@arnilo/prism 0.8.0 → 0.10.0
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/CHANGELOG.md +62 -1
- package/README.md +13 -12
- package/dist/agent-approval.d.ts +17 -2
- package/dist/agent-approval.js +15 -6
- package/dist/agent-event-source.d.ts +9 -1
- package/dist/agent-event-source.js +10 -3
- package/dist/agent-loops.js +7 -4
- package/dist/agent-run-lifecycle.d.ts +15 -1
- package/dist/agent-run-lifecycle.js +82 -11
- package/dist/agent-run-state.d.ts +47 -6
- package/dist/agent-run-state.js +154 -6
- package/dist/agent-session/event-subscriber.d.ts +2 -0
- package/dist/agent-session/event-subscriber.js +3 -0
- package/dist/agent-session/helpers.js +14 -0
- package/dist/agent-session/session/assemble.js +281 -32
- package/dist/agent-session/session/persist.d.ts +11 -0
- package/dist/agent-session/session/persist.js +48 -16
- package/dist/agent-session/session/provider-round.d.ts +14 -4
- package/dist/agent-session/session/provider-round.js +226 -19
- package/dist/agent-session/session/tool-round.d.ts +2 -2
- package/dist/agent-session/session/tool-round.js +78 -6
- package/dist/agent-session/session/types.d.ts +44 -3
- package/dist/agent-session/session.d.ts +100 -5
- package/dist/agent-session/session.js +224 -13
- package/dist/attention-compiler.d.ts +51 -2
- package/dist/attention-compiler.js +282 -21
- package/dist/cache-helpers.d.ts +4 -2
- package/dist/cache-helpers.js +8 -6
- package/dist/checkpoint-restore.d.ts +45 -0
- package/dist/checkpoint-restore.js +54 -0
- package/dist/context-budget.d.ts +13 -1
- package/dist/context-budget.js +57 -4
- package/dist/contracts-core/agent.d.ts +52 -1
- package/dist/contracts-core/attention.d.ts +95 -0
- package/dist/contracts-core/content.d.ts +10 -0
- package/dist/contracts-core/extensions.d.ts +3 -0
- package/dist/contracts-core/guardrail-packs.d.ts +46 -0
- package/dist/contracts-core/guardrail-packs.js +2 -0
- package/dist/contracts-core/loop.d.ts +36 -0
- package/dist/contracts-core/provider.d.ts +30 -0
- package/dist/contracts-core/run-limits.d.ts +29 -1
- package/dist/contracts-core/session.d.ts +23 -5
- package/dist/contracts-core/session.js +21 -2
- package/dist/contracts-core/usage.d.ts +40 -0
- package/dist/contracts-core/usage.js +8 -0
- package/dist/contracts-core.d.ts +2 -0
- package/dist/contracts-core.js +2 -0
- package/dist/contracts-protocol.d.ts +81 -5
- package/dist/contracts-run-state.d.ts +91 -2
- package/dist/contributions.d.ts +2 -1
- package/dist/contributions.js +1 -0
- package/dist/extensions.d.ts +15 -1
- package/dist/extensions.js +68 -0
- package/dist/guardrail-packs/coding-standard.d.ts +3 -0
- package/dist/guardrail-packs/coding-standard.js +63 -0
- package/dist/guardrail-packs/destructive-commands.d.ts +3 -0
- package/dist/guardrail-packs/destructive-commands.js +46 -0
- package/dist/guardrail-packs/errors.d.ts +7 -0
- package/dist/guardrail-packs/errors.js +9 -0
- package/dist/guardrail-packs/index.d.ts +4 -0
- package/dist/guardrail-packs/index.js +15 -0
- package/dist/guardrail-packs/secrets-hygiene.d.ts +3 -0
- package/dist/guardrail-packs/secrets-hygiene.js +23 -0
- package/dist/guardrail-packs/types.d.ts +26 -0
- package/dist/guardrail-packs/types.js +2 -0
- package/dist/guardrail-packs/validation-respect.d.ts +3 -0
- package/dist/guardrail-packs/validation-respect.js +69 -0
- package/dist/guardrails.d.ts +61 -1
- package/dist/guardrails.js +377 -0
- package/dist/index.d.ts +16 -11
- package/dist/index.js +10 -7
- package/dist/input.d.ts +8 -1
- package/dist/input.js +68 -6
- package/dist/middleware.d.ts +37 -2
- package/dist/middleware.js +41 -0
- package/dist/node/session-store-jsonl.js +18 -3
- package/dist/observability.js +6 -0
- package/dist/provider-events.d.ts +8 -2
- package/dist/provider-events.js +60 -2
- package/dist/providers/openai-compatible.js +6 -3
- package/dist/run-bundle.d.ts +6 -1
- package/dist/run-bundle.js +5 -1
- package/dist/run-limits.d.ts +11 -1
- package/dist/run-limits.js +59 -0
- package/dist/session-stores.d.ts +12 -1
- package/dist/session-stores.js +21 -4
- package/dist/testing/agent-event-source-conformance.js +41 -2
- package/dist/testing/prefix-stability-conformance.d.ts +59 -0
- package/dist/testing/prefix-stability-conformance.js +172 -0
- package/dist/testing/session-store-conformance.d.ts +3 -2
- package/dist/testing/session-store-conformance.js +48 -0
- package/dist/tools.d.ts +5 -0
- package/dist/tools.js +21 -6
- package/dist/usage-estimation.d.ts +29 -0
- package/dist/usage-estimation.js +79 -0
- package/docs/agent-events.md +75 -4
- package/docs/agent-session-runtime.md +10 -6
- package/docs/attention-compiler.md +89 -8
- package/docs/caveman.md +1 -1
- package/docs/coding-agent-tools.md +1 -1
- package/docs/compaction-and-retry.md +1 -1
- package/docs/compaction-llm.md +2 -0
- package/docs/compaction-observational-memory.md +54 -7
- package/docs/durable-runs.md +46 -3
- package/docs/embeddings.md +9 -0
- package/docs/evaluations.md +5 -0
- package/docs/execution-timeline.md +79 -1
- package/docs/extensions.md +20 -3
- package/docs/guardrails.md +50 -4
- package/docs/hooks.md +282 -0
- package/docs/index.md +37 -15
- package/docs/input-and-prompt-assembly.md +4 -4
- package/docs/instruction-injection.md +1 -0
- package/docs/knowledge-sync.md +4 -0
- package/docs/live-testing.md +3 -1
- package/docs/memory-fabric.md +28 -0
- package/docs/middleware-hooks.md +90 -4
- package/docs/migrate-to-0.9.md +210 -0
- package/docs/migration.md +26 -0
- package/docs/multi-agent-patterns.md +25 -2
- package/docs/node-jsonl-session-store.md +7 -1
- package/docs/observability.md +7 -3
- package/docs/options-index.md +4 -1
