docs: add NEWSTART.md and NEWSTART.llm.md capturing all governance, process, and blueprint for a clean repo restart
Some checks are pending
ci / test (push) Waiting to run

This commit is contained in:
2025-09-17 11:40:15 -05:00
parent 11a611e072
commit 6fa024ddfa
10 changed files with 200 additions and 5 deletions

18
NEWSTART.llm.md Normal file
View File

@@ -0,0 +1,18 @@
# NEW START — LLM
- Non-negotiables: zero-debt; TDD; plan-first; quiet chat; containers-first; CI parity; clean root; audits.
- Gates (human .md phrases): Questions→“Approved for Proposal”; Proposal→“Approved for Plan”; Plan→“Approved to Implement”.
- Chat: announce collab files only; ≤5 lines; no content/diff streaming.
- Quiet tooling: use scripts/toolwrap.sh; success → no chat; failure → append stdout/stderr to docs/devlog/tool.log.
- Audits: run scripts/audit.sh regularly and pre-release; store in docs/audits/.
- CI/containers: Gitea Actions + local Docker Compose; dependencies via images; explicit names; cleanup; mount host config if needed.
- .gitignore: include runs/ + OS files; keep current.
- Layout: collab/, docs/, prompts/global/, modes/, templates/project/_shared/, scripts/, .gitea/.
- Wrapper (MVP): subcommands new-mode (repo-only), new-project (outside), run (outside). Guardrails enforced.
- Prompt order: global system → mode system? → mode rules → project narrative.
- Config via yq: global < mode < project < ENV < CLI.
- Project scaffold: AGENTS.md; prompts/{project.md,style.md?}; prompts/_mode; codex.yaml; codex.sh; CI (gitea); docker/compose.yml + test/Dockerfile; scripts/{test.sh,test.docker.sh,audit.sh}; .gitignore; runs/.
- Tests: bats/internal; cover CLI/guardrails/scaffold/composition/precedence.
- Branching (recommended): trunk-based; protected main; short-lived branches; tags YYYY-MM-DD-HHMM; required checks; Conventional Commits.
- Start checklist: commit clean skeleton; add NEWSTART docs; start with Questions 00; use toolwrap + DevLogs; add audits early.

