dkh（DKH 搜索）

dkod Harness — Autonomous Parallel Build System

What This Is

A fully autonomous build harness. The user provides a single prompt ("build a webapp that..."). The harness does everything else — planning, parallel implementation, testing, fixing, and shipping — without any further user interaction.

This is an implementation of Anthropic's Planner → Generator → Evaluator harness pattern, purpose-built for dkod's parallel execution capabilities. Where Anthropic's reference architecture runs generators sequentially, this harness runs N generators simultaneously because dkod's AST-level merge eliminates false conflicts.

When This Skill Activates

• User says "build a...", "create a...", "make a..."
• User invokes /dkh <prompt>
• User describes a complete application or feature set to build from scratch
• Any task complex enough to benefit from parallel decomposition + evaluation

Handling /dkh continue

When the user sends /dkh continue (or just "continue"):

═══ MANDATORY: RECOVER STATE AND CLEAN UP BEFORE RESUMING ═══

Before re-dispatching ANY generators, recover the state from the interrupted session and clean up only what's incomplete. Do NOT bulk-close everything — submitted changesets represent completed work that must be preserved.

Step 1: Query dkod for existing changesets Call dk_status or list changesets via the API to see what the interrupted session left behind. Categorize each changeset:

• submitted state → KEEP. This generator finished its work. Record its changeset_id. Do NOT close it. Do NOT re-dispatch this unit.
• approved state → KEEP. This changeset passed review and is ready to merge. Record its changeset_id. Proceed to merge in Phase 3.
• draft state → INCOMPLETE. This generator was interrupted before dk_submit. Mark this unit for re-dispatch.
• conflicted state → STUCK. Mark this unit for re-dispatch.
• rejected state → FAILED. Mark this unit for re-dispatch.

Step 2: Close incomplete/failed changesets and release their symbol claims

# Close draft/conflicted/rejected — preserve submitted and approved
Bash: curl -sf -X POST "https://api.dkod.io/api/repos/<owner>/<repo>/changesets/bulk-close" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DKOD_API_KEY" \
  -d '{"states": ["draft", "conflicted", "rejected"], "created_before": "'$(date -u +%Y-%m-%dT%H:%M:%SZ)'"}'  \
  || { echo "Bulk-close failed — aborting resume. Check DKOD_API_KEY and repo path."; exit 1; }

Step 3: Reconstruct harness state

• changeset_ids = list of submitted + approved changeset_ids from Step 1
• active_units = units whose generators were incomplete (draft/conflicted/rejected/missing)
• Units with submitted or approved changesets are DONE — skip them

Step 4: Resume with only the incomplete units Re-dispatch only the generators for active_units (incomplete ones). The submitted changesets from completed generators are preserved — 20 minutes of work is not lost.

Output: "Resuming harness — N/M generators completed before interruption. Re-dispatching K incomplete units."

Then proceed:

1. If an active harness session exists (the agent has state from a prior turn):

- Output the current harness state and phase - Show which agents completed (submitted) vs need re-dispatch - Resume the harness loop, skipping completed units

1. If no active harness session exists (fresh context after app restart):

- Acknowledge the command: "No active harness session found in this context." - Check for any active dkod sessions via dk_status - If active sessions found, recover state as described above - If no sessions found, tell the user to start a new build with /dkh <prompt>

Never ignore a /dkh continue silently. Always acknowledge with current status.

Prerequisites Check

Before starting, verify these are available:

1. dkod MCP tools: dk_connect, dk_context, dk_file_write, dk_submit, dk_verify, dk_review, dk_approve, dk_merge, dk_push, dk_status, dk_watch
2. Browser testing (pick one — Playwright preferred):

- playwright-cli (preferred): Check with playwright-cli --version. Standalone CLI for skills-less browser automation — screenshots, script execution, PDF generation. No Node.js scripts needed, no MCP server, runs headless by default. See: https://github.com/microsoft/playwright-cli - chrome-devtools MCP (fallback): navigate_page, take_screenshot, click, evaluate_script, list_console_messages, lighthouse_audit. Used only if Playwright is not installed. - If not found, output install instructions and proceed with fallback. Do NOT ask the user or install. - If NEITHER Playwright nor chrome-devtools is available, evaluation falls to dk_verify + code review (no live UI testing).

1. Design system (pick one — DESIGN.md preferred):

- DESIGN.md (preferred): A design system file in the project root, sourced from awesome-design-md. If present, it becomes the authoritative design reference — the planner incorporates it into the spec, generators follow it directly (no skill invocation needed), and the evaluator scores against it. This produces more distinctive, brand-aligned UI than the generic skill. - frontend-design skill (fallback): If no DESIGN.md exists, generators invoke Skill(skill: "frontend-design") before implementing UI components. The planner still generates a Design Direction section in the spec. The evaluator still scores design quality. If not found, output install instructions and proceed with fallback. Do NOT ask the user or install. - If NEITHER is available, generators follow the planner's Design Direction section manually.