- package/docs/policy-and-audit.md +26 -1
- package/docs/prefix-stability-conformance.md +143 -0
- package/docs/provider-caching.md +4 -4
- package/docs/provider-conformance.md +16 -0
- package/docs/provider-packages.md +20 -20
- package/docs/public-contracts.md +3 -2
- package/docs/rag.md +188 -3
- package/docs/release-and-install.md +45 -40
- package/docs/runs-and-usage.md +56 -10
- package/docs/scoped-agent-memory.md +270 -0
- package/docs/scoped-memory.md +138 -0
- package/docs/session-store-conformance.md +1 -2
- package/docs/session-stores.md +17 -17
- package/docs/supervisors.md +32 -12
- package/docs/tools.md +18 -1
- package/docs/wiki.md +4 -2
- package/docs/workflows.md +5 -0
- package/package.json +8 -2
package/dist/input.js
CHANGED
|
@@ -5,7 +5,7 @@ import { runInstructionInjectors } from "./instruction-injection.js";
|
|
|
5
5
|
import { applyDefaultProviderRequestOptions } from "./provider-request-policy.js";
|
|
6
6
|
import { redactMessage } from "./redaction.js";
|
|
7
7
|
import { loadTextResource } from "./resources.js";
|
|
8
|
-
import { skillMessages as buildSkillMessages } from "./skill-disclosure.js";
|
|
8
|
+
import { capSkillCatalog, selectSkillsForPrompt, skillHasRenderableBody, skillMessages as buildSkillMessages, } from "./skill-disclosure.js";
|
|
9
9
|
import { composeSystemPrompt } from "./system-prompts.js";
|
|
10
10
|
import { foldToolResultHistory, foldToolResults } from "./tool-result-fold.js";
|
|
11
11
|
import { selectDisclosedTools } from "./tool-search.js";
|
|
@@ -51,11 +51,16 @@ export function createDefaultPromptBuilder() {
|
|
|
51
51
|
// text-only (or unknown-capability) models — duplicating it doubles tool tokens per turn.
|
|
52
52
|
const tools = request.model?.capabilities?.tools === true ? undefined : request.tools;
|
|
53
53
|
const context = contextMessages(request.context);
|
|
54
|
-
const skills = buildSkillMessages(request.skills,
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
54
|
+
const skills = buildSkillMessages(request.skills, request.tailSkillBodies
|
|
55
|
+
? {
|
|
56
|
+
disclosure: "progressive",
|
|
57
|
+
demotedBodies: request.demotedSkillBodies?.length ? new Set(request.demotedSkillBodies) : undefined,
|
|
58
|
+
}
|
|
59
|
+
: {
|
|
60
|
+
disclosure: request.skillsDisclosure,
|
|
61
|
+
loaded: request.loadedSkills,
|
|
62
|
+
demotedBodies: request.demotedSkillBodies?.length ? new Set(request.demotedSkillBodies) : undefined,
|
|
63
|
+
});
|
|
59
64
|
const declarations = toolMessages(tools);
|
|
60
65
|
if ((request.inputLayout ?? "cache_aware") === "legacy") {
|
|
61
66
|
return [...context, ...skills, ...declarations, ...request.messages];
|
|
@@ -156,9 +161,11 @@ export async function assembleProviderInput(options) {
|
|
|
156
161
|
tools,
|
|
157
162
|
fold: options.toolResultFold,
|
|
158
163
|
frontier: options.attentionSticky,
|
|
164
|
+
attentionFold: options.attentionFold,
|
|
159
165
|
redactor: options.redactor,
|
|
160
166
|
signal: options.signal,
|
|
161
167
|
turn,
|
|
168
|
+
runInputTokens: options.runInputTokens,
|
|
162
169
|
sessionId: options.sessionId,
|
|
163
170
|
runId: options.runId,
|
|
164
171
|
});
|
|
@@ -213,6 +220,14 @@ export async function assembleProviderInput(options) {
|
|
|
213
220
|
...baseContext,
|
|
214
221
|
});
|
|
215
222
|
}
|
|
223
|
+
const tailSegments = options.tailSegments;
|
|
224
|
+
const tailSkillBodies = tailSegments !== undefined;
|
|
225
|
+
if (tailSegments) {
|
|
226
|
+
const activeTailSegments = new Set();
|
|
227
|
+
messages = moveResourceMessagesToTail(messages, options, tailSegments, activeTailSegments);
|
|
228
|
+
appendSkillTailSegments(tailSegments, activeTailSegments, skills, options.skillsDisclosure, options.loadedSkills, demotedSkillBodies);
|
|
229
|
+
messages = [...messages, ...tailMessages(tailSegments, activeTailSegments)];
|
|
230
|
+
}
|
|
216
231
|
const promptBuilder = options.promptBuilder ?? createDefaultPromptBuilder();
|
|
217
232
|
const promptRequest = options.middleware
|
|
218
233
|
? await options.middleware.run("prompt_build", {
|
|
@@ -222,6 +237,7 @@ export async function assembleProviderInput(options) {
|
|
|
222
237
|
skills,
|
|
223
238
|
skillsDisclosure: options.skillsDisclosure,
|
|
224
239
|
loadedSkills: options.loadedSkills,
|
|
240
|
+
tailSkillBodies,
|
|
225
241
|
demotedSkillBodies,
|
|
226
242
|
tools,
|
|
227
243
|
metadata: options.metadata,
|
|
@@ -234,6 +250,7 @@ export async function assembleProviderInput(options) {
|
|
|
234
250
|
skills,
|
|
235
251
|
skillsDisclosure: options.skillsDisclosure,
|
|
236
252
|
loadedSkills: options.loadedSkills,
|
|
253
|
+
tailSkillBodies,
|
|
237
254
|
demotedSkillBodies,
|
|
238
255
|
tools,
|
|
239
256
|
metadata: options.metadata,
|
|
@@ -389,6 +406,51 @@ function textMessage(role, text, metadata) {
|
|
|
389
406
|
function contextMessages(context) {
|
|
390
407
|
return (context ?? []).map((block) => textMessage("system", `${block.title ? `${block.title}:\n` : "Context:\n"}${blockText(block)}`, block.metadata));
|
|
391
408
|
}
|
|
409
|
+
function moveResourceMessagesToTail(messages, options, tailSegments, activeTailSegments) {
|
|
410
|
+
const resourceUris = new Set([
|
|
411
|
+
...(options.resourceUris ?? []),
|
|
412
|
+
...(options.attachments ?? []).flatMap((attachment) => attachment.uri !== undefined && attachment.text === undefined && attachment.content === undefined ? [attachment.uri] : []),
|
|
413
|
+
]);
|
|
414
|
+
if (resourceUris.size === 0)
|
|
415
|
+
return messages;
|
|
416
|
+
return messages.filter((message) => {
|
|
417
|
+
const uri = message.metadata?.uri;
|
|
418
|
+
if (message.role !== "user" || typeof uri !== "string" || !resourceUris.has(uri))
