@devflow-tools/database 0.17.8 → 0.18.0
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 +21 -0
- package/dist/data/LocalDataProvider.d.ts +63 -0
- package/dist/data/LocalDataProvider.js +332 -0
- package/dist/data/default-enforcer-rules.d.ts +3 -0
- package/dist/data/default-enforcer-rules.js +170 -0
- package/dist/data/devflow-schema.d.ts +2 -0
- package/dist/data/devflow-schema.js +80 -0
- package/dist/database.d.ts +45 -7
- package/dist/database.js +583 -125
- package/dist/index.d.ts +8 -4
- package/dist/index.js +11 -5
- package/dist/retrieval-ledger.d.ts +16 -1
- package/dist/retrieval-ledger.js +6 -0
- package/dist/retrieval-runtime.d.ts +39 -0
- package/dist/retrieval-runtime.js +2 -0
- package/dist/task-aggregate.d.ts +49 -0
- package/dist/task-aggregate.js +2 -0
- package/dist/task-semantic-control.d.ts +8 -1
- package/dist/task-semantic-control.js +2 -4
- package/package.json +15 -3
- package/CHANGELOG.md +0 -844
- package/__tests__/database.failure-category.test.ts +0 -44
- package/__tests__/database.host-actions.test.ts +0 -405
- package/__tests__/database.learning-candidates.test.ts +0 -93
- package/__tests__/database.memory-turn-receipts.test.ts +0 -98
- package/__tests__/database.retrieval-ledger.test.ts +0 -68
- package/__tests__/database.retrieval-sessions.test.ts +0 -79
- package/__tests__/database.semantic-resolution.test.ts +0 -87
- package/__tests__/database.skill-executions.test.ts +0 -456
- package/__tests__/database.task-runtime.test.ts +0 -73
- package/__tests__/database.test.ts +0 -177
- package/__tests__/database.work-queue.test.ts +0 -274
- package/__tests__/database.workflow-workers.test.ts +0 -60
- package/__tests__/node-sqlite.test.ts +0 -68
- package/src/database.ts +0 -5853
- package/src/host-actions.ts +0 -211
- package/src/index.ts +0 -153
- package/src/learning-candidates.ts +0 -181
- package/src/node-sqlite.ts +0 -60
- package/src/obligation-ledger.ts +0 -57
- package/src/retrieval-ledger.ts +0 -35
- package/src/retrieval-sessions.ts +0 -196
- package/src/semantic-resolution.ts +0 -181
- package/src/task-runtime.ts +0 -169
- package/src/task-semantic-control.ts +0 -270
- package/src/types.ts +0 -17
- package/src/work-queue.ts +0 -123
- package/src/workflow-workers.ts +0 -198
- package/tsconfig.json +0 -19
- package/tsconfig.tsbuildinfo +0 -1
- package/vitest.config.ts +0 -41
|
@@ -1,68 +0,0 @@
|
|
|
1
|
-
import { mkdtempSync, rmSync } from 'node:fs';
|
|
2
|
-
import { tmpdir } from 'node:os';
|
|
3
|
-
import { join } from 'node:path';
|
|
4
|
-
import { afterEach, describe, expect, it } from 'vitest';
|
|
5
|
-
import { DevFlowDatabase, type RetrievalLedgerEventRecord } from '../src/index.js';
|
|
6
|
-
|
|
7
|
-
describe('retrieval ledger', () => {
|
|
8
|
-
const directories: string[] = [];
|
|
9
|
-
|
|
10
|
-
afterEach(() => {
|
|
11
|
-
for (const directory of directories.splice(0)) rmSync(directory, { recursive: true, force: true });
|
|
12
|
-
});
|
|
13
|
-
|
|
14
|
-
it('is append-only and idempotent per immutable event identity', () => {
|
|
15
|
-
const database = createDatabase();
|
|
16
|
-
const selected = event({ id: 'selected-a', stage: 'selected' });
|
|
17
|
-
expect(database.appendRetrievalLedgerEvent(selected)).toBe(true);
|
|
18
|
-
expect(database.appendRetrievalLedgerEvent({ ...selected, reason: 'changed' })).toBe(false);
|
|
19
|
-
expect(database.appendRetrievalLedgerEvent(event({ id: 'exposed-a', stage: 'exposed' }))).toBe(true);
|
|
20
|
-
expect(database.listRetrievalLedgerEvents({ projectRoot: '/project' })).toEqual(expect.arrayContaining([
|
|
21
|
-
expect.objectContaining({ id: 'selected-a', stage: 'selected', reason: undefined }),
|
|
22
|
-
expect.objectContaining({ id: 'exposed-a', stage: 'exposed' }),
|
|
23
|
-
]));
|
|
24
|
-
database.close();
|
|
25
|
-
});
|
|
26
|
-
|
|
27
|
-
it('rejects score loss between returned evidence and persisted selected evidence', () => {
|
|
28
|
-
const database = createDatabase();
|
|
29
|
-
expect(() => database.appendRetrievalLedgerEvent(event({
|
|
30
|
-
id: 'score-loss',
|
|
31
|
-
stage: 'selected',
|
|
32
|
-
finalScore: 0,
|
|
33
|
-
payload: { returnedFinalScore: 0.81 },
|
|
34
|
-
}))).toThrow('RETRIEVAL_LEDGER_INVALID_FINAL_SCORE');
|
|
35
|
-
database.close();
|
|
36
|
-
});
|
|
