sorokin.oleg.edu/learnharness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 6 Wiki Activity

fix: align compaction cuts and native usage accounting

Browse Source
This commit is contained in:
amosblomqvist
2026-10-24 19:05:02 +03:00
parent 50a2db775d
commit 624c91a437
45 changed files with 775 additions and 2329 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/core/compaction.ts
+62 -51
View File
@@ -1,7 +1,8 @@
/**
* COMPACTION (owner round 14, 2026-10-05): the context-fitting mechanic,
* mirroring pi's design (docs aside, this is OUR implementation — pi's
* compaction lives in the coding-agent package we don't depend on).
* with a bounded local port of Pi coding-agent 1.4.0 cutoff selection.
* No coding-agent dependency. See docs/compaction-accounting.md for the
* behavior boundary, native golden fixtures, and upstream MIT notice.
*
* THE CONTRACT: full history stays persisted in the record — extraction
* and the user's transcript keep reading everything. A compaction marker
@@ -38,7 +39,7 @@ import { COMPACTED_HISTORY_HEADER } from "./contracts/chat.ts";
import type { LhMessage, LhMarkerMessage } from "./native-messages.ts";
import { newMessageId } from "./native-messages.ts";
import type { AgentMessage } from "@earendil-works/pi-agent-core";
import type { AssistantMessage, ToolResultMessage } from "@earendil-works/pi-ai";
import type { AssistantMessage, ToolResultMessage, SystemMessage } from "@earendil-works/pi-ai";
// ─── constants (the tunables, owner-approved) ────────────────────────────────
@@ -77,9 +78,8 @@ export function calculateContextTokens(usage: {
);
}
/** Estimate a message's context tokens from its WIRE-relevant text only
* (pi's estimator): text + thinking + toolCall name/args for assistants,
* content text for user/toolResult/custom. NEVER JSON.stringify of the
/** Pi 1.4.0's heuristic: text + thinking + toolCall name/args for
* assistants, content text plus a fixed image allowance for other roles. NEVER JSON.stringify of the
* whole message — record scaffolding (ids, usage, responseId, app-side
* details, escaped quotes/backslashes) inflates ~3x (measured on the
* 2026-10-05 relational-databases incident: 84k estimated vs ~30k real). */
@@ -100,23 +100,32 @@ export function estimateMessageTokens(message: LhMessage | AgentMessage): number
}
} else {
chars += textCharsOfContent(content);
if (message.role === "system") {
const system = message as SystemMessage;
for (const section of Object.values(system.sections ?? {})) chars += section?.length ?? 0;
if (system.toolsAdded) chars += JSON.stringify(system.toolsAdded).length;
}
}
return Math.ceil(chars / 4);
}
// Pi coding-agent 1.4.0 estimateTokens: fixed image allowance, NOT a
// measurement of encrypted reasoning or actual provider vision tokens.
const ESTIMATED_IMAGE_CHARS = 4_800;
function textCharsOfContent(content: unknown): number {
if (typeof content === "string") return content.length;
if (!Array.isArray(content)) return 0;
let chars = 0;
for (const block of content as Array<string | { type?: string; text?: string }>) {
if (typeof block === "string") chars += block.length;
else if (block?.type === "text" || "text" in block) chars += block.text?.length ?? 0;
else if (block?.type === "image") chars += ESTIMATED_IMAGE_CHARS;
else if (block?.type === "text") chars += block.text?.length ?? 0;
}
return chars;
}
/** The newest assistant message carrying VALID provider-reported usage
* (pi's rule: skip aborted messages and all-zero usage — they carry no
* (pi's rule: skip aborted/error messages and all-zero usage — they carry no
* truth about the wire). The scan STOPS at the newest compaction
* marker: usage captured before a compaction described the pre-cut
* context — the next request carries the summary + kept tail instead,
@@ -131,7 +140,7 @@ export function lastAssistantUsageIndex(
if ((m as { role?: string }).role === "custom" && (m as LhMarkerMessage).customType === "compaction") return null;
if ((m as { role?: string }).role !== "assistant") continue;
const a = m as AssistantMessage & { stopReason?: string };
if (a.stopReason === "aborted") continue;
if (a.stopReason === "aborted" || a.stopReason === "error") continue;
if (calculateContextTokens((a.usage ?? {}) as Parameters<typeof calculateContextTokens>[0]) === 0) continue;
return i;
}
@@ -180,51 +189,53 @@ export function estimateTokens(messages: readonly (LhMessage | AgentMessage)[]):
// ─── the cut walk ────────────────────────────────────────────────────────────
/**
* Find the compaction cut index: walk BACKWARDS from the end accumulating
* estimates until ~keepRecentTokens are kept; the cut lands there.
