writer".into()); // not in [test-runner, explore]
1519        assert!(!route0.match_.matches(&f));
1520        f.subagent = Some("explore".into());
1521        assert!(route0.match_.matches(&f));
1522    }
1523
1524    #[test]
1525    fn model_ref_parsing() {
1526        let m = ModelRef::parse("anthropic/claude-haiku-4-5").unwrap();
1527        assert_eq!(m.provider, "anthropic");
1528        assert_eq!(m.model, "claude-haiku-4-5");
1529        assert!(ModelRef::parse("no-slash").is_err());
1530        assert!(ModelRef::parse("/model").is_err());
1531        assert!(ModelRef::parse("a/b/c").is_err());
1532    }
1533
1534    #[test]
1535    fn window_parsing_units_and_errors() {
1536        assert_eq!(parse_window("90s").unwrap(), Duration::from_secs(90));
1537        assert_eq!(parse_window("30m").unwrap(), Duration::from_secs(1800));
1538        assert_eq!(parse_window("2h").unwrap(), Duration::from_secs(7200));
1539        assert_eq!(parse_window("1d").unwrap(), Duration::from_secs(86_400));
1540        assert!(parse_window("30x").is_err());
1541        assert!(parse_window("abc").is_err());
1542    }
1543
1544    #[test]
1545    fn empty_config_defaults() {
1546        let c = Config::parse("").unwrap();
1547        assert!(c.routes.is_empty());
1548        assert_eq!(c.escalation.max_rungs_per_request, 3);
1549        assert_eq!(c.budget.on_exhausted, OnExhausted::ServeBestAttempt);
1550    }
1551
1552    // ── ExplorationConfig parse / validation ──────────────────────────────────
1553
1554    #[test]
1555    fn parses_exploration_config() {
1556        let c = Config::parse("[escalation.exploration]\nepsilon = 0.1\n").unwrap();
1557        let exp = c
1558            .escalation
1559            .exploration
1560            .expect("[escalation.exploration] should parse");
1561        assert!((exp.epsilon - 0.1).abs() < 1e-12);
1562        // Absent => None (deterministic policy, byte-identical behavior).
1563        assert!(Config::parse("").unwrap().escalation.exploration.is_none());
1564    }
1565
1566    #[test]
1567    fn exploration_epsilon_boundary_valid() {
1568        // Upper bound 0.5 is allowed.
1569        let c = Config::parse("[escalation.exploration]\nepsilon = 0.5\n").unwrap();
1570        assert!((c.escalation.exploration.unwrap().epsilon - 0.5).abs() < 1e-12);
1571    }
1572
1573    #[test]
1574    fn exploration_epsilon_above_half_rejected() {
1575        let bad = "[escalation.exploration]\nepsilon = 0.51\n";
1576        assert!(
1577            matches!(Config::parse(bad), Err(Error::InvalidConfig(_))),
1578            "epsilon > 0.5 must be rejected"
1579        );
1580    }
1581
1582    #[test]
1583    fn exploration_epsilon_zero_rejected() {
1584        let bad = "[escalation.exploration]\nepsilon = 0.0\n";
1585        assert!(
1586            matches!(Config::parse(bad), Err(Error::InvalidConfig(_))),
1587            "epsilon = 0 must be rejected (must be strictly positive)"
1588        );
1589    }
1590
1591    #[test]
1592    fn exploration_epsilon_negative_rejected() {
1593        let bad = "[escalation.exploration]\nepsilon = -0.1\n";
1594        assert!(
1595            matches!(Config::parse(bad), Err(Error::InvalidConfig(_))),
1596            "epsilon < 0 must be rejected"
1597        );
1598    }
1599
1600    #[test]
1601    fn gate_def_schema_and_on_abstain_parse() {
1602        let toml = r#"
1603[[route]]
1604match = {}
1605mode = "enforce"
1606ladder = ["anthropic/claude-haiku-4-5"]
1607gates = ["extract-shape"]
1608
1609[[gate]]
1610id = "extract-shape"
1611schema = { type = "object", required = ["name"] }
1612on_abstain = "fail_closed"
1613"#;
1614        let config = Config::parse(toml).expect("schema gate def must parse");
1615        let def = &config.gate_defs[0];
1616        assert_eq!(def.id, "extract-shape");
1617        let schema = def.schema.as_ref().expect("schema captured");
1618        assert_eq!(schema["type"], "object");
1619        assert_eq!(def.on_abstain, AbstainPolicy::FailClosed);
1620    }
1621
1622    #[test]
1623    fn gate_def_on_abstain_defaults_fail_open() {