|
|
419
|
+
return true;
|
|
420
|
+
appendTailSegment(tailSegments, activeTailSegments, `resource:${uri}`, message);
|
|
421
|
+
return false;
|
|
422
|
+
});
|
|
423
|
+
}
|
|
424
|
+
function appendSkillTailSegments(tailSegments, activeTailSegments, skills, disclosure, loaded, demotedSkillBodies) {
|
|
425
|
+
const renderContext = {
|
|
426
|
+
disclosure,
|
|
427
|
+
loaded,
|
|
428
|
+
demotedBodies: demotedSkillBodies?.length ? new Set(demotedSkillBodies) : undefined,
|
|
429
|
+
};
|
|
430
|
+
const selected = capSkillCatalog(selectSkillsForPrompt(skills ?? [], renderContext));
|
|
431
|
+
const byName = new Map(selected.map((skill) => [skill.name, skill]));
|
|
432
|
+
const ordered = disclosure === "eager"
|
|
433
|
+
? selected
|
|
434
|
+
: (loaded?.list() ?? []).flatMap((name) => {
|
|
435
|
+
const skill = byName.get(name);
|
|
436
|
+
return skill === undefined ? [] : [skill];
|
|
437
|
+
});
|
|
438
|
+
for (const skill of ordered) {
|
|
439
|
+
if (!skillHasRenderableBody(skill, renderContext))
|
|
440
|
+
continue;
|
|
441
|
+
const message = buildSkillMessages([skill], renderContext)[0];
|
|
442
|
+
if (message)
|
|
443
|
+
appendTailSegment(tailSegments, activeTailSegments, `skill:${skill.name}`, message);
|
|
444
|
+
}
|
|
445
|
+
}
|
|
446
|
+
/** Map#set retains first-insertion order, so re-derivation replaces only this segment's bytes. */
|
|
447
|
+
function appendTailSegment(tailSegments, activeTailSegments, id, message) {
|
|
448
|
+
tailSegments.set(id, message);
|
|
449
|
+
activeTailSegments.add(id);
|
|
450
|
+
}
|
|
451
|
+
function tailMessages(tailSegments, activeTailSegments) {
|
|
452
|
+
return [...tailSegments].flatMap(([id, message]) => (activeTailSegments.has(id) ? [message] : []));
|
|
453
|
+
}
|
|
392
454
|
function toolMessages(tools) {
|
|
393
455
|
if (!tools?.length)
|
|
394
456
|
return [];
|
package/dist/middleware.d.ts
CHANGED
|
@@ -1,5 +1,40 @@
|
|
|
1
|
-
import type { ExtensionEvent } from "./contracts.js";
|
|
2
|
-
export type MiddlewareHookName = "provider_request" | "input_assembly" | "prompt_build" | "context" | "tool_call" | "tool_result" | "retry" | "compaction" | "session_start" | "session_shutdown";
|
|
1
|
+
import type { ContentBlock, ExtensionEvent } from "./contracts.js";
|
|
2
|
+
export type MiddlewareHookName = "beforeProviderTurn" | "provider_request" | "input_assembly" | "prompt_build" | "context" | "tool_call" | "tool_result" | "retry" | "compaction_request" | "compaction" | "session_start" | "session_shutdown";
|
|
3
|
+
/** Provenance of a host-answered turn (plan 096). An id, never free host code. */
|
|
4
|
+
export interface DeterministicTurnProvenance {
|
|
5
|
+
/** Answering middleware id; bounded, replay-stable, and auditable. */
|
|
6
|
+
readonly middleware: string;
|
|
7
|
+
}
|
|
8
|
+
/** Host middleware answer that completes a turn without any provider request (plan 096). */
|
|
9
|
+
export interface DeterministicTurnAnswer {
|
|
10
|
+
/** Assistant content with provider content-block shape; tool calls are rejected (no provider ran). */
|
|
11
|
+
readonly content: readonly ContentBlock[];
|
|
12
|
+
/** Mandatory: a deterministic turn can never masquerade as model output. */
|
|
13
|
+
readonly provenance: DeterministicTurnProvenance;
|
|
14
|
+
}
|
|
15
|
+
/**
|
|
16
|
+
* Payload for the `beforeProviderTurn` hook (plan 096). Host middleware sets `answer` to complete the
|
|
17
|
+
* turn deterministically; leaving it absent/undefined sends the turn to the provider as usual.
|
|
18
|
+
*/
|
|
19
|
+
export interface BeforeProviderTurnPayload {
|
|
20
|
+
readonly sessionId: string;
|
|
21
|
+
readonly runId: string;
|
|
22
|
+
readonly turn: number;
|
|
23
|
+
/** Concatenated text blocks of the latest user-role message in the assembled request. */
|
|
24
|
+
readonly userText: string;
|
|
25
|
+
/** Set by host middleware to answer without a provider call. */
|
|
26
|
+
readonly answer?: DeterministicTurnAnswer;
|
|
27
|
+
}
|
|
28
|
+
/** Host answered with a malformed deterministic turn. Fails the run closed — never falls through. */
|
|
29
|
+
export declare class DeterministicTurnError extends Error {
|
|
30
|
+
readonly code = "ERR_PRISM_DETERMINISTIC_TURN";
|
|
31
|
+
constructor(message: string);
|
|
32
|
+
}
|
|
33
|
+
/**
|
|
34
|
+
* Validate a host deterministic answer at the trust boundary: non-empty assistant-visible content and
|
|
35
|
+
* mandatory bounded provenance. Throws instead of degrading to a provider call.
|
|
36
|
+
*/
|
|
37
|
+
export declare function validateDeterministicTurnAnswer(value: unknown): DeterministicTurnAnswer;
|
|
3
38
|
export type MiddlewareNext<T> = (value: T) => Promise<T>;
|
|
4
39
|
export type Middleware<T = unknown> = (value: T, next: MiddlewareNext<T>) => T | Promise<T>;
|
|
5
40
|
export interface MiddlewareRegistryOptions {
|
package/dist/middleware.js
CHANGED
|
@@ -1,4 +1,45 @@
|
|
|
1
1
|
import { errorToErrorInfo } from "./redaction.js";
|
|
2
|
+
/** Host answered with a malformed deterministic turn. Fails the run closed — never falls through. */
|
|
3
|
+
export class DeterministicTurnError extends Error {
|
|
4
|
+
code = "ERR_PRISM_DETERMINISTIC_TURN";
|
|
5
|
+
constructor(message) {
|
|
6
|
+
super(message);
|
|
7
|
+
this.name = "DeterministicTurnError";
|
|
8
|
+
}
|
|
9
|
+
}
|
|
10
|
+
/** Provenance ids stay ids: bounded, no whitespace or separators outside the id alphabet. */
|
|
11
|
+
const DETERMINISTIC_MIDDLEWARE_ID = /^[A-Za-z0-9][A-Za-z0-9._:-]{0,63}$/;
|
|
12
|
+
// Assistant-visible blocks only: a deterministic answer cannot smuggle tool execution/authority.
