dsh-continual-evolve 0.11.0 → 0.12.1
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/README.md +42 -115
- package/README.zh.md +33 -93
- package/client/client.js +253 -0
- package/cordis.patch.yml +2 -2
- package/lib/card-routes.d.ts +32 -0
- package/lib/card-routes.js +92 -0
- package/lib/index.d.ts +65 -217
- package/lib/index.js +74 -234
- package/lib/known-workspaces.d.ts +22 -0
- package/lib/known-workspaces.js +42 -0
- package/lib/memory-guide.d.ts +13 -14
- package/lib/memory-guide.js +34 -21
- package/lib/memory-section.d.ts +77 -0
- package/lib/memory-section.js +189 -0
- package/lib/memory-snapshot.d.ts +54 -0
- package/lib/memory-snapshot.js +125 -0
- package/package.json +8 -17
- package/lib/apply.d.ts +0 -22
- package/lib/apply.js +0 -157
- package/lib/approval.d.ts +0 -73
- package/lib/approval.js +0 -127
- package/lib/audit-log.d.ts +0 -19
- package/lib/audit-log.js +0 -52
- package/lib/auto.d.ts +0 -236
- package/lib/auto.js +0 -773
- package/lib/autocase.d.ts +0 -38
- package/lib/autocase.js +0 -76
- package/lib/benchmark-command.d.ts +0 -9
- package/lib/benchmark-command.js +0 -357
- package/lib/benchmark.d.ts +0 -151
- package/lib/benchmark.js +0 -250
- package/lib/command-util.d.ts +0 -70
- package/lib/command-util.js +0 -123
- package/lib/command.d.ts +0 -13
- package/lib/command.js +0 -134
- package/lib/consolidate.d.ts +0 -90
- package/lib/consolidate.js +0 -180
- package/lib/copy.d.ts +0 -37
- package/lib/copy.js +0 -58
- package/lib/declines.d.ts +0 -103
- package/lib/declines.js +0 -179
- package/lib/evaluate.d.ts +0 -74
- package/lib/evaluate.js +0 -271
- package/lib/evolve-event.d.ts +0 -40
- package/lib/evolve-event.js +0 -60
- package/lib/extraction-phase.d.ts +0 -132
- package/lib/extraction-phase.js +0 -236
- package/lib/failures.d.ts +0 -42
- package/lib/failures.js +0 -171
- package/lib/fate.d.ts +0 -130
- package/lib/fate.js +0 -350
- package/lib/goal-command.d.ts +0 -7
- package/lib/goal-command.js +0 -37
- package/lib/goal.d.ts +0 -72
- package/lib/goal.js +0 -72
- package/lib/inject.d.ts +0 -90
- package/lib/inject.js +0 -164
- package/lib/injection-caps.d.ts +0 -4
- package/lib/injection-caps.js +0 -4
- package/lib/injection-directory.d.ts +0 -29
- package/lib/injection-directory.js +0 -78
- package/lib/injection-rank.d.ts +0 -85
- package/lib/injection-rank.js +0 -161
- package/lib/listener.d.ts +0 -45
- package/lib/listener.js +0 -199
- package/lib/llm-text.d.ts +0 -104
- package/lib/llm-text.js +0 -140
- package/lib/logfile.d.ts +0 -71
- package/lib/logfile.js +0 -159
- package/lib/maintenance-commands.d.ts +0 -23
- package/lib/maintenance-commands.js +0 -136
- package/lib/memory-agent.d.ts +0 -93
- package/lib/memory-agent.js +0 -300
- package/lib/memory-apply.d.ts +0 -44
- package/lib/memory-apply.js +0 -202
- package/lib/memory-benchmark.d.ts +0 -76
- package/lib/memory-benchmark.js +0 -170
- package/lib/memory-index.d.ts +0 -114
- package/lib/memory-index.js +0 -197
- package/lib/memory-manifest.d.ts +0 -16
- package/lib/memory-manifest.js +0 -70
- package/lib/message-source.d.ts +0 -23
- package/lib/message-source.js +0 -12
- package/lib/mount-command.d.ts +0 -10
- package/lib/mount-command.js +0 -48
- package/lib/mount.d.ts +0 -42
- package/lib/mount.js +0 -212
- package/lib/notify.d.ts +0 -48
- package/lib/notify.js +0 -57
- package/lib/plan-command.d.ts +0 -12
- package/lib/plan-command.js +0 -38
- package/lib/plan.d.ts +0 -16
- package/lib/plan.js +0 -126
- package/lib/planner.d.ts +0 -61
- package/lib/planner.js +0 -144
- package/lib/pool.d.ts +0 -7
- package/lib/pool.js +0 -25
- package/lib/prefix-cache.d.ts +0 -79
- package/lib/prefix-cache.js +0 -168
- package/lib/project.d.ts +0 -28
- package/lib/project.js +0 -66
- package/lib/projection.d.ts +0 -22
- package/lib/projection.js +0 -112
- package/lib/promotion.d.ts +0 -94
- package/lib/promotion.js +0 -171
- package/lib/recall.d.ts +0 -85
- package/lib/recall.js +0 -142
- package/lib/record-language.d.ts +0 -85
- package/lib/record-language.js +0 -152
- package/lib/render.d.ts +0 -13
- package/lib/render.js +0 -80
- package/lib/review-scheduler.d.ts +0 -38
- package/lib/review-scheduler.js +0 -183
- package/lib/review.d.ts +0 -60
- package/lib/review.js +0 -145
- package/lib/rollback.d.ts +0 -9
- package/lib/rollback.js +0 -68
- package/lib/rubric.d.ts +0 -29
- package/lib/rubric.js +0 -119
- package/lib/runtime-commands.d.ts +0 -14
