planning-with-files

用任务计划、发现记录和进度文件支撑跨上下文的长周期编码任务。

主要能力

Manus-style persistent file-based planning for AI coding agents: keeps task_plan.md, findings.md, and progress.md on disk so work survives context loss and /clear. Use when asked to plan out, break down, or organize a multi-step project, research task, or any work requiring 5+ tool calls. Supports automatic session recovery after /clear.

使用方式

仓库级安装会保留完整目录；本次核验的入口是 .codebuddy/skills/planning-with-files/SKILL.md。使用前先阅读英文 README 与原始 SKILL.md，并按当前 Agent 的目录规范安装。

适用边界

简单任务会产生不必要的文件维护成本，五步以内通常无需启用。

许可与来源

来源：GitHub 公开仓库。核验许可：MIT。本页是贴近原仓库结构的中文导读，具体参数、依赖和更新记录以英文 README 为准。

<div align="center"> <img src="media/v3-banner.png" alt="planning-with-files" width="100%"> </div>

Planning with Files

📣 New in v3.0.0: opt-in autonomous and gated modes for long-running agent runs, with a completion gate that holds the agent until the plan is actually done. Existing setups change nothing.

Work like Manus — the AI agent company Meta acquired for $2 billion. planning-with-files is a persistent file-based planning skill for AI coding agents. It keeps task_plan.md, findings.md, and progress.md on disk so the agent survives context loss, /clear, and crashes, with an opt-in completion gate that holds the agent until the plan is actually done. It installs across 60+ agents via the SKILL.md standard.

-brightgreen)

<details> <summary><strong>💬 A Note from the Author</strong></summary>

To everyone who starred, forked, and shared this skill — thank you. This project blew up in less than 24 hours, and the support from the community has been incredible.

If this skill helps you work smarter, that's all I wanted.

</details>

<details> <summary><strong>🌍 What the community shipped</strong></summary>

Forks & Extensions

Fork	Author	What They Built
devis	@st01cs	Interview-first workflow, /devis:intv and /devis:impl commands, guaranteed activation
multi-manus-planning	@kmichels	Multi-project support, SessionStart git sync
plan-cascade	@Taoidle	Multi-level task orchestration, parallel execution, multi-agent collaboration
agentfund-skill	@RioTheGreat-ai	Crowdfunding for AI agents with milestone-based escrow on Base
openclaw-github-repo-commander	@wd041216-bit	7-stage GitHub repo audit, optimization, and cleanup workflow for OpenClaw

Used in the Wild

Project	What It Is
lincolnwan/Planning-with-files-copilot-agent	Entire Copilot agent repo built around the planning-with-files skill
cooragent/ClarityFinance	AI finance agent framework — Planning-with-Files approach directly credited
oeftimie/vv-claude-harness	Claude Code harness built on Manus-style persistent markdown planning
jessepwj/CCteam-creator	Multi-agent team orchestration skill using file-based planning

Skill Registries & Hubs

Registry	What It Is
buzhangsan/skill-manager	Bilingual (EN/中文) skill hub indexing 31,000+ Claude Code skills — planning-with-files installable one-click.

Built something? Open an issue to get listed!

</details>

<details> <summary><strong>🤝 Contributors</strong></summary>

See the full list of everyone who made this project better in CONTRIBUTORS.md.

