@martinloop/mcp 0.3.2 → 0.3.4
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 +112 -107
- package/dist/discovery-support.js +2 -2
- package/dist/prompts.d.ts +2 -1
- package/dist/resources.d.ts +2 -1
- package/dist/resources.js +147 -147
- package/dist/server-validation.js +5 -2
- package/dist/server.js +5 -4
- package/dist/tools/eval.js +29 -5
- package/dist/tools/execution-policy.d.ts +11 -0
- package/dist/tools/execution-policy.js +47 -0
- package/dist/tools/run-dossier.js +2 -2
- package/dist/tools/run-loop.d.ts +2 -1
- package/dist/tools/run-loop.js +15 -2
- package/dist/tools/run-store.js +166 -6
- package/dist/tools/tool-support.d.ts +3 -2
- package/dist/tools/tool-support.js +112 -18
- package/dist/tools/triage-runs.js +31 -6
- package/dist/vendor/adapters/claude-cli.d.ts +1 -1
- package/dist/vendor/adapters/claude-cli.js +71 -20
- package/dist/vendor/adapters/cli-bridge.d.ts +6 -0
- package/dist/vendor/adapters/cli-bridge.js +91 -7
- package/dist/vendor/adapters/codex-launcher.js +65 -8
- package/dist/vendor/adapters/openai-compatible.d.ts +1 -1
- package/dist/vendor/adapters/openai-compatible.js +7 -7
- package/dist/vendor/adapters/verifier-only.js +1 -1
- package/dist/vendor/contracts/adapter-capabilities.d.ts +17 -0
- package/dist/vendor/contracts/adapter-capabilities.js +35 -0
- package/dist/vendor/contracts/context-graph.d.ts +48 -0
- package/dist/vendor/contracts/context-graph.js +26 -0
- package/dist/vendor/contracts/execution-policy.d.ts +62 -0
- package/dist/vendor/contracts/execution-policy.js +34 -0
- package/dist/vendor/contracts/governance.d.ts +6 -3
- package/dist/vendor/contracts/identity.d.ts +29 -0
- package/dist/vendor/contracts/identity.js +30 -0
- package/dist/vendor/contracts/index.d.ts +13 -2
- package/dist/vendor/contracts/index.js +4 -0
- package/dist/vendor/contracts/observation.d.ts +21 -0
- package/dist/vendor/contracts/observation.js +27 -0
- package/dist/vendor/contracts/trajectory.d.ts +26 -0
- package/dist/vendor/contracts/trajectory.js +27 -0
- package/dist/vendor/contracts/verification.d.ts +28 -0
- package/dist/vendor/contracts/verification.js +7 -0
- package/dist/vendor/core/context-graph.d.ts +3 -0
- package/dist/vendor/core/context-graph.js +371 -0
- package/dist/vendor/core/context-integrity.js +1 -1
- package/dist/vendor/core/grounding.js +1 -1
- package/dist/vendor/core/identity.d.ts +20 -0
- package/dist/vendor/core/identity.js +86 -0
- package/dist/vendor/core/index.d.ts +7 -0
- package/dist/vendor/core/index.js +63 -9
- package/dist/vendor/core/leash.js +1 -1
- package/dist/vendor/core/persistence/integrity.d.ts +1 -0
- package/dist/vendor/core/persistence/integrity.js +7 -3
- package/dist/vendor/core/persistence/runs-reader.d.ts +14 -0
- package/dist/vendor/core/persistence/runs-reader.js +61 -4
- package/dist/vendor/core/policy-compiler.d.ts +2 -0
- package/dist/vendor/core/policy-compiler.js +195 -0
- package/dist/vendor/core/policy.d.ts +2 -0
- package/dist/vendor/core/policy.js +71 -4
- package/dist/vendor/core/rollback.js +32 -12
- package/dist/vendor/core/trajectory.d.ts +18 -0
- package/dist/vendor/core/trajectory.js +295 -0
- package/package.json +76 -75
- package/server.json +21 -21
|
@@ -5,6 +5,10 @@ import { dirname, join } from "node:path";
|
|
|
5
5
|
const RECEIPT_INTEGRITY_SCHEMA_VERSION = "martin.receipt-integrity.v1";
|
|
6
6
|
const RECEIPT_INTEGRITY_KEY_DIR_MODE = 0o700;
|
|
7
7
|
const RECEIPT_INTEGRITY_KEY_FILE_MODE = 0o600;
|
|
8
|
+
export function resolveReceiptIntegrityRoot(env = process.env) {
|
|
9
|
+
return env["MARTIN_INTEGRITY_KEY_DIR"]?.trim() ??
