arckit-cli 6.1.1__py3-none-any.whl

arckit_cli-6.1.1.data/data/share/arckit/scripts/autoresearch/program-selfharness.md

@@ -0,0 +1,748 @@
+# ArcKit Self-Harness Enhanced Autoresearch Program
+
+<!-- markdownlint-disable MD009 MD022 MD031 MD032 MD040 -->
+
+**Status**: DRAFT - Self-Harness Integration  
+**Based on**: "Self-Harness: Harnesses That Improve Themselves" (Zhang et al., 2026, arXiv:2606.09498v1)  
+**Version**: 1.0  
+**Date**: 2026-06-17
+
+---
+
+## Overview
+
+This document extends `scripts/autoresearch/program.md` with **Self-Harness** capabilities based on the paper by Zhang et al. (2026). It adds:
+
+1. **Multi-dimensional optimization**: Beyond prompt text to full harness (prompts + tools + runtime + verification)
+2. **Verifier-grounded weakness mining**: Automated clustering of failure signatures from execution traces
+3. **Held-in/held-out validation**: Conservative acceptance rule to prevent overfitting
+4. **Agent and hook optimization**: Extends beyond commands to optimize agents and hooks
+
+**Reference**: All Self-Harness concepts are based on Zhang et al. (2026, arXiv:2606.09498v1)
+
+---
+
+## Quick Start
+
+To use Self-Harness enhanced autoresearch:
+
+```text
+read scripts/autoresearch/program-selfharness.md and optimize the requirements command
+```
+
+Specify optimization mode:
+- `optimize command requirements` - Command prompt optimization (default, backward compatible)
+- `optimize command requirements mode:full` - Full harness optimization for command
+- `optimize agent research` - Agent definition optimization
+- `optimize hook graph-inject` - Hook behavior optimization
+
+---
+
+## 1. Setup
+
+### 1.1 Standard Setup (Same as program.md)
+
+1. **Agree on target**: Command, agent, or hook name
+2. **Create worktree**: 
+   ```bash
+   git worktree add ../autoresearch-<target> -b autoresearch/<target>-<tag>
+   cd ../autoresearch-<target>
+   ```
+3. **Determine optimization mode**:
+   - `mode: prompt` (default) - Optimize command `.md` only (backward compatible)
+   - `mode: full` - Full harness optimization (prompt + tools + runtime + verification)
+   - `mode: agent` - Optimize agent definition
+   - `mode: hook` - Optimize hook behavior
+
+4. **Read in-scope files** (varies by mode):
+   - **prompt mode**: `plugins/arckit-claude/commands/<command>.md` (ONLY file you modify)
+   - **full mode**: commands/ + hooks/ + .mcp.json + templates/ (multiple files)
+   - **agent mode**: `plugins/arckit-claude/agents/<agent>.md` (ONLY file you modify)
+   - **hook mode**: `plugins/arckit-claude/hooks/<hook>.mjs` (ONLY file you modify)
+
+5. **Read templates and checklists** (read-only):
+   - Template file for expected output structure
+   - `plugins/arckit-claude/references/quality-checklist.md`
+
+### 1.2 Enhanced Setup (Self-Harness)
+
+6. **Create held-in/held-out fixtures**:
+   ```bash
+   # Copy base fixtures
+   cp -r scripts/autoresearch/fixtures/001-test-project scripts/autoresearch/fixtures/held-in/001-test-project
+   