1624        let toml = r#"
1625[[route]]
1626match = {}
1627mode = "enforce"
1628ladder = ["anthropic/claude-haiku-4-5"]
1629
1630[[gate]]
1631id = "tests"
1632cmd = ["true"]
1633"#;
1634        let config = Config::parse(toml).expect("parse");
1635        assert_eq!(config.gate_defs[0].on_abstain, AbstainPolicy::FailOpen);
1636    }
1637
1638    #[test]
1639    fn gate_def_rejects_schema_plus_cmd() {
1640        let toml = r#"
1641[[route]]
1642match = {}
1643mode = "enforce"
1644ladder = ["anthropic/claude-haiku-4-5"]
1645
1646[[gate]]
1647id = "both"
1648cmd = ["true"]
1649schema = { type = "object" }
1650"#;
1651        let err = Config::parse(toml).unwrap_err();
1652        assert!(
1653            err.to_string().contains("exactly one"),
1654            "two kinds must be rejected: {err}"
1655        );
1656    }
1657
1658    #[test]
1659    fn price_overrides_parse_and_validate() {
1660        let toml = r#"
1661[[route]]
1662match = {}
1663mode = "observe"
1664ladder = ["anthropic/claude-haiku-4-5"]
1665
1666[[price]]
1667model = "anthropic/claude-haiku-4-5"
1668input_per_mtok = 0.8
1669output_per_mtok = 4.0
1670"#;
1671        let config = Config::parse(toml).expect("price override must parse");
1672        assert_eq!(config.price_defs[0].model, "anthropic/claude-haiku-4-5");
1673        assert!((config.price_defs[0].input_per_mtok - 0.8).abs() < 1e-12);
1674
1675        let bad = toml.replace("input_per_mtok = 0.8", "input_per_mtok = -1.0");
1676        assert!(Config::parse(&bad).is_err(), "negative price rejected");
1677    }
1678
1679    // ── RoutingMode / ModePreset ──────────────────────────────────────────────
1680
1681    /// The Balanced preset must be a strict no-op: all overrides None/false.
1682    /// This is the invariant that guarantees byte-identical behaviour when no mode is set.
1683    #[test]
1684    fn balanced_preset_is_strict_noop() {
1685        let p = RoutingMode::Balanced.preset();
1686        assert!(
1687            p.speculation.is_none(),
1688            "Balanced must not override speculation"
1689        );
1690        assert!(
1691            p.max_rungs_delta.is_none(),
1692            "Balanced must not override max_rungs"
1693        );
1694        assert!(!p.start_at_top, "Balanced must not set start_at_top");
1695    }
1696
1697    #[test]
1698    fn cost_preset_disables_speculation() {
1699        let p = RoutingMode::Cost.preset();
1700        assert_eq!(p.speculation, Some(0));
1701        assert!(p.max_rungs_delta.is_none());
1702        assert!(!p.start_at_top);
1703    }
1704
1705    #[test]
1706    fn quality_preset_bumps_max_rungs_and_disables_speculation() {
1707        let p = RoutingMode::Quality.preset();
1708        assert_eq!(p.max_rungs_delta, Some(1));
1709        assert_eq!(p.speculation, Some(0));
1710        assert!(!p.start_at_top);
1711    }
1712
1713    #[test]
1714    fn latency_preset_enables_speculation() {
1715        let p = RoutingMode::Latency.preset();
1716        assert_eq!(p.speculation, Some(1));
1717        assert!(p.max_rungs_delta.is_none());
1718        assert!(!p.start_at_top);
1719    }
1720
1721    #[test]
1722    fn max_preset_sets_start_at_top_and_disables_speculation() {
1723        let p = RoutingMode::Max.preset();
1724        assert!(p.start_at_top);
1725        assert_eq!(p.speculation, Some(0));
1726        assert!(p.max_rungs_delta.is_none());
1727    }
1728
1729    #[test]
1730    fn routing_mode_as_str_roundtrips() {
1731        for mode in RoutingMode::ALL {
1732            let s = mode.as_str();
1733            let back: RoutingMode = serde_json::from_str(&format!("\"{s}\"")).unwrap();
1734            assert_eq!(*mode, back, "as_str/serde roundtrip for {s}");
1735        }
1736    }
1737
1738    #[test]
1739    fn routing_mode_defaults_to_balanced() {
1740        assert_eq!(RoutingMode::default(), RoutingMode::Balanced);
1741    }
1742
1743    #[test]
1744    fn route_routing_mode_parses_and_defaults_to_none() {
1745        // Absent → None (byte-identical, no behavioral change).
