cortex-router 0.1.0

Fail-closed, model-free routing policy that decides between deterministic tools, local models, and an upstream agent.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
#![doc = include_str!("../README.md")]

mod classifier;
mod fold;
mod lexicon;

use serde::{Deserialize, Serialize};

pub use classifier::{classify, detector_disagreement};
pub use cortex_domain::{ExecutionTarget, RiskLevel};
pub use fold::{fold_text, fold_words, mixed_script};

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum TaskClass {
    Deterministic,
    RepositoryAnalysis,
    StructuredExtraction,
    ContextCompression,
    AdvisoryDraft,
    Implementation,
    Security,
    Authentication,
    Concurrency,
    Migration,
    Release,
    Deployment,
    Publication,
    Ambiguous,
}

impl TaskClass {
    #[must_use]
    pub const fn requires_verified_evidence(self) -> bool {
        matches!(self, Self::ContextCompression)
    }
}

impl TaskClass {
    #[must_use]
    pub const fn is_high_risk(self) -> bool {
        matches!(
            self,
            Self::Security
                | Self::Authentication
                | Self::Concurrency
                | Self::Migration
                | Self::Release
                | Self::Deployment
                | Self::Publication
        )
    }
}

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum EvidenceStatus {
    NotRequired,
    Verified,
    Missing,
    Contradictory,
}

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum MutationStatus {
    None,
    Approved,
    ApprovalRequired,
}

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "camelCase")]
pub struct TokenBudget {
    pub estimated_input_tokens: u32,
    pub estimated_output_tokens: u32,
    pub max_input_tokens: u32,
    pub max_output_tokens: u32,
}

impl Default for TokenBudget {
    fn default() -> Self {
        Self {
            estimated_input_tokens: 0,
            estimated_output_tokens: 0,
            max_input_tokens: 8_192,
            max_output_tokens: 1_024,
        }
    }
}

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "camelCase")]
pub struct LocalAvailability {
    pub weavatrix: bool,
    pub ollama: bool,
}

impl Default for LocalAvailability {
    fn default() -> Self {
        Self {
            weavatrix: true,
            ollama: true,
        }
    }
}

#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "camelCase")]
pub struct RoutingRequest {
    pub task: String,
    pub evidence: EvidenceStatus,
    pub schema_valid: bool,
    pub budget: TokenBudget,
    pub mutation: MutationStatus,
    pub availability: LocalAvailability,
}

impl RoutingRequest {
    #[must_use]
    pub fn new(task: impl Into<String>) -> Self {
        Self {
            task: task.into(),
            evidence: EvidenceStatus::NotRequired,
            schema_valid: true,
            budget: TokenBudget::default(),
            mutation: MutationStatus::None,
            availability: LocalAvailability::default(),
        }
    }

    /// A caller may report missing or contradictory evidence to escalate.
    /// They cannot mint `Verified` to take a cheaper path.
    pub fn distrust_self_reported_verified(&mut self) {
        if self.evidence == EvidenceStatus::Verified {
            self.evidence = EvidenceStatus::NotRequired;
        }
    }
}

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "camelCase")]
pub struct Classification {
    pub class: TaskClass,
    pub risk: RiskLevel,
    pub mutation_likely: bool,
}

#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case", tag = "code", content = "detail")]
pub enum RoutingReason {
    HighRiskTask(TaskClass),
    MissingEvidence,
    ContradictoryEvidence,
    SchemaValidationFailed,
    InputBudgetExceeded { estimated: u32, limit: u32 },
    OutputBudgetExceeded { estimated: u32, limit: u32 },
    MutationApprovalRequired,
    MutationReservedForUpstream,
    AmbiguousRequest,
    ImplementationReservedForUpstream,
    DeterministicRule,
    RepositoryGraphRule,
    AdvisoryDraftRule,
    WeavatrixUnavailable,
    OllamaUnavailable,
}

#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "camelCase")]
pub struct RoutingDecision {
    pub target: ExecutionTarget,
    pub class: TaskClass,
    pub risk: RiskLevel,
    pub reasons: Vec<RoutingReason>,
    /// True only for a bounded local draft which cannot mutate project state.
    pub advisory_only: bool,
    pub model_tier: ModelTier,
    pub context: ContextPlan,
}

impl RoutingDecision {
    #[must_use]
    pub const fn approves_local_model(&self) -> bool {
        matches!(self.target, ExecutionTarget::Ollama) && self.advisory_only
    }
}

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum ModelTier {
    None,
    LocalSmall,
    LocalMedium,
    UpstreamStrong,
}

