pi-crew 0.9.20 → 0.9.21
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/dist/build-meta.json +4725 -4677
- package/dist/index.mjs +2060 -1760
- package/dist/index.mjs.map +4 -4
- package/package.json +1 -1
- package/src/extension/cross-extension-rpc.ts +15 -2
- package/src/extension/team-tool/run.ts +33 -4
- package/src/runtime/background-runner.ts +7 -0
- package/src/runtime/dynamic-workflow-context.ts +3 -3
- package/src/runtime/task-runner.ts +4 -4
- package/src/runtime/team-runner.ts +130 -1
- package/src/runtime/tool-output-pruner.ts +5 -3
- package/src/state/event-log.ts +18 -0
- package/src/state/types.ts +5 -0
- package/src/utils/token-counter.ts +67 -0
- package/src/worktree/worktree-manager.ts +365 -1
package/package.json
CHANGED
|
@@ -1,9 +1,9 @@
|
|
|
1
1
|
import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
|
|
2
2
|
import type { TeamToolParamsValue } from "../schema/team-tool-schema.ts";
|
|
3
3
|
import { resolveContainedPath } from "../utils/safe-paths.ts";
|
|
4
|
+
import { withHmacVerification } from "./rpc-hmac.ts";
|
|
4
5
|
// Lazy-loaded to avoid pulling team-tool.ts (and its entire runtime chain) into module load.
|
|
5
6
|
import type { handleTeamTool as HandleTeamToolFn } from "./team-tool.ts";
|
|
6
|
-
import { withHmacVerification } from "./rpc-hmac.ts";
|
|
7
7
|
|
|
8
8
|
let _cachedHandleTeamTool: typeof HandleTeamToolFn | undefined;
|
|
9
9
|
async function handleTeamTool(
|
|
@@ -184,7 +184,20 @@ export function registerPiCrewRpc(
|
|
|
184
184
|
const ctx = getCtx();
|
|
185
185
|
if (!ctx) throw new Error("No active pi-crew session context.");
|
|
186
186
|
// Validate payload: only allow known fields from TeamToolParamsValue
|
|
187
|
-
const ALLOWED_RPC_RUN_KEYS = new Set([
|
|
187
|
+
const ALLOWED_RPC_RUN_KEYS = new Set([
|
|
188
|
+
"goal",
|
|
189
|
+
"team",
|
|
190
|
+
"workflow",
|
|
191
|
+
"async",
|
|
192
|
+
"cwd",
|
|
193
|
+
"config",
|
|
194
|
+
"skill",
|
|
195
|
+
"model",
|
|
196
|
+
"budgetTotal",
|
|
197
|
+
"budgetWarning",
|
|
198
|
+
"budgetAbort",
|
|
199
|
+
"budgetUnlimited",
|
|
200
|
+
]);
|
|
188
201
|
let params: TeamToolParamsValue;
|
|
189
202
|
if (raw && typeof raw === "object" && !Array.isArray(raw)) {
|
|
190
203
|
const filtered: Record<string, unknown> = {
|
|
@@ -12,7 +12,7 @@ import { createRunManifest, loadRunManifestById, updateRunStatus } from "../../s
|
|
|
12
12
|
import { allTeams, discoverTeams } from "../../teams/discover-teams.ts";
|
|
13
13
|
import { allWorkflows, discoverWorkflows } from "../../workflows/discover-workflows.ts";
|
|
14
14
|
import { validateWorkflowForTeam } from "../../workflows/validate-workflow.ts";
|
|
15
|
-
import {
|
|
15
|
+
import { assertCleanLeaderAsync, findGitRootAsync } from "../../worktree/worktree-manager.ts";
|
|
16
16
|
|
|
17
17
|
// eslint-disable-next-line @typescript-eslint/no-unused-vars -- type-only import for TS inference
|
|
18
18
|
const _typeCheck: typeof ExecuteTeamRunFn = null as never as typeof ExecuteTeamRunFn;
|
|
@@ -226,7 +226,7 @@ export async function handleRun(params: TeamToolParamsValue, ctx: TeamContext):
|
|
|
226
226
|
let resolvedCtx = ctx;
|
|
227
227
|
if (workingDir) {
|
|
228
228
|
try {
|
|
229
|
-
const gitRoot =
|
|
229
|
+
const gitRoot = await findGitRootAsync(workingDir);
|
|
230
230
|
if (gitRoot && gitRoot !== workingDir) {
|
|
231
231
|
resolvedCtx = { ...ctx, cwd: gitRoot };
|
|
232
232
|
}
|
|
@@ -241,7 +241,7 @@ export async function handleRun(params: TeamToolParamsValue, ctx: TeamContext):
|
|
|
241
241
|
if (params.workspaceMode === "worktree") {
|
|
242
242
|
let gitRoot: string | undefined;
|
|
243
243
|
try {
|
|
244
|
-
gitRoot =
|
|
244
|
+