</details>

<details> <summary><strong>📦 Releases & Session Recovery</strong></summary>

Current Version: v3.4.1

Version	Highlights
v3.5.1	Codex Windows shell resolver skips WSL bash launchers, pwf-hook.cmd hardens Python discovery, and Pi recitations are delivered as nextTurn so interactive tools are not broken.
v3.5.0	Codex Windows hooks emit valid JSON and survive Unicode (PR #205 by @yolo0731, closes #204), the Pi extension stops re-nagging closed and complete plans (#203 by @ziyu4huang), and the plan lifecycle is documented (#202 by @kcinzgg). On Windows the Codex front door forwarded plain stdout where Codex expects hookSpecificOutput.additionalContext or PreCompact JSON, and UTF-8 broke on both decode and encode; the fix serializes each event in its supported shape with ASCII-safe output, decodes shell output as UTF-8, and writes the .active_plan pointer without a BOM, with the containment resolver hardened to fail closed (audited clean before merge). The Pi resolveNewestPlanDir now ranks by task_plan.md file mtime instead of directory mtime and honors the close marker, so a finished plan no longer loses to an incomplete sibling. Four language commands that referenced a non-existent skill namespace were corrected, a Traditional Chinese /plan-zht command was added, and the README now documents the full v3 command, hook, and mode surface that had shipped but was never listed.
v3.4.1	Codex hooks now run on Windows (closes #201, reported by @mahdiit). The .codex/hooks.json commands were POSIX only, so on Windows the native command interpreter choked on python3 (the Store alias), 2>/dev/null, and the `\	\	true success guard whose true is not a Windows command, and every hook reported hook exited with code 1. Each hook now carries a per-hook commandWindows override (the POSIX command is untouched, so macOS and Linux are unchanged) that routes through a new pwf-hook.cmd launcher (real Python via py -3, falling back to python, never the Store python3 alias) and, for the three shell hooks, a new run_sh.py front door. codex_hook_adapter resolves the Git for Windows sh.exe by anchoring on git.exe, so the shell scripts run even when Git's usr\bin` is off PATH. Codex hook suites green (17 tests) plus new cross-platform and Windows-only coverage.
v3.4.0	PLANNING_DISABLED=1 per-invocation opt-out for one-shot sessions (closes #195, reported by @marcmuon). One-shot codex exec runs (CI review bots, read-only research agents, nested orchestrators) that merely shared a cwd with an incomplete plan were hijacked: plan context injected, the actual output redirected into progress.md, a fabricated completed phase appended to task_plan.md. All Codex hook entry points, the Python adapter route, and the canonical dispatchers (inject-plan.sh, gate-stop.sh, check-complete.sh/.ps1) now exit before reading the plan when PLANNING_DISABLED=1 is set; PreToolUse keeps emitting its allow decision so tool calls proceed. The guard ships in every distributed copy including the five language variants. Also corrected docs/codex.md's stale claim that the Stop hook blocks (non-blocking since v3.1.0). Suite at 200 passed.
v3.3.0	Pi extension hooks now wait for explicit user approval via /plan-execute (PR #193 by @Dikshj, closes #190, requested by @lazyst). The Pi extension previously activated all hooks as soon as task_plan.md existed, so plan injection, pre-tool recitation, post-write reminders, and auto-continue could start while the user was still reviewing a draft plan. The extension now stays passive with a status line until the active plan is approved with /plan-execute; approval is scoped to the current session and plan path, resets on session lifecycle events, and /plan-execute reset returns to passive review. A plan with a tampered SHA-256 attestation cannot be approved. Suite at 188 passed plus 21 Pi extension vitest tests.
v3.2.0	Repository health audit: session-catchup.py was non-functional on Windows, plus the "0/0 phases" false status (closes #191, addresses #103, closes #188). session-catchup.py, the mechanism behind "resume after /clear," never sanitized Windows-style paths correctly and had no explicit encoding on three reads, so it silently did nothing on Windows with no error. inject-plan.sh's containment guard also silently dropped plan injection and tamper detection under 8.3 short-name or /tmp-aliased paths. Both fixed. Separately, check-complete.sh/.ps1 and three IDE-specific Stop hooks reported a false "0/0 phases complete" for any task_plan.md without ### Phase headings (#191, reported by @mixian939); fixed everywhere the pattern appeared, canonical scripts included. The --template analytics flag (v2.29.0) was silently falling back to default templates in every installed skill package; the analytics templates now actually ship where init-session.sh reads them. Merged PR #187 (@Stephen-abc: Windows test-encoding fixes, stale install-path docs) and PR #192 (@igorcosta: Autohand Code setup docs). Added SECURITY.md and enabled private vulnerability reporting (@AvitalAviv). Corrected AGENTS.md's contributor-PR-merge guidance, which told agents to squash-merge in a way that reassigns authorship. Suite at 186 passed, 5 skipped, 0 failed.
v3.1.3	Hotfix: SKILL.md frontmatter was invalid YAML in v3.1.2. The v3.1.2 description refresh added a colon, and the English SKILL.md keep description unquoted, so YAML rejected the frontmatter ("mapping values are not allowed here"), which could break skill loading and the model-triggering description. v3.1.3 quotes the description (matching the already-quoted translated variants; the parsed value is identical) across the canonical file and the seven English IDE variants, and adds tests/test_skill_frontmatter_valid.py to validate every SKILL.md frontmatter as YAML. Suite at 184 passed.
v3.1.2	Session-catchup command works outside the plugin runtime, .hermes parity, refreshed skill description (PR #186 by @shunfeng8421 closes #185, reported by @xwang118). The documented Restore Context command used ${CLAUDE_PLUGIN_ROOT}, which the plugin runtime sets only inside hook execution, so a skill-only install (npx skills add, Codex, Cursor) running it in a shell got an empty variable and a broken /scripts/... path. Now SKILL_DIR="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/skills/planning-with-files}" on the canonical file, .codebuddy, and the five language variants; the .hermes variant got the same fallback for $HERMES_HOME. The eight English SKILL.md descriptions were refreshed to lead with planning for AI coding agents and context-loss survival, with the Use when trigger and translated variants unchanged. Documentation only; suite at 180 passed.