|
|
13
|
+
const DETERMINISTIC_BLOCK_TYPES = new Set(["text", "image", "audio", "file", "document", "video", "thinking"]);
|
|
14
|
+
/**
|
|
15
|
+
* Validate a host deterministic answer at the trust boundary: non-empty assistant-visible content and
|
|
16
|
+
* mandatory bounded provenance. Throws instead of degrading to a provider call.
|
|
17
|
+
*/
|
|
18
|
+
export function validateDeterministicTurnAnswer(value) {
|
|
19
|
+
if (typeof value !== "object" || value === null)
|
|
20
|
+
throw new DeterministicTurnError("answer must be an object");
|
|
21
|
+
const answer = value;
|
|
22
|
+
if (!Array.isArray(answer.content) || answer.content.length === 0) {
|
|
23
|
+
throw new DeterministicTurnError("answer.content must be a non-empty content block array");
|
|
24
|
+
}
|
|
25
|
+
for (const block of answer.content) {
|
|
26
|
+
const type = block?.type;
|
|
27
|
+
if (typeof block !== "object" || block === null || typeof type !== "string" || !DETERMINISTIC_BLOCK_TYPES.has(type)) {
|
|
28
|
+
throw new DeterministicTurnError("answer.content blocks must be assistant-visible content blocks (text, image, audio, file, document, video, thinking)");
|
|
29
|
+
}
|
|
30
|
+
if (type === "text" && typeof block.text !== "string") {
|
|
31
|
+
throw new DeterministicTurnError("answer.content text blocks require a string text");
|
|
32
|
+
}
|
|
33
|
+
if (type === "thinking" && typeof block.text !== "string") {
|
|
34
|
+
throw new DeterministicTurnError("answer.content thinking blocks require a string text");
|
|
35
|
+
}
|
|
36
|
+
}
|
|
37
|
+
const middleware = answer.provenance?.middleware;
|
|
38
|
+
if (typeof middleware !== "string" || !DETERMINISTIC_MIDDLEWARE_ID.test(middleware)) {
|
|
39
|
+
throw new DeterministicTurnError("answer.provenance.middleware must be a bounded id (1-64 chars: letters, digits, . _ : -)");
|
|
40
|
+
}
|
|
41
|
+
return { content: answer.content, provenance: { middleware } };
|
|
42
|
+
}
|
|
2
43
|
function middlewareError(error, hook, secrets) {
|
|
3
44
|
return { type: "extension_error", extension: `middleware:${hook}`, error: errorToErrorInfo(error, secrets) };
|
|
4
45
|
}
|
|
@@ -1,6 +1,7 @@
|
|
|
1
1
|
import { appendFile, mkdir, readFile } from "node:fs/promises";
|
|
2
2
|
import { dirname } from "node:path";
|
|
3
|
-
import { isSessionEntryKind, SESSION_APPEND_CONFLICT_CODE, SESSION_ENTRY_SCHEMA_VERSION, SessionAppendConflictError,
|
|
3
|
+
import { isSessionEntryKind, SESSION_APPEND_CONFLICT_CODE, SESSION_ENTRY_SCHEMA_VERSION, SessionAppendConflictError, } from "../contracts.js";
|
|
4
|
+
import { searchLinearSessions } from "../session-stores.js";
|
|
4
5
|
import { isNodeErrorCode } from "./config.js";
|
|
5
6
|
export function createJsonlSessionStore(pathOrOptions) {
|
|
6
7
|
const options = typeof pathOrOptions === "string" ? { path: pathOrOptions, createDirectory: true } : pathOrOptions;
|
|
@@ -50,8 +51,22 @@ export function createJsonlSessionStore(pathOrOptions) {
|
|
|
50
51
|
async get(id) {
|
|
51
52
|
return findEntry(path, id);
|
|
52
53
|
},
|
|
53
|
-
async searchSessions() {
|
|
54
|
-
|
|
54
|
+
async searchSessions(query) {
|
|
55
|
+
// ponytail: no index - every search reads and parses the file (O(corpus) time and memory), the
|
|
56
|
+
// recommended indexed paths are the SQLite/Postgres adapters. Corrupt lines are quarantined
|
|
57
|
+
// exactly as in list()/get(), and the contract linear caps bound entries/text scanned.
|
|
58
|
+
const { entries } = await readJsonlSessionEntries(path);
|
|
59
|
+
const bySession = new Map();
|
|
60
|
+
const leafBySession = new Map();
|
|
61
|
+
for (const entry of entries) {
|
|
62
|
+
const sessionEntries = bySession.get(entry.sessionId);
|
|
63
|
+
if (sessionEntries)
|
|
64
|
+
sessionEntries.push(entry);
|
|
65
|
+
else
|
|
66
|
+
bySession.set(entry.sessionId, [entry]);
|
|
67
|
+
leafBySession.set(entry.sessionId, entry.id);
|
|
68
|
+
}
|
|
69
|
+
return searchLinearSessions(bySession, leafBySession, query);
|
|
55
70
|
},
|
|
56
71
|
};
|
|
57
72
|
}
|
package/dist/observability.js
CHANGED
|
@@ -1,9 +1,15 @@
|
|
|
1
|
+
import { createHash } from "node:crypto";
|
|
1
2
|
export function createProviderTurnMetadata(request, providerId, fields = {}) {
|
|
3
|
+
const names = (request.tools ?? []).map((tool) => tool.name);
|
|
2
4
|
return {
|
|
3
5
|
providerId,
|
|
4
6
|
model: request.model,
|
|
5
7
|
requestId: readRequestId(request),
|
|
6
8
|
...fields,
|
|
9
|
+
tools: {
|
|
10
|
+
count: names.length,
|
|
11
|
+
idsHash: `sha256:${createHash("sha256").update(JSON.stringify(names), "utf8").digest("hex")}`,
|
|
12
|
+
},
|
|
7
13
|
};
|
|
8
14
|
}
|
|
9
15
|
export function readProviderHttpStatus(error) {
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import type { ContentBlock, JsonObject, ProviderEvent, ToolCallContent, ToolCallDeltaContent, Usage } from "./contracts.js";
|
|
1
|
+
import type { ContentBlock, JsonObject, ProviderEvent, ProviderStopReason, ToolCallContent, ToolCallDeltaContent, Usage } from "./contracts.js";
|
|
2
2
|
export declare function providerTextDelta(text: string): ProviderEvent;
|
|
3
3
|
export declare function providerThinkingDelta(text: string, signature?: string): ProviderEvent;
|
|
4
4
|
export declare function providerContentDelta(content: ContentBlock): ProviderEvent;
|
|
@@ -13,7 +13,13 @@ export declare function providerToolCallDeltaContent(delta: Omit<ToolCallDeltaCo
|
|
|
13
13
|
export declare function providerContinuationRequired(cursor: string, reason?: string): ProviderEvent;
|
|
14
14
|
export declare function reconstructToolCallDeltas(events: readonly ProviderEvent[]): readonly ToolCallContent[];
|
|
15
15
|
export declare function providerUsage(usage: Usage): ProviderEvent;
|
|
16
|
-
|
|
16
|
+
/**
|
|
17
|
+
* Map a native provider stop/finish reason onto the closed taxonomy. A missing, non-string, or
|
|
18
|
+
* unmapped value returns `"unknown"` rather than throwing, so a new wire value can never fail a
|
|
19
|
+
* run (plan 087 T1).
