reeval 0.0.1
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/LICENSE +202 -0
- package/THIRD_PARTY_NOTICES.md +2160 -0
- package/assets/package.json +9 -0
- package/assets/pnpm-lock.yaml +1336 -0
- package/dist/agent-DVHkx8aT.js +532 -0
- package/dist/agent-DVHkx8aT.js.map +1 -0
- package/dist/analysis-CxSCt1YY.js +16 -0
- package/dist/analysis-CxSCt1YY.js.map +1 -0
- package/dist/assemble-B-nmz8oH.js +695 -0
- package/dist/assemble-B-nmz8oH.js.map +1 -0
- package/dist/assemble-CmRVDZ2Q.js +2 -0
- package/dist/attempt-BdRQ9jKX.js +176 -0
- package/dist/attempt-BdRQ9jKX.js.map +1 -0
- package/dist/bridge/index.mjs +1911 -0
- package/dist/bridge/package.json +13 -0
- package/dist/bridge/pnpm-lock.yaml +1041 -0
- package/dist/bridge/pnpm-workspace.yaml +2 -0
- package/dist/budget-DBUrH4jX.js +8 -0
- package/dist/budget-DBUrH4jX.js.map +1 -0
- package/dist/cli--CQ3Iy9q.js +1865 -0
- package/dist/cli--CQ3Iy9q.js.map +1 -0
- package/dist/cli-spec-DKAu09Gi.js +478 -0
- package/dist/cli-spec-DKAu09Gi.js.map +1 -0
- package/dist/cli.d.ts +1 -0
- package/dist/cli.js +8 -0
- package/dist/cli.js.map +1 -0
- package/dist/colors-QV9dFVhN.js +34 -0
- package/dist/colors-QV9dFVhN.js.map +1 -0
- package/dist/command-BSJzHMtP.js +280 -0
- package/dist/command-BSJzHMtP.js.map +1 -0
- package/dist/command-CEm8PZ1N.js +1043 -0
- package/dist/command-CEm8PZ1N.js.map +1 -0
- package/dist/command-D-jnQWVG.js +673 -0
- package/dist/command-D-jnQWVG.js.map +1 -0
- package/dist/command-DQAvbG7d.js +432 -0
- package/dist/command-DQAvbG7d.js.map +1 -0
- package/dist/commands-A4udfyKZ.js +2 -0
- package/dist/commands-CF8mOfiu.js +207 -0
- package/dist/commands-CF8mOfiu.js.map +1 -0
- package/dist/config-Cuj0kVYj.js +415 -0
- package/dist/config-Cuj0kVYj.js.map +1 -0
- package/dist/context-loi9OCZv.js +106 -0
- package/dist/context-loi9OCZv.js.map +1 -0
- package/dist/core-tLN-z1Be.js +96 -0
- package/dist/core-tLN-z1Be.js.map +1 -0
- package/dist/db-Cjj2ZsfQ.js +2 -0
- package/dist/db-DxDzMhtB.js +827 -0
- package/dist/db-DxDzMhtB.js.map +1 -0
- package/dist/debug-D24NOQH1.js +89 -0
- package/dist/debug-D24NOQH1.js.map +1 -0
- package/dist/debug-SEmW4w8f.js +2 -0
- package/dist/dist-BZXYHN2X.js +1673 -0
- package/dist/dist-BZXYHN2X.js.map +1 -0
- package/dist/docker-DQDZWXF9.js +20 -0
- package/dist/docker-DQDZWXF9.js.map +1 -0
- package/dist/docker-DnWAVdEm.js +2 -0
- package/dist/docker-oSbCExHL.js +153 -0
- package/dist/docker-oSbCExHL.js.map +1 -0
- package/dist/doctor-CxlH1pDM.js +525 -0
- package/dist/doctor-CxlH1pDM.js.map +1 -0
- package/dist/egress-73SQfKdx.js +619 -0
- package/dist/egress-73SQfKdx.js.map +1 -0
- package/dist/env-BQqj93gu.js +21 -0
- package/dist/env-BQqj93gu.js.map +1 -0
- package/dist/environment-Mz7wvrtD.js +552 -0
- package/dist/environment-Mz7wvrtD.js.map +1 -0
- package/dist/errors-C9w_ud7g.js +160 -0
- package/dist/errors-C9w_ud7g.js.map +1 -0
- package/dist/errors-CSxfJ4qu.js +11 -0
- package/dist/errors-CSxfJ4qu.js.map +1 -0
- package/dist/files--6CEUXm7.js +2 -0
- package/dist/files-BLlrhU4-.js +62 -0
- package/dist/files-BLlrhU4-.js.map +1 -0
- package/dist/files-Bd25OAQH.js +2 -0
- package/dist/files-BvHCFWgD.js +2 -0
- package/dist/files-D2-qn_Dl.js +275 -0
- package/dist/files-D2-qn_Dl.js.map +1 -0
- package/dist/files-bF2fPuT9.js +99 -0
- package/dist/files-bF2fPuT9.js.map +1 -0
- package/dist/git-D6hMeieS.js +44 -0
- package/dist/git-D6hMeieS.js.map +1 -0
- package/dist/harness-ls5aX1UO.js +131 -0
- package/dist/harness-ls5aX1UO.js.map +1 -0
- package/dist/harness.d.ts +74 -0
- package/dist/harness.js +2 -0
- package/dist/header-DzFDoMCI.js +78 -0
- package/dist/header-DzFDoMCI.js.map +1 -0
- package/dist/ids-Bp6ut7ZD.js +49 -0
- package/dist/ids-Bp6ut7ZD.js.map +1 -0
