symforge 8.3.0

SymForge — in-memory code intelligence MCP server for Claude Code, Codex, and Gemini CLI
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
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
//! STEL L2 economics controller — evaluate [`StelPlan`] → [`StelDecision`] / [`StelEstimate`].
//!
//! Phase 2 P2-S3: normative admission states (`serve | degrade | bypass | cache_hit`).

use crate::protocol::format::{competent_manual_baseline_chars, estimate_tokens_from_chars};
use crate::protocol::session::SessionContext;

use super::types::{
    AdmissionDecision, GoldenRouteRow, IndexRef, RouteConfidence, StelBypassBody, StelCacheBody,
    StelDecision, StelEstimate, StelPlan, StelPlanStep, StelRequest,
};

/// Compact-3 worst-case schema tax per call (A-006 conservative path; no amortization credit).
pub const COMPACT_SCHEMA_TOKENS: u32 = 45;
/// Compact `symforge` invoke overhead per call (schema example + Phase 0 doctrine).
pub const COMPACT_INVOKE_TOKENS: u32 = 80;
/// Minimum predicted net vs manual before `serve` (schema `margin_high`).
pub const SERVE_MARGIN_TOKENS: i32 = 50;
/// Predicted net at or below this (but above zero) triggers `degrade` (`margin_low`).
pub const DEGRADE_MARGIN_LOW_TOKENS: i32 = SERVE_MARGIN_TOKENS;
/// Default capped response budget applied on degrade when request omits `max_tokens`.
pub const DEGRADE_DEFAULT_MAX_TOKENS: u32 = 400;

/// Numerator of the grounded structured-response fraction: SymForge's structured
/// serve (symbol body / outline / windowed slice) is smaller than the competent
/// windowed manual read it replaces. Calibrated conservatively at ~60% of the
/// competent-manual baseline so a grounded response prediction never *over*-claims
/// savings (a smaller predicted response would inflate `predicted_net`).
const GROUNDED_RESPONSE_FRACTION_NUM: u64 = 3;
/// Denominator of the grounded structured-response fraction (see numerator).
const GROUNDED_RESPONSE_FRACTION_DEN: u64 = 5;

/// Token economics breakdown for one planned invocation.
#[derive(Clone, Debug, Eq, PartialEq)]
pub struct EconomicsBreakdown {
    pub predicted_response_tokens: u32,
    pub predicted_manual_tokens: u32,
    pub predicted_schema_tokens: u32,
    pub predicted_invoke_tokens: u32,
    pub predicted_symforge_tokens: u32,
    pub predicted_net_vs_manual: i32,
}

/// Evaluate L2 admission for a draft plan (does not execute L3).
pub fn evaluate_plan(request: &StelRequest, plan: &StelPlan) -> StelDecision {
    evaluate_plan_with_session(request, plan, None)
}

/// Evaluate L2 admission with optional in-process session context for cache-hit detection.
pub fn evaluate_plan_with_session(
    request: &StelRequest,
    plan: &StelPlan,
    session: Option<&SessionContext>,
) -> StelDecision {
    if let Some(bypass) = detect_pff_bypass(request) {
        return decision_from_pff_bypass(plan, request, bypass);
    }

    if let Some(session) = session
        && let Some(cache) = detect_session_cache_hit(plan, session)
    {
        return decision_from_cache_hit(plan, request, cache);
    }

    let economics = estimate_economics(plan);
    let net = economics.predicted_net_vs_manual;

    if net <= 0 {
        let bypass = economics_bypass_body(plan, &economics);
        return StelDecision {
            plan_id: plan.plan_id.clone(),
            decision: AdmissionDecision::Bypass,
            decision_reason: format!(
                "predicted_net={net} <= 0; economics bypass via host_read `{}`",
                bypass.path
            ),
            effective_max_tokens: request.max_tokens,
            degrade_flags: vec![],
            steps: None,
            bypass: Some(bypass),
            cache: None,
        };
    }

    let mandatory_degrade =
        plan.confidence == RouteConfidence::Fallback && net < SERVE_MARGIN_TOKENS;
    let economics_degrade = net <= DEGRADE_MARGIN_LOW_TOKENS;