- package/lib/runtime-commands.js +0 -115
- package/lib/runtime.d.ts +0 -24
- package/lib/runtime.js +0 -87
- package/lib/score.d.ts +0 -64
- package/lib/score.js +0 -191
- package/lib/search.d.ts +0 -83
- package/lib/search.js +0 -136
- package/lib/service.d.ts +0 -63
- package/lib/service.js +0 -253
- package/lib/skill-render.d.ts +0 -23
- package/lib/skill-render.js +0 -69
- package/lib/skill.d.ts +0 -15
- package/lib/skill.js +0 -80
- package/lib/skillquality.d.ts +0 -85
- package/lib/skillquality.js +0 -320
- package/lib/source.d.ts +0 -30
- package/lib/source.js +0 -43
- package/lib/state.d.ts +0 -26
- package/lib/state.js +0 -151
- package/lib/store.d.ts +0 -61
- package/lib/store.js +0 -164
- package/lib/token-usage.d.ts +0 -68
- package/lib/token-usage.js +0 -211
- package/lib/tool.d.ts +0 -34
- package/lib/tool.js +0 -320
- package/lib/turn-snapshot.d.ts +0 -78
- package/lib/turn-snapshot.js +0 -252
- package/lib/types.d.ts +0 -218
- package/lib/types.js +0 -133
- package/lib/usage.d.ts +0 -36
- package/lib/usage.js +0 -98
- package/lib/validate.d.ts +0 -28
- package/lib/validate.js +0 -181
- package/lib/wrapup-candidates.d.ts +0 -127
- package/lib/wrapup-candidates.js +0 -161
- package/lib/wrapup-command.d.ts +0 -10
- package/lib/wrapup-command.js +0 -243
- package/lib/wrapup-proposals.d.ts +0 -101
- package/lib/wrapup-proposals.js +0 -281
- package/lib/wrapup.d.ts +0 -56
- package/lib/wrapup.js +0 -105
package/lib/store.d.ts
DELETED
|
@@ -1,61 +0,0 @@
|
|
|
1
|
-
import type { HarnessScope, RefinementResult } from "./types.js";
|
|
2
|
-
export declare const EVOLVE_DIR = "evolve";
|
|
3
|
-
/**
|
|
4
|
-
* Storage-hygiene defaults (#20): history used to grow append-only without
|
|
5
|
-
* bound (dozens of full-state snapshot copies, unbounded JSONL). Retention
|
|
6
|
-
* is count-based and generous — rollback only ever needs recent snapshots
|
|
7
|
-
* and readers (`failures.ts` over reviews) need a working window, not the
|
|
8
|
-
* full past. Tunable via the `historyRetain` plugin config.
|
|
9
|
-
*/
|
|
10
|
-
export declare const DEFAULT_SNAPSHOT_RETAIN = 20;
|
|
11
|
-
export { DEFAULT_REFINEMENTS_RETAIN, DEFAULT_REVIEWS_RETAIN } from "./types.js";
|
|
12
|
-
export declare const DEFAULT_TOKEN_USAGE_RETAIN = 500;
|
|
13
|
-
/** Resolved retention set: how many snapshots / JSONL tail lines to keep. */
|
|
14
|
-
export interface HistoryRetention {
|
|
15
|
-
/** Full-state snapshots kept per store (oldest pruned first). */
|
|
16
|
-
snapshots: number;
|
|
17
|
-
/** Tail lines kept in each store's refinements.jsonl. */
|
|
18
|
-
refinements: number;
|
|
19
|
-
/** Tail lines kept in the shared reviews.jsonl audit trail. */
|
|
20
|
-
reviews: number;
|
|
21
|
-
/** Tail lines kept in the direct-call token-usage.jsonl ledger. */
|
|
22
|
-
tokenUsage: number;
|
|
23
|
-
}
|
|
24
|
-
/** Defaults applied for absent / non-positive retention fields (fail loud never: clamp, don't throw). */
|
|
25
|
-
export declare function resolveHistoryRetention(raw?: Partial<HistoryRetention>): HistoryRetention;
|
|
26
|
-
export interface StorePaths {
|
|
27
|
-
/** Directory holding harness_state.json. */
|
|
28
|
-
stateDir: string;
|
|
29
|
-
/** Directory holding snapshots for this store. */
|
|
30
|
-
snapshotsDir: string;
|
|
31
|
-
/** JSONL path for applied refinement results. */
|
|
32
|
-
resultsPath: string;
|
|
33
|
-
}
|
|
34
|
-
/**
|
|
35
|
-
* Resolve store paths for a scope. For `project` the `sessionId` parameter
|
|
36
|
-
* carries the project key (see {@link resolveProjectKey} in project.ts) —
|
|
37
|
-
* it is sanitized to a single path segment so an externally supplied key
|
|
38
|
-
* can never traverse out of `evolve/projects/`.
|
|
39
|
-
*/
|
|
40
|
-
export declare function storePaths(baseDir: string, scope: HarnessScope, sessionId?: string): StorePaths;
|
|
41
|
-
/** Snapshot the current state file before a mutation, if one exists. */
|
|
42
|
-
export declare function snapshotBefore(paths: StorePaths, refinementId: string): void;
|
|
43
|
-
/** Append an applied result to the store's JSONL history. */
|
|
44
|
-
export declare function appendResult(paths: StorePaths, result: RefinementResult): void;
|
|
45
|
-
/**
|
|
46
|
-
* Prune a store's snapshots to the newest `retain` files (by mtime, oldest
|
|
47
|
-
* first). Best-effort: a missing directory or an unlink failure never breaks
|
|
48
|
-
* the apply path — the next apply retries.
