Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/scottconverse/dev-rigor-stack-lite

Portable hook-free evidence-first development and release skills for Codex, Claude Code, Antigravity, and Agent Skills hosts
https://github.com/scottconverse/dev-rigor-stack-lite

Last synced: 9 days ago
JSON representation

Portable hook-free evidence-first development and release skills for Codex, Claude Code, Antigravity, and Agent Skills hosts

Awesome Lists containing this project

README 

          
# dev-rigor-stack-lite



A portable, evidence-first development and release workflow for AI coding agents. It

contains the complete 19-skill workflow from `codex-dev-rigor-stack`, adapted to run

without lifecycle hooks, a background runtime, trust activation, Stop interception, or a

private evidence ledger — plus two drift-resistance layers that **install by default

with everything else**: a small **anchor block** for the host's persistent instructions

file, and **rigor-goals**, a stdlib-only CLI whose exit gate refuses to close a

multi-story job without verification evidence.



These are part of the stack, not extras. Skills alone are advice — a model can drift

off them in a long session and nothing pushes back; the anchor and the goals gate are

what make the discipline hold. There is no on-switch because they are never off. There

IS an off-switch (`--no-anchor` / `--no-goals`, or deleting the anchor block) — and it

belongs to the human owner alone. An agent never passes the opt-outs or disables the

block on its own initiative; the anchor text itself carries that rule.



## The three tiers



Model attention decays over a long session and dies at context compaction; hooks fight

that with per-turn injection but are host-specific. Lite v2 instead moves the

discipline's memory into places that do not decay:



| Tier | What | Force | Why it resists drift |

|---|---|---|---|

| 1 | The 19 skills | none — invoked knowledge | — |

| 2 | Anchor block in `CLAUDE.md` / `AGENTS.md` / `GEMINI.md` | reminder every turn | the host re-reads its instructions file each turn |

| 3 | `rigor-goals` CLI | one hard gate at "done" | state lives in `./.rigor/` on disk — survives compaction and session death; the *refusal* is a program, not a prompt (what it checks is that evidence is named and recorded — see the precision note below) |



The anchor's one enforced habit — "multi-step work: run `rigor-goals create` FIRST" — is

what feeds Tier 3. One reminder feeding one gate, instead of continuous forcing.



## What changes in Lite



- Skills, references, templates, and deterministic helper scripts remain.

- Evidence comes from ordinary commands, logs, CI, screenshots, traces, hashes, and reports.

- Host policies control delegation, approvals, merges, publishing, and available tools.

- Passing a gate establishes readiness; it never grants the agent authority to merge or publish.

- Missing capabilities are reported as blocked or unverifiable rather than silently passed.



## Compatibility



The bundle uses the portable Agent Skills layout: one directory per skill with a

`SKILL.md` containing `name` and `description` YAML frontmatter. It is designed for:



- OpenAI Codex: `~/.codex/skills`

- Claude Code: project `.claude/skills` or user `~/.claude/skills`

- Google Antigravity: project `.agents/skills` or user `~/.gemini/config/skills`

- Other Agent Skills-compatible hosts



The Markdown workflows are portable; tool availability and instruction adherence vary by

host and model. This repository does not claim mechanical enforcement.



## Install



PowerShell:



```powershell

.\install.ps1 -Target "$HOME\.codex\skills"

.\install.ps1 -Target ".claude\skills"

.\install.ps1 -Target ".agents\skills"

```



Bash:



```sh

./install.sh "$HOME/.codex/skills"

./install.sh .claude/skills

./install.sh .agents/skills

```



Installation copies only the 19 directories under `skills/`. Existing directories with

the same names are refused unless `-Force` or `--force` is supplied.



### The anchor block and rigor-goals install by default



A plain `./install.sh .claude/skills` (or `.\install.ps1 -Target ".claude\skills"`)

installs all three tiers: the 19 skills, the `rigor-goals` tool (default:

`/../tools/`), and the anchor block (default: `CLAUDE.md`, `GEMINI.md`, or

`AGENTS.md` in the current directory, inferred from the target path). Run the installer

from the project you are setting up.



`--anchor FILE` / `-Anchor` and `--goals DIR` / `-Goals` override the default locations.

The anchor install is idempotent: on upgrade the marker-fenced block is replaced in

place, hand edits outside the markers survive, and a diff is printed.



`--no-anchor` / `-NoAnchor` and `--no-goals` / `-NoGoals` are **owner-only opt-outs**.

They exist so the human who owns the machine can turn the discipline off; an agent

running the installer must never pass them on its own initiative.



### rigor-goals in 30 seconds



```sh

python3 tools/rigor_goals.py create --brief "ship feature X" \

  --goal "api::add the endpoint" --goal "docs::update the manual"

python3 tools/rigor_goals.py next

python3 tools/rigor_goals.py checkpoint --id G001 --status complete --evidence "test_api.py: 4 passed"

python3 tools/rigor_goals.py next

python3 tools/rigor_goals.py checkpoint --id G002 --status complete --evidence "manual updated" \

  --verify-cmd "pytest && python tools/validate_bundle.py" --verify-evidence "12 passed; BUNDLE_VALID"

python3 tools/rigor_goals.py status

```



The final story refuses to complete without `--verify-cmd` and `--verify-evidence` —

that refusal is the point. State lives in `./.rigor/` (add it to `.gitignore` or commit

it; your choice). A fresh session resumes with `status`.



Be precise about what the gate is: `rigor-goals` **records** the verification command

and its result — it does not run the command or check the result is true. It is a

workflow-completeness gate (no story closes without named evidence), not independent

proof enforcement. The honesty of the evidence is the agent's obligation and the

reviewer's to check. One active plan per working tree: concurrent tasks sharing a

checkout will fight over `./.rigor/` — use separate worktrees.



**Known limitation — the state is a file, not a fortress.** Any process that can delete

files in the workspace can destroy the plan, and nothing in-repo can detect a deletion

that also removes the ledger. Replacing a plan is loud (`create --force` prints what it

destroys, and every ledger event carries a `plan_id`), but if other agents or tools share

the same checkout, commit `./.rigor/` or back it up externally — the gate is only as

durable as the files it lives in.



## Main entrypoints



- `dev-rigor-stack-lite` — coordinate PLAN → BUILD → VERIFY → REVIEW → MERGE

- `dev-rigor-stack-lite-plan`

- `dev-rigor-stack-lite-build`

- `dev-rigor-stack-lite-proof-gate`

- `dev-rigor-stack-lite-audit-lite`

- `dev-rigor-stack-lite-audit-team`

- `dev-rigor-stack-lite-walkthrough`

- `dev-rigor-stack-lite-visitor-audit`

- `dev-rigor-stack-lite-gauntletgate`

- `dev-rigor-stack-lite-merge-gate`

- `dev-rigor-stack-lite-docs-gate`

- `dev-rigor-stack-lite-continuity`

- `dev-rigor-stack-lite-release`



Six short-name entrypoints are also included: `coder-tdd-qa-lite`, `proof-gate-lite`,

`quick-audit-lite`, `audit-team-lite`, `gauntletgate-lite`, and `visitor-audit-lite`.

(`quick-audit-lite` was `audit-lite` before 0.3.1 — renamed because it was the one name

that collided with, and could silently overwrite, a full dev-rigor-stack install sharing

the same skills directory.)



## Validate



```sh

python tools/validate_bundle.py

```



## Provenance



This is a hook-free adaptation of

[`scottconverse/codex-dev-rigor-stack`](https://github.com/scottconverse/codex-dev-rigor-stack),

originally released under the MIT License. See [NOTICE.md](NOTICE.md).



## License



MIT