|
|
10
|
+
join(homedir(), ".martin", "receipt-integrity");
|
|
11
|
+
}
|
|
8
12
|
export async function writeReceiptIntegrityMaterial(input) {
|
|
9
13
|
const signedAt = input.signedAt ?? new Date().toISOString();
|
|
10
14
|
const keyMaterial = await ensureReceiptIntegrityKey(input.runsRoot, input.runId);
|
|
@@ -45,9 +49,9 @@ export async function verifyReceiptIntegrityFromFiles(input) {
|
|
|
45
49
|
]);
|
|
46
50
|
if (!rawMaterial || !rawLoopRecord || rawLedger === null || !key) {
|
|
47
51
|
return {
|
|
48
|
-
state: "
|
|
52
|
+
state: "material_missing",
|
|
49
53
|
reason: "Receipt integrity material is missing for this run.",
|
|
50
|
-
warnings: ["
|
|
54
|
+
warnings: ["Receipt integrity material is incomplete; trust claims are unavailable."]
|
|
51
55
|
};
|
|
52
56
|
}
|
|
53
57
|
let material;
|
|
@@ -232,7 +236,7 @@ async function readReceiptIntegrityKey(runsRoot, runId) {
|
|
|
232
236
|
}
|
|
233
237
|
function resolveReceiptIntegrityKeyPath(runsRoot, runId) {
|
|
234
238
|
const rootHash = sha256(runsRoot).slice(0, 16);
|
|
235
|
-
return join(
|
|
239
|
+
return join(resolveReceiptIntegrityRoot(), rootHash, `${runId}.key`);
|
|
236
240
|
}
|
|
237
241
|
function serializeStoredJson(value) {
|
|
238
242
|
return `${JSON.stringify(value, null, 2)}\n`;
|
|
@@ -39,6 +39,19 @@ export interface LoopRunRecord {
|
|
|
39
39
|
objective: string;
|
|
40
40
|
};
|
|
41
41
|
}
|
|
42
|
+
export interface LoopRecordsRollup {
|
|
43
|
+
generatedAt: string;
|
|
44
|
+
totalRuns: number;
|
|
45
|
+
statusBreakdown: Record<string, number>;
|
|
46
|
+
lifecycleBreakdown: Record<string, number>;
|
|
47
|
+
latestByLoopId: Record<string, {
|
|
48
|
+
status: string;
|
|
49
|
+
lifecycleState: string;
|
|
50
|
+
updatedAt: string;
|
|
51
|
+
costUsd: number;
|
|
52
|
+
attempts: number;
|
|
53
|
+
}>;
|
|
54
|
+
}
|
|
42
55
|
export declare function readLoopRecordsFromFile(file: string): Promise<LoopRunRecord[]>;
|
|
43
56
|
export declare function readLatestLoopRecordFromFile(file: string): Promise<LoopRunRecord | null>;
|
|
44
57
|
/**
|
|
@@ -50,3 +63,4 @@ export declare function readAllLoopRecords(runsDir?: string): Promise<LoopRunRec
|
|
|
50
63
|
* Returns the most recently updated loop record, or null if none exist.
|
|
51
64
|
*/
|
|
52
65
|
export declare function readLatestLoopRecord(runsDir?: string): Promise<LoopRunRecord | null>;
|
|
66
|
+
export declare function buildLoopRecordsRollup(records: LoopRunRecord[]): LoopRecordsRollup;
|
|
@@ -45,13 +45,16 @@ export async function readAllLoopRecords(runsDir) {
|
|
|
45
45
|
catch {
|
|
46
46
|
return [];
|
|
47
47
|
}
|
|
48
|
-
const
|
|
48
|
+
const recordsByLoopId = new Map();
|
|
49
49
|
const jsonlFiles = entries
|
|
50
50
|
.filter((entry) => entry.isFile() && entry.name.endsWith(".jsonl"))
|
|
51
51
|
.map((entry) => entry.name);
|
|
52
52
|
for (const file of jsonlFiles) {
|
|
53
53
|
try {
|
|
54
|
-
|
|
54
|
+
const fromFile = await readLoopRecordsFromFile(join(dir, file));
|
|
55
|
+
for (const record of fromFile) {
|
|
56
|
+
ingestRecord(recordsByLoopId, record, "legacy_jsonl");
|
|
57
|
+
}
|
|
55
58
|
}
|
|
56
59
|
catch {
|
|
57
60
|
// skip malformed files or lines
|
|
@@ -60,13 +63,16 @@ export async function readAllLoopRecords(runsDir) {
|
|
|
60
63
|
const runDirectories = entries.filter((entry) => entry.isDirectory());
|
|
61
64
|
for (const entry of runDirectories) {
|
|
62
65
|
try {
|
|
63
|
-
|
|
66
|
+
const canonical = await readLoopRecordsFromFile(join(dir, entry.name, "loop-record.json"));
|
|
67
|
+
for (const record of canonical) {
|
|
68
|
+
ingestRecord(recordsByLoopId, record, "canonical_tree");
|
|
69
|
+
}
|
|
64
70
|
}
|
|
65
71
|
catch {
|
|
66
72
|
// skip missing or malformed canonical records
|
|
67
73
|
}
|
|
68
74
|
}
|
|
69
|
-
return
|
|
75
|
+
return [...recordsByLoopId.values()].map((entry) => entry.record);
|
|
70
76
|
}
|
|
71
77
|
/**
|
|
72
78
|
* Returns the most recently updated loop record, or null if none exist.
