test: add tests for session pair harvester (#91)

GENOME.md

@@ -1,239 +0,0 @@
-# GENOME.md — compounding-intelligence
-
-*Auto-generated codebase genome. See timmy-home#676.*
-
----
-
-## Project Overview
-
-**What:** A system that turns 1B+ daily agent tokens into durable, compounding fleet intelligence.
-
-**Why:** Every agent session starts at zero. The same mistakes get made repeatedly — the same HTTP 405 is rediscovered as a branch protection issue, the same token path is searched for from scratch. Intelligence evaporates when the session ends.
-
-**How:** Three pipelines form a compounding loop:
-
-```
-SESSION ENDS → HARVESTER → KNOWLEDGE STORE → BOOTSTRAPPER → NEW SESSION STARTS SMARTER
-                              ↓
-                         MEASURER → Prove it's working
-```
-
-**Status:** Early stage. Template and test scaffolding exist. Core pipeline scripts (harvester.py, bootstrapper.py, measurer.py, session_reader.py) are planned but not yet implemented. The knowledge extraction prompt is complete and validated.
-
----
-
-## Architecture
-
-```mermaid
-graph TD
-    A[Session Transcript<br/>.jsonl] --> B[Harvester]
-    B --> C{Extract Knowledge}
-    C --> D[knowledge/index.json]
-    C --> E[knowledge/global/*.md]
-    C --> F[knowledge/repos/{repo}.md]
-    C --> G[knowledge/agents/{agent}.md]
-    D --> H[Bootstrapper]
-    H --> I[Bootstrap Context<br/>2k token injection]
-    I --> J[New Session<br/>starts smarter]
-    J --> A
-    D --> K[Measurer]
-    K --> L[metrics/dashboard.md]
-    K --> M[Velocity / Hit Rate<br/>Error Reduction]
-```
-
-### Pipeline 1: Harvester
-
-**Status:** Prompt designed. Script not implemented.
-
-Reads finished session transcripts (JSONL). Uses `templates/harvest-prompt.md` to extract durable knowledge into five categories:
-
-| Category | Description | Example |
-|----------|-------------|---------|
-| `fact` | Concrete, verifiable information | "Repository X has 5 files" |
-| `pitfall` | Errors encountered, wrong assumptions | "Token is at ~/.config/gitea/token, not env var" |
-| `pattern` | Successful action sequences | "Deploy: test → build → push → webhook" |
-| `tool-quirk` | Environment-specific behaviors | "URL format requires trailing slash" |
-| `question` | Identified but unanswered | "Need optimal batch size for harvesting" |
-
-Output schema per knowledge item:
-```json
-{
-  "fact": "One sentence description",
-  "category": "fact|pitfall|pattern|tool-quirk|question",
-  "repo": "repo-name or 'global'",
-  "confidence": 0.0-1.0
-}
-```
-
-### Pipeline 2: Bootstrapper
-
-**Status:** Not implemented.
-
-Queries knowledge store before session start. Assembles a compact 2k-token context from relevant facts. Injects into session startup so the agent begins with full situational awareness.
-
-### Pipeline 3: Measurer
-
-**Status:** Not implemented.
-
-Tracks compounding metrics: knowledge velocity (facts/day), error reduction (%), hit rate (knowledge used / knowledge available), task completion improvement.
