pilot-shell vs WMDP

Analyzes user intent via the /spec command, automatically classifies tasks as features or bugfixes, and generates structured implementation plans using a state machine dispatcher that routes to feature or bugfix workflows. The planning phase uses Claude to decompose requirements into atomic steps with estimated complexity, then presents a human-reviewable plan before implementation begins. This enforces upfront design thinking and prevents Claude Code from diverging into ad-hoc implementations.

Unique: Uses a dispatcher-based state machine that routes feature and bugfix tasks through separate workflows (feature: plan → implement → verify; bugfix: plan → implement → regression test), with mandatory human approval gates between planning and implementation phases. This architectural pattern prevents Claude from skipping the planning phase entirely.

vs alternatives: Unlike Claude Code alone (which implements immediately) or generic AI agents (which lack project context), Pilot Shell enforces structured planning with automatic task classification and blocks implementation until a human approves the plan.

During the implementation phase of /spec workflows, generates test cases before code is written, then validates that all generated code passes those tests before marking tasks complete. The system uses a verification agent that runs test suites and blocks code merges if coverage or assertions are insufficient. This is enforced via hooks that intercept code changes and validate test presence before allowing commits.

Unique: Integrates test generation into the implementation phase via a hooks pipeline that intercepts code changes and validates test presence before allowing progression. Uses a verification agent that runs test suites and blocks code merges if tests fail or coverage is insufficient, making TDD non-optional rather than optional.

vs alternatives: Standard Claude Code has no built-in test enforcement; Pilot Shell's hooks pipeline and verification agent make test-first development automatic and mandatory, preventing developers from skipping tests even if they wanted to.

Pilot Shell injects project-specific context into Claude's system prompt at session start, including extracted conventions, relevant code patterns, and project rules from the semantic index. The context injection is selective and respects Claude's token budget — only the most relevant patterns are injected based on the current task, preventing context window overflow. The system uses a context monitor to track which files are most relevant to the current task and prioritizes injection of related patterns.

Unique: Uses a context monitor to selectively inject the most relevant project patterns into Claude's system prompt based on task scope, respecting token budgets by prioritizing high-impact patterns. This enables codebase awareness without exceeding context window limits, making large-codebase support practical.

vs alternatives: Unlike RAG systems that inject all matching documents (risking token overflow) or manual context setup (which is tedious), Pilot Shell's selective context injection uses task-aware heuristics to inject only the most relevant patterns, balancing context richness with token efficiency.

The verification phase includes an automated code review agent that checks for style violations, architectural inconsistencies, and deviations from project conventions. The agent uses the extracted project rules and conventions to validate that generated code follows established patterns. Code that violates style or architectural rules is flagged and can block merges, providing automated enforcement of code quality standards without requiring manual review.

Unique: Implements an automated code review agent that validates generated code against extracted project rules and conventions, providing architectural and style enforcement without manual review. The agent uses the same rules extracted by /sync and /learn, making reviews consistent with project standards.

vs alternatives: Unlike manual code review (which is slow and subjective) or linting tools alone (which only check syntax), Pilot Shell's code review agent understands project conventions and architectural patterns, providing semantic-level code quality assurance.

Pilot Shell persists session state (current task, implementation progress, test results, verification status) to disk, enabling recovery if a session crashes or is interrupted. The worker service maintains a session state file that tracks the current /spec task, implementation phase, and verification results. If a session is interrupted, the next session can resume from the last checkpoint, preventing loss of work and enabling recovery from failures.

Unique: Persists session state to disk via the worker service, enabling recovery from crashes and interruptions. Session state includes current task, implementation progress, test results, and verification status, allowing seamless resumption from the last checkpoint.

vs alternatives: Unlike Claude Code alone (which has no session persistence) or manual checkpointing (which is error-prone), Pilot Shell's automatic session persistence enables recovery from crashes without user intervention, making long-running tasks more reliable.

The /sync command builds a semantic search index of the entire codebase using embeddings, then stores project-specific context (architecture patterns, naming conventions, dependencies, test patterns) in a persistent memory store that survives across sessions. This context is automatically injected into Claude's context window at the start of each session, enabling Claude to understand project conventions without requiring manual context setup. The context monitor continuously tracks changes to key files and updates the index incrementally.

