[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-684":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":16,"forks30d":16,"starsTrendScore":16,"compositeScore":19,"rankGlobal":10,"rankLanguage":10,"license":20,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":23,"hasPages":21,"topics":24,"createdAt":10,"pushedAt":10,"updatedAt":25,"readmeContent":26,"aiSummary":27,"trendingCount":16,"starSnapshotCount":16,"syncStatus":15,"lastSyncTime":28,"discoverSource":29},684,"agentic-stack","codejunkie99\u002Fagentic-stack","codejunkie99","One brain, many harnesses. Portable .agent\u002F folder (memory + skills + protocols) that plugs into Claude Code, Cursor, Windsurf, OpenCode, OpenClaw, Hermes, or DIY Python — and keeps its knowledge when you switch.","",null,"Python",2101,258,15,2,0,18,160,78.24,"Apache License 2.0",false,"master",true,[],"2026-06-11 04:00:33","# agentic-stack\n\n**Keep one portable memory-and-skills layer across coding-agent harnesses, so switching tools doesn't reset how your agent works.**\n\nA portable `.agent\u002F` folder (memory + skills + protocols) that plugs into Claude Code, Cursor, Windsurf, OpenCode, OpenClaw, Hermes, Pi Coding Agent, Codex, Antigravity, or a DIY Python loop — and keeps its knowledge when you switch.\n\nIt also includes a local data layer so you can monitor the whole suite of\nagents from one place: harness activity, cron runs, active agents, token\u002Fcost\nestimates, KPI summaries, user-defined resource categories, and\nscreenshot-ready daily dashboards.\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"docs\u002Fdata-layer.svg\" alt=\"agentic-stack data layer dashboard flow\" width=\"880\"\u002F>\n\u003C\u002Fp>\n\nAnd it can turn approved, redacted runs into local flywheel artifacts:\ntrace records, context cards, eval cases, training-ready JSONL, and readiness\nmetrics without training a model or sending telemetry.\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"docs\u002Fdemo.gif\" alt=\"agentic-stack demo\" width=\"880\"\u002F>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"docs\u002Fdiagram.svg\" alt=\"agentic-stack architecture\" width=\"880\"\u002F>\n\u003C\u002Fp>\n\n### New in v0.15.0 — dashboard TUI\n\nMinor release. Adds `agentic-stack dashboard` as the production front door for\ninstalled projects: one terminal screen for health, adapters, verification,\nmemory, team brain, skills, instances, transfer, and local dashboard exports.\n\n- **Dashboard command.** Run `agentic-stack dashboard` or `.\u002Finstall.sh dashboard`\n  to open the TUI; use `dash` or `--plain` for a compact script-safe view.\n- **Trust-console parity.** The dashboard includes a per-harness verify matrix,\n  accepted\u002Frejected memory, `memory_why()` evidence lookup, team brain\n  status\u002Finit, skills, and managed instances.\n- **Safer installed-project default.** Bare interactive `.\u002Finstall.sh` opens the\n  dashboard once `.agent\u002Finstall.json` exists; non-TTY shells still print\n  script-safe command guidance instead of launching a TUI.\n- **Production coverage.** Renderer, CLI, parity, non-TTY fallback, and\n  interactive keypress navigation are covered by local tests.\n\nSee [CHANGELOG.md](CHANGELOG.md) for the full list.\n\n### v0.12.0 — tldraw visual canvas\n\nMinor release. Adds an opt-in `tldraw` seed skill for live canvas diagrams and\na skill-local snapshot store. It is beta and off by default.\n\n- **`tldraw` seed skill.** Draw, diagram, sketch, wireframe, flowchart, and\n  whiteboard on a live canvas at `http:\u002F\u002Flocalhost:3030` through an MCP server.\n- **Skill-local snapshots.** Save worthwhile canvases with\n  `.agent\u002Fskills\u002Ftldraw\u002Fstore.py snapshot`; list, load, and archive them later\n  without treating them as a fifth memory layer.