fix(core): leave headroom when fitting output to the context window - #59934

Issue for this PR

Closes #58463

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Once the prompt is large enough, outputLimit lowers max_tokens to exactly the room left in the window: context - measured - estimated * 1.15. Our count runs a little below the provider's (chat template tokens, tool schemas), so the request can overshoot the window by a few tokens, and strict providers reject it. The issue thread has real examples: Kimi K2.6 through OpenRouter ("maximum context length is 262144 tokens. However, you requested about 262223") and DeepSeek (6 tokens over).

This is easy to hit on models whose output limit is close to their window. Kimi K2.6 is 262k context / 236k output, so the output limit starts being fitted from about a 26k prompt, and every request after that sits right at the edge.

The fix leaves 1% of the window unused when a normal turn's output limit is fitted (2.6k on Kimi, about 10k on a 1M model). Requests that aren't up against the window are unchanged.

Compaction requests are left as they are. The summary is sized against its own reserve, and the existing tests pin that it gets the whole reserve, so I kept this change to normal turns.

The issue also says the compaction ceiling is above the request ceiling on 1M models. max_tokens already shrinks past context - 256k, so that alone doesn't fail a request. The real failure is the missing headroom, which is what this changes.

How did you verify your code works?

• Updated fits the output limit to the prompt size in session-runner.test.ts so it checks for the 1% gap through the real runner path.
• bun test test/session-runner.test.ts test/session-compaction.test.ts test/session-generate.test.ts test/session-model-request.test.ts from packages/core: 282 pass.
• bun typecheck in packages/core passes.

Visual

<svg xmlns="http://www.w3.org/2000/svg" width="240" height="72" viewBox="0 0 240 72" fill="none">
  <style>
    .p{animation:p 4s ease-in-out infinite}
    .o{animation:o 4s ease-in-out infinite}
    @keyframes p{0%,15%{width:96px}55%,85%{width:150px}100%{width:96px}}
    @keyframes o{0%,15%{x:116px;width:96px}55%,85%{x:170px;width:42px}100%{x:116px;width:96px}}
    @media (prefers-reduced-motion:reduce){.p,.o{animation:none}}
  </style>
  <rect x="20" y="24" width="200" height="24" rx="5" stroke="currentColor" stroke-opacity=".35"/>
  <rect class="p" x="20" y="24" width="150" height="24" rx="5" fill="currentColor" fill-opacity=".55"/>
  <rect class="o" x="170" y="24" width="42" height="24" rx="3" fill="currentColor" fill-opacity=".22"/>
  <line x1="212" y1="18" x2="212" y2="54" stroke="currentColor" stroke-opacity=".6" stroke-dasharray="2 2"/>
  <text x="20" y="64" font-family="system-ui,sans-serif" font-size="9" fill="currentColor" fill-opacity=".6">prompt</text>
  <text x="200" y="14" font-family="system-ui,sans-serif" font-size="9" fill="currentColor" fill-opacity=".6">headroom</text>
</svg>

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

Added by the opencode bot (AI-generated).

Related: #50574