37
|
-
|
|
38
|
-
it('persists v1 provenance and reads legacy adopted rows as consumed', () => {
|
|
39
|
-
const database = createDatabase();
|
|
40
|
-
database.appendRetrievalLedgerEvent(event({
|
|
41
|
-
id: 'legacy-adopted', stage: 'adopted', taskSpecHash: 'spec:1', actor: 'tool:read',
|
|
42
|
-
sourceVersion: 'v1', sourceContentHash: 'sha256:1', evidenceIds: ['memory-a'],
|
|
43
|
-
reasonCode: 'structural_evidence_ref',
|
|
44
|
-
}));
|
|
45
|
-
expect(database.listRetrievalLedgerEvents({ projectRoot: '/project' })[0]).toMatchObject({
|
|
46
|
-
schemaVersion: 'retrieval-ledger-event.v1', stage: 'consumed', taskSpecHash: 'spec:1',
|
|
47
|
-
actor: 'tool:read', sourceVersion: 'v1', sourceContentHash: 'sha256:1',
|
|
48
|
-
evidenceIds: ['memory-a'], reasonCode: 'structural_evidence_ref',
|
|
49
|
-
});
|
|
50
|
-
database.close();
|
|
51
|
-
});
|
|
52
|
-
|
|
53
|
-
function createDatabase(): DevFlowDatabase {
|
|
54
|
-
const directory = mkdtempSync(join(tmpdir(), 'devflow-retrieval-ledger-'));
|
|
55
|
-
directories.push(directory);
|
|
56
|
-
return new DevFlowDatabase(join(directory, 'devflow.db'));
|
|
57
|
-
}
|
|
58
|
-
|
|
59
|
-
function event(overrides: Partial<RetrievalLedgerEventRecord>): RetrievalLedgerEventRecord {
|
|
60
|
-
return {
|
|
61
|
-
id: 'event-a', projectRoot: '/project', sessionId: 'session-a', executionId: 'execution-a',
|
|
62
|
-
requestId: 'request-a', contextReceipt: 'context-a', sourceType: 'memory', sourceId: 'memory-a',
|
|
63
|
-
stage: 'candidate', rank: 1, rawScore: 0.7, normalizedScore: 0.75, finalScore: 0.8,
|
|
64
|
-
applicability: ['formatting'], toolEvidence: [], payload: { returnedFinalScore: 0.8 }, createdAt: 100,
|
|
65
|
-
...overrides,
|
|
66
|
-
};
|
|
67
|
-
}
|
|
68
|
-
});
|
|
@@ -1,79 +0,0 @@
|
|
|
1
|
-
import { mkdtempSync, rmSync } from 'node:fs';
|
|
2
|
-
import { tmpdir } from 'node:os';
|
|
3
|
-
import { join } from 'node:path';
|
|
4
|
-
import { afterEach, beforeEach, describe, expect, it } from 'vitest';
|
|
5
|
-
import { DevFlowDatabase } from '../src/database';
|
|
6
|
-
|
|
7
|
-
describe('DevFlowDatabase retrieval sessions', () => {
|
|
8
|
-
let root: string;
|
|
9
|
-
let database: DevFlowDatabase;
|
|
10
|
-
|
|
11
|
-
beforeEach(() => {
|
|
12
|
-
root = mkdtempSync(join(tmpdir(), 'devflow-retrieval-session-'));
|
|
13
|
-
database = new DevFlowDatabase(root);
|
|
14
|
-
});
|
|
15
|
-
|
|
16
|
-
afterEach(() => {
|
|
17
|
-
database.close();
|
|
18
|
-
rmSync(root, { recursive: true, force: true });
|
|
19
|
-
});
|
|
20
|
-
|
|
21
|
-
const create = () => database.createRetrievalSession({
|
|
22
|
-
id: 'retrieval-1', requestId: 'request-1', projectRoot: '/project-a',
|
|
23
|
-
sessionId: 'session-1', executionId: 'execution-1', query: 'review player',
|
|
24
|
-
intent: 'performance', tokenBudget: 1000, baselineReceipt: 'baseline-1',
|
|
25
|
-
expiresAt: Date.now() + 60_000,
|
|
26
|
-
});
|
|
27
|
-
|
|
28
|
-
it('persists idempotent cycles and rejects changed evidence', () => {
|
|
29
|
-
expect(create()).toMatchObject({ id: 'retrieval-1', cycle: 0, remainingTokenBudget: 1000 });
|
|
30
|
-
expect(create()).toMatchObject({ id: 'retrieval-1' });
|
|
31
|
-
const input = {
|
|
32
|
-
retrievalSessionId: 'retrieval-1', projectRoot: '/project-a', sessionId: 'session-1',
|
|
33
|
-
baselineReceipt: 'baseline-1', cycle: 1,
|
|
34
|
-
gaps: [{ kind: 'caller' as const, reason: 'missing caller', evidence: [], resolver: 'codegraph' as const, priority: 80 }],
|
|
35
|
-
selectedIds: ['code:src/player.tsx'], rejectedIds: [], tokenCost: 200,
|
|
36
|
-
remainingTokenBudget: 800, quality: { status: 'healthy' }, receipt: 'cycle-1', evidenceHash: 'hash-1',
|
|
37
|
-
};
|
|
38
|
-
expect(database.appendRetrievalCycle(input)).toMatchObject({ cycle: 1, tokenCost: 200 });
|
|
39
|
-
expect(database.appendRetrievalCycle(input)).toMatchObject({ evidenceHash: 'hash-1' });
|
|
40
|
-
expect(() => database.appendRetrievalCycle({ ...input, evidenceHash: 'changed' }))
|
|
41
|
-
.toThrow('RETRIEVAL_CYCLE_CONFLICT:1');
|
|
42
|
-
expect(database.getRetrievalSession('/project-a', 'session-1', 'retrieval-1'))
|
|
43
|
-
.toMatchObject({ cycle: 1, remainingTokenBudget: 800 });
|
|
44
|
-
});
|
|
45
|
-
|
|
46
|
-