|
|
@@ -81,3 +87,54 @@ export async function readLatestLoopRecord(runsDir) {
|
|
|
81
87
|
return a > b ? r : latest;
|
|
82
88
|
}, records[0]);
|
|
83
89
|
}
|
|
90
|
+
export function buildLoopRecordsRollup(records) {
|
|
91
|
+
const statusBreakdown = {};
|
|
92
|
+
const lifecycleBreakdown = {};
|
|
93
|
+
const latestByLoopId = {};
|
|
94
|
+
for (const record of records) {
|
|
95
|
+
statusBreakdown[record.status] = (statusBreakdown[record.status] ?? 0) + 1;
|
|
96
|
+
lifecycleBreakdown[record.lifecycleState] = (lifecycleBreakdown[record.lifecycleState] ?? 0) + 1;
|
|
97
|
+
latestByLoopId[record.loopId] = {
|
|
98
|
+
status: record.status,
|
|
99
|
+
lifecycleState: record.lifecycleState,
|
|
100
|
+
updatedAt: record.updatedAt,
|
|
101
|
+
costUsd: record.cost.actualUsd,
|
|
102
|
+
attempts: record.attempts.length
|
|
103
|
+
};
|
|
104
|
+
}
|
|
105
|
+
return {
|
|
106
|
+
generatedAt: new Date().toISOString(),
|
|
107
|
+
totalRuns: records.length,
|
|
108
|
+
statusBreakdown,
|
|
109
|
+
lifecycleBreakdown,
|
|
110
|
+
latestByLoopId
|
|
111
|
+
};
|
|
112
|
+
}
|
|
113
|
+
function ingestRecord(recordsByLoopId, record, source) {
|
|
114
|
+
const existing = recordsByLoopId.get(record.loopId);
|
|
115
|
+
if (!existing) {
|
|
116
|
+
recordsByLoopId.set(record.loopId, { record, source });
|
|
117
|
+
return;
|
|
118
|
+
}
|
|
119
|
+
const candidateTimestamp = resolveRecordTimestamp(record);
|
|
120
|
+
const existingTimestamp = resolveRecordTimestamp(existing.record);
|
|
121
|
+
if (candidateTimestamp > existingTimestamp) {
|
|
122
|
+
recordsByLoopId.set(record.loopId, { record, source });
|
|
123
|
+
return;
|
|
124
|
+
}
|
|
125
|
+
if (candidateTimestamp === existingTimestamp &&
|
|
126
|
+
sourcePrecedence(source) > sourcePrecedence(existing.source)) {
|
|
127
|
+
recordsByLoopId.set(record.loopId, { record, source });
|
|
128
|
+
}
|
|
129
|
+
}
|
|
130
|
+
function resolveRecordTimestamp(record) {
|
|
131
|
+
const updated = Date.parse(record.updatedAt ?? "");
|
|
132
|
+
if (Number.isFinite(updated)) {
|
|
133
|
+
return updated;
|
|
134
|
+
}
|
|
135
|
+
const created = Date.parse(record.createdAt ?? "");
|
|
136
|
+
return Number.isFinite(created) ? created : 0;
|
|
137
|
+
}
|
|
138
|
+
function sourcePrecedence(source) {
|
|
139
|
+
return source === "canonical_tree" ? 2 : 1;
|
|
140
|
+
}
|
|
@@ -0,0 +1,195 @@
|
|
|
1
|
+
import { cloneExecutionPolicy } from "../contracts/index.js";
|
|
2
|
+
export function compileExecutionPolicy(input) {
|
|
3
|
+
const provenance = [];
|
|
4
|
+
const budget = {
|
|
5
|
+
maxUsd: resolveBudgetField("maxUsd", input, provenance),
|
|
6
|
+
softLimitUsd: resolveBudgetField("softLimitUsd", input, provenance),
|
|
7
|
+
maxIterations: resolveBudgetField("maxIterations", input, provenance),
|
|
8
|
+
maxTokens: resolveBudgetField("maxTokens", input, provenance)
|
|
9
|
+
};
|
|
10
|
+
if (budget.softLimitUsd >= budget.maxUsd) {
|
|
11
|
+
budget.softLimitUsd = Math.round(budget.maxUsd * 0.75 * 100) / 100;
|
|
12
|
+
provenance.push({
|
|
13
|
+
field: "budget.softLimitUsd",
|
|
14
|
+
source: "config",
|
|
15
|
+
value: String(budget.softLimitUsd)
|
|
16
|
+
});
|
|
17
|
+
}
|
|
18
|
+
const governance = {
|
|
19
|
+
policyProfile: resolveEnumField({
|
|
20
|
+
field: "governance.policyProfile",
|
|
21
