bug(runtime): history compaction keeps earlier oversized tool results in a multi-result tool message

[liuxiaocs7]

What happened

fitHistoryCompactMessages (packages/runtime/src/history-compact-input-fit.ts) bounds a compaction request by replacing old, oversized tool-result payloads with a placeholder, keeping every call/result pair intact.

When a single tool message carries more than one tool-result part — which happens whenever the model issues parallel tool calls, so the provider batches their results into one message — only the last replaced part survives. Each replacement copies from the original message.content:

const content = [...message.content]; // line 65 — always the ORIGINAL parts
content[partIndex] = replacement;
bounded[messageIndex] = { ...message, content };

so an earlier part's placeholder is overwritten and its payload stays at full size in the returned history — even though the budget accounting (estimatedChars) has already been decremented for it.

Result: the function can return a history that still far exceeds maxInputEstimatedTokens while treating it as "fit", so the downstream compaction / recap model call is sent over its intended input budget (risking a provider context-window error). With other sizes the same accounting mismatch instead throws input_too_large for input that would actually fit. Either way the helper fails at exactly the case it exists to handle.

The sibling active-tool-result-prune.ts handles the same multi-part shape correctly by accumulating replacements (nextContent = originalContent.slice(0, index) then push), so multi-result tool messages are a known, supported shape.

How to reproduce

Build first (npm run build), then run this against the compiled function on main:

import { fitHistoryCompactMessages } from './packages/runtime/dist/history-compact-input-fit.js';
import { stableJsonLength } from './packages/runtime/dist/context-budget-helpers.js';

const big = 'raw-tool-output-'.repeat(1024); // ~16 KB each
const messages = [
  { role: 'assistant', content: [
    { type: 'tool-call', toolCallId: 'call-1', toolName: 'shell', input: { command: 'a' } },
    { type: 'tool-call', toolCallId: 'call-2', toolName: 'shell', input: { command: 'b' } },
  ]},
  { role: 'tool', content: [ // provider batches BOTH parallel results into one message
    { type: 'tool-result', toolCallId: 'call-1', toolName: 'shell', output: { type: 'text', value: big } },
    { type: 'tool-result', toolCallId: 'call-2', toolName: 'shell', output: { type: 'text', value: big } },
  ]},
  { role: 'assistant', content: [{ type: 'text', text: 'done' }] },
];

const bounded = fitHistoryCompactMessages(messages, { maxInputEstimatedTokens: 1000, charsPerToken: 1 });
console.log('returned "fit" chars:', stableJsonLength(bounded), '(budget 1000)');
console.log('full payloads left  :', JSON.stringify(bounded).split(big).length - 1, '(expected 0)');

Observed on main:

returned "fit" chars: 16974 (budget 1000)
full payloads left  : 1 (expected 0)

Expected: both payloads replaced, result within the requested budget.

Environment

• Maka commit: origin/main 920d7142d
• OS and version: macOS 15 (arm64)
• Surface: Runtime
• Node.js: 24

Logs, screenshots, or additional context

• Buggy line: packages/runtime/src/history-compact-input-fit.ts:65 copies from the original message.content on every replacement instead of carrying earlier replacements forward.
• Affected callers: history-compact-summarizer.ts, session-recap.ts, openai-codex-history-compactor.ts.
• The loop has been unchanged since it was introduced in fix(runtime): bound text history summarizer inputs #3113.

[liuxiaocs7]

take