it('enforces ownership, sequence, terminal state, and expiration', () => {
|
|
47
|
-
create();
|
|
48
|
-
expect(database.getRetrievalSession('/project-b', 'session-1', 'retrieval-1')).toBeNull();
|
|
49
|
-
expect(() => database.appendRetrievalCycle({
|
|
50
|
-
retrievalSessionId: 'retrieval-1', projectRoot: '/project-a', sessionId: 'session-1',
|
|
51
|
-
baselineReceipt: 'baseline-1', cycle: 2, gaps: [], selectedIds: [], rejectedIds: [],
|
|
52
|
-
tokenCost: 1, remainingTokenBudget: 999, quality: {}, receipt: 'cycle-2', evidenceHash: 'hash-2',
|
|
53
|
-
})).toThrow('RETRIEVAL_CYCLE_SEQUENCE:0->2');
|
|
54
|
-
database.finalizeRetrievalSession({
|
|
55
|
-
id: 'retrieval-1', projectRoot: '/project-a', sessionId: 'session-1',
|
|
56
|
-
state: 'exhausted', finalReceipt: 'final-1',
|
|
57
|
-
});
|
|
58
|
-
expect(() => database.finalizeRetrievalSession({
|
|
59
|
-
id: 'retrieval-1', projectRoot: '/project-a', sessionId: 'session-1',
|
|
60
|
-
state: 'satisfied', finalReceipt: 'different',
|
|
61
|
-
})).toThrow('RETRIEVAL_SESSION_TERMINAL:exhausted');
|
|
62
|
-
|
|
63
|
-
database.createRetrievalSession({
|
|
64
|
-
id: 'retrieval-expire', requestId: 'request-expire', projectRoot: '/project-a',
|
|
65
|
-
sessionId: 'session-1', query: 'x', intent: 'debug', tokenBudget: 1,
|
|
66
|
-
baselineReceipt: 'baseline-expire', expiresAt: Date.now() + 10,
|
|
67
|
-
});
|
|
68
|
-
expect(database.expireRetrievalSessions(Date.now() + 20)).toBe(1);
|
|
69
|
-
expect(database.getRetrievalSession('/project-a', 'session-1', 'retrieval-expire')?.state).toBe('expired');
|
|
70
|
-
});
|
|
71
|
-
|
|
72
|
-
it('keeps retrieval rows stable across reopen', () => {
|
|
73
|
-
create();
|
|
74
|
-
database.close();
|
|
75
|
-
database = new DevFlowDatabase(root);
|
|
76
|
-
expect(database.getRetrievalSession('/project-a', 'session-1', 'retrieval-1'))
|
|
77
|
-
.toMatchObject({ requestId: 'request-1', baselineReceipt: 'baseline-1' });
|
|
78
|
-
});
|
|
79
|
-
});
|
|
@@ -1,87 +0,0 @@
|
|
|
1
|
-
import { mkdtempSync, rmSync } from 'node:fs';
|
|
2
|
-
import { tmpdir } from 'node:os';
|
|
3
|
-
import { join } from 'node:path';
|
|
4
|
-
import { afterEach, beforeEach, describe, expect, it } from 'vitest';
|
|
5
|
-
import { DevFlowDatabase } from '../src/database';
|
|
6
|
-
import type { SemanticResolutionRecord } from '../src/semantic-resolution';
|
|
7
|
-
|
|
8
|
-
describe('DevFlowDatabase semantic resolution persistence', () => {
|
|
9
|
-
let directory: string;
|
|
10
|
-
let database: DevFlowDatabase;
|
|
11
|
-
|
|
12
|
-
beforeEach(() => {
|
|
13
|
-
directory = mkdtempSync(join(tmpdir(), 'devflow-semantic-resolution-'));
|
|
14
|
-
database = new DevFlowDatabase(directory);
|
|
15
|
-
});
|
|
16
|
-
|
|
17
|
-
afterEach(() => {
|
|
18
|
-
database.close();
|
|
19
|
-
rmSync(directory, { recursive: true, force: true });
|
|
20
|
-
});
|
|
21
|
-
|
|
22
|
-
it('appends, lists, and returns the latest immutable revision', () => {
|
|
23
|
-
const provisional = resolution();
|
|
24
|
-
const resolved = resolution({
|
|
25
|
-
frameHash: 'b'.repeat(64),
|
|
26
|
-
artifactHash: 'artifact:2',
|
|
27
|
-
artifactRevision: 2,
|
|
28
|
-
supersedesFrameHash: provisional.frameHash,
|
|
29
|
-
state: 'resolved',
|
|
30
|
-
frame: { frameHash: 'b'.repeat(64), state: 'resolved' },
|
|
31
|
-
createdAt: 2,
|
|
32
|
-
});
|
|
33
|
-
expect(database.appendSemanticResolution(provisional)).toBe(true);
|
|
34
|
-
expect(database.appendSemanticResolution(resolved)).toBe(true);
|
|
35
|
-
expect(database.listSemanticResolutions('/project', 'session:1', 'turn:1'))
|
|
36
|
-
.toEqual([provisional, resolved]);
|
|
37
|
-
expect(database.getLatestSemanticResolution('/project', 'session:1', 'turn:1'))
|
|
38
|
-
.toEqual(resolved);
|
|
39
|
-
});
|
|
40
|
-
|
|
41
|
-
it('makes identical replay idempotent and rejects hash conflicts', () => {
|
|
42
|
-
const record = resolution();
|
|
43
|
-
expect(database.appendSemanticResolution(record)).toBe(true);
|
|
44
|
-
expect(database.appendSemanticResolution(record)).toBe(false);
|
|
45
|
-
expect(() => database.appendSemanticResolution({
|
|
46
|
-
...record,
|
|
47
|
-
durationMs: 9,
|
|
48
|
-
})).toThrow(`SEMANTIC_RESOLUTION_HASH_CONFLICT:${record.frameHash}`);
|
|
49
|
-
});
|
|
50
|
-
|
|
51
|
-
