NightShift

Auditable local-first AI coding pipelines.

NightShift is a deterministic pipeline runner for AI-assisted coding work. It reads markdown tasks, builds bounded context, asks configured agents for plans or patches, validates and applies those patches through explicit stages, runs checks, and leaves a human-reviewable artifact trail.

NightShift is not an autonomous software engineer. It is an orchestration layer that treats AI agents as unreliable workers inside bounded, testable, auditable workflows.

Current Status

NightShift now supports the full local patch workflow:

• nightshift init, validate, status, run, run --task, run --all, and web.
• Markdown task parsing with dependencies.
• Command, Ollama, and OpenAI-compatible agent backends.
• Per-agent model settings such as temperature.
• Repo lookup tools: scoped list_files, read_file, and grep.
• Planner lookup requests with files-inspected.md artifacts.
• repo_context stage for context-pack.md.
• Project context chart generation at .nightshift/project-context-chart.md.
• code_writer stage for direct unified diff output.
• file_writer stage for model-written complete file blocks with deterministic diff generation.
• patch_normalizer, patch_validator, and patch_apply stages.
• Patch dry-run and apply modes.
• Test/static failure repair loops through existing retry routing.
• Run logs, dashboard log tails, git status artifacts, diffs, stage summaries, and final reports.

The default posture remains local-first and review-first: agents propose; NightShift validates, applies, tests, and records.

What NightShift Is

NightShift is built for reviewable automation:

• local-first execution
• declarative pipeline stages
• markdown task files
• command-backed and model-backed agent wrappers
• explicit retry limits
• scoped repository lookup
• patch validation before mutation
• command allowlists
• durable markdown/text artifacts
• compact context handoff
• final reports for human review

The goal is to wake up to useful artifacts and a repository state you can inspect.

What NightShift Is Not

NightShift does not push branches, deploy software, run unbounded task swarms, or grant agents unlimited repository access. Human review remains the final authority.

Install

Repo setup scripts can install NightShift in editable mode, check for Ollama, and offer to add the Python scripts directory to PATH.

Windows PowerShell:

.\setup.ps1

macOS/Linux:

sh ./setup.sh

Development install:

pip install -e .

You can also run the CLI module directly from a checkout:

python -m nightshift.cli --help

NightShift uses the Python standard library for runtime behavior where practical. PyYAML is used automatically if installed, but starter configs work with the built-in YAML subset parser.

Getting Started

Start with the Quickstart. It uses deterministic fake agents so you can verify lookup, context generation, patch validation, patch apply, tests, and artifacts without installing a model.

After that works, continue with Tutorial 01: Building A Small Imageboard With Real Local Models. It swaps the fake agents for Ollama-backed agents such as qwen2.5-coder:14b and walks through a small Flask/SQLite project with ordinary web-app tasks.

Quickstart Commands

Validate the included end-to-end patch example:

python -m nightshift.cli validate --config examples/quickstart-lisp/nightshift.yaml

Run the first task against a copy of the example project. The pipeline uses patch_apply mode: apply, so running it directly against examples/quickstart-lisp/ will modify those files.

cp -r examples/quickstart-lisp /tmp/nightshift-quickstart
python -m nightshift.cli run --config /tmp/nightshift-quickstart/nightshift.yaml --task TASK-001

For a new project:

nightshift init
nightshift validate
nightshift status
nightshift run --task TASK-001

For the first real-model tutorial target:

nightshift init --template tutorial-imageboard --root nightshift-imageboard

Other built-in real-model templates:

nightshift init --template real-simple --root bookmarks-demo
nightshift init --template real-long-running --root incident-service
nightshift init --template tutorial-deaddrop --root nightshift-deaddrop

Create an isolated integration sandbox for a template:

python -m nightshift.cli integ-run --template tutorial-deaddrop

To create the sandbox and run the Python setup immediately:

python -m nightshift.cli integ-run --template tutorial-deaddrop --setup

Then run the Python project setup helper. It finds the generated venv, installs this NightShift checkout into it, installs the target project, installs pytest by default, and runs nightshift validate:

python -m nightshift.cli integ-setup --project integ_runs/<timestamp>/project

integ-setup cannot activate the venv for your current shell. If you want plain python and nightshift to resolve to the integration venv in PowerShell, run:

integ_runs\<timestamp>\.venv\Scripts\Activate.ps1

After setup, you can also run from the generated project with the explicit venv Python:

integ_runs\<timestamp>\.venv\Scripts\python.exe -m nightshift.cli run --task TASK-001

Bash:

integ_runs/<timestamp>/.venv/bin/python -m nightshift.cli run --task TASK-001

After a run, explain the latest pass or failure from artifacts:

nightshift what-happened

Open the read-only artifact dashboard:

pip install flask
nightshift web

Task File Example

Tasks live in markdown checklist format:

# Tasks

- [ ] TASK-001: Add parser support

Description:
Implement parsing for the target language.

