菜单
Git-track AI agent learnings so they survive machine loss — memory snapshots, corrections, patterns, insights across Codex, Hermes, Codex, and Gemini.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/vamseeachanta/workspace-hub --skill agent-learnings-portability命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | agent-learnings-portability |
| version | 1.1 |
| description | Git-track AI agent learnings so they survive machine loss — memory snapshots, corrections, patterns, insights across Codex, Hermes, Codex, and Gemini. |
| tags | ["harness","git","portability","memory","agent-state","gitignore"] |
| trigger | When auditing agent state portability, setting up a new machine, or investigating why gitignore exceptions aren't working |
| effort | medium |
AI agents (Codex, Hermes, Codex, Gemini) accumulate valuable learnings in machine-local directories (~/.Codex/, ~/.hermes/, ~/.codex/, ~/.gemini/). Without explicit git tracking, these are lost on machine failure.
.Codex/memory/ (survives git clone, PRIMARY): Memory facts, conventions, user context — travels with the repo.Codex/memory/ Pattern (issue #1886)The primary mechanism for cross-machine context parity is writing canonical memory facts directly into .Codex/memory/ inside the repo. This directory is already git-tracked and read by Codex at session start on every machine.
How it works:
~/.hermes/memories/MEMORY.md + USER.md) is the authoritative sourcescripts/memory/bridge-hermes-Codex.sh) reads Hermes memory + Codex auto-memory.Codex/memory/agents.md and context.mdgit commit + push → every machine that pulls gets the same contextgit pull → done. No manual copying, no tarballs, no installers.Key files:
.Codex/memory/context.md — Machine conventions, paths, Python commands, workspace layout.Codex/memory/agents.md — User profile, AI subscriptions, workflow rules, GSD facts.Codex/memory/Codex-auto-memory.md — Snapshot of Codex's auto-generated MEMORY.mdscripts/memory/bridge-hermes-Codex.sh — The bridge script, run with --commit flagWhy not tarballs or export/import scripts? One-time snapshots go stale immediately. Git gives version history, diffs, rollback, and automatic updates on every git pull. The bridge should run daily via cron, not manually.
Return enrichment flow: Non-Hermes machines (Windows) enrich KNOWLEDGE.md or topic files → git commit/push → Linux pulls → bridge script picks up on next run via Codex auto-memory snapshot capture.
This pattern REPLACES the older config/agents/ snapshot approach for memory (keep config/agents/ for large state files like session exports, but use .Codex/memory/ for human-readable context).
memories/MEMORY.md + USER.md → snapshot to config/agents/hermes/memories/projects/<encoded-path>/memory/*.md → snapshot to config/agents/Codex/memory-snapshots/
feedback_*.md (user corrections) and project_*.md (context).Codex/state/ directories (in-repo): corrections/, patterns/, reflect-history/, cc-insights/, trends/, candidates/, session-signals/, skill-eval-results/.Codex/state/ files: learned-patterns.json, skill-scores.yaml, cc-user-insights.yaml, hermes-insights.yaml, cross-agent-memory.yaml, drift-summary.yaml, portfolio-signals.yamlrules/default.rules (learned permissions — CRITICAL)history.jsonl, session_index.jsonl
→ snapshot to config/agents/codex/state-snapshots/state.json, projects.json
→ snapshot to config/agents/gemini/state-snapshots/This is the #1 gotcha that makes gitignore exceptions silently fail.
# THIS IS BROKEN — exceptions below are DEAD CODE
.Codex/state/ # <-- trailing slash = ignores the DIRECTORY
!.Codex/state/corrections/ # <-- NEVER WORKS
When git ignores a directory (trailing /), it never looks inside it, so child exceptions cannot un-ignore anything.
# Layer 1: Top-level ignore with glob (not directory)
.Codex/*
# Layer 2: Un-ignore the parent directory itself
!.Codex/state/
# Layer 3: Re-ignore contents, THEN add specific exceptions
.Codex/state/*
!.Codex/state/corrections/
!.Codex/state/patterns/
!.Codex/state/learned-patterns.json
# Check if a file is ignored (should show NO output for tracked files)
git check-ignore -v .Codex/state/corrections/foo.jsonl
# If it shows a rule, the exception is NOT working
# The -v flag shows WHICH rule is blocking
# Step 1: Check which rule is blocking
git check-ignore -v .Codex/state/corrections/foo.jsonl
# Output shows the EXACT line: ".gitignore:127:.Codex/*"
# Step 2: Check if there are MULTIPLE blocking rules
git check-ignore -v .Codex/state/corrections/ .Codex/state/learned-patterns.json
# Step 3: After fixing, verify files CAN be staged
git add .Codex/state/corrections/ 2>&1
# If you see "ignored by .gitignore" there's still a blocking rule
/ instead of /* — children can never be excepted.Codex/* at line 127 AND .Codex/state/ at line 157. You fix one but the other still blocks. Always run git check-ignore -v AFTER each fix to verifygit add-ed — the gitignore exception alone doesn't track filesScript: scripts/cron/commit-learning-artifacts.sh
Steps:
config/agents/*/git add all state directories and snapshot directorieslegal-sanity-scan.sh --diff-only (MANDATORY — corrections and session data can contain client names)Wire into: comprehensive-learning-nightly.sh as final step.
session-signals/*.jsonl files contain a last_assistant_message field with full LLM response text. This text often references client project names (from doc-intelligence classification outputs). The field must be redacted before git-add.
Script: scripts/cron/redact-session-signals.sh
.Codex/state/session-signals/*.jsonllast_assistant_message with [REDACTED] using Python JSON parsinggit add in the nightly pipelineCodex sessions live at ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl in a different format than Codex/Hermes.
Key Codex JSONL types:
session_meta — session ID, cwd, model providerresponse_item with payload.type == "function_call" — tool callsresponse_item with payload.type == "function_call_output" — resultsexec_command, read_file, write_file, apply_diff, write_stdin, list_directoryScript: scripts/cron/codex-session-export.sh
logs/orchestrator/codex/session_YYYYMMDD.jsonl.last-export-ts timestamp fileSession signals, corrections, and memory files can contain client project names. The legal scan MUST gate all commits. If it finds violations, skip the commit rather than pushing sensitive data.
Key mitigation: run redact-session-signals.sh before staging. The session-signals directory is the #1 source of legal violations (522 hits in first attempt without redaction).
scripts/cron/memory-backup.sh rsyncs to ace-linux-2 daily:
~/.Codex/projects/)~/.hermes/memories/, ~/.hermes/sessions/)~/.codex/sessions/, ~/.codex/rules/)~/.gemini/tmp/)Use ssh -o ConnectTimeout=10 -o BatchMode=yes for non-interactive cron. Non-critical agents use || true to not abort.
Script: scripts/_core/sync-agent-configs.sh (restore section)
Behavior: only copies if target files are MISSING (safe on existing machines). Use --force to overwrite.
.snapshot files → ~/.hermes/memories/ (strips .snapshot suffix)config/agents/Codex/memory-snapshots/*.md → encoded project path, additive (skips existing)default.rules, history.jsonl, session_index.jsonlstate.json, projects.jsonCodex project path encoding: /mnt/local-analysis/workspace-hub → -mnt-local-analysis-workspace-hub