Memetics research — Codex starter bundle

This repo is a stateful workspace for a long-horizon memetics research sprint (outputs as files, resumable after interruptions).

What’s inside

• /spec/ — authoritative specs (your memetics bundle + FPF spec)
• /context/ — project context + extracted notes from the 2026-02-01 talk deck
• /work/ — where the agent writes deliverables + checkpoints
• /prompts/ — copy/paste prompts for Codex App / CLI + a resume prompt
• /templates/ — ADR + parity + acceptance templates (lightweight)

Quick start (recommended)

1. git init (or use a new branch/worktree in an existing repo).
2. Open Codex app and select this folder as the project.
3. Paste prompts/00_codex_thread_prompt.md into the thread.
4. Let it run. Review diffs; commit checkpoints.

Long-running mode (multi-hour)

Tip: prefer resumable slices + external memory (work/STATE.md) over one giant context window.

Use Codex app Automations to “re-trigger” progress periodically:

• Create an Automation scheduled hourly (or as frequently as the Automations UI allows).
• Use prompts/02_codex_automation_prompt.txt as the automation prompt.
• Each run should end by invoking $checkpoint (writes work/STATE.md + exactly one new work/checkpoints/YYYYMMDD-HHMM.md).

Repo-scoped Codex assets:

• Skill: .codex/skills/checkpoint/SKILL.md
• Rules allowlist: .codex/rules/safe-default.rules
• Cloud setup script: scripts/codex_setup.sh

Included usage limits (ChatGPT plans)

Codex included usage is measured in rolling 5-hour (18,000 s) windows and varies with task size/context. Per the Codex pricing page (accessed 2026-02-06):

Plan	Local messages / 5h	Cloud tasks / 5h	Code reviews / week
ChatGPT Plus	45–225	10–60	10–25
ChatGPT Pro	300–1500	50–400	100–250

To stretch included usage:

• keep prompts/context small (write state to files),
• prefer gpt-5.1-codex-mini for routine work (up to ~4x higher local-message limits), reserve gpt-5.2-codex for hard steps,
• disable MCP servers you don’t need (they add context and burn budget).

Rationale: automation runs create durable progress via repo artifacts, so you can stop/restart without losing state.

References:

• Codex pricing: https://developers.openai.com/codex/pricing/ (accessed 2026-02-06)
• Codex models: https://developers.openai.com/codex/models/ (accessed 2026-02-06)
• Codex Quickstart: https://developers.openai.com/codex/quickstart/ (accessed 2026-02-06)
• Codex web overview: https://developers.openai.com/codex/cloud/ (accessed 2026-02-06)
• Cloud environments: https://developers.openai.com/codex/cloud/environments/ (accessed 2026-02-06)
• Agent internet access: https://developers.openai.com/codex/cloud/internet-access/ (accessed 2026-02-06)
• Skills: https://developers.openai.com/codex/skills/ (accessed 2026-02-06)
• Team config: https://developers.openai.com/codex/team-config (accessed 2026-02-06)

Cloud environments (optional)

Codex Cloud tasks are usually 1–30 min, so “long-running” work is best done as many short cloud tasks chained together via repo state (work/STATE.md + work/checkpoints/).

One-time setup:

1. In Codex web, connect GitHub and select venikman/swam as the repo.
2. Create/select a Cloud environment for the repo.
3. Set the environment setup script to scripts/codex_setup.sh (Debian/Ubuntu apt-get; installs pandoc/ripgrep/git if missing).

Notes that matter for autonomy and safety:

• Setup scripts run with internet access; agent internet access is off by default and you should keep it off unless you explicitly want live sources (configurable per-environment).
• If you enable agent internet access, use a domain allowlist and restrict HTTP methods to GET/HEAD/OPTIONS.
• Codex caches container state for up to 12 hours. When caching, it clones the repo and checks out the default branch; when resuming a cached container, it checks out the branch specified for the task.

Cloud prompts (copy/paste into Codex web tasks):

• Bootstrap run: prompts/03_codex_cloud_bootstrap_prompt.txt
• Resume loop: prompts/04_codex_cloud_resume_prompt.txt

CLI trigger (one command after you push the default branch):

1. Set your env id (one-time): cp .codex/local.env.example .codex/local.env then edit it.
2. Load it: source .codex/local.env
3. Trigger: scripts/codex_cloud_trigger.sh resume

Autopilot loop (start once, run many slices):

• Start it: source .codex/local.env && scripts/codex_cloud_autopilot.sh --max-runs 0
• It will keep submitting Cloud tasks and pushing results to the remote default branch until:
  • a task fails (e.g., usage limits), or
  • you press Ctrl-C, or
  • you set a finite --max-runs.
• Safety: it refuses to commit if the cloud diff touches anything outside work/, and it enforces exactly one new work/checkpoints/* file per slice.

Quick sanity checks after Run 1:

• A PR exists with edits under work/.
• Exactly one new file exists under work/checkpoints/.
• work/STATE.md was updated with “what changed” + “next actions”.

Parallel cloud tasks without merge hell (pattern):

• Run 2–3 cloud tasks concurrently, each constrained to one file (or disjoint file sets) and explicitly not editing work/STATE.md.
• Each task still creates exactly one checkpoint file under work/checkpoints/ and opens a PR.
• Do NOT invoke $checkpoint in parallel tasks (it edits work/STATE.md); write the single checkpoint file manually, and let the integrator update work/STATE.md.
• Run an “integrator” cloud task to merge PR branches logically and update work/STATE.md with the integrated next actions.

If you prefer the CLI

• cd into this folder
• run codex and paste prompts/00_codex_thread_prompt.md

What you should edit first

• context/project_brief.md — confirm scope, outputs, and “stop conditions”.

Output contract (where to look)

• Checkpoints: /work/checkpoints/
• Core deliverables: /work/ (see file list in the main prompt)

Safety / hygiene

• Keep secrets out of this repo.
• Use git checkpoints before/after risky steps.
• If the agent proposes destructive actions, require explicit human confirmation.
• If you enable agent internet access, start with a domain allowlist and treat retrieved content as untrusted input (prompt injection is real).