\n- **Opt-in beta.** Onboarding writes `tldraw.enabled: false` by default. After\n  enabling it, users manually merge `adapters\u002F_shared\u002Ftldraw-mcp.json` into\n  their harness MCP config.\n\n### v0.11.0 — data layer + data flywheel\n\nAdded two local-first data capabilities for teams running multiple agent\nharnesses against the same `.agent\u002F` brain.\n\n- **`data-layer` seed skill.** Generate local dashboard exports across Claude\n  Code, Hermes, OpenClaw, Codex, Cursor, OpenCode, and custom loops:\n  harness events, cron timelines, KPI summaries, token\u002Fcost estimates,\n  categories, `dashboard.html`, and `daily-report.md`. The skill also acts as\n  the injected natural-language surface for showing the terminal dashboard.\n- **`data-flywheel` seed skill.** Export approved, redacted runs into trace\n  records, context cards, eval cases, training-ready JSONL, and flywheel\n  metrics. It is local-only and model-agnostic; it prepares artifacts but\n  does not train models or call external APIs.\n\n### v0.10.0 — design-md skill + Python 3.9 fix\n\nAdded the `design-md` seed skill for root `DESIGN.md` \u002F Google Stitch\nworkflows, and fixed the Python 3.9 crash that hit macOS-default brew users\non first run.\n\n### v0.9.1 — pi adapter fixes + tz correctness\n\nClosed the gap between v0.9.0 and a working pi adapter, plus a timezone\nsweep across every Python writer\u002Freader so the dream cycle stops drifting\nagainst the UTC decay window.\n\n### v0.9.0 — harness manager\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"docs\u002Fharness-manager.svg\" alt=\"harness manager v0.9.0\" width=\"880\"\u002F>\n\u003C\u002Fp>\n\nManifest-driven adapter system: every harness is now declared by an\n`adapter.json`, applied by a shared Python backend, and managed via\nverb subcommands or an interactive TUI. Cross-platform (POSIX +\nWindows) with concurrent-write protection, pre-v0.9 migration via\n`.\u002Finstall.sh doctor`, and shared-file ownership tracking so removing\none adapter never orphans another.\n\n[![GitHub release](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fv\u002Frelease\u002Fcodejunkie99\u002Fagentic-stack)](https:\u002F\u002Fgithub.com\u002Fcodejunkie99\u002Fagentic-stack\u002Freleases)\n[![License: Apache 2.0](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flicense-Apache%202.0-blue.svg)](LICENSE)\nMade by https:\u002F\u002Fx.com\u002FAv1dlive\n\n## Quickstart\n\n### macOS \u002F Linux\n\n```bash\n# tap + install (one-time — both lines required)\nbrew tap codejunkie99\u002Fagentic-stack https:\u002F\u002Fgithub.com\u002Fcodejunkie99\u002Fagentic-stack\nbrew install agentic-stack\n\n# drop the brain into any project — the onboarding wizard runs automatically\ncd your-project\nagentic-stack claude-code\n# or: cursor | windsurf | opencode | openclaw | hermes | pi | codex | standalone-python | antigravity\n```\n\n### Windows (PowerShell)\n\n```powershell\n# clone + run the native installer\ngit clone https:\u002F\u002Fgithub.com\u002Fcodejunkie99\u002Fagentic-stack.git\ncd agentic-stack\n.\\install.ps1 claude-code C:\\path\\to\\your-project\n```\n\n### Already installed?\n\n```bash\nbrew update && brew upgrade agentic-stack\nagentic-stack dashboard\n```\n\n### Clone instead?\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fcodejunkie99\u002Fagentic-stack.git\ncd agentic-stack && .\u002Finstall.sh claude-code         # mac \u002F linux \u002F git-bash\n# or on Windows PowerShell: .\\install.ps1 claude-code\n# adapters: claude-code | cursor | windsurf | opencode | openclaw | hermes | pi | codex | standalone-python | antigravity\n```\n\n### Once installed: manage what's wired\n\nAfter the first `.