training/build_4b_v43_patch4.py

#!/usr/bin/env python3
"""
Patch4 corpus for prism-coder:4b-v43.
Targets swe-bench failures not fixable by Layer 3:
  - Group C: session_task_route false negatives (implicit routing questions)
  - Group D: param extraction from natural/informal phrasing

Output format: {"text": "<ChatML string>"} to match existing train.jsonl
"""
import json, pathlib

OUT = pathlib.Path("/tmp/4b_v43_patch4.jsonl")

SYS = (
    "You are Synalux, a memory-augmented coding and clinical reasoning assistant. "
    "You have access to Prism Memory tools (session_save_ledger, session_load_context, "
    "session_search_memory, session_save_handoff, session_forget_memory, session_health_check, "
    "session_compact_ledger, session_export_memory, session_task_route, session_save_experience, "
    "session_synthesize_edges, session_backfill_links, knowledge_search, knowledge_forget, "
    "knowledge_upvote, knowledge_downvote, knowledge_set_retention) and 13 multimodal tool "
    "modules (image_gen, office, web_scraper, browser, tts, ocr, git, terminal, deps_scanner, "
    "hipaa, data_graph, templates, pdf_parser). "
    "Think step-by-step before answering. When the user references past work, prior decisions, "
    "or stored context, use the appropriate Prism Memory tool. "
    "Format tool calls inside <tool_call>...</tool_call> JSON blocks with fields 'name' and 'arguments'. "
    "If no tool is needed, answer directly in plain text. "
    "ABSTAIN for general programming questions, CS concepts, greetings, and capability questions."
)

THINK_OPEN = "<|synalux_think|>"
THINK_CLOSE = "</|synalux_think|>"


def render_chatml(user_text, assistant_text):
    return (
        f"<|im_start|>system\n{SYS}<|im_end|>\n"
        f"<|im_start|>user\n{user_text}<|im_end|>\n"
        f"<|im_start|>assistant\n{assistant_text}<|im_end|>"
    )


def msg(user, tool_name, args):
    call = json.dumps({"name": tool_name, "arguments": args})
    think = f"{THINK_OPEN}The user wants {tool_name.replace('_', ' ')}. I'll call the tool with the right parameters.{THINK_CLOSE}"
    assistant = f"{think}\n<tool_call>\n{call}\n</tool_call>"
    return {"text": render_chatml(user, assistant)}


def no_tool_msg(user, reply):
    think = f"{THINK_OPEN}This is a general programming/CS question. I should answer directly without calling any Prism tool.{THINK_CLOSE}"
    assistant = f"{think}\n{reply}"
    return {"text": render_chatml(user, assistant)}


rows = []

# ── GROUP C: session_task_route — implicit "should I/local/cloud?" questions ──
task_route_cases = [
    ("Should I handle this CSS grid refactor myself or punt it to the cloud model?",
     {"task_description": "CSS grid refactor", "prefer_local": False}),
    ("Is this bug fix simple enough for the local model to tackle?",
     {"task_description": "bug fix"}),
    ("Can the on-device model handle parsing this 50MB XML dump, or should I escalate to cloud?",
     {"task_description": "parsing 50MB XML dump", "prefer_local": False}),
    ("Should I let the local model handle this TypeScript refactor or route it to the bigger model?",
     {"task_description": "TypeScript refactor", "prefer_local": False}),
    ("Is writing a full test suite for the billing module something the local model can do well?",
     {"task_description": "writing test suite for billing module"}),
    ("This SQL query optimization seems complex — local or cloud?",
     {"task_description": "SQL query optimization"}),
    ("Route this: rewriting the authentication middleware from scratch.",
     {"task_description": "rewriting authentication middleware from scratch"}),
    ("Should the local agent handle this performance profiling task?",
     {"task_description": "performance profiling task"}),
    ("Is a full-text semantic search implementation doable on-device?",
     {"task_description": "full-text semantic search implementation"}),
    ("Local or cloud for this multi-file refactor of the payments module?",
     {"task_description": "multi-file refactor of payments module", "prefer_local": False}),
    ("Can the on-device model handle a 3000-line code review?",
     {"task_description": "3000-line code review"}),
    ("Route this debugging task: tracing intermittent race conditions in async code.",
     {"task_description": "tracing intermittent race conditions in async code"}),
    ("Should I send this architecture review to the cloud model or keep it local?",
     {"task_description": "architecture review", "prefer_local": False}),
    ("Is this data pipeline migration too complex for local inference?",
     {"task_description": "data pipeline migration"}),
    ("Handle or escalate? — writing API documentation for 40 endpoints.",
     {"task_description": "writing API documentation for 40 endpoints"}),
]
for user, args in task_route_cases:
    rows.append(msg(user, "session_task_route", args))

print(f"After GROUP C (task_route): {len(rows)} rows")