Detection flow (run once during PRE-FLIGHT):

# 1. Detect playwright-cli (standalone CLI, skills-less operation)
HAS_PLAYWRIGHT=false
playwright-cli --version 2>/dev/null && HAS_PLAYWRIGHT=true

# 2. Detect DESIGN.md (check all paths the planner searches)
HAS_DESIGN_MD=false
( [ -f DESIGN.md ] || [ -f design.md ] || [ -f docs/DESIGN.md ] || [ -f docs/design.md ] ) && HAS_DESIGN_MD=true

# Pass both flags to all agent dispatches:
# - Planner: HAS_DESIGN_MD
# - Generators: HAS_DESIGN_MD
# - Evaluators: HAS_PLAYWRIGHT
# - Smoke test: HAS_PLAYWRIGHT

If dkod is missing, guide installation:

claude mcp add --transport http dkod https://api.dkod.io/mcp

Model Profiles

Active profile: quality

Each agent runs on a model appropriate to its task. The orchestrator reads the active profile and passes model: AND effort: on every Agent dispatch call.

• The orchestrator model is set by the invoking Claude Code session, not by this table. This row is a recommendation for the session model, not enforced by the harness.

Effort levels are mandatory. Planner and Evaluator use max (complex reasoning — decomposition, scoring). Generator uses high (fast execution — file writes, not deep analysis). Always pass effort: when dispatching agents.

• quality — All Opus. Maximum capability. Use for complex or high-stakes builds.
• balanced (default) — Opus for planning and orchestration, Sonnet for implementation and evaluation. Best cost/quality trade-off.
• budget — Sonnet for planning and implementation, Haiku for evaluation. Fastest and cheapest. Use for simple builds or iteration.

To switch profiles, change Active profile: above. The orchestrator reads this value at the start of each run.

The Autonomous Loop

The harness runs a strict phase-gated loop: PLAN → BUILD → LAND → File Sync → Smoke Test → EVAL → SHIP/FIX/REPLAN. Each phase has entry/exit gates that block progression until artifacts exist.

agents/orchestrator.md is the single source of truth for all phase details, gate checks, state tracking, dispatch templates, and transition logic. Read it before starting.

Key constraints:

• No dk_push(mode:"pr") without completed eval reports (Phase 5 only)
• No evaluators without landed + smoke-tested code
• No generators without a validated plan
• dk_verify is NOT evaluation — Phase 4 tests the live app via chrome-devtools
• All code changes go through dkod — never use Write/Edit/Bash on source files
• Never ask the user anything — every decision is autonomous
• Max 3 eval rounds, then ship with documented issues

Critical Design Principles

0. Maximize parallelism — THE PRIME DIRECTIVE

Default to parallel execution; only serialize when there is a hard data dependency. Use Claude Code agent teams (multiple Agent calls in one message) + dkod session isolation (each agent gets its own overlay). Together they turn serial builds into parallel builds.

Applies to every phase: generators dispatch simultaneously, dk_verify runs in parallel, failed generators re-dispatch in parallel. The ONE exception: evaluators run sequentially (shared chrome-devtools browser session).

1. Decompose by symbol, not file

dkod merges at the AST level. Two generators editing different functions in the same file is not a conflict.

2. One dkod session per generator

Each generator calls dk_connect once. Sessions are isolated until merge.

3. The Evaluator is standalone and skeptical

From Anthropic's research: "tuning a standalone evaluator to be skeptical turns out to be far more tractable than making a generator critical of its own work." Default to FAIL unless proven PASS with evidence.

4. File-based communication

The plan and eval reports are structured artifacts that survive context resets.

5. Autonomy is non-negotiable

The harness NEVER asks the user for input. Conflicts → auto-resolve. Eval failures → auto-fix. Framework → infer from prompt. Package manager → bun.

6. Max 3 eval rounds

Prevents infinite loops. After 3 rounds, ship whatever works and document what doesn't.

Work Unit Schema

The Planner produces work units in this structure (embedded in the plan artifact):

## Work Unit: <id>
**Title:** <descriptive title>
**OWNS (exclusive):** <list of qualified symbol names this unit solely owns>
**Creates:** <list of new symbols with file paths>
**Acceptance criteria:**
- <testable criterion 1>
- <testable criterion 2>
**Complexity:** low | medium | high

Agent Definitions

• Planner: agents/planner.md — expands prompt into spec + parallel work units
• Generator: agents/generator.md — implements a single work unit via dkod session
• Evaluator: agents/evaluator.md — tests merged result via chrome-devtools + dk_verify
• Orchestrator: agents/orchestrator.md — drives the autonomous loop (this is you)

Reference Guides

• references/planning-guide.md — deep guide for symbol-level decomposition
• references/evaluation-guide.md — skeptical evaluation techniques and chrome-devtools patterns
• references/dkod-patterns.md — dkod session lifecycle and merge patterns