Dev Tools¶
Version: 0.3.0 | Skills: 9 | Agents: 4
Dev Tools provides the development lifecycle toolkit — from feature implementation through code review, documentation, and changelog management.
Skills¶
/feature-dev — Feature Development Workflow¶
The primary skill for implementing new features or significant changes. Runs a structured 7-phase workflow:
- Discovery — Understand feature requirements from user input
- Codebase Exploration — Deep-analysis maps relevant code areas (loads Core Tools deep-analysis)
- Clarifying Questions — Resolve ambiguities with the user
- Architecture Design — Spawn code-architect agents (Opus) to design implementation blueprints
- Implementation — Build the feature following the approved architecture
- Quality Review — Spawn code-reviewer agents (Opus) for correctness, security, and maintainability review
- Summary — Document what was accomplished
# Usage
/feature-dev Add dark mode toggle to the settings page
/feature-dev Implement WebSocket support for real-time notifications
Composition Chain
feature-dev orchestrates multiple agent teams:
- Phase 2: Loads deep-analysis for codebase exploration (code-explorer x N + code-synthesizer x 1)
- Phase 4: Spawns 2-3 code-architect agents (from core-tools) for blueprint design
- Phase 6: Spawns 3 code-reviewer agents for quality review
/docs-manager — Documentation Management¶
Manages MkDocs sites, standalone markdown files, and change summaries through a 6-phase interactive workflow. Supports generating new documentation, updating existing pages, and creating changelogs.
/docs-manager # Interactive discovery
/docs-manager README # Generate a README.md
/docs-manager mkdocs # Set up MkDocs site
/docs-manager changelog # Create change summary
/bug-killer — Hypothesis-Driven Debugging¶
A systematic debugging skill that enforces investigation before fixes. Every bug gets a hypothesis journal, evidence gathering, and root cause confirmation before any code changes are applied.
/bug-killer TypeError: Cannot read property 'id' of undefined
/bug-killer tests/test_auth.py::test_login_fails --deep
5-Phase Workflow:
- Triage & Reproduction — Reproduce the bug, form an initial hypothesis, route to quick or deep track
- Investigation — Gather evidence with language-specific debugging techniques
- Root Cause Analysis — Confirm root cause through hypothesis testing
- Fix & Verify — Apply minimal fix, run tests, write regression test
- Wrap-up & Report — Document the investigation trail, capture project learnings
Track Routing¶
The triage phase routes bugs to one of two investigation tracks:
| Track | Trigger | Investigation | Agents |
|---|---|---|---|
| Quick | Localized bug (1-2 files), clear error, obvious fix | Read error location + immediate callers | None |
| Deep | Multi-file bug, unclear root cause, concurrency/state issues, --deep flag |
2-3 code-explorer agents (core-tools) + 1-3 bug-investigator agents | Sonnet workers |
Auto-Escalation
If 2 hypotheses are rejected during the quick track, the workflow automatically escalates to the deep track. All hypothesis journal entries are preserved across escalation.
Hypothesis Journal¶
The core artifact of every debugging session. Each hypothesis follows a structured format with evidence for/against, a test plan, and status (Pending/Confirmed/Rejected). Rejected hypotheses are never deleted — they provide context for the investigation trail.
Deep Track Agent Teams¶
On the deep track, bug-killer orchestrates two types of agent teams:
- Phase 2: Spawns 2-3
code-exploreragents (from core-tools, Sonnet) to explore focus areas related to the bug - Phase 3: Spawns 1-3
bug-investigatoragents (Sonnet) to test specific hypotheses in parallel - Phase 4: Loads
code-qualityskill for fix validation and scans for related issues across the codebase
After the fix, bug-killer loads the project-learnings skill to evaluate whether the bug reveals project-specific knowledge worth capturing in CLAUDE.md.
/document-changes — Session Change Reports¶
Generates a structured markdown report documenting codebase changes from the current working session. Captures files added, modified, and deleted, along with commit history and a human-readable summary.
/document-changes # Auto-infer scope from changes
/document-changes auth refactor # Specify scope for the report
4-Step Workflow:
- Validate — Confirm the directory is a git repository
- Gather — Collect uncommitted changes, staged changes, and recent commits via git
- Locate — Determine report path (default:
internal/reports/<scope>-YYYY-MM-DD.md) - Generate — Write a structured markdown report with metadata, file table, change details, git status, and commit log
The report is suitable for team reviews, handoff documentation, or personal session records.
/release — Python Package Release¶
Automates Python package releases using uv and ruff. Handles version calculation, changelog updates, and tag creation. Runs on the Haiku model for speed.
Supporting Skills (Agent-loaded)¶
These skills are not directly invoked — they're loaded by agents as reference knowledge:
| Skill | Purpose | Used By |
|---|---|---|
architecture-patterns |
MVC, event-driven, microservices, CQRS pattern knowledge | code-architect agent |
code-quality |
SOLID, DRY, testing strategies, review best practices | code-reviewer agent, bug-killer (deep track) |
changelog-format |
Keep a Changelog specification and entry writing guidelines | changelog-manager agent |
project-learnings |
Captures project-specific patterns and anti-patterns into CLAUDE.md | bug-killer (Phase 5), feature-dev |
Agents¶
| Agent | Model | Tools | Purpose |
|---|---|---|---|
| code-reviewer | Opus | Read, Glob, Grep (read-only) | Reviews code for correctness, security, maintainability with confidence scores |
| bug-investigator | Sonnet | Read, Glob, Grep, Bash (read-only) | Executes diagnostic investigation tasks to test debugging hypotheses |
| changelog-manager | Sonnet | Bash, Read, Edit, Glob, Grep | Analyzes git history and updates CHANGELOG.md |
| docs-writer | Opus | Read, Glob, Grep, Bash | Generates MkDocs or GitHub-flavored markdown documentation |
Note:
code-architect(Opus, read-only) has moved to Core Tools. Feature-dev spawns it cross-plugin using the qualified nameagent-alchemy-core-tools:code-architect.
Read-Only Enforcement
The code-reviewer agent intentionally has no write tools. This enforces separation of concerns — reviewers audit but don't modify code directly. The code-architect agent (now in core-tools) follows the same pattern. Only the lead (feature-dev) applies changes.
Composition Chains¶
graph LR
FD["/feature-dev"] --> DA["deep-analysis"]
DA --> CE["code-explorer x N"]
DA --> CS["code-synthesizer x 1"]
FD --> CA["code-architect x 2-3<br/>(core-tools)"]
FD --> CR["code-reviewer x 3"]
DM["/docs-manager"] --> DA
DM --> DW["docs-writer x N"]
BK["/bug-killer"] --> CE2["code-explorer x 2-3<br/>(core-tools)"]
BK --> BI["bug-investigator x 1-3"]
BK --> CQ["code-quality"]
BK --> PL["project-learnings"]