-
----
-
-## Directory Structure
-
-```
-compounding-intelligence/
-├── README.md                           # Project overview and architecture
-├── GENOME.md                           # This file (codebase genome)
-├── knowledge/                          # [PLANNED] Knowledge store
-│   ├── index.json                      # Machine-readable fact index
-│   ├── global/                         # Cross-repo knowledge
-│   ├── repos/{repo}.md                 # Per-repo knowledge
-│   └── agents/{agent}.md               # Agent-type notes
-├── scripts/
-│   ├── test_harvest_prompt.py          # Basic prompt validation (2.5KB)
-│   └── test_harvest_prompt_comprehensive.py  # Full prompt structure test (6.8KB)
-├── templates/
-│   └── harvest-prompt.md               # Knowledge extraction prompt (3.5KB)
-├── test_sessions/
-│   ├── session_success.jsonl           # Happy path test data
-│   ├── session_failure.jsonl           # Failure path test data
-│   ├── session_partial.jsonl           # Incomplete session test data
-│   ├── session_patterns.jsonl          # Pattern extraction test data
-│   └── session_questions.jsonl         # Question identification test data
-└── metrics/                            # [PLANNED] Compounding metrics
-    └── dashboard.md
-```
-
----
-
-## Entry Points and Data Flow
-
-### Entry Point 1: Knowledge Extraction (Harvester)
-
-```
-Input:  Session transcript (JSONL)
-        ↓
-        templates/harvest-prompt.md (LLM prompt)
-        ↓
-        Knowledge items (JSON array)
-        ↓
-Output: knowledge/index.json + per-repo/per-agent markdown files
-```
-
-### Entry Point 2: Session Bootstrap (Bootstrapper)
-
-```
-Input:  Session context (repo, agent type, task type)
-        ↓
-        knowledge/index.json (query relevant facts)
-        ↓
-        2k-token bootstrap context
-        ↓
-Output: Injected into session startup
-```
-
-### Entry Point 3: Measurement (Measurer)
-
-```
-Input:  knowledge/index.json + session history
-        ↓
-        Velocity, hit rate, error reduction calculations
-        ↓
-Output: metrics/dashboard.md
-```
-
----
-
-## Key Abstractions
-
-### Knowledge Item
-The atomic unit. One sentence, one category, one confidence score. Designed to be small enough that 1000 items fit in a 2k-token bootstrap context.
-
-### Knowledge Store
-A directory structure that mirrors the fleet's mental model:
-- `global/` — knowledge that applies everywhere (tool quirks, environment facts)
-- `repos/` — knowledge specific to each repo
-- `agents/` — knowledge specific to each agent type
-
-### Confidence Score
-0.0–1.0 scale. Defines how certain the harvester is about each extracted fact:
-- 0.9–1.0: Explicitly stated with verification
-- 0.7–0.8: Clearly implied by multiple data points
-- 0.5–0.6: Suggested but not fully verified
-- 0.3–0.4: Inferred from limited data
-- 0.1–0.2: Speculative or uncertain
-
-### Bootstrap Context
-The 2k-token injection that a new session receives. Assembled from the most relevant knowledge items for the current task, filtered by confidence > 0.7, deduplicated, and compressed.
-
----
-
-## API Surface
-
-### Internal (scripts not yet implemented)
-
-| Script | Input | Output | Status |
-|--------|-------|--------|--------|
-| `harvester.py` | Session JSONL path | Knowledge items JSON | PLANNED |
-| `bootstrapper.py` | Repo + agent type | 2k-token context string | PLANNED |
-| `measurer.py` | Knowledge store path | Metrics JSON | PLANNED |
-| `session_reader.py` | Session JSONL path | Parsed transcript | PLANNED |
-
-### Prompt (templates/harvest-prompt.md)
-
-The extraction prompt is the core "API." It takes a session transcript and returns structured JSON. It defines:
-- Five extraction categories
-- Output format (JSON array of knowledge items)
-- Confidence scoring rubric
-- Constraints (no hallucination, specificity, relevance, brevity)
-- Example input/output pair
-
----
-
-## Test Coverage
-
-### What Exists
-
-| File | Tests | Coverage |
-|------|-------|----------|
-| `scripts/test_harvest_prompt.py` | 2 tests | Prompt file existence, sample transcript |
-| `scripts/test_harvest_prompt_comprehensive.py` | 5 tests | Prompt structure, categories, fields, confidence scoring, size limits |
-| `test_sessions/*.jsonl` | 5 sessions | Success, failure, partial, patterns, questions |
-
-### What's Missing
-
-1. **Harvester integration test** — Does the prompt actually extract correct knowledge from real transcripts?
-2. **Bootstrapper test** — Does it assemble relevant context correctly?
-3. **Knowledge store test** — Does the index.json maintain consistency?
-4. **Confidence calibration test** — Do high-confidence facts actually prove true in later sessions?
-5. **Deduplication test** — Are duplicate facts across sessions handled?
