Publicist Pipeline

This repository is a Memo-driven pipeline workspace for building AI-assisted workflows without giving the agent direct operational power.

The current primary example is publicist_draft_agent, a draft-first publicist system that can:

• read campaign source text from a selected workspace
• generate audience suggestions, target candidates, and reviewed draft messages
• preserve human review decisions and manual edits across reruns
• run controlled, approved research steps
• track manual outreach status without sending anything automatically

Intent

The important design boundary in this repo is the combination of the pipeline runner and Memo.

pipeline_runner.coffee is the execution layer. It loads a top-level recipe from config/, expands it into experiment.yaml, schedules steps, and materializes declared artifacts.

Memo is the control boundary. Steps do not talk directly to the OS, browser, or external systems as an agent shortcut. Instead they:

• read declared inputs through Memo-backed artifacts
• write declared outputs through Memo-backed artifacts
• rely on meta devices for durable file materialization

That keeps state explicit, replayable, and reviewable.

In practice:

• recipe code lives under shared repo EXEC
• run state and artifacts live under active workspace CWD
• steps are expected to use declared needs and makes
• human review can edit selected artifacts in place, and later steps preserve those edits when ids and source hashes still match

Workspace Model

This repo distinguishes two locations:

• EXEC: the shared code root for recipes, agents, runner, UI server, and docs
• CWD: the active campaign workspace, usually under pipes/<campaign>

For the publicist workflow, source input and outputs are workspace-local:

• source/publicist_source.txt
• source/publicist_campaign.yaml
• out/*.yaml
• out/review_packet.md
• runtime/publicist.sqlite

This means one shared system can run many separate campaign workspaces without duplicating code.

Setting Up A New Pipe

To create a new campaign workspace:

1. Create a new directory under pipes/.
2. Add source/publicist_source.txt.
3. Optionally add source/publicist_campaign.yaml.
4. Run the top-level recipe publicist_draft_agent from that workspace.

Minimal example:

pipes/mycampaign/
  source/
    publicist_source.txt
    publicist_campaign.yaml

Suggested starter source file:

Describe this campaign here.

Suggested starter campaign config:

priority_audiences:
  - technical_press
  - industry_partners
  - research_labs
  - pilot_customers

Then run from the workspace:

cd pipes/mycampaign
printf 'pipeline: publicist_draft_agent\n' > override.yaml
coffee ../../pipeline_runner.coffee

Expected outputs will appear in:

out/source_material.yaml
out/audience_profiles.yaml
out/contact_ledger.yaml
out/message_drafts.yaml
out/review_decisions.yaml
out/outreach_log.yaml
out/review_packet.md

If you use the local UI:

• the UI server code still runs from shared EXEC
• the active pipe determines CWD
• the UI should read and write source files and artifacts in the active workspace only

Recipes

Recipes are top-level YAML files in config/.

The publicist recipe is:

• config/publicist_draft_agent.yaml

That recipe is intentionally additive and review-oriented. It is designed to:

• avoid changing other pipelines
• avoid direct send behavior
• keep human approval visible in artifacts and UI

GPT Working Memory

The GPT/ subdirectory is assistant-owned working memory for Codex and future pipeline work.

It is not meant to be product output. It exists so Codex can recover important local knowledge quickly, especially when a pipeline has:

• step contracts
• proven invariants
• workspace rules
• known failure modes

For the publicist workflow, the most useful files are under:

• GPT/publicist_draft_agent/

Those notes describe:

• recipe intent
• EXEC vs CWD
• step-specific preservation rules
• common regressions and how to diagnose them

If a future Codex session seems to be missing pipeline context, GPT/ is the first place it should look.

Related Files

• pipeline_runner.coffee: runner and Memo integration
• ui_server.coffee: local UI server
• ui/: browser UI
• docs/publicist_workspaces.md: workspace-specific publicist notes
• docs/agents.md: agent-layer boundary and intent

Practical Rule

If you are extending this repo, prefer:

• new declared artifacts
• normal pipeline steps
• UI-only human controls
• preservation of human-edited files

Avoid:

• bypassing Memo
• hidden side effects
• shared root outputs for workspace-specific campaigns
• automatic send behavior