orobsonn
UserHarness de entrega determinística para Claude Code: toda tarefa passa por triagem → spec → plano → execução → review adversarial → harvest. Orquestrador Sonnet barato delega aos modelos premium só nos gates de fronteira. Hooks de entrada determinísticos, memória nativa no repo e rotinas headless na nuvem.
Categories
Indexed Skills (50)
oc-authoring-rules
Meta-skill for authoring and editing folder-scoped rules in the native OpenCode tree — nested AGENTS.md (or an existing folder CLAUDE.md router), root AGENTS.md/CLAUDE.md, never a Claude-style paths-gated rule file. Use when creating a new folder/subsystem rule, adding a Convention/Pattern/Gotcha to an existing one, or deciding whether something is a rule, a router-table row, or a one-off. Routes harness-wide rules to proposing-improvements via kaizen.md.
oc-brainstorming
You MUST load and follow this before ANY creative/dev work in LIGHT or FULL mode — before writing a spec, dispatching the planner, or any implementation. It elicits the operator's non-codifiable decisions one question at a time, proposes approaches, and HARD-GATES on an approved design. Skipping it makes the model FABRICATE product decisions the operator never made — the single biggest source of silently-wrong output.
oc-canonical-critical-classes
Reference loaded by the adversary (to HUNT) and compliance (to know which classes, when a diff puts one in play, REQUIRE a test). Carries the canonical taxonomy of critical failure classes — the known killers from the operator's systems-primitives model — and the irreversibility-first criticality ranking. Role-neutral: HOW each role acts lives in that agent's prompt.
oc-committing-changes
Analyzes changes, creates a semantic commit (Conventional Commits) and pushes. For a standalone commit OUTSIDE the full delivery loop — when you just want to commit a change without PR/review/docs. Conforms to AGENTS.md §4 git rules. For full delivery use the shipper agent.
oc-configuring-model-routing
Interactive skill to reconfigure harness.routing.json via dual-safe presets or custom slots. Deterministic apply rewrites routing + all agent frontmatter models + AGENTS.md §8 + opencode.json model/small_model. Validates before any write. Never disables dual without explicit operator override. Never invents roles or commits secrets.
oc-distilling-learnings
Reads the transient findings buffer, applies the durability test, then routes each durable insight by blast-radius to the right NATIVE destination — project pattern → native project-root MEMORY.md (durable index, one note = one concept), law of one folder → that folder's nested AGENTS.md (or its existing CLAUDE.md router) + root router row, global convention → kaizen.md proposal. No learnings store is written. Use when the harvester runs after recording-findings completes; never during active implementation.
oc-grill
Pre-implementation interview — turns an idea that is still a shadow into a written PRD at docs/prd/<slug>.md through relentless, one-at-a-time, consequence-framed questioning, then hands off to creating-issues. Runs BEFORE the delivery pipeline (before triage, before any plan) and is LOCAL/interactive only — it refuses in any headless or cron session. Suggest it when the operator arrives with a big or vague idea and there is no written spec yet. It is NOT the session entry gate (triaging-requests owns that) and its output is NOT a source of truth.
oc-importing-claude-memory
One-shot, OPERATOR-INVOKED migration that imports a repo's legacy Claude Code native memory (~/.claude/projects/<slug>/memory/) into this OpenCode harness's project-root MEMORY.md. The SINGLE sanctioned exception to the never-read-~/.claude rule — read-only, scoped to projects/<slug>/memory/ only, NEVER auto-run in the delivery loop, NEVER invoked by surveying-codebase or build. Run it by hand once when onboarding a repo that previously used Claude Code. Run the LLM-aided classification on a STRONG model (e.g. openai/gpt-5.5) — one-shot + security-critical, so the cheap open-source default is the wrong tradeoff here.
oc-orchestrating-delivery
Drives the LIGHT and FULL delivery loop — spec, plan, per-task executor/compliance/adversary/sniper cycle, final review, demo, and harvest. Dispatches one subagent per role via the task tool; never writes code itself. Invoked by triaging-requests for LIGHT/FULL; QUICK runs inline and never reaches this skill.