gitRoot = await findGitRootAsync(resolvedCtx.cwd);
|
|
245
245
|
} catch {
|
|
246
246
|
// not a git repo
|
|
247
247
|
}
|
|
@@ -256,7 +256,7 @@ export async function handleRun(params: TeamToolParamsValue, ctx: TeamContext):
|
|
|
256
256
|
const preCheckConfig = loadConfig(resolvedCtx.cwd);
|
|
257
257
|
if (preCheckConfig.config.requireCleanWorktreeLeader !== false) {
|
|
258
258
|
try {
|
|
259
|
-
|
|
259
|
+
await assertCleanLeaderAsync(gitRoot);
|
|
260
260
|
} catch (err) {
|
|
261
261
|
const msg = err instanceof Error ? err.message : String(err);
|
|
262
262
|
return result(
|
|
@@ -573,10 +573,31 @@ export async function handleRun(params: TeamToolParamsValue, ctx: TeamContext):
|
|
|
573
573
|
const executedConfig = effectiveRunConfig(loadedConfig.config, params.config);
|
|
574
574
|
const runtime = await resolveCrewRuntime(executedConfig);
|
|
575
575
|
const runtimeResolution = runtimeResolutionState(runtime);
|
|
576
|
+
// DEBUG: log what we received (gated to avoid stdout pollution in production)
|
|
577
|
+
if (process.env.PI_CREW_DEBUG_BUDGET === "1") {
|
|
578
|
+
console.log(
|
|
579
|
+
"[DEBUG budget] params keys:",
|
|
580
|
+
Object.keys(params),
|
|
581
|
+
"budgetTotal:",
|
|
582
|
+
params.budgetTotal,
|
|
583
|
+
"budgetWarning:",
|
|
584
|
+
params.budgetWarning,
|
|
585
|
+
"budgetAbort:",
|
|
586
|
+
params.budgetAbort,
|
|
587
|
+
);
|
|
588
|
+
}
|
|
576
589
|
const executionManifest = {
|
|
577
590
|
...updatedManifest,
|
|
578
591
|
runtimeResolution,
|
|
579
592
|
runConfig: executedConfig,
|
|
593
|
+
// Persist budget config on the manifest so it's observable post-run
|
|
594
|
+
// (events.jsonl, status reads, audits). The team-runner reads these
|
|
595
|
+
// from the input, but persisting them means consumers can verify
|
|
596
|
+
// enforcement was armed without re-parsing the original call params.
|
|
597
|
+
...(params.budgetTotal !== undefined ? { budgetTotal: params.budgetTotal } : {}),
|
|
598
|
+
...(params.budgetWarning !== undefined ? { budgetWarning: params.budgetWarning } : {}),
|
|
599
|
+
...(params.budgetAbort !== undefined ? { budgetAbort: params.budgetAbort } : {}),
|
|
600
|
+
...(params.budgetUnlimited !== undefined ? { budgetUnlimited: params.budgetUnlimited } : {}),
|
|
580
601
|
updatedAt: new Date().toISOString(),
|
|
581
602
|
};
|
|
582
603
|
atomicWriteJson(paths.manifestPath, executionManifest);
|
|
@@ -862,6 +883,10 @@ export async function handleRun(params: TeamToolParamsValue, ctx: TeamContext):
|
|
|
862
883
|
metricRegistry: ctx.metricRegistry,
|
|
863
884
|
onJsonEvent: ctx.onJsonEvent,
|
|
864
885
|
workspaceId: ctx.sessionId ?? ctx.cwd,
|
|
886
|
+
budgetTotal: params.budgetTotal,
|
|
887
|
+
budgetWarning: params.budgetWarning,
|
|
888
|
+
budgetAbort: params.budgetAbort,
|
|
889
|
+
budgetUnlimited: params.budgetUnlimited,
|
|
865
890
|
});
|
|
866
891
|
} finally {
|
|
867
892
|
unregisterActiveRun(updatedManifest.runId);
|
|
@@ -1006,6 +1031,10 @@ export async function handleRun(params: TeamToolParamsValue, ctx: TeamContext):
|
|
|
1006
1031
|
metricRegistry: ctx.metricRegistry,
|
|
1007
1032
|
onJsonEvent: ctx.onJsonEvent,
|
|
1008
1033
|
workspaceId: ctx.cwd,
|
|
1034
|
+
budgetTotal: params.budgetTotal,
|
|
1035
|
+
budgetWarning: params.budgetWarning,
|
|
1036
|
+
budgetAbort: params.budgetAbort,
|
|
1037
|
+
budgetUnlimited: params.budgetUnlimited,
|
|
1009
1038
|
});
|
|
1010
1039
|
} finally {
|
|
1011
1040
|
unregisterActiveRun(updatedManifest.runId);
|
|
@@ -714,6 +714,9 @@ async function main(): Promise<void> {
|
|
|
714
714
|
switch (manifest.runKind ?? "team-run") {
|
|
715
715
|
default: {
|
|
716
716
|
// Existing "team-run" path — unchanged behavior.
