Bad default: engine falls back to 128K token budget for model

Summary

When a routed model has no capability limits installed (or reports a zero context window), the agent engine silently falls back to a hardcoded 128,000-token prompt budget and drives context compaction against it. For large-context models (e.g. 1M-token Anthropic models) whose ids don't match a host-provided capabilities entry, this makes compaction fire ~8x too early (at ~102K tokens instead of ~800K+), causing avoidable summarization churn, latency, token cost, and quality loss.

We hit this integrating the SDK/engine into a host app. We can work around it host-side by always installing capabilities, but the engine's default is not sane and the fallback operator makes it worse.

Where it lives

In the bundled engine (@github/copilot, observed in 1.0.63 and 1.0.73; app.js, minified). The relevant compiled forms:

// module init
sen = .8, aen = .95, Nk = 128e3;            // Nk == 128000
class LM { static DEFAULT_TOKEN_LIMIT = Nk; /* ... */ }

// CompactionProcessor.preRequest
s = r.capabilities?.limits?.max_prompt_tokens
    || r.capabilities?.limits?.max_context_window_tokens
    || Nk;                                   // <-- falsy fallback to 128000
u = promptTokens + toolTokens;
d = u / s;                                   // utilization
// compaction triggers when d >= 0.8 (background) / 0.95 (buffer exhaustion)

The same ... || Nk fallback is repeated in the session.usage_info emit, contextInfo, and getTokenLimits paths.

Two distinct problems

1. Falsy fallback (||) instead of nullish (??).
   A model that reports max_context_window_tokens: 0 (a legitimately "unknown" signal) collapses to 128000 rather than being treated as unknown. This is compounded upstream: the @github/copilot-sdk client's models.list handler backfills missing limits with { max_context_window_tokens: 0 }, so an un-capped model arrives at the engine with 0 and || Nk turns it into 128K. 0 should not be coerced to the default via a truthiness check.

2. No model-aware default.
   Every un-registered model — including known large-window models — gets the same 128K budget. There is no per-family/default table and no way to distinguish "small model, 128K is right" from "1M-window model, 128K is catastrophically low."

Impact

• Large-context models compact at ~0.8 * 128000 ≈ 102K tokens regardless of their true window.
• Symptoms: premature/repeated context compaction, extra summarization round-trips, higher token spend and latency, degraded answer quality on long tasks.
• Silent: nothing in the default event payload surfaces the effective window that was used, so the "capped at 128K" cause has to be inferred from the model id. (We had to add host-side telemetry — reconstructing max_prompt_tokens ?? max_context_window_tokens ?? 128000 — to see it.)

Steps to reproduce

1. Create a session with a large-context model whose id is not matched by any host-installed capabilities entry (so no limits reach the engine, or they arrive as max_context_window_tokens: 0).
2. Send a turn whose prompt+tool tokens exceed ~102K but are well under the model's real window (e.g. 300K on a 1M-window model).
3. Observe session.compaction_start firing even though the real window is nowhere near exhausted.

Expected

• 0 / missing limits should be treated as "unknown," not coerced to 128K (use ??, or validate > 0).
• Provide a sane, model-aware default (or at minimum a host-configurable default budget) so large-window models are not capped at 128K.
• Surface the effective token limit the engine used in the compaction / usage_info telemetry so the applied budget is observable without host-side reconstruction.

Suggested fixes

• Change the fallback chain from a || b || Nk to nullish/> 0 validation so a real 0 isn't silently replaced.
• Add a model-aware default table, or accept an explicit host-supplied default token budget on session config.
• Emit the resolved effective prompt-token limit alongside compaction and session.usage_info events.

Environment

• Engine: @github/copilot 1.0.63 (also reproduced against 1.0.73 via @github/copilot-next).
• SDK: @github/copilot-sdk 1.0.0 (zero-window backfill also present in 1.0.7).

Related

The SDK-side zero-window backfill (@github/copilot-sdk models.list handler) contributes to problem (1). Happy to cross-file there if the SDK is the preferred owner for that half.

Commencez par la app.js fournie et suivez CompactionProcessor.preRequest, le fallback répété dans session.usage_info, contextInfo et getTokenLimits, ainsi que le handler models.list du SDK. Vérifiez le comportement avec une limite nulle ou manquante à l’aide des étapes de reproduction, puis définissez comment la limite effective de tokens du prompt doit être résolue et exposée dans compaction et la télémétrie d’utilisation.