Unique: Uses a context monitor hook that tracks file changes and incrementally updates the semantic index, combined with a memory & console system that persists extracted conventions across sessions. The index is injected into Claude's context at session start, eliminating the need for manual context setup while staying within token budgets via selective injection of relevant patterns.

vs alternatives: Unlike Claude Code alone (which has no persistent memory between sessions) or generic RAG systems (which require manual indexing), Pilot Shell's /sync command automatically indexes the codebase and injects relevant context at session start, making project knowledge persistent without manual effort.

The /learn command captures non-obvious discoveries from the current session (e.g., 'this project uses a custom logger instead of console.log', 'all async functions must have timeout handling') and converts them into reusable skill files stored in ~/.pilot/skills/. These skills are automatically loaded into Claude's context for future sessions on the same project, and can be shared across teams via the /vault command. The system uses Claude to extract generalizable patterns from session interactions and format them as structured rules.

Unique: Converts session discoveries into structured skill files that are automatically loaded into Claude's context for future sessions, with a /vault integration for team-wide sharing. Unlike generic documentation, skills are machine-readable and directly injected into Claude's reasoning, making them immediately actionable.

vs alternatives: Standard Claude Code has no mechanism to capture and reuse project-specific patterns; Pilot Shell's /learn command converts ephemeral session insights into persistent, shareable skills that improve Claude's performance on future tasks in the same project.

The /vault command shares rules, commands, skills, hooks, and agents across a team by syncing them to a private Git repository. Each team member's local ~/.pilot/ and ~/.claude/ directories can be configured to pull from a shared vault repository, enabling centralized management of project conventions, custom hooks, and reusable agents. The system uses Git as the backing store and provides conflict resolution via simple merge strategies (last-write-wins or manual resolution).

Unique: Uses Git as the backing store for team knowledge, enabling decentralized sync with version history and audit trails. Rules, skills, hooks, and agents are stored as files in the vault repository and pulled into each team member's local ~/.pilot/ directory, making team knowledge portable and version-controlled.

vs alternatives: Unlike centralized knowledge bases (which require a server) or manual documentation (which gets out of sync), Pilot Shell's /vault uses Git for decentralized, version-controlled sharing of project-specific rules and agents, making team knowledge portable and auditable.

Evaluates LLM outputs against curated question sets spanning three distinct hazard domains (biosecurity, cybersecurity, chemical security) using domain-expert-validated benchmarks. The assessment framework maps model responses to risk levels within each domain, enabling quantitative measurement of dangerous capability presence. Responses are scored against rubrics developed by security domain experts to identify whether models can produce actionable harmful information.

Unique: Combines expert-validated questions across three distinct security domains (biosecurity, cybersecurity, chemical) into a unified benchmark framework, rather than treating each domain separately. Uses domain-expert rubrics for scoring rather than automated classifiers, ensuring nuanced assessment of harmful capability presence.

vs alternatives: More comprehensive than single-domain safety benchmarks (e.g., ToxiGen for toxicity) because it measures dangerous knowledge across multiple hazard categories simultaneously, enabling holistic safety evaluation.

Provides standardized evaluation infrastructure to measure the effectiveness of unlearning techniques (methods that remove dangerous capabilities from trained models) by comparing model performance before and after unlearning interventions. The framework isolates the impact of unlearning by holding the benchmark constant while varying the model state, enabling quantitative assessment of whether dangerous knowledge has been successfully suppressed.

Unique: Provides a standardized evaluation harness specifically designed for unlearning research, with built-in comparison logic and side-effect detection. Unlike generic benchmarks, it explicitly measures delta between model states and flags unintended capability loss.

vs alternatives: More rigorous than ad-hoc unlearning evaluation because it enforces consistent benchmark administration, statistical testing, and side-effect measurement across all methods being compared.