\u002Finstall.sh \u003Cadapter>`, manage your project with\nverb-style subcommands (works with both `install.sh` and `install.ps1`):\n\n```bash\n.\u002Finstall.sh dashboard           # TUI dashboard: health, verify, memory, team, skills, instances\n.\u002Finstall.sh add cursor          # add a second adapter (Claude Code + Cursor in same repo)\n.\u002Finstall.sh status              # one-screen view: which adapters, brain stats\n.\u002Finstall.sh doctor              # read-only audit; green \u002F yellow \u002F red per adapter\n.\u002Finstall.sh manage              # interactive TUI: header pane + menu loop for add\u002Fremove\u002Faudit\n.\u002Finstall.sh transfer            # onboarding-style wizard: export\u002Fimport memory as a curl bridge\n.\u002Finstall.sh remove cursor       # confirm prompt + delete; no quarantine, no undo\n```\n\nPowerShell uses the same verbs, for example `.\\install.ps1 dashboard`.\n\nBare `.\u002Finstall.sh` (no arguments) opens a **multi-select wizard** on\na fresh project — check every harness you actually use, hit enter,\neach one gets installed. The wizard auto-detects harnesses already on\ndisk and pre-checks them. On a project that already has an\n`install.json`, bare interactive `.\u002Finstall.sh` opens the dashboard.\nIn non-TTY shells (CI), it stays script-safe and prints the available\nsubcommands instead of opening a TUI.\n\nUpgrading from pre-v0.9? Run `.\u002Finstall.sh doctor` first — it\nsynthesizes `install.json` from on-disk adapter signals so the new\nbackend can track them. Installing on top without migration would\norphan the prior installs.\n\n## Onboarding wizard\n\nIf you ran bare `.\u002Finstall.sh` (no adapter name), the wizard starts\nwith a **multi-select harness step**: it lists all 10 adapters, pre-\nchecks any it detects on disk, and installs each one you confirm with\nspace + enter. After the install(s), the preferences flow runs.\n\nIf you ran `.\u002Finstall.sh \u003Cadapter>` directly, only the preferences\nflow runs.\n\nEither way, the preferences step populates\n`.agent\u002Fmemory\u002Fpersonal\u002FPREFERENCES.md` — the **first file your AI reads\nat the start of every session** — and writes a feature-toggle file at\n`.agent\u002Fmemory\u002F.features.json`.\n\nSix preference questions (each skippable with Enter):\n\n| Question | Default |\n|---|---|\n| What should I call you? | *(skip)* |\n| Primary language(s)? | `unspecified` |\n| Explanation style? | `concise` |\n| Test strategy? | `test-after` |\n| Commit message style? | `conventional commits` |\n| Code review depth? | `critical issues only` |\n\nPlus one **Optional features** step (opt-in, off by default):\n\n| Feature | Default |\n|---|---|\n| Enable FTS memory search `[BETA]` | `no` |\n| Enable tldraw visual canvas `[BETA]` | `no` |\n\n**Flags:**\n\n```bash\nagentic-stack claude-code --yes          # accept all defaults, beta off (CI\u002Fscripted)\nagentic-stack claude-code --reconfigure  # re-run the wizard on an existing project\n```\n\nEdit `.agent\u002Fmemory\u002Fpersonal\u002FPREFERENCES.md` any time to refine your\nconventions, or `.agent\u002Fmemory\u002F.features.json` to flip feature toggles.\n\n## Transfer wizard\n\nMove the portable parts of one project brain into Codex, Cursor, Windsurf,\nor a terminal-only project with the onboarding-style TUI:\n\n```bash\n.\u002Finstall.sh transfer\n```\n\nThe wizard turns a plain-language intent into a transfer plan, lets you\nreview target harnesses and memory scopes, blocks secret-like content before\nexport, and emits a one-line curl command the next environment can run.\nFor `move my memory`, it includes preferences, accepted lessons, skills,\nworking memory, episodic\u002Fhistory logs, and candidate lessons. The importer\nunpacks the bundle, verifies its SHA-256 digest, merges preferences and\naccepted lessons, copies selected skills, restores selected memory files,\nand installs the matching adapter files.