…st) (#1323)

The system prompt was 24,387 bytes (≈ 6,100 tokens) — much of it
overlapping with the tool descriptions sitting right next to it in
the cache prefix. Sections like "When to propose a plan", "When to
ask the user to pick", and "When to track multi-step intent" each
recited rules that 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. The full how-to lives
  in the run_command / run_background tool descriptions; the prompt
  only needs the picking rule.
- Compress the audit-mode rails (#610) prose around the six rails
  themselves. 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 PR #1320 / #1321 the cache-prefix tax per request is
now ~16k tokens instead of ~36k.

Behaviour unchanged — every rail / gate / mode constraint is still
asserted by the existing prompt tests.

Co-authored-by: reasonix <reasonix@deepseek.com>

Every byte of the system prompt and tool spec list ships in every
request — paid at full miss price the first time and ~10% on every
subsequent cache hit. Recent PRs have grown both without anyone
tracking the cumulative cost.

Lock the current values:
- CODE_SYSTEM_PROMPT.length ≤ 24_500 bytes
- code-mode tool spec list ≤ 40_000 bytes
- no single tool description > 8 KiB

A failing assertion forces the PR author to either compress an
equivalent section or raise the budget explicitly with a
justification in the commit message — visible in code review.

Co-authored-by: reasonix <reasonix@deepseek.com>

…st) (esengine#1323)

The system prompt was 24,387 bytes (≈ 6,100 tokens) — much of it
overlapping with the tool descriptions sitting right next to it in
the cache prefix. Sections like "When to propose a plan", "When to
ask the user to pick", and "When to track multi-step intent" each
recited rules that 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. The full how-to lives
  in the run_command / run_background tool descriptions; the prompt
  only needs the picking rule.
- Compress the audit-mode rails (esengine#610) prose around the six rails
  themselves. 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 PR esengine#1320 / esengine#1321 the cache-prefix tax per request is
now ~16k tokens instead of ~36k.

Behaviour unchanged — every rail / gate / mode constraint is still
asserted by the existing prompt tests.

Co-authored-by: reasonix <reasonix@deepseek.com>