1746        let no_mode = Config::parse(
1747            "[[route]]\nmatch = {}\nmode = \"enforce\"\nladder = [\"anthropic/claude-haiku-4-5\"]\n",
1748        )
1749        .unwrap();
1750        assert_eq!(no_mode.routes[0].routing_mode, None);
1751
1752        // Explicit cost mode.
1753        let with_mode = Config::parse(
1754            "[[route]]\nmatch = {}\nmode = \"enforce\"\nladder = [\"anthropic/claude-haiku-4-5\"]\nrouting_mode = \"cost\"\n",
1755        )
1756        .unwrap();
1757        assert_eq!(with_mode.routes[0].routing_mode, Some(RoutingMode::Cost));
1758    }
1759
1760    #[test]
1761    fn all_routing_modes_have_non_empty_description_and_tradeoff() {
1762        for mode in RoutingMode::ALL {
1763            let p = mode.preset();
1764            assert!(
1765                !p.description.is_empty(),
1766                "mode {} has empty description",
1767                mode.as_str()
1768            );
1769            assert!(
1770                !p.tradeoff.is_empty(),
1771                "mode {} has empty tradeoff",
1772                mode.as_str()
1773            );
1774        }
1775    }
1776
1777    #[test]
1778    fn bandit_thompson_and_band_parse_and_validate() {
1779        let toml = r#"
1780[[route]]
1781match = {}
1782mode = "enforce"
1783ladder = ["anthropic/claude-haiku-4-5"]
1784
1785[escalation]
1786speculation = 1
1787speculation_band = [0.3, 0.7]
1788
1789[escalation.bandit]
1790algorithm = "thompson"
1791discount = 0.98
1792"#;
1793        let config = Config::parse(toml).expect("thompson + band must parse");
1794        let b = config.escalation.bandit.as_ref().unwrap();
1795        assert_eq!(b.algorithm, BanditAlgorithm::Thompson);
1796        assert!((b.discount - 0.98).abs() < 1e-12);
1797        assert_eq!(config.escalation.speculation_band, Some([0.3, 0.7]));
1798
1799        // Defaults: ucb1, discount 1.0, no band.
1800        let plain = Config::parse(
1801            "[[route]]\nmatch = {}\nmode = \"enforce\"\nladder = [\"anthropic/claude-haiku-4-5\"]\n[escalation.bandit]\n",
1802        )
1803        .unwrap();
1804        let b = plain.escalation.bandit.as_ref().unwrap();
1805        assert_eq!(b.algorithm, BanditAlgorithm::Ucb1);
1806        assert!((b.discount - 1.0).abs() < 1e-12);
1807
1808        // Bad discount and bad band are rejected.
1809        assert!(Config::parse(&toml.replace("discount = 0.98", "discount = 0.0")).is_err());
1810        assert!(
1811            Config::parse(&toml.replace(
1812                "speculation_band = [0.3, 0.7]",
1813                "speculation_band = [0.9, 0.2]"
1814            ))
1815            .is_err()
1816        );
1817    }
1818
1819    // ── ProbeConfig parse / validation ───────────────────────────────────────
1820
1821    #[test]
1822    fn probe_absent_defaults_to_none() {
1823        // Default off = byte-identical to today.
