fix(opencode): use native instructions config to load CLAUDE.md and fragments

[glifocat]

Type of Change

• Feature skill - adds a channel or integration (source code changes + SKILL.md)
• Utility skill - adds a standalone tool (code files in .claude/skills/<name>/, no source changes)
• Operational/container skill - adds a workflow or agent skill (SKILL.md only, no source changes)
• Fix - bug fix or security fix to source code
• Simplification - reduces or simplifies source code
• Documentation - docs, README, or CONTRIBUTING changes only

Description

Closes #2150.

wrapPromptWithContext previously concatenated /workspace/agent/CLAUDE.md verbatim into a <system>...</system> block in the user-message text. That file is the host-composed entry point that contains only @./... includes (the .claude-shared.md symlink to /app/CLAUDE.md, plus the module fragments). OpenCode does not expand @ in instruction files — the syntax is a Claude Code convention for the model's own Read tool to lazy-load. Result: the model received the literal lines @./.claude-shared.md\n@./.claude-fragments/module-...md as text and saw none of the actual fragment content (workspace, memory, conversation history, agents/core/interactive/scheduling/self-mod modules, OneCLI, etc.). Strong-prior models (Claude, GPT-4) masked the gap; weaker local models visibly lacked NanoClaw conventions.

This PR routes the concrete files through OpenCode's native instructions config field. Per packages/opencode/src/session/instruction.ts on the upstream dev branch (already shipped in @opencode-ai/sdk@1.4.17, the version pinned by /add-opencode), absolute paths and globs are resolved with fs.glob(basename, { cwd: dirname, absolute: true, include: 'file' }), files are read raw, and the resulting strings are concatenated into the LLM's system prompt at packages/opencode/src/session/prompt.ts:1442. That is the canonical channel — same one OpenCode uses for AGENTS.md / CLAUDE.md auto-discovery.

Configured set

const instructions = [
  '/app/CLAUDE.md',                              // shared base
  '/workspace/agent/.claude-fragments/*.md',     // per-skill fragments
  '/workspace/agent/CLAUDE.local.md',            // per-group memory
];
if (process.env.NANOCLAW_IS_MAIN !== '1') {
  instructions.push('/workspace/global/CLAUDE.md');  // cross-group memory (non-main)
}

Removed: readClaudeMdForPrompt, the manual <system> wrap of its output in wrapPromptWithContext, and the now-unused fs import. The dynamic systemInstructions wrap (assistant name + destinations) stays as-is — that's per-call data, not file content.

Verification

Empirically tested with gemma4:31b on local Ollama before the fix: agent replied "I haven't received instructions" to "what self-modification tools do you have". Post-fix: agent listed the exact MCP tool names (install_packages, add_mcp_server) and structured response from module-self-mod.md content. Base model has no priors for those tools, so the only place that knowledge can come from is the fragment now reaching the system prompt.

Note on composeGroupClaudeMd

src/claude-md-compose.ts still produces /workspace/agent/CLAUDE.md with @./... includes. That file remains required for the Claude provider (whose SDK does expand @ correctly) but is now redundant for OpenCode users — OpenCode also auto-discovers it via findUp and adds the literal text again, but instructions takes precedence for actual content. Cleanup is out of scope for this fix; flagging so future work knows.

Related

• Follow-up to #1843: latency improvements available + branch target question #1955 raised a perf-driven angle on instructions duplication ("similar duplication may exist for Claude/OpenCode — worth investigating"). This PR fixes the inverse symptom (silent context loss); both efforts are complementary.

For Skills

• SKILL.md contains instructions, not inline code (code goes in separate files)
• SKILL.md is under 500 lines
• I tested this skill on a fresh clone

Not a skill PR — section N/A.