-6. **Staleness test** — How does the system handle outdated knowledge?
-
----
-
-## Security Considerations
-
-1. **No secrets in knowledge store** — The harvester must filter out API keys, tokens, and credentials from extracted facts. The prompt constraints mention this but there is no automated guard.
-
-2. **Knowledge poisoning** — A malicious or corrupted session could inject false facts. Confidence scoring partially mitigates this, but there is no verification step.
-
-3. **Access control** — The knowledge store has no access control. Any process that can read the directory can read all facts. In a multi-tenant setup, this is a concern.
-
-4. **Transcript privacy** — Session transcripts may contain user data. The harvester must not extract personally identifiable information into the knowledge store.
-
----
-
-## The 100x Path (from README)
-
-```
-Month 1:  15,000 facts, sessions 20% faster
-Month 2:  45,000 facts, sessions 40% faster, first-try success up 30%
-Month 3:  90,000 facts, fleet measurably smarter per token
-```
-
-Each new session is better than the last. The intelligence compounds.
-
----
-
-*Generated by codebase-genome pipeline. Ref: timmy-home#676.*

scripts/session_pair_harvester.py

@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+"""
+Session Transcript → Training Pair Harvester
+
+Scans Hermes session JSONL files for Q&A patterns and extracts
+terse→rich training pairs. Outputs JSONL matching the timmy-config
+training pairs spec.
+
+Usage:
+  python3 scripts/session_pair_harvester.py ~/.hermes/sessions/
+  python3 scripts/session_pair_harvester.py session.jsonl --output pairs.jsonl
+  python3 scripts/session_pair_harvester.py --dir ~/.hermes/sessions/ --min-ratio 2.0
+
+Output format:
+  {"terse": "user short prompt", "rich": "ai detailed response", "source": "session_id", "model": "..."}
+"""
+
+import argparse
+import hashlib
+import json
+import sys
+from pathlib import Path
+from typing import Optional
+
+
+def compute_hash(text: str) -> str:
+    """Content hash for deduplication."""
+    return hashlib.sha256(text.encode()).hexdigest()[:16]
+
+
+def extract_pairs_from_session(session_data: dict, min_ratio: float = 1.5,
+                                min_response_words: int = 20) -> list:
+    """Extract terse→rich pairs from a single session object."""
+    pairs = []
+    conversations = session_data.get("conversations", [])
+    session_id = session_data.get("id", "unknown")
+    model = session_data.get("model", "unknown")
+
+    seen_hashes = set()
+
+    for i, msg in enumerate(conversations):
+        # Look for assistant/gpt responses
+        if msg.get("from") not in ("gpt", "assistant"):
+            continue
+
+        response_text = msg.get("value", "")
+        if not response_text or len(response_text.split()) < min_response_words:
+            continue
+
+        # Find the preceding human message
+        prompt_text = ""
+        for j in range(i - 1, -1, -1):
+            if conversations[j].get("from") == "human":
+                prompt_text = conversations[j].get("value", "")
+                break
+
+        if not prompt_text:
+            continue
+
+        # Filter: skip tool results, system messages embedded as human
+        if prompt_text.startswith("{") and "output" in prompt_text[:100]:
+            continue  # likely a tool result
+        if prompt_text.startswith("# SOUL.md") or prompt_text.startswith("You are"):
+            continue  # system prompt leak
+
+        # Quality filters
+        prompt_words = len(prompt_text.split())
+        response_words = len(response_text.split())
+
+        # Must have meaningful length ratio
+        if prompt_words == 0 or response_words == 0:
+            continue
+        ratio = response_words / prompt_words
+        if ratio < min_ratio:
+            continue
+
+        # Skip responses that are mostly code
+        code_blocks = response_text.count("```")
+        if code_blocks >= 4 and len(response_text.replace("```", "").strip()) < 50:
+            continue
+
+        # Skip responses with tool call artifacts
+        if "tool_call" in response_text[:100] or "function_call" in response_text[:100]:
+            continue
+
+        # Deduplicate by content hash
+        content_hash = compute_hash(prompt_text + response_text[:200])
+        if content_hash in seen_hashes:
+            continue
+        seen_hashes.add(content_hash)
+
+        # Clean up response: remove markdown headers if too many
+        clean_response = response_text
+
+        pairs.append({
+            "terse": prompt_text.strip(),
+            "rich": clean_response.strip(),
+            "source": session_id,
+            "model": model,
+            "prompt_words": prompt_words,
+            "response_words": response_words,
+            "ratio": round(ratio, 2),
+        })
+
+    return pairs
+
+
+def extract_from_jsonl_file(filepath: str, **kwargs) -> list:
+    """Extract pairs from a session JSONL file."""
