33 KiB
NightShift
Auditable Local-First AI Coding Pipelines
Version: v0.1 Draft Author: K455 Status: Design Proposal
1. Executive Summary
NightShift is a local-first AI pipeline runner designed to execute long-running coding workflows against a constrained project workspace.
The system is intended to run overnight or unattended for extended periods while remaining:
- Cheap
- Correct
- Auditable
- Safe
- Reviewable
NightShift is not designed to be a fully autonomous "AI software engineer." Instead, it is a deterministic orchestration system that allows fallible AI agents to operate within constrained, test-driven, auditable workflows.
The core philosophy is:
Treat LLMs like unreliable distributed systems.
Agents are bounded by:
- Scoped repository access
- Structured stage contracts
- Explicit retry behavior
- Tests and static checks
- Review stages
- Context compaction
- Artifact logging
The intended workflow is:
-
User provides:
- Repository
- Task list
- Pipeline configuration
- Agent definitions
-
NightShift:
- Selects the next task
- Generates a plan
- Reviews the plan
- Implements changes
- Runs tests/static analysis
- Reviews results
- Retries if necessary
- Produces an overnight report
The result is a reviewable repository state and a full audit trail of AI behavior.
2. Goals
2.1 Primary Goals
Local-first execution
The system should work primarily with local models and local execution environments.
Examples:
- Ollama
- Local transformers
- Local agent runtimes
- Claude Code
- Codex CLI
Long-running unattended workflows
NightShift should support:
- Overnight execution
- Large task chains
- Multi-stage workflows
- Automated retries
- Context handoff between stages
Auditability
Every important action should be recorded.
Users should be able to inspect:
- Prompts
- Plans
- Reviews
- Command outputs
- Diffs
- Test results
- Retry reasoning
- Final summaries
Cheapness-first execution
The orchestration layer should assume:
- Cheap local models handle most work
- Expensive models are escalation layers
- Context size matters
- Token usage matters
- Retry cost matters
Safe repository boundaries
The system should:
- Restrict file access
- Restrict shell commands
- Avoid destructive operations
- Minimize repository damage
2.2 Non-Goals (v1)
The following are intentionally out of scope for v1:
- Fully autonomous software development
- Parallel distributed execution
- Automatic deployment
- Cloud-native orchestration
- Dynamic self-modifying pipelines
- Autonomous internet access
- Agent swarms
- Arbitrary Python execution hooks
- Automatic git pushes
- Full DAG orchestration
3. Design Philosophy
NightShift is built around several core principles.
3.1 Deterministic orchestration
Agents are nondeterministic.
The orchestration system should not be.
Pipeline behavior should be:
- Predictable
- Reproducible
- Configurable
- Explicit
3.2 Structured state transitions
NightShift uses a state-machine workflow model.
A task moves through defined stages:
Task Queue
-> Plan
-> Plan Review
-> Implement
-> Test
-> Static Check
-> Review
-> Retry / Complete
Each stage produces:
status: pass | fail | retry | escalate
reason: string
next_stage: optional
context_update: optional
This allows the pipeline runner to remain deterministic even while agents are probabilistic.
3.3 Context compaction
Agents should not inherit unlimited history.
Instead:
- Project-level context is persistent and compact
- Task-level context is scoped
- Retry context is summarized
- Stage context is minimized
This reduces:
- Token costs
- Context poisoning
- Hallucination drift
- Recursive confusion
3.4 Reviewability over autonomy
NightShift is optimized to produce:
- Reviewable code
- Reviewable reports
- Reviewable reasoning
The primary output is:
A useful morning review state.
Not:
Fully autonomous shipping.
