[BUG] Context [Execution] truncated — Capsule Preview diff bloats prompt to 70KB+

[zhuyankarl]

Environment

• Evolver version: 1.86.0
• Node.js: v22.22.1
• Platform: Linux x64

Problem

The gep_prompt_Cycle_*.txt files generated by dispatch.js consistently have Context [Execution]: truncated to just ...[TRUNCATED]... (39 bytes). This means the subagent receives NO execution instructions.

Root Cause

The Context [Injection Hint] section is bloated to ~54KB due to:

1. Capsule Preview diff field — each capsule embeds a full git diff (~8.5KB), 3 capsules = 25KB of diffs not useful for strategy decisions
2. Gene Preview learning_history and anti_patterns fields — verbose text
3. Total prompt reaches 77KB, exceeding dispatch.js internal size limit → Execution truncated

Evidence

Context [Injection Hint]:: 54,850 bytes  (74% of total)
Context [Execution]:: 39 bytes  ← TRUNCATED
Total: 76,815 bytes

After stripping verbose fields: 35,971 bytes (53% reduction).

Suggested Fix

1. Strip diff from Capsule Preview (or make opt-in)
2. Strip learning_history/anti_patterns from Gene Preview
3. Or: increase/remove internal prompt size limit

Workaround

Post-processing script strips verbose fields. Reduces tokens but cannot restore already-truncated Execution section.

[imxyanua]

Thanks for the byte-accurate breakdown, @zhuyankarl — having the Hint vs Execution split called out in bytes pointed straight at the truncator math, and your strip-and-remeasure (54KB → 36KB, ~53% reduction) made the root cause unambiguous.

Confirmed in src/gep/prompt.js. There are two coupled defects here:

1. formatAssetPreview currently inlines every field of every Capsule / Gene entry, including:

   • capsule.diff (~8.5KB each)
   • compact_diff
   • execution_trace
   • gene.learning_history
   • gene.evolution_history
   • gene.anti_patterns
   • failed_attempts

   With three capsules and a couple of genes, Context [Injection Hint] alone can exceed ~54KB.

2. The truncator computes allowed execution bytes with:

Math.max(0, maxChars - prefix.length - 100)

Once prefix.length exceeds maxChars (default 50_000), the expression goes negative, Math.max clamps to 0, and Context [Execution] is truncated to an empty string — exactly matching the 39-byte header + ...[TRUNCATED]... output you observed.

Internal patch is already in flight as a two-layer fix:

• Introduce PREVIEW_STRIP_FIELDS before formatAssetPreview() runs, removing the bloated fields from preview rendering while preserving:

  • ids
  • summaries
  • scores

  The resulting byte reduction matches your measurements.

• Add an execution-context floor (EXEC_FLOOR = 8000) in the truncator.

  If the prefix still exceeds the limit, the prefix is truncated instead of Execution. That's the safer trade-off:

  • losing reference context is recoverable
  • losing execution instructions breaks the cycle entirely

Regression coverage was added in test/prompt.test.js:

• a 9KB HEAVY_DIFF marker that must never appear in the rendered prompt
• an EXEC_MARKER that must always survive truncation

Verified by surgically reverting only the src changes — all new assertions fail with the expected regression messages.

We’ll ship the patch through the normal release pipeline shortly and close this issue once the public mirror lands.