Summary

• Adds speculative upgrade probing for models that support higher context on premium plans (e.g., Claude Opus 4.6: 200K default → 1M on Anthropic Max)
• Default context remains 200K — Max/Team/Enterprise users are automatically detected and upgraded to 1M without any configuration
• Standard-plan users experience no disruption — the probe is transparent and the result is cached for future sessions

Why not just hardcode 1M?

The 1M context window for Claude 4.6 is not universal — it depends on the API plan:

See also #1820 which proposes hardcoding 1M.

How it works

1. When the compression threshold (50% of 200K = 100K) is first reached, the compressor speculatively raises the context limit to the upgrade tier (1M) instead of compressing
2. If the next API call succeeds past the old threshold → premium tier confirmed, 1M cached
3. If the API call fails with a context error → standard tier confirmed, reverts to 200K, compresses normally, 200K cached
4. Future sessions load from cache — no re-probing

Files changed

Design decisions

• Probe at compression threshold, not session start: avoids wasted API calls for short sessions
• Cache suppresses re-probing: get_cached_context_length check prevents standard-plan users from re-probing every session
• Fuzzy match only checks key in model (not reverse): prevents claude-opus-4 from falsely matching claude-opus-4-6's upgrade tier
• Error handler uses separate branch for probe revert: avoids incorrect step-down (200K → 128K) that would occur if the parsed limit equals the pre-upgrade default

Adding new models

One line in UPGRADE_CONTEXT_TIERS:

UPGRADE_CONTEXT_TIERS = {
    "anthropic/claude-opus-4.6": 1_000_000,
    "claude-opus-4-6": 1_000_000,
    # Add new models here
}

Related

• Supersedes feat(agent): set claude-sonnet-4.6 and claude-opus-4.6 context to 1M tokens #1820 (hardcoding 1M breaks non-Max users)
• Works well with fix: auto-invalidate stale context length cache when defaults change #1852 (cache invalidation when defaults change) — independent but complementary

Test plan

• 118 tests passing (model_metadata + context_compressor + context overflow)
• Manual test: verify Max-plan user sees 1M context cached after first long session
• Manual test: verify standard-plan user gets transparent compression at ~200K