4. Architecture Overview
4.1 High-Level Components
+-------------------+
| Task Parser |
+-------------------+
|
v
+-------------------+
| Pipeline Runner |
+-------------------+
|
v
+-------------------+
| Stage Executor |
+-------------------+
| |
| +----------------+
| |
v v
+-----------+ +----------------+
| Agent API | | Command Runner |
+-----------+ +----------------+
| |
v v
+-----------+ +----------------+
| LLM Model | | Test/Lint/etc |
+-----------+ +----------------+
4.2 Core Components
Task Parser
Responsible for:
- Reading markdown task files
- Parsing acceptance criteria
- Tracking completion state
- Determining dependencies
Pipeline Runner
Responsible for:
- Stage orchestration
- Retry logic
- State transitions
- Artifact management
- Context propagation
Stage Executor
Responsible for:
- Executing stage definitions
- Calling agents
- Running commands
- Collecting outputs
Agent Layer
Responsible for:
- Prompt construction
- Model backend integration
- Structured output parsing
- Context injection
Command Runner
Responsible for:
- Executing tests
- Static analysis
- Formatting
- Shell command restrictions
- Sandboxing
5. Workflow Model
5.1 State Machine Model
NightShift uses a configurable state-machine workflow.
This was selected over:
- DAG orchestration
- Arbitrary scripting
because:
- v1 executes one task at a time
- Retry loops are first-class
- Auditability is easier
- Deterministic transitions are simpler
5.2 Default Pipeline
PLAN
↓
REVIEW_PLAN
↓
IMPLEMENT
↓
TEST
↓
STATIC_ANALYSIS
↓
REVIEW
↓
DECISION
Decision outcomes:
- COMPLETE
- RETRY_IMPLEMENTATION
- RETRY_PLANNING
- FAIL
5.3 Configurable Pipelines
Pipelines are defined declaratively.
Users may:
- Swap stage orders
- Add/remove stages
- Define retry behavior
- Use different models
- A/B test prompts
- Experiment with reasoning structures
6. Configuration System
6.1 Configuration Format
NightShift uses YAML configuration files.
Reasons:
- Human-readable
- Good nested structure support
- Easier workflow representation than TOML
- Safer than arbitrary Python execution
6.2 Example Configuration
project:
name: my-project
root: .
task_file: tasks.md
artifact_dir: .nightshift
safety:
require_clean_worktree: true
scoped_paths:
- src/
- tests/
forbidden_commands:
- rm -rf
- git push
allowed_commands:
- cargo test
- cargo fmt
- cargo clippy
agents:
planner:
backend: ollama
model: qwen2.5-coder:14b
system_prompt: agents/planner.md
implementer:
backend: claude-code
model: sonnet
system_prompt: agents/implementer.md
reviewer:
backend: ollama
model: deepseek-r1:32b
system_prompt: agents/reviewer.md
pipeline:
max_task_retries: 3
stages:
- id: plan
type: agent
agent: planner
- id: review_plan
type: review
agent: reviewer
on_fail: plan
- id: implement
type: agent
agent: implementer
- id: test
type: command
commands:
- cargo test
- id: static
type: command
commands:
- cargo fmt --check
- cargo clippy -- -D warnings
- id: review
type: review
agent: reviewer
on_fail: implement
7. Task System
7.1 Task Format
Tasks are defined in markdown.
Example:
- [ ] TASK-001: Add retry support to pipeline runner
Acceptance Criteria:
- Retries configurable per stage
- Retry summaries persisted
- Retry count visible in final report
7.2 Task Lifecycle
Each task:
- Is parsed
- Is assigned a workspace
- Receives planning
- Receives implementation
- Is validated
- Is reviewed
- Produces artifacts
- Is marked complete or failed
7.3 Task Dependencies
Future versions may support:
TASK-003 depends on TASK-001
However:
- Tasks should remain independently testable when possible
- Pipelines should maintain a buildable repository state
8. Agent Model
8.1 Agent Roles
Agents are specialized.
Example roles:
- planner
- implementer
- reviewer
- summarizer
- test-writer
8.2 Agent Definitions
Agents are configurable.
Each agent defines:
- Backend
- Model
- System prompt
- Constraints
- Output schema
8.3 Multi-Backend Support
NightShift should support:
- Ollama
- Claude Code
- Codex CLI
- Future local runners
This allows:
- Cheap local planning
- Expensive selective escalation
- Hybrid pipelines
8.4 Structured Outputs
Agents should emit machine-readable results.
Example:
status: pass
summary: |
Tests succeeded.
issues:
- None
next_stage: review
9. Context System
9.1 Context Layers
NightShift uses layered context.
Project Context
Long-lived information:
- Architecture
- Coding standards
- Constraints
- Previous summaries
Task Context
Task-specific information:
- Acceptance criteria
- Relevant files
- Prior retries
- Implementation notes
Retry Context
Compact summaries of:
- Previous failures
- Previous reviews
- Previous test errors
9.2 Context Compaction
Every stage should summarize output.
