Pin OpenCode context limits for gateway Claude and OSS models

[mats16]

Fixes #905

What

ug opencode now gives OpenCode the real context windows of the gateway models, plus a limit.input that makes auto-compaction start at about 90% of the window. Token limits are applied to every managed OpenCode provider (databricks-anthropic, databricks-google, databricks-oss), not only OSS.

Model	limit after (context / input / output)	Auto-compaction before	Auto-compaction after (default buffer)
claude-opus-* / claude-sonnet-* 4.6+ (incl. 5.5)	1,000,000 / 900,000 / 32,000	180,000 (OpenCode fallback)	880,000
claude-haiku-4-5	200,000 / 180,000 / 32,000	180,000 (fallback)	160,000
glm-5-*	1,048,576 / 943,718 / 32,000	175,000 (ug GLM-4.6 pin)	923,718
deepseek-v4*, kimi-k3*	1,048,576 / 943,718 / 32,000	180,000 (fallback)	923,718
glm-4-*	200,000 / 180,000 / 25,000	175,000 (ug pin)	160,000
Gemini and other unverified models	not pinned	180,000 (fallback)	unchanged

Why

OpenCode has no catalog entry for the gateway's system.ai.* model ids. Any model without a configured limit falls back to OpenCode's hardcoded {context: 200000, output: 32000}, and that fallback can't be changed through config. As a result, sessions on 1M-context models auto-compacted at about 180K tokens instead of near the real window. The only pin ug wrote was a stale GLM-4.6 {200000, 25000} entry, and it also matched GLM-5.x.

Pinning context alone is not enough to get a sensible threshold. OpenCode 2.0.14 auto-compacts once usage reaches:

min(limit.input - buffer, limit.context - max(min(limit.output, 32000), buffer))

buffer is the compaction buffer (20,000 by default, configurable). Without limit.input, a 1M-context model compacts only at 968,000 tokens (~97%). limit.input is read only by this check, so setting it changes when compaction starts and nothing else.

How

• databricks.py
  • _MODEL_TOKEN_LIMITS is keyed by model-id substring, and the first match wins. The single "glm" entry is replaced by glm-5, deepseek-v4 and kimi-k3 (1,048,576), glm-4 (the existing 200K/25K pin), and claude-haiku-4-5 (200K).
  • New claude_has_1m_context() covers Claude Opus/Sonnet 4.6+ by version comparison. model_token_limits() returns {1000000, 32000} for those models.
• agents/opencode.py
  • _oss_model_overlay becomes _model_overlay(model, overlay), and render_overlay uses it for all three providers. Each Anthropic model now gets its own entry dict, where before they all shared one dict via dict.fromkeys. The per-model headers and toolStreaming: false are unchanged.
  • _model_overlay sets limit.input to 90% of limit.context. This is a flat ratio and is not offset by OpenCode's buffer, so a user-configured buffer still applies on top of it. It lives in the OpenCode agent, not in the shared model_token_limits table, because it encodes OpenCode's compaction policy rather than a gateway limit.
• agents/claude.py: _maybe_add_1m_suffix now uses claude_has_1m_context, so Claude Code's [1m] suffix and OpenCode's 1M limit share one rule. Behavior is unchanged.
• Output is 32,000 because OpenCode clamps max_tokens to min(limit.output, OUTPUT_TOKEN_MAX = 32000), so a higher value would change nothing by default. It is also below every listed gateway cap from OpenCode: gateway models fall back to OpenCode's 200K context (Claude 5.5 and OSS are ~1M) #905 (GLM-5.x 131,072; DeepSeek V4 flash 384,000; Kimi K3 1,048,576; Claude Opus 5.5 128,000).

Not changed:

• Gemini stays unpinned because its gateway limits have not been verified (OpenCode: gateway models fall back to OpenCode's 200K context (Claude 5.5 and OSS are ~1M) #905 notes the same). Adding a _MODEL_TOKEN_LIMITS entry is enough to pin it later, since render_overlay now applies limits to the Google provider too.
• Model discovery (/api/2.1/unity-catalog/model-services) only returns names, so the limits stay in a table.
• 200K-context models compact at 80% with the default buffer. Raising that would need a lower output, which is out of scope here.

Testing

• uv run --frozen pytest tests/test_agent_opencode.py tests/test_databricks.py tests/test_agent_claude.py: 640 passed (after merging the latest main). New and updated cases cover:
  • Claude Opus/Sonnet 5.5 and 4.6 (both system.ai. and databricks- forms), and Haiku 4.5.
  • GLM-5.x, DeepSeek V4 flash and Kimi K3.
  • limit.input at 90% of the context for every pinned model.
  • The GLM-4.x pin kept.
  • Models without known limits (Sonnet 4.5, Fable, Kimi K2, Gemini) left unpinned.
  • Per-model toolStreaming still set on every Anthropic entry.
• uv run --frozen pytest: 2965 passed, 40 skipped, 5 failed on the first revision of this PR, before limit.input and before merging main. The same 5 fail on main in this environment, so they are unrelated to this change: 2 in tests/test_claude_smart_routing_v2.py and 3 in tests/test_e2e_user_agent.py. The full suite has not been re-run since.
• ruff check src/ tests/, ruff format --check src/ tests/, and ty check src/ pass.
• Manual check with OpenCode 2.0.14:
  • Read the auto-compaction check and the model fallback from the installed binary to confirm the formula above, and that limit.input is optional and read only by compaction.
  • Rendered the overlay and ran opencode serve with isolated XDG_*/HOME directories. /api/model reported the pinned context/output for every model above, and limit.input was passed through as configured. This was checked with an earlier input formula (90% of context plus the 20K default buffer); the final flat-90% values were checked by unit tests only.
  • No model requests were sent.

Environment: macOS arm64, OpenCode 2.0.14 (@opencode/cli).

This pull request and its description were written by Isaac.