Implements a structured scoring framework where model responses to dangerous knowledge questions are evaluated against expert-developed rubrics that assess the degree of hazard (e.g., specificity, actionability, completeness of harmful information). Responses are scored on multi-point scales (typically 0-4 or 0-5) rather than binary pass/fail, capturing nuance in how dangerous a model's output actually is. Rubrics are domain-specific (biosecurity, cybersecurity, chemical) and developed by subject matter experts to ensure validity.

Unique: Uses domain-expert-developed multi-point rubrics rather than automated classifiers or binary labels, enabling nuanced assessment of dangerous knowledge severity. Rubrics are calibrated to distinguish between vague, incomplete, and highly actionable harmful information.

vs alternatives: More interpretable and defensible than black-box classifiers because rubric criteria are explicit and expert-validated; enables stakeholders to understand why a response received a particular score.

Analyzes patterns in how dangerous knowledge correlates across the three benchmark domains (biosecurity, cybersecurity, chemical security), identifying whether models that excel at suppressing one type of hazard tend to suppress others. The analysis uses statistical correlation and clustering techniques to reveal whether dangerous capabilities are independent or coupled in model behavior. This enables understanding of whether unlearning interventions have domain-specific or global effects.

Unique: Explicitly analyzes relationships between dangerous knowledge across domains rather than treating each domain independently. Enables discovery of whether hazards are coupled or independent in model behavior.

vs alternatives: Provides deeper insight than single-domain benchmarks by revealing how safety properties interact across different hazard categories, informing more effective unlearning strategies.

Manages the creation, validation, and versioning of benchmark questions and rubrics through a structured curation pipeline involving domain experts, adversarial testing, and iterative refinement. The pipeline ensures questions are sufficiently difficult to elicit dangerous knowledge without being unrealistic, and rubrics are calibrated through inter-rater agreement studies. Version control enables tracking of benchmark evolution and ensures reproducibility across research papers.

Unique: Implements a formal curation pipeline with expert validation and inter-rater agreement checks, rather than ad-hoc question collection. Versioning enables reproducible research and transparent tracking of benchmark evolution.

vs alternatives: More rigorous than informal benchmarks because it enforces expert review, inter-rater validation, and version control, reducing bias and enabling reproducible comparisons across papers.

Provides a unified interface for evaluating diverse LLM architectures (open-source models, API-based models, fine-tuned variants) by abstracting away implementation differences. The abstraction handles API calls (OpenAI, Anthropic, etc.), local inference (Hugging Face, Ollama), and custom model serving, enabling consistent benchmark administration across heterogeneous model types. This enables fair comparison between models with different deployment modalities.

Unique: Abstracts away differences between API-based, local, and custom-deployed models through a unified interface, enabling fair comparison without reimplementing benchmark logic for each model type.

vs alternatives: More flexible than model-specific benchmarks because it supports any LLM architecture without code changes, reducing friction for researchers evaluating new models.

Implements rigorous statistical testing to determine whether differences in dangerous knowledge scores between models or unlearning methods are statistically significant or due to random variation. Uses techniques like bootstrap confidence intervals, permutation tests, and effect size estimation to quantify uncertainty in benchmark results. This prevents overconfident claims about safety improvements that may not be robust.

Unique: Integrates formal statistical testing into the benchmark evaluation pipeline rather than relying on point estimates, ensuring claims about safety improvements are statistically justified.

vs alternatives: More rigorous than informal comparisons because it quantifies uncertainty and prevents overconfident claims about safety improvements that may not be robust to sampling variation.

Employs adversarial testing techniques to validate that benchmark questions reliably elicit dangerous knowledge and cannot be easily circumvented by prompt engineering. Red-teamers attempt to find questions that fail to elicit dangerous knowledge or rubric edge cases, and the benchmark is iteratively refined based on findings. This ensures the benchmark is robust to adversarial adaptation and captures genuine dangerous capabilities rather than surface-level patterns.

Unique: Incorporates formal red-teaming into the benchmark validation pipeline rather than assuming questions are robust, ensuring the benchmark remains effective against adversarial adaptation.

vs alternatives: More robust than static benchmarks because it actively searches for evasion techniques and iteratively refines questions, reducing the risk that models can circumvent the benchmark through prompt engineering.