This prevents:
- Infinite context growth
- Token explosion
- Recursive hallucination
- Low-signal history accumulation
10. Safety Model
10.1 Repository Scope Restrictions
NightShift should restrict:
- Accessible directories
- Writable paths
- Executable commands
10.2 Command Restrictions
Commands are allowlisted.
Potentially dangerous commands are forbidden.
Examples:
Forbidden:
- rm -rf
- git push
- curl | bash
10.3 Clean Worktree Requirement
v1 may optionally require:
git status == clean
before execution.
This simplifies:
- Auditability
- Recovery
- Diff inspection
11. Testing and Validation
11.1 Validation Pipeline
Validation occurs in multiple stages:
Tests
↓
Static Analysis
↓
Review Agent
↓
Decision
11.2 Global Test Suite
Tests are global.
Rationale:
- New changes must not break old functionality
- Pipeline should maintain cumulative stability
11.3 Generated Tests
Agents may generate tests for features.
Generated tests become part of the persistent suite.
12. Artifact System
12.1 Artifact Goals
Artifacts provide:
- Auditability
- Replayability
- Debugging
- Historical inspection
- Prompt experimentation
12.2 Example Layout
.nightshift/
project-context.md
runs/
2026-05-16-overnight/
run-summary.md
config.snapshot.yaml
tasks/
TASK-001/
task.md
plan.md
plan-review.md
implementation-log.md
test-output.txt
static-output.txt
review.md
final-notes.md
diff.patch
context-out.md
13. Overnight Report
At completion NightShift generates:
- Completed tasks
- Failed tasks
- Retry counts
- Files modified
- Test results
- Reviewer summaries
- Remaining issues
- Suggested follow-up work
The goal is:
Wake up to a review package.
14. Future Directions
Potential future features:
- Parallel task execution
- DAG workflows
- Distributed workers
- Sandboxed containers
- Git branch isolation
- Agent tournaments
- Constraint language experimentation
- Prompt A/B testing
- Semantic memory systems
- Multi-repo orchestration
- Web dashboard
- Cost telemetry
- Human approval gates
15. Risks
15.1 Context poisoning
Mitigation:
- Context compaction
- Retry summarization
- Structured stage boundaries
15.2 Agent loops
Mitigation:
- Explicit retry counts
- Deterministic transitions
- Timeout handling
15.3 Repository damage
Mitigation:
- Scoped directories
- Command restrictions
- Validation stages
15.4 Cost explosion
Mitigation:
- Local-first execution
- Context minimization
- Escalation-only expensive models
16. Implemented Baseline
The MVP and the patch-capable local runner are implemented.
NightShift currently provides:
nightshift initfor starter project generationnightshift validatefor config, prompt, task, dependency, path, and command validationnightshift statusfor read-only project inspectionnightshift runfor the next runnable incomplete tasknightshift run --task TASK-IDfor a specific tasknightshift run --allfor sequential multi-task executionnightshift webfor a read-only artifact dashboard- Operational run logging to the CLI, per-run logs, and aggregate logs
- Markdown task parsing with descriptions, acceptance criteria, completion state, and dependency bullets
- Dependency validation for missing references and simple cycles
- Dependency-aware task selection and task blocking
- Declarative YAML pipeline execution
- Command, agent, agent-review, review, summarize, repo-context, code-writer, file-writer, patch-normalizer, patch-validator, and patch-apply stage handling
- Retry redirection with a configured task retry limit
- Command-backed agents
- Ollama-backed local model agents through the local HTTP API
- OpenAI-compatible local/server model agents
- Per-agent temperature settings
- Cost, runtime, retry, and estimated token telemetry summaries
- Scoped repo lookup tools:
list_files,read_file, andgrep - Lightweight semantic repository indexing for files, symbols, imports, tests, and compact task context
- Planner lookup requests,
files-inspected.md, and planner reruns with retrieved context - Project context chart generation
- Context pack generation
- Unified diff code-writing contract
- Deterministic diff generation from model-supplied complete file blocks
- Patch normalization, deterministic hunk-count repair, validation, dry-run, and apply modes
- Per-attempt retry patch artifacts such as
repair-1.patch,normalized-1.patch, andpatch-validation-1.md - Test/static failure repair loops via bounded stage retries
- Prompt bundle construction with project, task, retry, and previous-stage context
- Prompt snapshots and run metadata for experiment comparison
- Optional experiment labels and prompt variant metadata
- Command allowlists and forbidden-fragment checks
- Optional shell-free command execution
- Per-stage command timeouts
- Project-root-restricted command working directories
- Environment variable allowlists for command stages
- Scoped path and artifact path safety checks
- Optional clean-worktree enforcement
- Pre-run and post-run git status artifacts
- Per-task
diff.patchartifacts - Task completion mutation for successful runs
- Per-run and per-task markdown/text artifacts
- Project, task, retry, and context-out files
- Final task notes, stage summaries, task completion artifacts, and run summaries
- Documentation for config, artifact review, troubleshooting, quickstart, and patch workflows
- A complete fake-agent patch-mode quickstart Lisp example under
examples/quickstart-lisp/ - A deterministic pastebin tutorial template with model fallback configuration
The system remains sequential and local-first. It is designed to produce reviewable artifacts and repository state, not to deploy, push, or autonomously ship changes.