\n\nFor scripted handoff:\n\n```bash\n.\u002Finstall.sh transfer export --intent \"move my preferences and lessons into Codex\" --print-curl\n.\u002Finstall.sh transfer import --payload-file transfer.txt --sha256 \u003Cdigest> --target codex\n```\n\n## Review protocol (host-agent CLI)\n\nThe nightly `auto_dream.py` cycle only **stages** candidate lessons. It\ndoes not mark anything accepted or modify semantic memory. Your host\nagent does the review in-session:\n\n```bash\n# list pending candidates, sorted by priority\npython3 .agent\u002Ftools\u002Flist_candidates.py\n\n# accept with rationale (required)\npython3 .agent\u002Ftools\u002Fgraduate.py \u003Cid> --rationale \"evidence holds, matches PREFERENCES\"\n\n# reject with reason (required); preserves decision history\npython3 .agent\u002Ftools\u002Freject.py \u003Cid> --reason \"too specific to generalize\"\n\n# requeue a previously-rejected candidate\npython3 .agent\u002Ftools\u002Freopen.py \u003Cid>\n```\n\nGraduated lessons land in `semantic\u002Flessons.jsonl` (source of truth) and\nare rendered to `semantic\u002FLESSONS.md`. Rejected candidates retain full\ndecision history so recurring churn is visible, not fresh.\n\nSee [`docs\u002Farchitecture.md`](docs\u002Farchitecture.md) for the full lifecycle.\n\n---\n\n## What this is\n\nEvery guide shows the folder structure. This repo gives you the folder\nstructure **plus the files that actually go inside**: a working portable\nbrain with nine seed skills, four memory layers, enforced permissions, a\nnightly staging cycle, host-agent review tools, and adapters for multiple\nharnesses.\n\n- **Memory** — `working\u002F`, `episodic\u002F`, `semantic\u002F`, `personal\u002F`. Each\n  layer has its own retention policy. Query-aware retrieval (salience ×\n  relevance); nightly compression into reviewable candidates.\n- **Review protocol** — `auto_dream.py` stages candidate lessons\n  mechanically. Your host agent reviews them via CLI tools\n  (`graduate.py`, `reject.py`, `reopen.py`) and commits decisions with\n  a required rationale. No unattended reasoning, no provider coupling.\n- **Skills** — progressive disclosure. A lightweight manifest always\n  loads; full `SKILL.md` files only load when triggers match the task.\n  Every skill ships with a self-rewrite hook. The bundled `design-md`\n  skill teaches agents to use a root `DESIGN.md` as the visual source of\n  truth for UI and Google Stitch workflows.\n- **Protocols** — typed tool schemas, a `permissions.md` that the\n  pre-tool-call hook enforces, and a delegation contract for sub-agents.\n- **Data layer** — local-only dashboard exports across every harness sharing\n  `.agent\u002F`: agent events, cron timelines, KPI summaries, tokens\u002Fcost\n  estimates, task categories, harness mix, `dashboard.html`, and daily report\n  handoff.\n- **Data flywheel** — approved, redacted runs can become trace records,\n  context cards, eval cases, training-ready JSONL, and readiness metrics\n  without training a model or sending telemetry.\n\n## Releases & changelog\n\nPer-version release notes live in [CHANGELOG.md](CHANGELOG.md). The\nlatest release, what broke, what's new, upgrade path, all there.\n\n## Memory search `[BETA]`\n\nOpt-in FTS5 keyword search over all memory documents:\n\n```bash\n# enable during onboarding (or set manually in .agent\u002Fmemory\u002F.features.json)\npython3 .agent\u002Fmemory\u002Fmemory_search.py \"deploy failure\"\npython3 .agent\u002Fmemory\u002Fmemory_search.py --status\npython3 .agent\u002Fmemory\u002Fmemory_search.py --rebuild\n```\n\nFalls back to **ripgrep** (`rg`) if installed, then to `grep` — both\nrestricted to `.md` \u002F `.jsonl` so source files never pollute results.