|
|
49
|
-
*/
|
|
50
|
-
export declare function pruneSnapshots(snapshotsDir: string, retain: number): void;
|
|
51
|
-
/**
|
|
52
|
-
* Truncate a JSONL file to its last `retain` non-empty lines. No-op when the
|
|
53
|
-
* file is missing or already within budget — the common case costs one
|
|
54
|
-
* read and no write.
|
|
55
|
-
*/
|
|
56
|
-
export declare function pruneJsonlFile(path: string, retain: number): void;
|
|
57
|
-
/** Reviews audit-trail path shared by the gate, fate, and tool events. */
|
|
58
|
-
export declare function reviewsPath(baseDir: string): string;
|
|
59
|
-
/** Read the applied results history; malformed lines are skipped, never fatal. */
|
|
60
|
-
export declare function loadResults(paths: StorePaths): RefinementResult[];
|
|
61
|
-
//# sourceMappingURL=store.d.ts.map
|
package/lib/store.js
DELETED
|
@@ -1,164 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Store layout and snapshot discipline for the evolution plugin.
|
|
3
|
-
*
|
|
4
|
-
* Layout (self-contained under the DSH home; no dependency on session
|
|
5
|
-
* persistence internals):
|
|
6
|
-
*
|
|
7
|
-
* <dshHome>/evolve/global/harness_state.json cross-project store
|
|
8
|
-
* <dshHome>/evolve/global/refinements.jsonl applied results (rollback source)
|
|
9
|
-
* <dshHome>/evolve/projects/<slug-hash>/... per-project store
|
|
10
|
-
* <dshHome>/evolve/local/<sessionId>/... per-session staging store
|
|
11
|
-
*
|
|
12
|
-
* Snapshot discipline is code-enforced: before any mutating apply, the
|
|
13
|
-
* pre-apply state is copied to `snapshots/<refinementId>.json`. The model has
|
|
14
|
-
* no way to skip it — it runs inside the service, not in a prompt.
|
|
15
|
-
*/
|
|
16
|
-
import { existsSync, mkdirSync, readFileSync, readdirSync, statSync, unlinkSync, writeFileSync } from "node:fs";
|
|
17
|
-
import { DEFAULT_REFINEMENTS_RETAIN, DEFAULT_REVIEWS_RETAIN } from "./types.js";
|
|
18
|
-
import { join } from "node:path";
|
|
19
|
-
import { sanitizeProjectKey } from "./project.js";
|
|
20
|
-
import { stateFilePath } from "./state.js";
|
|
21
|
-
export const EVOLVE_DIR = "evolve";
|
|
22
|
-
/**
|
|
23
|
-
* Storage-hygiene defaults (#20): history used to grow append-only without
|
|
24
|
-
* bound (dozens of full-state snapshot copies, unbounded JSONL). Retention
|
|
25
|
-
* is count-based and generous — rollback only ever needs recent snapshots
|
|
26
|
-
* and readers (`failures.ts` over reviews) need a working window, not the
|
|
27
|
-
* full past. Tunable via the `historyRetain` plugin config.
|
|
28
|
-
*/
|
|
29
|
-
export const DEFAULT_SNAPSHOT_RETAIN = 20;
|
|
30
|
-
export { DEFAULT_REFINEMENTS_RETAIN, DEFAULT_REVIEWS_RETAIN } from "./types.js";
|
|
31
|
-
export const DEFAULT_TOKEN_USAGE_RETAIN = 500;
|
|
32
|
-
/** Defaults applied for absent / non-positive retention fields (fail loud never: clamp, don't throw). */
|
|
33
|
-
export function resolveHistoryRetention(raw) {
|
|
34
|
-
const pick = (value, fallback) => typeof value === "number" && Number.isFinite(value) && value > 0 ? Math.floor(value) : fallback;
|
|
35
|
-
return {
|
|
36
|
-
snapshots: pick(raw?.snapshots, DEFAULT_SNAPSHOT_RETAIN),
|
|
37
|
-
refinements: pick(raw?.refinements, DEFAULT_REFINEMENTS_RETAIN),
|
|
38
|
-
reviews: pick(raw?.reviews, DEFAULT_REVIEWS_RETAIN),
|
|
39
|
-
tokenUsage: pick(raw?.tokenUsage, DEFAULT_TOKEN_USAGE_RETAIN),
|
|
40
|
-
};
|
|
41
|
-
}
|
|
42
|
-
/**
|
|
43
|
-
* Resolve store paths for a scope. For `project` the `sessionId` parameter
|
|
44
|
-
* carries the project key (see {@link resolveProjectKey} in project.ts) —
|
|
45
|
-
* it is sanitized to a single path segment so an externally supplied key
|
|
46
|
-
* can never traverse out of `evolve/projects/`.