+    pairs = []
+    path = Path(filepath)
+
+    if not path.exists():
+        print(f"Warning: {filepath} not found", file=sys.stderr)
+        return pairs
+
+    content = path.read_text()
+    lines = content.strip().split("\n")
+
+    for line in lines:
+        line = line.strip()
+        if not line:
+            continue
+        try:
+            session = json.loads(line)
+        except json.JSONDecodeError:
+            continue
+
+        session_pairs = extract_pairs_from_session(session, **kwargs)
+        pairs.extend(session_pairs)
+
+    return pairs
+
+
+def deduplicate_pairs(pairs: list) -> list:
+    """Remove duplicate pairs across files."""
+    seen = set()
+    unique = []
+    for pair in pairs:
+        key = compute_hash(pair["terse"] + pair["rich"][:200])
+        if key not in seen:
+            seen.add(key)
+            unique.append(pair)
+    return unique
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Harvest training pairs from session transcripts")
+    parser.add_argument("input", nargs="?", help="Session JSONL file or directory")
+    parser.add_argument("--dir", "-d", help="Directory to scan for session files")
+    parser.add_argument("--output", "-o", default="harvested_pairs.jsonl", help="Output file")
+    parser.add_argument("--min-ratio", type=float, default=1.5, help="Min response/prompt word ratio")
+    parser.add_argument("--min-words", type=int, default=20, help="Min response word count")
+    parser.add_argument("--dry-run", action="store_true", help="Print stats without writing")
+    args = parser.parse_args()
+
+    all_pairs = []
+    files_scanned = 0
+
+    scan_dir = args.dir or args.input
+    if not scan_dir:
+        parser.print_help()
+        sys.exit(1)
+
+    scan_path = Path(scan_dir)
+    if scan_path.is_dir():
+        jsonl_files = sorted(scan_path.rglob("*.jsonl"))
+        print(f"Scanning {len(jsonl_files)} files in {scan_dir}...", file=sys.stderr)
+        for fpath in jsonl_files:
+            pairs = extract_from_jsonl_file(
+                str(fpath),
+                min_ratio=args.min_ratio,
+                min_response_words=args.min_words
+            )
+            all_pairs.extend(pairs)
+            files_scanned += 1
+    else:
+        pairs = extract_from_jsonl_file(
+            str(scan_path),
+            min_ratio=args.min_ratio,
+            min_response_words=args.min_words
+        )
+        all_pairs.extend(pairs)
+        files_scanned = 1
+
+    # Deduplicate
+    unique_pairs = deduplicate_pairs(all_pairs)
+
+    # Stats
+    if unique_pairs:
+        avg_prompt = sum(p["prompt_words"] for p in unique_pairs) / len(unique_pairs)
+        avg_response = sum(p["response_words"] for p in unique_pairs) / len(unique_pairs)
+        avg_ratio = sum(p["ratio"] for p in unique_pairs) / len(unique_pairs)
+    else:
+        avg_prompt = avg_response = avg_ratio = 0
+
+    stats = {
+        "files_scanned": files_scanned,
+        "raw_pairs": len(all_pairs),
+        "unique_pairs": len(unique_pairs),
+        "duplicates_removed": len(all_pairs) - len(unique_pairs),
+        "avg_prompt_words": round(avg_prompt, 1),
+        "avg_response_words": round(avg_response, 1),
+        "avg_ratio": round(avg_ratio, 2),
+    }
+
+    print(json.dumps(stats, indent=2), file=sys.stderr)
+
+    if args.dry_run:
+        # Print sample pairs
+        for pair in unique_pairs[:3]:
+            print(f"\n--- Source: {pair['source']} (ratio: {pair['ratio']}) ---", file=sys.stderr)
+            print(f"TERSE: {pair['terse'][:100]}...", file=sys.stderr)
+            print(f"RICH: {pair['rich'][:150]}...", file=sys.stderr)
+        return
+
+    # Write output
+    output_path = Path(args.output)
+    with open(output_path, "w") as f:
+        for pair in unique_pairs:
+            # Strip internal fields for output
+            output = {
+                "terse": pair["terse"],
+                "rich": pair["rich"],
+                "source": pair["source"],
+                "model": pair["model"],
+            }
+            f.write(json.dumps(output) + "\n")
+
+    print(f"\nWrote {len(unique_pairs)} pairs to {output_path}", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()

scripts/test_session_pair_harvester.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+"""Tests for session_pair_harvester."""