it('rejects another frame at the same artifact revision', () => {
|
|
52
|
-
const record = resolution();
|
|
53
|
-
database.appendSemanticResolution(record);
|
|
54
|
-
expect(() => database.appendSemanticResolution({
|
|
55
|
-
...record,
|
|
56
|
-
frameHash: 'c'.repeat(64),
|
|
57
|
-
frame: { frameHash: 'c'.repeat(64), state: 'provisional' },
|
|
58
|
-
})).toThrow('SEMANTIC_RESOLUTION_REVISION_CONFLICT:turn:1:1');
|
|
59
|
-
});
|
|
60
|
-
});
|
|
61
|
-
|
|
62
|
-
function resolution(overrides: Partial<SemanticResolutionRecord> = {}): SemanticResolutionRecord {
|
|
63
|
-
const frameHash = overrides.frameHash ?? 'a'.repeat(64);
|
|
64
|
-
return {
|
|
65
|
-
frameHash,
|
|
66
|
-
projectRoot: '/project',
|
|
67
|
-
projectId: 'project:1',
|
|
68
|
-
hostId: 'claude-code',
|
|
69
|
-
sessionId: 'session:1',
|
|
70
|
-
turnId: 'turn:1',
|
|
71
|
-
requestId: 'request:1',
|
|
72
|
-
sourceHash: 'source:1',
|
|
73
|
-
artifactHash: 'artifact:1',
|
|
74
|
-
artifactRevision: 1,
|
|
75
|
-
state: 'provisional',
|
|
76
|
-
generatedBy: 'deterministic',
|
|
77
|
-
routeCatalogVersion: 'catalog.v1',
|
|
78
|
-
thresholdVersion: 'threshold.v1',
|
|
79
|
-
route: { status: 'unavailable' },
|
|
80
|
-
sampling: { attempted: false, status: 'not_needed' },
|
|
81
|
-
conflicts: [],
|
|
82
|
-
frame: { frameHash, state: 'provisional' },
|
|
83
|
-
durationMs: 0,
|
|
84
|
-
createdAt: 1,
|
|
85
|
-
...overrides,
|
|
86
|
-
};
|
|
87
|
-
}
|
|
@@ -1,456 +0,0 @@
|
|
|
1
|
-
import { describe, it, expect, beforeEach, afterEach } from 'vitest';
|
|
2
|
-
import { DevFlowDatabase } from '../src/database';
|
|
3
|
-
import { tmpdir } from 'os';
|
|
4
|
-
import { join } from 'path';
|
|
5
|
-
import { mkdtempSync, rmSync } from 'fs';
|
|
6
|
-
|
|
7
|
-
describe('DevFlowDatabase - Skill Executions', () => {
|
|
8
|
-
let db: DevFlowDatabase;
|
|
9
|
-
let tmpDir: string;
|
|
10
|
-
|
|
11
|
-
beforeEach(() => {
|
|
12
|
-
tmpDir = mkdtempSync(join(tmpdir(), 'devflow-test-'));
|
|
13
|
-
db = new DevFlowDatabase(tmpDir);
|
|
14
|
-
});
|
|
15
|
-
|
|
16
|
-
afterEach(() => {
|
|
17
|
-
db.close();
|
|
18
|
-
rmSync(tmpDir, { recursive: true, force: true });
|
|
19
|
-
});
|
|
20
|
-
|
|
21
|
-
it('should insert and retrieve a skill execution', () => {
|
|
22
|
-
db.insertSkillExecution({
|
|
23
|
-
executionId: 'exec_001',
|
|
24
|
-
skillName: 'devflow:react',
|
|
25
|
-
startedAt: Date.now(),
|
|
26
|
-
status: 'running',
|
|
27
|
-
});
|
|
28
|
-
|
|
29
|
-
const execution = db.getSkillExecution('exec_001');
|
|
30
|
-
expect(execution).not.toBeNull();
|
|
31
|
-
expect(execution!.executionId).toBe('exec_001');
|
|
32
|
-
expect(execution!.skillName).toBe('devflow:react');
|
|
33
|
-
expect(execution!.status).toBe('running');
|
|
34
|
-
expect(execution!.totalToolCalls).toBe(0);
|
|
35
|
-
});
|
|
36
|
-
|
|
37
|
-
it('should list skill executions with limit', () => {
|
|
38
|
-
db.insertSkillExecution({
|
|
39
|
-
executionId: 'exec_001',
|
|
40
|
-
skillName: 'devflow:react',
|
|
41
|
-
startedAt: Date.now() - 1000,
|
|
42
|
-
status: 'completed',
|
|
43
|
-
});
|
|
44
|
-
db.insertSkillExecution({
|
|
45
|
-
executionId: 'exec_002',
|
|
46
|
-
skillName: 'devflow:vue',
|
|
47
|
-
startedAt: Date.now(),
|
|
48
|
-
status: 'running',
|
|
49
|
-
});
|
|
50
|
-
|
|
51
|
-
const executions = db.listSkillExecutions(10);
|
|
52
|
-
expect(executions).toHaveLength(2);
|
|
53
|
-
// Most recent first
|
|
54
|
-
expect(executions[0].executionId).toBe('exec_002');
|
|
55
|
-
});
|
|
56
|
-
|
|
57
|
-
it('should update skill execution on complete', () => {
|
|
58
|
-
db.insertSkillExecution({
|
|
59
|
-
executionId: 'exec_001',
|
|
60
|
-
skillName: 'devflow:react',
|
|
61
|
-
startedAt: Date.now() - 5000,
|
|
62
|
-
status: 'running',
|
|
63
|
-
});
|
|
64
|
-
|
|
65
|
-
db.updateSkillExecution('exec_001', {
|
|
66
|
-
status: 'completed',
|
|
67
|
-
finishedAt: Date.now(),
|
|
68
|
-
totalDuration: 5000,
|
|
69
|
-
totalTokens: 1500,
|
|
70
|
-
mcpComplianceRate: 85.5,
|
|
71
|
-
});
|
|
72
|
-
|
|
73
|
-
const execution = db.getSkillExecution('exec_001');
|
|
74
|
-
expect(execution!.status).toBe('completed');
|
|
75
|
-
expect(execution!.totalDuration).toBe(5000);
|
|
76
|
-
expect(execution!.mcpComplianceRate).toBe(85.5);
|
|
77
|
-
});
|
|