- package/dist/images-Cc1-coFc.js +286 -0
- package/dist/images-Cc1-coFc.js.map +1 -0
- package/dist/index.d.ts +97 -0
- package/dist/index.js +1959 -0
- package/dist/index.js.map +1 -0
- package/dist/init-CFljS8BJ.js +344 -0
- package/dist/init-CFljS8BJ.js.map +1 -0
- package/dist/init-z2ZPn9Tj.js +2 -0
- package/dist/invocations-rXktgSAn.js +157 -0
- package/dist/invocations-rXktgSAn.js.map +1 -0
- package/dist/jiti-BOLXpKyG.js +46 -0
- package/dist/jiti-BOLXpKyG.js.map +1 -0
- package/dist/judge-D8Uf4dA9.js +146 -0
- package/dist/judge-D8Uf4dA9.js.map +1 -0
- package/dist/judge.d.ts +21 -0
- package/dist/judge.js +19 -0
- package/dist/judge.js.map +1 -0
- package/dist/live-Bs3LReMV.js +98 -0
- package/dist/live-Bs3LReMV.js.map +1 -0
- package/dist/matchers-DyLvm7EG.js +374 -0
- package/dist/matchers-DyLvm7EG.js.map +1 -0
- package/dist/model-BdO79gWz.js +44 -0
- package/dist/model-BdO79gWz.js.map +1 -0
- package/dist/model-CDL98xue.js +725 -0
- package/dist/model-CDL98xue.js.map +1 -0
- package/dist/model-v-0_Ik7S.js +2 -0
- package/dist/node-kG64XxdE.js +506 -0
- package/dist/node-kG64XxdE.js.map +1 -0
- package/dist/opencode-providers-DZuD_oYt.js +3149 -0
- package/dist/opencode-providers-DZuD_oYt.js.map +1 -0
- package/dist/options-CAQhu--C.js +452 -0
- package/dist/options-CAQhu--C.js.map +1 -0
- package/dist/patch-BkKje4cp.js +69 -0
- package/dist/patch-BkKje4cp.js.map +1 -0
- package/dist/paths-CBvm6pap.js +72 -0
- package/dist/paths-CBvm6pap.js.map +1 -0
- package/dist/plan-gwk6dJ8c.js +237 -0
- package/dist/plan-gwk6dJ8c.js.map +1 -0
- package/dist/preflight-BtubXKQ8.js +2 -0
- package/dist/preflight-BypZMNBG.js +185 -0
- package/dist/preflight-BypZMNBG.js.map +1 -0
- package/dist/prepare-Cwv0bHs_.js +3 -0
- package/dist/prepare-D6Uzcubb.js +213 -0
- package/dist/prepare-D6Uzcubb.js.map +1 -0
- package/dist/provenance-Dfb8YY5Y.js +129 -0
- package/dist/provenance-Dfb8YY5Y.js.map +1 -0
- package/dist/provider-BDbNOQhQ.js +2 -0
- package/dist/provider-CmAYTQQs.js +14 -0
- package/dist/provider-CmAYTQQs.js.map +1 -0
- package/dist/providers-2p3ylMHI.js +107 -0
- package/dist/providers-2p3ylMHI.js.map +1 -0
- package/dist/providers-BUOxitcT.js +2 -0
- package/dist/query-CFATHZmV.js +91 -0
- package/dist/query-CFATHZmV.js.map +1 -0
- package/dist/query-DR4m6CXf.js +2 -0
- package/dist/record-DQ8wzyco.js +99 -0
- package/dist/record-DQ8wzyco.js.map +1 -0
- package/dist/redact-C4fL_bZP.js +167 -0
- package/dist/redact-C4fL_bZP.js.map +1 -0
- package/dist/results-CeXb8Wde.js +37 -0
- package/dist/results-CeXb8Wde.js.map +1 -0
- package/dist/root-CFDjpWuc.js +110 -0
- package/dist/root-CFDjpWuc.js.map +1 -0
- package/dist/sandbox-By2gkdkC.js +133 -0
- package/dist/sandbox-By2gkdkC.js.map +1 -0
- package/dist/sandbox.d.ts +41 -0
- package/dist/sandbox.js +2 -0
- package/dist/server-Bi61uiY0.js +350 -0
- package/dist/server-Bi61uiY0.js.map +1 -0
- package/dist/server-CsSBL8NC.js +2 -0
- package/dist/source-BXzrmof5.js +2 -0
- package/dist/source-Bsfa3LZH.js +40 -0
- package/dist/source-Bsfa3LZH.js.map +1 -0
- package/dist/src-55j1zNHH.js +3 -0
- package/dist/src-AEwxd1UP.js +602 -0
- package/dist/src-AEwxd1UP.js.map +1 -0
- package/dist/src-BDwSuyIo.js +3 -0
- package/dist/src-BQyMiWri.js +490 -0
- package/dist/src-BQyMiWri.js.map +1 -0
- package/dist/src-C-LHyTSV.js +3 -0
- package/dist/src-Cmp-7YrD.js +1136 -0
- package/dist/src-Cmp-7YrD.js.map +1 -0
- package/dist/src-D0RdqpNb.js +252 -0
- package/dist/src-D0RdqpNb.js.map +1 -0
- package/dist/src-D8KYzxBX.js +793 -0
- package/dist/src-D8KYzxBX.js.map +1 -0
- package/dist/src-DEgff9Z7.js +4 -0
- package/dist/src-DXx_29pq.js +567 -0
- package/dist/src-DXx_29pq.js.map +1 -0
- package/dist/src-DwwiVLqv.js +2937 -0