    if mandatory_degrade || economics_degrade {
        let mut degrade_flags = Vec::new();
        if mandatory_degrade {
            degrade_flags.push("fallback_mandatory".to_string());
        }
        if plan
            .steps
            .iter()
            .any(|step| step.tool == "get_file_context")
        {
            degrade_flags.push("outline_only".to_string());
        } else {
            degrade_flags.push("max_tokens_cap".to_string());
        }
        degrade_flags.sort();
        degrade_flags.dedup();

        let effective_max_tokens = request
            .max_tokens
            .or(Some(DEGRADE_DEFAULT_MAX_TOKENS))
            .map(|cap| cap.min(DEGRADE_DEFAULT_MAX_TOKENS));

        let reason = if mandatory_degrade {
            format!(
                "predicted_net={net} < margin={}; fallback confidence requires degrade",
                SERVE_MARGIN_TOKENS
            )
        } else {
            format!(
                "predicted_net={net} <= margin_low={}",
                DEGRADE_MARGIN_LOW_TOKENS
            )
        };

        return StelDecision {
            plan_id: plan.plan_id.clone(),
            decision: AdmissionDecision::Degrade,
            decision_reason: reason,
            effective_max_tokens,
            degrade_flags,
            steps: Some(plan.steps.clone()),
            bypass: None,
            cache: None,
        };
    }

    StelDecision {
        plan_id: plan.plan_id.clone(),
        decision: AdmissionDecision::Serve,
        decision_reason: format!("predicted_net={net} > margin={}", SERVE_MARGIN_TOKENS),
        effective_max_tokens: request.max_tokens,
        degrade_flags: vec![],
        steps: Some(plan.steps.clone()),
        bypass: None,
        cache: None,
    }
}

fn decision_from_pff_bypass(
    plan: &StelPlan,
    request: &StelRequest,
    bypass: StelBypassBody,
) -> StelDecision {
    let economics = economics_for_bypass(&bypass);
    StelDecision {
        plan_id: plan.plan_id.clone(),
        decision: AdmissionDecision::Bypass,
        decision_reason: format!(
            "{}; predicted_net={}",
            bypass.reason, economics.predicted_net_vs_manual
        ),
        effective_max_tokens: request.max_tokens,
        degrade_flags: vec![],
        steps: None,
        bypass: Some(bypass),
        cache: None,
    }
}

fn decision_from_cache_hit(
    plan: &StelPlan,
    request: &StelRequest,
    cache: StelCacheBody,
) -> StelDecision {
    StelDecision {
        plan_id: plan.plan_id.clone(),
        decision: AdmissionDecision::CacheHit,
        decision_reason: format!(
            "session cache hit for {} `{}` (prior_tokens={})",
            cache.kind, cache.path, cache.prior_tokens
        ),
        effective_max_tokens: request.max_tokens,
        degrade_flags: vec![],
        steps: None,
        bypass: None,
        cache: Some(cache),
    }
}

fn detect_session_cache_hit(plan: &StelPlan, session: &SessionContext) -> Option<StelCacheBody> {
    let step = plan.steps.first()?;
    session
        .try_cache_hit_from_stel_step(&step.tool, &step.args)
        .map(|meta| StelCacheBody {
            kind: meta.kind.to_string(),
            path: meta.path,
            name: meta.name,
            prior_tokens: meta.prior_tokens,
            session_age_secs: meta.session_age_secs,
        })
}

fn economics_bypass_body(plan: &StelPlan, economics: &EconomicsBreakdown) -> StelBypassBody {
    let path = plan_primary_path(plan).unwrap_or_else(|| "unknown".to_string());
    StelBypassBody {
        action: "host_read".to_string(),
        path: path.clone(),
        start_line: 1,
        end_line: Some(80),
        predicted_manual_tokens: economics.predicted_manual_tokens,
        predicted_symforge_tokens: economics.predicted_symforge_tokens,
        reason: format!(
            "predicted_net={} <= 0; host_read `{path}` cheaper than serve",
            economics.predicted_net_vs_manual
        ),
    }
}

fn plan_primary_path(plan: &StelPlan) -> Option<String> {
    plan.steps.first().and_then(primary_path_from_step)
}

fn primary_path_from_step(step: &StelPlanStep) -> Option<String> {
    step.args
        .get("path")
        .and_then(|value| value.as_str())
        .map(str::trim)
        .filter(|path| !path.is_empty())
        .map(str::to_string)
        .or_else(|| {
            step.args
                .get("name")
                .and_then(|value| value.as_str())
                .map(str::trim)
                .filter(|name| !name.is_empty())
                .map(str::to_string)
        })
}