78
|
-
|
|
79
|
-
it('should insert and list tool call events', () => {
|
|
80
|
-
db.insertSkillExecution({
|
|
81
|
-
executionId: 'exec_001',
|
|
82
|
-
skillName: 'devflow:react',
|
|
83
|
-
startedAt: Date.now(),
|
|
84
|
-
status: 'running',
|
|
85
|
-
});
|
|
86
|
-
|
|
87
|
-
db.insertToolCallEvent({
|
|
88
|
-
eventId: 'evt_001',
|
|
89
|
-
executionId: 'exec_001',
|
|
90
|
-
timestamp: Date.now(),
|
|
91
|
-
toolName: 'mcp__devflow__react_diagnose_bug',
|
|
92
|
-
toolType: 'mcp',
|
|
93
|
-
isMcpTool: true,
|
|
94
|
-
mcpToolName: 'react_diagnose_bug',
|
|
95
|
-
mcpEnforced: true,
|
|
96
|
-
mcpFallback: false,
|
|
97
|
-
input: { bug: 'white screen' },
|
|
98
|
-
tokensUsed: 500,
|
|
99
|
-
duration: 2000,
|
|
100
|
-
blocked: false,
|
|
101
|
-
});
|
|
102
|
-
|
|
103
|
-
db.insertToolCallEvent({
|
|
104
|
-
eventId: 'evt_002',
|
|
105
|
-
executionId: 'exec_001',
|
|
106
|
-
timestamp: Date.now() + 100,
|
|
107
|
-
toolName: 'Read',
|
|
108
|
-
toolType: 'direct',
|
|
109
|
-
isMcpTool: false,
|
|
110
|
-
mcpEnforced: false,
|
|
111
|
-
mcpFallback: true,
|
|
112
|
-
input: { file_path: '/src/App.tsx' },
|
|
113
|
-
tokensUsed: 100,
|
|
114
|
-
duration: 50,
|
|
115
|
-
blocked: false,
|
|
116
|
-
});
|
|
117
|
-
|
|
118
|
-
const events = db.listToolCallEvents('exec_001');
|
|
119
|
-
expect(events).toHaveLength(2);
|
|
120
|
-
expect(events[0].toolName).toBe('mcp__devflow__react_diagnose_bug');
|
|
121
|
-
expect(events[0].isMcpTool).toBe(true);
|
|
122
|
-
expect(events[1].toolName).toBe('Read');
|
|
123
|
-
expect(events[1].isMcpTool).toBe(false);
|
|
124
|
-
});
|
|
125
|
-
|
|
126
|
-
it('should filter executions by skill name', () => {
|
|
127
|
-
db.insertSkillExecution({
|
|
128
|
-
executionId: 'exec_001',
|
|
129
|
-
skillName: 'devflow:react',
|
|
130
|
-
startedAt: Date.now(),
|
|
131
|
-
status: 'completed',
|
|
132
|
-
});
|
|
133
|
-
db.insertSkillExecution({
|
|
134
|
-
executionId: 'exec_002',
|
|
135
|
-
skillName: 'devflow:vue',
|
|
136
|
-
startedAt: Date.now(),
|
|
137
|
-
status: 'completed',
|
|
138
|
-
});
|
|
139
|
-
|
|
140
|
-
const reactExecs = db.listSkillExecutions(10, 'devflow:react');
|
|
141
|
-
expect(reactExecs).toHaveLength(1);
|
|
142
|
-
expect(reactExecs[0].skillName).toBe('devflow:react');
|
|
143
|
-
});
|
|
144
|
-
|
|
145
|
-
it('computes required MCP obligation compliance independently of direct calls', () => {
|
|
146
|
-
// React requires successful context plus at least one React domain tool.
|
|
147
|
-
db.insertSkillExecution({
|
|
148
|
-
executionId: 'exec_001',
|
|
149
|
-
skillName: 'devflow:react',
|
|
150
|
-
startedAt: Date.now(),
|
|
151
|
-
status: 'completed',
|
|
152
|
-
});
|
|
153
|
-
db.insertToolCallEvent({
|
|
154
|
-
eventId: 'evt_context_1', executionId: 'exec_001', timestamp: Date.now(),
|
|
155
|
-
toolName: 'mcp__devflow__get_project_context', toolType: 'mcp',
|
|
156
|
-
isMcpTool: true, mcpToolName: 'get_project_context', mcpEnforced: true,
|
|
157
|
-
mcpFallback: false, input: {}, output: { files: ['src/App.tsx'] }, duration: 20, blocked: false,
|
|
158
|
-
});
|
|
159
|
-
db.insertToolCallEvent({
|
|
160
|
-
eventId: 'evt_001', executionId: 'exec_001', timestamp: Date.now(),
|
|
161
|
-
toolName: 'mcp__devflow__react_diagnose_bug', toolType: 'mcp',
|
|
162
|
-
isMcpTool: true, mcpEnforced: true, mcpFallback: false,
|
|
163
|
-
input: {}, output: { findings: [{ id: 'finding-1' }] }, tokensUsed: 100, duration: 500, blocked: false,
|
|
164
|
-
});
|
|
165
|
-
db.insertToolCallEvent({
|
|
166
|
-
eventId: 'evt_002', executionId: 'exec_001', timestamp: Date.now(),
|
|
167
|
-
toolName: 'mcp__devflow__react_review_hooks', toolType: 'mcp',
|
|
168
|
-
isMcpTool: true, mcpEnforced: true, mcpFallback: false,
|
|
169
|
-
input: {}, output: { findings: [{ id: 'finding-2' }] }, tokensUsed: 100, duration: 500, blocked: false,
|
|
170
|
-
});
|
|
171
|
-
db.insertToolCallEvent({
|
|
172
|
-
eventId: 'evt_003', executionId: 'exec_001', timestamp: Date.now(),
|
|
173
|
-
toolName: 'Read', toolType: 'direct',
|
|
174
|
-
isMcpTool: false, mcpEnforced: false, mcpFallback: false,
|
|
175
|
-
input: {}, tokensUsed: 50, duration: 100, blocked: false,
|
|
176
|
-
});
|
|
177
|
-
|
|
178
|
-
// Vue has the same context + domain obligations.
