Harness-Neutral
Agentic SDLC
Turn an approved Spec into verified software deliveries inside your coding harness.
Gantry is a skill pack, not an execution engine. Workflow scripts decide readiness, acceptance, gates, and roadmap state; agents plan, implement, review, and refute; the operator holds approval authority.
npx skills add JulioBorges/gantrynpx @julioborges/gantry setup --harness claude
Pipeline Simulator
Explore the 6 deterministic execution phases from Spec verification to serial gate integration.
One-Time Repository Setup
Adapts the repository to Gantry once, before any feature planning or implementation Run.
/gantry-setup- ■Runs once per repository, then remains versioned as repository policy.
- ■Maps existing artifacts, templates, checks, Git policy, roles, and supported hooks.
- ■Shows the complete proposed policy and waits for operator approval before writing.
- ›Inspects repository conventions instead of creating duplicate structures.
- ›Explains each configuration choice, benefit, and trade-off.
- ›Writes policy only after explicit operator confirmation.
/gantry-setup{
"event": "setup.completed",
"repository": "otp-service",
"policy": ".gantry/config.json",
"status": "ready-for-planning",
"timestamp": "2026-09-17T18:00:00Z"
}Why Raw Agents Fail — And How Gantry Fixes It
Prose instructions in prompts are probabilistic. Gantry replaces wishful prompting with deterministic code, adversarial checks, and operator approval.
"Agents + code > agents alone."
Every decision that must not be improvised — what is ready, what must be proven, whether quality gates pass, and when an Issue is done — is made by a workflow script. Agents propose; scripts verify; the operator retains final approval.
Context Degradation & Hallucination
Long-running agent chat sessions accumulate tokens, lose instruction precision, drop critical constraints, and hallucinate missing architecture.
Preflight validates tokens with budget.py (refuting packages > 15% window). Every phase (Implementer, Reviewer, Critic) spawns a fresh subagent with zero memory pollution.
Mock Escapes & Test Tampering
When code fails, agents casually mock out complex modules, weaken assertions, delete failing tests, or claim victory on untested stubs.
The adversarial Critic directly inspects the git diff for mock escapes, deleted tests, or weakened checks. No delivery is accepted if tests were bypassed or watered down.
Wishful Self-Approval & Premature Done
When the builder grades its own homework, the implementer hallucinates that edge cases pass, summarizes away defects, and marks its own tasks as completed.
Zero self-approval: roadmap.py done is the sole authority to mark issues complete, guarded by differential quality gates (RFC 6901 mappings) and clean test exits. Critic refutation enforces proof.
Big-Bang Divergence & Integration Hell
Parallel multi-agent branches create silent regressions when merged in bulk, hiding broken dependencies until after the repository is corrupted.
frontier.py computes DAG execution waves. Accepted slices merge serially, running differential quality gates (gates.py) after every merge.
The Four Skills
Gantry ships four focused skills: one-time repository setup, standalone planning, automatic implementation, and read-only Run visibility.
The orchestrator executing the full SDLC loop: spec validation, context budget checks, TDD implementation loops, two-axis reviews (standards + spec), and adversarial verification.
- › TDD Cycles: Red-green-refactor with strictly minimal diffs.
- › Adversarial Critic: Fresh agent disproves claims before integration.
- › Quality Gates: Differential gates.py run serially.
Conversational repository setup and governance assistant. The sole authorized writer of repository policy (.gantry/config.json) and marked AGENTS.md policy sections.
- › Policy Generation: Interactive questionnaire configuring gates and budgets.
- › Harness Detection: Inspects active IDE harness and configures hooks.
- › Drift Protection: Never writes unvalidated or conflicting settings.
Transforms a free-text goal or existing Spec into a verified Spec, tracer-bullet Issues, a budgeted breakdown, and an explicit operator approval gate.
- › Socratic Gate: Clarifies scope, seams, non-goals, and scenarios.
- › Spec & Issues: Writes the Spec and vertical Issue breakdown.
- › Approval Stop: Never starts implementation before operator approval.
Zero-dependency, read-only multi-run kanban observer. Binds locally to loopback (localhost:4600) with strict zero-mutation guarantees.
- › Multi-Run Kanban: Visual swimlanes tracking issues across all phases.
- › Append-Only Log: Reads ~/.gantry/state/ without database overhead.
- › Strict Loopback: Binds exclusively to 127.0.0.1 for local privacy.
Harness Compatibility Matrix
Gantry declares harness capabilities in machine-readable JSON files (.agents/skills/gantry/capabilities/) and proves them against real test fixtures. Gantry never advertises guarantees it cannot measure.
Full native harness support: lifecycle hooks guard, structured output contracts, isolated worktrees, per-role model selection, and complete end-to-end fixture proof.
Multi-role execution, hook guard integration (via harness hooks or runner wrap), and independent Critic verification proven in automated test suites.
Executable workflow through process runner, skill discovery, and prompt instructions. Guard hooks fall back to repository policy and manual operator gates.
| Harness | Tier | Hook Guardrails | Role Execution | Worktree Isolation | Structured Output | Verification Evidence & Notes |
|---|---|---|---|---|---|---|
| Claude Code | Reference | Native Pre/PostToolUse hooks | Full Multi-Role (Subagents) | Native Git Worktrees | JSON Schema (result.py) | Native subagents, settings hooks, result contract validation, and full end-to-end fixture run recorded in CI. |
| Gemini / Antigravity | Supported | .agents/hooks.json guard | Subagents & agy --print | Dedicated Worktrees | JSON Schema (result.py) | Supported tier. Hook guard integration via .agents/hooks.json and agy execution verified in local test suite. |
| Codex CLI | Supported | Direct Process Runner Guard | Independent Critic & Roles | Worktree Supported | CLI JSON Contract | Supported tier. Direct CLI runner with per-role model arguments, independent Critic verification proven on real runs. |
| OpenCode | Compatible | Plugin Hooks Wiring (opencode.json) | Per-role subagents | Host Worktree | Contract Verified | Plugin hooks wiring for guard.py, sequential and parallel rounds, verified against OpenCode test harnesses. |
| Cursor | Compatible | Rules & Commands Policy | Composer Workflows | Shared Worktree | Text Protocol Output | Compatible tier. Rules and commands integration for Cursor composer workflows with manual operator gates. |
A declared tier does not prove every harness version has been exercised. Gantry tests specific integrations against reproducible fixture suites and logs the active tier in every Run report. No unearned claims, no marketing hyperbole.