|
|
717
|
+
// Forward budget fields from manifest (set by team-tool/run.ts
|
|
718
|
+
// from params.budgetTotal/etc.) so the team-runner's
|
|
719
|
+
// checkPerTaskBudget guard actually arms.
|
|
717
720
|
result = await executeTeamRun({
|
|
718
721
|
manifest,
|
|
719
722
|
tasks,
|
|
@@ -728,6 +731,10 @@ async function main(): Promise<void> {
|
|
|
728
731
|
reliability: runConfig.reliability,
|
|
729
732
|
workspaceId: manifest.ownerSessionId ?? manifest.cwd,
|
|
730
733
|
signal: abortController.signal,
|
|
734
|
+
...(manifest.budgetTotal !== undefined ? { budgetTotal: manifest.budgetTotal } : {}),
|
|
735
|
+
...(manifest.budgetWarning !== undefined ? { budgetWarning: manifest.budgetWarning } : {}),
|
|
736
|
+
...(manifest.budgetAbort !== undefined ? { budgetAbort: manifest.budgetAbort } : {}),
|
|
737
|
+
...(manifest.budgetUnlimited !== undefined ? { budgetUnlimited: manifest.budgetUnlimited } : {}),
|
|
731
738
|
});
|
|
732
739
|
break;
|
|
733
740
|
}
|
|
@@ -32,7 +32,7 @@ import { appendMailboxMessage, readMailbox } from "../state/mailbox.ts";
|
|
|
32
32
|
import type { TeamRunManifest } from "../state/types.ts";
|
|
33
33
|
import type { TeamConfig } from "../teams/team-config.ts";
|
|
34
34
|
import { logInternalError } from "../utils/internal-error.ts";
|
|
35
|
-
import {
|
|
35
|
+
import { cleanupAgentWorktreeAsync, prepareAgentWorktreeAsync } from "../worktree/worktree-manager.ts";
|
|
36
36
|
import { runChildPi } from "./child-pi.ts";
|
|
37
37
|
import type { DwfCheckpointState } from "./dwf-state-store.ts";
|
|
38
38
|
import { mapConcurrent } from "./parallel-utils.ts";
|
|
@@ -342,7 +342,7 @@ export function makeWorkflowCtx(manifest: TeamRunManifest, opts: MakeWorkflowCtx
|
|
|
342
342
|
// when worktree creation is unavailable (no git repo, dirty leader).
|
|
343
343
|
let agentCwd = manifest.cwd;
|
|
344
344
|
if (call.worktree === true) {
|
|
345
|
-
const wt =
|
|
345
|
+
const wt = await prepareAgentWorktreeAsync(manifest, `dwf-agent-${Date.now()}-${randomBytes(4).toString("hex")}`);
|
|
346
346
|
if (wt?.worktreePath) {
|
|
347
347
|
agentCwd = wt.cwd;
|
|
348
348
|
worktreePath = wt.worktreePath;
|
|
@@ -431,7 +431,7 @@ export function makeWorkflowCtx(manifest: TeamRunManifest, opts: MakeWorkflowCtx
|
|
|
431
431
|
// — a leak must never crash the workflow.
