fix(hyperagent): default 1M context for fable/opus/sonnet

[artickc]

Summary

Follow-up related to HyperAgent (#7994 / sticky tool-loop work).

getTokenLimit('hyperagent', 'fable-latest') was falling through to the generic 128k default. Real agentic tool-loop prompts (catalog + history) easily exceed that and fail with:

Input exceeds the context window for hyperagent/fable: estimated 137042 input tokens, limit 128000

Claude-family agents on HyperAgent (Fable / Opus / Sonnet) should use a 1M context window by default.

Changes

• hyperagent registry: defaultContextLength: 1_000_000 + per-model contextLength
• hyperagentModels.ts: HYPERAGENT_DEFAULT_CONTEXT_LENGTH = 1_000_000 on catalog models
• contextManager.ts: resolve hyperagent/ha (and fable/opus wire ids) to 1M before models.dev DB / generic default

Tests

node --import tsx -e "import { getTokenLimit } from './open-sse/services/contextManager.ts'; console.log(getTokenLimit('hyperagent','fable-latest'))"
# → 1000000

node --import tsx --test tests/unit/context-manager.test.ts tests/unit/service-context-manager.test.ts
# 30/30 PASS

Risk

Low — only raises the pre-flight token budget for HyperAgent; does not change chat protocol. Env override CONTEXT_LENGTH_HYPERAGENT still wins if set.

[diegosouzapw]

Thanks for chasing this down — getTokenLimit('hyperagent', 'fable-latest') really was falling through to the generic 128k default, and the registry-based fix (defaultContextLength on the provider entry + per-model contextLength in hyperagentModels.ts) is exactly the right pattern; it matches how dozens of other providers (windsurf, devin, github, etc.) already declare per-model context windows.

One thing we need adjusted before merge: the new block added at the top of resolveTokenLimit in contextManager.ts checks model.toLowerCase().includes(...) for strings like claude-opus-4, claude-fable, sonnet-latest, claude-sonnet-5 — but that check isn't scoped to the hyperagent provider, so it also fires for any other provider whose model id happens to contain those substrings (we verified this hits windsurf, anthropic, github, ghe-copilot, bedrock and more). Since it runs before the existing models.dev/registry per-model lookup, it silently overrides those providers' correctly-configured (smaller) context windows with 1M.

We tested and confirmed that the registry-only part of your change (hyperagent/index.ts + hyperagentModels.ts) is actually sufficient on its own to fix getTokenLimit for both hyperagent and its ha alias — no changes to contextManager.ts are needed at all. Could you drop the new p === "hyperagent" || p === "ha" / model-substring block from contextManager.ts (or, if you'd like to keep an explicit short-circuit there for clarity, scope it strictly to the provider check and drop the substring matching), and add a small regression test asserting getTokenLimit('hyperagent'|'ha', 'fable-latest') === 1_000_000? Happy to help iterate on this in your branch.

[diegosouzapw]

Thanks @artickc — merged into release/v3.8.49 via the local merge-train (validated as one combined tree: full test:unit + test:vitest 274/274 on the 32-core box, tip d4b9ce6016). Your commit keeps its authorship. 🚀