+   # Create additional held-in fixtures (2-4 more)
+   cp -r scripts/autoresearch/fixtures/001-test-project scripts/autoresearch/fixtures/held-in/002-test-project
+   cp -r scripts/autoresearch/fixtures/001-test-project scripts/autoresearch/fixtures/held-in/003-test-project
+   
+   # Create held-out fixtures (2-3)
+   mkdir -p scripts/autoresearch/fixtures/held-out/004-test-project
+   mkdir -p scripts/autoresearch/fixtures/held-out/005-test-project
+   # Note: held-out fixtures should have variations to test generalization
+   ```
+
+7. **Set task lists**:
+   ```bash
+   export HELD_IN_TASKS=("001" "002" "003")
+   export HELD_OUT_TASKS=("004" "005")
+   ```
+
+8. **Initialize trace storage**:
+   ```bash
+   mkdir -p .arckit/autoresearch-traces/<target>
+   ```
+
+9. **Initialize results.tsv** with extended header:
+   ```text
+   commit  structural  score  effort  model  status  description  held_in_score  held_out_score  cluster  trace_id
+   ```
+
+10. **Set optimization parameters**:
+    ```bash
+    export MIN_DELTA=0.3           # Minimum improvement to accept
+    export MAX_ITERATIONS=50       # Iteration budget
+    export PLATEAU_THRESHOLD=15    # Discards before plateau detection
+    export PARALLEL_CANDIDATES=3   # Number of parallel proposals (for full mode)
+    ```
+
+11. **Run baseline on all tasks**:
+    - For each task in HELD_IN_TASKS + HELD_OUT_TASKS:
+      - Execute with current harness
+      - Capture trace
+      - Score output
+    - Compute average held-in and held-out scores
+    - Log baseline with status `baseline`
+
+---
+
+## 2. How Commands/Agents/Hooks Are Executed
+
+Execution method varies by optimization mode:
+
+### Mode: Prompt (Default - Backward Compatible)
+Same as original `program.md`:
+1. Read `plugins/arckit-claude/commands/<command>.md`
+2. Follow instructions directly
+3. Apply substitutions: `$ARGUMENTS`, `${CLAUDE_PLUGIN_ROOT}`
+4. Write to `scratch/projects/`
+
+### Mode: Full (Harness Optimization)
+1. Read current harness configuration (multiple files)
+2. Apply harness to execution environment
+3. Execute command with configured harness
+4. Capture full execution trace
+
+### Mode: Agent
+1. Launch agent with current definition
+2. Provide test input based on fixtures
+3. Capture agent execution trace (subagent calls, tool uses, decisions)
+4. Score final output
+
+### Mode: Hook
+1. Trigger hook with test input
+2. Capture hook execution trace (modifications, context injections)
+3. Measure impact on downstream command/agent
+4. Score based on effectiveness and side effects
+
+---
+
+## 3. Evaluation
+
+### 3.1 Structural Checks (Layer 1 - Gate)
+
+Same as original `program.md` - 8 checks must all pass:
+1. Document Control table with all 14 required fields
+2. Document ID follows `ARC-NNN-TYPE-vX.Y` pattern
+3. Revision History table present
+4. Standard footer present
+5. All major template sections present
+6. File written to correct path
+7. Domain-specific IDs correct (BR-xxx, FR-xxx, etc.)
