name	discover-workflows
description	Recommend GitHub agent workflows (gh-aw) tailored to this repo. Use when the user asks which automations to add to their repo, what agentic workflows would help, or types /discover-workflows.

discover-workflows

Recommend 1-3 workflows from the upstream githubnext/agentics catalog that fit the user's current repo. No local catalog — every call fetches the latest state from upstream.

Flow

Detect repo shape: language, framework, test presence, CI presence, size (files, LOC), activity (recent commits). Use git and filesystem inspection only — no external calls yet.
Fetch the upstream workflow list:
- Prefer gh aw list (or equivalent listing verb) if it enumerates upstream workflows.
- Otherwise, gh api repos/githubnext/agentics/contents/workflows to list available .md workflow files, and fetch githubnext/agentics's README or workflows/README.md for one-line descriptions.
Short-list ~5 candidates by matching names + descriptions to the detected repo shape. Do NOT read every workflow's body at this stage — that's too many tokens.
For each short-listed candidate, fetch the workflow's frontmatter only (gh api repos/githubnext/agentics/contents/workflows/<name>.md → base64-decode → parse YAML frontmatter). Confirm triggers, required secrets, and fit signals.
Pick 1-3 whose frontmatter genuinely matches the repo shape. The recommendation IS the product — don't fall back to "here's the full list."
For each recommendation, show: name, one-line purpose, why it fits THIS repo (one sentence, specific), estimated setup friction.
Ask which (if any) the user wants to install. Hand off to /install-workflow <name>.

Hard rules

Never recommend a workflow whose required secrets the repo clearly can't produce (e.g. don't recommend a Slack-notification workflow in a repo with no Slack references anywhere).
Never recommend more than 3 at once. Two is usually right.
Never draft a custom workflow. If the user asks for one, point them at zircote/github-agentic-workflows and stop.
Never fall back to a stale or inline list if the upstream fetch fails. Surface the error plainly and stop — a broken network beats a stale recommendation.

Out of scope for v0.1

Auditing already-installed workflows
Recommending across multiple repos
Persisting recommendations or caches between sessions

discover-workflows

Recommend 1-3 workflows from the upstream githubnext/agentics catalog that fit the user's current repo. No local catalog — every call fetches the latest state from upstream.

Flow

Detect repo shape: language, framework, test presence, CI presence, size (files, LOC), activity (recent commits). Use git and filesystem inspection only — no external calls yet.

Fetch the upstream workflow list:

Prefer gh aw list (or equivalent listing verb) if it enumerates upstream workflows.
Otherwise, gh api repos/githubnext/agentics/contents/workflows to list available .md workflow files, and fetch githubnext/agentics's README or workflows/README.md for one-line descriptions.

Short-list ~5 candidates by matching names + descriptions to the detected repo shape. Do NOT read every workflow's body at this stage — that's too many tokens.

For each short-listed candidate, fetch the workflow's frontmatter only (gh api repos/githubnext/agentics/contents/workflows/<name>.md → base64-decode → parse YAML frontmatter). Confirm triggers, required secrets, and fit signals.

Pick 1-3 whose frontmatter genuinely matches the repo shape. The recommendation IS the product — don't fall back to "here's the full list."

For each recommendation, show: name, one-line purpose, why it fits THIS repo (one sentence, specific), estimated setup friction.

Ask which (if any) the user wants to install. Hand off to /install-workflow <name>.

Hard rules

Never recommend a workflow whose required secrets the repo clearly can't produce (e.g. don't recommend a Slack-notification workflow in a repo with no Slack references anywhere).

Never recommend more than 3 at once. Two is usually right.

Never draft a custom workflow. If the user asks for one, point them at zircote/github-agentic-workflows and stop.

Never fall back to a stale or inline list if the upstream fetch fails. Surface the error plainly and stop — a broken network beats a stale recommendation.

Out of scope for v0.1

Auditing already-installed workflows

Recommending across multiple repos

Persisting recommendations or caches between sessions

discover-workflows