\nThe index is stored at `.agent\u002Fmemory\u002F.index\u002F` and gitignored.\n\n## Repo layout\n\n```\n.agent\u002F                         # the portable brain (same across harnesses)\n├── AGENTS.md                   # the map\n├── harness\u002F                    # conductor + hooks (standalone path)\n│   └── hooks\u002F\n│       ├── claude_code_post_tool.py  # rich PostToolUse logging (v0.8+)\n│       ├── pre_tool_call.py    # permissions enforcement\n│       ├── post_execution.py   # log_execution() entry point\n│       └── on_failure.py       # failure write + repeated-failure rewrite flag\n├── memory\u002F                     # working \u002F episodic \u002F semantic \u002F personal\n│   ├── auto_dream.py           # staging-only dream cycle\n│   ├── cluster.py              # content clustering + pattern extraction\n│   ├── promote.py              # stage candidates\n│   ├── validate.py             # heuristic prefilter (length + exact duplicate)\n│   ├── review_state.py         # candidate lifecycle + decision log\n│   ├── render_lessons.py       # lessons.jsonl → LESSONS.md\n│   └── memory_search.py        # [BETA] FTS5 search (opt-in)\n├── skills\u002F                     # _index.md + _manifest.jsonl + SKILL.md files\n├── protocols\u002F                  # permissions + tool schemas + delegation\n│   └── hook_patterns.json      # user-owned high\u002Fmedium-stakes regex (v0.8+)\n└── tools\u002F                      # host-agent CLI + memory_reflect + skill_loader\n    ├── learn.py                # one-shot lesson teaching (stage + graduate)\n    ├── recall.py               # surface lessons relevant to an intent\n    ├── show.py                 # colorful brain-state dashboard\n    ├── data_layer_export.py    # local cross-harness dashboard\u002Fdata export\n    ├── data_flywheel_export.py # approved runs -> traces\u002Fcards\u002Fevals\u002FJSONL\n    ├── list_candidates.py\n    ├── graduate.py\n    ├── reject.py\n    └── reopen.py\n\nadapters\u002F                       # one small shim per harness, each with adapter.json manifest\n├── claude-code\u002F   (CLAUDE.md + settings.json hooks — $CLAUDE_PROJECT_DIR wired, closes #18)\n├── cursor\u002F        (.cursor\u002Frules\u002F*.mdc)\n├── windsurf\u002F      (.windsurf\u002Frules\u002F*.md + legacy .windsurfrules)\n├── opencode\u002F      (AGENTS.md + opencode.json)\n├── openclaw\u002F      (AGENTS.md + system-prompt include; auto-registers per-project agent)\n├── hermes\u002F        (AGENTS.md)\n├── pi\u002F            (AGENTS.md + .pi\u002Fskills symlink)\n├── codex\u002F         (AGENTS.md + .agents\u002Fskills symlink)\n├── standalone-python\u002F  (DIY conductor entrypoint)\n└── antigravity\u002F   (ANTIGRAVITY.md)\n\nharness_manager\u002F                # v0.9.0 manifest-driven Python backend\n├── schema.py                   # adapter.json validator (path-safe on POSIX + Windows)\n├── install.py                  # applies file entries per merge_policy\n├── state.py                    # install.json read\u002Fwrite with fcntl\u002Fmsvcrt locking\n├── doctor.py                   # read-only audit + pre-v0.9 migration synthesis\n├── remove.py                   # safe uninstall with shared-file detection + ownership handoff\n├── dashboard_tui.py            # project dashboard for health\u002Fverify\u002Fmemory\u002Fteam\u002Fskills\u002Finstances\n├── post_install.py             # named built-ins (openclaw_register_workspace)\n├── manage_tui.py               # interactive menu loop for add\u002Fremove\u002Faudit\n├── transfer_tui.py             # onboarding-style memory transfer wizard\n├── transfer_plan.py            # natural-language target\u002Fscope planning\n├── transfer_bundle.py          # export\u002Fimport bundle codec + merge logic\n└── cli.py                      # argparse dispatcher for install.sh \u002F install.ps1\n\ndocs\u002F                           # architecture, getting-started, per-harness\nschemas\u002Fdata-layer\u002F             # local dashboard\u002Fevent schemas\nexamples\u002Fdata-layer\u002F            # sanitized data-layer shapes\nschemas\u002Fflywheel\u002F               # data-flywheel artifact schemas\nexamples\u002Fflywheel\u002F              # sanitized approved-run examples\ninstall.sh                      # mac \u002F linux \u002F git-bash installer (thin Python dispatcher)\ninstall.ps1                     # Windows PowerShell installer (thin Python dispatcher)\nFormula\u002Fagentic-stack.rb        # Homebrew formula\nCHANGELOG.md                    # per-version release notes (v0.1.0 onward)\nonboard.py                      # onboarding wizard entry point\nonboard_features.py             # .features.json read\u002Fwrite\nonboard_ui.py                   # ANSI palette, banner, clack-style layout\nonboard_widgets.py              # arrow-key prompts (text, select, confirm)\nonboard_render.py               # answers → PREFERENCES.md content\nonboard_write.py                # atomic file write with backup\ntest_claude_code_hook.py        # hook validation suite (54 checks)\nverify_codex_fixes.py           # v0.8.0 regression checks (33 checks)\n```\n\n## Supported harnesses\n\n| Harness | Config file it reads | Hook support |\n|---|---|---|\n| **Claude Code** | `CLAUDE.md` + `.claude\u002Fsettings.json` | yes (PostToolUse, Stop) |\n| **Cursor** | `.cursor\u002Frules\u002F*.mdc` | no (manual reflect calls) |\n| **Windsurf** | `.windsurfrules` | no (manual reflect calls) |\n| **OpenCode** | `AGENTS.md` + `opencode.json` | partial (permission rules) |\n| **OpenClaw** | `AGENTS.md` (auto-injected) + per-project `openclaw agents add --workspace` | varies by fork |\n| **Hermes Agent** | `AGENTS.md` (agentskills.io compatible) | partial (own memory) |\n| **Pi Coding Agent** | `AGENTS.md` + `.pi\u002Fskills\u002F` + `.pi\u002Fextensions\u002F` | yes (`tool_result` event) |\n| **Codex** | `AGENTS.md` + `.agents\u002Fskills\u002F` | no (manual reflect calls) |\n| **Standalone Python** | `run.py` (any LLM) | yes (full control) |\n| **Antigravity** | `ANTIGRAVITY.md` | yes (system context) |\n\n## Seed skills\n\n- **skillforge** — creates new skills from recurring patterns\n- **memory-manager** — runs reflection cycles, surfaces candidate lessons\n- **git-proxy** — all git ops, with safety constraints\n- **debug-investigator** — reproduce → isolate → hypothesize → verify\n- **deploy-checklist** — the fence between staging and production\n- **design-md** — uses Google Stitch-style `DESIGN.md` files as portable\n  design-system context for UI, frontend, and component work\n- **data-layer** — exports local dashboard data, cron timelines, KPIs, and\n  daily reports across harnesses\n- **data-flywheel** — approved runs into context cards, evals, redacted traces,\n  training-ready JSONL, and flywheel metrics\n- **tldraw** — opt-in beta skill for live canvas diagrams with a local\n  snapshot store under `.agent\u002Fskills\u002Ftldraw\u002F`\n\n## How it compounds\n\n1. Skills log every action to episodic memory.\n2. `auto_dream.py` clusters recurring patterns into candidate lessons.\n3. The host agent reviews candidates with `graduate.py` \u002F `reject.py`.\n4. Graduated lessons append to `lessons.jsonl`; `LESSONS.md` re-renders.\n5. Future sessions load query-relevant accepted lessons automatically.\n6. `on_failure` flags skills that fail 3+ times in 14 days for rewrite.\n7. `git log .agent\u002Fmemory\u002F` becomes the agent's autobiography.