gcontext/docs/workflows.md

Workflows

A context-based workflow is a module with a fixed shape. It is a series of steps the agent executes with judgment, where every run leaves a persistent trace on disk. The workflow remembers what happened last run, accumulates knowledge, and gets better over time. A skill or prompt runs and forgets; a workflow holds state.

This document is the template spec: the contract a workflow folder must follow to be distributable. The CLI (gcontext add), the site directory, and the authoring tooling all build against it. It is one standard for all workflows; there are no per-domain variants.

Folder anatomy

<workflow-id>/
  index.md            # required: frontmatter manifest + objective, parameters, context
  steps/              # required: the procedure
    index.md          #   map: one line per step
    0-preflight.md    #   numbered step files, executed in order
    1-init.md
    ...
  commands/           # required: entry points
    setup.md          #   required: the install interview (see setup contract)
    start-workflow.md #   optional: the run driver, added when it earns its place
  functions/          # optional: per-step helper library
    2-transform/
      index.md        #   which helper applies to which case (the switch)
      from-x.md
      from-y.md
  runs/               # one folder per run
    example/          #   ships with the template: a fabricated run (see example run)
    2026-08-07/       #   the user's own runs, generated locally, never shared

The only code-enforced requirement is index.md with valid frontmatter. Everything else is convention: step files state what they need from previous steps in prose; nothing checks it mechanically. The process is AI-driven, so requirements live in the text the agent reads, not in validation code.

Manifest: frontmatter in index.md

The manifest is YAML frontmatter at the top of the workflow's index.md. There is no separate manifest file, and no version field: installs are snapshots, and a version mechanism comes only when updates exist.

---
id: coolify-ops              # unique, url-safe slug; the argument to `gcontext add`
                             # and the site path /workflows/<id>
name: Coolify Ops            # human display name, shown in the directory
description: >               # one or two sentences; the directory card text
  Mirror of a Coolify instance with operational playbooks that
  accumulate as incidents are resolved.
parameters:                  # what a run starts with; bound at setup or per run
  - name: instance-url
    description: Base URL of the instance to operate
    required: true
  - name: scope
    description: Limit operations to one project
    required: false
connections:                 # service capabilities the workflow needs
  - kind: http-api           # generic kind, not a product name
    description: The hosting panel API (Coolify, Dokploy, or similar)
tags: [ops, infrastructure]  # directory filtering
---

Field notes:

• id is the identity everywhere: the install argument, the folder name, the site slug. Lowercase letters, digits, hyphens.
• name is the display name for the directory and the workflow page; the id stays the machine identity.
• parameters are slots, not values. The setup interview or the run start binds them. Never ship bound values.
• connections entries are structured (kind plus description) so the site can render them as requirement badges. They name capability kinds, not products. The body of index.md may mention concrete services as examples; the steps must not depend on one (see docs/modules.md on connection-agnostic modules). The agent maps kinds to its own connections/ at run time.
• tags is a flat list for the directory. Keep it short.

After the frontmatter, the body of index.md carries: the objective in the first paragraph, what each parameter means in practice, the workflow's run naming scheme (see runs/), and the general context the agent needs across all steps. Context specific to one step belongs in that step's file.

steps/

steps/index.md is the map: one line per step, in order. Each step is one numbered file (0-preflight.md, 1-init.md, ...). Number from 0 when there is a gate or check before real work starts.

Each step file states:

• Purpose: what this step achieves and why it exists.
• Input: what it needs, and from where (parameters, a previous step's results file, a connection).
• Output: what it writes into the run folder, with the schema when the output is tabular (column list for a CSV, field list for JSON).
• How to execute: the procedure, in enough detail that an agent without prior context can do it. Include known blockers and how to classify or route them.
• Done when: the condition that closes the step.

Steps that pause for the user (approval, manual action, batching) say so explicitly: what to present, what to wait for.

runs/

Every execution of the workflow is one folder in runs/.

The run folder name is workflow-defined. Each workflow states its own run naming scheme in its index.md: whatever identifies one run in that domain. A gym migration names runs by gym id, an invoicing workflow by plant and period, a research pipeline by batch name. The ISO date (2026-08-07; second run the same day 2026-08-07-b) is only the default for workflows with no better key. The run name should carry meaning; the date is the fallback.

