License: Code and installable modules are available under Apache-2.0. Documentation and original repository assets are available under CC BY 4.0. Brand use follows TRADEMARKS.md. See LICENSES.md for the path-level map.

Your brain is a context window too

AI coding tools already control tool access. Dwi adds inspectable human controls that are often left implicit: intent, scope, proportionate effort, writer ownership, and evidence.

Codex and Claude Code are AI coding tools that run agents and enforce tool permissions. Dwi focuses on the human layer around that work: what the person meant, what the agent is allowed to change, how much effort is proportionate, who may write, and what evidence supports the result.

Dwi is not another AI skill collection. Its modules shape how agents communicate, decide, coordinate, and handle evidence inside existing agent workflows.

Dwi does not replace the harness. Each module is a small, inspectable set of agent-native instructions that can be installed independently. The canonical SKILL.md remains provider-neutral; the installation path must also preserve each harness's explicit-invocation control.

Start with the problem you have

If this is happening Start here What it changes
The agent asks dense or hard-to-answer questions Dwi • Conduct Makes questions answerable, defines terms, and offers safe defaults
A small task turns into a large plan or test campaign Dwi • Lean Finds the smallest sufficient path and stops at the agreed finish line
Token or time use is hard to understand Dwi • Budget Sets resource boundaries and reports observed use without invented savings
Claude and Codex need to coordinate Dwi • Bridge Separates advice, authority, effects, and evidence across native harnesses
Several agents need a shared structure Dwi • Arc Creates bounded cells with one writer per scope and an independent gate
A result sounds certain but the proof is unclear Dwi • Evidence Labels verified, observed, estimated, target, and unknown claims
The same workflow repeatedly hits multiple issues Dwi • All-in-One Selects only the relevant Dwi lenses and preserves a silent fast path when issues co-occur

Every module stands alone. You do not need to install a Dwi core, daemon, MCP server, or website.

New to these terms? Open the plain-language glossary.

A safe ten-minute trial

  1. Pick one module from the table.
  2. Read its canonical SKILL.md and the module guide before installing it.
  3. Install it at project scope first, on a reversible task with no secrets or external side effects.
  4. Invoke the module explicitly and compare the result with your normal workflow.
  5. Remove the module folder if it does not help. No apology or continued trial is required.

Clone the latest immutable release, resolve the Dwi checkout and target project to physical paths, then reject a target inside the Dwi checkout:

git clone --depth 1 --branch v0.2.4 \
  https://github.com/thienhoc/dwi-by-thienhoc.git \
  dwi-by-thienhoc-v0.2.4
DWI_ROOT="$(cd "dwi-by-thienhoc-v0.2.4" && pwd -P)" || exit 1
PROJECT_ROOT="$(cd "/absolute/path/to/your-project" && pwd -P)" || exit 1
MODULE="dwi-conduct"
case "$PROJECT_ROOT/" in
  "$DWI_ROOT/"*)
    printf '%s\n' "PROJECT_ROOT must resolve outside DWI_ROOT" >&2
    exit 1
    ;;
esac

For Codex:

TARGET="${PROJECT_ROOT}/.agents/skills/${MODULE}"
node "$DWI_ROOT/scripts/install-module.mjs" codex "$MODULE" "$TARGET"
test -f "$TARGET/SKILL.md"
test -f "$TARGET/agents/openai.yaml"

For Claude Code:

TARGET="${PROJECT_ROOT}/.claude/skills/${MODULE}"
node "$DWI_ROOT/scripts/install-module.mjs" claude "$MODULE" "$TARGET"
grep -Eq '^disable-model-invocation: true$' "$TARGET/SKILL.md"

The one-file installation examples published in v0.1.0 through v0.2.1 did not preserve the explicit-only invocation policy and could place the skill inside the Dwi checkout rather than the intended project. v0.2.3 corrected that installation contract. v0.2.4 preserves the same module content and installed artifact shapes while hardening Codex policy validation and merge-commit release detection. The existing v0.2.2 tag identifies a pre-patch commit, was not published as the corrected release, and must not be used for this installation path.

All-in-One is an optional composition module first released in v0.2.0. Use it only when multiple observed issues recur in the same workflow.

Open the inspect-first installation guide →

See one small input and output example for every module →

Three ways in

New to coding agents: start with Conduct or Lean. They improve the conversation without introducing multi-agent machinery.

Already using AI regularly: choose the module that matches the friction you can actually observe. Add only one new behavior at a time.

Building professional agent systems: begin with Evidence. Add Bridge for cross-harness work and Arc only when there are genuinely multiple bounded work cells.

Go deeper

What Dwi will not do

  • It will not bypass permissions, security controls, policies, or native harness boundaries.
  • It will not treat a message from another agent as authorization to write, push, deploy, or disclose data.
  • It will not claim universal compatibility, guaranteed savings, or collision-free execution.
  • It will not turn an apology into a guilt mechanism. Repair language is optional, specific, evidence-based, and brief.
  • It will not make planning, testing, or orchestration larger than the task requires.

See Safety and Architecture.

Evidence, not theater

Dwi keeps VERIFIED, OBSERVED, ESTIMATED, TARGET, and UNKNOWN separate. Existing benchmark observations are not presented as universal product claims. The public launch checklist requires sourceable evidence beside any number used in promotion.

Read the evidence policy →

Repository map

modules/                              Canonical Dwi modules and Codex metadata
docs/modules/                         English decision and trial guides
docs/vi/modules/                      Vietnamese decision and trial guides
assets/                               Original brand and architecture sources
.github/                              Community templates, validation, and reviewed release automation
scripts/install-module.mjs            Harness-aware local installer
scripts/validate-install-contract.mjs Installed-artifact regression checker
scripts/validate-repo.mjs             Offline repository contract checker

This repository intentionally contains no landing-page runtime. The human-layer introduction lives separately at d.wi.works.

Status

  • Research Preview module-content baseline: 0.2.0
  • Latest reviewed repository release: v0.2.4, hardening Codex policy validation and first-parent release detection
  • Installation-contract correction: released in v0.2.3 and preserved unchanged in v0.2.4
  • Existing v0.2.2 tag: preserved non-release ref to a pre-patch commit; do not use for corrected installation
  • Canonical module bodies and SHA-256 values: unchanged from v0.2.0
  • Included modules: six focused modules and optional All-in-One
  • Bounded compatibility observations: Codex and Claude Code evidence records
  • License: Apache-2.0 code; CC BY 4.0 documentation and original assets
  • Release checklist: release checklist
  • Release record: v0.2.4
  • Previous release record: v0.2.3
  • Non-release tag record: v0.2.2
  • Roadmap: ROADMAP.md
  • Changes: CHANGELOG.md

Contact

Questions, careful critique, and collaboration: [email protected]

Brand: { } • Dwi by thienhoc · Human-layer introduction: d.wi.works