1824        let c = Config::parse("").unwrap();
1825        assert!(c.escalation.probe.is_none());
1826    }
1827
1828    #[test]
1829    fn parses_valid_probe_config() {
1830        let c = Config::parse("[escalation.probe]\nk = 5\nsample_rate = 0.1\n").unwrap();
1831        let p = c.escalation.probe.expect("[escalation.probe] should parse");
1832        assert_eq!(p.k, 5);
1833        assert!((p.sample_rate - 0.1).abs() < 1e-12);
1834    }
1835
1836    #[test]
1837    fn probe_sample_rate_boundaries_accepted() {
1838        // 0.0 (never probe) and 1.0 (always probe) are both valid.
1839        let c0 = Config::parse("[escalation.probe]\nk = 2\nsample_rate = 0.0\n").unwrap();
1840        assert!((c0.escalation.probe.unwrap().sample_rate - 0.0).abs() < 1e-12);
1841        let c1 = Config::parse("[escalation.probe]\nk = 8\nsample_rate = 1.0\n").unwrap();
1842        assert!((c1.escalation.probe.unwrap().sample_rate - 1.0).abs() < 1e-12);
1843    }
1844
1845    #[test]
1846    fn probe_rejects_k_below_2() {
1847        let bad = "[escalation.probe]\nk = 1\nsample_rate = 0.5\n";
1848        assert!(
1849            matches!(Config::parse(bad), Err(Error::InvalidConfig(_))),
1850            "k=1 must be rejected"
1851        );
1852    }
1853
1854    #[test]
1855    fn probe_rejects_k_above_8() {
1856        let bad = "[escalation.probe]\nk = 9\nsample_rate = 0.5\n";
1857        assert!(
1858            matches!(Config::parse(bad), Err(Error::InvalidConfig(_))),
1859            "k=9 must be rejected"
1860        );
1861    }
1862
1863    #[test]
1864    fn probe_rejects_sample_rate_above_1() {
1865        let bad = "[escalation.probe]\nk = 5\nsample_rate = 1.5\n";
1866        assert!(
1867            matches!(Config::parse(bad), Err(Error::InvalidConfig(_))),
1868            "sample_rate=1.5 must be rejected"
1869        );
1870    }
1871
1872    #[test]
1873    fn probe_rejects_negative_sample_rate() {
1874        let bad = "[escalation.probe]\nk = 5\nsample_rate = -0.1\n";
1875        assert!(
1876            matches!(Config::parse(bad), Err(Error::InvalidConfig(_))),
1877            "negative sample_rate must be rejected"
1878        );
1879    }
1880
1881    #[test]
1882    fn predictor_config_parses_and_validates() {
1883        let base = "[[route]]\nmatch = {}\nmode = \"enforce\"\nladder = [\"anthropic/claude-haiku-4-5\"]\n[escalation.predictor]\nlr = 0.05\nl2 = 0.001\n";
1884        let cfg = Config::parse(base).expect("valid predictor must parse");
1885        let pred = cfg.escalation.predictor.unwrap();
1886        assert!((pred.lr - 0.05).abs() < 1e-12 && (pred.l2 - 0.001).abs() < 1e-12);
1887        // defaults when omitted
1888        let d = Config::parse("[[route]]\nmatch = {}\nmode = \"enforce\"\nladder = [\"anthropic/claude-haiku-4-5\"]\n[escalation.predictor]\n").unwrap();
1889        assert!(d.escalation.predictor.is_some());
1890        // bad lr / l2 rejected
1891        assert!(Config::parse(&base.replace("lr = 0.05", "lr = 0.0")).is_err());
1892        assert!(Config::parse(&base.replace("lr = 0.05", "lr = 1.5")).is_err());
1893        assert!(Config::parse(&base.replace("l2 = 0.001", "l2 = -1.0")).is_err());
1894    }
1895
1896    /// An unpriced ladder rung must be rejected at parse. The router prices a call with
1897    /// `cost_usd(..).unwrap_or(0.0)`, so accepting one would put a spend nobody measured into a
1898    /// tamper-evident receipt and leave every `[budget]` cap permanently un-trippable — the two
1899    /// failures this product can least afford.