/// Evaluate L2 admission for a structural edit plan (no NL P-FF bypass path).
pub fn evaluate_edit_plan(plan: &StelPlan) -> StelDecision {
    let economics = estimate_economics(plan);
    let net = economics.predicted_net_vs_manual;
    if net <= 0 {
        let bypass = economics_bypass_body(plan, &economics);
        return StelDecision {
            plan_id: plan.plan_id.clone(),
            decision: AdmissionDecision::Bypass,
            decision_reason: format!("predicted_net={net} <= 0; edit economics bypass"),
            effective_max_tokens: None,
            degrade_flags: vec![],
            steps: Some(plan.steps.clone()),
            bypass: Some(bypass),
            cache: None,
        };
    }
    if net <= DEGRADE_MARGIN_LOW_TOKENS {
        return StelDecision {
            plan_id: plan.plan_id.clone(),
            decision: AdmissionDecision::Degrade,
            decision_reason: format!(
                "predicted_net={net} <= margin_low={}",
                DEGRADE_MARGIN_LOW_TOKENS
            ),
            effective_max_tokens: Some(DEGRADE_DEFAULT_MAX_TOKENS),
            degrade_flags: vec!["max_tokens_cap".to_string()],
            steps: Some(plan.steps.clone()),
            bypass: None,
            cache: None,
        };
    }
    StelDecision {
        plan_id: plan.plan_id.clone(),
        decision: AdmissionDecision::Serve,
        decision_reason: format!("predicted_net={net} > margin={}", SERVE_MARGIN_TOKENS),
        effective_max_tokens: None,
        degrade_flags: vec![],
        steps: Some(plan.steps.clone()),
        bypass: None,
        cache: None,
    }
}

/// Build preview economics for `preview: true` (L1+L2 only).
pub fn build_estimate(
    request: &StelRequest,
    plan: &StelPlan,
    decision: &StelDecision,
) -> StelEstimate {
    let economics = if decision.decision == AdmissionDecision::Bypass {
        decision
            .bypass
            .as_ref()
            .map(economics_for_bypass)
            .unwrap_or_else(|| estimate_economics(plan))
    } else {
        estimate_economics(plan)
    };
    let recommended = decision.decision == AdmissionDecision::Serve
        && economics.predicted_net_vs_manual > SERVE_MARGIN_TOKENS;
    let _ = request;
    StelEstimate {
        plan_id: plan.plan_id.clone(),
        decision: decision.decision,
        predicted_response_tokens: economics.predicted_response_tokens,
        predicted_manual_tokens: economics.predicted_manual_tokens,
        predicted_schema_tokens: economics.predicted_schema_tokens,
        predicted_invoke_tokens: economics.predicted_invoke_tokens,
        predicted_net_vs_manual: economics.predicted_net_vs_manual,
        recommended,
    }
}