|
|
20
|
+
*/
|
|
21
|
+
export declare function mapProviderStopReason(native: string | null | undefined): ProviderStopReason;
|
|
22
|
+
export declare function providerDone(usage?: Usage, stopReason?: ProviderStopReason): ProviderEvent;
|
|
17
23
|
export declare function providerError(error: unknown, secrets?: readonly (string | undefined)[]): Extract<ProviderEvent, {
|
|
18
24
|
type: "error";
|
|
19
25
|
}>;
|
package/dist/provider-events.js
CHANGED
|
@@ -46,8 +46,66 @@ export function reconstructToolCallDeltas(events) {
|
|
|
46
46
|
export function providerUsage(usage) {
|
|
47
47
|
return { type: "usage", usage };
|
|
48
48
|
}
|
|
49
|
-
|
|
50
|
-
|
|
49
|
+
/**
|
|
50
|
+
* One shared native → taxonomy table (plan 087 T1). Every adapter routes its wire reason
|
|
51
|
+
* through `mapProviderStopReason`, so hosts switch on one closed union instead of per-provider
|
|
52
|
+
* strings. Keys are lowercased wire values; `unknown` never appears here — it is the fallback.
|
|
53
|
+
*/
|
|
54
|
+
const PROVIDER_STOP_REASONS = Object.freeze({
|
|
55
|
+
// OpenAI Chat Completions (`finish_reason`) and generic OpenAI-compatible routes.
|
|
56
|
+
stop: "end_turn",
|
|
57
|
+
length: "max_output_tokens",
|
|
58
|
+
tool_calls: "tool_calls",
|
|
59
|
+
function_call: "tool_calls",
|
|
60
|
+
content_filter: "content_filter",
|
|
61
|
+
// Messages-style `stop_reason` routes and Bedrock Converse (`stopReason`).
|
|
62
|
+
end_turn: "end_turn",
|
|
63
|
+
stop_sequence: "end_turn",
|
|
64
|
+
pause_turn: "end_turn",
|
|
65
|
+
tool_use: "tool_calls",
|
|
66
|
+
max_tokens: "max_output_tokens",
|
|
67
|
+
refusal: "content_filter",
|
|
68
|
+
// Google generateContent (`finishReason`).
|
|
69
|
+
safety: "content_filter",
|
|
70
|
+
recitation: "content_filter",
|
|
71
|
+
blocklist: "content_filter",
|
|
72
|
+
prohibited_content: "content_filter",
|
|
73
|
+
spii: "content_filter",
|
|
74
|
+
image_safety: "content_filter",
|
|
75
|
+
language: "content_filter",
|
|
76
|
+
malformed_function_call: "provider_error",
|
|
77
|
+
unexpected_tool_call: "provider_error",
|
|
78
|
+
// OpenAI Responses (status / `incomplete_details.reason`).
|
|
79
|
+
completed: "end_turn",
|
|
80
|
+
failed: "provider_error",
|
|
81
|
+
cancelled: "abort",
|
|
82
|
+
canceled: "abort",
|
|
83
|
+
// Bedrock Converse guarded routes.
|
|
84
|
+
content_filtered: "content_filter",
|
|
85
|
+
guardrail_intervened: "content_filter",
|
|
86
|
+
malformed_model_output: "provider_error",
|
|
87
|
+
malformed_tool_use: "provider_error",
|
|
88
|
+
// AI SDK unified finish reasons (hyphenated).
|
|
89
|
+
"content-filter": "content_filter",
|
|
90
|
+
"tool-calls": "tool_calls",
|
|
91
|
+
error: "provider_error",
|
|
92
|
+
other: "unknown",
|
|
93
|
+
unknown: "unknown",
|
|
94
|
+
abort: "abort",
|
|
95
|
+
aborted: "abort",
|
|
96
|
+
});
|
|
97
|
+
/**
|
|
98
|
+
* Map a native provider stop/finish reason onto the closed taxonomy. A missing, non-string, or
|
|
99
|
+
* unmapped value returns `"unknown"` rather than throwing, so a new wire value can never fail a
|
|
100
|
+
* run (plan 087 T1).
