yellow-core

Reference for agent-native architecture principles: action parity, context parity, shared workspace, primitives over workflows, and dynamic context injection. Use when authoring agent integrations or designing tool surfaces.

Audit checklist for evaluating an existing codebase against agent-native architecture principles to identify parity gaps. Use when auditing for agent-native readiness, before adding a new agent capability, or when triaging agent-behavior regressions.

Reference guide for iterative brainstorm dialogues — question techniques, YAGNI, approach exploration, and research escalation. Use when running /flow:brainstorm or authoring brainstorm-style agents.

Audit, refresh, and consolidate `docs/solutions/` to keep the institutional knowledge catalog from rotting. Use when a `docs/solutions/` sweep is needed or after `knowledge-compounder` flags an older entry as superseded.

Expert guidance for creating Claude Code skills and agents. Use when working with SKILL.md files, authoring new skills, creating slash commands, or designing agent workflows.

Find root causes systematically before fixing — trace the causal chain and form hypotheses with predictions for uncertain links. Use when debugging errors, test failures, stack traces, or when stuck after failed fix attempts. For delegating a stuck investigation to an independent Codex session, use /codex:rescue.

Git worktree management for isolated parallel development. Use when reviewing PRs in isolation, working on multiple features simultaneously, or when workflows offer worktree option.

Generate 3 grounded approaches to a soft problem, then hand the chosen one off to `brainstorm-orchestrator`. Use when the problem statement is still vague or `/flow:brainstorm` narrows too quickly.

Defines the yellow-plugins.local.md per-project config file schema. Use when authoring commands that should accept per-project overrides for review pipeline behavior, reviewer set narrowing, depth controls, or focus-area filtering.

Canonical MCP server health classification (OFFLINE / DEGRADED / HEALTHY plus PRESENT (untested) refinement) for `/<plugin>:status` commands. Use when authoring a status command that needs to distinguish 'MCP never started' from 'MCP running but upstream API degraded'.

Shared patterns for ruvector recall/remember and morph discovery integration (the ruvector protocol constants are a replica of yellow-ruvector's canonical memory-query skill). Use when authoring commands or agents that should leverage institutional memory and advanced editing tools.

Recall-Before-Act pattern for ruvector — query past learnings via hooks_recall at workflow start. Use when authoring commands or agents that should seed a workflow with relevant institutional memory before executing.

Tiered-Remember-After-Act pattern for ruvector — record learnings via hooks_remember at workflow completion with signal-strength tiers (Auto/Prompted/Skip). Use when authoring commands that should compound institutional memory from workflow outcomes.

Morph discovery + fallback pattern — discover edit_file and warpgrep at runtime via ToolSearch, prefer them for large edits and intent-based search, silently fall back to Edit/Grep when yellow-morph is missing. Use when authoring agents that edit files or run semantic searches.

Multi-host plugin credential and config reference for yellow-plugins. Use when setting up a new workstation, configuring CI/devcontainer, replicating credentials across a fleet, or wiring a secrets manager. Documents the canonical shell env var names for every credential-bearing plugin.

Run a metric-driven optimization pass: parallel candidate variants scored against an LLM-as-judge analytic rubric. Use when comparing approaches against a measurable goal — anywhere 'better' can be expressed as a per-criterion rubric. For session-level plan-adherence and scope-drift review, use /flow:review instead.

Show a per-file checkbox progress dashboard of plans/ (open) and plans/complete/ (archived) plan documents. Use when reviewing which plans are ready to archive or checking work-in-flight at a glance.

Canonical prompt-injection hardening block for agents that analyze untrusted content (source code, CI logs, workflow files). Use when authoring an agent that reads untrusted input — copy the appropriate variant verbatim into the agent body; only pure output-only agents (no file reading, no untrusted input) are exempt.

Write a validated session-handoff note at plans/handoff/<YYYY-MM-DD>-<slug>.md, or resume from an explicitly named one after a read-only preflight. Use when the user says "create a handoff", "save session state", "handoff before compact", "pick up where we left off", or names a plans/handoff/ file to resume. Shell code measures repository, worktree, HEAD, dirty fingerprint and source session into the note; the narrative is model-authored reference data. Not the shell halt pattern — /flow:pick-next-shell halts by design after writing its expansion artifact and needs no handoff; use this for free-form session state only.

Search prior sessions across Claude Code, Devin, and Codex for the same problem or topic. Use when the current session cannot see prior context — after agent compaction, session boundary, or when reconstructing a multi-session decision trail. This searches raw prior-session transcripts; for distilled learnings, use ruvector recall (/ruvector:memory).

Enforce the stacked-PR provider invariants before any provider-changing action — exactly one enabled, managed scopes fail closed, no direct settings edits, no silent fallback. Use when validating a proposed provider switch or auditing current provider state.

Resolve which stacked-PR provider is active and route provider-specific work to it, or refuse when the state is ambiguous. Use when a workflow needs to know whether Graphite or GitHub native stacks is in effect before running a stack operation.