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum ContextStrategy {
    Direct,
    DeterministicExtraction,
    WeavatrixEvidence,
    CitationCompression,
    UpstreamEvidence,
}

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "camelCase")]
pub struct ContextPlan {
    pub strategy: ContextStrategy,
    pub max_input_tokens: u32,
    pub require_evidence_ids: bool,
}

/// Select a target through deterministic, fail-closed routing rules.
#[must_use]
pub fn route(request: &RoutingRequest) -> RoutingDecision {
    route_with_classification(request, classify(&request.task))
}

/// Same policy as [`route`], with a precomputed classification.
///
/// Keeps every guard and target rule. Callers that ask a gated local model for
/// a tier still go through this path so lexical under-calls cannot bypass
/// high-risk or mutation floors.
#[must_use]
pub fn route_with_classification(
    request: &RoutingRequest,
    classification: Classification,
) -> RoutingDecision {
    let mut reasons = guard_reasons(request);
    if classification.class.requires_verified_evidence()
        && request.evidence == EvidenceStatus::NotRequired
    {
        reasons.push(RoutingReason::MissingEvidence);
    }
    if classification.class.is_high_risk() {
        reasons.push(RoutingReason::HighRiskTask(classification.class));
    }
    if request.mutation == MutationStatus::ApprovalRequired {
        reasons.push(RoutingReason::MutationApprovalRequired);
    }
    if request.mutation != MutationStatus::None || classification.mutation_likely {
        reasons.push(RoutingReason::MutationReservedForUpstream);
    }

    if !reasons.is_empty() {
        return upstream(classification, reasons, request);
    }

    match classification.class {
        TaskClass::Deterministic => decision(
            ExecutionTarget::Deterministic,
            classification,
            RoutingReason::DeterministicRule,
            false,
            ModelTier::None,
            ContextStrategy::DeterministicExtraction,
            request,
        ),
        TaskClass::RepositoryAnalysis if request.availability.weavatrix => decision(
            ExecutionTarget::Weavatrix,
            classification,
            RoutingReason::RepositoryGraphRule,
            false,
            ModelTier::None,
            ContextStrategy::WeavatrixEvidence,
            request,
        ),
        TaskClass::RepositoryAnalysis => upstream(
            classification,
            vec![RoutingReason::WeavatrixUnavailable],
            request,
        ),
        TaskClass::StructuredExtraction if request.availability.ollama => decision(
            ExecutionTarget::Ollama,
            classification,
            RoutingReason::AdvisoryDraftRule,
            true,
            ModelTier::LocalSmall,
            ContextStrategy::DeterministicExtraction,
            request,
        ),
        TaskClass::ContextCompression | TaskClass::AdvisoryDraft if request.availability.ollama => {
            decision(
                ExecutionTarget::Ollama,
                classification,
                RoutingReason::AdvisoryDraftRule,
                true,
                ModelTier::LocalMedium,
                ContextStrategy::CitationCompression,
                request,
            )
        }
        TaskClass::StructuredExtraction
        | TaskClass::ContextCompression
        | TaskClass::AdvisoryDraft => upstream(
            classification,
            vec![RoutingReason::OllamaUnavailable],
            request,
        ),
        TaskClass::Implementation => upstream(
            classification,
            vec![RoutingReason::ImplementationReservedForUpstream],
            request,
        ),
        TaskClass::Ambiguous => upstream(
            classification,
            vec![RoutingReason::AmbiguousRequest],
            request,
        ),
        _ => upstream(
            classification,
            vec![RoutingReason::HighRiskTask(classification.class)],
            request,
        ),
    }
}

/// Map a task class to the model tier the routing policy grants it.
#[must_use]
pub const fn policy_tier(class: TaskClass) -> ModelTier {
    match class {
        TaskClass::Deterministic | TaskClass::RepositoryAnalysis => ModelTier::None,
        TaskClass::StructuredExtraction => ModelTier::LocalSmall,
        TaskClass::ContextCompression | TaskClass::AdvisoryDraft => ModelTier::LocalMedium,
        TaskClass::Implementation
        | TaskClass::Security
        | TaskClass::Authentication
        | TaskClass::Concurrency
        | TaskClass::Migration
        | TaskClass::Release
        | TaskClass::Deployment
        | TaskClass::Publication
        | TaskClass::Ambiguous => ModelTier::UpstreamStrong,
    }
}

/// Rank tiers so under-calls are detectable without floating point.
#[must_use]
pub const fn tier_rank(tier: ModelTier) -> u8 {
    match tier {
        ModelTier::None => 0,
        ModelTier::LocalSmall => 1,
        ModelTier::LocalMedium => 2,
        ModelTier::UpstreamStrong => 3,
    }
}