|
|
432
432
|
if (worktreePath) {
|
|
433
433
|
try {
|
|
434
|
-
|
|
434
|
+
await cleanupAgentWorktreeAsync(manifest, worktreePath, worktreeBranch);
|
|
435
435
|
} catch (cleanupError) {
|
|
436
436
|
logInternalError("dynamic-workflow-context.worktree-cleanup", cleanupError, `worktreePath=${worktreePath}`);
|
|
437
437
|
}
|
|
@@ -20,7 +20,7 @@ import type {
|
|
|
20
20
|
import { logInternalError } from "../utils/internal-error.ts";
|
|
21
21
|
import { resolveRealContainedPath } from "../utils/safe-paths.ts";
|
|
22
22
|
import type { WorkflowStep } from "../workflows/workflow-config.ts";
|
|
23
|
-
import {
|
|
23
|
+
import { captureWorktreeDiffAsync, captureWorktreeDiffStatAsync, prepareTaskWorkspaceAsync } from "../worktree/worktree-manager.ts";
|
|
24
24
|
import { reserveControlChannel } from "./agent-control.ts";
|
|
25
25
|
import { appendTaskAttentionEvent } from "./attention-events.ts";
|
|
26
26
|
import { buildSyntheticTerminalEvidence, cancellationReasonFromSignal } from "./cancellation.ts";
|
|
@@ -122,7 +122,7 @@ export async function runTeamTask(input: TaskRunnerInput): Promise<{ manifest: T
|
|
|
122
122
|
let streamBridge: ReturnType<typeof registerStreamBridge> | undefined;
|
|
123
123
|
try {
|
|
124
124
|
streamBridge = registerStreamBridge(manifest.runId);
|
|
125
|
-
const workspace =
|
|
125
|
+
const workspace = await prepareTaskWorkspaceAsync(manifest, input.task, input.step.seedPaths);
|
|
126
126
|
const worktree =
|
|
127
127
|
workspace.worktreePath && workspace.branch
|
|
128
128
|
? {
|
|
@@ -905,7 +905,7 @@ export async function runTeamTask(input: TaskRunnerInput): Promise<{ manifest: T
|
|
|
905
905
|
? writeArtifact(manifest.artifactsRoot, {
|
|
906
906
|
kind: "diff",
|
|
907
907
|
relativePath: `diffs/${task.id}.diff`,
|
|
908
|
-
content:
|
|
908
|
+
content: await captureWorktreeDiffAsync(workspace.worktreePath),
|
|
909
909
|
producer: task.id,
|
|
910
910
|
})
|
|
911
911
|
: undefined;
|
|
@@ -913,7 +913,7 @@ export async function runTeamTask(input: TaskRunnerInput): Promise<{ manifest: T
|
|
|
913
913
|
? writeArtifact(manifest.artifactsRoot, {
|
|
914
914
|
kind: "metadata",
|
|
915
915
|
relativePath: `metadata/${task.id}.diff-stat.json`,
|
|
916
|
-
content: `${JSON.stringify({ ...
|
|
916
|
+
content: `${JSON.stringify({ ...(await captureWorktreeDiffStatAsync(workspace.worktreePath)), syntheticPaths: workspace.syntheticPaths ?? [], nodeModulesLinked: workspace.nodeModulesLinked ?? false }, null, 2)}\n`,
|
|
917
917
|
producer: task.id,
|
|
918
918
|
})
|
|
919
919
|
: undefined;
|
|
@@ -13,7 +13,7 @@ import { HealthStore } from "../state/health-store.ts";
|
|
|
13
13
|
import { withRunLock } from "../state/locks.ts";
|
|
14
14
|
import { loadRunManifestById, saveRunManifest, saveRunManifestAsync, saveRunTasksAsync, updateRunStatus } from "../state/state-store.ts";
|
|
15
15
|
import type { ArtifactDescriptor, PolicyDecision, TaskAttemptState, TeamRunManifest, TeamTaskState } from "../state/types.ts";
|
|
16
|
-
import { aggregateUsage, formatUsage } from "../state/usage.ts";
|
|
16
|
+
import { aggregateUsage, formatTokens, formatUsage } from "../state/usage.ts";
|
|
17
17
|
import type { TeamConfig } from "../teams/team-config.ts";
|
|
18
18
|
import { logInternalError } from "../utils/internal-error.ts";
|
|
19
19
|
import type { WorkflowConfig, WorkflowStep } from "../workflows/workflow-config.ts";
|
|
@@ -130,6 +130,60 @@ export interface ExecuteTeamRunInput {
|
|
|
130
130
|
onJsonEvent?: (taskId: string, runId: string, event: unknown) => void;
|
|
131
131
|
/** Workspace where this run was initiated — used for session-scoped live-agent visibility. */
|
|
132
132
|
workspaceId: string;
|
|
133
|
+
/** Total token budget for the run. When set, enables per-task budget enforcement. */
|
|
134
|
+
budgetTotal?: number;
|
|
135
|
+
/** Budget warning threshold as a fraction (0-1). Default: 0.8 (80%). */
|
|
136
|
+
budgetWarning?: number;
|
|
137
|
+
/** Budget abort threshold as a fraction (0-1). Default: 0.95 (95%). */
|
|
138
|
+
budgetAbort?: number;
|
|
139
|
+
/** When true, skip budget enforcement entirely. */
|
|
140
|
+
budgetUnlimited?: boolean;
|
|
141
|
+
}
|
|
142
|
+
|
|
143
|
+
/**
|
|
144
|
+
* Result of a per-task budget check against cumulative run usage.
|
|
145
|
+
*/
|
|
146
|
+
export interface PerTaskBudgetCheckResult {
|
|
147
|
+
/** Whether the abort threshold was exceeded. */
|
|
148
|
+
abort: boolean;
|
|
149
|
+
/** Whether the warning threshold was exceeded. */
|
|
150
|
+
warning: boolean;
|
|
151
|
+
/** IDs of tasks that exceeded their fair share (>50% of remaining budget). */
|
|
152
|
+
fairShareViolators: string[];
|
|
153
|
+
/** Total tokens used so far. */
|
|
154
|
+
totalUsed: number;
|
|
155
|
+
}
|
|
156
|
+
|
|
157
|
+
/**
|
|
158
|
+
* Check cumulative token usage against per-task budget thresholds.
