backpass 0.1.13 → 0.1.14

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -313,7 +313,10 @@ not duplicate content.
313
313
 
314
314
  `backpass apply` is the only command that writes. It serves a review surface through
315
315
  [`lavish-axi`](https://github.com/kunchenguid/lavish-axi): one card per edit with the diff,
316
- the evidence quotes and their sources, a live budget gauge, and ACCEPT / REJECT.
316
+ the evidence quotes and their sources, a live budget gauge, and ACCEPT / REJECT. A compact
317
+ gap funnel shows how accumulated sightings narrowed through domain filtering, clustering,
318
+ and the corroboration floor to the gaps eligible for a proposal; older proposals without
319
+ recorded funnel counts omit it.
317
320
 
318
321
  The surface is a static template shipped in the package - the CLI injects one JSON payload,
319
322
  so it is instant, deterministic, and identical every run. Nothing there is model-generated.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "backpass",
3
- "version": "0.1.13",
3
+ "version": "0.1.14",
4
4
  "packageManager": "pnpm@11.5.0",
5
5
  "description": "Gradient descent for your agent memory - analyzes past agent session transcripts and proposes evidence-backed edits to AGENTS.md / CLAUDE.md",
6
6
  "type": "module",
package/src/fold.js CHANGED
@@ -143,6 +143,10 @@ export function foldEvidence(evidenceRecords, { minGapEvidence = 2, memoryFile =
143
143
  totals: {
144
144
  positive: positiveCount,
145
145
  negative: negativeCount,
146
+ // The gap funnel's top: every sighting this fold clustered over (the pruned ledger
147
+ // when one is passed, this run's records otherwise). Orchestration sightings are the
148
+ // slice excluded before clustering; project-domain is the difference.
149
+ gapSightings: allObservations.length,
146
150
  gapClusters: gaps.length,
147
151
  droppedGapSingletons,
148
152
  orchestrationGapSightings,
package/src/proposal.js CHANGED
@@ -489,6 +489,13 @@ export function buildProposal(rawResult, context) {
489
489
  positive: summary?.totals?.positive ?? 0,
490
490
  negative: summary?.totals?.negative ?? 0,
491
491
  gapClusters: summary?.totals?.gapClusters ?? 0,
492
+ // The gap funnel, for the apply surface: sightings -> project-domain (sightings
493
+ // minus orchestration) -> distinct gaps (clusters + dropped) -> eligible
494
+ // (clusters). `null`, never 0, when the summary predates a count - the surface
495
+ // hides what was not recorded rather than showing an invented zero.
496
+ gapSightings: summary?.totals?.gapSightings ?? null,
497
+ orchestrationGapSightings: summary?.totals?.orchestrationGapSightings ?? null,
498
+ droppedGapSingletons: summary?.totals?.droppedGapSingletons ?? null,
492
499
  skillExtractions: accepted.reduce((n, e) => n + editSkills(e).length, 0),
493
500
  },
494
501
  edits: accepted,
@@ -126,6 +126,124 @@
126
126
  margin-top: 3px;
127
127
  }
128
128
 