/// Sum step estimates + compact surface overhead into a manual-vs-symforge comparison.
///
/// US5 grounding (010 FR-014, D2): when a step carries real [`IndexRef`] byte
/// sizes (populated by the index-aware serve path before the gate), the
/// per-step response/manual token estimates are derived from those real bytes
/// via the byte-grounded estimator ([`competent_manual_baseline_chars`]) rather
/// than the planner's `400/800` plan-only constants. This is why predictions now
/// vary with the actual work and why the adaptive economics branches
/// (`bypass`/`degrade`) become reachable for small/cheap inputs (a tiny file's
/// manual baseline is below SymForge's fixed schema+invoke overhead, so
/// `predicted_net` correctly goes non-positive). A step with no `index_refs`
/// (plan-only callers, preview, unit fixtures) keeps the deterministic per-step
/// `est_response_tokens`/`est_manual_tokens` floor unchanged. The result is still
/// an *estimate* (predicted from real bytes, not a measured token count) — the
/// envelope keeps the `est_`/`heuristic` label (relabel != measure).
pub fn estimate_economics(plan: &StelPlan) -> EconomicsBreakdown {
    let (predicted_response_tokens, predicted_manual_tokens) = plan
        .steps
        .iter()
        .map(grounded_step_tokens)
        .fold((0u32, 0u32), |(resp, manual), (step_resp, step_manual)| {
            (
                resp.saturating_add(step_resp),
                manual.saturating_add(step_manual),
            )
        });
    let predicted_symforge_tokens = predicted_response_tokens
        .saturating_add(COMPACT_SCHEMA_TOKENS)
        .saturating_add(COMPACT_INVOKE_TOKENS);
    let predicted_net_vs_manual = predicted_manual_tokens as i32 - predicted_symforge_tokens as i32;
    EconomicsBreakdown {
        predicted_response_tokens,
        predicted_manual_tokens,
        predicted_schema_tokens: COMPACT_SCHEMA_TOKENS,
        predicted_invoke_tokens: COMPACT_INVOKE_TOKENS,
        predicted_symforge_tokens,
        predicted_net_vs_manual,
    }
}

/// Per-step `(predicted_response_tokens, predicted_manual_tokens)`.
///
/// Grounded path: when `step.index_refs` is non-empty the manual baseline is the
/// competent windowed read of the real target bytes, and the predicted response
/// is the structured-serve fraction of that baseline (SymForge's symbol/outline/
/// slice is smaller than a windowed manual read). Falls back to the plan-only
/// `est_*` constants when no real size is known (preserves determinism for
/// plan-only callers and existing fixtures).
fn grounded_step_tokens(step: &StelPlanStep) -> (u32, u32) {
    if step.index_refs.is_empty() {
        return (step.est_response_tokens, step.est_manual_tokens);
    }
    let total_raw_chars: usize = step
        .index_refs
        .iter()
        .map(|index_ref| index_ref.raw_chars as usize)
        .sum();
    grounded_tokens_from_raw_chars(total_raw_chars)
}

/// Map real target byte length to `(predicted_response_tokens, predicted_manual_tokens)`.
///
/// `manual` = competent windowed read of the real bytes (what a disciplined
/// agent would read by hand); `response` = the structured-serve fraction of that
/// baseline. Both flow through the existing byte-grounded estimator so there is a
/// single source of truth for the manual baseline (no duplicate estimator).
fn grounded_tokens_from_raw_chars(raw_chars: usize) -> (u32, u32) {
    let manual_chars = competent_manual_baseline_chars(raw_chars);
    let manual_tokens = estimate_tokens_from_chars(manual_chars);
    let response_tokens = manual_tokens.saturating_mul(GROUNDED_RESPONSE_FRACTION_NUM)
        / GROUNDED_RESPONSE_FRACTION_DEN;
    (clamp_u32(response_tokens), clamp_u32(manual_tokens))
}

/// Saturating `u64` -> `u32` (token counts never legitimately exceed `u32::MAX`).
fn clamp_u32(value: u64) -> u32 {
    value.min(u32::MAX as u64) as u32
}

/// Build a grounded [`IndexRef`] for a resolved target. Used by the index-aware
/// serve path to stamp real byte sizes onto plan steps before the L2 gate.
pub fn index_ref_for_target(path: impl Into<String>, raw_chars: u64) -> IndexRef {
    IndexRef {
        path: path.into(),
        raw_chars,
    }
}

pub fn economics_for_bypass(bypass: &StelBypassBody) -> EconomicsBreakdown {
    let predicted_symforge_tokens = bypass
        .predicted_symforge_tokens
        .saturating_add(COMPACT_SCHEMA_TOKENS)
        .saturating_add(COMPACT_INVOKE_TOKENS);
    let predicted_net_vs_manual =
        bypass.predicted_manual_tokens as i32 - predicted_symforge_tokens as i32;
    EconomicsBreakdown {
        predicted_response_tokens: 0,
        predicted_manual_tokens: bypass.predicted_manual_tokens,
        predicted_schema_tokens: COMPACT_SCHEMA_TOKENS,
        predicted_invoke_tokens: COMPACT_INVOKE_TOKENS,
        predicted_symforge_tokens,
        predicted_net_vs_manual,
    }
}