|
|
101
|
+
*/
|
|
102
|
+
export function mapProviderStopReason(native) {
|
|
103
|
+
if (typeof native !== "string")
|
|
104
|
+
return "unknown";
|
|
105
|
+
return PROVIDER_STOP_REASONS[native.trim().toLowerCase()] ?? "unknown";
|
|
106
|
+
}
|
|
107
|
+
export function providerDone(usage, stopReason) {
|
|
108
|
+
return { type: "done", usage, ...(stopReason === undefined ? {} : { stopReason }) };
|
|
51
109
|
}
|
|
52
110
|
export function providerError(error, secrets = []) {
|
|
53
111
|
const info = errorToErrorInfo(error, secrets);
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { resolveCredentialValue } from "../credentials.js";
|
|
2
|
-
import { providerDone, providerError, providerTextDelta, providerThinkingDelta, providerToolCall, providerToolCallDelta, providerUsage, toolCallFromArgumentsText, } from "../provider-events.js";
|
|
2
|
+
import { mapProviderStopReason, providerDone, providerError, providerTextDelta, providerThinkingDelta, providerToolCall, providerToolCallDelta, providerUsage, toolCallFromArgumentsText, } from "../provider-events.js";
|
|
3
3
|
import { assertStructuredOutputRequestSupported } from "../structured-output.js";
|
|
4
4
|
import { trimTrailingSlashes } from "../trim-trailing-slashes.js";
|
|
5
5
|
import { applyOpenAIChatStructuredOutput, assertOpenAIChatMessage, mapOpenAIChatUsage, serializeOpenAIChatMessage, serializeOpenAITool, } from "./openai-primitives.js";
|
|
@@ -13,6 +13,7 @@ export async function* openAIChatEvents(body, options = {}) {
|
|
|
13
13
|
let usage;
|
|
14
14
|
let sawDoneMarker = false;
|
|
15
15
|
let sawFinishReason = false;
|
|
16
|
+
let finishReason;
|
|
16
17
|
for await (const sseEvent of readSseEvents(body, { signal: options.signal })) {
|
|
17
18
|
if (options.onComment && sseEvent.comments?.length) {
|
|
18
19
|
for (const text of sseEvent.comments) {
|
|
@@ -43,8 +44,10 @@ export async function* openAIChatEvents(body, options = {}) {
|
|
|
43
44
|
yield providerUsage(mapped);
|
|
44
45
|
}
|
|
45
46
|
for (const choice of parsed.choices ?? []) {
|
|
46
|
-
if (choice.finish_reason)
|
|
47
|
+
if (choice.finish_reason) {
|
|
47
48
|
sawFinishReason = true;
|
|
49
|
+
finishReason = choice.finish_reason;
|
|
50
|
+
}
|
|
48
51
|
const delta = choice.delta ?? {};
|
|
49
52
|
if (typeof delta.content === "string" && delta.content)
|
|
50
53
|
yield providerTextDelta(delta.content);
|
|
@@ -83,7 +86,7 @@ export async function* openAIChatEvents(body, options = {}) {
|
|
|
83
86
|
for (const call of tools.values()) {
|
|
84
87
|
yield providerToolCall(toolCallFromArgumentsText(call.id, call.name, call.argumentsText));
|
|
85
88
|
}
|
|
86
|
-
yield providerDone((options.strictCompletion ?? true) || options.doneUsage ? usage : undefined);
|
|
89
|
+
yield providerDone((options.strictCompletion ?? true) || options.doneUsage ? usage : undefined, finishReason === undefined ? undefined : mapProviderStopReason(finishReason));
|
|
87
90
|
}
|
|
88
91
|
export function createOpenAICompatibleProvider(options) {
|
|
89
92
|
const providerId = options.id ?? "openai-compatible";
|
package/dist/run-bundle.d.ts
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import type { Agent, AgentSessionConfig, GuardrailStage, RunOptions } from "./contracts.js";
|
|
1
|
+
import type { Agent, AgentSessionConfig, GuardrailPackRef, GuardrailStage, RunOptions } from "./contracts.js";
|
|
2
2
|
/** Report format revision. Any shape change bumps this so pinned digests cannot compare across formats. */
|
|
3
3
|
export declare const RUN_BUNDLE_SCHEMA_VERSION = 1;
|
|
4
4
|
/** Inspectable projection of the inputs a run actually resolves to. Frozen JSON, safe to persist and diff. */
|
|
@@ -81,6 +81,11 @@ export interface RunBundleSnapshotInput {
|
|
|
81
81
|
readonly run?: RunOptions;
|
|
82
82
|
/** Optional memory store instance; only its kind/durability label is read, never its contents. */
|
|
83
83
|
readonly memory?: unknown;
|
|
84
|
+
/**
|
|
85
|
+
* Plan 104 Task 2: effective pack refs (`session.guardrailPackRefs`) to report instead of the
|
|
86
|
+
* caller-supplied `config.guardrailPacks` — a resumed session's enforced rows.
|
|
87
|
+
*/
|
|
88
|
+
readonly packs?: readonly GuardrailPackRef[];
|
|
84
89
|
}
|
|
85
90
|
/**
|
|
86
91
|
* Snapshots the effective run bundle: synchronous, in-memory, zero network and zero store reads.
|
package/dist/run-bundle.js
CHANGED
|
@@ -1,5 +1,6 @@
|
|
|
1
1
|
import { createHash } from "node:crypto";
|
|
2
2
|
import { agentFingerprint, BUILT_IN_LOOP_REVISIONS } from "./agent-run-state.js";
|
|
3
|
+
import { describeGuardrailPacks } from "./guardrails.js";
|
|
3
4
|
import { describeStorage } from "./host-composition.js";
|
|
4
5
|
import { canonicalizeJsonSchema } from "./providers/schema.js";
|
|
5
6
|
import { resolveRunLimits } from "./run-limits.js";
|
|
@@ -51,7 +52,10 @@ export function snapshotRunBundle(input) {
|
|
|
51
52
|
effect: tool.effect === undefined ? null : typeof tool.effect === "function" ? "classifier" : tool.effect.kind,
|
|
52
53
|
})),
|
|
53
54
|
activeSkills: run?.activeSkills ?? null,
|
|
54
|
-
guardrails:
|
|
55
|
+
guardrails: [
|
|
56
|
+
...guardrailRows(config.guardrails, run?.guardrails),
|
|
57
|
+
...describeGuardrailPacks(input.packs ?? input.config?.guardrailPacks),
|
|
58
|
+
],
|
|
55
59
|
loop: loopIdentity(effectiveLoop),
|
|
56
60
|
thinkingLevel: run?.thinkingLevel ?? config.thinkingLevel ?? null,
|
|
57
61
|
limits: resolveRunLimits(config.limits, run?.limits),
|
package/dist/run-limits.d.ts
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import type { ResolvedRunLimits, RunLimitBreach, RunLimitCounters, RunLimitName, RunLimits, Usage } from "./contracts.js";
|
|
1
|
+
import type { BudgetAxisUsage, BudgetConsumedCounters, ResolvedRunLimits, RunLimitBreach, RunLimitCounters, RunLimitName, RunLimits, ToolCallSummary, Usage } from "./contracts.js";
|
|
2
2
|
export declare const DEFAULT_RUN_LIMITS: Readonly<{
|
|
3
3
|
maxTurns: 16;
|
|
4
4
|
maxProviderAttempts: 24;
|
|
@@ -51,3 +51,13 @@ export declare class RunLimitTracker {
|
|
|
51
51
|
private exceed;
|
|
52
52
|
}
|
|
53
53
|
export declare function createRunLimitTracker(limits: RunLimits | undefined, options?: RunLimitTrackerOptions): RunLimitTracker;
|
|
54
|
+
/**
|
|
55
|
+
* Build the `budget_exhausted` payload (plan 087 T2): which axis fired, the counters a host reads
|
|
56
|
+
* first, the closest other axes, and the last dispatched tool calls (hashes only).
