world-model-ledger

One-time SETUP: installs a persistent SQLite-backed WORLD MODEL for a coding agent into a project (or global ~/.claude). Use when the user wants an agent to remember a codebase across sessions, detect contradictions, propose fixes, and improve correctness over time — "give the agent a world model", "track what's verified vs assumed", "persist codebase knowledge with confidence". Tracks entities (symbols/files/modules/external referents), interactions, and constraints with TWO confidence axes (observed vs normative), validation status, and PROV-style evidence. Load-bearing rule: code observation is NOT ground truth — only oracle evidence (tests/CI/docs/human) raises normative confidence, flagging observed-but-unverified. Every triple is checked against a predicate ontology (RDFS-style domain/range) before insert — hallucinated verbs and impossible pairings are rejected, not stored. Zero-config capture via four hooks (PreToolUse pre-edit summaries; UNIVERSAL PostToolUse observing every tool call; Stop/SessionSt

インストールへ移動

リポジトリ: dhanesh/agent-skills
ソースの最終更新活動: 2026年7月27日 18:21
検出された SKILL.md の言語: 英語
スター: 1
フォーク: 0

インストール方法

デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。

npx skills add https://github.com/dhanesh/agent-skills --skill world-model-ledger

コマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。

ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。

ソースファイルを確認

インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。

ファイルエクスプローラー

25 ファイル

SKILL.md

ソースの指示 · 読み取り専用プレビュー

検索に戻るインストールに戻る

feynman-walkthrough

Walk a person through a subject, codebase, whitepaper, or complex system until they genuinely understand it. Use whenever the user wants to understand or learn something: "explain X", "walk me through this repo/paper", "help me understand how Y works", "onboard me onto Z" — even if they never mention a learning method. Explains with Feynman-grade simplicity (plain language, analogy, concrete example before abstraction) and research-backed structure (big picture first, one segment at a time, checks matched to the goal), gauges understanding only AFTER the walkthrough to see where more coaching helps, and produces a standalone, sharable reference explainer — persisted as an Open Knowledge Format (OKF) bundle with source fingerprints pinned, so later sessions can review the knowledge and refresh it diff-aware when the codebase, document, or topic changes. A guided walkthrough, not a study regimen: retention tooling (spaced retrieval, self-quizzing, review schedules) exists but is strictly opt-in.

2026-07-28

crafting-self-prompting-loops

Design and build sound self-prompting loops — loops where a model iterates toward a goal across turns (draft→critique→revise, autonomous task loops, multi-agent orchestration, human-checkpointed loops). Use whenever the user wants to make a model "keep improving X until Y", build a self-refinement or reflexion loop, an autonomous agent that works toward a goal with little supervision, a loop that orchestrates or fans out to other agents, a recurring self-driven/scheduled task, or any agentic loop — even if they don't say the word "loop"; also use it to audit or fix an existing loop that runs away, never stops, oscillates, drifts off-goal, or is vulnerable to prompt injection. It produces a filled loop spec, a runnable scaffold, and bakes in the mandatory safety properties (a hard stop backstop and the trusted/untrusted two-channel boundary) that loops fail without.

2026-07-27

context-hygiene-kit

One-time SETUP that installs durable context management into a Claude Code project (or global ~/.claude): a bounded, scored, tiered cache that keeps working memory lean (anti-bloat) and rot-proof across compactions (anti-rot). Use when setting up, installing, configuring, verifying, or uninstalling context hygiene for a project or globally; when the user asks to "set up context management", "stop my context from bloating", "avoid context rot", "make Claude remember decisions across compaction", "persist working memory across sessions", or to replace a lossy local-model session-summariser with a deterministic lossless one. It drops in a stdlib-only Python ledger, a deterministic per-turn transcript harvester, and three lifecycle hooks (Stop / PreCompact / SessionStart) wired into settings.json — after install those hooks run automatically every turn, so you invoke this ONCE. Not an ongoing per-turn skill; not for ordinary app caching, RAG vector stores, or LLM prompt-caching config.

2026-07-27

このリポジトリのすべてを見る

world-model-ledger

A persistent world model for a coding agent, backed by SQLite. It records what the agent learns about a codebase — entities (symbols / files / modules / real-world referents), the interactions between them, and the constraints that should hold — and, for every interaction and constraint, how sure we are it exists, how sure we are it is correct, whether it has been validated, and what evidence backs it.

The defining rule, enforced as an invariant: code-observed relationships are not ground truth. A file sighting can drive observed_conf to 1.0 while normative_conf stays 0 and status stays unverified. Only oracle evidence — a passing test, CI, a doc, a human — raises normative confidence. See references/confidence-model.md for the epistemics and references/schema.md for the data model. The design rationale is in docs/superpowers/specs/2026-07-01-world-model-ledger-design.md.

Install & operate

Detect the install state first — before asking anything. The skill is already installed if ~/.claude/world-model-ledger/wm.py exists (GLOBAL) or ./wm.py + ./world_model.py exist in the project (PROJECT). Then route by how you were invoked; the full flag table is in references/parameters.md.