|
|
47
|
-
*/
|
|
48
|
-
export function storePaths(baseDir, scope, sessionId) {
|
|
49
|
-
const scopeDir = scope === "global" ? "global" : scope === "project" ? join("projects", sanitizeProjectKey(sessionId ?? "project")) : join("local", sessionId ?? "anonymous");
|
|
50
|
-
const stateDir = join(baseDir, EVOLVE_DIR, scopeDir);
|
|
51
|
-
return {
|
|
52
|
-
stateDir,
|
|
53
|
-
snapshotsDir: join(stateDir, "snapshots"),
|
|
54
|
-
resultsPath: join(stateDir, "refinements.jsonl"),
|
|
55
|
-
};
|
|
56
|
-
}
|
|
57
|
-
/** Snapshot the current state file before a mutation, if one exists. */
|
|
58
|
-
export function snapshotBefore(paths, refinementId) {
|
|
59
|
-
const statePath = stateFilePath(paths.stateDir);
|
|
60
|
-
if (!existsSync(statePath)) {
|
|
61
|
-
return;
|
|
62
|
-
}
|
|
63
|
-
mkdirSync(paths.snapshotsDir, { recursive: true });
|
|
64
|
-
// 0600: the snapshot is a full copy of harness_state.json — same
|
|
65
|
-
// permission discipline as the store itself (review audit 2026-08-28 S5).
|
|
66
|
-
writeFileSync(join(paths.snapshotsDir, `${refinementId}.json`), readFileSync(statePath, "utf8"), { encoding: "utf8", mode: 0o600 });
|
|
67
|
-
}
|
|
68
|
-
/** Append an applied result to the store's JSONL history. */
|
|
69
|
-
export function appendResult(paths, result) {
|
|
70
|
-
mkdirSync(paths.stateDir, { recursive: true });
|
|
71
|
-
writeFileSync(paths.resultsPath, `${JSON.stringify(result)}\n`, { encoding: "utf8", flag: "a" });
|
|
72
|
-
}
|
|
73
|
-
/**
|
|
74
|
-
* Prune a store's snapshots to the newest `retain` files (by mtime, oldest
|
|
75
|
-
* first). Best-effort: a missing directory or an unlink failure never breaks
|
|
76
|
-
* the apply path — the next apply retries.
|
|
77
|
-
*/
|
|
78
|
-
export function pruneSnapshots(snapshotsDir, retain) {
|
|
79
|
-
const keep = Math.floor(retain);
|
|
80
|
-
if (!(keep > 0) || !existsSync(snapshotsDir))
|
|
81
|
-
return;
|
|
82
|
-
let files;
|
|
83
|
-
try {
|
|
84
|
-
files = readdirSync(snapshotsDir).filter((f) => f.endsWith(".json"));
|
|
85
|
-
}
|
|
86
|
-
catch {
|
|
87
|
-
return; // unreadable directory — leave it for the next attempt
|
|
88
|
-
}
|
|
89
|
-
if (files.length <= keep)
|
|
90
|
-
return;
|
|
91
|
-
const withTime = [];
|
|
92
|
-
for (const file of files) {
|
|
93
|
-
try {
|
|
94
|
-
withTime.push({ file, mtime: statSync(join(snapshotsDir, file)).mtimeMs });
|
|
95
|
-
}
|
|
96
|
-
catch {
|
|
97
|
-
// unstatable file — keep it rather than delete blindly
|
|
98
|
-
}
|
|
99
|
-
}
|
|
100
|
-
withTime.sort((a, b) => a.mtime - b.mtime);
|
|
101
|
-
for (const victim of withTime.slice(0, Math.max(0, withTime.length - keep))) {
|
|
102
|
-
try {
|
|
103
|
-
unlinkSync(join(snapshotsDir, victim.file));
|
|
104
|
-
}
|
|
105
|
-
catch {
|
|
106
|
-
// one bad unlink must not stop the sweep or the apply path
|
|
107
|
-
}
|
|
108
|
-
}
|
|
109
|
-
}
|
|
110
|
-
/**
|
|
111
|
-
* Truncate a JSONL file to its last `retain` non-empty lines. No-op when the
|
|
112
|
-
* file is missing or already within budget — the common case costs one
|
|
113
|
-
* read and no write.
|
|
114
|
-
*/
|
|
115
|
-
export function pruneJsonlFile(path, retain) {
|
|
116
|
-
const keep = Math.floor(retain);
|
|
117
|
-
if (!(keep > 0) || !existsSync(path))
|
|
118
|
-
return;
|
|
119
|
-
let lines;
|
|
120
|
-
try {
|
|
121
|
-
lines = readFileSync(path, "utf8").split("\n").filter((l) => l.trim().length > 0);
|
|
122
|
-
}
|
|
123
|
-
catch {
|
|
124
|
-
return; // unreadable file — leave it for the next attempt
|
|
125
|
-
}
|
|
126
|
-
if (lines.length <= keep)
|
|
127
|
-
return;
|
|
128
|
-
try {
|
|
129
|
-
writeFileSync(path, `${lines.slice(-keep).join("\n")}\n`, "utf8");
|
|
130
|
-
}
|
|
131
|
-
catch {
|
|
132
|
-
// truncation failure must not break the write path that called it
|
|
133
|
-
}
|
|
134
|
-
}
|
|
135
|
-
/** Reviews audit-trail path shared by the gate, fate, and tool events. */
|
|
136
|
-
export function reviewsPath(baseDir) {
|
|
137
|
-
return join(baseDir, EVOLVE_DIR, "reviews.jsonl");
|
|
138
|
-
}
|
|
139
|
-
/** Read the applied results history; malformed lines are skipped, never fatal. */
|
|
140
|
-
export function loadResults(paths) {
|
|
141
|
-
if (!existsSync(paths.resultsPath)) {
|
|
142
|
-
return [];
|
|
143
|
-
}
|
|
144
|
-
const results = [];
|
|
145
|
-
for (const line of readFileSync(paths.resultsPath, "utf8").split("\n")) {
|
|
146
|
-
const trimmed = line.trim();
|
|
147
|
-
if (!trimmed)
|
|
148
|
-
continue;
|
|
149
|
-
try {
|
|
150
|
-
const parsed = JSON.parse(trimmed);
|
|
151
|
-
if (isResult(parsed)) {
|
|
152
|
-
results.push(parsed);