/// Policy P-FF: whole-file review queries bypass SymForge serve (A-012 / golden corpus).
pub fn detect_pff_bypass(request: &StelRequest) -> Option<StelBypassBody> {
    let query = request.query.trim();
    let lower = query.to_ascii_lowercase();
    if !is_pff_query(&lower) {
        return None;
    }
    let path = extract_pff_path(query, &lower)?;
    Some(StelBypassBody {
        action: "host_read".to_string(),
        path: path.clone(),
        start_line: 1,
        end_line: None,
        predicted_manual_tokens: 320,
        predicted_symforge_tokens: COMPACT_SCHEMA_TOKENS + COMPACT_INVOKE_TOKENS + 256,
        reason: format!("policy=P-FF whole-file review; host_read `{path}`"),
    })
}

fn is_pff_query(lower: &str) -> bool {
    const PHRASES: &[&str] = &[
        "entire ",
        "whole ",
        "full file",
        "complete ",
        "line by line",
        "audit full ",
        "review entire ",
        "read complete ",
    ];
    PHRASES.iter().any(|phrase| lower.contains(phrase))
}

fn extract_pff_path(query: &str, lower: &str) -> Option<String> {
    for token in query.split_whitespace().rev() {
        let cleaned = token.trim_matches(|c: char| !c.is_alphanumeric() && c != '.' && c != '/');
        if cleaned.contains('.') && !cleaned.eq_ignore_ascii_case("line") {
            return Some(cleaned.to_string());
        }
    }
    for token in lower.split_whitespace() {
        if token.contains('.') {
            return Some(token.to_string());
        }
    }
    None
}

/// Golden-row helper: L2 decision should match fixture `expected_decision` for representative rows.
pub fn decision_matches_golden(
    row: &GoldenRouteRow,
    plan: &StelPlan,
    request: &StelRequest,
) -> bool {
    let decision = evaluate_plan(request, plan);
    decision.decision == row.expected_decision
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::path::Path;

    use crate::stel::golden_replay::{GOLDEN_ROUTES_FIXTURE, load_golden_rows};
    use crate::stel::planner::build_plan;
    use crate::stel::types::{IntentBucket, RouteConfidence, StelPlan, StelPlanStep};

    fn row_by_id<'a>(rows: &'a [GoldenRouteRow], id: &str) -> &'a GoldenRouteRow {
        rows.iter()
            .find(|row| row.id == id)
            .unwrap_or_else(|| panic!("missing golden row {id}"))
    }

    fn request_for_row(row: &GoldenRouteRow) -> StelRequest {
        let mut request = row.to_request();
        request.intent = row.intent;
        request
    }

    fn low_net_plan() -> StelPlan {
        StelPlan {
            plan_id: "low-net".to_string(),
            intent: IntentBucket::Read,
            confidence: RouteConfidence::Inferred,
            confidence_rationale: "test".to_string(),
            steps: vec![StelPlanStep {
                order: 1,
                tool: "get_file_context".to_string(),
                args: serde_json::json!({ "path": "src/lib.rs" }),
                est_response_tokens: 900,
                est_manual_tokens: 100,
                index_refs: vec![],
            }],
            suggested_followup: None,
        }
    }

    #[test]
    fn serve_rows_classify_as_serve() {
        let path = Path::new(env!("CARGO_MANIFEST_DIR")).join(GOLDEN_ROUTES_FIXTURE);
        let rows = load_golden_rows(&path).expect("golden fixture");
        for id in [
            "cfg-if/t1_search",
            "cfg-if/t4_refs",
            "records/t4_refs",
            "compression/t5_dependents",
        ] {
            let row = row_by_id(&rows, id);
            let request = request_for_row(row);
            let plan = build_plan(&request);
            let decision = evaluate_plan(&request, &plan);
            assert_eq!(
                decision.decision,
                AdmissionDecision::Serve,
                "{id}: {:?}",
                decision.decision_reason
            );
            assert!(row.eligible_h6.unwrap_or(false));
            let economics = estimate_economics(&plan);
            assert!(
                economics.predicted_net_vs_manual > SERVE_MARGIN_TOKENS,
                "{id} net {}",
                economics.predicted_net_vs_manual
            );
        }
    }