oc-proposing-deepening
Architectural retrofit analysis for an EXISTING codebase — walks the code, finds DIRECTORIES of shallow modules, and proposes at most 5 ranked deepening candidates to docs/architecture/deepening-candidates.md. Runs in the sole `build` primary agent and never writes source, opens an issue, or dispatches delivery. LOCAL/interactive only — it refuses in any headless or cron session. A candidate is a model deduction, never a locked decision, and the issue derived from it is NEVER created harness:ready: a blind restructure of working code is delivered locally, never auto-merged. Use when a legacy/existing project needs its architecture retrofitted toward deep modules, or when asked to find where the codebase became hard to change.
oc-proposing-improvements
Scans the transient findings buffer for systemic patterns that indicate a harness agent or skill could be improved, and appends proposals (never auto-applies them) to the project-root kaizen.md. Because the findings buffer is deleted at harvest end, kaizen.md is the durable cross-run signal — check it for precedent. Use when the harvester runs after distilling-learnings completes; never during active implementation.
oc-recording-findings
Consolidates executor, compliance, adversary, and sniper finding blocks from a completed feature run into the transient findings buffer at the project root. The findings buffer is ephemeral — the input to distilling-learnings, deleted by the harvester at the end of the run; durable audit lives in git. Use when the harvester runs after final review — never during active implementation.
oc-releasing-versions
Versioned release. If the project has release-please configured, the flow IS release-please — conventional commits on main, the action opens the release PR itself, and merging that PR creates the tag + GitHub Release. Without release-please, it falls back to the manual release-PR flow. Use when changes merged into main are ready to become a version. Conforms to AGENTS.md §4 git rules.
oc-surveying-codebase
Surveys an onboarded project's codebase for durable, reusable knowledge (existing helpers to reuse, implicit conventions, architectural patterns, anti-patterns/gotchas), applies the durability test, and routes each insight by blast-radius to the right NATIVE destination — the same routing as distilling-learnings, but sourced from the code itself, not a run's findings buffer. Use when a legacy/existing project enters the harness with empty memory (no/empty MEMORY.md + unfilled nested AGENTS.md), or when build's Phase 0 finds the memory cold; one-time / on-demand, re-run when the codebase changes substantially. No learnings store is written.
creating-plans
INTERNAL to the planner agent — NOT a main-loop skill. Do NOT invoke this directly from the orchestrator or main loop; instead dispatch the `planner` agent (always Opus), which runs this skill in isolation. Describes how the planner decomposes an approved spec/PRD into a validated execution-plan.json (atomic tasks, locked tests, severity tiers, adversarial flags, scope_paths) consumed by orchestrating-delivery.
oc-creating-plans
INTERNAL to the planner agent — NOT a main-loop skill. Do NOT invoke this directly from the orchestrator or main loop; instead dispatch the `planner` agent (always the planner model from routing), which runs this skill in isolation. Describes how the planner decomposes an approved spec/PRD into a validated execution-plan.json (atomic tasks, locked tests, severity tiers, adversarial flags, scope_paths) consumed by orchestrating-delivery.
authoring-rules
Meta-skill para criar e editar rules do Claude Code (.claude/rules/*.md). Use quando for criar uma nova rule no projeto, adicionar pattern/gotcha/convention a uma rule existente, ou revisar rules.
canonical-critical-classes
Load when ATTACKING (adversary) or VERIFYING (compliance) an implementation. Carries the canonical taxonomy of critical failure classes — the known killers from the operator's systems-primitives model — plus the irreversibility-first criticality ranking. Use it to hunt the bug that makes the system UNVIABLE, never to fill a finding quota. Role-neutral: HOW each role acts lives in that agent's prompt.
committing-changes
Analisa mudancas, cria commit semantico (Conventional Commits) e faz push. Para commit avulso fora do pipeline /ship — quando voce so quer commitar uma mudanca sem PR/review/docs.