- package/dist/src-DwwiVLqv.js.map +1 -0
- package/dist/src-N7E_mfcZ.js +2 -0
- package/dist/src-W9g-y5X5.js +3 -0
- package/dist/src-daS0wKH2.js +3 -0
- package/dist/state-Cyw3QTsN.js +2 -0
- package/dist/state-DZ7eqb1F.js +167 -0
- package/dist/state-DZ7eqb1F.js.map +1 -0
- package/dist/stats-BNAHlNMD.js +114 -0
- package/dist/stats-BNAHlNMD.js.map +1 -0
- package/dist/studio-app/assets/index-D6XyfFSH.js +58 -0
- package/dist/studio-app/assets/index-xwzWdcE_.css +2 -0
- package/dist/studio-app/favicon.svg +1 -0
- package/dist/studio-app/index.html +15 -0
- package/dist/summary-D7Ttky0z.js +166 -0
- package/dist/summary-D7Ttky0z.js.map +1 -0
- package/dist/text-gEcKjAyp.js +29 -0
- package/dist/text-gEcKjAyp.js.map +1 -0
- package/dist/types-C5vI3g1C.js +9 -0
- package/dist/types-C5vI3g1C.js.map +1 -0
- package/dist/types-DLec-6Cx.d.ts +1535 -0
- package/dist/usage-DsYUqIUM.js +24 -0
- package/dist/usage-DsYUqIUM.js.map +1 -0
- package/dist/view/index.d.ts +645 -0
- package/dist/view/index.js +8 -0
- package/dist/views-BrHOv6wg.js +220 -0
- package/dist/views-BrHOv6wg.js.map +1 -0
- package/dist/views-CwCWLWqo.js +3 -0
- package/dist/worker-entry-DnYkANPG.js +506 -0
- package/dist/worker-entry-DnYkANPG.js.map +1 -0
- package/dist/workspace-C2S0Z6gl.js +327 -0
- package/dist/workspace-C2S0Z6gl.js.map +1 -0
- package/dist/workspace-CyoXG6CY.js +2 -0
- package/docker/Dockerfile +19 -0
- package/package.json +79 -0
|
@@ -0,0 +1,220 @@
|
|
|
1
|
+
import { r as formatDuration } from "./text-gEcKjAyp.js";
|
|
2
|
+
import { M as effortChangeLabel, V as cellsChanged, W as stackChanges } from "./model-CDL98xue.js";
|
|
3
|
+
import { o as warnings } from "./state-DZ7eqb1F.js";
|
|
4
|
+
import { c as summaryCounts, n as cellCounts, o as repeatsVerdict, s as stateRows, t as OUTCOME_MARK } from "./summary-D7Ttky0z.js";
|
|
5
|
+
//#region src/history/views.ts
|
|
6
|
+
var RefError = class extends Error {
|
|
7
|
+
name = "RefError";
|
|
8
|
+
};
|
|
9
|
+
/**
|
|
10
|
+
* `latest`, `latest~N`, a run id or unique prefix of one, or a git SHA prefix (the newest run
|
|
11
|
+
* at that commit). `entries` are sorted oldest first.
|
|
12
|
+
*/
|
|
13
|
+
function resolveRun(entries, ref = "latest") {
|
|
14
|
+
if (entries.length === 0) throw new RefError("No runs yet. Run reeval run.");
|
|
15
|
+
const back = /^latest(?:~(\d+))?$/.exec(ref);
|
|
16
|
+
if (back) {
|
|
17
|
+
const entry = entries[entries.length - 1 - Number(back[1] ?? 0)];
|
|
18
|
+
if (!entry) throw new RefError(`There are only ${entries.length} runs`);
|
|
19
|
+
return entry;
|
|
20
|
+
}
|
|
21
|
+
const exact = entries.find((e) => e.runId === ref);
|
|
22
|
+
if (exact) return exact;
|
|
23
|
+
const byId = entries.filter((e) => e.runId.startsWith(ref));
|
|
24
|
+
if (byId.length === 1 && byId[0]) return byId[0];
|
|
25
|
+
if (byId.length > 1) throw new RefError(`"${ref}" matches ${byId.length} runs; use more of the id`);
|
|
26
|
+
if (/^[0-9a-f]{4,40}$/i.test(ref)) {
|
|
27
|
+
const atSha = entries.filter((e) => e.git?.sha.startsWith(ref.toLowerCase()));
|
|
28
|
+
const newest = atSha[atSha.length - 1];
|
|
29
|
+
if (newest) return newest;
|
|
30
|
+
}
|
|
31
|
+
throw new RefError(`No run matches "${ref}"`);
|
|
32
|
+
}
|
|
33
|
+
/**
|
|
34
|
+
* The run to compare `head` against: a complete, earlier run on the first commit of `ancestry` that
|
|
35
|
+
* has one, newest run first; otherwise the previous complete run by time. `ancestry` is `git
|
|
36
|
+
* rev-list <head's commit>`: head's own commit first, then every ancestor through all parents
|
|
37
|
+
* (merged branches too, not only first parents), newest commit first. Ancestry beats wall-clock
|
|
38
|
+
* time, which can be skewed.