v3.1.1	Codex verification command checks the canonical hooks feature flag (PR #184 by @Fat-Jan). The docs/codex.md verify block ran `codex features list \	rg '^codex_hooks\s', but Codex moved its canonical feature key from codex_hooks to hooks in 0.129.0 (openai/codex#20522). The alias still resolves in config, yet codex features list prints only hooks, so the old pattern matched nothing on current Codex and routed correctly configured users to the upgrade path. The command now greps ^(hooks\	codex_hooks)\s and the troubleshooting sentence covers both names, matching the hooks = true` guidance carried in the file since v2.39.0. Documentation only; suite at 180 passed.
v3.1.0	Codex Stop hook no longer blocks on an incomplete plan, native Codex PreCompact parity, Pi extension test suite, and accurate SHA-cache docs (PR #180 by @2023Anita closes #178, PR #181 by @GongYuanCaiJi, PR #175 and PR #174 by @mvanhorn close #163 and #164). The .codex Stop adapter dropped the {"decision":"block"} path that pushed Codex agents to auto-continue unfinished phases; it now emits an advisory progress-sync reminder only, matching the v3 principle that an incomplete plan alone never blocks a stop. The native .codex/hooks.json route gained the PreCompact hook (pre-compact.sh) it was missing relative to the canonical SKILL.md, dormant on runtimes that never fire the event. The Pi extension gained a TypeScript integration suite covering all eight lifecycle handlers, the four runtime modes, and the attestation gate. docs/perf-notes.md documents the attestation SHA cache, corrected to the v3 $XDG_CACHE_HOME/pwf-sha location. Suite at 180 passed.
v3.0.0	Autonomous and gated modes for long-running agentic runs, structured run ledger, opt-in completion gate (no breaking changes: with no mode marker the hooks produce byte-identical v2.43 output). init-session --autonomous drops the per-tool-call plan re-injection for strong models and keeps the turn-start injection; --gated adds a deliberate Stop-hook completion gate that blocks only when five conditions hold (gated mode, an in_progress phase, stop_hook_active false, block count under the cap, ledger progressed since the last block), so an incomplete plan alone never traps a session. New append-only JSONL run ledger (ledger-append, ledger-summary, phase-status, sh + ps1) replaces the raw progress.md tail in v3 modes with a fixed-shape summary. Attestation is default-on in v3 modes and unattested plan bodies are refused at injection. Per-session nonce delimiters, SHA cache moved to $XDG_CACHE_HOME/pwf-sha, realpath containment in the plan-dir resolver. Hook scalars replaced by thin dispatchers (inject-plan.sh, gate-stop.sh) shipped in both scripts/ locations. New templates/task_plan_autonomous.md with DependsOn/Owner/AcceptanceCheck fields, v2-to-v3 migration guide in MIGRATION.md, host capability tiers documented (hard block, follow-up inject, notify only). Suite at 178 passed plus location-parity and gate/ledger/init-mode/containment tests.
v2.43.0	CONTRIBUTING.md + OpenCode docs fix + .continue/.gemini/.kiro variant sync to parity (PR #171 by @Skulli485, issue #172 by @luyanfeng, issues #159/#160/#161): first CONTRIBUTING.md at repo root, auto-surfaced by GitHub in the PR creation flow. docs/opencode.md Quick Install switched from \git clone\ to \npx skills add\ after the manual-install block was found referencing a doubled path (planning-with-files/planning-with-files/SKILL.md). Three historically lagging IDE SKILL.md variants brought to v2.43.0 parity: .continue from v2.34.0 (9 versions behind), .gemini from v2.34.0 (9 versions behind), .kiro from v2.32.0-kiro (11 versions behind), preserving IDE-specific frontmatter, hook shapes, and Kiro Agent Skill layout.
v2.42.0	POSIX init-session.sh portability + plugin-vs-skill install transparency + Topic Handoff docs (PR #169 and PR #170 by @carterusedulm2-maker): init-session.sh and its 7 mirrors swap the [[ ]] bashism for POSIX [ ] so tests/test_init_session_slug.py runs cleanly under dash (Ubuntu) when the test invokes the script via sh rather than the bash shebang. Canonical SKILL.md gains an install-scope clarification: /plugin install ships the commands/ folder with /plan-goal and /plan-loop, but npx skills add (and ClawHub) do not. A manual fallback procedure for both wrappers is documented inline so skill-only sessions can produce the same effect by invoking Claude Code's native /goal and /loop primitives directly. docs/quickstart.md and docs/workflow.md add an optional Topic Handoff Pattern for very long-running operational topics (handoffs/<topic>.md alongside progress.md).
v2.41.0	Windows exec-bit test skip + attestation-locking docs (PR #167 by @gauravvojha, Issue #166; PR #168 by @CleanDev-Fix, Issue #165): test_script_permissions.py now skips on Windows with a class-level pytest.mark.skipif(sys.platform == "win32") since NTFS does not store POSIX executable bits; the 2 pre-existing Windows exec-bit failures (present since v2.34.1) are resolved. New dedicated docs/attestation-locking.md page documents the attest-plan.sh write path, the atomic temp-rename guarantee, the optional flock advisory lock, and the recommended slug-mode workflow for parallel sessions.
v2.40.1	Pi adapter SKILL.md sync gap + npm scope correction (PR #158 by @TomXPRIME): the .pi SKILL.md lagged the canonical Claude Code copy after v2.39.0; v2.40.1 backports Rule 7 (Continue After Completion), the Security Boundary section, the expanded Scripts section covering set-active-plan.sh/resolve-plan-dir.sh/attest-plan.sh plus the parallel task workflow, and the "Write web content to task_plan.md" anti-pattern row. The Pi npm package is renamed from the unscoped pi-planning-with-files to @tomxprime/planning-with-files, matching the package author's namespace; install docs updated accordingly. Author, repository, license, and bugs URLs preserved.
