Commands

Global commands available in OpenCode after installing the harness. Each command runs a global skill or diagnostic script. This reference separates the shared contract from commands specific to the OpenCode adapter.

Names start with / when invoked in the OpenCode chat (for example, /prevc).

Commands shared by the PREVC contract

Session

Command	When to use	What it does
/harness-session-start	At the start of every session	Discovers root instructions, progress state, feature state, and startup path; runs the baseline if safe; declares the active task
/harness-clean-handoff	At the end of every session	Collects evidence, blockers, progress notes, handoff, and git status; returns a handoff report

Project

Command	When to use	What it does
/harness-init	First time in a project	Audits the harness layers without writing anything; produces a gap report with recommended fixes
/harness-bootstrap	Install the full harness package	Proposes the full package (AGENTS.md, feature_list.json, docs/harness, agent-os) and writes only after confirmation
/harness-standards	Before planning or implementing	Detects the stack and loads project and global standards; returns applicable risk, validation, permission, and skill profiles
/harness-spec	Medium or large feature	Creates or proposes an Agent OS spec in agent-os/specs/YYYY-MM-DD-HHMM-slug/

Work

Command	When to use	What it does
/prevc	Any significant work	Controls the lifecycle grouped into Prepare, Execute, Validate, and Confirm

Commands specific to OpenCode

Command	When to use	What it does
/plan	Plan significant work in this distribution	OpenCode's size-gated pipeline; ends at awaiting_plan_approval
/goal	Track a long-term objective	Optional OpenCode plugin; keeps objective memory separate from PREVC

Agent OS

Command	When to use	What it does
/plan-product	Start product documentation	Creates the base product documentation in agent-os/product/
/shape-spec	Manual escape hatch for heavy specs (the default pipeline is /plan)	Models a lean spec in planning mode; does not implement until approved
/discover-standards	Document code patterns	Discovers existing patterns and creates standards in agent-os/standards/
/index-standards	Rebuild the standards index	Scans agent-os/standards/ and regenerates the alphabetized index.yml
/inject-standards	Bring standards into context	Reads index.yml, suggests relevant standards, and injects content into the conversation, skill, or plan

Security and diagnostics

Command	When to use	What it does
/harness-security-scan	Audit the security surface	Scans OpenCode, Cursor, the repo harness, MCPs, and memory for supply-chain risks and secrets; summarizes PASS/WARN/FAIL without printing values
/harness-mcp-inventory	Inventory MCPs	Lists OpenCode/Cursor MCP servers, detects drift, and flags keys holding secrets without exposing values
/harness-context-budget	Audit context load	Reports the context budget of skills, plugins, commands, and MCPs; recommends lazy-loading
/harness-status	Check readiness	Prints git, PREVC, goal, handoff, context, and security status; points to the next best action
/harness-worktree-lifecycle	Parallel agents with worktrees	Reports worktree lifecycle safety (dirty, clean, stale) without deleting files

Quality and Refine

Command	When to use	What it does
/quality	Before a commit (or any time)	Runs the measured quality gate and risk router; unavailable metrics or missing/stale reports are named gaps, never passes
/refine	After Judge, or to record an operator note	Runs the read-only refiner and routes proposals by blast radius; it writes nothing

Interface and investigation commands

Local commands in command/, focused on interface design, investigation, and PRs in the OpenCode adapter.

Command	When to use	What it does
/init	Build UI with craft	Builds an interface following the interface-design skill
/audit	Check code against the design system	Checks design-system violations
/critique	Review build craft	Reviews the build as a design lead
/extract	Create system.md from code	Extracts design patterns into system.md
/status	View design system state	Shows direction, tokens, and patterns
/investigate	Discovery before planning	Asks focused questions and analyzes the code
/investigate-batch	Batch discovery	Asks up to five questions, then analyzes the code
/trim	Shrink a PR description	Reduces the current PR description while preserving essentials

Step-by-step for the main commands

1. Work on something significant — /prevc

/prevc prepare <objective>

• Runs the global prevc-workflow skill, the single lifecycle controller for significant work.
• prepare <objective> performs discovery, classifies risk, and creates a plan, stopping at awaiting_plan_approval in the conversation.
• Never invokes /goal or the goal tools; lifecycle labels stay in the conversation and in the existing harness files.
• Routes low-risk routine operations (session-start, status checks, explicit dependency installation, formatting, narrow test repairs) through a fast path with objective verification, without the PREVC ceremony.