129
+ /* ---------- gap funnel ---------- */
130
+ /* The template's own display rules (.statrow's grid) would beat the UA
131
+ stylesheet's [hidden]; this keeps the hidden attribute authoritative. */
132
+ [hidden] {
133
+ display: none !important;
134
+ }
135
+ .run-context {
136
+ border: 1px solid var(--line);
137
+ background: var(--line);
138
+ display: grid;
139
+ grid-template-columns: minmax(0, 3fr) minmax(280px, 2fr);
140
+ gap: 1px;
141
+ margin-bottom: 36px;
142
+ }
143
+ .ctx-funnel {
144
+ background: var(--panel);
145
+ padding: 13px 18px 14px;
146
+ min-width: 0;
147
+ }
148
+ .ctx-stats {
149
+ display: grid;
150
+ grid-template-columns: 1fr 1fr;
151
+ grid-auto-rows: 1fr;
152
+ gap: 1px;
153
+ min-width: 0;
154
+ }
155
+ .funnel-head {
156
+ display: flex;
157
+ justify-content: space-between;
158
+ align-items: baseline;
159
+ flex-wrap: wrap;
160
+ gap: 8px;
161
+ margin-bottom: 10px;
162
+ }
163
+ .funnel-head .t {
164
+ font-size: 13px;
165
+ color: var(--dim);
166
+ }
167
+ .funnel-head .facts {
168
+ font-family: var(--mono);
169
+ font-size: 11px;
170
+ color: var(--faint);
171
+ }
172
+ .funnel-head .facts b {
173
+ color: var(--text);
174
+ font-weight: 500;
175
+ }
176
+ .frow {
177
+ display: flex;
178
+ align-items: center;
179
+ gap: 12px;
180
+ margin: 4px 0;
181
+ }
182
+ .frow .fl {
183
+ width: 148px;
184
+ flex: none;
185
+ font-family: var(--mono);
186
+ font-size: 10px;
187
+ letter-spacing: 0.11em;
188
+ text-transform: uppercase;
189
+ color: var(--faint);
190
+ white-space: nowrap;
191
+ overflow: hidden;
192
+ text-overflow: ellipsis;
193
+ }
194
+ .frow .fb {
195
+ flex: 1;
196
+ min-width: 0;
197
+ display: flex;
198
+ align-items: center;
199
+ height: 16px;
200
+ background: var(--panel-2);
201
+ border: 1px solid var(--line-soft);
202
+ }
203
+ .frow .fb i {
204
+ display: block;
205
+ height: 100%;
206
+ background: rgba(110, 168, 255, 0.38);
207
+ border-right: 1px solid rgba(110, 168, 255, 0.75);
208
+ }
209
+ .frow.lit .fb i {
210
+ background: rgba(79, 227, 193, 0.4);
211
+ border-right-color: var(--accent);
212
+ }
213
+ .frow.zero .fb i {
214
+ border-right: none;
215
+ }
216
+ .frow .fv {
217
+ flex: none;
218
+ width: 34px;
219
+ font-family: var(--mono);
220
+ font-size: 12.5px;
221
+ color: var(--text);
222
+ }
223
+ .frow.zero .fv {
224
+ color: var(--faint);
225
+ }
226
+ .frow.lit .fv {
227
+ color: var(--accent);
228
+ }
229
+ .fdrop {
230
+ margin: 1px 0 6px 160px;
231
+ font-size: 12px;
232
+ color: var(--dim);
233
+ }
234
+ .fdrop b {
235
+ font-family: var(--mono);
236
+ font-weight: 500;
237
+ color: var(--faint);
238
+ }
239
+ .funnel-note {
240
+ margin-top: 10px;
241
+ padding-top: 9px;
242
+ border-top: 1px dashed var(--line-soft);
243
+ font-size: 12.5px;
244
+ color: var(--dim);
245
+ }
246
+
129
247
  /* ---------- budget gauge ---------- */
130
248
  .gauge-block {
131
249
  border: 1px solid var(--line);
@@ -564,7 +682,7 @@
564
682
  cursor: not-allowed;
565
683
  }
566
684
 
567
- @media (max-width: 640px) {
685
+ @media (max-width: 760px) {
568
686
  .wrap {
569
687
  padding: 28px 16px 170px;
570
688
  }
@@ -572,6 +690,15 @@
572
690
  padding: 12px 16px;
573
691
  gap: 12px;
574
692
  }
693
+ .run-context {
694
+ grid-template-columns: minmax(0, 1fr);
695
+ }
696
+ .frow .fl {
697
+ width: 92px;
698
+ }
699
+ .fdrop {
700
+ margin-left: 0;
701
+ }
575
702
  }
576
703
  </style>
577
704
  </head>
@@ -584,7 +711,19 @@
584
711
  </div>
585
712
  <div class="runline" id="runline"></div>
586
713
 
587
- <div class="statrow" id="statrow"></div>
714
+ <div class="statrow" id="statrow" hidden></div>
715
+
716
+ <div class="run-context" id="ctx" hidden>
717
+ <div class="ctx-funnel">
718
+ <div class="funnel-head">
719
+ <span class="t">Gap funnel · counted across runs</span>
720
+ <span class="facts num" id="funnel-floor"></span>
721
+ </div>
722
+ <div id="funnel-bars"></div>
723
+ <div class="funnel-note" id="funnel-note" hidden></div>
724
+ </div>
725
+ <div class="ctx-stats" id="ctx-stats"></div>
726
+ </div>
588
727
 
589
728
  <div class="gauge-block">
590
729
  <div class="gauge-head">
@@ -668,26 +807,142 @@
668
807
  document.createTextNode(" transcripts" + (harnessBits.length ? " (" + harnessBits.join(" · ") + ")" : "")),
669
808
  );
670
809
 