|
|
159
|
+
* Returns a structured result — callers decide how to act (warn vs abort).
|
|
160
|
+
*
|
|
161
|
+
* Exported for unit testing.
|
|
162
|
+
*/
|
|
163
|
+
export function checkPerTaskBudget(
|
|
164
|
+
tasks: TeamTaskState[],
|
|
165
|
+
budgetTotal: number,
|
|
166
|
+
budgetWarning: number,
|
|
167
|
+
budgetAbort: number,
|
|
168
|
+
fairShareFraction = 0.5,
|
|
169
|
+
): PerTaskBudgetCheckResult {
|
|
170
|
+
const usage = aggregateUsage(tasks);
|
|
171
|
+
const totalUsed = (usage?.input ?? 0) + (usage?.output ?? 0) + (usage?.cacheWrite ?? 0);
|
|
172
|
+
const abort = totalUsed >= budgetAbort * budgetTotal;
|
|
173
|
+
const warning = !abort && totalUsed >= budgetWarning * budgetTotal;
|
|
174
|
+
const remainingBudget = Math.max(0, budgetTotal - totalUsed);
|
|
175
|
+
const fairShareThreshold = remainingBudget * fairShareFraction;
|
|
176
|
+
const fairShareViolators: string[] = [];
|
|
177
|
+
for (const task of tasks) {
|
|
178
|
+
if (!task.usage) continue;
|
|
179
|
+
const taskTotal = (task.usage.input ?? 0) + (task.usage.output ?? 0) + (task.usage.cacheWrite ?? 0);
|
|
180
|
+
// Only flag tasks that individually consumed a significant portion of the
|
|
181
|
+
// budget (>10% of total) AND exceeded the fair share of remaining budget.
|
|
182
|
+
if (fairShareThreshold > 0 && taskTotal > fairShareThreshold && taskTotal > budgetTotal * 0.1) {
|
|
183
|
+
fairShareViolators.push(task.id);
|
|
184
|
+
}
|
|
185
|
+
}
|
|
186
|
+
return { abort, warning, fairShareViolators, totalUsed };
|
|
133
187
|
}
|
|
134
188
|
|
|
135
189
|
function findStep(workflow: WorkflowConfig, task: TeamTaskState): WorkflowStep {
|
|
@@ -586,6 +640,18 @@ export async function executeTeamRun(input: ExecuteTeamRunInput): Promise<{ mani
|
|
|
586
640
|
input.executeWorkers ? "Executing team workflow." : "Creating workflow prompts and placeholder results.",
|
|
587
641
|
);
|
|
588
642
|
|
|
643
|
+
// Persist budget fields on the manifest so all subsequent saveRunManifest
|
|
644
|
+
// calls (there are many in executeTeamRunCore) preserve the budget config.
|
|
645
|
+
// Without this, the spread in updateRunStatus and other transformations
|
|
646
|
+
// can drop budget fields before they reach the persistence layer.
|
|
647
|
+
if (input.budgetTotal !== undefined) manifest.budgetTotal = input.budgetTotal;
|
|
648
|
+
if (input.budgetWarning !== undefined) manifest.budgetWarning = input.budgetWarning;
|
|
649
|
+
if (input.budgetAbort !== undefined) manifest.budgetAbort = input.budgetAbort;
|
|
650
|
+
if (input.budgetUnlimited !== undefined) manifest.budgetUnlimited = input.budgetUnlimited;
|
|
651
|
+
if (manifest.budgetTotal !== undefined && manifest.budgetTotal > 0 && !manifest.budgetUnlimited) {
|
|
652
|
+
saveRunManifest(manifest);
|
|
653
|
+
}
|
|
654
|
+
|
|
589
655
|
void registerRunPromise(manifest.runId);
|
|
590
656
|
|
|
591
657
|
// FIX (Round 15, regression): Start a team-level heartbeat so the stale
|
|
@@ -1444,6 +1510,69 @@ async function executeTeamRunCore(
|
|
|
1444
1510
|
wfMachine = { ...wfMachine, currentPhaseIndex: pi + 1 };
|
|
1445
1511
|
}
|
|
1446
1512
|
|
|
1513
|
+
// Per-task budget enforcement: check cumulative usage after each batch merge.
|
|
1514
|
+
// This prevents a single task from consuming 100% of the budget before
|
|
1515
|
+
// abort triggers (the goal-loop only checks at turn boundaries).
