Summary

The system prompt had grown to 24,387 bytes (≈ 6,100 tokens) — every byte paid on every request. Much of it overlapped with tool descriptions sitting right next to it in the cache prefix: "When to propose a plan", "When to ask the user to pick", and "When to track multi-step intent" each recited rules the tool's own description already carried.

Aggressive dedup pass:

• Drop the redundant "you have these filesystem tools" opening sentence — the API ships the tool list separately.
• Merge the three independent submit_plan / ask_choice / todo_write sections into one short "Picking the right tool" block.
• Fold "Exploration", "Trust what you already know", and "When the user wants to switch project" into shorter equivalents — same rules, no narrative.
• Collapse the foreground/background section — full how-to lives in the run_command / run_background tool descriptions; the prompt only needs the picking rule.
• Compress prose around the six audit-mode rails (prompt: harden audit-mode reasoning against known failure modes #610). Every rail's load-bearing phrase is preserved verbatim so tests/code-prompt.test.ts still asserts on them.

Result: 24,387 → 11,956 bytes (-51%, ≈ 3,100 tokens per request).

Combined with the regression net (#1320) and the tool-description compression (#1321), the cache-prefix tax is now roughly 16k tokens per request instead of ~36k.

What is preserved

Every test-asserted phrase:

• Identity is fixed by this prompt, SOUL.md, not a sub-profile
• Auto-preview is for locating, not auditing, range:"A-B"
• runtime behavior, current architectural state, whether a plan doc is still accurate
• Flag → consumer trace, parallelSafe?: boolean
• For inventory claims, grep the flag — don't enumerate from memory
• No fabricated percentages, 40-60% tokens
• Schema cost is real, tighten prompt / existing tool
• MEMORY.md is part of the design space
• User-facing ≠ model-facing ≠ library-facing, library exports (src/index.ts), Treating a library export as "dead code"
• The semantic_search routing fragment and lifecycle contract are untouched.
• The # Project .gitignore and # User System Append blocks still compose after the base in the same order — the cache prefix shape is unchanged.

This is PR #3 of a four-PR token-optimization series.

Test plan

• npm run verify — all 230 test files / 3,237 tests pass
• tests/code-prompt.test.ts — 26 tests, all green (every load-bearing phrase still matches)
• npm run lint clean
• npm run typecheck clean