671
- // ---- stats ----
672
- var stats = [
810
+ // ---- run context: one frame, the gap funnel beside the run stats ----
811
+ // A single bordered band. The left pane tells the gap story as proportional
812
+ // bars - sightings recorded by analysis -> project-domain (orchestration
813
+ // excluded) -> distinct gaps after clustering -> eligible to propose - with
814
+ // each drop-off explained in plain words. The right pane holds the run facts
815
+ // (edits, evidence, skills) as stat cells; a separate stat row would duplicate
816
+ // the funnel's last stage. Every count is recorded by the fold (`totals` in
817
+ // evidence-summary.json); a proposal saved before the counts existed falls
818
+ // back to the classic stat row rather than showing zeros nobody measured.
819
+ function funnelCounts() {
820
+ if (typeof (P.stats || {}).gapSightings !== "number") return null;
821
+ var eligible = P.stats.gapClusters || 0;
822
+ var dropped = P.stats.droppedGapSingletons || 0;
823
+ var orch = P.stats.orchestrationGapSightings || 0;
824
+ var project = Math.max(0, P.stats.gapSightings - orch);
825
+ return {
826
+ floor: Number((P.config || {}).minGapEvidence) || 2,
827
+ sighted: P.stats.gapSightings,
828
+ orch: orch,
829
+ project: project,
830
+ distinct: eligible + dropped,
831
+ merged: Math.max(0, project - (eligible + dropped)),
832
+ dropped: dropped,
833
+ eligible: eligible,
834
+ };
835
+ }
836
+ function funnelBar(host, value, label, max, lit) {
837
+ var row = el("div", "frow" + (value === 0 ? " zero" : lit ? " lit" : ""));
838
+ row.appendChild(el("span", "fl", label));
839
+ var track = el("div", "fb");
840
+ var fill = el("i");
841
+ fill.style.width = max > 0 ? Math.max((value / max) * 100, value > 0 ? 2 : 0) + "%" : "0%";
842
+ track.appendChild(fill);
843
+ row.appendChild(track);
844
+ row.appendChild(el("span", "fv num", fmt(value)));
845
+ host.appendChild(row);
846
+ }
847
+ function funnelDrop(host, n, text) {
848
+ if (!n) return;
849
+ var line = el("div", "fdrop");
850
+ line.appendChild(el("b", null, "− " + fmt(n) + " "));
851
+ line.appendChild(document.createTextNode(text));
852
+ host.appendChild(line);
853
+ }
854
+ var funnelNote = null;
855
+ var funnel = funnelCounts();
856
+ if (funnel) {
857
+ var bars = document.getElementById("funnel-bars");
858
+ funnelBar(bars, funnel.sighted, "sightings", funnel.sighted);
859
+ funnelDrop(
860
+ bars,
861
+ funnel.orch,
862
+ "orchestration: mistakes in how the task was run, not in this project - excluded from proposals by design",
863
+ );
864
+ funnelBar(bars, funnel.project, "project-domain", funnel.sighted);
865
+ funnelDrop(
866
+ bars,
867
+ funnel.merged,
868
+ "merged: repeat sightings of a gap already counted - they corroborate it rather than add a new one",
869
+ );
870
+ funnelBar(bars, funnel.distinct, "distinct gaps", funnel.sighted);
871
+ funnelDrop(
872
+ bars,
873
+ funnel.dropped,
874
+ "below the corroboration floor: seen in fewer than " +
875
+ funnel.floor +
876
+ " sessions so far - more transcripts can corroborate them",
877
+ );
878
+ funnelBar(bars, funnel.eligible, "eligible to propose", funnel.sighted, true);
879
+
880
+ document.getElementById("funnel-floor").textContent = "corroboration floor · " + funnel.floor + " sessions";
881
+ var statHost = document.getElementById("ctx-stats");
673
882
  [
674
- String(edits.length) + "/" + String((P.config || {}).maxEditsPerRun || edits.length),
675
- "edits proposed / cap",
676
- null,
677
- ],
678
- [fmt(P.stats.positive), "positive evidence", "accent"],
679
- [fmt(P.stats.negative), "negative evidence", "reject"],
680
- [fmt(P.stats.gapClusters), "gap clusters", "defer"],
681
- [fmt(P.stats.skillExtractions), "skill extractions", null],
682
- ];
