swe-workbench

SWE Workbench

A senior engineer's toolkit for Claude Code — principled design, language expertise, pragmatic workflows.

What it is

swe-workbench bundles the reasoning a careful senior engineer does every day: architectural judgement (Clean Architecture, DDD, SOLID), test discipline (TDD, F.I.R.S.T.), pattern fluency (GoF and beyond), and idiomatic expertise in Bash, C#, Dart, Go, Java, Kotlin, Python, Ruby, Rust, SQL, Swift, and TypeScript. Principle and language skills auto-hint by trigger (a non-blocking hint fires when a matching file is touched); commands and subagents are there when you want explicit help.

Install

From the marketplace:

/plugin marketplace add lugassawan/swe-workbench
/plugin install swe-workbench

For local development:

git clone https://github.com/lugassawan/swe-workbench
cd swe-workbench
/plugin marketplace add $(pwd)
/plugin install swe-workbench

For the Pi Coding Agent (a separate harness from Claude Code — not a /plugin install):

pi install git:github.com/lugassawan/swe-workbench

This loads the same skills/, commands/, and agents/ trees Claude Code uses, via the runtime adapter in pi/extensions/. See docs/pi-port.md for the Pi-specific rulings.

The GitHub-facing commands shell out to the GitHub CLI. /swe-workbench:report-issue is verified against gh 2.45.0 or newer.

What's inside

• Commands — /swe-workbench:review, /swe-workbench:design, /swe-workbench:architect, /swe-workbench:document, /swe-workbench:refactor, /swe-workbench:migrate, /swe-workbench:debug, /swe-workbench:implement, /swe-workbench:hotfix, /swe-workbench:extend, /swe-workbench:test, /swe-workbench:security-review, /swe-workbench:capture, /swe-workbench:report-issue, /swe-workbench:cleanup-merged, /swe-workbench:sync, /swe-workbench:address-feedback, /swe-workbench:audit-codebase, /swe-workbench:codebase-knowledge, /swe-workbench:doctor, /swe-workbench:converge, /swe-workbench:handoff — see docs/catalog.md.
• Subagents — swe-workbench:accessibility-auditor, swe-workbench:alignment-assessor, swe-workbench:architect, swe-workbench:auditor, swe-workbench:code-impl, swe-workbench:conflict-resolver, swe-workbench:contributor-auditor, swe-workbench:debugger, swe-workbench:dependency-auditor, swe-workbench:e2e-test-verifier, swe-workbench:e2e-test-writer, swe-workbench:migrator, swe-workbench:performance-tuner, swe-workbench:product-designer, swe-workbench:product-manager, swe-workbench:redundancy-assessor, swe-workbench:refactorer, swe-workbench:reviewer, swe-workbench:security-auditor, swe-workbench:senior-engineer, swe-workbench:tech-writer, swe-workbench:test-reviewer, swe-workbench:test-writer — see docs/catalog.md.
• Principles — Clean Architecture, DDD, SOLID, TDD, design patterns, clean code, observability, API design, concurrency, data modeling, error handling, security, product design — auto-hint by trigger keyword.
• Languages — Bash, C#, Dart, Go, Java, Kotlin, Python, Ruby, Rust, SQL, Swift, TypeScript — auto-hint by file extension (subagents load deterministically via catalog injection).
• Integrations — external-service context skills (swe-workbench:ticket-context, swe-workbench:observability-context, swe-workbench:comms-context) — auto-load on ticket references, Sentry links, Slack/PagerDuty links to feed the full context into commands — see docs/catalog.md.
• Workflows — development orchestrator wrapping the full 5-phase implementation lifecycle; swe-workbench:workflow-audit-emit-issues files grouped GitHub issues from codebase audit findings.

Full reference tables → docs/catalog.md. Extending guide and philosophy → docs/extending.md. Runtime dependencies → docs/dependencies.md.

Secret detection

Every Write and Edit tool call is scanned for hardcoded secrets (GitHub tokens, AWS keys, .env-style assignments) before the file is written. Detected secrets are blocked with a BLOCKED: message naming the pattern, line number, and file. Use # nosecret on a line to suppress intentional fixtures. See docs/secret-detection.md for the full pattern list, suppression options, and security notes.

Workflow state persistence

When Claude Code auto-compacts a long conversation, any in-progress swe-workbench:workflow-development, swe-workbench:workflow-bug-triage, or swe-workbench:workflow-pr-review state is saved to a sidecar JSON file under .claude/cache/workflow-state/. A SessionStart hook detects this file and injects a resume preamble so the workflow continues at the correct phase — no manual restart needed. The hook fires on compaction as well as on plain session startup/resume, and also nudges a worktree re-anchor when the session's cwd has drifted from a linked worktree. See shared/docs/workflow-state.md for the schema, lifecycle table, and a manual smoke test.

.claude/cache/ is this repo's ephemeral-state directory — add it to your own repo's .gitignore when installing swe-workbench, or these sidecars become committable. /swe-workbench:doctor checks for this and warns if it's missing.

Skill-usage telemetry

When the orchestrator dispatches a subagent, the skills that subagent invokes are surfaced in the transcript:

Skills used by reviewer: swe-workbench:principle-code-review, swe-workbench:principle-clean-code

Top-level skill calls and zero-skill runs produce no output. Individual agents can opt out via skill_telemetry: false in their frontmatter. See docs/skill-usage-telemetry.md for full details.

Pi subagent quota profiles

Use /swe-workbench-profile economy|balanced|full to choose a user-global Pi subagent profile. Economy reduces routine child cost while retaining verification and independent review; Balanced keeps declared effort and complete preloads with a 16k success-output cap; Full uses complete preloads and raises known effort levels to max. See task dispatch.

Contributing

See CONTRIBUTING.md.

License

MIT.