Configuration¶

Every peggbot run is configured through action inputs. Nothing is hard-coded: each workflow can pin different models, prompts, permissions, branches, and timeouts.

Inputs¶

Mode & authentication¶

Input	Required	Default	Description
`mode`	yes	—	`solve-issue`, `review-pr`, `respond` or `release-notes`
`token`	yes	—	`${{ secrets.GITHUB_TOKEN }}` or a fine-grained PAT of the bot account

Which token?

GITHUB_TOKEN is scoped to the current workflow permissions and works for all modes. Use a PAT when you want the bot's actions attributed to the bot account (e.g. reviews) or when the repository needs a persistent identity.

LLM¶

Input	Required	Default	Description
`llm-provider`	yes	—	Provider name as registered in pegg: `anthropic`, `openai`, `google`, …
`llm-model`	yes	—	Model id, e.g. `claude-sonnet-4-5`, `gpt-4o`
`llm-api-key`	yes	—	Provider API key — always pass as a secret
`llm-base-url`	no	—	Provider/base URL override (proxy, self-hosted endpoint)

Prompting¶

Input	Default	Description
`prompt-file`	—	Markdown file inside the repo (e.g. `.github/prompts/solve-issue.md`) whose content is appended to the mode's system prompt
`skills-dir`	—	Directory of `SKILL.md` skill folders loaded into the agent via the SDK skill engine

See Prompts.

Trigger context¶

These override the values derived from the workflow event. In the example workflows they are always bound to event fields, so they rarely need editing.

Input	Used by	Description
`issue-number`	`solve-issue`, `respond`	Issue or PR conversation number
`pull-request-number`	`review-pr`, `respond`	Pull request number
`release-id`	`release-notes`	Numeric release id
`comment-id`	`respond`	Trigger comment id (informational)

Behavior¶

Input	Default	Description
`required-permission`	`admin`	Minimum actor permission for `respond`: `admin`, `write` or `read`. Owners and org admins count as `admin`
`unauthorized-action`	`skip`	What happens when the actor lacks permission: `skip` (silent) or `comment` (post a polite decline)
`bot-login`	`peggbot`	The bot's login. Its own comments are never answered (loop guard)
`branch-prefix`	`peggbot`	`solve-issue` branches become `<prefix>/issue-<n>`
`base-branch`	`main`	Base branch for PRs opened by `solve-issue`
`max-iterations`	`80`	Cap on agent loop iterations
`timeout-minutes`	`30`	Cap on agent runtime before the run is aborted
`max-attachments`	`5`	Max image attachments downloaded from a thread
`workspace`	repo root	Agent workspace root relative to the checkout
`dry-run`	`false`	When `true`, nothing is written to GitHub

Enterprise¶

Input	Default	Description
`github-server-url`	`https://github.com`	Used to build the git push auth URL (set to your GHES host)
`ghes-api-base-url`	—	GitHub Enterprise API root, e.g. `https://ghe.example.com/api/v3/`

Environment¶

The action reads the standard GitHub Actions environment automatically:

Variable	Used for
`GITHUB_REPOSITORY`	Owner/repo of the current repository
`GITHUB_ACTOR`	Who triggered the run (permission gate, loop guard)
`GITHUB_EVENT_PATH`	The webhook payload JSON
`GITHUB_WORKSPACE`	Default workspace root
`GITHUB_ACTION_PATH`	Action source directory (the compiled binary is built here)

Permission model¶

The respond mode is the only mode with an interactive trigger, so it is the only one gated by default:

If the comment author is bot-login, the run stops immediately.
The actor's permission is resolved via GET /repos/{owner}/{repo}/collaborators/{actor}/permission.
admin < write < read < none; the run proceeds only when the actor's level satisfies required-permission.
On denial: skip logs and does nothing; comment posts a short decline.

solve-issue (label trigger) and review-pr (review request) are triggered by the workflow's if: conditions; restrict them further with additional conditions in the workflow, e.g. requiring a specific labeler.

Provider keys¶

The engine is opened with Options.APIKeys and Options.Providers set from llm-provider, llm-model, llm-api-key and llm-base-url, so a run works with any provider the pegg SDK supports — no local config file needed.