|
|
39
|
+
*/
|
|
40
|
+
function pickBaseline(entries, head, ancestry) {
|
|
41
|
+
const earlier = entries.filter((e) => e.runId < head.runId && e.status === "complete");
|
|
42
|
+
if (ancestry && ancestry.length > 0) {
|
|
43
|
+
const rank = new Map(ancestry.map((sha, i) => [sha, i]));
|
|
44
|
+
let best = null;
|
|
45
|
+
let bestRank = Number.POSITIVE_INFINITY;
|
|
46
|
+
for (const e of earlier) {
|
|
47
|
+
const r = e.git ? rank.get(e.git.sha) : void 0;
|
|
48
|
+
if (r !== void 0 && (r < bestRank || r === bestRank && best && e.runId > best.runId)) {
|
|
49
|
+
best = e;
|
|
50
|
+
bestRank = r;
|
|
51
|
+
}
|
|
52
|
+
}
|
|
53
|
+
if (best) return best;
|
|
54
|
+
}
|
|
55
|
+
return earlier[earlier.length - 1] ?? null;
|
|
56
|
+
}
|
|
57
|
+
const shortSha = (e) => e.git ? e.git.sha.slice(0, 7) : "no-git";
|
|
58
|
+
const unreachable = (entry, reach) => !!reach && !!entry.git && !reach.has(entry.git.sha);
|
|
59
|
+
function history(entries, options = {}) {
|
|
60
|
+
const runs = entries.slice(-(options.limit ?? 20));
|
|
61
|
+
const ids = [];
|
|
62
|
+
for (const run of [...runs].reverse()) for (const e of run.evals) if (!ids.includes(e.id)) ids.push(e.id);
|
|
63
|
+
const needle = options.filter?.toLowerCase();
|
|
64
|
+
return {
|
|
65
|
+
runs,
|
|
66
|
+
rows: ids.filter((id) => !needle || id.toLowerCase().includes(needle)).map((id) => {
|
|
67
|
+
const row = {
|
|
68
|
+
id,
|
|
69
|
+
outcomes: [],
|
|
70
|
+
drift: [],
|
|
71
|
+
pass: 0,
|
|
72
|
+
counted: 0
|
|
73
|
+
};
|
|
74
|
+
const seen = /* @__PURE__ */ new Map();
|
|
75
|
+
runs.forEach((run, i) => {
|
|
76
|
+
const e = run.evals.find((x) => x.id === id);
|
|
77
|
+
row.outcomes.push(e?.outcome ?? null);
|
|
78
|
+
if (!e) return;
|
|
79
|
+
if (cellsChanged(seen, id, e.cells)) row.drift.push(i);
|
|
80
|
+
if (e.outcome === "pass" || e.outcome === "fail") row.counted++;
|
|
81
|
+
if (e.outcome === "pass") row.pass++;
|
|
82
|
+
});
|
|
83
|
+
return row;
|
|
84
|
+
}),
|
|
85
|
+
orphaned: runs.filter((r) => unreachable(r, options.reach ?? null)).map((r) => r.runId)
|
|
86
|
+
};
|
|
87
|
+
}
|
|
88
|
+
function formatHistory({ runs, rows, orphaned }) {
|
|
89
|
+
if (runs.length === 0) return "No recorded runs yet. Run reeval run --record, or reeval record latest.";
|
|
90
|
+
const first = runs[0];
|
|
91
|
+
const last = runs[runs.length - 1];
|
|
92
|
+
const lines = [`${runs.length} run${runs.length === 1 ? "" : "s"}, oldest → newest: ${first.runId} (${shortSha(first)}) … ${last.runId} (${shortSha(last)})`, ""];
|
|
93
|
+
const width = Math.max(...rows.map((r) => r.outcomes.length + r.drift.length), 0);
|
|
94
|
+
for (const row of rows) {
|
|
95
|
+
const strip = row.outcomes.map((o, i) => `${row.drift.includes(i) ? "¦" : ""}${o ? OUTCOME_MARK[o] : "·"}`).join("");
|
|
96
|
+
const rate = row.counted ? `${row.pass}/${row.counted}` : "–";
|
|
97
|
+
lines.push(` ${strip.padEnd(width)} ${rate.padStart(5)} ${row.id}`);
|
|
98
|
+
}
|
|
99
|
+
lines.push("", "✓ pass ✗ fail ! infra error ↓ skipped · not run ¦ eval or stack changed");
|
|
100
|
+
if (orphaned.length > 0) lines.push("", `${orphaned.length} run${orphaned.length === 1 ? " was" : "s were"} recorded on commits that aren't in HEAD's history (rebased or dropped?):`, ...orphaned.map((id) => ` ${id}`));
|
|
101
|
+
return lines.join("\n");
|
|
102
|
+
}
|
|
103
|
+
const STATUS_LABEL = {
|
|
104
|
+
regressed: "regressed",
|
|
105