+8. Wardley Map math validation (WARD commands only)
+
+If ANY check fails: score = `FAIL 0.0`, skip to logging
+
+### 3.2 LLM-as-Judge (Layer 2 - Quality)
+
+Same 5 dimensions, 1-10 scale:
+- **Completeness**: All sections substantively filled
+- **Specificity**: References actual project context
+- **Traceability**: Cross-references present and correct
+- **Actionability**: Could be used as-is by vendor/review board
+- **Clarity**: Well-structured, no contradictions
+
+Combined score = arithmetic mean, rounded to 1 decimal
+
+### 3.3 Trace Collection (NEW - Self-Harness)
+
+After each execution, capture:
+```json
+{
+  "iteration": N,
+  "target": "<command|agent|hook>",
+  "mode": "<prompt|full|agent|hook>",
+  "timestamp": "ISO8601",
+  "environment": {
+    "model": "claude-3-5-sonnet",
+    "effort": "high"
+  },
+  "execution": {
+    "toolCalls": [
+      {"name": "Read", "path": "...", "tokensIn": 100, "tokensOut": 500},
+      {"name": "WebSearch", "query": "...", "tokensIn": 20, "tokensOut": 300}
+    ],
+    "totalTokens": 1500,
+    "durationMs": 45000,
+    "artifactsCreated": ["projects/001/ARC-001-REQ-v1.0.md"]
+  },
+  "output": "...",
+  "verifier": {
+    "passed": true,
+    "failures": []
+  }
+}
+```
+
+Save to: `.arckit/autoresearch-traces/<target>/<mode>/iteration-N.json`
+
+---
+
+## 4. Weakness Mining (NEW - Self-Harness Section 3.2)
+
+After scoring, mine execution trace for failure patterns:
+
+### 4.1 Extract Failure Signature
+
+```javascript
+// Based on Zhang et al. (2026, Section 3.2)
+function extractFailureSignature(trace, verifierOutput, score) {
+  // If structural failure
+  if (score === 0.0) {
+    return {
+      verifierCause: getVerifierCause(verifierOutput),
+      agentBehavior: getAgentBehavior(trace),
+      mechanism: "structural_violation"
+    };
+  }
+  
+  // If score < MIN_DELTA improvement or discard
+  if (score < previousBest || (score - previousBest) < MIN_DELTA) {
+    return {
+      verifierCause: "quality_insufficient",
+      agentBehavior: analyzeAgentBehavior(trace),
+      mechanism: inferMechanism(trace, verifierOutput)
+    };
+  }
+  
+  return null; // No weakness if accepted
+}
+```
+
+**Verifier Causes** (from structural checks):
+- `missing_document_control`
+- `invalid_id_format`
+- `missing_revision_history`
+- `missing_footer`
+- `missing_sections`
+- `wrong_path`
+- `invalid_ids`
+- `wardley_math_error`
+
+**Agent Behaviors** (from trace analysis):
+- `early_termination`: Stopped before completing all sections
+- `excessive_tool_use`: Too many tool calls
+- `insufficient_exploration`: Didn't explore enough
+- `wrong_tool_selection`: Used incorrect tool
+- `missing_context`: Didn't reference project context
+- `repetitive_actions`: Repeated same action without progress
+- `ignored_guidance`: Didn't follow prompt instructions
+
+**Mechanisms** (inferred):
+- `insufficient_context`: Not enough project-specific info
+- `missing_guidance`: Prompt didn't provide enough direction
+- `unbounded_exploration`: No limits on tool use
+- `format_violation`: Output didn't match template
+- `cross_ref_missing`: Didn't link to other artifacts
+- `generic_output`: Too much boilerplate
+
+### 4.2 Cluster Failures
+
+Group failures by signature:
+```json
+{
+  "cluster_001": {
+    "signature": {
+      "verifierCause": "missing_sections",
+      "agentBehavior": "early_termination",
+      "mechanism": "missing_guidance"
+    },
+    "count": 5,
+    "frequency": 0.45,  // 45% of all failures
+    "traces": ["iteration-1", "iteration-3", "iteration-7", "iteration-12", "iteration-15"],
