<< All versions

Skill v1.0.1

currentAutomated scan100/100
maicongambini/omp-harness-guide/harness-bootstrap

+1 new

──Details
PublishedOctober 27, 2026 at 08:18 PM
Content Hashsha256:4763688361cdb376...
Git SHAe535151c6050
Bump Typepatch
Compare with v1.0.0
──Files
Files (1 file, 8.3 KB)
SKILL.md8.3 KBactive
SKILL.md · 120 lines · 8.3 KB

name: harness-bootstrap description: "Installs or merges the complete Agent OS harness into a repo. Use when adopting the full harness suite after approval. Do NOT use for read-only harness audits (use harness-initializer)." metadata: version: "1.0.1"


Harness Bootstrap

Install a full harness into any project, but only after explicit user approval (or pre-authorized when invoked via novo-projeto).

Portable Onboarding Integration

Full bootstrap installs the entire package while preserving project-specific files, then delegates environment detection, baseline setup, and service drafting to the shared onboarding CLI:

bash
node "<harnessHome>/scripts/harness-project-onboarding.mjs" --root "<path-to-repo>" --write [--json]

Path & Environment Resolution

  • <harnessHome> dynamically resolves to the installed harness home (~/.omp/agent in Oh My Pi, or canonical ~/.config/opencode in OpenCode).
  • Commands must be absolute and quoted to safely support paths with spaces. Never hardcode machine-specific usernames.
  • Dry-run is the default when --write is omitted.

Automated Onboarding Execution

The shared onboarding CLI automates four essential artifacts without inventing data:

  1. Project Profile & Cérebro (`agent-os/project-profile.json`): Detects test components, containerization, CI, and delivery config via harness-project-profile.mjs, publishing automatically to the Obsidian project cérebro (--publish-cerebro).
  2. Bounded Eval Mining (`evals/dataset.json`): Mines bugfix commits and corresponding test pairs from git history via harness-eval-runner.mjs --mine --write. If the repository is fresh or lacks bugfix commits, it leaves an explicit empty dataset/gap; it never fabricates synthetic tests or fake passing suites.
  3. Evidence-Backed Security Contract (`agent-os/security-contract.json`): Scans for routes and policy files to draft tenant isolation, authz, and IDOR targets. Entries specify { "path": "<file>", "command"?: "<runner>", "requires"?: [{ "host": "127.0.0.1", "port": 5432 }], "envFile"?: "<path>", "requireNoSkips"?: true }. If no contract or security tests exist, it records an explicit gap with the status returned by the tool (status: "unavailable" when the contract file is absent, or status: "not_applicable" when no executable tests are declared); it never invents placeholder tests.
  4. Quality Thresholds (`agent-os/quality-thresholds.json`): Configures the repository in Phase A, where all metrics are set to observe except regression_suite (which is blocking). Existing blocking settings, tighter ratchets, and custom operator fields survive repeat onboarding unchanged; onboarding never downgrades an existing blocker.

Full Package

Propose these files:

  • AGENTS.md
  • feature_list.json
  • docs/harness/progress.md
  • docs/harness/session-handoff.md
  • docs/harness/sprint-contract.md
  • docs/harness/matt-pocock-tracker.md
  • docs/harness/refine-log.md
  • docs/harness/findings/README.md
  • docs/harness/licoes-aprendidas.md
  • docs/ARCHITECTURE.md
  • docs/PRODUCT.md
  • docs/RELIABILITY.md
  • agent-os/judges/project-judge.md
  • agent-os/standards/*
  • agent-os/specs/_template/*
  • agent-os/quality-thresholds.json
  • agent-os/learned-rules.json
  • agent-os/prevc-autonomy.json
  • agent-os/skills-proposals/
  • agent-os/dependency-ledger.json
  • agent-os/redteam-allowlist.json
  • agent-os/quality-decisions.md
  • docs/review.md
  • docs/pull-request.md
  • .github/pull_request_template.md
  • docs/review-area.md