v2.40.0	Slug-mode resolution fixes + perf cache + KV-cache hygiene + Pi false-positive fix (9 items from the v2.40 R&D experiment): hook resolution order inverted so slug-mode wins over legacy root, .active_plan target dir + content validated against a safe-identifier regex, check-complete.sh honors $PLAN_ID and .active_plan, Pi extension isDangerousBashCommand swapped to a word-boundary regex array so benign git push origin <branch> no longer fires the warning, mtime-keyed SHA-256 cache cuts attestation-hook latency on Windows Git Bash, progress.md tail timestamps normalized for KV-cache prefix stability, resolve-plan-dir.sh mtime resolution made portable across GNU/BSD/macOS/Alpine/Git Bash with python+perl fallbacks, attest-plan.sh uses atomic temp-rename with optional flock to close the concurrent-writer race. 130 pass / 2 pre-existing Windows exec-bit fails, +20 new tests.
v2.39.0	Pi Coding Agent full hook parity extension + Codex hooks flag fix (PR #157 by @TomXPRIME, Issue #154 by @DLI1996): the .pi adapter ships a bundled TypeScript extension mapping eight Pi lifecycle events to the same behavior the skill provides on Claude Code, with a four-mode system (auto/parity/cache-safe/notify) that auto-detects DeepSeek and keeps the KV-cache prefix stable. Pi runtime reads the same .attestation file the canonical v2.37 attest-plan.sh writes, so attesting once locks the plan across both runtimes. Four slash commands (/plan-status, /plan-attest, /plan-goal, /plan-loop) mirror their Claude Code counterparts. Separately, docs/codex.md swaps from codex_hooks = true to hooks = true to match the current OpenAI canonical key, with an alias note so users on older configs are not pushed to migrate.
v2.38.1	Description field garbled in Claude Code skill picker (surfaced via Discussion #153 by @bmyury): hook commands embedded '---BEGIN PLAN DATA---' plan-injection delimiters; Claude Code's skill-discovery loader split frontmatter on the first --- and read the truncated value as the description. Swapped to ===BEGIN PLAN DATA=== / ===END PLAN DATA=== across canonical SKILL.md, all five language variants, the .codebuddy/.codex/.cursor adapter mirrors, and clawhub-upload. Hook execution and tamper attestation never affected; only the displayed metadata.
v2.38.0	Claude Code turn-loop integration + OpenCode SQLite fix: new PreCompact hook fires on /compact and autoCompact, surfaces a reminder to flush progress before compaction completes and prints the active Plan-SHA256 when attested. New /plan-goal slash command composes with Claude Code's /goal (v2.1.139, May 12 2026): derives a termination condition from the active plan. New /plan-loop composes with /loop (v2.1.72+): default 10-minute tick re-reads planning files and runs check-complete. New templates/loop.md for the bare /loop planning-aware default. Session-catchup rewritten for OpenCode's SQLite migration. Codex gets a PermissionRequest adapter that surfaces plan context at permission prompts.
v2.37.0	Hash attestation + parity bumper (closes #150, #151): /plan-attest locks task_plan.md with a SHA-256; hooks block injection on tamper. scripts/bump-version.py + parity test kill the "missed one variant" regression class behind v2.34.1, v2.36.0, v2.36.2, and v2.36.3. (thanks @oaabahussain!)
v2.36.3	Parallel planning scripts now ship in the skill: resolve-plan-dir.sh and set-active-plan.sh were missing from the installed skill in v2.36.0; now in canonical + all IDE mirrors + SKILL.md docs updated
v2.36.2	Canonical script sync (PR #149): skills/planning-with-files/scripts/init-session.sh was missing slug mode from v2.36.0; now synced with IDE mirrors + regression test. (thanks @voidborne-d!)
v2.36.1	Security hardening: Stop hook cache search removed, ExecutionPolicy Bypass changed to RemoteSigned, prompt injection delimiters added. (Gen Agent Trust Hub FAIL resolved)
v2.36.0	Parallel plan isolation + Codex session isolation (closes #146, #148): init-session.sh slug mode, set-active-plan.sh, resolve-plan-dir.sh, all Codex hooks route through resolver, session attachment gating. Hermes docs (closes #147): integration notes added to docs/hermes.md. 34 new tests. (thanks @githubYiheng, @09ashishkapoor, @shawnli1874!)
v2.35.1	Shebang portability fix: changed /bin/bash to /usr/bin/env bash in hook scripts, fixing compatibility on NixOS and other systems where bash is not at /bin/bash. (thanks @Emin017!)
v2.35.0	Hermes adapter + NLPM audit hardening: Hermes platform 17 support (thanks @bailob!), NLPM audit fixed Python PATH resolution, session-catchup injection cap, Pi PowerShell syntax (thanks @xiaolai!)
v2.34.1	Stop hook Windows portability fix (closes #133): export SD= failed in Windows Git Bash hook context; fallback path was wrong for plugin cache structure. Fixed across all 13 SKILL.md variants. (thanks @nazeshinjite!)
v2.34.0	Codex hooks fully restored (closes #132): .codex/hooks.json + lifecycle scripts back — SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop. Tessl CI for SKILL.md quality reviews. Exec bit fix. 4 missing contributors added. (thanks @Leon-Algo, @popey!)
v2.33.0	Multi-language expansion: Arabic, German, and Spanish skill variants added (thanks to community contributors!)
v2.32.0	Codex session catchup rewrite (thanks @ebrevdo!), Loaditout A-grade security badge, Stop hook Git Bash fix
v2.31.0	Codex hooks.json integration with full lifecycle hooks (thanks @Leon-Algo!)
v2.30.1	Fix: Codex script executable bits restored (thanks @Leon-Algo!)
v2.30.0	CLAUDE_SKILL_DIR variable, IDE configs moved to per-IDE branches, plugin.json bumped from 2.23.0
v2.29.0	Analytics workflow template: --template analytics flag for data exploration sessions (thanks @mvanhorn!)
v2.28.0	Traditional Chinese (zh-TW) skill variant (thanks @waynelee2048!)
v2.26.2	Fix: --- in hook commands broke YAML frontmatter parsing, hooks now register correctly
v2.26.1	Fix: session catchup after /clear, path sanitization on Windows + content injection (thanks @tony-stark-eth!)
v2.26.0	IDE audit: Factory hooks, Copilot errorOccurred hook, Gemini hooks, bug fixes
v2.18.2	Mastra Code hooks fix (hooks.json + docs accuracy)
v2.18.1	Copilot garbled characters complete fix
v2.18.0	BoxLite sandbox runtime integration
v2.17.0	Mastra Code support + all IDE SKILL.md spec fixes
v2.16.1	Copilot garbled characters fix: PS1 UTF-8 encoding + bash ensure_ascii (thanks @Hexiaopi!)
v2.16.0	GitHub Copilot hooks support (thanks @lincolnwan!)
v2.27.0	Kiro Agent Skill layout (thanks @EListenX!)
v2.15.1	Session catchup false-positive fix (thanks @gydx6!)
v2.15.0	/plan:status command, OpenCode compatibility fix
v2.14.0	Pi Agent support, OpenClaw docs update, Codex path fix
v2.11.0	/plan command for easier autocomplete
v2.10.0	Kiro steering files support
v2.7.0	Gemini CLI support
v2.2.0	Session recovery, Windows PowerShell, OS-aware hooks