|
+
fixed: "fixed",
|
|
106
|
+
same: "same",
|
|
107
|
+
drift: "changed (not comparable)",
|
|
108
|
+
added: "new",
|
|
109
|
+
removed: "removed",
|
|
110
|
+
infra: "infra error"
|
|
111
|
+
};
|
|
112
|
+
/**
|
|
113
|
+
* `: evals/helpers.ts changed, harness config changed` for a row that isn't comparable, and
|
|
114
|
+
* ` · cells removed: …` when the runs ran different cells of the eval.
|
|
115
|
+
*/
|
|
116
|
+
const why = (r) => [
|
|
117
|
+
r.changes?.length ? `: ${r.changes.join(", ")}` : "",
|
|
118
|
+
r.cells?.added.length ? ` · cells added: ${r.cells.added.join(", ")}` : "",
|
|
119
|
+
r.cells?.removed.length ? ` · cells removed: ${r.cells.removed.join(", ")}` : ""
|
|
120
|
+
].join("");
|
|
121
|
+
const ORDER = [
|
|
122
|
+
"regressed",
|
|
123
|
+
"fixed",
|
|
124
|
+
"drift",
|
|
125
|
+
"infra",
|
|
126
|
+
"added",
|
|
127
|
+
"removed",
|
|
128
|
+
"same"
|
|
129
|
+
];
|
|
130
|
+
const describeRun = (e, reach) => `${e.runId} (${shortSha(e)}${e.recorded ? "" : ", not recorded"}${unreachable(e, reach) ? ", commit not in HEAD's history" : ""})`;
|
|
131
|
+
function formatCompare(base, head, rows, md, reach = null) {
|
|
132
|
+
const sorted = [...rows].sort((a, b) => ORDER.indexOf(a.status) - ORDER.indexOf(b.status));
|
|
133
|
+
const mark = (o) => o ? OUTCOME_MARK[o] : "·";
|
|
134
|
+
const count = (s) => rows.filter((r) => r.status === s).length;
|
|
135
|
+
const tally = `${count("regressed")} regressed · ${count("fixed")} fixed · ${count("drift")} not comparable · ${count("same")} same`;
|
|
136
|
+
const changes = stackChanges(base, head);
|
|
137
|
+
const changeText = changes.map((c) => `${c.label} ${c.base} → ${c.head}`).join("; ");
|
|
138
|
+
const esc = (s) => s.replace(/\|/g, "\\|");
|
|
139
|
+
const cells = sorted.flatMap((r) => (r.perCell ?? []).map((c) => ({
|
|
140
|
+
row: r,
|
|
141
|
+
cell: c
|
|
142
|
+
})));
|
|
143
|
+
if (md) return [
|
|
144
|
+
`### reeval: ${head.runId} vs ${base.runId}`,
|
|
145
|
+
"",
|
|
146
|
+
`Base ${describeRun(base, reach)} → head ${describeRun(head, reach)}. ${tally}.`,
|
|
147
|
+
"",
|
|
148
|
+
...changes.length ? [`**Stack changed:** ${changeText}.`, ""] : [],
|
|
149
|
+
"| Eval | Base | Head | Change |",
|
|
150
|
+
"|---|---|---|---|",
|
|
151
|
+
...sorted.map((r) => `| ${esc(r.id)} | ${mark(r.base)} | ${mark(r.head)} | ${STATUS_LABEL[r.status]}${why(r)} |`),
|
|
152
|
+
"",
|
|
153
|
+
...cells.length ? [
|
|
154
|
+
"| Eval | Cell | Lift, base → head | Effort, base → head |",
|
|
155
|
+
"|---|---|---|---|",
|
|
156
|
+
...cells.map(({ row, cell }) => `| ${esc(row.id)} | ${esc(cell.key)} | ${liftChange(cell)} | ${effortChange(cell)} |`),
|
|
157
|
+
""
|
|
158
|
+
] : []
|
|
159
|
+
].join("\n");
|
|
160
|
+
return [
|
|
161
|
+
`base ${describeRun(base, reach)} → head ${describeRun(head, reach)}`,
|
|
162
|
+
tally,
|
|
163
|
+
...changes.length ? [`stack changed: ${changeText}`] : [],
|
|
164
|
+
"",
|
|
165
|
+
...sorted.flatMap((r) => [` ${mark(r.base)} → ${mark(r.head)} ${STATUS_LABEL[r.status].padEnd(24)} ${r.id}${why(r)}`, ...(r.perCell ?? []).map((c) => ` ${c.key} ${[c.base.lift || c.head.lift ? `lift ${liftChange(c)}` : "", effortChange(c)].filter(Boolean).join(" · ")}`)])
|
|
166
|
+
].join("\n");
|
|
167
|
+
}
|
|
168
|
+
/** `reeval compare` when head has no earlier recorded run: every eval is new. */
|
|
169
|
+
function formatNoBaseline(head, rows, md) {
|
|
170
|
+