+    "addressableSurfaces": ["prompt", "templates", "verification_rules"],
+    "severity": "high"
+  }
+}
+```
+
+Save to: `.arckit/autoresearch-traces/<target>/<mode>/clusters.json`
+
+### 4.3 Addressable Surfaces Mapping
+
+Based on failure signature, determine what can be changed:
+
+| Verifier Cause | Agent Behavior | Mechanism | Addressable Surfaces |
+|---------------|----------------|-----------|---------------------|
+| missing_sections | early_termination | missing_guidance | prompt, templates, verification_rules |
+| wrong_path | * | * | prompt, runtime_policy |
+| invalid_ids | * | format_violation | prompt, verification_rules, templates |
+| * | excessive_tool_use | unbounded_exploration | runtime_policy, tool_restrictions, prompt |
+| * | * | insufficient_context | prompt, context_injection, bootstrap_instruction |
+| * | * | cross_ref_missing | prompt, context_injection, verification_rules |
+| quality_insufficient | * | generic_output | prompt, context_injection |
+
+---
+
+## 5. The Experiment Loop
+
+### LOOP UNTIL STOP CONDITION
+
+**Stop conditions** (extended from Zhang et al., 2026, Algorithm 1):
+1. Score target hit: best >= 9.5
+2. Iteration budget exhausted: iter >= MAX_ITERATIONS
+3. Double plateau detected: two plateau markers within 10 iterations
+4. Manual stop by user
+
+### 5.1 Standard Loop Steps (Extended)
+
+1. **READ**: Current harness configuration + results history + clusters
+2. **IDENTIFY**: ONE specific improvement based on:
+   - Low-scoring dimensions from LLM-as-judge
+   - Structural failures
+   - Template gaps
+   - Quality checklist criteria
+   - **NEW**: Top failure clusters from weakness mining
+   - **NEW**: Previous near-misses in results history
+   - **NEW**: For full mode: consider all addressable surfaces
+
+3. **GENERATE CANDIDATES** (NEW for full mode):
+   - If mode = `prompt`: Generate 1 candidate (original behavior)
+   - If mode = `full`: Generate PARALLEL_CANDIDATES (default: 3)
+     - Each candidate targets different cluster OR different surface
+     - Ensure diversity: different clusters, surfaces, or mechanisms
+
+4. **EDIT**: Apply change(s) to harness
+   - For prompt mode: Edit command `.md` file
+   - For full mode: Edit appropriate harness file(s)
+   - For agent mode: Edit agent `.md` file
+   - For hook mode: Edit hook `.mjs` file
+
+5. **COMMIT**: Git commit with description of changes
+   - Format: "[harness] <description> | cluster: <cluster_id> | surface: <surface>"
+
+6. **CLEAN**: Remove previously generated artifacts, keep fixtures
+
+7. **EXECUTE**: 
+   - For each task in HELD_IN_TASKS:
+     - Clean scratch project
+     - Copy fixtures
+     - Execute with candidate harness
+     - Capture trace
+     - Score output
+   - Compute average held-in score
+
+8. **MINE**: Run weakness mining on all held-in execution traces
+   - Extract failure signatures
+   - Update clusters
+   - Identify dominant patterns
+
+9. **VALIDATE** (Zhang et al., 2026, Section 3.4):
+   - For each task in HELD_OUT_TASKS:
+     - Clean scratch project
+     - Copy fixtures
+     - Execute with candidate harness
+     - Score output
+   - Compute average held-out score
+   - Compute deltas:
+     - Δ_in = avg_held_in - prev_avg_held_in
+     - Δ_out = avg_held_out - prev_avg_held_out
+
+10. **ACCEPT/REJECT** (Zhang et al., 2026, Algorithm 1):
+    ```
+    IF Δ_in >= 0 AND Δ_out >= 0 AND max(Δ_in, Δ_out) > MIN_DELTA:
+        STATUS = "keep"
+        Merge changes into active harness
+        Update previous best scores
+    ELSE:
+        STATUS = "discard"
+        Revert to previous best harness
+    ```
+
