菜单
Harden memory_heist.py against token blowouts, unbounded growth, cross-platform path bugs, stale Claude references, and continuation/parser marker drift. Use when modifying memory_heist.py, fixing compaction logic, capping diffs, or adding safety limits to the Context Engine.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | context-engine-hardening |
| description | Harden memory_heist.py against token blowouts, unbounded growth, cross-platform path bugs, stale Claude references, and continuation/parser marker drift. Use when modifying memory_heist.py, fixing compaction logic, capping diffs, or adding safety limits to the Context Engine. |
src/memory_heist.py_format_continuation() or _extract_prior_summary()src/memory_heist.py before making changes.IGNORE_DIRS includes .sessions and .hermes.SWARM.md, .ham).discover_config() uses .ham.json / .ham/ paths only.git_diff() output is capped by MAX_DIFF_CHARS (when implemented)._summarize() caps the Timeline to the last N messages (e.g. 20)._merge_summaries() truncates the previous summary to prevent nesting growth._extract_key_files() does NOT gate on "/" in token only — extension-based detection for cross-platform paths._format_continuation() uses agent-appropriate language, not chatbot phrasing._format_continuation() text changes, update _extract_prior_summary() end_markers to match; keep legacy markers for backward compatibility with existing session JSON until migrated.src/memory_heist.py after changes.| Constant | Default | Purpose |
|---|---|---|
MAX_INSTRUCTION_FILE_CHARS | 4,000 | Per-file instruction cap |
MAX_TOTAL_INSTRUCTION_CHARS | 12,000 | Total instruction budget |
MAX_DIFF_CHARS | 8,000 | Git diff output cap |
MAX_SUMMARY_CHARS | 4,000 | Compaction summary cap |
All of these should be overridable via ContextBuilder constructor params once wiring is complete.
Review executor prompts for red flags, alignment issues, blockers, scope drift, and evidence gaps before any non-trivial execution.
Structured diagnosis loop: reproduce, minimize, hypothesize, instrument, fix, and regression-test.
Grilling session for plan alignment against domain language and ADRs, with inline doc updates when decisions are finalized.
Summarize the current thread into a compact handoff artifact for a fresh agent session.
PARKED/MANUAL ONLY: Identify architecture deepening opportunities and refactor candidates.
Repo-local setup for Matt Pocock engineering skills in HAM. Configures issue tracker, triage labels, and domain doc locations without touching runtime or deploy settings.