Skills descriptions bloat system prompt (~600KB / ~150K tokens), breaking Bedrock and small-context models

[eugenio]

Problem

OpenCode injects ALL installed skill descriptions (name + description XML blocks) into the system prompt on every request, regardless of whether they're needed. With ~1,500 skills from opencode-skills-antigravity, this adds ~580KB / ~147K tokens to every system prompt.

This causes:

• BadRequestError on AWS Bedrock models (Kimi K2.5, Qwen3 480B, Qwen3 30B) when the total prompt exceeds their 262K token context window
• 56% of the context window consumed before the user even types anything
• All models in the fallback chain exhausted, resulting in "All configured models are currently unavailable"
• Multi-turn conversations with tool results fail even faster as context accumulates

Evidence

From NadirClaw raw request logs:

• system_prompt_length: 595,269 chars (595KB)
• OpenCode's own instructions: ~9KB
• Skill descriptions: ~580KB (1,539 <skill> XML blocks)
• CLAUDE.md user config: ~5KB
• Approximate tokens: ~148,817 (56% of 262K context window)

Expected behavior

Skill descriptions should NOT be injected into the system prompt. Skills should be loaded on-demand via the skill tool, with only a lightweight index (skill names) available for discovery.

Environment

• OpenCode with opencode-skills-antigravity plugin (~1,330 skills)
• NadirClaw proxy routing to AWS Bedrock (eu-west-2)
• Models: Kimi K2.5, Qwen3 480B, Qwen3 30B (all 262K context)
• LiteLLM 1.83.0

Workaround

Move skills out of ~/.config/opencode/skills/ to a directory that OpenCode's built-in scanner doesn't check. Use @zenobius/opencode-skillful plugin for lazy discovery from the alternate path.

Related issues

• Large AGENTS.md files consume entire context window, triggering immediate compaction #18037 — Large AGENTS.md files consume entire context window
• Prompt too long when using Claude Opus 4.6 (Global) Amazon Bedrock #12688 — Prompt too long when using Bedrock
• Excessive token consumption since ~Jan 12, 2026 - apparently sending entire project context on every request #8234 — Excessive token consumption, sending entire project context

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• [FEATURE]: Lazy-load agent/skill lists instead of inlining into tool descriptions #13188: Describes the same root problem — skill and agent descriptions being statically inlined into tool descriptions on every turn, with the same proposed solution of lazy/on-demand loading. That issue includes code-level analysis of skill.ts and task.ts, and proposes a two-step discovery pattern identical to what you're requesting.
• Prompt too long when using Claude Opus 4.6 (Global) Amazon Bedrock #12688: Prompt too long on AWS Bedrock (referenced in your issue body, but worth confirming as the same class of issue).
• [FEATURE]: ToolSearch for lazy/deferred tool discovery to reduce per-turn context size #20489: Related feature request for deferred tool discovery to reduce per-turn context size.

If your issue is specifically about the scale of the problem (1,500+ skills from a large skill pack exhausting the context window entirely vs. the more modest overhead described in #13188), it may be worth adding a comment there with your evidence and environment details rather than tracking it separately.

[eugenio]

Closing as duplicate of #13188 — added evidence and measurements there.