|
|
153
|
-
}
|
|
154
|
-
}
|
|
155
|
-
catch {
|
|
156
|
-
// skip malformed line
|
|
157
|
-
}
|
|
158
|
-
}
|
|
159
|
-
return results;
|
|
160
|
-
}
|
|
161
|
-
function isResult(data) {
|
|
162
|
-
return typeof data === "object" && data !== null && "id" in data && "appliedEdits" in data;
|
|
163
|
-
}
|
|
164
|
-
//# sourceMappingURL=store.js.map
|
package/lib/token-usage.d.ts
DELETED
|
@@ -1,68 +0,0 @@
|
|
|
1
|
-
import type { TokenUsage } from "@deepseek-ai/dsh-llm";
|
|
2
|
-
import type { StreamTextObservation, StreamTextUsageObserver } from "./llm-text.js";
|
|
3
|
-
export declare const TOKEN_USAGE_LEDGER_VERSION = 1;
|
|
4
|
-
/** Plugin-owned direct call paths represented in the ledger. */
|
|
5
|
-
export declare const TOKEN_USAGE_PHASES: readonly ["review", "memory", "planner", "wrapup", "fate"];
|
|
6
|
-
export type TokenUsagePhase = (typeof TOKEN_USAGE_PHASES)[number];
|
|
7
|
-
/** Whether the provider supplied a usage sample; missing is never treated as zero. */
|
|
8
|
-
export type TokenUsageStatus = "reported" | "missing";
|
|
9
|
-
/** How the exact full-call total was obtained. */
|
|
10
|
-
export type TokenTotalSource = "provider" | "derived" | "unavailable";
|
|
11
|
-
/** One retained direct-call observation. */
|
|
12
|
-
export interface TokenUsageRecord {
|
|
13
|
-
version: typeof TOKEN_USAGE_LEDGER_VERSION;
|
|
14
|
-
timestamp: string;
|
|
15
|
-
sessionId: string;
|
|
16
|
-
phase: TokenUsagePhase;
|
|
17
|
-
provider: string;
|
|
18
|
-
model: string;
|
|
19
|
-
outcome: StreamTextObservation["outcome"];
|
|
20
|
-
usageStatus: TokenUsageStatus;
|
|
21
|
-
/** Absent exactly when {@link usageStatus} is `missing`. */
|
|
22
|
-
usage?: TokenUsage;
|
|
23
|
-
totalSource: TokenTotalSource;
|
|
24
|
-
}
|
|
25
|
-
/** Caller-owned destination and failure hook; phase/provider/model are supplied by the call site. */
|
|
26
|
-
export interface TokenUsageTarget {
|
|
27
|
-
baseDir: string;
|
|
28
|
-
sessionId: string;
|
|
29
|
-
retain: number;
|
|
30
|
-
onError?: (cause: unknown) => void;
|
|
31
|
-
}
|
|
32
|
-
/** Loaded retained window plus corrupt-line count for honest coverage reporting. */
|
|
33
|
-
export interface LoadedTokenUsage {
|
|
34
|
-
records: TokenUsageRecord[];
|
|
35
|
-
corruptLines: number;
|
|
36
|
-
}
|
|
37
|
-
/** Aggregate provider-reported buckets over a retained ledger window. */
|
|
38
|
-
export interface TokenUsageTotals {
|
|
39
|
-
calls: number;
|
|
40
|
-
reportedCalls: number;
|
|
41
|
-
missingUsageCalls: number;
|
|
42
|
-
uncachedInputTokens: number;
|
|
43
|
-
cacheReadTokens: number;
|
|
44
|
-
cacheWriteTokens: number;
|
|
45
|
-
outputTokens: number;
|
|
46
|
-
reasoningTokens: number;
|
|
47
|
-
totalTokens: number;
|
|
48
|
-
}
|
|
49
|
-
/** Absolute path of the direct-call token ledger. */
|
|
50
|
-
export declare function tokenUsagePath(baseDir: string): string;
|
|
51
|
-
/** Exact full-call total, with the old-peer fallback only when `totalTokens` is absent. */
|
|
52
|
-
export declare function exactTokenTotal(usage: TokenUsage): number;
|
|
53
|
-
/** Build the observer attached to one `streamText` call. Errors are diagnostic-only. */
|
|
54
|
-
export declare function createTokenUsageObserver(target: TokenUsageTarget, phase: TokenUsagePhase, provider: string, model: string): StreamTextUsageObserver;
|
|
55
|
-
/**
|
|
56
|
-
* Append one call and enforce the dedicated tail budget. Append + conditional
|
|
57
|
-
* prune stay in one synchronous block, so concurrent async callers in one DSH
|
|
58
|
-
* process cannot interleave the read/truncate/write. Multiple DSH processes
|
|
59
|
-
* sharing one baseDir are outside the plugin's storage contract.
|
|
60
|
-
*/
|
|
61
|
-
export declare function appendTokenUsage(target: TokenUsageTarget, record: TokenUsageRecord): void;
|
|
62
|
-
/** Load valid records from the retained ledger; malformed lines never break `/evolve usage`. */
|
|
63
|
-
export declare function loadTokenUsage(baseDir: string): LoadedTokenUsage;
|
|
64
|
-
/** Sum exact provider-reported buckets without double-counting reasoning or cache fields. */
|
|
65
|
-
export declare function aggregateTokenUsage(records: readonly TokenUsageRecord[]): TokenUsageTotals;
|
|
66
|
-
/** Render the human-facing direct-call ledger section. */
|
|
67
|
-
export declare function renderTokenUsageReport(ledger: LoadedTokenUsage, retain: number): string[];
|
|
68
|
-
//# sourceMappingURL=token-usage.d.ts.map
|
package/lib/token-usage.js
DELETED
|
@@ -1,211 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Provider-reported token accounting for plugin-owned direct LLM calls.