|
+
defaultsTo: input.defaults.policyProfile,
|
|
22
|
+
configValue: input.config?.policyProfile,
|
|
23
|
+
requestValue: input.request.policyProfile,
|
|
24
|
+
allowed: ["strict", "balanced", "overnight", "debug"],
|
|
25
|
+
provenance
|
|
26
|
+
}),
|
|
27
|
+
telemetryDestination: resolveEnumField({
|
|
28
|
+
field: "governance.telemetryDestination",
|
|
29
|
+
defaultsTo: input.defaults.telemetryDestination,
|
|
30
|
+
configValue: input.config?.governance?.telemetryDestination,
|
|
31
|
+
requestValue: input.request.telemetryDestination,
|
|
32
|
+
allowed: ["local-only", "control-plane"],
|
|
33
|
+
provenance
|
|
34
|
+
}),
|
|
35
|
+
destructiveActionPolicy: resolveEnumField({
|
|
36
|
+
field: "governance.destructiveActionPolicy",
|
|
37
|
+
defaultsTo: input.defaults.destructiveActionPolicy,
|
|
38
|
+
configValue: input.config?.governance?.destructiveActionPolicy,
|
|
39
|
+
allowed: ["never", "approval", "allowed"],
|
|
40
|
+
provenance
|
|
41
|
+
})
|
|
42
|
+
};
|
|
43
|
+
const verificationPlan = resolveVerificationPlan(input, provenance);
|
|
44
|
+
const task = {
|
|
45
|
+
verificationPlan
|
|
46
|
+
};
|
|
47
|
+
if (input.request.mutationMode) {
|
|
48
|
+
task.mutationMode = input.request.mutationMode;
|
|
49
|
+
provenance.push({
|
|
50
|
+
field: "task.mutationMode",
|
|
51
|
+
source: "request",
|
|
52
|
+
value: input.request.mutationMode
|
|
53
|
+
});
|
|
54
|
+
}
|
|
55
|
+
if (input.request.repoRoot) {
|
|
56
|
+
task.repoRoot = input.request.repoRoot;
|
|
57
|
+
provenance.push({
|
|
58
|
+
field: "task.repoRoot",
|
|
59
|
+
source: "request",
|
|
60
|
+
value: input.request.repoRoot
|
|
61
|
+
});
|
|
62
|
+
}
|
|
63
|
+
if ((input.request.allowedPaths?.length ?? 0) > 0) {
|
|
64
|
+
const allowedPaths = input.request.allowedPaths ?? [];
|
|
65
|
+
task.allowedPaths = [...allowedPaths];
|
|
66
|
+
provenance.push({
|
|
67
|
+
field: "task.allowedPaths",
|
|
68
|
+
source: "request",
|
|
69
|
+
value: allowedPaths.join(", ")
|
|
70
|
+
});
|
|
71
|
+
}
|
|
72
|
+
if ((input.request.deniedPaths?.length ?? 0) > 0) {
|
|
73
|
+
const deniedPaths = input.request.deniedPaths ?? [];
|
|
74
|
+
task.deniedPaths = [...deniedPaths];
|
|
75
|
+
provenance.push({
|
|
76
|
+
field: "task.deniedPaths",
|
|
77
|
+
source: "request",
|
|
78
|
+
value: deniedPaths.join(", ")
|
|
79
|
+
});
|
|
80
|
+
}
|
|
81
|
+
if ((input.request.acceptanceCriteria?.length ?? 0) > 0) {
|
|
82
|
+
const acceptanceCriteria = input.request.acceptanceCriteria ?? [];
|
|
83
|
+
task.acceptanceCriteria = [...acceptanceCriteria];
|
|
84
|
+
provenance.push({
|
|
85
|
+
field: "task.acceptanceCriteria",
|
|
86
|
+
source: "request",
|
|
87
|
+
value: acceptanceCriteria.join(", ")
|
|
88
|
+
});
|
|
89
|
+
}
|
|
90
|
+
if (input.request.approvalPolicy) {
|
|
91
|
+
task.approvalPolicy = { ...input.request.approvalPolicy };
|
|
92
|
+
provenance.push({
|
|
93
|
+
field: "task.approvalPolicy",
|
|
94
|
+
source: "request",
|
|
95
|
+
value: JSON.stringify(input.request.approvalPolicy)
|
|
96
|
+
});
|
|
97
|
+
}
|
|
98
|
+
return cloneExecutionPolicy({
|
|
99
|
+
configPath: input.configPath,
|
|
100
|
+
budget,
|
|
101
|
+
governance,
|
|
102
|
+
task,
|
|
103
|
+
provenance
|
|
104
|
+
});
|
|
105
|
+
}
|
|
106
|
+
function resolveBudgetField(field, input, provenance) {