View all releases · CHANGELOG

Parallel plan isolation (.planning/YYYY-MM-DD-slug/ directories) and Codex session isolation shipped in v2.36.0. The experimental/isolated-planning branch was the earlier prototype; master is now the canonical location.

---

Session Recovery

When your context fills up and you run /clear, this skill automatically recovers your previous session.

How it works:

1. Checks for previous session data in the active IDE's session store (~/.claude/projects/ for Claude Code, ~/.codex/sessions/ for Codex)
2. Finds when planning files were last updated
3. Extracts conversation that happened after (potentially lost context)
4. Shows a catchup report so you can sync

Pro tip: Disable auto-compact to maximize context before clearing:

{ "autoCompact": false }

</details>

<details> <summary><strong>🛠️ Supported IDEs (18+ Platforms)</strong></summary>

Enhanced Support (hooks + lifecycle automation)

These IDEs have dedicated hook configurations that automatically re-read your plan before tool use, remind you to update progress, and verify completion before stopping:

IDE	Installation Guide	Integration
Claude Code	Installation	Plugin + SKILL.md + Hooks
Cursor	Cursor Setup	Skills + hooks.json
GitHub Copilot	Copilot Setup	Hooks (incl. errorOccurred)
Mastra Code	Mastra Setup	Skills + Hooks
Gemini CLI	Gemini Setup	Skills + Hooks
Kiro	Kiro Setup	Agent Skills
Codex	Codex Setup	Skills + Hooks
Hermes Agent	Hermes Setup	Skill + Project Plugin
CodeBuddy	CodeBuddy Setup	Skills + Hooks
FactoryAI Droid	Factory Setup	Skills + Hooks
OpenCode	OpenCode Setup	Skills + Custom session storage

Standard Agent Skills Support

These IDEs implement the Agent Skills open specification. Install with npx skills add — the installer places the skill in each IDE's discovery path automatically:

IDE	Installation Guide	Skill Discovery Path
Continue	Continue Setup	.continue/skills/ + .prompt files
Pi Agent	Pi Agent Setup	.pi/skills/ (npm package)
OpenClaw	OpenClaw Setup	.openclaw/skills/ (docs)
Autohand Code	Autohand Code Setup	~/.autohand/skills/ or .autohand/skills/
Antigravity	Antigravity Setup	.agent/skills/ (docs)
Kilocode	Kilocode Setup	.kilocode/skills/ (docs)
AdaL CLI (Sylph AI)	AdaL Setup	.adal/skills/ (docs)

Note: If your IDE uses the legacy Rules system instead of Skills, see the legacy-rules-support branch.

</details>

<details> <summary><strong>🧱 Sandbox Runtimes (1 Platform)</strong></summary>

Runtime	Status	Guide	Notes
BoxLite	✅ Documented	BoxLite Setup	Run Claude Code + planning-with-files inside hardware-isolated micro-VMs

Note: BoxLite is a sandbox runtime, not an IDE. Skills load via ClaudeBox — BoxLite’s official Claude Code integration layer.

</details>

---

A Claude Code plugin that transforms your workflow to use persistent markdown files for planning, progress tracking, and knowledge storage — the exact pattern that made Manus worth billions.

Quick Install

npx skills add OthmanAdi/planning-with-files --skill planning-with-files -g

<details> <summary><strong>🌐 Available in 5 other languages</strong></summary>

🇸🇦 العربية / Arabic

npx skills add OthmanAdi/planning-with-files --skill planning-with-files-ar -g

🇩🇪 Deutsch / German

npx skills add OthmanAdi/planning-with-files --skill planning-with-files-de -g

🇪🇸 Español / Spanish

npx skills add OthmanAdi/planning-with-files --skill planning-with-files-es -g

🇨🇳 中文版 / Chinese (Simplified)

npx skills add OthmanAdi/planning-with-files --skill planning-with-files-zh -g

🇹🇼 正體中文版 / Chinese (Traditional)

npx skills add OthmanAdi/planning-with-files --skill planning-with-files-zht -g

</details>

Works with Claude Code, Cursor, Codex, Gemini CLI, Autohand Code, and 40+ agents supporting the Agent Skills spec.