|
|
179
|
-
db.insertSkillExecution({
|
|
180
|
-
executionId: 'exec_002',
|
|
181
|
-
skillName: 'devflow:vue',
|
|
182
|
-
startedAt: Date.now(),
|
|
183
|
-
status: 'completed',
|
|
184
|
-
});
|
|
185
|
-
db.insertToolCallEvent({
|
|
186
|
-
eventId: 'evt_context_2', executionId: 'exec_002', timestamp: Date.now(),
|
|
187
|
-
toolName: 'mcp__devflow__get_project_context', toolType: 'mcp',
|
|
188
|
-
isMcpTool: true, mcpToolName: 'get_project_context', mcpEnforced: true,
|
|
189
|
-
mcpFallback: false, input: {}, output: { files: ['src/App.vue'] }, duration: 20, blocked: false,
|
|
190
|
-
});
|
|
191
|
-
db.insertToolCallEvent({
|
|
192
|
-
eventId: 'evt_004', executionId: 'exec_002', timestamp: Date.now(),
|
|
193
|
-
toolName: 'mcp__devflow__vue_diagnose_bug', toolType: 'mcp',
|
|
194
|
-
isMcpTool: true, mcpEnforced: true, mcpFallback: false,
|
|
195
|
-
input: {}, output: { findings: [{ id: 'finding-vue' }] }, tokensUsed: 100, duration: 500, blocked: false,
|
|
196
|
-
});
|
|
197
|
-
|
|
198
|
-
const compliance = db.getMcpCompliance();
|
|
199
|
-
expect(compliance.overall).toBe(100);
|
|
200
|
-
expect(compliance.bySkill['devflow:react']).toBe(100);
|
|
201
|
-
expect(compliance.bySkill['devflow:vue']).toBeCloseTo(100, 0);
|
|
202
|
-
expect(compliance.applicableObligations).toBe(4);
|
|
203
|
-
expect(compliance.satisfiedObligations).toBe(4);
|
|
204
|
-
expect(compliance.mcpCallShare).toBeCloseTo(83.33, 1);
|
|
205
|
-
});
|
|
206
|
-
|
|
207
|
-
it('persists missed tools, actual failures, and fallback reasons on reconciliation', () => {
|
|
208
|
-
db.ensureSession({ id: 'session-facts', projectRoot: '/project', startedAt: 100 });
|
|
209
|
-
db.insertSkillExecution({
|
|
210
|
-
executionId: 'exec-facts',
|
|
211
|
-
sessionId: 'session-facts',
|
|
212
|
-
skillName: 'devflow:react',
|
|
213
|
-
startedAt: 100,
|
|
214
|
-
status: 'running',
|
|
215
|
-
});
|
|
216
|
-
db.insertToolCallEvent({
|
|
217
|
-
eventId: 'event-failed', executionId: 'exec-facts', sessionId: 'session-facts', timestamp: 150,
|
|
218
|
-
toolName: 'mcp__devflow__get_project_context', toolType: 'mcp', isMcpTool: true,
|
|
219
|
-
mcpToolName: 'get_project_context', mcpEnforced: true, mcpFallback: false,
|
|
220
|
-
input: {}, error: 'route failed', duration: 20, blocked: false,
|
|
221
|
-
});
|
|
222
|
-
db.insertToolCallEvent({
|
|
223
|
-
eventId: 'event-fallback', executionId: 'exec-facts', sessionId: 'session-facts', timestamp: 160,
|
|
224
|
-
toolName: 'Read', toolType: 'direct', isMcpTool: false, mcpEnforced: false,
|
|
225
|
-
mcpFallback: true, input: {}, output: { content: 'fallback' }, duration: 10, blocked: false,
|
|
226
|
-
});
|
|
227
|
-
db.insertToolCallEvent({
|
|
228
|
-
eventId: 'event-memory-receipt', executionId: 'exec-facts', sessionId: 'session-facts', timestamp: 165,
|
|
229
|
-
toolName: 'mcp__devflow__memory_commit_turn', toolType: 'mcp', isMcpTool: true,
|
|
230
|
-
mcpToolName: 'memory_commit_turn', mcpEnforced: true, mcpFallback: false,
|
|
231
|
-
input: {}, output: { receiptId: 'memory-receipt:host-commit' }, duration: 5, blocked: false,
|
|
232
|
-
});
|
|
233
|
-
db.insertToolCallEvent({
|
|
234
|
-
eventId: 'event-distill-receipt', executionId: 'exec-facts', sessionId: 'session-facts', timestamp: 166,
|
|
235
|
-
toolName: 'mcp__devflow__memory_save_distilled', toolType: 'mcp', isMcpTool: true,
|
|
236
|
-
mcpToolName: 'memory_save_distilled', mcpEnforced: true, mcpFallback: false,
|
|
237
|
-
input: {}, output: { receipt: { receiptId: 'distill-receipt:batch-1' } }, duration: 5, blocked: false,
|
|
238
|
-
});
|
|
239
|
-
db.insertHookFallback({
|
|
240
|
-
id: 'hook-fallback-1', projectRoot: '/project', sessionId: 'session-facts',
|
|
241
|
-
requestType: 'post-tool-use', tool: 'Read', reason: 'timeout', durationMs: 1_000,
|
|
242
|
-
attempts: 1, createdAt: 170,
|
|
243
|
-
});
|
|
244
|
-
|
|
245
|
-
db.reconcileSkillExecution('exec-facts', 'completed', 200, {
|
|
246
|
-
missedTools: [],
|
|
247
|
-
actualFailureCount: 0,
|
|
248
|
-
blockedCount: 99,
|
|
249
|
-
fallbackCount: 0,
|
|
250
|
-
fallbackReasons: ['stale_supplied_reason'],