|
|
107
|
+
if (input.request.budgetOverrides?.[field]) {
|
|
108
|
+
const value = input.request.budget[field];
|
|
109
|
+
provenance.push({
|
|
110
|
+
field: `budget.${String(field)}`,
|
|
111
|
+
source: "request",
|
|
112
|
+
value: String(value)
|
|
113
|
+
});
|
|
114
|
+
return value;
|
|
115
|
+
}
|
|
116
|
+
const configValue = input.config?.budget?.[field];
|
|
117
|
+
if (typeof configValue === "number" && Number.isFinite(configValue)) {
|
|
118
|
+
provenance.push({
|
|
119
|
+
field: `budget.${String(field)}`,
|
|
120
|
+
source: "config",
|
|
121
|
+
value: String(configValue)
|
|
122
|
+
});
|
|
123
|
+
return configValue;
|
|
124
|
+
}
|
|
125
|
+
const value = input.defaults.budget[field];
|
|
126
|
+
provenance.push({
|
|
127
|
+
field: `budget.${String(field)}`,
|
|
128
|
+
source: "default",
|
|
129
|
+
value: String(value)
|
|
130
|
+
});
|
|
131
|
+
return value;
|
|
132
|
+
}
|
|
133
|
+
function resolveVerificationPlan(input, provenance) {
|
|
134
|
+
const requestVerificationPlan = normalizeStringList(input.request.verificationPlan);
|
|
135
|
+
if (requestVerificationPlan.length > 0) {
|
|
136
|
+
provenance.push({
|
|
137
|
+
field: "task.verificationPlan",
|
|
138
|
+
source: "request",
|
|
139
|
+
value: requestVerificationPlan.join(", ")
|
|
140
|
+
});
|
|
141
|
+
return requestVerificationPlan;
|
|
142
|
+
}
|
|
143
|
+
const configVerificationPlan = normalizeStringList(input.config?.governance?.verifierRules);
|
|
144
|
+
if (configVerificationPlan.length > 0) {
|
|
145
|
+
provenance.push({
|
|
146
|
+
field: "task.verificationPlan",
|
|
147
|
+
source: "config",
|
|
148
|
+
value: configVerificationPlan.join(", ")
|
|
149
|
+
});
|
|
150
|
+
return configVerificationPlan;
|
|
151
|
+
}
|
|
152
|
+
const defaultVerificationPlan = normalizeStringList(input.defaults.verifierRules);
|
|
153
|
+
provenance.push({
|
|
154
|
+
field: "task.verificationPlan",
|
|
155
|
+
source: "default",
|
|
156
|
+
value: defaultVerificationPlan.join(", ")
|
|
157
|
+
});
|
|
158
|
+
return defaultVerificationPlan;
|
|
159
|
+
}
|
|
160
|
+
function resolveEnumField(input) {
|
|
161
|
+
const normalizedRequest = normalizeEnumCandidate(input.requestValue, input.allowed);
|
|
162
|
+
if (normalizedRequest) {
|
|
163
|
+
input.provenance.push({
|
|
164
|
+
field: input.field,
|
|
165
|
+
source: "request",
|
|
166
|
+
value: normalizedRequest
|
|
167
|
+
});
|
|
168
|
+
return normalizedRequest;
|
|
169
|
+
}
|
|
170
|
+
const normalizedConfig = normalizeEnumCandidate(input.configValue, input.allowed);
|
|
171
|
+
if (normalizedConfig) {
|
|
172
|
+
input.provenance.push({
|
|
173
|
+
field: input.field,
|
|
174
|
+
source: "config",
|
|
175
|
+
value: normalizedConfig
|
|
176
|
+
});
|
|
177
|
+
return normalizedConfig;
|
|
178
|
+
}
|
|
179
|
+
input.provenance.push({
|
|
180
|
+
field: input.field,
|
|
181
|
+
source: "default",
|
|
182
|
+
value: input.defaultsTo
|
|
183
|
+
});
|
|
184
|
+
return input.defaultsTo;
|
|
185
|
+
}
|
|
186
|
+
function normalizeEnumCandidate(value, allowed) {
|
|
187
|
+
if (typeof value !== "string") {
|
|
188
|
+
return undefined;
|
|
189
|
+
}
|
|
190
|
+
const trimmed = value.trim();
|
|
191
|
+
return allowed.find((candidate) => candidate === trimmed);
|
|
192
|
+
}
|
|
193
|
+
function normalizeStringList(values) {