Inside a run folder:

runs/<run-key>/
  index.md           # map and status: scope of the run, per-step status table
  0-parameters.csv   # the parameters this run started with (.csv, .json, or .md)
  1-init/
    results.csv      # the step's output, schema per the step file
    script.py        # optional: a generated script saved to avoid regenerating it
  2-transform/
    results.json
  done/
    info.md          # written when the run closes: what was achieved, what was learned

Conventions:

• 0-parameters.* records what the run started with, always, even when trivial. It is what makes a run reproducible and auditable.
• One folder per executed step, named like the step file without the extension. Its main artifact is results.*. When the agent generated code worth keeping, it saves the script next to the results.
• The run's index.md is the resume point: a session picking up a half-finished run reads it and continues from the first step that is not done.
• done/ closes the run: info.md summarizes what was achieved and anything learned that should change the steps, plus any final deliverable files. A run without done/ is open.
• Learnings that outlive the run (a new blocker type, a better procedure) get folded back into the step files or functions/. That is how the workflow improves with use.

functions/ (optional)

Some steps do the same transformation from ever-varying inputs. functions/ is a mini library the agent picks from, organized per step: functions/<step>/index.md describes the cases (the switch), one file per case describes the procedure or code for it. The step file points to its functions folder. Add functions/ only when a step has proven to need it; most workflows ship without it.

What ships vs what is generated

A workflow is distributed as a template. The template is the procedure plus one fabricated demonstration; the state is born empty on the user's machine.

Ships in the template:

• index.md with the frontmatter manifest
• steps/
• commands/setup.md (and any other generic commands)
• functions/ when the workflow has them
• runs/example/: the example run

Never ships, generated locally at setup and use:

• the user's own run folders in runs/
• every personalized file: configs, credentials references, scripts bound to the user's systems, playbooks learned from the user's own work

Installs are snapshots. The user's copy is theirs: personalized, growing, never overwritten by an update. gcontext add on an existing module warns and stops instead of overwriting.

The example run

Every template ships one fabricated run at runs/example/, in the exact runs/ shape: index.md, 0-parameters.*, one folder per step with plausible results, done/info.md. All names and data in it are fake, made up by the author; it must contain nothing personal.

The example run has two jobs:

• On the site, it is the centerpiece of the workflow's page: the visitor browses it file by file and sees exactly what each step produces before installing anything.
• In the installed folder, it is the reference: the agent reads it to see what a correct run looks like before executing its first real one.

The folder name example (instead of a run key) is what marks it fabricated. The setup command leaves it in place.

The setup command contract

commands/setup.md is the bridge from template to personal instance. It is an interview: the agent asks, the user answers in plain words, the agent builds and confirms. The contract:

1. Read first: the command starts by instructing the agent to read the workflow's index.md and steps/index.md so the interview is informed.
2. Bind every parameter slot: ask for each manifest parameter that is bound at setup time (per-run parameters are only explained, not bound).
3. Map connections: for each connections entry, find a matching service in the agent's environment or help the user create one. In gcontext that means connections/; standalone it means whatever access the user's agent has.
4. Generate the personal state: create the files this workflow needs locally (config, scripts against the user's systems, an empty runs/ besides the example). What gets generated is listed in the setup command itself.
5. Smoke test: verify the critical path (a read against the user's system, a dry run of the first step) before declaring setup done.
6. Never rewrite the procedure: setup personalizes state; it does not edit steps/.

The same file must work on both install paths:

• In gcontext: commands/setup.md carries the standard command frontmatter (description, optional parameters), so the server exposes it as an MCP prompt and the user runs it as a slash command.
• Standalone: the user downloads the plain folder, opens any agent in it, and says "run the setup in commands/setup.md". The agent reads the file and executes the same interview. Therefore the body must be self-contained prose that assumes only file access, not gcontext tools.

Sharing a workflow

Authors turn a lived workflow into a template with the share-workflow instructions: docs/share-workflow.md. It strips the personal specifics into parameter slots and connection requirements, generates the setup command, fabricates the example run, and verifies the result against this spec.

Relation to modules

A workflow is a module (see docs/modules.md): installed into modules/, connection-agnostic, growing with use. The workflow spec adds the fixed shape on top: manifest frontmatter, steps/, runs/, the setup contract, the example run. Everything modules.md says about growth and portability applies unchanged.