+11. **LOG**: Append row to results.tsv with all metrics
+
+12. **PRINT STATUS**:
+    ```text
+    [mode:<mode> iter N] score: X.X/X.X (best: Y.Y/Y.Y) | effort: high model: inherit | status: keep/discard | keeps: K discards: D | cluster: <top_cluster> | Δ_in: +A.A Δ_out: +B.B | streak: S/PL to plateau
+    ```
+
+13. **CHECK STOP CONDITIONS** (before step 1)
+
+### 5.2 Parallel Candidate Evaluation (Full Mode Only)
+
+For full mode with PARALLEL_CANDIDATES > 1:
+
+1. Generate N candidates (N = PARALLEL_CANDIDATES)
+2. For each candidate:
+   - Apply candidate changes
+   - Execute on held-in tasks
+   - Score and mine
+   - Validate on held-out tasks
+   - Determine accept/reject
+3. **Merge accepted candidates**:
+   - If multiple candidates accepted and compatible (no conflicts):
+     - Merge all changes
+     - Single commit with merged changes
+   - If candidates conflict:
+     - Accept highest-scoring candidate
+     - Reject others
+4. **Compatible**: Changes to different files OR non-overlapping changes to same file
+
+### 5.3 Plateau Detection (Extended)
+
+**Trigger**: Last PLATEAU_THRESHOLD (default: 15) consecutive iterations discarded
+
+**Strategy shift** (Zhang et al., 2026, Section 4.3):
+1. Re-read the template line by line for unaddressed sections
+2. Review quality checklist for uncovered criteria
+3. **NEW**: Review top failure clusters:
+   - Focus on highest-frequency cluster
+   - Generate candidates specifically for this cluster
+4. Try prompt simplification
+5. Try combining ideas from previous near-misses
+6. Try changing effort: or model: if not already tested
+7. **NEW**: Try different optimization surface:
+   - If only prompt changes tried: try tool configuration
+   - If only tool changes tried: try runtime mechanism
+8. Reset discard streak to 0 after strategy shift
+
+**Double plateau**: If second plateau within 10 iterations → exit with status `complete`
+
+---
+
+## 6. Output Format
+
+### 6.1 Results TSV (Extended)
+
+Tab-separated, header:
+```text
+commit  structural  score  effort  model  status  description  held_in_score  held_out_score  cluster  trace_id
+```
+
+Example:
+```text
+commit structural score effort model status description held_in_score held_out_score cluster trace_id
+a1b2c3d PASS 8.4 high inherit keep baseline 8.4 8.3 - iter-001
+b2c3d4e PASS 8.8 high inherit keep add cross-ref guidance 8.8 8.7 missing_cross_refs iter-002
+c3d4e5f PASS 9.0 high inherit discard simplify prompt 8.9 8.6 - iter-003
+d4e5f6g PASS 9.2 high inherit keep add context injection 9.2 9.1 generic_boilerplate iter-004
+e5f6g7h FAIL 0.0 high inherit discard remove doc control 0.0 0.0 structural_invalid iter-005
+```
+
+### 6.2 Trace Files
+
+Each iteration produces:
+- `.arckit/autoresearch-traces/<target>/<mode>/iteration-N.json` - Full execution trace
+- `.arckit/autoresearch-traces/<target>/<mode>/clusters.json` - All clusters (updated each iteration)
+
+---
+
+## 7. Optimization Modes Deep Dive
+
+### 7.1 Mode: Prompt (Default)
+
+**Backward compatible** with original `program.md`
+
+**Target**: `commands/<command>.md`
+
+**Editable**:
+- Instruction text
+- Examples
+- Formatting
+- effort: YAML field
+- model: YAML field
+- Section ordering
+
+**Read-only**:
+- Template
+- Quality checklist
+- Fixtures
+- Evaluation rubric
+
+**Proposal types**:
+- Add/remove instruction
+- Reword guidance
+- Add/remove example
+- Change effort/model
+- Reorder sections
+
+### 7.2 Mode: Full
+
+**Target**: Full harness for a command
+
+**Editable surfaces**:
+- `commands/<command>.md` (prompt, effort, model)
+- `hooks/*.mjs` that affect this command