|
|
194
|
+
return (values ?? []).map((value) => value.trim()).filter(Boolean);
|
|
195
|
+
}
|
|
@@ -4,6 +4,8 @@ export interface FailureAssessment {
|
|
|
4
4
|
rationale: string;
|
|
5
5
|
retryable: boolean;
|
|
6
6
|
recommendedIntervention: InterventionType;
|
|
7
|
+
/** Actionable diagnosis injected into the next attempt's prompt context. */
|
|
8
|
+
diagnosticHint?: string;
|
|
7
9
|
}
|
|
8
10
|
export interface CostGovernorState {
|
|
9
11
|
pressure: "healthy" | "soft_limit" | "hard_limit";
|
|
@@ -78,7 +78,8 @@ export function classifyFailure(input) {
|
|
|
78
78
|
failureClass: "environment_mismatch",
|
|
79
79
|
rationale: "The adapter could not access the required local runtime or CLI tooling.",
|
|
80
80
|
retryable: false,
|
|
81
|
-
recommendedIntervention: "switch_adapter"
|
|
81
|
+
recommendedIntervention: "switch_adapter",
|
|
82
|
+
diagnosticHint: diagnoseEnvironmentIssue(message)
|
|
82
83
|
};
|
|
83
84
|
}
|
|
84
85
|
if (containsPositive(message, ["syntax error", "parse error", "typescript error"])) {
|
|
@@ -86,7 +87,8 @@ export function classifyFailure(input) {
|
|
|
86
87
|
failureClass: "syntax_error",
|
|
87
88
|
rationale: "The attempt failed on a parser or compiler-style issue.",
|
|
88
89
|
retryable: true,
|
|
89
|
-
recommendedIntervention: "compress_context"
|
|
90
|
+
recommendedIntervention: "compress_context",
|
|
91
|
+
diagnosticHint: "Check for malformed edits — the verifier found a syntax or parse error in the changed files."
|
|
90
92
|
};
|
|
91
93
|
}
|
|
92
94
|
if (containsPositive(message, [
|
|
@@ -104,7 +106,8 @@ export function classifyFailure(input) {
|
|
|
104
106
|
failureClass: "repo_grounding_failure",
|
|
105
107
|
rationale: "The attempt referenced modules, symbols, or files that are not grounded in the repo or approved docs.",
|
|
106
108
|
retryable: true,
|
|
107
|
-
recommendedIntervention: repeatedFailure === "repo_grounding_failure" ? "tighten_task" : "run_verifier"
|
|
109
|
+
recommendedIntervention: repeatedFailure === "repo_grounding_failure" ? "tighten_task" : "run_verifier",
|
|
110
|
+
diagnosticHint: diagnoseGroundingIssue(message)
|
|
108
111
|
};
|
|
109
112
|
}
|
|
110
113
|
if (containsPositive(message, [
|
|
@@ -118,7 +121,8 @@ export function classifyFailure(input) {
|
|
|
118
121
|
failureClass: "verification_failure",
|
|
119
122
|
rationale: "The proposed fix did not satisfy the verification gate.",
|
|
120
123
|
retryable: true,
|
|
121
|
-
recommendedIntervention: repeatedFailure === "verification_failure" ? "tighten_task" : "run_verifier"
|
|
124
|
+
recommendedIntervention: repeatedFailure === "verification_failure" ? "tighten_task" : "run_verifier",
|
|
125
|
+
diagnosticHint: diagnoseVerificationIssue(message)
|
|
122
126
|
};
|
|
123
127
|
}
|
|
124
128
|
if (containsPositive(message, [
|
|
@@ -622,3 +626,66 @@ function computeDiffRiskScore(input) {
|
|
|
622
626
|
function roundScore(value) {
|
|
623
627
|
return Math.round(value * 100) / 100;
|
|
624
628
|
}
|
|
629
|
+
// ---------------------------------------------------------------------------
|
|
630
|
+
// Diagnostic hint generators
|
|
631
|
+
//
|
|
632
|
+
// These extract actionable context from failure messages so the next attempt
|
|
633
|
+
// gets a concrete diagnosis, not just "verification failed."