|
|
57
|
+
*/
|
|
58
|
+
export declare function describeBudgetExhaustion(tracker: RunLimitTracker, breach: RunLimitBreach, recentToolCalls: readonly ToolCallSummary[]): {
|
|
59
|
+
limit: RunLimitName;
|
|
60
|
+
consumed: BudgetConsumedCounters;
|
|
61
|
+
closestOtherAxes: BudgetAxisUsage[];
|
|
62
|
+
recentToolCalls: ToolCallSummary[];
|
|
63
|
+
};
|
package/dist/run-limits.js
CHANGED
|
@@ -10,6 +10,8 @@ export const DEFAULT_RUN_LIMITS = Object.freeze({
|
|
|
10
10
|
maxOutputTokens: 10_000,
|
|
11
11
|
maxTotalTokens: 50_000,
|
|
12
12
|
});
|
|
13
|
+
/** Stop-hook continuation cap when no layer configures one (plan 106 R1). Not a counter axis. */
|
|
14
|
+
const DEFAULT_MAX_STOP_CONTINUATIONS = 3;
|
|
13
15
|
/**
|
|
14
16
|
* Process-safety ceilings that exist so a bug cannot OOM the host via JSON.parse of giant
|
|
15
17
|
* provider frames. Product axes (turns, wall time, tokens, …) have no hard cap: hosts set
|
|
@@ -83,8 +85,13 @@ export function resolveRunLimits(agent, run) {
|
|
|
83
85
|
resolved.maxProviderAttempts = turns;
|
|
84
86
|
}
|
|
85
87
|
const maxCost = override?.maxCost ?? base?.maxCost;
|
|
88
|
+
// Stop-hook continuation cap (plan 106 R1): no counters-table row — the wrapper turns it into a
|
|
89
|
+
// clean `hook_limit` stop, not a breach — so it resolves outside the counter-backed axes and
|
|
90
|
+
// keeps the same narrowing-only law (min, `null` = uncapped).
|
|
91
|
+
const stopContinuations = minCap(base?.maxStopContinuations, override?.maxStopContinuations);
|
|
86
92
|
return Object.freeze({
|
|
87
93
|
...resolved,
|
|
94
|
+
maxStopContinuations: stopContinuations !== undefined ? stopContinuations : DEFAULT_MAX_STOP_CONTINUATIONS,
|
|
88
95
|
...(maxCost
|
|
89
96
|
? {
|
|
90
97
|
maxCost: base?.maxCost && override?.maxCost
|
|
@@ -121,6 +128,12 @@ function validateLimits(input) {
|
|
|
121
128
|
if (!Number.isFinite(amount) || amount < 0 || !currency.trim())
|
|
122
129
|
throw new TypeError("maxCost requires a finite non-negative amount and currency");
|
|
123
130
|
}
|
|
131
|
+
const stopContinuations = input.maxStopContinuations;
|
|
132
|
+
if (stopContinuations !== undefined &&
|
|
133
|
+
stopContinuations !== null &&
|
|
134
|
+
(!Number.isSafeInteger(stopContinuations) || stopContinuations < 0)) {
|
|
135
|
+
throw new TypeError("maxStopContinuations must be a non-negative safe integer or null to disable the cap");
|
|
136
|
+
}
|
|
124
137
|
return input;
|
|
125
138
|
}
|
|
126
139
|
export class RunLimitTracker {
|
|
@@ -263,4 +276,50 @@ export class RunLimitTracker {
|
|
|
263
276
|
export function createRunLimitTracker(limits, options) {
|
|
264
277
|
return new RunLimitTracker(resolveRunLimits(undefined, limits), options);
|
|
265
278
|
}
|
|
279
|
+
/** Product axes reported as "how close was everything else" (plan 087 T2). Request/response bytes
|
|
280
|
+
* stay out: their caps are per-frame, so a run-lifetime ratio would be meaningless. */
|
|
281
|
+
const ATTRIBUTION_AXES = [
|
|
282
|
+
{ axis: "maxTurns", counter: "turns" },
|
|
283
|
+
{ axis: "maxProviderAttempts", counter: "providerAttempts" },
|
|
284
|
+
{ axis: "maxToolRounds", counter: "toolRounds" },
|
|
285
|
+
{ axis: "maxToolCalls", counter: "toolCalls" },
|
|
286
|
+
{ axis: "maxWallTimeMs", counter: "wallTimeMs" },
|
|
287
|
+
{ axis: "maxInputTokens", counter: "inputTokens" },
|
|
288
|
+
{ axis: "maxOutputTokens", counter: "outputTokens" },
|
|
289
|
+
{ axis: "maxTotalTokens", counter: "totalTokens" },
|
|
290
|
+
{ axis: "maxCost", counter: "cost" },
|
|
291
|
+
];
|
|
292
|
+
const CLOSEST_AXIS_COUNT = 3;
|
|
293
|
+
function axisCap(limits, axis) {
|
|
294
|
+
if (axis === "maxCost")
|
|
295
|
+
return limits.maxCost?.amount;
|
|
296
|
+
const value = limits[axis];
|
|
297
|
+
return typeof value === "number" ? value : null;
|
|
298
|
+
}
|
|
299
|
+
/**
|
|
300
|
+
* Build the `budget_exhausted` payload (plan 087 T2): which axis fired, the counters a host reads
|
|
301
|
+
* first, the closest other axes, and the last dispatched tool calls (hashes only).
|
|
302
|
+
*/
|
|
303
|
+
export function describeBudgetExhaustion(tracker, breach, recentToolCalls) {
|
|
304
|
+
const counters = tracker.snapshot();
|
|
305
|
+
const closestOtherAxes = ATTRIBUTION_AXES.filter(({ axis }) => axis !== breach.limit)
|
|
306
|
+
.flatMap(({ axis, counter }) => {
|
|
307
|
+
const cap = axisCap(tracker.limits, axis);
|
|
308
|
+
return typeof cap === "number" && cap > 0 ? [{ axis, cap, used: counters[counter] }] : [];
|
|
309
|
+
})
|
|
310
|
+
.map(({ axis, cap, used }) => ({ axis, usedRatio: Math.round(Math.min(1, used / cap) * 10_000) / 10_000 }))
|
|
311
|
+
.sort((a, b) => b.usedRatio - a.usedRatio)
|
|
312
|
+
.slice(0, CLOSEST_AXIS_COUNT);
|
|
313
|
+
return {
|
|
314
|
+
limit: breach.limit,
|
|
315
|
+
consumed: {
|
|
316
|
+
turns: counters.turns,
|
|
317
|
+
inputTokens: counters.inputTokens,
|
|
318
|
+
providerAttempts: counters.providerAttempts,
|
|
319
|
+
requestBytes: counters.requestBytes,
|
|
320
|
+
},
|
|
321
|
+
closestOtherAxes,
|
|
322
|
+
recentToolCalls: [...recentToolCalls],
|
|
323
|
+
};
|
|
324
|
+
}
|
|
266
325
|
//# sourceMappingURL=run-limits.js.map
|
package/dist/session-stores.d.ts
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { type BranchReader, type Message, type SessionBranchRead, type SessionEntry, type SessionStore } from "./contracts.js";
|
|
1
|
+
import { type BranchReader, type Message, type PersistencePage, type SessionBranchRead, type SessionEntry, type SessionSearchHit, type SessionSearchQuery, type SessionStore } from "./contracts.js";
|
|
2
2
|
export interface CreateSessionEntryOptions extends Omit<SessionEntry, "id" | "timestamp"> {
|
|
3
3
|
readonly id?: string;
|
|
4
4
|
readonly timestamp?: string;
|
|
@@ -40,4 +40,15 @@ export interface CreateMemorySessionStoreOptions {
|
|
|
40
40
|
readonly maxLinearBytes?: number;
|
|
41
41
|
};
|
|
42
42
|
}
|
|
43
|
+
/** Resolved bounds for the capped linear session scan. */
|
|
44
|
+
export interface LinearSearchCaps {
|
|
45
|
+
readonly sessions: number;
|
|
46
|
+
readonly entries: number;
|
|
47
|
+
readonly bytes: number;
|
|
48
|
+
}
|
|
43
49
|
export declare function createMemorySessionStore(initialEntries?: readonly SessionEntry[], options?: CreateMemorySessionStoreOptions): SessionStore;
|
|
50
|
+
/**
|
|
51
|
+
* Shared linear (unindexed) session search over already-grouped entries: the memory store's default
|
|
52
|
+
* mode and the JSONL store's implementation. `caps` defaults to the contract linear caps.
