ralph-claude-code

Ralph's per-loop execution contract — one task from fix_plan.md, the RALPH_STATUS exit block, epic-boundary QA deferral, and the dual-condition EXIT_SIGNAL gate. Invoke at the start of every Ralph loop so the response follows the contract the harness depends on.

Eval-first principles and cost-aware model routing for the Ralph loop. This skill should be used before committing budget to a MEDIUM+ task or when choosing a sub-agent: articulate the behavior to verify first, then route by task type (docs/tools → Haiku, code → Sonnet, arch → Opus, with an Opus escalation after repeated QA failures). Integrates with lib/complexity.sh and the optional sdk/ralph_sdk/cost.py::select_model.

Token-hygiene pass for the Ralph loop. This skill should be used before reading another large file or near the end of a Continue-As-New window: it drops stale file reads, prefers targeted Grep over full-file Read, and consolidates repeated file-scan patterns to stop the drift where each iteration re-reads the same file and burns prompt cache. Works with the SDK ContextManager (sdk/ralph_sdk/context.py) and the Continue-As-New trigger.

Run, monitor, and continuously improve Ralph (autonomous dev loop for Claude Code) against a Linear backlog. This skill should be used when the user says "run ralph", "start ralph", "ralph campaign", "monitor ralph", "restart ralph", "ralph against TAP-### / ENG-### / <ISSUE-ID>", or wants Ralph to chew through filed epics autonomously with the orchestrator filing follow-up Linear issues for the friction patterns it observes. Encodes proven startup, monitor-filter, kill-safety, and subagent-delegation patterns. Keeps the orchestrator's context window tiny by delegating Linear writes, commit verification, and bulk log analysis to subagents.

Research-before-coding workflow for the Ralph loop. This skill should be used before writing a new utility, helper, or abstraction: it searches the repo, installed libraries, and public registries for an existing solution first, delegating deep search to the ralph-explorer sub-agent (Haiku). Use it whenever the next task — a fix_plan item or a Linear issue — says "add/implement/build/wrap/integrate" something that sounds generic.

Pre-commit / epic-boundary quality pass for the Ralph loop. This skill should be used at an epic boundary to prune dead code, duplicated logic, unused imports, redundant comments, and speculative error handling introduced during the loop's Green phase. It removes, never adds — the diff should end smaller than it started.

Test-first discipline for the Ralph loop. This skill should be used when fixing a bug or adding behavior: write a failing test that reproduces the target behavior before implementing, then make it pass with the smallest change. Stack-agnostic (BATS for shell, pytest for the SDK, jest where present). Integrates with the ralph-tester sub-agent at epic boundaries.

Ralph's per-loop execution contract — one task from fix_plan.md, the RALPH_STATUS exit block, epic-boundary QA deferral, and the dual-condition EXIT_SIGNAL gate. Invoke at the start of every Ralph loop so the response follows the contract the harness depends on.

Read-only Python (and sibling-interpreter) introspection without hitting the validate-command.sh `-c | -e` block. Use whenever you want to check an import, parse an AST, print a module version, or evaluate a tiny snippet — write the snippet to /tmp/snippet.py and run `python3 /tmp/snippet.py`. Or, for plain "does symbol X live in module Y" questions, prefer Read or Grep over a script entirely.

Execute one Ralph development loop iteration. Reads fix_plan.md, implements the first unchecked task, verifies, and commits.

Manually re-optimize fix_plan.md task ordering. Rebuilds the import graph, resolves vague tasks via exploration, and reorders for dependency order, module locality, and phase ordering. Shows what changed.

Research the codebase before implementing a task. Use Grep/Glob, semantic search, or a read-only explore subagent to find relevant files, patterns, existing code, and test files.