|
|
634
|
+
// ---------------------------------------------------------------------------
|
|
635
|
+
const KNOWN_TOOLS = {
|
|
636
|
+
bun: "npm install -g bun",
|
|
637
|
+
pnpm: "npm install -g pnpm",
|
|
638
|
+
yarn: "npm install -g yarn",
|
|
639
|
+
eslint: "npm install eslint",
|
|
640
|
+
vitest: "npm install vitest",
|
|
641
|
+
jest: "npm install jest",
|
|
642
|
+
tsc: "npm install typescript",
|
|
643
|
+
prettier: "npm install prettier",
|
|
644
|
+
biome: "npm install @biomejs/biome"
|
|
645
|
+
};
|
|
646
|
+
function diagnoseEnvironmentIssue(message) {
|
|
647
|
+
// Try to extract the missing command name
|
|
648
|
+
const cmdMatch = message.match(/(?:command not found|enoent)[:\s]*['"]?(\w[\w.-]*)/i)
|
|
649
|
+
?? message.match(/['"](\w[\w.-]*)['"].*(?:not found|enoent)/i);
|
|
650
|
+
const missingCmd = cmdMatch?.[1];
|
|
651
|
+
if (missingCmd && KNOWN_TOOLS[missingCmd.toLowerCase()]) {
|
|
652
|
+
return `The verifier needs "${missingCmd}" which is not installed. Fix: ${KNOWN_TOOLS[missingCmd.toLowerCase()]}`;
|
|
653
|
+
}
|
|
654
|
+
if (missingCmd) {
|
|
655
|
+
return `The verifier could not find "${missingCmd}". Ensure it is installed in the workspace before retrying.`;
|
|
656
|
+
}
|
|
657
|
+
return "A required tool or command was not found in the environment. Check that all dependencies referenced by the verify command are installed.";
|
|
658
|
+
}
|
|
659
|
+
function diagnoseGroundingIssue(message) {
|
|
660
|
+
// Extract the missing module/symbol name
|
|
661
|
+
const moduleMatch = message.match(/cannot find module ['"]([^'"]+)['"]/i)
|
|
662
|
+
?? message.match(/module not found[:\s]*['"]?([^\s'"]+)/i);
|
|
663
|
+
const nameMatch = message.match(/cannot find name ['"]?(\w+)/i);
|
|
664
|
+
if (moduleMatch?.[1]) {
|
|
665
|
+
const mod = moduleMatch[1];
|
|
666
|
+
const isRelative = mod.startsWith(".") || mod.startsWith("/");
|
|
667
|
+
if (isRelative) {
|
|
668
|
+
return `Import "${mod}" points to a file that does not exist. Check the path and filename.`;
|
|
669
|
+
}
|
|
670
|
+
return `Module "${mod}" is not installed. Run npm install or check package.json dependencies.`;
|
|
671
|
+
}
|
|
672
|
+
if (nameMatch?.[1]) {
|
|
673
|
+
return `Symbol "${nameMatch[1]}" is not defined. Check that the import exists and the name is spelled correctly.`;
|
|
674
|
+
}
|
|
675
|
+
return "The code references modules or symbols that do not exist in the repository. Verify imports and file paths.";
|
|
676
|
+
}
|
|
677
|
+
function diagnoseVerificationIssue(message) {
|
|
678
|
+
// Count assertion failures for severity signal
|
|
679
|
+
const assertionCount = (message.match(/assertionerror/gi) ?? []).length;
|
|
680
|
+
if (assertionCount > 3) {
|
|
681
|
+
return `${String(assertionCount)} assertion failures detected — the change may have introduced a regression. Focus on the specific test expectations that broke.`;
|
|
682
|
+
}
|
|
683
|
+
// Check for common patterns
|
|
684
|
+
if (/timeout/i.test(message) && !/budget/i.test(message)) {
|
|
685
|
+
return "Tests timed out. The change may have introduced an infinite loop or unresolved async operation.";
|
|
686
|
+
}
|
|
687
|
+
if (/exit code (?!0)\d+/i.test(message) || /exited with code/i.test(message)) {
|
|
688
|
+
return "The verifier command exited non-zero. Read the output above to identify which specific check failed and why.";
|
|
689
|
+
}
|
|
690
|
+
return "Verification did not pass. Review the verifier output for the specific failing assertion or check.";
|
|
691
|
+
}
|
|
@@ -119,17 +119,24 @@ function readRepoState(repoRoot) {
|
|
|
119
119
|
};
|
|
120
120
|
}
|
|
121
121
|
function readGitLines(repoRoot, args) {
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
|
|
122
|
+
for (let attempt = 0; attempt < 2; attempt++) {
|
|
123
|
+
const result = spawnSync("git", args, {
|
|
124
|
+
cwd: repoRoot,
|
|
125
|
+
encoding: "utf8"
|
|
126
|
+
});
|
|
127
|
+
if (result.status === 0 && typeof result.stdout === "string") {
|
|
128
|
+
return uniqueSorted(result.stdout
|
|
129
|
+
.split(/\r?\n/u)
|
|
130
|
+
.map((line) => normalizeRepoPath(line))
|
|
131
|
+
.filter(Boolean));
|
|
132
|
+
}
|
|
133
|
+
// Retry once after a short delay — handles git lock-file contention
|
|
134
|
+
// that occurs when another process holds .git/index.lock.