113
NEWSTART.md Normal file
View File

@@ -0,0 +1,113 @@
# New Start — CodexHelper Project Blueprint
Purpose
- Capture everything we learned so far and define a clean, production-ready blueprint for a fresh repository using CodexHelper. This document is human-focused; see NEWSTART.llm.md for the compact machine version.
Guiding Principles (Non-Negotiable)
- Safety first, speed second. Zero technical debt at all times; every commit is production-ready.
- TDD by default. Write failing tests first; make them pass; refactor.
- Plan first, implement second. Strict Questions → Proposal → Plan → Implement.
- Quiet, file-first collaboration. Chat is minimal; files and logs carry the detail.
- Containers-first, with CI parity between local and Gitea.
- Clean repository roots and predictable scaffolding.
Process & Governance
- Sequencing (Strict Gates)
- Questions → Proposal → Plan → Implement (one-way progression; no backsteps after approval).
- Approvals must use exact phrases in the human .md file:
- Questions approval: “Approved for Proposal”
- Proposal approval: “Approved for Plan”
- Plan approval: “Approved to Implement”
- Agent reads `.llm.md`; humans edit `.md`. Agent writes both siblings for each artifact.
- Chat Output Policy
- Default: “Updated <filepath>. Read/edit and let me know.”
- ≤5 lines; do not stream file contents or diffs in chat.
- Announce only collab files (questions/proposals/plan). Everything else goes to DevLog.
- Quiet Shell, Tool Logging
- Quiet is mandatory (no toggle). Use silent checks and avoid printing command output.
- Use `scripts/toolwrap.sh` to run host commands quietly.
- On success: no chat output; brief SUCCESS line is logged to `docs/devlog/tool.log`.
- On failure (not sandbox): append full stdout/stderr + timestamp + command to `docs/devlog/tool.log`.
- TDD & Quality
- Tests live under `tests/`; provide `scripts/test.sh` and `scripts/test.docker.sh`.
- Add tests for every feature; keep tests fast and focused.
- No unrelated refactors; fix root causes.
- Architecture & Planning
- Maintain `docs/architecture.md` and ADRs for significant decisions.
- Design module boundaries up front to avoid refactors; if assumptions change, update Questions/Proposal/Plan and docs before changing code.
- Clean Roots & .gitignore Housekeeping
- Keep root minimal; organize assets under `docs/`, `templates/`, `collab/`, `prompts/`, `modes/`, `scripts/`, `meta/`, `.gitea/`.
- Include `.gitignore` with `runs/` and common OS files; keep it current.
- CI/Containers (Gitea + Docker)
- CI uses Gitea Actions (`.gitea/workflows/`) and must mirror local Docker Compose.
- Do work in containers when appropriate; host is for git/tea and Docker orchestration only.
- Dependencies (e.g., bats, yq) are provided via Docker images; avoid host installs.
- Naming hygiene for containers/networks; explicit names; ensure cleanup.
- Where host auth/config is required (e.g., codex), mount config dirs into the container securely.
- Audits (Regular and Pre-Release)
- Run `scripts/audit.sh` regularly and before any release tag.
- Audit checks: governance compliance (gates, quiet policy), structure, .gitignore, CI parity, tests present.
- Record audits in `docs/audits/` and summarize in DevLogs.
Branching & Releases (Recommended)
- Trunk-based development:
- `main` is protected, always green, production-ready.
- Short-lived branches by type: `feat/*`, `fix/*`, `chore/*`, `docs/*`, `ci/*`, `refactor/*`.
- Optional `next` only if batching risky work.
- Protections: required checks (audit + Docker tests), PRs required, Conventional Commits, linear history or squash.
- Tags: `YYYY-MM-DD-HHMM` for release-ready commits on `main` only.
Repository Layout (Clean Start)
- collab/ — questions/, proposals/, plan/ (each step has `NN-<subject>.md` + `.llm.md`).
- docs/ — devlog/, audits/, architecture.md, feature docs, ADRs.
- prompts/ — `global/system.md` and `.llm.md` (canonical rules); modes/ live in `modes/`.
- modes/ — `<ModeName>/{mode.md, system.md?, defaults.yaml}`.
- templates/ — project scaffolding, including `_shared` with AGENTS.md, CI, docker, scripts.
- scripts/ — `test.sh`, `test.docker.sh`, `audit.sh`, `toolwrap.sh`.
- .gitea/ — workflows for CI.
CodexHelper (Wrapper) — MVP Blueprint
- Goals
- Provide modes (global/mode/project prompts), project scaffolding, prompt composition, config precedence, and safe defaults around `codex`/`codex-cli`.
- CLI (subcommands)
- `new-mode` (repo-only): scaffold `modes/<Name>/{mode.md, defaults.yaml, system.md?}`.
- `new-project` (outside repo): scaffold project structure (see below).
- `run` (outside repo): compose prompts and invoke codex, writing to `runs/<ts>/`.
- Binary Detection
- `CODEX_BIN` env > `which codex` > `which codex-cli`; fail with a helpful message otherwise.
- Prompt Composition
- Order: Global system → Mode system overlay (optional) → Mode rules → Project narrative.
- Config Precedence (YAML + yq)
- global defaults < mode defaults < project config < ENV < CLI.
- Safety
- Require `--force` to overwrite; never run `git push` for user projects.
- Project Scaffolding (Generated Project Layout)
- `AGENTS.md` (from template)
- `prompts/project.md` (+ optional `prompts/style.md`)
- `prompts/_mode/` (read-only copies of global/mode prompts)
- `codex.yaml` (project settings)
- `codex.sh` (entrypoint to compose and call codex)
- `.gitea/workflows/ci.yml` (CI) + `docker/compose.yml` + `docker/test/Dockerfile`
- `scripts/test.sh` + `scripts/test.docker.sh` + `scripts/audit.sh`
- `.gitignore` (includes `runs/`)
- `runs/` (created on first run; ignored by VCS)
Acceptance (Phase 1 Crawl”)
- `new-mode` creates mode skeleton; refuses overwrites unless `--force`.
- `new-project` scaffolds all files above without overwrites; includes CI and Docker artifacts.
- `run` composes prompts in the right order and invokes the detected codex binary; artifacts under `runs/<timestamp>/`.
- Config precedence enforced via yq.
- Guardrails: inside the helper repo only `new-mode` is allowed; elsewhere `new-project`/`run` allowed.
- Tests (bats or internal): cover CLI, guardrails, scaffolding, composition, precedence.
- CI: Gitea workflow executes audit and Docker-based tests; local Docker run mirrors CI.
Start-From-Scratch Checklist (New Repo)
1) Commit clean skeleton: collab/, docs/, prompts/global/, templates/_shared, scripts/, .gitea/.
2) Add `NEWSTART.md` and `NEWSTART.llm.md` (this pair) and `docs/architecture.md`.
3) Begin with `collab/questions/00-bootstrap.{md,llm.md}`; do not proceed without Approved for Proposal”.
4) Keep chat minimal; use toolwrap and DevLogs. Add audits early.
5) Implement Phase 1 via TDD, following QuestionsProposalPlan gate for each subject.
Notes
- This blueprint is the canonical source for repository setup and workflow. Keep it in sync with prompts/global/system.md.