|
|
3
|
-
*
|
|
4
|
-
* The review gate, planner, dedicated memory agent, wrap-up command, and
|
|
5
|
-
* local-fate classifier all use the shared model-call boundary. These calls do
|
|
6
|
-
* not belong to an Agent turn (they pass no sessionId), so the host session
|
|
7
|
-
* token-meter cannot attribute them. This ledger records the exact
|
|
8
|
-
* `TokenUsage` emitted by the adapter at the one shared stream boundary.
|
|
9
|
-
*
|
|
10
|
-
* Storage: `<baseDir>/evolve/token-usage.jsonl`. The file contains route and
|
|
11
|
-
* count metadata only — never prompts, responses, or tool arguments — and is
|
|
12
|
-
* trimmed to its dedicated retention budget after each synchronous append.
|
|
13
|
-
*/
|
|
14
|
-
import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
|
|
15
|
-
import { join } from "node:path";
|
|
16
|
-
import { EVOLVE_DIR, pruneJsonlFile } from "./store.js";
|
|
17
|
-
const TOKEN_USAGE_FILE = "token-usage.jsonl";
|
|
18
|
-
export const TOKEN_USAGE_LEDGER_VERSION = 1;
|
|
19
|
-
/** Plugin-owned direct call paths represented in the ledger. */
|
|
20
|
-
export const TOKEN_USAGE_PHASES = ["review", "memory", "planner", "wrapup", "fate"];
|
|
21
|
-
/** Absolute path of the direct-call token ledger. */
|
|
22
|
-
export function tokenUsagePath(baseDir) {
|
|
23
|
-
return join(baseDir, EVOLVE_DIR, TOKEN_USAGE_FILE);
|
|
24
|
-
}
|
|
25
|
-
/** Exact full-call total, with the old-peer fallback only when `totalTokens` is absent. */
|
|
26
|
-
export function exactTokenTotal(usage) {
|
|
27
|
-
if (usage.totalTokens !== undefined)
|
|
28
|
-
return usage.totalTokens;
|
|
29
|
-
return usage.inputTokens + usage.outputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0);
|
|
30
|
-
}
|
|
31
|
-
/** Build the observer attached to one `streamText` call. Errors are diagnostic-only. */
|
|
32
|
-
export function createTokenUsageObserver(target, phase, provider, model) {
|
|
33
|
-
return (observation) => {
|
|
34
|
-
try {
|
|
35
|
-
appendTokenUsage(target, {
|
|
36
|
-
version: TOKEN_USAGE_LEDGER_VERSION,
|
|
37
|
-
timestamp: new Date().toISOString(),
|
|
38
|
-
sessionId: target.sessionId,
|
|
39
|
-
phase,
|
|
40
|
-
provider,
|
|
41
|
-
model,
|
|
42
|
-
outcome: observation.outcome,
|
|
43
|
-
usageStatus: observation.usage ? "reported" : "missing",
|
|
44
|
-
...(observation.usage ? { usage: observation.usage } : {}),
|
|
45
|
-
totalSource: observation.usage
|
|
46
|
-
? observation.usage.totalTokens !== undefined
|
|
47
|
-
? "provider"
|
|
48
|
-
: "derived"
|
|
49
|
-
: "unavailable",
|
|
50
|
-
});
|
|
51
|
-
}
|
|
52
|
-
catch (cause) {
|
|
53
|
-
target.onError?.(cause);
|
|
54
|
-
}
|
|
55
|
-
};
|
|
56
|
-
}
|
|
57
|
-
/**
|
|
58
|
-
* Append one call and enforce the dedicated tail budget. Append + conditional
|
|
59
|
-
* prune stay in one synchronous block, so concurrent async callers in one DSH
|
|
60
|
-
* process cannot interleave the read/truncate/write. Multiple DSH processes
|
|
61
|
-
* sharing one baseDir are outside the plugin's storage contract.