|
|
1516
|
+
if (input.budgetTotal !== undefined && input.budgetTotal > 0 && input.budgetUnlimited !== true) {
|
|
1517
|
+
const warnThreshold = input.budgetWarning ?? 0.8;
|
|
1518
|
+
const abortThreshold = input.budgetAbort ?? 0.95;
|
|
1519
|
+
const budgetCheck = checkPerTaskBudget(tasks, input.budgetTotal, warnThreshold, abortThreshold);
|
|
1520
|
+
|
|
1521
|
+
if (budgetCheck.abort) {
|
|
1522
|
+
const message = `Per-task budget abort threshold exceeded: ${formatTokens(budgetCheck.totalUsed)}/${formatTokens(input.budgetTotal)} (${Math.round((budgetCheck.totalUsed / input.budgetTotal) * 100)}%)`;
|
|
1523
|
+
console.warn(`[team-runner] ${message}`);
|
|
1524
|
+
await appendEventAsync(manifest.eventsPath, {
|
|
1525
|
+
type: "run.budget_abort",
|
|
1526
|
+
runId: manifest.runId,
|
|
1527
|
+
message,
|
|
1528
|
+
data: {
|
|
1529
|
+
budgetTotal: input.budgetTotal,
|
|
1530
|
+
budgetUsed: budgetCheck.totalUsed,
|
|
1531
|
+
threshold: "abort",
|
|
1532
|
+
},
|
|
1533
|
+
});
|
|
1534
|
+
tasks = markBlocked(tasks, `Budget abort threshold exceeded: ${message}`);
|
|
1535
|
+
await saveRunTasksAsync(manifest, tasks);
|
|
1536
|
+
manifest = updateRunStatus(manifest, "failed", message);
|
|
1537
|
+
return { manifest, tasks };
|
|
1538
|
+
}
|
|
1539
|
+
|
|
1540
|
+
if (budgetCheck.warning) {
|
|
1541
|
+
const message = `Per-task budget warning threshold crossed: ${formatTokens(budgetCheck.totalUsed)}/${formatTokens(input.budgetTotal)} (${Math.round((budgetCheck.totalUsed / input.budgetTotal) * 100)}%)`;
|
|
1542
|
+
console.warn(`[team-runner] ${message}`);
|
|
1543
|
+
await appendEventAsync(manifest.eventsPath, {
|
|
1544
|
+
type: "run.budget_warning",
|
|
1545
|
+
runId: manifest.runId,
|
|
1546
|
+
message,
|
|
1547
|
+
data: {
|
|
1548
|
+
budgetTotal: input.budgetTotal,
|
|
1549
|
+
budgetUsed: budgetCheck.totalUsed,
|
|
1550
|
+
threshold: "warning",
|
|
1551
|
+
},
|
|
1552
|
+
});
|
|
1553
|
+
}
|
|
1554
|
+
|
|
1555
|
+
// Fair-share warning: flag tasks that consumed >50% of remaining budget
|
|
1556
|
+
// without killing them mid-execution.
|
|
1557
|
+
for (const violatorId of budgetCheck.fairShareViolators) {
|
|
1558
|
+
const violator = tasks.find((t) => t.id === violatorId);
|
|
1559
|
+
if (!violator) continue;
|
|
1560
|
+
const taskTotal = (violator.usage?.input ?? 0) + (violator.usage?.output ?? 0) + (violator.usage?.cacheWrite ?? 0);
|
|
1561
|
+
const message = `Task '${violatorId}' consumed ${formatTokens(taskTotal)} (${Math.round((taskTotal / input.budgetTotal) * 100)}% of total budget) — exceeds fair share`;
|
|
1562
|
+
console.warn(`[team-runner.fair-share] ${message}`);
|
|
1563
|
+
await appendEventAsync(manifest.eventsPath, {
|
|
1564
|
+
type: "task.budget_fair_share",
|
|
1565
|
+
runId: manifest.runId,
|
|
1566
|
+
taskId: violatorId,
|
|
1567
|
+
message,
|
|
1568
|
+
data: {
|
|
1569
|
+
budgetTotal: input.budgetTotal,
|
|
1570
|
+
taskUsage: taskTotal,
|
|
1571
|
+
},
|
|
1572
|
+
});
|
|
1573
|
+
}
|
|
1574
|
+
}
|
|
1575
|
+
|
|
1447
1576
|
const cancelledResult = results.find((item) => item.manifest.status === "cancelled");
|
|
1448
1577
|
if (cancelledResult || input.signal?.aborted) {
|
|
1449
1578
|
const reason = input.signal?.aborted ? cancellationReasonFromSignal(input.signal) : undefined;
|
|
@@ -90,8 +90,10 @@ export interface PruneResult {