const failing = rows.filter((r) => r.head === "fail").length;
|
|
171
|
+
const note = `No earlier recorded run to compare ${head.runId} with: every eval is new${failing ? `, and ${failing} fail${failing === 1 ? "s" : ""}` : ""}.`;
|
|
172
|
+
const mark = (o) => o ? OUTCOME_MARK[o] : "·";
|
|
173
|
+
if (md) return [
|
|
174
|
+
`### reeval: ${head.runId}`,
|
|
175
|
+
"",
|
|
176
|
+
note,
|
|
177
|
+
"",
|
|
178
|
+
"| Eval | Head |",
|
|
179
|
+
"|---|---|",
|
|
180
|
+
...rows.map((r) => `| ${r.id.replace(/\|/g, "\\|")} | ${mark(r.head)} |`),
|
|
181
|
+
""
|
|
182
|
+
].join("\n");
|
|
183
|
+
return [
|
|
184
|
+
note,
|
|
185
|
+
"",
|
|
186
|
+
...rows.map((r) => ` ${mark(r.head)} new ${r.id}`)
|
|
187
|
+
].join("\n");
|
|
188
|
+
}
|
|
189
|
+
const liftShort = (l) => l ? `${signed(l.points)} (${signed(l.lo)}…${signed(l.hi)})${l.withinNoise ? " noise" : ""}` : "–";
|
|
190
|
+
const signed = (x) => `${x >= 0 ? "+" : "−"}${Math.round(Math.abs(x) * 100)}`;
|
|
191
|
+
const liftChange = (c) => `${liftShort(c.base.lift)} → ${liftShort(c.head.lift)}`;
|
|
192
|
+
const effortChange = (c) => c.base.effort && c.head.effort ? effortChangeLabel(c.base.effort, c.head.effort) : "effort –";
|
|
193
|
+
function formatShow(run) {
|
|
194
|
+
const rows = stateRows(run);
|
|
195
|
+
const width = Math.max(...rows.map(([k]) => k.length));
|
|
196
|
+
const lines = [
|
|
197
|
+
`run ${run.runId} ${summaryCounts(run.summary)}${run.status === "interrupted" ? " (interrupted)" : ""}`,
|
|
198
|
+
` started ${run.startedAt} (${formatDuration(run.durationMs)}) · artifacts ${run.artifacts}`,
|
|
199
|
+
"",
|
|
200
|
+
...rows.map(([k, v]) => ` ${k.padEnd(width)} ${v}`),
|
|
201
|
+
...warnings(run).map((w) => ` ! ${w}`),
|
|
202
|
+
""
|
|
203
|
+
];
|
|
204
|
+
for (const e of run.evals) {
|
|
205
|
+
const single = e.attempts.length === 1;
|
|
206
|
+
const first = e.attempts[0];
|
|
207
|
+
lines.push(` ${OUTCOME_MARK[e.outcome]} ${e.id}${single && first ? ` ${formatDuration(first.durationMs)}` : ` ${cellCounts(e)}`}`);
|
|
208
|
+
for (const t of e.attempts) {
|
|
209
|
+
const indent = single ? " " : " ";
|
|
210
|
+
if (!single) lines.push(` ${OUTCOME_MARK[t.outcome]} ${t.id} ${formatDuration(t.durationMs)}`);
|
|
211
|
+
for (const j of t.judgements) lines.push(`${indent}judge ${j.verdict} ${j.score.toFixed(2)} (≥ ${j.threshold}) ${j.question}`);
|
|
212
|
+
if (t.error && !repeatsVerdict(t)) lines.push(`${indent}${t.error.kind}: ${t.error.message.split("\n")[0]}`);
|
|
213
|
+
}
|
|
214
|
+
}
|
|
215
|
+
return lines.join("\n");
|
|
216
|
+
}
|
|
217
|
+
//#endregion
|
|
218
|
+
export { formatShow as a, resolveRun as c, formatNoBaseline as i, unreachable as l, formatCompare as n, history as o, formatHistory as r, pickBaseline as s, RefError as t };
|
|
219
|
+
|
|
220
|
+
//# sourceMappingURL=views-BrHOv6wg.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"views-BrHOv6wg.js","names":[],"sources":["../src/history/views.ts"],"mappings":";;;;;AAsBA,IAAa,WAAb,cAA8B,MAAM;CAClC,OAAgB;AAClB;;;;;AAMA,SAAgB,WAAW,SAAyB,MAAM,UAAwB;CAChF,IAAI,QAAQ,WAAW,GAAG,MAAM,IAAI,SAAS,8BAA8B;CAC3E,MAAM,OAAO,sBAAsB,KAAK,GAAG;CAC3C,IAAI,MAAM;EACR,MAAM,QAAQ,QAAQ,QAAQ,SAAS,IAAI,OAAO,KAAK,MAAM,CAAC;EAC9D,IAAI,CAAC,OAAO,MAAM,IAAI,SAAS,kBAAkB,QAAQ,OAAO,MAAM;EACtE,OAAO;CACT;CACA,MAAM,QAAQ,QAAQ,MAAM,MAAM,EAAE,UAAU,GAAG;CACjD,IAAI,OAAO,OAAO;CAClB,MAAM,OAAO,QAAQ,QAAQ,MAAM,EAAE,MAAM,WAAW,GAAG,CAAC;CAC1D,IAAI,KAAK,WAAW