|
|
62
|
-
*/
|
|
63
|
-
export function appendTokenUsage(target, record) {
|
|
64
|
-
if (!isTokenUsageRecord(record)) {
|
|
65
|
-
throw new Error("evolve: refusing to append an invalid token-usage record");
|
|
66
|
-
}
|
|
67
|
-
const dir = join(target.baseDir, EVOLVE_DIR);
|
|
68
|
-
mkdirSync(dir, { recursive: true, mode: 0o700 });
|
|
69
|
-
writeFileSync(tokenUsagePath(target.baseDir), `${JSON.stringify(record)}\n`, {
|
|
70
|
-
encoding: "utf8",
|
|
71
|
-
flag: "a",
|
|
72
|
-
mode: 0o600,
|
|
73
|
-
});
|
|
74
|
-
pruneJsonlFile(tokenUsagePath(target.baseDir), target.retain);
|
|
75
|
-
}
|
|
76
|
-
/** Load valid records from the retained ledger; malformed lines never break `/evolve usage`. */
|
|
77
|
-
export function loadTokenUsage(baseDir) {
|
|
78
|
-
const path = tokenUsagePath(baseDir);
|
|
79
|
-
if (!existsSync(path))
|
|
80
|
-
return { records: [], corruptLines: 0 };
|
|
81
|
-
let lines;
|
|
82
|
-
try {
|
|
83
|
-
lines = readFileSync(path, "utf8").split("\n");
|
|
84
|
-
}
|
|
85
|
-
catch {
|
|
86
|
-
return { records: [], corruptLines: 0 };
|
|
87
|
-
}
|
|
88
|
-
const records = [];
|
|
89
|
-
let corruptLines = 0;
|
|
90
|
-
for (const line of lines) {
|
|
91
|
-
if (line.trim().length === 0)
|
|
92
|
-
continue;
|
|
93
|
-
try {
|
|
94
|
-
const parsed = JSON.parse(line);
|
|
95
|
-
if (isTokenUsageRecord(parsed))
|
|
96
|
-
records.push(parsed);
|
|
97
|
-
else
|
|
98
|
-
corruptLines += 1;
|
|
99
|
-
}
|
|
100
|
-
catch {
|
|
101
|
-
corruptLines += 1;
|
|
102
|
-
}
|
|
103
|
-
}
|
|
104
|
-
return { records, corruptLines };
|
|
105
|
-
}
|
|
106
|
-
/** Sum exact provider-reported buckets without double-counting reasoning or cache fields. */
|
|
107
|
-
export function aggregateTokenUsage(records) {
|
|
108
|
-
const totals = {
|
|
109
|
-
calls: records.length,
|
|
110
|
-
reportedCalls: 0,
|
|
111
|
-
missingUsageCalls: 0,
|
|
112
|
-
uncachedInputTokens: 0,
|
|
113
|
-
cacheReadTokens: 0,
|
|
114
|
-
cacheWriteTokens: 0,
|
|
115
|
-
outputTokens: 0,
|
|
116
|
-
reasoningTokens: 0,
|
|
117
|
-
totalTokens: 0,
|
|
118
|
-
};
|
|
119
|
-
for (const record of records) {
|
|
120
|
-
if (record.usageStatus === "missing" || !record.usage) {
|
|
121
|
-
totals.missingUsageCalls += 1;
|
|
122
|
-
continue;
|
|
123
|
-
}
|
|
124
|
-
totals.reportedCalls += 1;
|
|
125
|
-
totals.uncachedInputTokens += record.usage.inputTokens;
|
|
126
|
-
totals.cacheReadTokens += record.usage.cacheReadTokens ?? 0;
|
|
127
|
-
totals.cacheWriteTokens += record.usage.cacheWriteTokens ?? 0;
|
|
128
|
-
totals.outputTokens += record.usage.outputTokens;
|
|
129
|
-
totals.reasoningTokens += record.usage.reasoningTokens ?? 0;
|
|
130
|
-
totals.totalTokens += exactTokenTotal(record.usage);
|
|
131
|
-
}
|
|
132
|
-
return totals;
|
|
133
|
-
}
|
|
134
|
-
/** Render the human-facing direct-call ledger section. */
|
|
135
|
-
export function renderTokenUsageReport(ledger, retain) {
|
|
136
|
-
const { records, corruptLines } = ledger;
|
|
137
|
-
if (records.length === 0) {
|
|
138
|
-
const corrupt = corruptLines > 0 ? ` ${corruptLines} corrupt line(s) ignored.` : "";
|
|
139
|
-
return [
|
|
140
|
-
`direct LLM token usage: no valid calls in the retained tail.${corrupt}`,
|
|
141
|
-
`window: last ${retain} call(s), not a lifetime total.`,
|
|
142
|
-
"scope: review/memory/planner/wrapup/fate only; host benchmark subagents, their agent-loop calls, and per-entry injection cost are excluded.",
|
|
143
|
-
];
|
|
144
|
-
}
|
|
145
|
-
const totals = aggregateTokenUsage(records);
|
|
146
|
-
const lines = [
|
|
147
|
-
`direct LLM token usage: ${totals.calls} valid calls (${totals.reportedCalls} reported usage, ${totals.missingUsageCalls} missing)${corruptLines > 0 ? `; ${corruptLines} corrupt line(s) ignored` : ""}`,
|
|
148
|
-
` uncached input: ${totals.uncachedInputTokens}`,
|
|
149
|
-
` cache read: ${totals.cacheReadTokens}`,
|
|
150
|
-
` cache write: ${totals.cacheWriteTokens}`,
|
|
151
|
-
` output: ${totals.outputTokens}`,
|
|
152
|
-
` reasoning: ${totals.reasoningTokens}${records.some((record) => record.usageStatus === "reported" && record.usage?.reasoningTokens === undefined) ? " (partial provider reports)" : ""}`,
|
|
153
|
-
` exact total: ${totals.totalTokens} (reported calls only)`,
|
|
154
|
-
`window: last ${retain} call(s), not a lifetime total.`,
|
|
155
|
-
"by phase:",
|
|
156
|
-
];
|
|
157
|
-
for (const phase of TOKEN_USAGE_PHASES) {
|
|
158
|
-
const phaseRecords = records.filter((record) => record.phase === phase);
|
|
159
|
-
if (phaseRecords.length === 0)
|
|
160
|
-
continue;
|
|
161
|
-
const phaseTotals = aggregateTokenUsage(phaseRecords);
|
|