    #[test]
    fn pff_rows_classify_as_bypass() {
        let path = Path::new(env!("CARGO_MANIFEST_DIR")).join(GOLDEN_ROUTES_FIXTURE);
        let rows = load_golden_rows(&path).expect("golden fixture");
        for id in [
            "cfg-if/pff_whole_lib",
            "records/pff_whole_module",
            "is-plain/pff_whole_index",
            "compression/pff_whole_service",
        ] {
            let row = row_by_id(&rows, id);
            let request = request_for_row(row);
            let plan = build_plan(&request);
            let decision = evaluate_plan(&request, &plan);
            assert_eq!(decision.decision, AdmissionDecision::Bypass, "{id}");
            assert!(decision.bypass.is_some(), "{id} bypass body");
            let bypass = decision.bypass.as_ref().expect("bypass body");
            assert_eq!(
                bypass.end_line, None,
                "{id} P-FF bypass must request whole-file host read"
            );
            assert_eq!(row.eligible_h6, Some(false));
            let estimate = build_estimate(&request, &plan, &decision);
            assert!(!estimate.recommended, "{id} should not recommend serve");
        }
    }

    #[test]
    fn negative_net_non_pff_bypass_has_host_read_body() {
        let plan = low_net_plan();
        let request = StelRequest::default();
        let decision = evaluate_plan(&request, &plan);
        assert_eq!(decision.decision, AdmissionDecision::Bypass);
        let bypass = decision.bypass.as_ref().expect("economics bypass body");
        assert_eq!(bypass.path, "src/lib.rs");
        assert!(decision.decision_reason.contains("predicted_net="));
    }

    #[test]
    fn marginal_net_degrades_with_outline_only_flag() {
        let plan = StelPlan {
            plan_id: "degrade".to_string(),
            intent: IntentBucket::Read,
            confidence: RouteConfidence::Inferred,
            confidence_rationale: "test".to_string(),
            steps: vec![StelPlanStep {
                order: 1,
                tool: "get_file_context".to_string(),
                args: serde_json::json!({ "path": "src/lib.rs" }),
                est_response_tokens: 400,
                est_manual_tokens: 530,
                index_refs: vec![],
            }],
            suggested_followup: None,
        };
        let request = StelRequest::default();
        let decision = evaluate_plan(&request, &plan);
        assert_eq!(decision.decision, AdmissionDecision::Degrade);
        assert!(decision.degrade_flags.contains(&"outline_only".to_string()));
        assert!(decision.effective_max_tokens.is_some());
    }

    #[test]
    fn fallback_confidence_forces_degrade_below_margin_high() {
        let plan = StelPlan {
            plan_id: "fallback".to_string(),
            intent: IntentBucket::Read,
            confidence: RouteConfidence::Fallback,
            confidence_rationale: "test".to_string(),
            steps: vec![StelPlanStep {
                order: 1,
                tool: "get_file_content".to_string(),
                args: serde_json::json!({ "path": "src/lib.rs" }),
                est_response_tokens: 400,
                est_manual_tokens: 530,
                index_refs: vec![],
            }],
            suggested_followup: None,
        };
        let request = StelRequest::default();
        let decision = evaluate_plan(&request, &plan);
        assert_eq!(decision.decision, AdmissionDecision::Degrade);
        assert!(
            decision
                .degrade_flags
                .contains(&"fallback_mandatory".to_string())
        );
    }

    #[test]
    fn session_cache_hit_for_prefetched_symbol() {
        let session = SessionContext::new();
        session.record_symbol_fetch(
            "src/lib.rs",
            "cfg_if",
            crate::protocol::session::hash_symbol_params(None, None, None),
            128,
        );
        let plan = StelPlan {
            plan_id: "cache".to_string(),
            intent: IntentBucket::Read,
            confidence: RouteConfidence::Exact,
            confidence_rationale: "test".to_string(),
            steps: vec![StelPlanStep {
                order: 1,
                tool: "get_symbol".to_string(),
                args: serde_json::json!({ "path": "src/lib.rs", "name": "cfg_if" }),
                est_response_tokens: 400,
                est_manual_tokens: 800,
                index_refs: vec![],
            }],
            suggested_followup: None,
        };
        let request = StelRequest::default();
        let decision = evaluate_plan_with_session(&request, &plan, Some(&session));
        assert_eq!(decision.decision, AdmissionDecision::CacheHit);
        let cache = decision.cache.as_ref().expect("cache body");
        assert_eq!(cache.kind, "symbol");
        assert_eq!(cache.prior_tokens, 128);
    }