|
|
135
|
+
if (attempt === 0 && result.status !== 0) {
|
|
136
|
+
spawnSync("sleep", ["0.5"], { encoding: "utf8" });
|
|
137
|
+
}
|
|
128
138
|
}
|
|
129
|
-
return
|
|
130
|
-
.split(/\r?\n/u)
|
|
131
|
-
.map((line) => normalizeRepoPath(line))
|
|
132
|
-
.filter(Boolean));
|
|
139
|
+
return [];
|
|
133
140
|
}
|
|
134
141
|
function readGitScalar(repoRoot, args) {
|
|
135
142
|
const result = spawnSync("git", args, {
|
|
@@ -175,9 +182,22 @@ function restoreTrackedFileFromHead(repoRoot, filePath) {
|
|
|
175
182
|
cwd: repoRoot,
|
|
176
183
|
encoding: "utf8"
|
|
177
184
|
});
|
|
178
|
-
if (result.status
|
|
179
|
-
|
|
185
|
+
if (result.status === 0) {
|
|
186
|
+
return;
|
|
187
|
+
}
|
|
188
|
+
// Retry once after a short delay (git lock-file contention)
|
|
189
|
+
spawnSync("sleep", ["0.5"], { encoding: "utf8" });
|
|
190
|
+
const retry = spawnSync("git", ["restore", "--staged", "--worktree", "--source=HEAD", "--", filePath], { cwd: repoRoot, encoding: "utf8" });
|
|
191
|
+
if (retry.status === 0) {
|
|
192
|
+
return;
|
|
193
|
+
}
|
|
194
|
+
// Fallback: try git checkout which works on older git versions
|
|
195
|
+
const fallback = spawnSync("git", ["checkout", "HEAD", "--", filePath], { cwd: repoRoot, encoding: "utf8" });
|
|
196
|
+
if (fallback.status === 0) {
|
|
197
|
+
return;
|
|
180
198
|
}
|
|
199
|
+
throw new Error(result.stderr?.trim() ||
|
|
200
|
+
`git restore failed for ${filePath} (retry and checkout fallback also failed)`);
|
|
181
201
|
}
|
|
182
202
|
async function removeRepoPath(repoRoot, filePath) {
|
|
183
203
|
await rm(resolveRepoPath(repoRoot, filePath), { recursive: true, force: true });
|
|
@@ -0,0 +1,18 @@
|
|
|
1
|
+
import { type CircuitBreakDecision, type TrajectoryAssessment } from "../contracts/index.js";
|
|
2
|
+
export declare function assessTrajectory(input: {
|
|
3
|
+
objective: string;
|
|
4
|
+
verificationPlan?: string[];
|
|
5
|
+
attempts: TrajectoryAttemptLike[];
|
|
6
|
+
}): TrajectoryAssessment;
|
|
7
|
+
export declare function decideCircuitBreak(input: {
|
|
8
|
+
objective: string;
|
|
9
|
+
verificationPlan?: string[];
|
|
10
|
+
attempts: TrajectoryAttemptLike[];
|
|
11
|
+
remainingIterations?: number;
|
|
12
|
+
}): CircuitBreakDecision;
|
|
13
|
+
interface TrajectoryAttemptLike {
|
|
14
|
+
index: number;
|
|
15
|
+
summary?: string;
|
|
16
|
+
failureClass?: string;
|
|
17
|
+
}
|
|
18
|
+
export {};
|