|
|
53
|
+
*/
|
|
54
|
+
export declare function searchLinearSessions(bySession: Map<string, SessionEntry[]>, leafBySession: Map<string, string>, query: SessionSearchQuery, caps?: LinearSearchCaps): PersistencePage<SessionSearchHit>;
|
package/dist/session-stores.js
CHANGED
|
@@ -133,7 +133,7 @@ export function createMemorySessionStore(initialEntries = [], options = {}) {
|
|
|
133
133
|
async searchSessions(query) {
|
|
134
134
|
if (mode === "unsupported")
|
|
135
135
|
throw new SessionSearchUnsupportedError();
|
|
136
|
-
return
|
|
136
|
+
return searchLinearSessions(bySession, leafBySession, query, searchCaps);
|
|
137
137
|
},
|
|
138
138
|
};
|
|
139
139
|
function add(entry, options) {
|
|
@@ -175,7 +175,11 @@ export function createMemorySessionStore(initialEntries = [], options = {}) {
|
|
|
175
175
|
leafBySession.set(entry.sessionId, entry.id);
|
|
176
176
|
}
|
|
177
177
|
}
|
|
178
|
-
|
|
178
|
+
/**
|
|
179
|
+
* Shared linear (unindexed) session search over already-grouped entries: the memory store's default
|
|
180
|
+
* mode and the JSONL store's implementation. `caps` defaults to the contract linear caps.
|
|
181
|
+
*/
|
|
182
|
+
export function searchLinearSessions(bySession, leafBySession, query, caps = resolveLinearSearchCaps(undefined)) {
|
|
179
183
|
const q = resolveSessionSearchQuery(query);
|
|
180
184
|
q.signal?.throwIfAborted();
|
|
181
185
|
let sessionsScanned = 0;
|
|
@@ -203,13 +207,20 @@ function searchMemorySessionsLinear(bySession, leafBySession, query, caps) {
|
|
|
203
207
|
let matchedQuery = false;
|
|
204
208
|
let matchedProvider = false;
|
|
205
209
|
let matchedModel = false;
|
|
210
|
+
let matchedKind = false;
|
|
211
|
+
let matchedEntry;
|
|
206
212
|
let snippetSource;
|
|
213
|
+
let turn = 0;
|
|
207
214
|
for (const entry of entries) {
|
|
208
215
|
if (entriesScanned >= caps.entries)
|
|
209
216
|
break;
|
|
210
217
|
if (bytesScanned >= caps.bytes)
|
|
211
218
|
break;
|
|
212
219
|
entriesScanned += 1;
|
|
220
|
+
turn += 1;
|
|
221
|
+
const kindMatches = q.kind === undefined || q.kind.includes(entry.kind);
|
|
222
|
+
if (kindMatches)
|
|
223
|
+
matchedKind = true;
|
|
213
224
|
const text = entrySearchText(entry);
|
|
214
225
|
bytesScanned += utf8Bytes(text) + utf8Bytes(entry.label) + utf8Bytes(entry.summary);
|
|
215
226
|
if (entry.timestamp > updatedAt)
|
|
@@ -234,11 +245,12 @@ function searchMemorySessionsLinear(bySession, leafBySession, query, caps) {
|
|
|
234
245
|
matchedLabel = true;
|
|
235
246
|
if (q.summary && entry.summary?.includes(q.summary))
|
|
236
247
|
matchedSummary = true;
|
|
237
|
-
if (q.query) {
|
|
248
|
+
if (q.query && kindMatches) {
|
|
238
249
|
const hay = `${entry.label ?? ""}\n${entry.summary ?? ""}\n${text}`;
|
|
239
250
|
if (hay.includes(q.query)) {
|
|
240
251
|
matchedQuery = true;
|
|
241
|
-
|
|
252
|
+
matchedEntry ??= { entry, turn };
|
|
253
|
+
snippetSource ??= [entry.label, entry.summary, text].find((part) => part?.includes(q.query)) ?? hay;
|
|
242
254
|
}
|
|
243
255
|
}
|
|
244
256
|
if (q.provider && (entry.model?.provider === q.provider || metaProvider(entry) === q.provider))
|
|
@@ -260,6 +272,8 @@ function searchMemorySessionsLinear(bySession, leafBySession, query, caps) {
|
|
|
260
272
|
continue;
|
|
261
273
|
if (q.query && !matchedQuery)
|
|
262
274
|
continue;
|
|
275
|
+
if (q.kind && (q.query ? matchedEntry === undefined : !matchedKind))
|
|
276
|
+
continue;
|
|
263
277
|
if (q.provider && !matchedProvider)
|
|
264
278
|
continue;
|
|
265
279
|
if (q.model && !matchedModel)
|
|
@@ -271,6 +285,9 @@ function searchMemorySessionsLinear(bySession, leafBySession, query, caps) {
|
|
|
271
285
|
matches.push({
|
|
272
286
|
sessionId,
|
|
273
287
|
leafId: leafBySession.get(sessionId),
|
|
288
|
+
entryId: matchedEntry?.entry.id,
|
|
289
|
+
runId: matchedEntry?.entry.runId,
|
|
290
|
+
turn: matchedEntry?.turn,
|
|
274
291
|
updatedAt: updatedAt || undefined,
|
|
275
292
|
label,
|
|
276
293
|
summary,
|