hermes-nightshift-glm

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 key

mkdir -p ~/nightshift-workspace ~/.nightshift
curl -sL https://raw.githubusercontent.com/Microck/hermes-nightshift-glm/main/nightshift.py > ~/nightshift-workspace/nightshift.py
curl -sL https://raw.githubusercontent.com/Microck/hermes-nightshift-glm/main/nightshift_tasks.json > ~/nightshift-workspace/nightshift_tasks.json
curl -sL https://raw.githubusercontent.com/Microck/hermes-nightshift-glm/main/glm_quota.py > ~/nightshift-workspace/glm_quota.py

set your key in ~/.hermes/.env:

GLM_API_KEY=your-key

check quota, preview, run:

python3 ~/nightshift-workspace/glm_quota.py --check
python3 ~/nightshift-workspace/nightshift.py --dry-run
python3 ~/nightshift-workspace/nightshift.py

schedule with hermes agent cron:

hermes cron create "nightshift" --skill nightshift --schedule "*/15 * * * *" --model glm-5.1 --deliver discord

tasks

63 tasks across 7 categories. 21 create pull requests with code changes, 42 create github issues with analysis findings.

PR tasks (21) — create pull requests

task	description	cost
lint-fix	fix linting errors and style issues	low
bug-finder	identify and fix potential bugs	high
auto-dry	refactor duplicate code	high
skill-groom	clean up and organize skill configs	medium
api-contract-verify	verify API contracts and schemas	medium
backward-compat	check backward compatibility	medium
build-optimize	optimize build configuration	high
docs-backfill	fill in missing documentation	medium
commit-normalize	normalize commit messages	low
changelog-synth	synthesize changelogs from git history	low
release-notes	generate release notes	low
adr-draft	draft architecture decision records	medium
ci-fixes	fix CI pipeline issues	low
dependency-updates	update dependencies	medium
readme-improvements	improve README quality	low
dead-code	remove dead/unused code	medium
code-quality	improve code quality patterns	medium
code-review	expert review + fixes (SOLID, security, perf)	medium
visibility-instrument	add observability instrumentation	high
perf-audit	audit and fix performance issues	high
autoresearch	automated research experiment loop	high

issue tasks (42) — create github issues

analysis (17 tasks)

doc-drift, semantic-diff, dependency-risk, test-gap, test-flakiness, logging-audit, metrics-coverage, perf-regression, cost-attribution, security-footgun, pii-scanner, privacy-policy, schema-evolution, event-taxonomy, roadmap-entropy, bus-factor, knowledge-silo

options (13 tasks)

task-groomer, guide-improver, tech-debt-classify, why-annotator, edge-case-enum, error-msg-improve, slo-suggester, ux-copy-sharpener, a11y-lint, service-advisor, ownership-boundary, oncall-estimator, idea-generator

safe (5 tasks) — expensive, use sparingly

migration-rehearsal, contract-fuzzer, golden-path, perf-profile, allocation-profile

map (7 tasks)

visibility-instrument (PR), repo-topology, permissions-mapper, data-lifecycle, feature-flag-monitor, ci-signal-noise, historical-context

emergency (3 tasks) — 4-week cooldown

runbook-gen, rollback-plan, postmortem-gen

quota & burn window

GLM operates on a 5-hour rolling quota. nightshift only runs in a burn window of 5–50 minutes before quota reset to maximize value from remaining budget:

• < 50% used → up to 3 tasks
• < 75% used → up to 2 tasks
• < 95% used → 1 task
• ≥ 95% used → scrape remaining with cheapest tasks only

the cron runs every 15 min but the quota gate means it only actually executes during the burn window. the reset time drifts, so the dynamic check adapts automatically.

runtime hygiene:

• stale clone directories are removed on startup before new work begins
• state entries older than the longest task cooldown are pruned automatically
• quota responses are cached for 5 minutes to avoid hitting the GLM monitor API on every cron poll

config

optional ~/.nightshift/config.yaml:

exclude_repos:
  - "*-backup"
tasks_per_run: 3
max_prs_per_repo: 2
max_cost_tier: "very_high"
public_only: true
max_inactive_days: 30
enabled_categories:
  - pr
  - analysis
  - options

nightshift validates config keys and types on load and fails fast with a clear error if config.yaml is malformed.

option	default	description
exclude_repos	[]	repo name patterns to skip
tasks_per_run	3	max tasks per cron invocation
max_prs_per_repo	2	max open nightshift PRs per repo
max_cost_tier	very_high	highest cost tier to run
public_only	true	only run on public repos
max_inactive_days	30	skip repos with no recent pushes
enabled_categories	all	restrict to specific categories

bot account setup

nightshift operates from a separate github account to keep your repos' "last updated" timestamps clean:

1. create a github account for the bot (e.g. nightshift-micr)
2. generate a personal access token with repo and read:org scopes
3. save the token: echo "ghp_..." > ~/.nightshift/.gh-token-nightshift
4. nightshift forks repos under the bot account, pushes branches there, and opens PRs from the fork

all PRs and issues will appear as authored by the bot account.

files

file	description
nightshift.py	repo discovery, task selection, output generation
nightshift_tasks.json	63 task definitions with output_mode, cost tier, cooldown
glm_quota.py	quota check via zhipu API with burn window calculation
SKILL.md	hermes agent skill documentation
references/*.md	code review checklists (SOLID, security, code quality)

for llm agents

if a user asks you to install hermes-nightshift-glm, follow these steps:

1. create directories:

   mkdir -p ~/nightshift-workspace ~/.nightshift ~/.hermes/skills/nightshift/references

2. download all files using curl (see quickstart section above)

3. ask user for GLM_API_KEY or check if it exists in ~/.hermes/.env. if missing, user needs to get one from open.bigmodel.cn

4. verify installation:

   python3 ~/nightshift-workspace/glm_quota.py --check

   should show quota status or error if key is missing

5. test dry run:

   python3 ~/nightshift-workspace/nightshift.py --dry-run

   should show selected repos/tasks

done. user can now run nightshift manually or schedule it via hermes cron.

license

mit license