16.5 Current Tasks Todo
- TASK-001: Failure classification pipeline
Dependencies:
- None
Description: Add a deterministic post-failure analysis stage that runs after every failed command or test execution. The classifier should inspect stdout/stderr, exit codes, modified files, and failing tests, then categorize the failure and recommend the next orchestration action.
Acceptance Criteria:
-
Captures stdout, stderr, exit code, modified files, and failing test names
-
Produces structured output containing:
- failure category
- probable root cause
- confidence
- recommended next action
- retry recommendation
-
Supports initial categories:
- syntax/import error
- missing dependency
- missing resource/fixture
- environment/config issue
- API misuse
- test expectation mismatch
- logic bug
- stuck/unclear
-
Integrates into orchestration pipeline before retries occur
-
Includes tests for classification behavior
-
TASK-002: Structured blocked/resource request system
Dependencies:
- TASK-001
Description: Allow agents to explicitly declare missing resources or environmental requirements instead of endlessly retrying implementation attempts. Add structured "blocked" responses and runtime support for generating common fixtures and test resources.
Acceptance Criteria:
-
Supports structured blocked responses such as:
- missing fixture
- missing config
- missing database
- missing asset
-
Includes fixture generators for:
- PNG/JPG images
- JSON fixtures
- sqlite databases
- text/blob files
-
Runtime can automatically satisfy supported requests
-
Generated fixtures are isolated to the active run directory
-
Includes tests for fixture generation and blocked flow handling
-
TASK-003: Dedicated debugger agent role
Dependencies:
- TASK-001
Description: Introduce a dedicated debugger agent responsible for diagnosis rather than implementation. The debugger reviews failed attempts and provides concise explanations and recommendations for the implementer.
Acceptance Criteria:
-
Debugger receives:
- task description
- current patch
- failure output
- recent attempt history
-
Debugger outputs:
- concise diagnosis
- recommended next action
- "do not modify" guidance
-
Debugger does not directly modify code initially
-
Implementer receives debugger output in retry context
-
Includes tests for debugger orchestration behavior
-
TASK-004: Stuck detection and escalation policy engine
Dependencies:
- TASK-001
- TASK-003
Description: Detect retry churn loops and automatically escalate to different models, debugger review, or human intervention when progress stalls.
Acceptance Criteria:
-
Tracks:
- repeated failures
- repeated file edits
- unchanged failing tests
- expanding diff size
- oscillating implementations
-
Supports configurable retry budgets
-
Supports escalation policies such as:
- debugger review
- larger local model
- cloud model
- human review
-
Stops infinite retry loops
-
Includes tests for churn detection and escalation behavior
-
TASK-005: Multi-model orchestration and escalation
Dependencies:
- TASK-004
Description: Add support for multiple implementation and debugging models with configurable routing, retry budgets, and escalation rules. Provide examples
Acceptance Criteria:
-
Supports separate model pools for:
- implementers
- debuggers
- escalation models
-
Allows configurable retry budgets per model
-
Supports configurable temperatures per role
-
Allows fallback ordering between models
-
Integrates with escalation policy engine
-
Includes tests for model routing and escalation flow
-
TASK-006: Dependency management agent
Dependencies:
- TASK-001
Description: Add a dependency management subsystem capable of detecting missing packages, understanding dependency manifests, and automatically resolving installation issues. Just for python now.