View File

@@ -245,3 +245,10 @@ This log is concise and structured for quick machine parsing and summarization.
- Updated system prompts and AGENTS templates to state quiet is required
- next:
- Continue with quiet operations and filesystem-first logging
## 2025-09-17T16:34Z
- context: Quiet tooling wrapper added; tool.log policy established
- actions:
- Added scripts/toolwrap.sh for quiet execution and failure logging to docs/devlog/tool.log
- Updated system prompts and AGENTS templates to use tool.log
- next:
- Use toolwrap for host-side orchestration commands to keep chat silent

5
docs/devlog/tool.log Normal file
View File

@@ -0,0 +1,5 @@
2025-09-17T16:40:06Z FAILURE: git add -A
--- STDOUT ---
--- STDERR ---
./scripts/toolwrap.sh: line 21: {git: command not found
--------------

View File

@@ -6,6 +6,7 @@
- Chat ≤5 lines; default “Updated <filepath>…”; no diffs; announce only collab file changes; log details in `docs/devlog/`.
- Quiet shell: avoid `cat`/`sed` outputs; use silent checks; log details to DevLog files.
- Quiet is mandatory (no toggle) for this tool.
- Tool logging: use quiet wrapper; log to `docs/devlog/tool.log` on failures; no chat output.
- Keep changes minimal and focused; adopt TDD (tests first); require unit/integration tests for all features; consistent style.
- Git: Conventional Commits; branch `main`; optional tags `YYYY-MM-DD-HHMM`.
- Tools: file-first; use `rg`; read ≤250 lines; respect sandbox/approvals; preface grouped commands.

View File

@@ -40,6 +40,7 @@ Note: This is a template copied into generated projects. Customize as needed for
- Only announce changes to collaboration files in `collab/`; log details in DevLog.
- Quiet shell: avoid streaming file contents/diffs; prefer silent checks and log details to DevLog.
- Quiet is mandatory (no toggle) for this tools workflows.
- Tool logging: use a quiet wrapper to log to `docs/devlog/tool.log` on failures; keep chat silent.
## Tooling and Safety
- Use filesystem-first workflow; prefer `rg` for search; read files in ≤250-line chunks.

View File

@@ -53,3 +53,7 @@
- Never stream file contents/diffs in chat. Avoid `cat`/`sed` outputs.
- Use silent checks (exit codes, `grep -q`); log details in DevLog files.
- Quiet is mandatory (no toggle); default to minimal, non-chat output.
-- Tool Logging --
- On success: no chat output; optional one-line summary to `docs/devlog/tool.log`.
- On failure (not sandbox): capture stdout/stderr and append to `docs/devlog/tool.log` with timestamp and command.

View File

@@ -37,6 +37,11 @@ You are a coding agent running in the Codex CLI (terminal-based). Be precise, sa
- If content must be inspected, avoid printing it to chat; summarize findings in DevLog and reference the file.
- Quiet is mandatory (no toggle). Tools and scripts should default to minimal, non-chat output.
## Tool Logging (tool.log)
- On command success: do not print to chat; optionally append a one-line summary to `docs/devlog/tool.log`.
- On command failure (not sandbox failures): write full stdout/stderr to `docs/devlog/tool.log` with timestamp and command.
- Use a quiet wrapper script to enforce this behavior.
## Dev Logs and Docs
- Maintain `docs/devlog/DEVLOG_LLM.md` and `docs/devlog/DEVLOG_HUMAN.md`. Add an entry for each meaningful change.
- Keep `README.md` up to date for quickstart and usage.

40
scripts/toolwrap.sh Executable file
View File

@@ -0,0 +1,40 @@
#!/usr/bin/env bash
set -euo pipefail
# Quiet command wrapper:
# - Runs the given command without emitting stdout/stderr to the console.
# - On success: prints nothing; logs an optional one-line summary to tool.log.
# - On failure: writes full stdout/stderr to tool.log and exits non-zero.
log_file="docs/devlog/tool.log"
mkdir -p "$(dirname "$log_file")"
ts() { date -u +"%Y-%m-%dT%H:%M:%SZ"; }
cmd=("$@")
# Capture stdout/stderr
out_file="$(mktemp)"
err_file="$(mktemp)"
rc=0
{"${cmd[@]}"} >"$out_file" 2>"$err_file" || rc=$?
if [ "$rc" -eq 0 ]; then
echo "$(ts) SUCCESS: ${cmd[*]}" >> "$log_file"
rm -f "$out_file" "$err_file"
exit 0
fi
{
echo "$(ts) FAILURE: ${cmd[*]}"
echo "--- STDOUT ---"
cat "$out_file"
echo "--- STDERR ---"
cat "$err_file"
echo "--------------"
} >> "$log_file"
rm -f "$out_file" "$err_file"
exit "$rc"

View File

@@ -21,6 +21,7 @@ This file is copied by scaffolding into new projects. Edit to suit the project w
- Keep chat ≤5 lines; no diffs; announce only collab file changes; log details in DevLog.
- Quiet shell: avoid printing file contents/diffs; prefer silent checks and DevLog summaries.
- Quiet is mandatory (no toggle) in this workflow.
- Tool logging: prefer a quiet wrapper that logs to `docs/devlog/tool.log` on failures; keep chat silent.
## Coding, Tests, and Git
- Minimal, focused changes; adopt TDD (write tests first) and require unit/integration tests for all features; consistent style.