# ── GROUP D1: session_export_memory — extract output_dir from casual phrasing ──
export_cases = [
    ("Dump everything to /tmp/backup so I can archive it. JSON format.",
     {"output_path": "/tmp/backup", "format": "json"}),
    ("Back up all my memories to /var/exports before I wipe the session.",
     {"output_path": "/var/exports"}),
    ("Export the prism-mcp project memory to /tmp/prism-export in markdown format.",
     {"project": "prism-mcp", "output_path": "/tmp/prism-export", "format": "markdown"}),
    ("I want to export a backup and then compact the old entries — save it to /tmp/mem-backup.",
     {"output_path": "/tmp/mem-backup"}),
    ("Export everything from the billing project to /tmp/billing-backup.json.",
     {"project": "billing", "output_path": "/tmp/billing-backup.json"}),
    ("Can you dump the analytics project sessions to /home/user/exports/analytics?",
     {"project": "analytics", "output_path": "/home/user/exports/analytics"}),
    ("Save a memory backup to /tmp/export before we start the migration.",
     {"output_path": "/tmp/export"}),
    ("Archive the prism-training project to /tmp/training-archive in JSON.",
     {"project": "prism-training", "output_path": "/tmp/training-archive", "format": "json"}),
]
for user, args in export_cases:
    rows.append(msg(user, "session_export_memory", args))

print(f"After GROUP D1 (export): {len(rows)} rows")

# ── GROUP D2: session_forget_memory — extract memory_id from prompt ──
forget_cases = [
    ("Delete the specific memory entry with ID mem-abc-123.",
     {"memory_id": "mem-abc-123"}),
    ("Remove memory entry 7f3a-bc21-d4e5 — it's completely wrong.",
     {"memory_id": "7f3a-bc21-d4e5"}),
    ("Get rid of entry ID: ent-2024-991. It was saved by mistake.",
     {"memory_id": "ent-2024-991"}),
    ("Forget memory ID mem-portal-007.",
     {"memory_id": "mem-portal-007"}),
    ("Delete that specific memory entry with ID 'sess-42-bad'.",
     {"memory_id": "sess-42-bad"}),
    ("The entry tagged as 'exp-001' is outdated. Please remove it.",
     {"memory_id": "exp-001"}),
    ("Wipe memory entry ID = abc123xyz.",
     {"memory_id": "abc123xyz"}),
    ("Remove the wrong entry — its ID is MEM-2025-0042.",
     {"memory_id": "MEM-2025-0042"}),
    ("That entry we saved earlier about the broken migration config — the ID is err-cfg-88, delete it.",
     {"memory_id": "err-cfg-88"}),
    ("Clear out memory ID: old-deploy-notes-19.",
     {"memory_id": "old-deploy-notes-19"}),
]
for user, args in forget_cases:
    rows.append(msg(user, "session_forget_memory", args))

print(f"After GROUP D2 (forget+id): {len(rows)} rows")

# ── GROUP D3: session_task_route — extract task_description from implicit phrasing ──
task_route_desc_cases = [
    ("Route this refactoring task — if local, proceed; if cloud, just tell me.",
     {"task_description": "refactoring task"}),
    ("Is the on-device model good enough for a full GraphQL schema migration?",
     {"task_description": "GraphQL schema migration"}),
    ("This linting fix across 200 files — local or cloud?",
     {"task_description": "linting fix across 200 files"}),
    ("Route: generating embeddings for 10k documents.",
     {"task_description": "generating embeddings for 10k documents"}),
    ("Should the local model handle writing unit tests for the auth module?",
     {"task_description": "writing unit tests for auth module"}),
]
for user, args in task_route_desc_cases:
    rows.append(msg(user, "session_task_route", args))

print(f"After GROUP D3 (task_route desc): {len(rows)} rows")

