Harness Moaiadk Best Practices

moai-adk-go Best Practices

TRUST 5 Quality Gates

Every change must pass all five dimensions before completion:

Coverage targets: 85% package minimum; 90%+ for critical packages (internal/cli, internal/template, internal/hook).

Test Isolation

• Always t.TempDir() for temp dirs — auto-cleanup, under os.TempDir().
• macOS path pitfall: t.TempDir() returns /var/folders/.... Go's filepath.Join(cwd, absPath) does NOT strip the leading /: filepath.Join("/a/b", "/var/folders/x") → "/a/b/var/folders/x" (WRONG). Use filepath.Abs() when resolving user-supplied paths in CLI commands.
• No OTEL env in parallel tests (CLAUDE.local.md §WARN): never t.Setenv("OTEL_EXPORTER_*", ...) in parallel tests — the OTEL SDK initializes global state from env vars on first use, causing data races. Use a fake/no-op exporter instead; or make the parent test non-parallel.
• No t.Setenv("HOME", tmpDir) in GLM integration tests — parallel-test pollution. Use t.TempDir() + explicit path construction.
• After fixing any test, run the FULL suite (go test ./...) to catch cascading failures. Use -count=1 to disable caching when debugging flaky tests; use -race for concurrency-safety checks.

Hardcoding Prevention

• URLs / model names / org names / API headers → extract to const.
• Environment variable names → define in internal/config/envkeys.go as constants; reference the constant everywhere. Never inline a raw env string.
• Thresholds → single source in internal/config/defaults.go. Never duplicate a threshold across packages.
• Cross-platform paths → prefer $HOME, HOMEBREW_PREFIX, etc. In .sh.tmpl fallback paths use $HOME (not .HomeDir), because .HomeDir freezes at moai init time and breaks for users with non-standard layouts.
• Hardcoding allowed only in CLAUDE.local.md, settings.local.json, and _test.go files inside t.TempDir().

AskUserQuestion Boundary (orchestrator-only)

• AskUserQuestion is the ONLY user-facing question channel, and it is reserved for the MoAI orchestrator (main session).
• Subagents (including these harness specialists) MUST NOT invoke AskUserQuestion. If user input is required, return a structured blocker report to the orchestrator (see .claude/rules/moai/core/askuser-protocol.md § Blocker Report Format).
• Deferred-tool preload: AskUserQuestion, TaskCreate, TaskUpdate, TaskList, TaskGet are deferred tools — schema not loaded at session start. The orchestrator MUST call ToolSearch(query: "select:AskUserQuestion,TaskCreate,...") before first use. Subagents inherit this constraint.
• Free-form prose questions in response text are prohibited — always route through AskUserQuestion (orchestrator) or a blocker report (subagent).

Archived-Agent Rejection Contract

12 agents are ARCHIVED and MUST NOT be referenced anywhere in generated harness files — no delegates-to, no prose, no examples. The full list of archived names lives in the canonical SSOT at .claude/rules/moai/workflow/archived-agent-rejection.md §B; this skill does not repeat the literal names (repeating them in every generated file would re-seed the exact tokens the rejection contract is meant to suppress).

The 8 RETAINED agents are the only valid delegation targets:

manager-spec, manager-develop, manager-docs, manager-git,
plan-auditor, sync-auditor, builder-harness, Explore (Anthropic built-in)

For domain expertise formerly provided by the archived domain-expert agents, use the per-spawn pattern: Agent(subagent_type: "general-purpose", model: "opus", tools: "<whitelist>", prompt: "...<domain> specialist: <conventions>...") at delegation time. See .claude/rules/moai/workflow/archived-agent-rejection.md §C for the full migration table (rows #1-#12), which maps each archived agent to its canonical retained-agent or per-spawn replacement.

Verification-Claim Integrity

Per .claude/rules/moai/core/verification-claim-integrity.md:

• No unobserved claims. A "tests pass" / "coverage 87%" / "lint clean" assertion is valid ONLY when the actor ran the command and observed the output. An unran command is a gap, never a pass.
• No unobserved defect claims. Inferring a defect/debt/drift from frontmatter text or grep matches alone — without the domain's dedicated tool (moai spec audit, go test -cover, golangci-lint) — is a hypothesis, not a verified defect. The 2026-06-17 incident (29 SPECs wrongly flagged as "Mx-close debt"; moai spec audit showed all 29 were grandfather-protected) is the canonical worked example.
• Baseline attribution. Every verification claim names the command run + the verbatim output observed, measured against this tree in this run. A number from a different SPEC/package/time is a carry-over, not a baseline.
• 5-section report format: Claim / Evidence / Baseline-attribution / Gaps / Residual-risk. The Gaps section is the defense — force yourself to enumerate what was NOT observed.

Cross-References

• CLAUDE.local.md §6 (testing), §14 (hardcoding), §19 (AskUserQuestion)
• .claude/rules/moai/core/verification-claim-integrity.md v1.1.0
• .claude/rules/moai/core/askuser-protocol.md
• .claude/rules/moai/workflow/archived-agent-rejection.md
• .claude/rules/moai/development/coding-standards.md