|
|
251
|
-
evidenceContractOutcomes: [{ status: 'resolved' }],
|
|
252
|
-
});
|
|
253
|
-
|
|
254
|
-
expect(db.getSkillExecution('exec-facts')).toMatchObject({
|
|
255
|
-
missedMcpTools: ['react:context', 'react:domain'],
|
|
256
|
-
failedToolCalls: 1,
|
|
257
|
-
blockedToolCalls: 0,
|
|
258
|
-
fallbackCount: 2,
|
|
259
|
-
fallbackReasons: ['direct_tool_during_context', 'daemon_timeout'],
|
|
260
|
-
metadata: expect.objectContaining({
|
|
261
|
-
missedTools: ['react:context', 'react:domain'],
|
|
262
|
-
actualFailureCount: 1,
|
|
263
|
-
blockedCount: 0,
|
|
264
|
-
fallbackCount: 2,
|
|
265
|
-
fallbackReasons: ['direct_tool_during_context', 'daemon_timeout'],
|
|
266
|
-
memoryReceiptIds: ['memory-receipt:host-commit'],
|
|
267
|
-
distillReceiptIds: ['distill-receipt:batch-1'],
|
|
268
|
-
evidenceContractOutcomes: [{ status: 'resolved' }],
|
|
269
|
-
}),
|
|
270
|
-
});
|
|
271
|
-
});
|
|
272
|
-
|
|
273
|
-
it('unions evidence metadata across retries without allowing stale counters', () => {
|
|
274
|
-
db.ensureSession({ id: 'session-evidence-retry', projectRoot: '/project', startedAt: 100 });
|
|
275
|
-
db.insertSkillExecution({
|
|
276
|
-
executionId: 'exec-evidence-retry',
|
|
277
|
-
sessionId: 'session-evidence-retry',
|
|
278
|
-
skillName: 'devflow:context',
|
|
279
|
-
startedAt: 100,
|
|
280
|
-
status: 'running',
|
|
281
|
-
});
|
|
282
|
-
db.insertToolCallEvent({
|
|
283
|
-
eventId: 'event-context', executionId: 'exec-evidence-retry',
|
|
284
|
-
sessionId: 'session-evidence-retry', timestamp: 120,
|
|
285
|
-
toolName: 'get_project_context', toolType: 'mcp', isMcpTool: true,
|
|
286
|
-
mcpToolName: 'get_project_context', mcpEnforced: true, mcpFallback: false,
|
|
287
|
-
input: {}, output: { files: ['src/App.tsx'] }, duration: 10, blocked: false,
|
|
288
|
-
});
|
|
289
|
-
const degradation = {
|
|
290
|
-
obligationId: 'obligation-1', contractId: 'contract-1', reason: 'bounded_fail_open',
|
|
291
|
-
degradedAt: 150, attempt: 2,
|
|
292
|
-
};
|
|
293
|
-
const outcome = {
|
|
294
|
-
obligationId: 'obligation-1', contractId: 'contract-1', status: 'degraded',
|
|
295
|
-
completedAt: 150, attempt: 2, reason: 'bounded_fail_open',
|
|
296
|
-
};
|
|
297
|
-
|
|
298
|
-
db.reconcileSkillExecution('exec-evidence-retry', 'completed', 200, {
|
|
299
|
-
evidenceDegradations: [degradation],
|
|
300
|
-
evidenceContractOutcomes: [outcome],
|
|
301
|
-
pendingEvidenceContractIds: ['contract-pending'],
|
|
302
|
-
});
|
|
303
|
-
db.reconcileSkillExecution('exec-evidence-retry', 'completed', 220, {
|
|
304
|
-
evidenceDegradations: [],
|
|
305
|
-
evidenceContractOutcomes: [
|
|
306
|
-
{ attempt: 2, completedAt: 150, reason: 'bounded_fail_open', status: 'degraded',
|
|
307
|
-
contractId: 'contract-1', obligationId: 'obligation-1' },
|
|
308
|
-
],
|
|
309
|
-
pendingEvidenceContractIds: [],
|
|
310
|
-
missedTools: ['stale'],
|
|
311
|
-
actualFailureCount: 99,
|
|
312
|
-
});
|
|
313
|
-
|
|
314
|
-
expect(db.getSkillExecution('exec-evidence-retry')?.metadata).toMatchObject({
|
|
315
|
-
evidenceDegradations: [degradation],
|
|
316
|
-
evidenceContractOutcomes: [outcome],
|
|
317
|
-
pendingEvidenceContractIds: ['contract-pending'],
|
|
318
|
-
missedTools: [],
|
|
319
|
-
actualFailureCount: 0,
|
|
320
|
-
});
|
|
321
|
-
});
|
|
322
|
-
|
|
323
|
-
it('aggregates an empty findings collection as an empty result', () => {
|
|
324
|
-
db.insertSkillExecution({
|
|
325
|
-
executionId: 'exec_empty_findings',
|
|
326
|
-
sessionId: 'session_empty_findings',
|
|
327
|
-
skillName: 'devflow:react',
|
|
328
|
-
startedAt: 1_000,
|
|
329
|
-
status: 'running',
|
|
330
|
-
});
|
|
331
|
-
db.insertToolCallEvent({
|
|
332
|
-
eventId: 'evt_empty_findings',
|
|
333
|
-
executionId: 'exec_empty_findings',
|
|
334
|
-
sessionId: 'session_empty_findings',
|
|
335
|
-
timestamp: 2_000,
|
|
336
|
-
toolName: 'react_review_hooks',
|
|
337
|
-
toolType: 'mcp',
|
|
338
|
-
isMcpTool: true,
|
|
339
|
-
mcpToolName: 'react_review_hooks',
|
|
340
|
-
mcpEnforced: true,
|
|
341
|
-
mcpFallback: false,
|
|
342
|
-
input: { query: 'review hooks' },