Acceptance Criteria:
-
Detects:
- missing imports
- missing packages
- dependency manifest drift
- invalid package references
-
Supports:
- pip
- uv
- poetry
- requirements.txt
- pyproject.toml
-
Can propose or apply dependency fixes
-
Can retry runs after dependency installation
-
Includes tests for dependency resolution flows
-
TASK-007: Patch governor and diff safety system
Dependencies:
- TASK-004
Description: Prevent runaway architectural rewrites and unrelated modifications during retry loops by analyzing diffs and rejecting unsafe patches.
Acceptance Criteria:
-
Detects:
- unrelated file modifications
- excessive diff growth
- deletion-heavy patches
- architecture drift
-
Can reject unsafe patches before commit/application
-
Produces actionable rejection feedback for implementers
-
Supports configurable thresholds and policies
-
Includes tests for diff analysis and patch rejection behavior
-
TASK-008: Integration sandbox runner
Dependencies:
- None
Description: Add a one-command integration environment runner that creates isolated timestamped run directories for NightShift testing and orchestration experiments. This is the equivalent of doing --template with the tutorials
Acceptance Criteria:
-
Adds command:
nightshift integ-run
-
Creates timestamped run directories under:
integ_runs/
-
Automatically:
- creates isolated venv
- installs project dependencies
- initializes clean template/project state
-
Adds
integ_runs/to.gitignore -
Persists:
- logs
- transcripts
- patches
- generated artifacts
-
Supports cleanup policies for old runs
-
Includes tests for sandbox creation and cleanup behavior
-
TASK-009: Structured retry memory system
Dependencies:
- TASK-001
- TASK-004
Description: Persist compact structured summaries of previous attempts to prevent retry amnesia and repeated failed approaches.
Acceptance Criteria:
-
Stores:
- attempted fixes
- failure causes
- rejected hypotheses
- successful observations
-
Produces compact retry summaries instead of raw log dumps
-
Retry summaries are injected into implementer context
-
Supports configurable memory compaction
-
Includes tests for retry memory summarization behavior
-
TASK-010: Environment-aware execution diagnostics
Dependencies:
- TASK-001
- TASK-006
Description: Improve orchestration awareness of environment-level failures versus implementation-level failures to reduce wasted retries and false debugging paths.
Acceptance Criteria:
-
Distinguishes:
- environment failures
- dependency failures
- fixture/resource failures
- implementation logic failures
-
Prevents implementation retries when environment is invalid
-
Surfaces actionable remediation guidance
-
Integrates with failure classifier and dependency manager
-
Includes tests for environment diagnostic behavior
-
TASK-011: Update tutorials to reflect the previous changes to the templates as needed
Description: Tutorials should have the newly added features when relevant.,
Acceptance Criteria:
-
Tutorials have features
-
TASK-012: Stage output should be more organized. Right now run/task/ produces many files and it is difficult to keep track of. Either sub folders for retries, appending for retries, or compacting, whichever makes sense for our use case.
-
TASK-013: Cost, token, and runtime telemetry
Dependencies:
- TASK-005
Description: Track orchestration cost, latency, retry counts, token usage, and success rates across agents and models. Generally telemetry for analyzing model efficiency and usage. Which model fixes bugs fastest?
Acceptance Criteria:
-
Tracks token usage per agent and run
-
Tracks runtime duration and retry counts
-
Records success/failure metrics
-
Supports per-model statistics
-
Exposes telemetry summaries and reports
-
Includes tests for telemetry aggregation
-
TASK-014: Repository semantic indexing system
Dependencies:
- None
Description: Build lightweight semantic indexing over repositories so agents can retrieve relevant files, symbols, tests, and architecture context without loading excessive raw context.