connecting-orca
Takes a project that already has the harness vendored and wires it to Orca end-to-end until autonomous delivery actually runs: diagnoses whether THIS session can reach the VPS at all (the failures that read as 'no access' but are not), confirms the runtime + pairing, registers the repo, installs the implementation queue (project JSON + selector cron) and the PR-review/conditional-merge automation, then proves it with one canary issue. Use after initializing-projects/updating-harness, or whenever an agent session concludes it cannot operate the VPS.
creating-issues
Authors well-formed harness issue(s) — a single issue or a dependency-ordered roadmap — from the operator's conversation, from a PRD written by grill (docs/prd/<slug>.md), or from a deepening candidate written by proposing-deepening (docs/architecture/deepening-candidates.md). Applies the small-delivery-unit sizing rule, coaches VERIFIABLE acceptance criteria (the locked_tests derive from them), keeps model assumptions attackable and blocks on open questions, wires harness-deps for order, creates issues harness:ready EXCEPT retrofit issues (local-only, unlabelled, never auto-merged), and runs the DAG lint. Use when the operator wants to create an issue or organize a roadmap.
distilling-learnings
Reads the transient findings.md, applies the durability test, then routes each durable insight by blast-radius to the right repo-committed destination — project pattern → repo memory (.claude/memory/<name>.md + .claude/memory/MEMORY.md index), law of one folder → that folder's nested CLAUDE.md + root router row, global convention → .claude/kaizen.md proposal. No learnings.md is written. Use when the harvester runs after recording-findings completes; never during active implementation.
grill
Pre-implementation interview — turns an idea that is still a shadow into a written PRD at docs/prd/<slug>.md through relentless, one-at-a-time, consequence-framed questioning, then hands off to creating-issues. Runs BEFORE the delivery pipeline (before triage, before any plan) and is LOCAL/interactive only — it refuses in any headless or cron session. Suggest it when the operator arrives with a big or vague idea and there is no written spec yet. It is NOT the session entry gate (triaging-requests owns that) and its output is NOT a source of truth.
initializing-projects
Installs (vendors) the Claude Harness core into a project's .claude/ so the pipeline works locally and in cloud routines. Resolves the source portably (local path or git URL), copies agents/skills/rules/CLAUDE.md/settings.json, seeds repo-relative memory, stamps .harness-version, and merges idempotently so re-running updates without clobbering project content. Use to onboard a project or to update its vendored harness.
measuring-cost
Use at the end of a delivery (invoked by the harvester) to report the API-equivalent cost of the session plus the weekly Claude Code consumption trend. Wraps the `ccusage` tool over the local transcript JSONL; fail-soft when ccusage is unreachable (offline / cloud headless).
orchestrating-delivery
Conducts the LIGHT and FULL delivery loop of the Claude Harness — runs spec, plan, the per-task executor/compliance/adversary/sniper/gates cycle, final dual review, demo, and harvest. Dispatches a fresh subagent per role and curates layered ICM context for each; never writes code itself. Invoked by triaging-requests for LIGHT/FULL (QUICK runs inline and never reaches this skill).
proposing-deepening
Architectural retrofit analysis for an EXISTING codebase — walks the code, finds DIRECTORIES of shallow modules, and proposes at most 5 ranked deepening candidates to docs/architecture/deepening-candidates.md. READ-ONLY on source and PROPOSE-ONLY: it never refactors, never opens an issue, never dispatches a delivery agent. LOCAL/interactive only — it refuses in any headless or cron session. A candidate is a model deduction, never a locked decision, and the issue derived from it is created WITHOUT harness:ready: a blind restructure of working code is delivered locally, never auto-merged. Use when a legacy/existing project needs its architecture retrofitted toward deep modules, or when asked to find where the codebase became hard to change.
proposing-improvements
Scans the transient findings.md for systemic patterns that indicate an agent or skill in the harness could be improved, and appends proposals (never auto-applies them) to kaizen.md. Because findings.md is deleted at harvest end, kaizen.md is the durable cross-run signal — check it for precedent. Use when the harvester runs after distilling-learnings completes.