/prevc run

• Accepts explicit operator approval in plain language or /prevc run; requires no goal ID or /goal confirm.
• An operator instruction to execute a named plan or spec end to end is itself the authorization for an autonomous run: no separate /prevc run, no per-task approval; the run stops once at awaiting_confirmation for the whole spec.
• Automates Review → Execute → Validate → Judge within the approved scope and the declared budgets (files, permissions, verification, retry, tools, duration).
• On low-risk work, allows a limited repair pass for a newly exposed verification prerequisite (a missing test, lint, typecheck, or dev-dependency build) and then runs full validation again.
• Preserves the low, medium, high, and untrusted risk paths: medium+ requires context/security evidence or a documented skip; high/untrusted requires explicit risk acknowledgment before executing.
• Never commits, pushes, deploys, switches branches, or runs remote Git operations automatically.

2. Audit a project — /harness-init

/harness-init

• Runs the global harness-initializer skill.
• Detects the repo root, root instructions, progress state, feature state, startup path, Judge/evaluation layer, harness skills, security posture, hook/plugin posture, MCP/plugin hygiene, context budget, goal memory hygiene, worktrees, agent permissions, and parity with Cursor.
• Produces a non-mutating gap report with recommended fixes.
• Does not create or edit files unless the user approves a specific fix.
• Does not assume that init.ps1, feature_list.json, or .specs/project/STATE.md exist.

3. Install the full harness — /harness-bootstrap

/harness-bootstrap

• Runs the global harness-bootstrap skill.
• Always proposes the full package: AGENTS.md, feature_list.json, docs/harness (progress, handoff, sprint-contract, security-policy, hook-policy, context-budget, eval-contract, agent-permission-matrix, status), docs/ARCHITECTURE.md, docs/PRODUCT.md, docs/RELIABILITY.md, agent-os/judges/project-judge.md, agent-os/standards, and spec templates including evals.md.
• First audits the current repo, detects the stack with harness-stack-router, and discovers the startup path.
• Shows the exact files to create, merge, or skip and asks for explicit confirmation before writing anything.
• Never assumes init.ps1, feature_list.json, .specs/project/STATE.md, or repo paths.

4. Start the session — /harness-session-start

/harness-session-start

• Runs the global harness-session-start skill as a direct routine operation, not as PREVC work.
• Never invokes /goal or the goal tools.
• Reads root instructions, if they exist.
• Locates the progress state in .specs/project/STATE.md, docs/harness/progress.md, or agent-progress.md.
• Locates the feature state in feature_list.json or .specs/features/*/tasks.md.
• Reads docs/harness/session-handoff.md, if present.
• Discovers the startup command with harness-startup-path and runs it if safe.
• Runs git log --oneline -3 in git repos.
• Applies WIP=1 when there is durable state.
• Ends by declaring: Active task, AC (observable criteria). If startup fails, it records the exact error in the progress file and stops.

5. Close the session — /harness-clean-handoff

/harness-clean-handoff

• Runs the global harness-clean-handoff skill as a PREVC handoff subroutine.
• Collects evidence, blockers, progress notes, a session handoff, passing verification results, and git status.
• Only PREVC writes lifecycle transitions after the operator's decision.
• When invoked directly, it returns only a handoff report: it does not mark feature state, infer confirmation, or assert completion.

6. Load standards and skills — /harness-standards

/harness-standards

• Runs harness-stack-router and then harness-standards-router.
• Detects the stack from the files before guessing.
• Loads project standards in agent-os/standards first, then the global standards templates.
• Returns a risk profile, validation profile, agent permission profile, security standards, minimum checks, and skills applicable to the current task.
• Does not implement code: it produces only the standards and skills routing report.

7. Create a spec — /harness-spec

/harness-spec <objective>

• Runs harness-agent-os-specs for the requested objective.
• Creates or proposes agent-os/specs/YYYY-MM-DD-HHMM-slug/ with spec.md, plan.md, tasks.md, verification.md, and decisions.md from the global templates.
• Replaces placeholders with concrete requirements, acceptance criteria, verification commands, and decisions.
• Asks before writing if the target repo does not yet have a harness.

Quick commands (cheat sheet)

# Unknown project: audit without writing
/harness-init

# Project without harness: propose full install
/harness-bootstrap

# Start of a normal session
/harness-session-start

# Prepare significant work
/prevc prepare <objective>

# Approve the prepared plan (approved ad-hoc plan): explicit approval in plain
# language; to run the approved plan, use
/prevc run

# Autonomous run: an instruction to execute a named plan/spec end to end is
# itself the authorization — no separate /prevc run, no per-task approval

# Pause or close with durable context
/harness-clean-handoff

# Quality and refine
/quality              # measured quality gate + risk router (--mode full before a commit)
/refine               # Refine phase after Judge (or --note "<what I fixed by hand>")

# Global diagnostics
/harness-status
/harness-security-scan
/harness-context-budget
/harness-mcp-inventory
/harness-worktree-lifecycle