Acceptance Criteria:
- Parses numbers
- Parses symbols
- Parses nested lists
- Includes unit tests

NightShift parses task id, title, completion state, description, acceptance criteria, dependency bullets, and raw task markdown.

Pipeline Example

pipeline:
  max_task_retries: 2
  continue_on_task_failure: false
  stages:
    - id: plan
      type: agent
      agent: planner
      output: plan.md

    - id: context
      type: repo_context
      output: context-pack.md

    - id: implement
      type: file_writer
      agent: implementer
      output: proposed.patch

    - id: normalize
      type: patch_normalizer
      output: normalized.patch

    - id: validate_patch
      type: patch_validator
      output: patch-validation.md
      max_files: 8
      max_lines: 800
      on_fail: implement

    - id: apply_patch
      type: patch_apply
      mode: apply
      output: patch-apply-output.txt
      on_fail: implement

    - id: test
      type: command
      commands:
        - python -m unittest discover -v
      output: test-output.txt
      on_fail: implement

    - id: review
      type: agent_review
      agent: reviewer
      on_fail: implement
      output: review.md

Use mode: dry_run for patch applicability checks without modifying files. Use mode: apply to write the validated patch to the target project.

Agent Backends

NightShift supports:

• backend: command
• backend: ollama
• backend: openai_compatible

Example Ollama agent:

agents:
  implementer:
    backend: ollama
    model: qwen2.5-coder:14b
    base_url: http://localhost:11434
    temperature: 0.2
    system_prompt: agents/implementer.md

The Ollama backend uses the local HTTP API instead of ollama run, which keeps exact patch output away from terminal rendering and line wrapping.

Example OpenAI-compatible agent:

agents:
  implementer:
    backend: openai_compatible
    model: local-model
    base_url: http://localhost:11434/v1
    api_key_env: OPENAI_API_KEY
    temperature: 0.2
    system_prompt: agents/implementer.md

NightShift passes prompt bundles to agents and persists stdout, stderr, exit code, duration, and prompt artifacts. code_writer agents return unified diffs directly. file_writer agents return complete file blocks, and NightShift generates the unified diff deterministically. On retries, patch artifacts are versioned by attempt, for example repair-1.patch, normalized-1.patch, and patch-validation-1.md.

Review agents should emit:

status: pass | fail | retry | escalate
reason: <short explanation>
next_stage: <optional stage id>
context_update: <compact useful note>

Safety Model

NightShift validates paths, commands, and patches before mutation.

Path safety:

• project roots are resolved with pathlib
• task and prompt files must stay inside the project root
• artifact paths cannot escape .nightshift/
• repo lookup tools are constrained by safety.scoped_paths

Command safety:

• command stages must match allowed_commands
• forbidden fragments are blocked before allowlist acceptance
• command output and exit codes are recorded
• command stages stop at the first failing or timed-out command

Patch safety:

• code changes are represented as unified diffs, either supplied directly or generated from complete file blocks
• patches are normalized and validated before apply
• path traversal and forbidden paths are rejected
• scoped paths, max files, and max changed lines are enforced
• patch_apply records apply output and git status artifacts

Artifact Layout

A run creates human-readable artifacts:

.nightshift/
  project-context.md
  project-context-chart.md
  nightshift.log
  runs/
    <run-id>/
      run.log
      run-summary.md
      config.snapshot.yaml
      run-metadata.md
      prompts/
        <agent-id>.md
      tasks/
        TASK-001/
          task.md
          context.md
          files-inspected.md
          context-pack.md
          plan.md
          proposed.patch
          repair-1.patch
          normalized.patch
          normalized-1.patch
          patch-validation.md
          patch-validation-1.md
          applied.patch
          applied-1.patch
          patch-apply-output.txt
          patch-apply-output-1.txt
          test-output.txt
          review.md
          stage-results.md
          context-out.md
          task-completion.md
          diff.patch
          final-notes.md

Exact artifact names depend on configured stage output values.

Development

Run tests:

python -m unittest discover -v

Compile-check modules:

python -m compileall nightshift tests

Additional docs:

• Quickstart
• Tutorial 01: imageboard with real local models
• Tutorial 02: Lisp with real local models
• Tutorial 03: DeadDrop with fixed tests and telemetry
• Config reference
• Artifact review workflow
• Troubleshooting
• Quickstart Lisp example

Roadmap

The active roadmap now lives in docs/design.md. Completed phase checklists are cleared from that document so it stays focused on the current platform shape and the next important work.