+
+import json
+import sys
+import os
+import tempfile
+
+sys.path.insert(0, os.path.dirname(__file__))
+from session_pair_harvester import extract_pairs_from_session, deduplicate_pairs, compute_hash
+
+
+def test_basic_extraction():
+    session = {
+        "id": "test_001",
+        "model": "test-model",
+        "conversations": [
+            {"from": "system", "value": "You are helpful."},
+            {"from": "human", "value": "What is Python?"},
+            {"from": "gpt", "value": "Python is a high-level programming language known for its readability and versatility. It supports multiple paradigms including procedural, object-oriented, and functional programming. Python is widely used in web development, data science, machine learning, and automation."},
+        ]
+    }
+    pairs = extract_pairs_from_session(session, min_ratio=1.5, min_response_words=10)
+    assert len(pairs) == 1
+    assert pairs[0]["terse"] == "What is Python?"
+    assert "programming language" in pairs[0]["rich"]
+    assert pairs[0]["source"] == "test_001"
+    print("PASS: test_basic_extraction")
+
+
+def test_filters_short_responses():
+    session = {
+        "id": "test_002",
+        "model": "test",
+        "conversations": [
+            {"from": "human", "value": "Hi"},
+            {"from": "gpt", "value": "Hello!"},
+        ]
+    }
+    pairs = extract_pairs_from_session(session, min_ratio=1.5, min_response_words=20)
+    assert len(pairs) == 0
+    print("PASS: test_filters_short_responses")
+
+
+def test_skips_tool_results():
+    session = {
+        "id": "test_003",
+        "model": "test",
+        "conversations": [
+            {"from": "human", "value": '{"output": "file content", "exit_code": 0}'},
+            {"from": "gpt", "value": "The file was read successfully. Now let me analyze the content and provide a detailed summary of what was found in the file system."},
+        ]
+    }
+    pairs = extract_pairs_from_session(session, min_ratio=1.5, min_response_words=10)
+    assert len(pairs) == 0
+    print("PASS: test_skips_tool_results")
+
+
+def test_deduplication():
+    pairs = [
+        {"terse": "What is X?", "rich": "X is Y.", "source": "s1", "model": "m"},
+        {"terse": "What is X?", "rich": "X is Y.", "source": "s2", "model": "m"},
+        {"terse": "What is Z?", "rich": "Z is W.", "source": "s1", "model": "m"},
+    ]
+    unique = deduplicate_pairs(pairs)
+    assert len(unique) == 2
+    print("PASS: test_deduplication")
+
+
+def test_ratio_filter():
+    session = {
+        "id": "test_005",
+        "model": "test",
+        "conversations": [
+            {"from": "human", "value": "Explain quantum computing in detail with examples and applications"},
+            {"from": "gpt", "value": "OK."},
+        ]
+    }
+    pairs = extract_pairs_from_session(session, min_ratio=1.5, min_response_words=10)
+    assert len(pairs) == 0  # response too short relative to prompt
+    print("PASS: test_ratio_filter")
+
+
+if __name__ == "__main__":
+    test_basic_extraction()
+    test_filters_short_responses()
+    test_skips_tool_results()
+    test_deduplication()
+    test_ratio_filter()
+    print("\nAll tests passed.")