|
|
|
90
90
|
// Token estimation (rough char/4 heuristic, matching gajae-code)
|
|
91
91
|
// ---------------------------------------------------------------------------
|
|
92
92
|
|
|
93
|
+
import { countTokens } from "../utils/token-counter";
|
|
94
|
+
|
|
93
95
|
function estimateTokens(text: string): number {
|
|
94
|
-
return
|
|
96
|
+
return countTokens(text);
|
|
95
97
|
}
|
|
96
98
|
|
|
97
99
|
// ---------------------------------------------------------------------------
|
|
@@ -146,7 +148,7 @@ function createPrunedNotice(tokens: number, entry: ToolResultEntry): string {
|
|
|
146
148
|
const generic = `[Output pruned — ${tokens} tokens]`;
|
|
147
149
|
const digest = resultDigest(entry.toolName, entry.content, entry.isError);
|
|
148
150
|
if (!digest) return generic;
|
|
149
|
-
const genericTokens =
|
|
151
|
+
const genericTokens = countTokens(generic);
|
|
150
152
|
const maxTokens = Math.max(genericTokens, Math.floor(genericTokens * DIGEST_NOTICE_TOKEN_CAP_MULTIPLIER));
|
|
151
153
|
const prefix = `[Output pruned — ${tokens} tokens; `;
|
|
152
154
|
const suffix = "]";
|
|
@@ -303,7 +305,7 @@ export function pruneToolOutputs(results: ToolResultEntry[], config: PruneConfig
|
|
|
303
305
|
entry,
|
|
304
306
|
tokens,
|
|
305
307
|
notice,
|
|
306
|
-
savings: Math.max(0, tokens -
|
|
308
|
+
savings: Math.max(0, tokens - countTokens(notice)),
|
|
307
309
|
});
|
|
308
310
|
accumulatedTokens += tokens;
|
|
309
311
|
}
|
package/src/state/event-log.ts
CHANGED
|
@@ -350,6 +350,11 @@ export function checkSequenceGaps(eventsPath: string): { missing: number }[] {
|
|
|
350
350
|
* from firing and degrading live-agent responsiveness.
|
|
351
351
|
*/
|
|
352
352
|
export function appendEvent(eventsPath: string, event: AppendTeamEvent): TeamEvent {
|
|
353
|
+
// NOTE: appendEvent is a sync function that uses withEventLogLockSync (sleepSync).
|
|
354
|
+
// It cannot route through appendEventBuffered because the buffer timer requires
|
|
355
|
+
// the event loop to fire, which sleepSync blocks. Both terminal and non-terminal
|
|
356
|
+
// events use the direct sync path here. For non-terminal events, callers should
|
|
357
|
+
// prefer appendEventAsync (which routes through the buffer for coalesced writes).
|
|
353
358
|
return withEventLogLockSync(eventsPath, () => appendEventInsideLock(eventsPath, event));
|
|
354
359
|
}
|
|
355
360
|
|
|
@@ -405,6 +410,19 @@ export function resetEventLogMode(): void {
|
|
|
405
410
|
* foreground-control, etc.), prefer this over the sync `appendEvent()`.
|
|
406
411
|
*/
|
|
407
412
|
export async function appendEventAsync(eventsPath: string, event: AppendTeamEvent): Promise<TeamEvent> {
|
|
413
|
+
// Non-terminal events: route through buffer for coalesced writes (20ms default).
|
|
414
|
+
// This eliminates per-event fsync + persistSequence overhead for high-frequency events.
|
|
415
|
+
if (!TERMINAL_EVENT_TYPES.has(event.type)) {
|
|
416
|
+
const result = appendEventBuffered(eventsPath, event);
|
|
417
|
+
// FIX: appendEventBuffered uses unref()'d timers. Keep the event loop alive
|
|
418
|
+
// so the promise resolves in standalone/test contexts where nothing else
|
|
419
|
+
// keeps the loop running. Without this, the event loop drains before the
|
|
420
|
+
// timer fires and the promise never resolves.
