| name | harness-steering |
| description | Owns the persistent behavior surfaces of the Claude Code harness and the placement law that picks one: the CLAUDE.md hierarchy, imports, path-scoped rules, the memory content law and scored audit, auto memory, settings scopes and precedence, permission evaluation, sandbox and auto-mode guardrails, model and effort routing (aliases, opusplan, advisor), skill listing budget, statusline, output styles, keybindings, plugin marketplaces, LSP plugins, env, plus the headless lanes — print and bare mode, session continuation, background sessions, the SDK. Use when deciding where an instruction lives (memory file, rule, skill, hook, setting, or subagent), auditing or slimming a CLAUDE.md, tuning model, effort, permission, or sandbox defaults, diagnosing what loaded into context, or scripting noninteractive calls. Mechanical settings.json edits belong to update-config; hook construction to hooks-builder; skill bundle authoring to skill-writer; work placement across subagents, teams, and workflows to agent-dispatch. |
[HARNESS_STEERING]
One instruction, one owner: every durable behavior rides exactly one steering surface, chosen by load timing, scope, authority, and context cost. Memory shapes behavior and hooks and permissions enforce it — an instruction that must never be violated is not prose.
Every Claude Code fact — settings keys, limits, defaults, command and flag surfaces, official values — resolves live through the claudeCodeDocs MCP, never from recall: mcp__claudeCodeDocs__search_claude_code_docs for concept queries, mcp__claudeCodeDocs__query_docs_filesystem_claude_code_docs for exact pages and key spellings under /en/*.mdx (rg, cat, tree against a virtual docs filesystem), and mcp__claudeCodeDocs__submit_feedback for doc defects found along the way.
[01]-[ROUTING]
- [01]-MEMORY: memory hierarchy, @path imports, path-scoped rules, auto memory, the memory content law and scored audit
- [02]-SETTINGS: scope precedence, permission evaluation, model and effort routing, skill listing budget, interface rows
- [03]-GUARDRAILS: sandbox schema, auto-mode classifier tuning, managed lockdown rows
- [04]-HEADLESS: print mode, structured output, session continuation, background sessions, worktree isolation, SDK boundary
- [05]-PLUGINS: marketplace registration, scope, enablement rows, cache staleness,
.lsp.json anatomy, LSP lifecycle, traps
[02]-[STEERING]
| [INDEX] | [SURFACE] | [LOADS] | [PLACE_WHEN] |
|---|
| [01] | Memory file | Every session, always in context | An always-true convention for the machine, project, or operator |
| [02] | Rule | Launch, or lazily on paths match | A convention true only for a subtree or file class |
| [03] | Skill | On task-shape selection | A reusable procedure or deliverable doctrine — skill-writer owns it |
| [04] | Subagent | On delegation | Work whose transcript belongs out of context — agent-dispatch owns it |
| [05] | Hook | On lifecycle event, deterministic | A gate that must fire regardless of model judgment — hooks-builder owns it |
| [06] | Setting | Startup and hot-reload | Enforcement, defaults, and budgets: permissions, model, effort, env |
| [07] | Output style | System prompt, rebuilt on /clear | A register change for the whole session, not a task procedure |
[03]-[MEMORY]
Memory loads broad to specific — managed policy, user, project, local — and every level is context, never enforcement. Load algorithm, @path imports with the four-hop recursion cap, .claude/rules/ with paths scoping, CLAUDE.local.md, auto memory limits, exclusion globs, the content law deciding what a memory file may carry, the recall mechanics that decide which one fires, and the scored audit are memory.md.
[04]-[SETTINGS]
Precedence runs managed, command line, local, project, user — scalars replace across scopes while arrays concatenate and deduplicate — and permission rules evaluate deny, ask, allow. Model aliases, effort levels and their per-model floors, the skill listing budget, interface rows (keybindings, statusline, output styles), and the high-leverage rows worth setting deliberately are settings.md. Editing mechanics for settings.json files belong to update-config; this skill decides which row, which scope, and which value.
[05]-[GUARDRAILS]
Guardrails bind beneath permissions: sandbox.* isolates Bash at the OS level, autoMode.* tunes the classifier that adjudicates commands no rule decided, and managed lockdown rows freeze what a fleet's users can re-open. Sandbox schema, classifier rule grammar, and the lockdown census are guardrails.md.
[06]-[HEADLESS]
claude -p runs the full harness noninteractively; --bare strips discovery for deterministic scripted calls; --bg detaches whole sessions; --agents and --agent define and mount agents at launch. Lane selection, structured output, session continuation and persistence, worktree isolation for background work, cache economics, and the SDK boundary are headless.md.
[07]-[PLUGINS]
Marketplace registration state is global per user and marketplace names are unique — registration and enablement are decoupled surfaces, and a cross-repo plugin set lives once at user scope while projects add rows only for repo-specific plugins. Plugin cache keys on resolved version, making unbumped master edits invisible until an explicit update. Registration mechanics, cache law, .lsp.json anatomy, and the LSP trap census are plugins.md.
[08]-[DIAGNOSTICS]
/context shows token allocation across system prompt, tools, memory, skills, and history — the first stop when context is mysteriously full./memory lists every loaded memory file and the auto memory directory — the first stop when an instruction is mysteriously absent.InstructionsLoaded logs which instruction files loaded, when, and why — the deterministic answer for lazy-loaded rules and subdirectory memory.ConfigChange fires on every settings reload./doctor validates installation, settings parse, hook health; claude --safe-mode starts with customization disabled to bisect a broken config./config surfaces the interactive toggles and sets any key as /config key=value; a behavior that survives --safe-mode is upstream, not local.
[09]-[COMPOSITION]
- update-config performs the mechanical
settings.json and settings.local.json edits this skill's decisions call for.
- hooks-builder constructs the hook a steering decision lands on; this skill only rules that a hook is the right surface.
- skill-writer owns bundle anatomy, triggers, and listing economics beyond the budget rows named here.
- agent-dispatch owns where work executes; this skill owns where instructions and defaults persist.