162
|
-
lines.push(` ${phase.padEnd(7)} ${phaseTotals.calls} calls · ${phaseTotals.totalTokens} tokens`);
|
|
163
|
-
}
|
|
164
|
-
const last = records.at(-1);
|
|
165
|
-
if (last) {
|
|
166
|
-
const lastTotal = last.usage ? exactTokenTotal(last.usage) : "usage unavailable";
|
|
167
|
-
lines.push(`last: ${last.phase} ${last.provider}/${last.model} · ${last.outcome} · ${lastTotal} tokens · ${last.timestamp}`);
|
|
168
|
-
}
|
|
169
|
-
lines.push("scope: review/memory/planner/wrapup/fate only; host benchmark subagents, their agent-loop calls, and per-entry injection cost are excluded.");
|
|
170
|
-
return lines;
|
|
171
|
-
}
|
|
172
|
-
function isTokenUsageRecord(value) {
|
|
173
|
-
if (typeof value !== "object" || value === null || Array.isArray(value))
|
|
174
|
-
return false;
|
|
175
|
-
const record = value;
|
|
176
|
-
if (record["version"] !== TOKEN_USAGE_LEDGER_VERSION ||
|
|
177
|
-
typeof record["timestamp"] !== "string" ||
|
|
178
|
-
typeof record["sessionId"] !== "string" ||
|
|
179
|
-
typeof record["provider"] !== "string" ||
|
|
180
|
-
typeof record["model"] !== "string" ||
|
|
181
|
-
typeof record["outcome"] !== "string" ||
|
|
182
|
-
!TOKEN_USAGE_PHASES.includes(record["phase"]) ||
|
|
183
|
-
!["success", "max-tokens", "error", "aborted", "empty"].includes(record["outcome"]) ||
|
|
184
|
-
!["reported", "missing"].includes(record["usageStatus"]) ||
|
|
185
|
-
!["provider", "derived", "unavailable"].includes(record["totalSource"])) {
|
|
186
|
-
return false;
|
|
187
|
-
}
|
|
188
|
-
if (record["usageStatus"] === "missing") {
|
|
189
|
-
return record["usage"] === undefined && record["totalSource"] === "unavailable";
|
|
190
|
-
}
|
|
191
|
-
if (!isTokenUsage(record["usage"]))
|
|
192
|
-
return false;
|
|
193
|
-
const expectedSource = record["usage"]["totalTokens"] !== undefined ? "provider" : "derived";
|
|
194
|
-
return record["totalSource"] === expectedSource;
|
|
195
|
-
}
|
|
196
|
-
function isCount(value) {
|
|
197
|
-
return typeof value === "number" && Number.isSafeInteger(value) && value >= 0;
|
|
198
|
-
}
|
|
199
|
-
function isTokenUsage(value) {
|
|
200
|
-
if (typeof value !== "object" || value === null || Array.isArray(value))
|
|
201
|
-
return false;
|
|
202
|
-
const usage = value;
|
|
203
|
-
if (!isCount(usage["inputTokens"]) || !isCount(usage["outputTokens"]))
|
|
204
|
-
return false;
|
|
205
|
-
for (const key of ["totalTokens", "cacheReadTokens", "cacheWriteTokens", "reasoningTokens"]) {
|
|
206
|
-
if (usage[key] !== undefined && !isCount(usage[key]))
|
|
207
|
-
return false;
|
|
208
|
-
}
|
|
209
|
-
return true;
|
|
210
|
-
}
|
|
211
|
-
//# sourceMappingURL=token-usage.js.map
|
package/lib/tool.d.ts
DELETED
|
@@ -1,34 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Model-facing evolve_* tools. The model supplies content; every guarantee
|
|
3
|
-
* (validation, snapshot, versioning, history, rollback) is code-enforced in
|
|
4
|
-
* the engine. `global: true` is required explicitly for cross-session edits.
|
|
5
|
-
*/
|
|
6
|
-
import type { Context } from "@deepseek-ai/cordis";
|
|
7
|
-
import type { RecordLanguagePreference } from "./record-language.js";
|
|
8
|
-
import type { HarnessScope } from "./types.js";
|
|
9
|
-
import type { EvolutionEngine } from "./service.js";
|
|
10
|
-
/**
|
|
11
|
-
* Accept the string form (`scope: "project"`) with the legacy boolean tool
|
|
12
|
-
* parameter (`global: true`) as fallback. `scope` wins when both are given.
|
|
13
|
-
*/
|
|
14
|
-
export declare function scopeOf(value: unknown, fallback: HarnessScope): HarnessScope;
|
|
15
|
-
/**
|
|
16
|
-
* Collect the target ids for an evolve_delete call: legacy single `id`
|
|
17
|
-
* plus the batch `ids` array, de-duplicated in first-seen order. Throws
|
|
18
|
-
* loudly when neither carries an id — a silent no-op delete would look
|
|
19
|
-
* like success while deleting nothing.
|
|
20
|
-
*
|
|
21
|
-
* @throws when both `id` and `ids` are absent or empty.
|
|
22
|
-
*/
|
|
23
|
-
export declare function collectDeleteIds(id: unknown, ids: unknown): string[];
|
|
24
|
-
export interface ToolGateOptions {
|
|
25
|
-
requireGlobalApproval: boolean;
|
|
26
|
-
/**
|
|
27
|
-
* Plugin `recordLanguage` preference for the approval dialog copy;
|
|
28
|
-
* `auto`/absent defers to the durable client preference and the session's
|
|
29
|
-
* own user text (see `approval.ts`).
|
|
30
|
-
*/
|
|
31
|
-
recordLanguage?: RecordLanguagePreference;
|
|
32
|
-
}
|
|
33
|
-
export declare function registerEvolveTools(ctx: Context, engine: EvolutionEngine, opts: ToolGateOptions): void;
|
|
34
|
-
//# sourceMappingURL=tool.d.ts.map
|