*
* HARD INVARIANT — a tool result is never separated from its call: if the
* walk stops on a toolResult, the KEPT region extends to the assistant
* message carrying its call (keeping the result means keeping the call).
* This is the provider protocol's shape law: dropping a call while
* keeping its result strands an orphan the provider rejects. A PENDING
* question/problem call (no result yet — the pose→answer window) sits at
* the transcript end and is always kept: the walk never crosses it.
*
* Returns 0 when everything must be summarized (the whole history is the
* span — valid), or messages.length when there is nothing meaningful to
* cut (kept budget covers everything — caller checks before compacting).
*/
export function compactionCutIndex(messages: readonly LhMessage[], keepRecentTokens = KEEP_RECENT_TOKENS): number {
let kept = 0;
for (let i = messages.length - 1; i >= 0; i--) {
if (kept >= keepRecentTokens) {
// The cut candidate: extend past orphaned tool results to the
// assistant message carrying their calls.
let cut = i + 1;
while (cut < messages.length && (messages[cut] as { role?: string }).role === "toolResult") {
const callIndex = assistantIndexOfCall(messages, (messages[cut] as ToolResultMessage).toolCallId, cut);
if (callIndex === -1) break; // no call found (defensive): stop extending
cut = callIndex;
}
return cut;
}
kept += estimateMessageTokens(messages[i]!);
}
return 0; // the whole history fits inside the kept budget — summarize all
/** Model-visible entry adapter for the local Pi 1.4.0 cutoff port.
* Compaction entries are boundaries, not ordinary cut candidates; UI-only
* markers have no context, just like native metadata entries. */
function cutMessage(message: LhMessage): LhMessage | null {
if (message.role !== "custom") return message;
if (["compaction", "plan-block", "problem-dismissed"].includes(message.customType)) return null;
return message;
}
/** The newest assistant message at or before `before` whose content
* carries the tool call `toolCallId` — the call a result answers. */
function assistantIndexOfCall(messages: readonly LhMessage[], toolCallId: string, before: number): number {
for (let i = before - 1; i >= 0; i--) {
const m = messages[i]!;
if ((m as { role?: string }).role !== "assistant") continue;
const assistant = m as AssistantMessage;
if (assistant.content.some((b) => b.type === "toolCall" && b.id === toolCallId)) return i;
/** Local port of pi-coding-agent 1.4.0 findCutPoint (MIT).
* Walk backwards to >= budget, then choose the nearest valid cut AT OR
* AFTER that entry. If oversized trailing results exhaust the budget,
* retain their preceding assistant call, not the entire session.
* No encrypted-signature sizing and no exact retained-token guarantee.
* Returns the first kept record index; a zero/no-progress cut drops nothing.
* Keep this and estimateMessageTokens aligned deliberately when upgrading Pi.
*/
export function compactionCutIndex(messages: readonly LhMessage[], keepRecentTokens = KEEP_RECENT_TOKENS): number {
const previous = activeCompactionOf(messages);
const previousKept = previous ? keptIndexFor(messages, previous) : -1;
const start = Math.max(0, previousKept);
const points: number[] = [];
for (let i = start; i < messages.length; i++) {
const message = cutMessage(messages[i]!);
if (message && (message.role === "user" || message.role === "assistant" || message.role === "custom")) points.push(i);
}
return -1;
if (!points.length) return start;
let cut = points[0]!;
let accumulated = 0;
for (let i = messages.length - 1; i >= start; i--) {
const message = cutMessage(messages[i]!);
const tokens = message ? estimateMessageTokens(message) : 0;
if (tokens === 0) continue;
accumulated += tokens;
if (accumulated >= keepRecentTokens) {
cut = points.find((candidate) => candidate >= i) ?? points.at(-1)!;
break;
}
}
// Native entries preserve adjacent non-context metadata, but never walk
// backwards across a previous compaction entry.
while (cut > start) {
const before = messages[cut - 1]!;
if ((before.role === "custom" && before.customType === "compaction") || cutMessage(before)) break;
cut--;
}
return cut;
}
// ─── the marker ──────────────────────────────────────────────────────────────