|
|
343
|
-
output: { findings: [] },
|
|
344
|
-
duration: 25,
|
|
345
|
-
blocked: false,
|
|
346
|
-
});
|
|
347
|
-
|
|
348
|
-
expect(db.aggregatePendingToolMetrics()).toBe(1);
|
|
349
|
-
expect(db.getToolMetrics('session_empty_findings')).toEqual([
|
|
350
|
-
expect.objectContaining({ status: 'empty', result_count: 0 }),
|
|
351
|
-
]);
|
|
352
|
-
});
|
|
353
|
-
|
|
354
|
-
it.each([
|
|
355
|
-
['success false', { success: false }],
|
|
356
|
-
['degraded data', { success: true, data: { degraded: true } }],
|
|
357
|
-
])('aggregates %s as degraded', (_caseName, output) => {
|
|
358
|
-
const suffix = output.success === false ? 'success_false' : 'data_degraded';
|
|
359
|
-
const executionId = `exec_${suffix}`;
|
|
360
|
-
const sessionId = `session_${suffix}`;
|
|
361
|
-
db.insertSkillExecution({
|
|
362
|
-
executionId,
|
|
363
|
-
sessionId,
|
|
364
|
-
skillName: 'devflow:react',
|
|
365
|
-
startedAt: 1_000,
|
|
366
|
-
status: 'running',
|
|
367
|
-
});
|
|
368
|
-
db.insertToolCallEvent({
|
|
369
|
-
eventId: `evt_${suffix}`,
|
|
370
|
-
executionId,
|
|
371
|
-
sessionId,
|
|
372
|
-
timestamp: 2_000,
|
|
373
|
-
toolName: 'react_audit_performance',
|
|
374
|
-
toolType: 'mcp',
|
|
375
|
-
isMcpTool: true,
|
|
376
|
-
mcpToolName: 'react_audit_performance',
|
|
377
|
-
mcpEnforced: true,
|
|
378
|
-
mcpFallback: false,
|
|
379
|
-
input: {},
|
|
380
|
-
output,
|
|
381
|
-
duration: 25,
|
|
382
|
-
blocked: false,
|
|
383
|
-
});
|
|
384
|
-
|
|
385
|
-
expect(db.aggregatePendingToolMetrics()).toBe(1);
|
|
386
|
-
expect(db.getToolMetrics(sessionId)).toEqual([
|
|
387
|
-
expect.objectContaining({ status: 'degraded' }),
|
|
388
|
-
]);
|
|
389
|
-
});
|
|
390
|
-
|
|
391
|
-
it('reconciles execution duration from wall-clock start to finish', () => {
|
|
392
|
-
db.insertSkillExecution({
|
|
393
|
-
executionId: 'exec_wall_clock',
|
|
394
|
-
skillName: 'devflow:react',
|
|
395
|
-
startedAt: 1_000,
|
|
396
|
-
status: 'running',
|
|
397
|
-
});
|
|
398
|
-
db.insertToolCallEvent({
|
|
399
|
-
eventId: 'evt_wall_clock',
|
|
400
|
-
executionId: 'exec_wall_clock',
|
|
401
|
-
timestamp: 4_000,
|
|
402
|
-
toolName: 'react_review_hooks',
|
|
403
|
-
toolType: 'mcp',
|
|
404
|
-
isMcpTool: true,
|
|
405
|
-
mcpToolName: 'react_review_hooks',
|
|
406
|
-
mcpEnforced: true,
|
|
407
|
-
mcpFallback: false,
|
|
408
|
-
input: {},
|
|
409
|
-
output: { findings: [] },
|
|
410
|
-
duration: 100,
|
|
411
|
-
blocked: false,
|
|
412
|
-
});
|
|
413
|
-
|
|
414
|
-
db.reconcileSkillExecution('exec_wall_clock', 'completed', 11_000);
|
|
415
|
-
|
|
416
|
-
expect(db.getSkillExecution('exec_wall_clock')).toEqual(
|
|
417
|
-
expect.objectContaining({
|
|
418
|
-
status: 'completed',
|
|
419
|
-
finishedAt: 11_000,
|
|
420
|
-
totalDuration: 10_000,
|
|
421
|
-
}),
|
|
422
|
-
);
|
|
423
|
-
});
|
|
424
|
-
|
|
425
|
-
it('reconciles every running execution owned by a closing session', () => {
|
|
426
|
-
for (const executionId of ['exec_implicit_a', 'exec_implicit_b']) {
|
|
427
|
-
db.insertSkillExecution({
|
|
428
|
-
executionId,
|
|
429
|
-
sessionId: 'session-closing',
|
|
430
|
-
skillName: 'devflow:implicit',
|
|
431
|
-
startedAt: 1_000,
|
|
432
|
-
status: 'running',
|
|
433
|
-
});
|
|
434
|
-
}
|
|
435
|
-
db.insertSkillExecution({
|
|
436
|
-
executionId: 'exec_other_session',
|
|
437
|
-
sessionId: 'session-other',
|
|
438
|
-
skillName: 'devflow:implicit',
|
|
439
|
-
startedAt: 1_000,
|
|
440
|
-
status: 'running',
|
|
441
|
-
});
|
|
442
|
-
|
|
443
|
-
const reconciled = db.reconcileRunningSkillExecutionsForSession(
|
|
444
|
-
'session-closing',
|
|
445
|
-
5_000,
|
|
446
|
-
{ closureReason: 'session_finalize' },
|
|
447
|
-
);
|
|
448
|
-
|
|
449
|
-
expect(reconciled.map(execution => execution.executionId)).toEqual([
|
|
450
|
-
'exec_implicit_a',
|
|
451
|
-
'exec_implicit_b',
|
|
452
|
-
]);
|
|
453
|
-
expect(reconciled.every(execution => execution.status === 'completed')).toBe(true);
|
|
454
|
-
expect(db.getSkillExecution('exec_other_session')).toMatchObject({ status: 'running' });
|
|
455
|
-
});
|
|
456
|
-
});
|