+- `.mcp.json` tool configuration
+- `templates/<template>.md` for output
+- `config/*.mjs` for document types
+
+**Proposal types**:
+- **Type A - Prompt**: Same as mode:prompt
+- **Type B - Tools**: Enable/disable MCP servers, adjust tool permissions
+- **Type C - Runtime**: Modify hook behavior (graph depth, context injection)
+- **Type D - Verification**: Add/strengthen validation rules
+- **Type E - Orchestration**: Modify project context injection
+
+**Constraints**:
+- One surface per candidate (for parallel evaluation)
+- Minimal, targeted changes
+- Grounded in failure cluster
+
+### 7.3 Mode: Agent
+
+**Target**: `agents/<agent>.md`
+
+**Editable**:
+- Agent prompt/instructions
+- tools: list
+- maxTurns:
+- effort:
+- model:
+- description:
+
+**Proposal types**:
+- Add/remove tools
+- Strengthen instructions
+- Add pre-checks/post-checks
+- Adjust maxTurns
+- Change effort/model
+
+**Evaluation**:
+- Run agent on test scenarios
+- Score output artifacts
+- Validate on held-out scenarios
+
+### 7.4 Mode: Hook
+
+**Target**: `hooks/<hook>.mjs`
+
+**Editable**:
+- Matcher patterns
+- Timeout values
+- Configuration options
+- Logic flow
+- Error handling
+
+**Proposal types**:
+- Adjust timeout
+- Modify matcher
+- Add/remove configuration
+- Optimize logic
+- Improve error handling
+
+**Evaluation**:
+- Test hook with various inputs
+- Measure impact on commands/agents
+- Validate no regressions
+
+---
+
+## 8. File Structure
+
+```
+autoresearch-<target>/
+├── plugins/
+│   └── arckit-claude/
+│       ├── commands/           # Modified in prompt/full mode
+│       │   └── <command>.md
+│       ├── agents/            # Modified in agent mode
+│       │   └── <agent>.md
+│       └── hooks/             # Modified in hook/full mode
+│           └── <hook>.mjs
+├── .arckit/
+│   └── autoresearch-traces/
+│       └── <target>/
+│           ├── prompt/        # mode:prompt traces
+│           │   ├── iteration-001.json
+│           │   ├── iteration-002.json
+│           │   └── clusters.json
+│           └── full/          # mode:full traces
+│               ├── iteration-001.json
+│               └── clusters.json
+├── scripts/
+│   └── autoresearch/
+│       └── fixtures/
+│           ├── held-in/
+│           │   ├── 001-test-project/
+│           │   ├── 002-test-project/
+│           │   └── 003-test-project/
+│           └── held-out/
+│               ├── 004-test-project/
+│               └── 005-test-project/
+├── scratch/                  # Test execution area
+│   └── projects/
+│       └── <project>/
+├── results.tsv               # Extended format
+└── program-selfharness.md    # This file
+```
+
+---
+
+## 9. Tips
+
+### 9.1 Running Overnight
+
+- Each iteration takes 3-5 minutes with Self-Harness (vs 2-3 with original)
+- ~12-20 experiments per hour
+- Set MAX_ITERATIONS accordingly
+- Enable prompt caching: `ENABLE_PROMPT_CACHING_1H=1`
+
+### 9.2 Strategy Selection
+
+- **New command**: Start with mode:prompt, then escalate to mode:full
+- **Mature command**: Use mode:full directly
+- **Agent optimization**: Use mode:agent with agent-specific fixtures
+- **Hook optimization**: Use mode:hook with hook-specific tests
+
+### 9.3 Cluster Analysis
+
+- Review clusters.json regularly
+- Focus on highest-frequency clusters first
+- Some clusters may indicate template issues (not harness issues)
+- Document cluster patterns for future reference
+
+### 9.4 Cleanup
+
+- Cherry-pick kept commits to clean branch
+- Remove worktree when done: `git worktree remove ../autoresearch-<target>`
+- Archive traces: `tar -czvf traces-<target>.tar.gz .arckit/autoresearch-traces/<target>`