<details> <summary><strong>🔧 Claude Code Plugin (Advanced Features)</strong></summary>

For Claude Code-specific features like /plan autocomplete commands:

/plugin marketplace add OthmanAdi/planning-with-files
/plugin install planning-with-files@planning-with-files

</details>

That's it! Now use one of these commands in Claude Code:

Command	Autocomplete	Description
/planning-with-files:plan	Type /plan	Start planning session (v2.11.0+)
/planning-with-files:status	Type /plan:status	Show planning progress at a glance (v2.15.0+)
/planning-with-files:start	Type /planning	Original start command
/planning-with-files:pwf	Type /pwf	Short alias for /plan; supports --autonomous / --gated init (v3.0.0+)
/planning-with-files:plan-goal	Type /plan-goal	Run until the plan reports complete by composing with Claude Code /goal (v2.38.0+)
/planning-with-files:plan-loop	Type /plan-loop	Planning-aware cadence on Claude Code /loop, default 10m tick (v2.38.0+)
/planning-with-files:plan-attest	Type /plan-attest	Lock task_plan.md with a SHA-256 attestation; --show / --clear (v2.37.0+)
/planning-with-files:plan-de	Type /plan-de	Start planning in German (also -ar, -es, -zh, -zht) (v2.33.0+)

Typing /plan prefix-matches every plan* command in autocomplete; /planning-with-files:status autocompletes as /status (the older /plan:status label predates the rename). On the Pi extension the status command is named plan-status.

Alternative: If you want /planning-with-files (without prefix), copy skills to your local folder:

macOS/Linux:

cp -r ~/.claude/plugins/cache/planning-with-files/planning-with-files/*/skills/planning-with-files ~/.claude/skills/

Windows (PowerShell):

Copy-Item -Recurse -Path "$env:USERPROFILE\.claude\plugins\cache\planning-with-files\planning-with-files\*\skills\planning-with-files" -Destination "$env:USERPROFILE\.claude\skills\"

See docs/installation.md for all installation methods.

Pi extension commands

Install the Pi extension with pi install npm:@tomxprime/planning-with-files; it registers these commands, and on Pi they are typed with no /planning-with-files: prefix.

Command	What it does	Version
/plan-execute	Pi only. Approve the active plan to ACTIVATE all Pi hooks; hooks stay passive until you run this; reset returns to passive review	v3.3.0+
/plan-status	Active plan path, scope, and phase totals	v2.39.0+
`/plan-goal <text\	default\	clear>`	Set or clear the goal string appended to auto-continue prompts	v2.39.0+
`/plan-loop [interval] [prompt\	stop]`	Start or stop a planning tick (default 10m) that re-reads the plan and nudges progress	v2.39.0+
`/plan-attest [--show\	--clear]`	Run the attest-plan helper; shares the .attestation file with Claude Code	v2.39.0+

On Pi there is no /plan command to create the files; the skill creates them, then /plan-execute approves and activates the hooks. Note that Pi plan-goal/plan-loop run their own logic, while the Claude Code commands of the same name forward to native /goal and /loop.

Command names vs skill names

The command you TYPE differs by platform. Claude Code uses /planning-with-files:<verb> (for example /plan, /plan-de, /plan-goal, /plan-attest, /pwf, /status). Pi uses the bare form (/plan-status, /plan-execute, /plan-goal, /plan-loop, /plan-attest), no prefix. Continue.dev uses /planning-with-files. The underlying model-invocable SKILLS are named planning-with-files:planning-with-files (and -de, -es, -ar, -zh, -zht); that doubled form is the skill id, not a command you type. There is no /pwf-de and no /planning-with-files:planning-with-files-goal; /pwf is just a short alias for /plan.

Why This Skill?

On December 29, 2025, Meta acquired Manus for $2 billion. In just 8 months, Manus went from launch to $100M+ revenue. Their secret? Context engineering.

"Markdown is my 'working memory' on disk. Since I process information iteratively and my active context has limits, Markdown files serve as scratch pads for notes, checkpoints for progress, building blocks for final deliverables." — Manus AI

The Problem

Claude Code (and most AI agents) suffer from:

• Volatile memory — TodoWrite tool disappears on context reset
• Goal drift — After 50+ tool calls, original goals get forgotten
• Hidden errors — Failures aren't tracked, so the same mistakes repeat
• Context stuffing — Everything crammed into context instead of stored

The Solution: 3-File Pattern

For every complex task, create THREE files:

task_plan.md      → Track phases and progress
findings.md       → Store research and findings
progress.md       → Session log and test results

The Core Principle

Context Window = RAM (volatile, limited)
Filesystem = Disk (persistent, unlimited)

→ Anything important gets written to disk.

The Manus Principles

Principle	Implementation
Filesystem as memory	Store in files, not context
Attention manipulation	Re-read plan before decisions (hooks)
Error persistence	Log failures in plan file
Goal tracking	Checkboxes show progress
Completion verification	Stop hook checks all phases

Usage

Once installed, the AI agent will:

1. Ask for your task if no description is provided
2. Create task_plan.md, findings.md, and progress.md in your project directory
3. Re-read plan before major decisions (via PreToolUse hook in legacy mode; autonomous mode injects at session start and phase transitions instead, dropping the per-tool-call tax on strong models)
4. Remind you to update status after file writes (via PostToolUse hook)
5. Store findings in findings.md instead of stuffing context
6. Log errors for future reference
7. Verify completion before stopping (via Stop hook)

Invoke with:

• /planning-with-files:plan - Type /plan to find in autocomplete (v2.11.0+)
• /planning-with-files:pwf - Type /pwf, short alias for /plan (v3.0.0+)
• /planning-with-files:start - Type /planning to find in autocomplete
• /planning-with-files - Only if you copied skills to ~/.claude/skills/

For the goal, loop, attest, language, and Pi commands, see the command tables above.

See docs/quickstart.md for the full 5-step guide.

v3 Long-Running Agent Features

The v3 line adds features aimed at long-running agentic runs. Each one is listed with the command or flag that turns it on. With no mode marker set, the hooks produce the same output as v2.43, so nothing changes for existing setups.