|
|
421
|
+
const keepAlive = setInterval(() => {}, 200);
|
|
422
|
+
void result.finally(() => clearInterval(keepAlive));
|
|
423
|
+
return result;
|
|
424
|
+
}
|
|
425
|
+
// Terminal events: direct async path for immediate durability guarantee
|
|
408
426
|
const queueKey = eventsPath;
|
|
409
427
|
const prev = asyncQueues.get(queueKey) ?? Promise.resolve();
|
|
410
428
|
const next = prev.then(async (): Promise<TeamEvent> => {
|
package/src/state/types.ts
CHANGED
|
@@ -208,6 +208,11 @@ export interface TeamRunManifest {
|
|
|
208
208
|
runKind?: "team-run" | "goal-loop" | "dynamic-workflow";
|
|
209
209
|
/** round-14 P1-5: typed workflow arguments accessible in .dwf.ts scripts via ctx.args<T>(). Any JSON value; default {} when unset. */
|
|
210
210
|
args?: unknown;
|
|
211
|
+
/** Per-run token budget snapshot. budgetTotal is the cap; budgetWarning/budgetAbort are fractions in [0,1] at which the team-runner emits run.budget_warning / run.budget_abort. budgetUnlimited=true opts out of enforcement. Optional for backward compat. */
|
|
212
|
+
budgetTotal?: number;
|
|
213
|
+
budgetWarning?: number;
|
|
214
|
+
budgetAbort?: number;
|
|
215
|
+
budgetUnlimited?: boolean;
|
|
211
216
|
summary?: string;
|
|
212
217
|
policyDecisions?: PolicyDecision[];
|
|
213
218
|
/** #2 (assessment): goal-achievement verdict — kills the silent false-green. */
|
|
@@ -0,0 +1,67 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Lightweight token counter for estimating token counts in text.
|
|
3
|
+
*
|
|
4
|
+
* Provides a more accurate estimate than the naive char/4 heuristic by
|
|
5
|
+
* distinguishing word characters from punctuation. Real LLM tokenizers
|
|
6
|
+
* (BPE) typically count alphanumeric runs as ~1 token per ~4 characters,
|
|
7
|
+
* while individual punctuation characters are often separate tokens.
|
|
8
|
+
*
|
|
9
|
+
* Performance: O(n) single-pass, ~1ms for 10KB text, no external deps.
|
|
10
|
+
*/
|
|
11
|
+
|
|
12
|
+
function isWhitespace(c: number): boolean {
|
|
13
|
+
return c === 0x20 || c === 0x09 || c === 0x0a || c === 0x0d;
|
|
14
|
+
}
|
|
15
|
+
|
|
16
|
+
function isAlphanumeric(c: number): boolean {
|
|
17
|
+
return (
|
|
18
|
+
(c >= 0x30 && c <= 0x39) || // 0-9
|
|
19
|
+
(c >= 0x41 && c <= 0x5a) || // A-Z
|
|
20
|
+
(c >= 0x61 && c <= 0x7a) || // a-z
|
|
21
|
+
c === 0x5f // _
|
|
22
|
+
);
|
|
23
|
+
}
|
|
24
|
+
|
|
25
|
+
/**
|
|
26
|
+
* Estimate token count for a string using a single-pass O(n) scan.
|
|
27
|
+
*
|
|
28
|
+
* Algorithm:
|
|
29
|
+
* 1. Walk text char-by-char with charCodeAt (fast, no allocations).
|
|
30
|
+
* 2. Count alphabetic chars (each ~1 token per 4 chars, like BPE prose).
|
|
31
|
+
* 3. Count punctuation chars separately (each = 1 token, since operators
|
|
32
|
+
* and punctuation typically tokenize as separate units in BPE).
|
|
33
|
+
* 4. Estimate: ceil(alpha / 4) + punct
|
|
34
|
+
*
|
|
35
|
+
* Why this beats char/4:
|
|
36
|
+
* - char/4 undercounts operators in code-heavy content (treats `=>` as ~3
|
|
37
|
+
* chars/token when BPE gives ~1-2 tokens per operator).
|
|
38
|
+
* - char/4 also miscounts short punctuation-only segments.
|
|
39
|
+
* - This formula weights punctuation at 1 token each, matching BPE's
|
|
40
|
+
* tendency to tokenize operators, brackets, and symbols individually.
|
|
41
|
+
*
|
|
42
|
+
* Accuracy: typically within ±10-15% of actual BPE token counts for both
|
|
43
|
+
* English prose and code, beating the char/4 heuristic (which can be
|
|
44
|
+
* 30%+ off for code-heavy content).
|
|
45
|
+
*
|
|
46
|
+
* @param text Input text to estimate tokens for.
|
|
47
|
+
* @returns Estimated token count.
|
|
48
|
+
*/
|
|
49
|
+
export function countTokens(text: string): number {
|
|
50
|
+
if (!text || text.length === 0) return 0;
|
|
51
|
+
|
|
52
|
+
let alpha = 0;
|
|
53
|
+
let punct = 0;
|
|
54
|
+
const len = text.length;
|
|
55
|
+
|
|
56
|
+
for (let i = 0; i < len; i++) {
|
|
57
|
+
const c = text.charCodeAt(i);
|
|
58
|
+
if (isWhitespace(c)) continue;
|
|
59
|
+
if (isAlphanumeric(c)) {
|
|
60
|
+
alpha++;
|
|
61
|
+
} else {
|
|
62
|
+
punct++;
|
|
63
|
+
}
|
|
64
|
+
}
|
|
65
|
+
|
|
66
|
+
return Math.ceil(alpha / 4) + punct;
|
|
67
|
+
}
|