Acceptance Criteria:
-
Indexes symbols, files, imports, and tests
-
Supports semantic and keyword search
-
Returns compact relevant context snippets
-
Reduces prompt context size
-
Includes tests for retrieval quality
-
TASK-015: Pastebin tutorial project template
Dependencies:
- TASK-008
- TASK-005
Description:
Add a new tutorial project template for NightShift based on a small Pastebin/snippet-hosting service. This should work like the existing imageboard tutorial, but be simpler, more deterministic, and easier to use for testing agent orchestration. The template should be creatable with --template.
Acceptance Criteria:
- Adds a new template named
pastebin - Supports creating the tutorial project with a command such as:
nightshift init --template tutorial-pastebin
- Template includes a small but realistic app with:
- snippet creation
- snippet viewing
- snippet listing
- optional expiration field
- tags or language field
- basic search/filtering
- Includes a test suite with multiple incremental tasks suitable for agent testing
- Avoids complex media/file-upload behavior from the imageboard tutorial
- Uses deterministic fixtures and simple dependencies
- Includes clear task descriptions for the agent to complete
- Includes README instructions explaining the tutorial goals
- Supports model fallback ordering for this template:
qwen2.5-coder:14bcarstenuhlig/omnicoder-9bdeepseek-coder-v2:16b
- If the first model fails or exceeds its retry budget, the next fallback model is attempted
- Records which model handled each attempt
- Includes tests for template creation and model fallback configuration
17. Current Product Shape
The implemented product is now a practical local runner rather than only a single-task MVP.
17.1 CLI Workflow
Common workflow:
nightshift init
nightshift validate
nightshift status
nightshift run
nightshift run --task TASK-001
nightshift run --all
nightshift web
The CLI can validate a project, select runnable tasks, enforce dependencies, run one or more tasks, and report artifact locations.
17.2 Artifact Workflow
Artifacts are still the primary audit surface.
Current run artifacts include:
.nightshift/
project-context.md
runs/
<run-id>/
run-summary.md
config.snapshot.yaml
run-metadata.md
prompts/
<agent-id>.md
tasks/
TASK-001/
task.md
context.md
plan.md
files-inspected.md
context-pack.md
proposed.patch
normalized.patch
patch-validation.md
applied.patch
patch-apply-output.txt
test-output.txt
review.md
stage-results.md
context-out.md
task-completion.md
git-status-before.txt
git-status-after.txt
diff.patch
final-notes.md
Exact task artifact names depend on configured stage output values.
17.3 Dashboard Workflow
The web dashboard is read-only and artifact-driven.
It currently:
- Lists runs from
.nightshift/runs/ - Shows run summaries
- Links to text and markdown artifacts
- Safely rejects artifact path traversal
- Auto-refreshes
It does not:
- Start or stop runs
- Mutate config or tasks
- Provide approval gates
- Stream live process output
- Authenticate users
17.4 Known Limitations
Current limitations:
- Execution is sequential; there is no parallel task runner.
- The web dashboard is read-only and artifact-oriented.
- Flask is optional;
nightshift webrequires it to be installed. - Model backends depend on the user's local model server, Ollama installation, or command wrappers.
- Git artifacts can be unavailable or degraded in non-git repositories or repositories blocked by Git safe-directory rules.
- Task mutation is intentionally minimal and only flips matching checklist lines.
- Patch application currently uses
git apply; non-git workflows are limited. - Command configuration remains string-first for compatibility.
- There is no branch isolation, resumable run state machine, approval workflow, or deployment integration.
18. Active Roadmap
Completed phase checklists are removed from this design document once they are reflected in the implemented baseline and user-facing docs. Track future phase work here only while it is active, using concise implementation notes when a decision needs durable context.
The next important additions are:
-
Branch isolation for patch runs Run each task on a dedicated branch or worktree, record branch metadata, and make rollback/review safer.
-
Resumable run state Persist machine-readable run state so interrupted runs can continue from the last completed stage instead of restarting.
-
Human approval gates Add optional approval stages before patch apply, after failed validation, or before task completion.
-
Structured patch policy config Move max files, max lines, forbidden paths, allowed file types, binary rejection, and protected files into a reusable project-level write policy.