• Autonomous mode (/pwf --autonomous, or init-session.sh --autonomous): drops the per-tool-call plan recitation, keeps the turn-start injection, and turns attestation on by default.
• Gated mode (--gated): adds a Stop completion gate that blocks only when all completion conditions hold at once, so an incomplete plan alone never traps a session.
• Auto-continue on Pi (agent_end handler): re-prompts the agent up to a limit of 3 to keep an unfinished plan moving, plus an optional /plan-goal string appended to the prompt.
• Pi approval gate (/plan-execute): Pi hooks stay passive with a status line until you approve the active plan for the current session.
• Session-catchup: resumes work after /clear by re-reading the planning files from the active IDE's session store.
• PreCompact progress flush (PreCompact hook): surfaces a reminder to flush progress before compaction completes, and prints the active Plan-SHA256 when attested.
• SHA-256 plan attestation (/plan-attest): locks task_plan.md; a tampered plan body is refused at injection.
• Run ledger: an append-only JSONL record of phase transitions that replaces the raw progress.md tail in v3 modes with a fixed-shape summary.
• Host capability tiers: hard block on Claude Code, Codex, and Continue; follow-up injection on Cursor, Pi, and Kiro; notify-only elsewhere.
• Per-invocation opt-out (PLANNING_DISABLED=1, v3.4.0): a one-shot session that merely shares a cwd with an incomplete plan skips all plan reading at every hook entry point.

Hooks and modes reference

Platform	Lifecycle hooks	Where registered
Claude Code	5: UserPromptSubmit, PreToolUse, PostToolUse, Stop, PreCompact	The skill's SKILL.md frontmatter (not plugin.json), so they ship with the bundled skill
Codex CLI	7: SessionStart, UserPromptSubmit, PreToolUse, PermissionRequest, PostToolUse, PreCompact, Stop	.codex/hooks.json, Windows-safe via commandWindows since v3.4.1
Pi	8 lifecycle handlers in the bundled extension	The injection and recitation handlers stay passive until /plan-execute

Pi runtime modes:

Pi mode	Behavior
auto	Detects the model and picks parity or cache-safe
parity	Full plan injection, mirrors the Claude Code skill
cache-safe	A stable reminder instead of full injection, for KV-cache-sensitive models like DeepSeek
notify	Status-line only, no model injection

Benchmark Results

Methodology note: the 96.7% figure comes from the v2.21.0 evaluation run on claude-sonnet-4-6 (2026-03-06). It measures file-pattern fidelity (does the agent create and maintain the 3-file structure), not goal-drift over long autonomous runs. Newer models and the autonomous-mode work are not yet covered by this number. Full methodology, dataset, and assertion list: docs/evals.md.

Formally evaluated using Anthropic's skill-creator framework (v2.22.0), model claude-sonnet-4-6, 2026-03-06. 10 parallel subagents, 5 task types, 30 objectively verifiable assertions, 3 blind A/B comparisons.

Test	with_skill	without_skill
Pass rate (30 assertions)	96.7% (29/30)	6.7% (2/30)
3-file pattern followed	5/5 evals	0/5 evals
Blind A/B wins	3/3 (100%)	0/3
Avg rubric score	10.0/10	6.8/10

Full methodology and results · Technical write-up

Key Rules

1. Create Plan First — Never start without task_plan.md
2. The 2-Action Rule — Save findings after every 2 view/browser operations
3. Log ALL Errors — They help avoid repetition
4. Never Repeat Failures — Track attempts, mutate approach

When to Use

Use this pattern for:

• Multi-step tasks (3+ steps)
• Research tasks
• Building/creating projects
• Tasks spanning many tool calls

Skip for:

• Simple questions
• Single-file edits
• Quick lookups

File Structure

planning-with-files/
├── commands/                # Plugin commands
│   ├── plan.md              # /planning-with-files:plan command (v2.11.0+)
│   ├── pwf.md               # /planning-with-files:pwf short alias for /plan (v3.0.0+)
│   ├── status.md            # /planning-with-files:status progress view (v2.15.0+)
│   ├── plan-goal.md         # /planning-with-files:plan-goal, composes with /goal (v2.38.0+)
│   ├── plan-loop.md         # /planning-with-files:plan-loop, composes with /loop (v2.38.0+)
│   ├── plan-attest.md       # /planning-with-files:plan-attest SHA-256 lock (v2.37.0+)
│   ├── plan-ar.md           # Arabic /plan command (v2.33.0+)
│   ├── plan-de.md           # German /plan command (v2.33.0+)
│   ├── plan-es.md           # Spanish /plan command (v2.33.0+)
│   ├── plan-zh.md           # Chinese Simplified /plan command (v2.33.0+)
│   ├── plan-zht.md          # Chinese Traditional /plan command (v2.33.0+)
│   └── start.md             # /planning-with-files:start command
├── templates/               # Root-level templates (for CLAUDE_PLUGIN_ROOT)
├── scripts/                 # Root-level scripts (for CLAUDE_PLUGIN_ROOT)
├── docs/                    # Documentation
│   ├── installation.md
│   ├── quickstart.md
│   ├── workflow.md
│   ├── troubleshooting.md
│   ├── gemini.md            # Gemini CLI setup
│   ├── cursor.md
│   ├── windows.md
│   ├── kilocode.md
│   ├── codex.md
│   ├── opencode.md
│   ├── mastra.md             # Mastra Code setup
│   └── boxlite.md            # BoxLite sandbox setup
├── examples/                # Integration examples
│   └── boxlite/             # BoxLite quickstart
│       ├── README.md
│       └── quickstart.py
├── planning-with-files/     # Plugin skill folder
│   ├── SKILL.md
│   ├── templates/
│   └── scripts/
├── skills/                  # Skill variants
│   ├── planning-with-files/     # English (default)
│   │   ├── SKILL.md
│   │   ├── examples.md
│   │   ├── reference.md
│   │   ├── templates/
│   │   └── scripts/
│   │       ├── init-session.sh
│   │       ├── check-complete.sh
│   │       ├── init-session.ps1   # Windows PowerShell
│   │       └── check-complete.ps1 # Windows PowerShell
│   ├── planning-with-files-ar/   # Arabic (v2.33.0+)
│   │   ├── SKILL.md
│   │   ├── templates/
│   │   └── scripts/
│   ├── planning-with-files-de/   # German (v2.33.0+)
│   │   ├── SKILL.md
│   │   ├── templates/
│   │   └── scripts/
│   ├── planning-with-files-es/   # Spanish (v2.33.0+)
│   │   ├── SKILL.md
│   │   ├── templates/
│   │   └── scripts/
│   ├── planning-with-files-zh/   # Chinese Simplified (v2.25.0+)
│   └── planning-with-files-zht/  # Chinese Traditional (v2.28.0+)
├── .gemini/                 # Gemini CLI skills + hooks
│   ├── settings.json        # Hook configuration (v2.26.0)
│   ├── hooks/               # Hook scripts (SessionStart, BeforeTool, AfterTool, BeforeModel, SessionEnd)
│   └── skills/
│       └── planning-with-files/
├── .codex/                  # Codex CLI skills + hooks
│   └── skills/
├── .opencode/               # OpenCode skills (custom session storage)
│   └── skills/
├── .claude-plugin/          # Plugin manifest
├── .cursor/                 # Cursor skills + hooks
│   ├── hooks.json           # Hook configuration
│   ├── hooks/               # Hook scripts (bash + PowerShell)
│   └── skills/
├── .codebuddy/              # CodeBuddy skills + hooks
│   └── skills/
├── .factory/                # FactoryAI Droid skills + hooks (v2.26.0)
│   └── skills/
├── .pi/                     # Pi Agent skills (npm package)
│   └── skills/
│       └── planning-with-files/
├── .continue/               # Continue.dev skills + prompt files
│   ├── prompts/             # .prompt file for slash commands
│   └── skills/
├── .github/                 # GitHub Copilot hooks (incl. errorOccurred)
│   └── hooks/
│       ├── planning-with-files.json  # Hook configuration
│       └── scripts/         # Hook scripts (bash + PowerShell)
├── .mastracode/             # Mastra Code skills + hooks
│   └── skills/
├── .kiro/                   # Kiro Agent Skills (v2.27.0+)
│   └── skills/
├── CHANGELOG.md
├── CITATION.cff
├── LICENSE
└── README.md

FAQ

How do I stop my coding agent from losing its plan after /clear or a crash?

The plan lives on disk in task_plan.md, findings.md, and progress.md, not only in the context window. At the start of each turn the UserPromptSubmit hook re-injects the active plan, and after a /clear or a new session the skill re-reads the files from disk (session recovery), so the agent recovers its goals and progress automatically.

What is the difference between planning-with-files and an agent memory tool?

Agent memory tools (vector stores, knowledge graphs) help an agent recall facts from past sessions. planning-with-files manages active execution state: the phases, status, dependencies, and completion check for the task the agent is working on right now. The problem it solves is planning continuity, not retrieval, and the two are complementary.

How does this prevent context rot?

Context rot is the drift that sets in as the context window fills and earlier instructions get crowded out. Because the plan is re-injected at the start of each turn from disk, the goals and phase status stay in the model's attention window as the conversation grows. This is an implementation of what Anthropic calls structured note-taking: write durable state to files outside the window, then read it back in when needed.

Which coding agents does this work with?

Claude Code, OpenAI Codex CLI, Cursor, GitHub Copilot, Kiro, OpenCode, Continue, Pi, CodeBuddy, Factory, Mastra, and 60+ others via the SKILL.md open standard. Installation is one command; see Quick Install above.

What happens to the plan files after a task is complete?

They are working memory, not a tracked deliverable. task_plan.md, findings.md, progress.md, and the .planning/ directory are gitignored by default and are not archived automatically: the next task overwrites the root plan, and a slug directory just stops being active. Anything worth keeping should be promoted into code, a commit, or a doc. See After Completion: What Happens to the Plan Files for the full lifecycle and how to retain a completed plan. This is a deliberate default, not a missing feature; a completion-triggered archive step is a welcome opt-in extension.

Documentation

All platform setup guides and documentation are in the docs/ folder.

Acknowledgments

• Manus AI — For pioneering context engineering patterns
• Anthropic — For Claude Code, Agent Skills, and the Plugin system
• Lance Martin — For the detailed Manus architecture analysis
• Based on Context Engineering for AI Agents

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Submit a pull request

License

MIT License — feel free to use, modify, and distribute.

---

Author: Ahmad Othman Ammar Adi

Star History

<a href="https://repostars.dev/?repos=OthmanAdi%2Fplanning-with-files&theme=copper"><img src="https://repostars.dev/api/embed?repo=OthmanAdi%2Fplanning-with-files&theme=copper" width="100%" alt="Star History Chart" /></a>