    #[test]
    fn eligible_h6_serve_rows_remain_eligible() {
        let path = Path::new(env!("CARGO_MANIFEST_DIR")).join(GOLDEN_ROUTES_FIXTURE);
        let rows = load_golden_rows(&path).expect("golden fixture");
        let eligible: Vec<_> = rows
            .iter()
            .filter(|row| {
                row.eligible_h6 == Some(true) && row.expected_decision == AdmissionDecision::Serve
            })
            .collect();
        assert!(eligible.len() >= 20);
        for row in eligible.iter().take(6) {
            let request = request_for_row(row);
            let plan = build_plan(&request);
            assert!(
                decision_matches_golden(row, &plan, &request),
                "eligible serve row {} decision mismatch",
                row.id
            );
        }
    }

    #[test]
    fn preview_estimate_carries_schema_and_invoke_costs() {
        let request = StelRequest {
            query: "who references cfg_if".to_string(),
            ..Default::default()
        };
        let plan = build_plan(&request);
        let decision = evaluate_plan(&request, &plan);
        let estimate = build_estimate(&request, &plan, &decision);
        assert_eq!(estimate.predicted_schema_tokens, COMPACT_SCHEMA_TOKENS);
        assert_eq!(estimate.predicted_invoke_tokens, COMPACT_INVOKE_TOKENS);
        assert_eq!(estimate.decision, AdmissionDecision::Serve);
        assert!(estimate.recommended);
    }

    #[test]
    fn fallback_confidence_still_serves_when_net_positive() {
        let plan = StelPlan {
            plan_id: "test".to_string(),
            intent: IntentBucket::Read,
            confidence: RouteConfidence::Fallback,
            confidence_rationale: "test".to_string(),
            steps: vec![StelPlanStep {
                order: 1,
                tool: "get_file_context".to_string(),
                args: serde_json::json!({}),
                est_response_tokens: 400,
                est_manual_tokens: 800,
                index_refs: vec![],
            }],
            suggested_followup: None,
        };
        let request = StelRequest::default();
        let decision = evaluate_plan(&request, &plan);
        assert_eq!(decision.decision, AdmissionDecision::Serve);
    }

    /// Build a single-step read plan grounded in a real target byte size.
    fn grounded_read_plan(raw_chars: u64, confidence: RouteConfidence) -> StelPlan {
        StelPlan {
            plan_id: format!("grounded-{raw_chars}"),
            intent: IntentBucket::Read,
            confidence,
            confidence_rationale: "test".to_string(),
            steps: vec![StelPlanStep {
                order: 1,
                tool: "get_file_context".to_string(),
                args: serde_json::json!({ "path": "src/lib.rs" }),
                // Plan-only floor that grounding must override when index_refs present.
                est_response_tokens: 400,
                est_manual_tokens: 800,
                index_refs: vec![IndexRef {
                    path: "src/lib.rs".to_string(),
                    raw_chars,
                }],
            }],
            suggested_followup: None,
        }
    }

    /// T035 (SC-005, US5 AC-1): the same operation over a small and a large real
    /// input yields DIFFERENT grounded predictions — never the pre-grounding
    /// fixed constant. The prediction scales with the actual target byte length.
    #[test]
    fn grounded_predictions_vary_with_real_input_size() {
        let small = estimate_economics(&grounded_read_plan(600, RouteConfidence::Inferred));
        let large = estimate_economics(&grounded_read_plan(40_000, RouteConfidence::Inferred));

