🤖 perf: reuse instruction sources during stream startup

[yanncabral]

Summary

Load structured instruction sources once during stream startup and reuse that snapshot for both system-message assembly and tool-scoped instruction extraction. Existing buildSystemMessage() and readToolInstructions() wrappers remain available for callers that need load-and-build behavior.

Background

Stream startup previously loaded the same AGENTS.md/instruction sources while building the system message, then loaded them again while extracting tool instructions. Besides duplicate I/O, the two consumers could observe different file contents if instructions changed between reads.

Implementation

• add pure helpers that assemble a system message and extract tool instructions from an existing InstructionSources value
• return the loaded snapshot from buildStreamSystemContext() and reuse it for tool extraction
• thread the snapshot through policy/model-driven system-context rebuilds within the same startup
• retain the existing loading APIs and source-ordering/scoping behavior

Validation

• TDD RED: the new stream-context regression initially failed because extractToolInstructionsFromSources was not exported
• TDD GREEN: focused regression now verifies one snapshot produces both the expected prompt content and bash tool instructions
• bun test src/node/services/systemMessage.test.ts src/node/services/streamContextBuilder.test.ts src/node/services/aiService.test.ts --timeout 30000 — 150 passed
• targeted ESLint on all five changed files — passed (split into two invocations after the combined process was OOM-killed)
• bun run tsgo --noEmit -p tsconfig.main.json — passed
• Prettier check/write and git diff --check — passed
• full make typecheck attempted: main-process typecheck passed, broad renderer typecheck was SIGKILL/OOM

Risks

The change touches startup context plumbing. Regression coverage retains the existing system-message matrix for sub-project, multi-project, global/compat, local, model/mode, Xum-only, and tool-scoped semantics, and asserts identity reuse at the AI service boundary.

---

Generated with xum • Model: openai:gpt-5.6-sol • Thinking: medium • Cost: unknown