/// Parse a closed-set label into a [`ModelTier`].
#[must_use]
pub fn parse_model_tier(label: &str) -> Option<ModelTier> {
    match label.trim().to_ascii_lowercase().as_str() {
        "none" => Some(ModelTier::None),
        "local_small" => Some(ModelTier::LocalSmall),
        "local_medium" => Some(ModelTier::LocalMedium),
        "upstream_strong" => Some(ModelTier::UpstreamStrong),
        _ => None,
    }
}

/// Turn a model-chosen tier into a [`Classification`], retaining lexical
/// mutation cues and preferring a more specific lexical class when the tier
/// matches.
#[must_use]
pub fn classification_for_tier(tier: ModelTier, lexical: Classification) -> Classification {
    let class = match tier {
        ModelTier::None => {
            if matches!(lexical.class, TaskClass::RepositoryAnalysis) {
                TaskClass::RepositoryAnalysis
            } else {
                TaskClass::Deterministic
            }
        }
        ModelTier::LocalSmall => TaskClass::StructuredExtraction,
        ModelTier::LocalMedium => {
            if matches!(lexical.class, TaskClass::ContextCompression) {
                TaskClass::ContextCompression
            } else {
                TaskClass::AdvisoryDraft
            }
        }
        ModelTier::UpstreamStrong => {
            if lexical.class.is_high_risk()
                || matches!(
                    lexical.class,
                    TaskClass::Implementation | TaskClass::Ambiguous
                )
            {
                lexical.class
            } else {
                TaskClass::Ambiguous
            }
        }
    };
    let risk = if class.is_high_risk() {
        RiskLevel::High
    } else if matches!(
        class,
        TaskClass::Implementation | TaskClass::RepositoryAnalysis
    ) {
        RiskLevel::Medium
    } else {
        RiskLevel::Low
    };
    Classification {
        class,
        risk,
        mutation_likely: lexical.mutation_likely,
    }
}

fn guard_reasons(request: &RoutingRequest) -> Vec<RoutingReason> {
    let mut reasons = Vec::new();
    match request.evidence {
        EvidenceStatus::Missing => reasons.push(RoutingReason::MissingEvidence),
        EvidenceStatus::Contradictory => reasons.push(RoutingReason::ContradictoryEvidence),
        EvidenceStatus::NotRequired | EvidenceStatus::Verified => {}
    }
    if !request.schema_valid {
        reasons.push(RoutingReason::SchemaValidationFailed);
    }
    if request.budget.estimated_input_tokens > request.budget.max_input_tokens {
        reasons.push(RoutingReason::InputBudgetExceeded {
            estimated: request.budget.estimated_input_tokens,
            limit: request.budget.max_input_tokens,
        });
    }
    if request.budget.estimated_output_tokens > request.budget.max_output_tokens {
        reasons.push(RoutingReason::OutputBudgetExceeded {
            estimated: request.budget.estimated_output_tokens,
            limit: request.budget.max_output_tokens,
        });
    }
    reasons
}

fn decision(
    target: ExecutionTarget,
    classification: Classification,
    reason: RoutingReason,
    advisory_only: bool,
    model_tier: ModelTier,
    strategy: ContextStrategy,
    request: &RoutingRequest,
) -> RoutingDecision {
    RoutingDecision {
        target,
        class: classification.class,
        risk: classification.risk,
        reasons: vec![reason],
        advisory_only,
        model_tier,
        context: ContextPlan {
            strategy,
            max_input_tokens: request.budget.max_input_tokens,
            require_evidence_ids: matches!(
                strategy,
                ContextStrategy::WeavatrixEvidence | ContextStrategy::CitationCompression
            ),
        },
    }
}

/// Default evidence ceiling for upstream work when the caller allows more.
pub const UPSTREAM_EVIDENCE_TOKENS: u32 = 8_192;

fn upstream(
    classification: Classification,
    reasons: Vec<RoutingReason>,
    request: &RoutingRequest,
) -> RoutingDecision {
    RoutingDecision {
        target: ExecutionTarget::Upstream,
        class: classification.class,
        risk: classification.risk.max(RiskLevel::Medium),
        reasons,
        advisory_only: false,
        model_tier: ModelTier::UpstreamStrong,
        context: ContextPlan {
            strategy: ContextStrategy::UpstreamEvidence,
            // Never exceed the caller's declared bound: a plan that reports
            // more than the request allows would break the budget contract.
            max_input_tokens: request
                .budget
                .max_input_tokens
                .min(UPSTREAM_EVIDENCE_TOKENS),
            require_evidence_ids: true,
        },
    }
}

#[cfg(test)]
mod tests;