        // Predictions are not the fixed 400/800-derived constant.
        assert_ne!(
            small.predicted_manual_tokens, large.predicted_manual_tokens,
            "manual baseline must vary with real size"
        );
        assert_ne!(
            small.predicted_net_vs_manual, large.predicted_net_vs_manual,
            "predicted net must vary with real size"
        );
        // Larger input ⇒ larger manual baseline ⇒ a more favorable (higher) net.
        assert!(
            large.predicted_manual_tokens > small.predicted_manual_tokens,
            "bigger file ⇒ bigger manual baseline (small={} large={})",
            small.predicted_manual_tokens,
            large.predicted_manual_tokens
        );
        assert!(
            large.predicted_net_vs_manual > small.predicted_net_vs_manual,
            "bigger file ⇒ higher predicted net (small={} large={})",
            small.predicted_net_vs_manual,
            large.predicted_net_vs_manual
        );

        // A step with NO index_refs keeps the deterministic plan-only floor
        // (preserves determinism for plan-only callers / preview / fixtures).
        let plan_only = StelPlan {
            steps: vec![StelPlanStep {
                index_refs: vec![],
                ..grounded_read_plan(600, RouteConfidence::Inferred).steps[0].clone()
            }],
            ..grounded_read_plan(600, RouteConfidence::Inferred)
        };
        let floor = estimate_economics(&plan_only);
        assert_eq!(floor.predicted_response_tokens, 400);
        assert_eq!(floor.predicted_manual_tokens, 800);
    }

    /// T036 (US5 AC-2, TR-04b): a sufficiently small grounded request makes a
    /// non-serve economics branch REACHABLE — a tiny file's competent-manual
    /// baseline falls below SymForge's fixed schema+invoke overhead, so
    /// `predicted_net <= 0` and the gate correctly BYPASSES (host-read is cheaper
    /// than serving). This branch was unreachable while every step carried the
    /// 400/800 constant (net was permanently ~275 ⇒ always serve).
    #[test]
    fn grounded_small_request_reaches_economics_bypass() {
        // 180-char file: below the 200-char small-file threshold ⇒ manual
        // baseline is the whole tiny file (~45 tokens), well under the 125-token
        // schema+invoke floor ⇒ predicted_net is non-positive.
        let tiny = grounded_read_plan(180, RouteConfidence::Inferred);
        let economics = estimate_economics(&tiny);
        assert!(
            economics.predicted_net_vs_manual <= 0,
            "tiny grounded request must predict non-positive net, got {}",
            economics.predicted_net_vs_manual
        );
        let decision = evaluate_plan(&StelRequest::default(), &tiny);
        assert_eq!(
            decision.decision,
            AdmissionDecision::Bypass,
            "tiny grounded request must reach economics bypass: {}",
            decision.decision_reason
        );
        assert!(
            decision.bypass.is_some(),
            "economics bypass must carry a host-read body"
        );
    }

    /// T036 / N-2: a low-confidence Fallback route with a grounded marginal net
    /// reaches the `mandatory_degrade` branch — the third dead branch the
    /// constant kept unreachable. Low-confidence routes now get an economic
    /// guardrail (degrade) instead of being served at full budget.
    #[test]
    fn grounded_fallback_marginal_reaches_mandatory_degrade() {
        // Choose a raw size whose grounded net lands in (0, SERVE_MARGIN): a
        // ~1600-char file ⇒ manual ~400 tokens, response ~240, net = 400 - (240+125)
        // = 35 ⇒ 0 < 35 < 50 = SERVE_MARGIN. On a Fallback route that triggers
        // mandatory_degrade (and economics_degrade), never serve, never bypass.
        let plan = grounded_read_plan(1600, RouteConfidence::Fallback);
        let economics = estimate_economics(&plan);
        assert!(
            economics.predicted_net_vs_manual > 0
                && economics.predicted_net_vs_manual < SERVE_MARGIN_TOKENS,
            "grounded net must be marginal (0 < net < {SERVE_MARGIN_TOKENS}), got {}",
            economics.predicted_net_vs_manual
        );
        let decision = evaluate_plan(&StelRequest::default(), &plan);
        assert_eq!(
            decision.decision,
            AdmissionDecision::Degrade,
            "marginal fallback must degrade: {}",
            decision.decision_reason
        );
        assert!(
            decision
                .degrade_flags
                .contains(&"fallback_mandatory".to_string()),
            "N-2: low-confidence fallback must carry the mandatory degrade guardrail: {:?}",
            decision.degrade_flags
        );
    }
}