,KAAK,KAAK,IAAI,OAAO,KAAK;CAC9C,IAAI,KAAK,SAAS,GAChB,MAAM,IAAI,SAAS,IAAI,IAAI,YAAY,KAAK,OAAO,0BAA0B;CAC/E,IAAI,oBAAoB,KAAK,GAAG,GAAG;EACjC,MAAM,QAAQ,QAAQ,QAAQ,MAAM,EAAE,KAAK,IAAI,WAAW,IAAI,YAAY,CAAC,CAAC;EAC5E,MAAM,SAAS,MAAM,MAAM,SAAS;EACpC,IAAI,QAAQ,OAAO;CACrB;CACA,MAAM,IAAI,SAAS,mBAAmB,IAAI,EAAE;AAC9C;;;;;;;;AASA,SAAgB,aACd,SACA,MACA,UACqB;CACrB,MAAM,UAAU,QAAQ,QAAQ,MAAM,EAAE,QAAQ,KAAK,SAAS,EAAE,WAAW,UAAU;CACrF,IAAI,YAAY,SAAS,SAAS,GAAG;EACnC,MAAM,OAAO,IAAI,IAAI,SAAS,KAAK,KAAK,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;EACvD,IAAI,OAA4B;EAChC,IAAI,WAAW,OAAO;EACtB,KAAK,MAAM,KAAK,SAAS;GACvB,MAAM,IAAI,EAAE,MAAM,KAAK,IAAI,EAAE,IAAI,GAAG,IAAI,KAAA;GACxC,IAAI,MAAM,KAAA,MAAc,IAAI,YAAa,MAAM,YAAY,QAAQ,EAAE,QAAQ,KAAK,QAAS;IACzF,OAAO;IACP,WAAW;GACb;EACF;EACA,IAAI,MAAM,OAAO;CACnB;CACA,OAAO,QAAQ,QAAQ,SAAS,MAAM;AACxC;AAEA,MAAM,YAAY,MAAqB,EAAE,MAAM,EAAE,IAAI,IAAI,MAAM,GAAG,CAAC,IAAI;AAkBvE,MAAa,eAAe,OAAqB,UAC/C,CAAC,CAAC,SAAS,CAAC,CAAC,MAAM,OAAO,CAAC,MAAM,IAAI,MAAM,IAAI,GAAG;AAEpD,SAAgB,QACd,SACA,UAAqE,CAAC,GACtE;CACA,MAAM,OAAO,QAAQ,MAAM,EAAE,QAAQ,SAAS,GAAG;CACjD,MAAM,MAAgB,CAAC;CACvB,KAAK,MAAM,OAAO,CAAC,GAAG,IAAI,CAAC,CAAC,QAAQ,GAClC,KAAK,MAAM,KAAK,IAAI,OAAO,IAAI,CAAC,IAAI,SAAS,EAAE,EAAE,GAAG,IAAI,KAAK,EAAE,EAAE;CAEnE,MAAM,SAAS,QAAQ,QAAQ,YAAY;CAiB3C,OAAO;EAAE;EAAM,MAhBY,IACxB,QAAQ,OAAO,CAAC,UAAU,GAAG,YAAY,CAAC,CAAC,SAAS,MAAM,CAAC,CAAC,CAC5D,KAAK,OAAO;GACX,MAAM,MAAkB;IAAE;IAAI,UAAU,CAAC;IAAG,OAAO,CAAC;IAAG,MAAM;IAAG,SAAS;GAAE;GAC3E,MAAM,uBAAO,IAAI,IAAoB;GACrC,KAAK,SAAS,KAAK,MAAM;IACvB,MAAM,IAAI,IAAI,MAAM,MAAM,MAAM,EAAE,OAAO,EAAE;IAC3C,IAAI,SAAS,KAAK,GAAG,WAAW,IAAI;IACpC,IAAI,CAAC,GAAG;IACR,IAAI,aAAa,MAAM,IAAI,EAAE,KAAK,GAAG,IAAI,MAAM,KAAK,CAAC;IACrD,IAAI,EAAE,YAAY,UAAU,EAAE,YAAY,QAAQ,IAAI;IACtD,IAAI,EAAE,YAAY,QAAQ,IAAI;GAChC,CAAC;GACD,OAAO;EACT,CAEgB;EAAG,UADJ,KAAK,QAAQ,MAAM,YAAY,GAAG,QAAQ,SAAS,IAAI,CAAC,CAAC,CAAC,KAAK,MAAM,EAAE,KAC5D;CAAE;AAChC;AAEA,SAAgB,cAAc,EAAE,MAAM,MAAM,YAAgD;CAC1F,IAAI,KAAK,WAAW,GAClB,OAAO;CACT,MAAM,QAAQ,KAAK;CACnB,MAAM,OAAO,KAAK,KAAK,SAAS;CAChC,MAAM,QAAQ,CACZ,GAAG,KAAK,OAAO,MAAM,KAAK,WAAW,IAAI,KAAK,IAAI,qBAAqB,MAAM,MAAM,IAAI,SAAS,KAAK,EAAE,MAAM,KAAK,MAAM,IAAI,SAAS,IAAI,EAAE,IAC3I,EACF;CACA,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,KAAK,MAAM,EAAE,SAAS,SAAS,EAAE,MAAM,MAAM,GAAG,CAAC;CAChF,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,IAAI,SACf,KAAK,GAAG,MAAM,GAAG,IAAI,MAAM,SAAS,CAAC,IAAI,MAAM,KAAK,IAAI,aAAa,KAAK,KAAK,CAAC,CAChF,KAAK,EAAE;EACV,MAAM,OAAO,IAAI,UAAU,GAAG,IAAI,KAAK,GAAG,IAAI,YAAY;EAC1D,MAAM,KAAK,KAAK,MAAM,OAAO,KAAK,EAAE,IAAI,KAAK,SAAS,CAAC,EAAE,IAAI,IAAI,IAAI;CACvE;CACA,MAAM,KAAK,IAAI,8EAA8E;CAC7F,IAAI,SAAS,SAAS,GACpB,MAAM,KACJ,IACA,GAAG,SAAS,OAAO,MAAM,SAAS,WAAW,IAAI,SAAS,SAAS,4EACnE,GAAG,SAAS,KAAK,OAAO,KAAK,IAAI,CACnC;CAEF,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,MAAM,eAA8C;CAClD,WAAW;CACX,OAAO;CACP,MAAM;CACN,OAAO;CACP,OAAO;CACP,SAAS;CACT,OAAO;AACT;;;;;AAMA,MAAM,OAAO,MACX;CACE,EAAE,SAAS,SAAS,KAAK,EAAE,QAAQ,KAAK,IAAI,MAAM;CAClD,EAAE,OAAO,MAAM,SAAS,mBAAmB,EAAE,MAAM,MAAM,KAAK,IAAI,MAAM;CACxE,EAAE,OAAO,QAAQ,SAAS,qBAAqB,EAAE,MAAM,QAAQ,KAAK,IAAI,MAAM;AAChF,CAAC,CAAC,KAAK,EAAE;AAEX,MAAM,QAAyB;CAAC;CAAa;CAAS;CAAS;CAAS;CAAS;CAAW;AAAM;AAElG,MAAM,eAAe,GAAiB,UACpC,GAAG,EAAE,MAAM,IAAI,SAAS,CAAC,IAAI,EAAE,WAAW,KAAK,mBAAmB,YAAY,GAAG,KAAK,IAAI,mCAAmC,GAAG;AAElI,SAAgB,cACd,MACA,MACA,MACA,IACA,QAAsB,MACtB;CACA,MAAM,SAAS,CAAC,GAAG,IAAI,CAAC,CAAC,MAAM,GAAG,MAAM,MAAM,QAAQ,EAAE,MAAM,IAAI,MAAM,QAAQ,EAAE,MAAM,CAAC;CACzF,MAAM,QAAQ,MAAuB,IAAI,aAAa,KAAK;CAC3D,MAAM,SAAS,MAAqB,KAAK,QAAQ,MAAM,EAAE,WAAW,CAAC,CAAC,CAAC;CACvE,MAAM,QAAQ,GAAG,MAAM,WAAW,EAAE,eAAe,MAAM,OAAO,EAAE,WAAW,MAAM,OAAO,EAAE,oBAAoB,MAAM,MAAM,EAAE;CAC9H,MAAM,UAAU,aAAa,MAAM,IAAI;CACvC,MAAM,aAAa,QAAQ,KAAK,MAAM,GAAG,EAAE,MAAM,GAAG,EAAE,KAAK,KAAK,EAAE,MAAM,CAAC,CAAC,KAAK,IAAI;CACnF,MAAM,OAAO,MAAc,EAAE,QAAQ,OAAO,KAAK;CACjD,MAAM,QAAQ,OAAO,SAAS,OAAO,EAAE,WAAW,CAAC,EAAA,CAAG,KAAK,OAAO;EAAE,KAAK;EAAG,MAAM;CAAE,EAAE,CAAC;CACvF,IAAI,IACF,OAAO;EACL,eAAe,KAAK,MAAM,MAAM,KAAK;EACrC;EACA,QAAQ,YAAY,MAAM,KAAK,EAAE,UAAU,YAAY,MAAM,KAAK,EAAE,IAAI,MAAM;EAC9E;EACA,GAAI,QAAQ,SAAS,CAAC,sBAAsB,WAAW,IAAI,EAAE,IAAI,CAAC;EAClE;EACA;EACA,GAAG,OAAO,KACP,MACC,KAAK,IAAI,EAAE,EAAE,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,KAAK,aAAa,EAAE,UAAU,IAAI,CAAC,EAAE,GAC5F;EACA;EACA,GAAI,MAAM,SACN;GACE;GACA;GACA,GAAG,MAAM,KACN,EAAE,KAAK,WACN,KAAK,IAAI,IAAI,EAAE,EAAE,KAAK,IAAI,KAAK,GAAG,EAAE,KAAK,WAAW,IAAI,EAAE,KAAK,aAAa,IAAI,EAAE,GACtF;GACA;EACF,IACA,CAAC;CACP,CAAC,CAAC,KAAK,IAAI;CAEb,OAAO;EACL,QAAQ,YAAY,MAAM,KAAK,EAAE,UAAU,YAAY,MAAM,KAAK;EAClE;EACA,GAAI,QAAQ,SAAS,CAAC,kBAAkB,YAAY,IAAI,CAAC;EACzD;EACA,GAAG,OAAO,SAAS,MAAM,CACvB,KAAK,KAAK,EAAE,IAAI,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,IAAI,aAAa,EAAE,OAAO,CAAC,OAAO,EAAE,EAAE,IAAI,EAAE,KAAK,IAAI,CAAC,KAC1F,IAAI,EAAE,WAAW,CAAC,EAAA,CAAG,KAClB,MACC,SAAS,EAAE,IAAI,IAAI,CAAC,EAAE,KAAK,QAAQ,EAAE,KAAK,OAAO,QAAQ,WAAW,CAAC,MAAM,IAAI,aAAa,CAAC,CAAC,CAAC,CAAC,OAAO,OAAO,CAAC,CAAC,KAAK,KAAK,GAC9H,CACF,CAAC;CACH,CAAC,CAAC,KAAK,IAAI;AACb;;AAGA,SAAgB,iBAAiB,MAAoB,MAAoB,IAAqB;CAC5F,MAAM,UAAU,KAAK,QAAQ,MAAM,EAAE,SAAS,MAAM,CAAC,CAAC;CACtD,MAAM,OAAO,sCAAsC,KAAK,MAAM,0BAA0B,UAAU,SAAS,QAAQ,OAAO,YAAY,IAAI,MAAM,OAAO,GAAG;CAC1J,MAAM,QAAQ,MAAuB,IAAI,aAAa,KAAK;CAC3D,IAAI,IACF,OAAO;EACL,eAAe,KAAK;EACpB;EACA;EACA;EACA;EACA;EACA,GAAG,KAAK,KAAK,MAAM,KAAK,EAAE,GAAG,QAAQ,OAAO,KAAK,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,GAAG;EACxE;CACF,CAAC,CAAC,KAAK,IAAI;CAEb,OAAO;EAAC;EAAM;EAAI,GAAG,KAAK,KAAK,MAAM,KAAK,KAAK,EAAE,IAAI,EAAE,SAAS,EAAE,IAAI;CAAC,CAAC,CAAC,KAAK,IAAI;AACpF;AAEA,MAAM,aAAa,MACjB,IAAI,GAAG,OAAO,EAAE,MAAM,EAAE,IAAI,OAAO,EAAE,EAAE,EAAE,GAAG,OAAO,EAAE,EAAE,EAAE,GAAG,EAAE,cAAc,WAAW,OAAO;AAChG,MAAM,UAAU,MAAc,GAAG,KAAK,IAAI,MAAM,MAAM,KAAK,MAAM,KAAK,IAAI,CAAC,IAAI,GAAG;AAClF,MAAM,cAAc,MAAmB,GAAG,UAAU,EAAE,KAAK,IAAI,EAAE,KAAK,UAAU,EAAE,KAAK,IAAI;AAC3F,MAAM,gBAAgB,MACpB,EAAE,KAAK,UAAU,EAAE,KAAK,SAAS,kBAAkB,EAAE,KAAK,QAAQ,EAAE,KAAK,MAAM,IAAI;AAErF,SAAgB,WAAW,KAAwB;CACjD,MAAM,OAAO,UAAU,GAAG;CAC1B,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC;CACrD,MAAM,QAAQ;EACZ,OAAO,IAAI,MAAM,IAAI,cAAc,IAAI,OAAO,IAAI,IAAI,WAAW,gBAAgB,oBAAoB;EACrG,aAAa,IAAI,UAAU,KAAK,eAAe,IAAI,UAAU,EAAE,gBAAgB,IAAI;EACnF;EACA,GAAG,KAAK,KAAK,CAAC,GAAG,OAAO,KAAK,EAAE,OAAO,KAAK,EAAE,IAAI,GAAG;EACpD,GAAG,SAAS,GAAG,CAAC,CAAC,KAAK,MAAM,OAAO,GAAG;EACtC;CACF;CACA,KAAK,MAAM,KAAK,IAAI,OAAO;EACzB,MAAM,SAAS,EAAE,SAAS,WAAW;EACrC,MAAM,QAAQ,EAAE,SAAS;EACzB,MAAM,KACJ,KAAK,aAAa,EAAE,SAAS,GAAG,EAAE,KAAK,UAAU,QAAQ,KAAK,eAAe,MAAM,UAAU,MAAM,KAAK,WAAW,CAAC,KACtH;EACA,KAAK,MAAM,KAAK,EAAE,UAAU;GAC1B,MAAM,SAAS,SAAS,WAAW;GACnC,IAAI,CAAC,QACH,MAAM,KAAK,SAAS,aAAa,EAAE,SAAS,GAAG,EAAE,GAAG,IAAI,eAAe,EAAE,UAAU,GAAG;GAExF,KAAK,MAAM,KAAK,EAAE,YAChB,MAAM,KACJ,GAAG,OAAO,QAAQ,EAAE,QAAQ,GAAG,EAAE,MAAM,QAAQ,CAAC,EAAE,MAAM,EAAE,UAAU,KAAK,EAAE,UAC7E;GAEF,IAAI,EAAE,SAAS,CAAC,eAAe,CAAC,GAC9B,MAAM,KAAK,GAAG,SAAS,EAAE,MAAM,KAAK,IAAI,EAAE,MAAM,QAAQ,MAAM,IAAI,CAAC,CAAC,IAAI;EAE5E;CACF;CACA,OAAO,MAAM,KAAK,IAAI;AACxB"}
|