683
- var statrow = document.getElementById("statrow");
684
- stats.forEach(function (s) {
685
- var box = el("div", "stat");
686
- var v = el("div", "v num", s[0]);
687
- box.appendChild(v);
688
- box.appendChild(el("div", "k", s[1]));
689
- statrow.appendChild(box);
690
- });
883
+ [
884
+ String(edits.length) + "/" + String((P.config || {}).maxEditsPerRun || edits.length),
885
+ "edits proposed / cap",
886
+ ],
887
+ [fmt(P.stats.positive), "positive evidence"],
888
+ [fmt(P.stats.negative), "negative evidence"],
889
+ [fmt(P.stats.skillExtractions), "skill extractions"],
890
+ ].forEach(function (s) {
891
+ var box = el("div", "stat");
892
+ box.appendChild(el("div", "v num", s[0]));
893
+ box.appendChild(el("div", "k", s[1]));
894
+ statHost.appendChild(box);
895
+ });
896
+
897
+ // The drop lines carry the takeaway; the visible note only speaks when there
898
+ // are no bars to read. funnelNote itself still feeds the zero-edit empty state.
899
+ if (!funnel.sighted) {
900
+ funnelNote = "Analysis reported no uncovered gaps.";
901
+ } else if (!funnel.eligible && funnel.dropped) {
902
+ funnelNote =
903
+ funnel.dropped +
904
+ " distinct gap" +
905
+ (funnel.dropped === 1 ? " is" : "s are") +
906
+ " still below the " +
907
+ funnel.floor +
908
+ "-session corroboration floor - more transcripts may corroborate " +
909
+ (funnel.dropped === 1 ? "it." : "them.");
910
+ } else if (!funnel.eligible && funnel.orch >= funnel.sighted) {
911
+ funnelNote =
912
+ "All " +
913
+ funnel.sighted +
914
+ " sighting" +
915
+ (funnel.sighted === 1 ? " was" : "s were") +
916
+ " orchestration-domain (caused by the orchestrating harness, not this repository) and " +
917
+ (funnel.sighted === 1 ? "is" : "are") +
918
+ " excluded by design.";
919
+ }
920
+ if (funnelNote && !funnel.sighted) {
921
+ var noteEl = document.getElementById("funnel-note");
922
+ noteEl.textContent = funnelNote;
923
+ noteEl.hidden = false;
924
+ }
925
+ document.getElementById("ctx").hidden = false;
926
+ } else {
927
+ // A proposal from before the funnel counts existed: the classic stat row.
928
+ var statrow = document.getElementById("statrow");
929
+ statrow.hidden = false;
930
+ [
931
+ [
932
+ String(edits.length) + "/" + String((P.config || {}).maxEditsPerRun || edits.length),
933
+ "edits proposed / cap",
934
+ ],
935
+ [fmt(P.stats.positive), "positive evidence"],
936
+ [fmt(P.stats.negative), "negative evidence"],
937
+ [fmt(P.stats.gapClusters), "eligible gaps"],
938
+ [fmt(P.stats.skillExtractions), "skill extractions"],
939
+ ].forEach(function (s) {
940
+ var box = el("div", "stat");
941
+ box.appendChild(el("div", "v num", s[0]));
942
+ box.appendChild(el("div", "k", s[1]));
943
+ statrow.appendChild(box);
944
+ });
945
+ }
691
946
 
692
947
  // ---- budget gauge ----
693
948
  document.getElementById("gauge-title").textContent = "Always-loaded budget · " + (P.memoryFile.path || "");
@@ -700,8 +955,21 @@
700
955
  edits.length + " of " + ((P.config || {}).maxEditsPerRun || edits.length) + " (learning-rate cap)";
701
956
 
702
957
  if (!edits.length) {
958
+ var eligibleWithoutEdits = Number((P.stats || {}).gapClusters) || 0;
959
+ var emptyReason = funnelNote;
960
+ if (!emptyReason && eligibleWithoutEdits > 0) {
961
+ emptyReason =
962
+ fmt(eligibleWithoutEdits) +
963
+ " eligible gap" +
964
+ (eligibleWithoutEdits === 1 ? "" : "s") +
965
+ " reached synthesis, but no edit was proposed this run.";
966
+ }
703
967
  host.appendChild(
704
- el("div", "empty", "No edits proposed. The evidence did not clear the thresholds this run."),
968
+ el(
969
+ "div",
970
+ "empty",
971
+ "No edits proposed. " + (emptyReason || "The evidence did not clear the thresholds this run."),
972
+ ),
705
973
  );
706
974
  }
707
975