Harness Engineering Readiness Roadmap

Based on Harness Engineering: Leveraging Codex in an Agent-First World (OpenAI, Feb 2026)

Last updated: 2026-03-14

Current readiness: ~25%

Pillar Assessment

Pillar 1: Repo Knowledge as System of Record

Principle: AGENTS.md is a table of contents, not the encyclopedia. Progressive disclosure. Anything outside the repo does not exist.

Done

• AGENTS.md at repo root (table of contents, progressive disclosure)
• ARCHITECTURE.md at repo root (system reference with source-file refs, ownership rule)
• docs/architecture.mdx renamed to "Design Philosophy" (why decisions were made)
• Legacy docs/Docs_To_Review/ARCHITECTURE.md deprecated with banner
• Cross-references between all architecture docs
• Proto contract system with CI freshness checks
• Comprehensive Mintlify docs site with API reference

Remaining

• Create docs/design-docs/ directory with index.md
• Create docs/exec-plans/active/ and docs/exec-plans/completed/
• Create docs/product-specs/ with index.md
• Migrate relevant .claude/memory/ entries into repo-visible docs (conventions that apply to all contributors, not just Claude)
• Add docs/generated/ for auto-generated reference docs (e.g., db-schema, cache-key inventory)

Pillar 2: Enforced Architecture

Principle: Documentation alone cannot maintain coherence. Custom linters enforce dependency direction, naming, file size, structured logging. Lint errors include remediation instructions for agents.

Done

• TypeScript strict mode (noUncheckedIndexedAccess, noUnusedLocals, noUnusedParameters)
• tsc --noEmit in CI and pre-push hook
• Edge function self-containment check (esbuild bundle + import guardrail test)
• Proto breaking-change detection (buf breaking)
• Markdown linting in CI

Remaining

• P0: Add JS/TS linter (Biome 2.4.7) — biome.json, npm run lint, CI workflow lint-code.yml, ~120 files auto-fixed
• P0: Architectural boundary lint — scripts/lint-boundaries.mjs, npm run lint:boundaries, CI enforced. Fixed 12 violations (moved types to proper layers). 3 pragmatic exceptions with boundary-ignore comments
• Encode .claude/memory/ conventions as lint rules:
  • Ban fetch.bind(globalThis) (use deferred lambda)
  • Require cachedFetchJson() in new RPC handlers
  • Require seed-meta:<key> write in seed scripts
  • Require User-Agent header in server-side fetch
  • Require cache key includes request-varying params
• File size limits with warnings
• Structured logging enforcement in API handlers

Pillar 3: Application Legibility (Agent Observability)

Principle: Agents must be able to launch the app, navigate UI, capture screenshots, inspect DOM, and query logs/metrics/traces.

Done

• Sentry error tracking in browser
• api/health.js with per-key freshness monitoring
• api/seed-health.js for seed loop monitoring
• Playwright E2E test infrastructure (config, specs, visual regression)
• Circuit breaker instrumentation

Remaining

• P1: Expand Playwright E2E harness for agent-driven validation (launch app, navigate, screenshot, assert)
• P1: Add structured JSON logging to API handlers (request ID, latency, error context)
• Expose logs in a queryable format (even grep on Railway logs is a start)
• Add performance budgets (startup time, critical path latency) as testable assertions
• Wire Chrome DevTools Protocol for agent DOM inspection (desktop)

Pillar 4: Agent-to-Agent Review Loops

Principle: Agent reviews its own work locally. Additional agents review. Feedback loops run until reviews pass. Humans sometimes review PRs.

Done

• Pre-push hook runs automated checks (typecheck, edge bundle, markdown lint)
• CI runs typecheck on all PRs

Remaining

• P2: Configure agent PR review in CI (check for architectural violations, convention adherence, test coverage)
• Start with advisory comments, not blocking
• Add self-review step: agent runs tests + lint before opening PR
• Multi-agent review: different agents check different aspects (security, performance, conventions)

Pillar 5: Self-Healing / Garbage Collection

Principle: Background agents scan for violations and open refactoring PRs. Technical debt becomes incremental maintenance instead of large refactors.

Done

• "Golden principles" partially encoded in AGENTS.md (key patterns, critical conventions)

Remaining

• P3: Create convention violation scanner (dead code, banned patterns, missing seed-meta, cache key issues)
• Background agent opens small refactoring PRs
• Track tech debt in docs/exec-plans/tech-debt-tracker.md
• Define "golden principles" document with shared utilities, data shape validation rules, anti-patterns

Pillar 6: Full Feature Loops

Principle: Given a prompt, agent can validate repo state, reproduce bug, record video, implement fix, validate fix, open PR, address feedback, merge.

Done

• Agents can open PRs via gh
• Agents can run tests via npm run test:data
• Git worktree support for isolated work

Remaining

• P4: Agent bug reproduction harness (receive bug report, reproduce, record screenshot/video)
• Agent self-merge pipeline for low-risk PRs (requires P0-P2 as safety net)
• Agent escalation protocol (when to ask human vs. proceed)
• Build failure auto-repair (agent detects CI failure, fixes, re-pushes)

Pillar 7: Doc Linters / Gardening

Principle: Dedicated linters validate documentation freshness, cross-links, structure. Background agent runs doc gardening tasks.

Done

• markdownlint-cli2 in CI and pre-push
• MDX lint for Mintlify compatibility
• Ownership rule in ARCHITECTURE.md ("update in same PR")

Remaining

• P3: Doc freshness linter (detect stale dates, broken internal links, orphaned docs)
• Cross-link validator (ensure all doc references resolve)
• Doc gardening agent (background task to fix stale docs, update counts, verify source-file refs)

Implementation Order

Phase 1 (P0) — Foundation          ← START HERE
├── Add Biome/ESLint linter
├── Add tests to CI
└── Architectural boundary rules

Phase 2 (P1) — Agent Observability
├── Expand Playwright harness
├── Structured logging
└── Encode memory conventions as lint rules

Phase 3 (P2) — Review Loops
├── Automated PR review agent
└── Golden patterns doc

Phase 4 (P3) — Self-Healing
├── Convention violation scanner
├── Doc freshness linter
└── Tech debt tracker

Phase 5 (P4) — Full Autonomy
├── Bug reproduction harness
├── Self-merge pipeline
└── Progressive disclosure doc tree

Progress Log