-
Better model backend support Expand OpenAI-compatible behavior, add request metadata artifacts, support response format hints, and document local server patterns. Machine-readable Ollama output now uses the HTTP API instead of the interactive
ollama runterminal path; keep this non-terminal capture policy for future model backends where exact patch text matters. -
Deterministic edit formats beyond full files The
file_writerstage now generates unified diffs from complete file blocks. Future work should add smaller structured edit descriptions for large files while preserving deterministic diff generation. -
Retry artifact versioning Continue improving per-attempt artifact preservation. Patch retries now preserve files such as
repair-1.patch,normalized-1.patch, andpatch-validation-1.md; future work should add richer latest-attempt indexes and dashboard navigation. -
Patch repair stage Hunk counts are now deterministically recomputed during normalization for direct unified diff output. Future work should add an explicit patch repair stage for malformed hunk bodies that receives the invalid patch, validation error, and relevant source excerpts, then returns a complete replacement patch. This stage should remain bounded by strict validation and should not silently guess intent for arbitrary malformed hunks.
-
Richer dashboard Add task/stage navigation, patch views, validation status, run log tail, and artifact links without adding mutation controls.
-
Project context chart improvements Use language-aware parsers where available, include import graphs, ownership hints, and stale-context detection.
-
Stronger repair feedback Feed compact test/static failure summaries, patch apply errors, and reviewer objections into repair attempts with clearer bounded policies.
-
End-to-end apply-mode examples Add more small target projects and fake-agent fixtures that exercise patch apply, repair, validation failure, and review retry paths.
-
Packaging and dependency extras Add optional extras such as
nightshift[web], document supported Python versions, and prepare the project for repeatable installation.
Implementation note:
Recent local-model patch experiments exposed repeated line-fragment artifacts where long generated lines were split and the tail was duplicated on the following line. This affected prose and unified diffs, producing malformed hunk lines that strict validation correctly rejected. Treat this as a backend/output-capture and patch-contract problem before adding editor or linter agents: avoid terminal streaming for machine output, preserve retry artifacts, and prefer deterministic diff generation when exact syntax matters.
Appendix A: Design Decisions and Rationale
A.1 Local-first architecture
Decision:
- Prefer local models and local execution
Reasoning:
- Cheapness-first design
- Better experimentation
- Better privacy
- Reduced vendor dependency
- Better overnight scalability
A.2 State machine over DAG
Decision:
- Use configurable state-machine workflows
Reasoning:
- One-task-at-a-time execution
- Retry loops are primary workflow behavior
- Easier auditing
- Easier debugging
- Simpler MVP
A.3 YAML configuration
Decision:
- Use declarative YAML config
Reasoning:
- Human-readable
- Easier nested workflow representation
- Safer than arbitrary Python
- Better portability
A.4 Cheapness-first model routing
Decision:
- Use expensive models selectively
Reasoning:
- Overnight pipelines can become token-expensive
- Local models are sufficient for many stages
- Review stages benefit more from premium models
A.5 Strict repository scoping
Decision:
- Limit writable paths and executable commands
Reasoning:
- Prevent accidental damage
- Maintain trust in unattended execution
- Improve auditability
A.6 Reviewable output over autonomy
Decision:
- Produce review packages rather than autonomous shipping
Reasoning:
- Human review remains critical
- Improves safety
- Improves correctness
- Keeps architecture grounded and practical
A.7 Layered context model
Decision:
- Separate project, task, and retry context
Reasoning:
- Reduces token usage
- Prevents context explosion
- Improves signal quality
- Prevents recursive drift
A.8 Artifact-heavy architecture
Decision:
- Persist plans, logs, reviews, outputs, and summaries
Reasoning:
- Debugging
- Prompt experimentation
- A/B testing
- Replayability
- Portfolio visibility
A.9 No parallelism in v1
Decision:
- Execute one task at a time
Reasoning:
- Simpler correctness model
- Easier debugging
- Easier repository safety
- Easier context management
A.10 Declarative pipelines first
Decision:
- No arbitrary Python hooks in v1
Reasoning:
- Safer execution
- Easier reproducibility
- Easier auditing
- Easier portability
Closing Statement
NightShift is intended to explore a practical middle ground between:
- Fully manual software engineering
- Reckless autonomous agent systems
The system assumes that AI agents are useful but unreliable.
NightShift therefore treats agents as bounded workers inside deterministic, auditable, test-driven workflows.
The primary output is not blind autonomy.
The primary output is trustworthy leverage.