--seed / --prune — operate on an existing install (NO scope question, NO re-install)

If the skill is already installed and you were asked to seed or prune, do exactly that against the current repo and stop — do not run the full installer and do not ask which scope to use (that decision was already made at install time):

scripts/install.sh --seed # build THIS repo's model (files + structural edges) scripts/install.sh --prune # build + soft-invalidate edges for deleted/renamed files

These short-circuit to python3 "$WM/wm.py" build . [--prune] (where $WM is . for a project install, else ~/.claude/world-model-ledger), create and gitignore .world-model/, and report entities_added / interactions_added. They are idempotent — safe to re-run anytime on an installed model.

world-model-ledger

Install & operate

--seed / --prune — operate on an existing install (NO scope question, NO re-install)

scripts/install.sh --seed # build THIS repo's model (files + structural edges) scripts/install.sh --prune # build + soft-invalidate edges for deleted/renamed files

name	world-model-ledger
description	One-time SETUP: installs a persistent SQLite-backed WORLD MODEL for a coding agent into a project (or global ~/.claude). Use when the user wants an agent to remember a codebase across sessions, detect contradictions, propose fixes, and improve correctness over time — "give the agent a world model", "track what's verified vs assumed", "persist codebase knowledge with confidence". Tracks entities (symbols/files/modules/external referents), interactions, and constraints with TWO confidence axes (observed vs normative), validation status, and PROV-style evidence. Load-bearing rule: code observation is NOT ground truth — only oracle evidence (tests/CI/docs/human) raises normative confidence, flagging observed-but-unverified. Every triple is checked against a predicate ontology (RDFS-style domain/range) before insert — hallucinated verbs and impossible pairings are rejected, not stored. Zero-config capture via four hooks (PreToolUse pre-edit summaries; UNIVERSAL PostToolUse observing every tool call; Stop/SessionStart consolidate + inject digest, auto-bootstrap); invoke ONCE; hooks run automatically. Not a linter, LSP server, RAG vector store, or model-driven fact extractor.
license	MIT
compatibility	Requires Claude Code lifecycle hooks (PreToolUse/PostToolUse/Stop/SessionStart), bash, and python3 with its stdlib sqlite3 (no pip, no network); jq optional for clean settings.json merging.
x-spec-version	1
metadata	{"author":"dhanesh","version":"1.1.0","tags":"claude-code,hooks,world-model,sqlite,memory,confidence,provenance,contradictions,ontology"}

name	world-model-ledger
description	One-time SETUP: installs a persistent SQLite-backed WORLD MODEL for a coding agent into a project (or global ~/.claude). Use when the user wants an agent to remember a codebase across sessions, detect contradictions, propose fixes, and improve correctness over time — "give the agent a world model", "track what's verified vs assumed", "persist codebase knowledge with confidence". Tracks entities (symbols/files/modules/external referents), interactions, and constraints with TWO confidence axes (observed vs normative), validation status, and PROV-style evidence. Load-bearing rule: code observation is NOT ground truth — only oracle evidence (tests/CI/docs/human) raises normative confidence, flagging observed-but-unverified. Every triple is checked against a predicate ontology (RDFS-style domain/range) before insert — hallucinated verbs and impossible pairings are rejected, not stored. Zero-config capture via four hooks (PreToolUse pre-edit summaries; UNIVERSAL PostToolUse observing every tool call; Stop/SessionStart consolidate + inject digest, auto-bootstrap); invoke ONCE; hooks run automatically. Not a linter, LSP server, RAG vector store, or model-driven fact extractor.
license	MIT
compatibility	Requires Claude Code lifecycle hooks (PreToolUse/PostToolUse/Stop/SessionStart), bash, and python3 with its stdlib sqlite3 (no pip, no network); jq optional for clean settings.json merging.
x-spec-version	1
metadata	{"author":"dhanesh","version":"1.1.0","tags":"claude-code,hooks,world-model,sqlite,memory,confidence,provenance,contradictions,ontology"}

Hook	Fires	Job
PreToolUse → `hooks/pretooluse.sh`	before an edit	retrieve & summarize the model's ✓validated / ?unverified / ✗contradicted items for the touched file(s)/symbol(s); inject as context. Read-only.
PostToolUse → `hooks/posttooluse-observe.sh`	after every tool call	the universal observer: register the entities/edges the call reveals — files any tool reads/edits → entities; `Bash` → `runtime` `executes`/`reads` edges + verifier oracle (green→validated, red→contradicted); URLs → referents. Tool input only (never output); no invented facts.
Stop → `hooks/stop.sh`	every turn	the summarization home: harvest any optional markers from the trusted channel, consolidate (re-derive confidence + evaluate constraints), refresh `.world-model/digest.md`.
SessionStart → `hooks/session_start.sh`	session start/resume	auto-bootstrap the model on first run (create + seed `.world-model/`, no manual step), then inject the digest so a resumed session starts aware of contradictions.

world-model-ledger

インストール方法

ソースファイルを確認