1900    #[test]
1901    fn unpriced_ladder_rung_is_rejected_and_an_explicit_price_fixes_it() {
1902        let unpriced =
1903            "[[route]]\nmatch = {}\nmode = \"enforce\"\nladder = [\"ollama/qwen2.5-coder:7b\"]\n";
1904        let err = Config::parse(unpriced).expect_err("an unpriced rung must not parse");
1905        let msg = err.to_string();
1906        assert!(
1907            msg.contains("ollama/qwen2.5-coder:7b") && msg.contains("[[price]]"),
1908            "the error must name the rung and the fix, got: {msg}"
1909        );
1910
1911        // A self-hosted model really is free — but it has to SAY so.
1912        let declared = format!(
1913            "[[price]]\nmodel = \"ollama/qwen2.5-coder:7b\"\ninput_per_mtok = 0.0\noutput_per_mtok = 0.0\n{unpriced}"
1914        );
1915        let cfg = Config::parse(&declared).expect("an explicit zero price is legal");
1916        assert_eq!(
1917            cfg.price_table()
1918                .get("ollama/qwen2.5-coder:7b")
1919                .unwrap()
1920                .input_per_mtok,
1921            0.0
1922        );
1923
1924        // The built-in first-party models still need no ceremony.
1925        assert!(
1926            Config::parse("[[route]]\nmatch = {}\nmode = \"enforce\"\nladder = [\"anthropic/claude-haiku-4-5\"]\n").is_ok()
1927        );
1928    }
1929
1930    /// Shadow is off unless asked for, and a nonsensical setting must fail loudly at parse rather
1931    /// than being clamped — a silently-corrected ceiling would spend a different amount than the
1932    /// operator authorised.
1933    #[test]
1934    fn shadow_config_parses_and_validates() {
1935        let base = "[[route]]\nmatch = {}\nmode = \"observe\"\nladder = [\"anthropic/claude-haiku-4-5\"]\n";
1936        let none = Config::parse(base).unwrap();
1937        assert!(
1938            none.routes[0].shadow.is_none(),
1939            "shadow must be off by default"
1940        );
1941
1942        let ok = Config::parse(&format!(
1943            "{base}[route.shadow]\nsample_rate = 0.25\nmax_usd_per_day = 5.0\n"
1944        ))
1945        .unwrap();
1946        let sh = ok.routes[0].shadow.expect("shadow parsed");
1947        assert!((sh.sample_rate - 0.25).abs() < f64::EPSILON);
1948        assert!((sh.max_usd_per_day - 5.0).abs() < f64::EPSILON);
1949
1950        for bad in [
1951            "sample_rate = 1.5\nmax_usd_per_day = 5.0",
1952            "sample_rate = -0.1\nmax_usd_per_day = 5.0",
1953            "sample_rate = 0.1\nmax_usd_per_day = -1.0",
1954        ] {
1955            assert!(
1956                Config::parse(&format!("{base}[route.shadow]\n{bad}\n")).is_err(),
1957                "invalid shadow config accepted: {bad}"
1958            );
1959        }
1960    }
1961
1962    /// Same for rollout: absent by default, invalid percent rejected at parse.
1963    #[test]
1964    fn rollout_config_parses_and_validates() {
1965        let base = "[[route]]\nmatch = {}\nmode = \"enforce\"\nladder = [\"anthropic/claude-haiku-4-5\"]\n";
1966        assert!(Config::parse(base).unwrap().routes[0].rollout.is_none());
1967        let ok = Config::parse(&format!(
1968            "{base}[route.rollout]\npercent = 5.0\nkey = \"session\"\n"
1969        ))
1970        .unwrap();
1971        assert!((ok.routes[0].rollout.unwrap().percent - 5.0).abs() < f64::EPSILON);
1972        assert!(Config::parse(&format!("{base}[route.rollout]\npercent = 140.0\n")).is_err());
1973    }
1974}