Install mapping — quality and continual-harness files

Global templateInstalls to
templates/agent-os/quality-thresholds.jsonagent-os/quality-thresholds.json
templates/agent-os/quality-decisions.mdagent-os/quality-decisions.md
templates/docs/review.mddocs/review.md
templates/docs/pull-request.mddocs/pull-request.md
templates/agent-os/learned-rules.jsonagent-os/learned-rules.json
templates/agent-os/prevc-autonomy.jsonagent-os/prevc-autonomy.json
templates/agent-os/skills-proposals/agent-os/skills-proposals/
templates/agent-os/dependency-ledger.jsonagent-os/dependency-ledger.json
templates/agent-os/redteam-allowlist.jsonagent-os/redteam-allowlist.json
templates/docs/harness/refine-log.mddocs/harness/refine-log.md
templates/docs/harness/findings/README.mddocs/harness/findings/README.md
templates/docs/harness/licoes-aprendidas.mddocs/harness/licoes-aprendidas.md
templates/github/pull_request_template.md.github/pull_request_template.md (somente GitHub; para projeto com tracker Azure não crie .azuredevops/ — registre como pendência futura)
templates/docs/review-area.mddocs/review.md (só se não existir; nunca sobrescrever arquivo existente)

The thresholds template now carries learned_rules (the loop's governing knobs), so the first row installs the knob block with the thresholds. The ledger itself is the separate row above. Bootstrap also installs the project-local dependency ledger as a fresh starting point. These ledgers contain no history, approvals, or rules from another project; they are owned by the target project after install. Bootstrap also provisions agent-os/skills-proposals/ for project-local skill proposals mined from failure traces.

Install the thresholds file at Phase A, with every metric observe except regression_suite. The memory_available metric is retired. Never carry another project's numbers into this one. The global template is a reusable source only and owns no project tree. After bootstrap, recommend harness-project-calibration: calibration is the project rollout boundary and derives the repository's own thresholds, high-risk paths, and tool-availability list from its layout and history, then writes that project's agent-os/quality-thresholds.json.

The redteam_findings and dependency_cooldown rows remain observe-only with null baselines until calibration replaces them with measurements from the target repository.

The bypass_findings metric is installed in observe with a null command by default (like skill_lint). The skill_lint metric is observe with a null command by default; projects that ship skills set commands.skill_lint to: node "<home>/.omp/agent/scripts/harness-skill-lint.mjs" --root . --json (the linter resolves agent/skills or .omp/skills; --global lints ~/.omp/agent/skills); red-team allowlist entries are content-hash keyed, require reason/allowedBy/expires, and are an operator decision, never an agent's.

Workflow

  1. Detect repo root and existing files.
  2. Run harness-stack-router to identify stack signals.
  3. Run harness-startup-path to find startup/verification commands.
  4. Compare target files with global templates in ~/.config/opencode/templates.
  5. Print a change plan: create, merge, skip, or preserve.
  6. Authorization check:
  • If invoked via novo-projeto: explicit authorization to create and onboard the project is already granted; proceed directly.
  • If invoked directly: Ask: Apply this full harness to the current repo? yes/no. If no, stop with the plan only.
  1. Create missing directories and write template files from templates/ into the repo, strictly preserving existing files and proposing merges where appropriate.
  2. Execute shared onboarding CLI: node "<harnessHome>/scripts/harness-project-onboarding.mjs" --root "<path-to-repo>" --write.
  3. Run discovered startup verification if safe.
  4. End with next recommendation: harness-project-calibration (to derive project-specific baselines). Recomendar cerebro-carga-inicial quando o projeto estiver registrado em ~/.task-pipeline/projects.json, e novo-projeto quando não estiver; ou /harness-session-start.

Rules

  • Never write before confirmation (unless pre-authorized by novo-projeto).
  • Always install the full package, not minimal modes. The shared CLI complements the package; it does not replace full template installation.
  • Preserve project-specific content, existing blockers, and operator configuration.
  • If a file exists, propose a merge instead of overwriting.
  • Do not assume init.ps1, feature_list.json, or .specs/project/STATE.md exist.
  • No tests or history means an explicit gap with the status returned by the tool (unavailable or not_applicable), never an invented test or fake passing suite.
← v1.0.0All versions