autonomous overnight code quality bot. 63 tasks, plan-implement-review, 21 PR tasks + 42 issue tasks.

---

`nightshift` runs on a cron schedule during GLM's quota burn window, selects from your public GitHub repos, and autonomously executes code quality tasks. it picks a repo, clones it, delegates to a coding agent that plans → implements → reviews, then opens a PR or issue depending on the task type.

how it works

1. quota gate — checks GLM token quota. only runs in the burn window (5–50 min before quota reset)
2. discover — lists your public, non-fork, non-archived repos with recent pushes
3. select — picks tasks based on cooldowns, cost tiers, and recent run history
4. execute — clones the repo, runs the task via a coding agent (GLM 5.1)
5. output — PR tasks: commit + fork + pull request. Issue tasks: analysis + github issue
6. cleanup — deletes the clone, logs the run to state

architecture

PR tasks use a **plan → implement → review** loop:

plan ──→ implement ──→ git diff check ──→ review ──→ PR
              ↑                              │
              └──── retry (max 3×) ←──────── failed

• plan — agent analyzes the repo, outputs JSON with steps/files
• implement — agent makes targeted code changes (no formatters, no new deps)
• review — agent checks correctness, security, unintended changes. if review fails, loops back to implement with feedback (up to 3 iterations)

issue tasks are **analysis only** — no branches, no commits. the agent reads the codebase and writes a structured markdown report with severity levels (P0–P3), file paths, and actionable recommendations.

quickstart

requires a GLM coding plan