\n8. Data-layer exports turn local activity into dashboard-ready monitoring.\n9. Approved, redacted runs can be exported into `.agent\u002Fflywheel\u002F` artifacts\n   for retrieval, evals, prompt shrinking, and optional future adapters.\n\n## Export approved runs into a data flywheel\n\nPut sanitized human-approved runs in:\n\n```text\n.agent\u002Fflywheel\u002Fapproved-runs.jsonl\n```\n\nThen run:\n\n```bash\npython3 .agent\u002Ftools\u002Fdata_flywheel_export.py\n```\n\nOutputs land in `.agent\u002Fflywheel\u002Fexports\u002F\u003Cdate>\u002F`:\n\n- `trace-records.jsonl`\n- `training-examples.jsonl`\n- `eval-cases.jsonl`\n- `context-cards\u002F\u003Cdomain>\u002F\u003Cworkflow>.md`\n- `flywheel-metrics.json`\n\nThis is local-only and model-agnostic. It creates training-ready artifacts; it\ndoes not train a model.\n\nSee [docs\u002Fdata-flywheel.md](docs\u002Fdata-flywheel.md).\n\n## Run the staging cycle nightly\n\n```bash\ncrontab -e\n0 3 * * * python3 \u002Fpath\u002Fto\u002Fproject\u002F.agent\u002Fmemory\u002Fauto_dream.py >> \u002Fpath\u002Fto\u002Fproject\u002F.agent\u002Fmemory\u002Fdream.log 2>&1\n```\n\n`auto_dream.py` resolves its paths absolutely and performs only mechanical\nfile operations (cluster, stage, prefilter, decay). No git commits, no\nnetwork, no reasoning — safe to run unattended.\n\n## Monitor your agent suite\n\nGenerate a local dashboard for all harnesses writing to the same `.agent\u002F`\nbrain:\n\n```bash\npython3 .agent\u002Ftools\u002Fdata_layer_export.py --window 30d --bucket day\n```\n\nOr let the injected `data-layer` skill pass the user's words through:\n\n```bash\npython3 .agent\u002Ftools\u002Fdata_layer_export.py show me last 7 days by hour\n```\n\nOutputs land in `.agent\u002Fdata-layer\u002Fexports\u002F\u003Cdate>\u002F`, including\n`dashboard.html`, `dashboard.tui.txt`, and `daily-report.md`. The command also\nprints the onboarding-style terminal dashboard directly inside your coding tool.\nOptional local inputs let you add scheduled runs and categories:\n\n```text\n.agent\u002Fdata-layer\u002Fcron-runs.jsonl\n.agent\u002Fdata-layer\u002Fcategory-rules.json\n.agent\u002Fdata-layer\u002Fharness-events.jsonl\n```\n\nUse this to track crons by day, active agents, token\u002Fcost estimates by\nhour\u002Fday\u002Fweek\u002Fmonth, harness mix across Claude\u002FHermes\u002FOpenClaw\u002FCodex\u002Fetc.,\nsuccess\u002Ferror rates, run cadence, workflow breadth, and user-defined categories\nlike personal, admin, work, financial, and coding. The data layer is local-only;\nscreenshot delivery requires explicit user approval and a user-configured\nchannel.\n\nSee [docs\u002Fdata-layer.md](docs\u002Fdata-layer.md).\n\n## License\n\nApache 2.0 — see [LICENSE](LICENSE).\n\n## Credits\n\nBased on the article **[\"The Agentic Stack\"](https:\u002F\u002Fx.com\u002FAv1dlive\u002Fstatus\u002F2044453102703841645?s=20)**\nby [@AV1DLIVE](https:\u002F\u002Ftwitter.com\u002FAV1DLIVE) — follow for updates and collabs.\nCoded using Minimax-M2.7 in the Claude Code harness; PR review by Macroscope and Codex.\nPatterns from Gstack, Claude Code's memory system, and conversations in the\nagent-engineering community. Built with the hypothesis that\n**harness-agnosticism is the point**.\n\n## Star History\n\n[![Star History Chart](https:\u002F\u002Fapi.star-history.com\u002Fsvg?repos=codejunkie99\u002Fagentic-stack&type=Date)](https:\u002F\u002Fstar-history.com\u002F#codejunkie99\u002Fagentic-stack&Date)\n","agentic-stack 是一个用于跨多个编码代理工具保持统一记忆和技能层的项目。它通过一个可移植的 `.agent\u002F` 文件夹（包含记忆、技能和协议）与Claude Code、Cursor、Windsurf等工具集成，并在切换工具时保留其知识。项目支持本地数据层，允许用户从单一界面监控所有代理活动、计划任务运行情况、活跃代理状态及成本估算等信息。此外，它能够将批准的、经过编辑的运行转化为本地飞轮工件，如跟踪记录、上下文卡片、评估案例等，无需训练模型或发送遥测数据。此项目适用于需要频繁切换不同编码辅助工具但希望保持一致工作环境的开发者或团队。","2026-06-11 02:38:38","CREATED_QUERY"]