+
+---
+
+## 10. Limitations
+
+### 10.1 Self-Evaluation Bias
+
+- Same model generates and judges (mitigated by adversarial scoring)
+- Consider using different model for scoring (future enhancement)
+
+### 10.2 Computational Cost
+
+- Full mode: PARALLEL_CANDIDATES x (|HELD_IN| + |HELD_OUT|) executions per iteration
+- Recommend: PARALLEL_CANDIDATES=2-3, |HELD_IN|=3, |HELD_OUT|=2 for balance
+
+### 10.3 Non-Determinism
+
+- LLM output is non-deterministic
+- Use MIN_DELTA threshold to filter noise
+- Consider multiple runs for statistical significance
+
+### 10.4 Overhead
+
+- Trace collection adds ~10-20% overhead
+- Weakness mining adds ~5-10 seconds per iteration
+- Held-out validation doubles execution time
+
+---
+
+## 11. Usage Examples
+
+### Example 1: Standard Command Optimization (Backward Compatible)
+
+```text
+read scripts/autoresearch/program-selfharness.md and optimize the requirements command
+```
+
+This uses mode:prompt (default), same as original program.md
+
+### Example 2: Full Harness Optimization
+
+```text
+read scripts/autoresearch/program-selfharness.md and optimize the requirements command mode:full
+```
+
+Optimizes prompt + tools + runtime + verification for requirements command
+
+### Example 3: Agent Optimization
+
+```text
+read scripts/autoresearch/program-selfharness.md and optimize agent research
+```
+
+Optimizes arckit-research.md agent definition
+
+### Example 4: Hook Optimization
+
+```text
+read scripts/autoresearch/program-selfharness.md and optimize hook graph-inject
+```
+
+Optimizes graph-inject.mjs hook behavior
+
+---
+
+## 12. Implementation Notes
+
+### 12.1 Based on Zhang et al. (2026)
+
+This enhanced program implements the **Self-Harness** paradigm from:
+
+Zhang, H., Zhang, S., Li, K., Chen, Y., Zhang, Y., Bai, L., Hu, S. (2026). *Self-Harness: Harnesses That Improve Themselves*. arXiv preprint arXiv:2606.09498v1. Shanghai Artificial Intelligence Laboratory.
+
+**Key concepts implemented:**
+- Three-stage loop: Weakness Mining → Harness Proposal → Proposal Validation
+- Conservative acceptance rule (Algorithm 1)
+- Verifier-grounded failure signatures (Section 3.2)
+- Held-in/held-out split validation (Section 3.4)
+- Diverse, minimal candidate generation (Section 3.3)
+
+### 12.2 Extensions Beyond Paper
+
+- **Multi-mode optimization**: prompt, full, agent, hook
+- **Parallel candidate evaluation**: Multiple candidates per iteration
+- **Backward compatibility**: mode:prompt matches original behavior
+- **ArcKit-specific**: Integration with ArcKit hooks, agents, commands
+
+### 12.3 Backward Compatibility
+
+To use original behavior (no Self-Harness):
+```text
+read scripts/autoresearch/program.md and optimize the requirements command
+```
+
+Or with Self-Harness enhanced version:
+```text
+read scripts/autoresearch/program-selfharness.md and optimize the requirements command mode:prompt
+```
+
+Both produce identical results for mode:prompt
+
+---
+
+## Document Control
+
+| Document Control | | |
+|---|---|---|
+| **Document ID** | program-selfharness-v1.0 | |
+| **Document Type** | Program Specification | |
+| **Project** | ArcKit Autoresearch | |
+| **Classification** | Internal | |
+| **Status** | DRAFT | |
+| **Version** | 1.0 | |
+| **Created Date** | 2026-06-17 | |
+| **Last Modified** | 2026-06-17 | |
+| **Cites** | Self-Harness: Harnesses That Improve Themselves (Zhang et al., 2026, arXiv:2606.09498v1) | |
+
+---
+
+*Based on Self-Harness paper by Zhang et al. (2026, arXiv:2606.09498v1). Generated by Mistral Vibe. Co-Authored-By: Mistral Vibe <vibe@mistral.ai>*