Run the LenserFighter Autonomous Contributor

LenserFighter is an autonomous AI Lenser that contributes to a GitHub repository on a schedule. Each run does one thing — review an open PR, triage and open one issue, or draft one small fix PR — and then stops. It never merges and never pushes to protected branches, so every contribution still goes through your normal CI and human review.

• On GitHub it acts as the lenserfighter[bot] GitHub App.
• In the LenserFight product it is the @lenserfighter AI Lenser.

The runtime is a GitHub Actions workflow (.github/workflows/lenserfighter.yml) that runs anthropics/claude-code-action — the same way Claude contributes to GitHub repositories. Its behavior is defined entirely by the persona file .lenserfight/lensers/lenserfighter/LENSER.MD.

The workflow ships inert: until you add the GitHub App secrets it skips every scheduled run with a notice, so you never get a daily failing cron before setup is complete.

---

What it does each run

Focus	Action	Writes?
review-prs	Posts one structured review comment on the most recently updated PR it has not already reviewed.	Comment only
triage-issues	Opens one well-formed, deduped issue using the matching issue template.	One issue
draft-fix	Picks one good-first-issue, implements the minimal fix on a lenserfighter/ branch, opens a PR into development.	One branch + PR
auto (cron default)	Rotates the three focuses by day-of-week so cost stays at one action per day.	Varies

It never merges, approves, force-pushes, runs migrations, or opens duplicates. See the persona file for the complete behavior contract.

---

Prerequisites

• A repository where you can add Actions secrets and (optionally) install a GitHub App.
• A Claude Code OAuth token (CLAUDE_CODE_OAUTH_TOKEN) — the repository already uses this for the Claude issue-fix workflow.

---

Step 1 — Create the GitHub App

A GitHub App gives the agent its own bot identity (lenserfighter[bot]) and short-lived, repo-scoped install tokens — more secure than a personal access token, and installable by anyone.

1. Go to Settings → Developer settings → GitHub Apps → New GitHub App (under the lenserfight account, your own account, or your org).
2. Name it (e.g. LenserFighter). The contribution actor becomes <app-name>[bot].
3. Webhook: uncheck Active — no webhook is needed.
4. Repository permissions:
   • Contents: Read and write
   • Pull requests: Read and write
   • Issues: Read and write
5. Create the App, then generate a private key (downloads a .pem) and note the App ID.

Step 2 — Install the App

From the App's page, Install App and select the target repository (connectlens-org/lenserfight or your fork). Installing it is what grants it write access — no separate collaborator step is needed.

Step 3 — Add the secrets

In the target repo: Settings → Secrets and variables → Actions → New repository secret.

Secret	Value
LENSERFIGHTER_APP_ID	The App ID from Step 1
LENSERFIGHTER_APP_PRIVATE_KEY	The full contents of the downloaded .pem
CLAUDE_CODE_OAUTH_TOKEN	Your Claude Code OAuth token (reused if already present)

The moment LENSERFIGHTER_APP_ID exists, the next scheduled run activates.

Step 4 — Run it

• Daily: the workflow runs automatically at 07:00 UTC with focus: auto.
• On demand: go to Actions → LenserFighter Autonomous Contributor → Run workflow, pick a focus, and optionally pass a free-text task to override the focus playbook for that run.

Watch a few manual runs (start with review-prs) before relying on the cron.

---

Tuning

• Change the focus mix: edit the day-of-week rotation in the Resolve focus for this run step of the workflow.
• Change cadence: edit the schedule.cron expression. Keep it coarse (daily) — one unit of work per run keeps token and CI cost bounded.
• Change the model / turn budget: edit claude_args (--model, --max-turns).
• Change behavior: edit the persona file .lenserfight/lensers/lenserfighter/LENSER.MD. The workflow reads it at runtime, so behavior changes do not require touching the workflow.

Cost

Each run is one Claude Code session bounded by --max-turns and a single task, scheduled at most once per day. There are no retries or fan-out, so spend is predictable. Raising the cadence or turn budget raises cost proportionally.

Safety and pausing

• Never merges. Every change lands as a PR for human review behind your existing CI gates.
• Pause instantly by disabling the workflow (Actions → ⋯ → Disable workflow) or uninstalling the GitHub App — either one is an immediate kill switch.
• All write actions are attributed to lenserfighter[bot], so its contributions are easy to audit and filter.

Related

• What is an Agent
• Agent Boundaries
• Manage Agent Settings