recording-findings
Consolidates executor, compliance, adversary, and sniper finding blocks from a completed feature run into the transient findings.md buffer at the project root. findings.md is ephemeral — the input to distilling-learnings, deleted by the harvester at the end of the run; durable audit lives in git. Use when the harvester agent runs after final dual review — never during active implementation.
releasing-versions
Release versionada. Se o projeto tem release-please configurado, o fluxo E release-please — commits convencionais na main, a action abre sozinha o PR de release, e mergear esse PR gera tag + GitHub Release. Sem release-please, cai no fallback manual de PR de release. Use quando ha mudancas em main prontas pra virar versao.
reviewing-pull-requests
Machine-only skill that runs a fresh-eyes review session over a PR diff — adversary, compliance (via the diff-adapter), and security eyes. Eyes only, Claude tier, no write-hand. Emits raw eye-outputs to the engine-controlled stateDir; Node (cron-review) derives the canonical merge verdict from them and orchestrates cross-family eligibility.
surveying-codebase
Surveys an onboarded project's codebase for durable, reusable knowledge (existing helpers to reuse, implicit conventions, architectural patterns, anti-patterns/gotchas), applies the durability test, and routes each insight by blast-radius to the right repo-committed destination — the same routing as distilling-learnings, but sourced from the code itself, not a run's findings.md. Use when a legacy/existing project enters the harness with empty memory (.claude/memory/MEMORY.md + unfilled nested CLAUDE.md), via /surveying-codebase or when orchestrating-delivery Phase 0 finds the memory cold; one-time / on-demand, re-run when the codebase changes substantially. No learnings.md is written.
triaging-requests
Classifies the operator's request into QUICK (fix or craft), LIGHT, or FULL delivery mode — or no ceremony at all — and dispatches accordingly. QUICK-craft is the fast lane for self-contained visual artifacts (page/quiz/landing) routed to an artisan skill. Runs on the first interaction of every session before any implementation begins.
updating-harness
Use to install or update the Claude Harness in the CURRENT project from its source repo — the source URL is baked in, the operator never pastes it. Detects install-vs-update, pins to the latest GitHub release tag, reports what changed, and re-vendors via vendor-core without clobbering project memory/kaizen/settings. Call it in any repo to onboard or sync the harness after a new release.
harness-planning
Use after a design is approved and before a multi-step implementation.
harness-releasing-versions
Use when a completed milestone is ready for a release-please release PR, tag, or GitHub Release.
oc-creating-issues
Cria uma issue harness-ready ou um roadmap de issues pequenas — a partir da conversa, de um PRD escrito pelo grill (docs/prd/<slug>.md) ou de um candidato do proposing-deepening (docs/architecture/deepening-candidates.md) — com campos verificáveis e submissão segura pelo GitHub CLI. Issue vinda de candidato a aprofundamento é entrega local: nunca harness:ready, nunca submetida pelo submit-issue.mjs. Use quando o operador pedir para criar issue ou organizar roadmap.
harness-brainstorming
Use before proposing or implementing a non-trivial product or architecture change.
harness-delivery
Use to execute an approved delivery plan with TDD, scoped subagents, and verification.
harness-discovery
Use for product discovery, requirements interviews, and pre-build challenge sessions.
harness-improvement
Use when proposing a measured improvement or a deeper follow-up to the harness or product.
harness-issues
Use when converting a validated request or finding into a GitHub issue.
harness-learning
Use when recording verified reusable findings or importing durable project knowledge.
harness-lifecycle
Use when installing, updating, or validating the project-vendored Codex harness.
harness-orca
Use when an operator asks to connect the delivery harness to Orca automation.
harness-review
Use for code review, plan review, adversarial review, compliance review, or security review.
harness-routing
Use when selecting models, reasoning effort, or cost controls for a delivery task.
harness-rules
Use when writing delivery rules, assessing critical failure classes, or changing policy.
harness-survey
Use when mapping an unfamiliar codebase, runtime, or harness surface before changing it.
harness-triage
Use at the start of a top-level request to select delivery ceremony and model cost tier.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.