# ── GROUP D4: session_save_ledger — extract project + summary from natural phrasing ──
ledger_natural_cases = [
    ("We just finished a big refactor of the billing module. Make sure it's written down. Project: billing, conv: conv-bill-2024.",
     {"project": "billing", "conversation_id": "conv-bill-2024", "summary": "Completed big refactor of billing module"}),
    ("Before I hand off, save what we did today: fixed the OAuth flow and updated tests. Project is portal.",
     {"project": "portal", "summary": "Fixed OAuth flow and updated tests"}),
    ("We're done for the day. Log what we accomplished on the analytics project.",
     {"project": "analytics", "summary": "Session progress logged"}),
    ("Jot this down for the prism-training project: we built and validated the v43 corpus.",
     {"project": "prism-training", "summary": "Built and validated v43 corpus"}),
    ("Log session for ios-app project: completed llama.cpp Swift integration.",
     {"project": "ios-app", "summary": "Completed llama.cpp Swift integration"}),
    ("Write it down — we migrated auth to OAuth2 in the portal project. conv-id: sess-99.",
     {"project": "portal", "conversation_id": "sess-99", "summary": "Migrated auth to OAuth2"}),
    ("End of session for prism-mcp. We deployed v3 MCP server and ran smoke tests.",
     {"project": "prism-mcp", "summary": "Deployed v3 MCP server and ran smoke tests"}),
    ("Record this session: we fixed the race condition in the payments module. project=payments.",
     {"project": "payments", "summary": "Fixed race condition in payments module"}),
]
for user, args in ledger_natural_cases:
    rows.append(msg(user, "session_save_ledger", args))

print(f"After GROUP D4 (ledger natural): {len(rows)} rows")

# ── GROUP D5: backfill_links + synthesize_edges — verifier distinctions ──
verifier_cases = [
    ("Backfill the missing cross-session links for the analytics project.",
     "session_backfill_links", {"project": "analytics"}),
    ("Reconnect the dangling session references for the billing project.",
     "session_backfill_links", {"project": "billing"}),
    ("Patch up the link gaps in our session history for prism-training.",
     "session_backfill_links", {"project": "prism-training"}),
    ("Fix the broken links between sessions in the portal project.",
     "session_backfill_links", {"project": "portal"}),
    ("There are dangling references in ios-app sessions. Backfill them.",
     "session_backfill_links", {"project": "ios-app"}),
    ("Reconnect missing session cross-references for prism-mcp.",
     "session_backfill_links", {"project": "prism-mcp"}),
    ("Run a synthesis pass on the prism-mcp project to make sure all edges are up to date.",
     "session_synthesize_edges", {"project": "prism-mcp"}),
    ("Before we close out, verify all the session links are consistent for the portal project.",
     "session_synthesize_edges", {"project": "portal"}),
    ("Synthesize the session graph for analytics to build fresh edge connections.",
     "session_synthesize_edges", {"project": "analytics"}),
    ("Verify graph integrity — synthesize edges for the ios-app project.",
     "session_synthesize_edges", {"project": "ios-app"}),
    ("Do a graph synthesis on the billing sessions to update the relationship map.",
     "session_synthesize_edges", {"project": "billing"}),
    ("Build the session edge graph for prism-training.",
     "session_synthesize_edges", {"project": "prism-training"}),
]
for user, tool, args in verifier_cases:
    rows.append(msg(user, tool, args))

print(f"After GROUP D5 (verifier): {len(rows)} rows")

# ── GROUP D6: session_search_memory — "remind me / did we ever decide" phrasing ──
remind_search_cases = [
    ("Remind me — did we ever decide between Redis and Memcached for caching?",
     {"query": "Redis vs Memcached caching decision"}),
    ("What did we decide about the database schema for the billing module?",
     {"query": "database schema billing module decision"}),
    ("Did we settle on which logging library to use for the portal?",
     {"query": "logging library portal decision"}),
    ("Remind me what we concluded about microservices vs monolith.",
     {"query": "microservices vs monolith conclusion"}),
    ("What was our decision on the auth token expiry?",
     {"query": "auth token expiry decision"}),
    ("Did we ever document our stance on using TypeScript strict mode?",
     {"query": "TypeScript strict mode decision"}),
    ("Can you remind me what we figured out about the WebSocket reconnection strategy?",
     {"query": "WebSocket reconnection strategy"}),
    ("What did we decide to do about the rate limiting approach?",
     {"query": "rate limiting approach decision"}),
    ("Remind me of our agreed-on approach for handling API errors.",
     {"query": "API error handling approach"}),
    ("Did we ever land on a pagination strategy for the search API?",
     {"query": "pagination strategy search API"}),
]
for user, args in remind_search_cases:
    rows.append(msg(user, "session_search_memory", args))

print(f"After GROUP D6 (remind/search): {len(rows)} rows")

# Write output
with OUT.open("w") as f:
    for row in rows:
        f.write(json.dumps(row) + "\n")

print(f"\nWrote {len(rows)} patch4 rows to {OUT}")

# Verify format
with OUT.open() as f:
    first = json.loads(f.readline())
print(f"First row keys: {list(first.keys())}")
print(f"Text starts with: {first['text'][:50]}")
print(f"Think close tag present: {'</|synalux_think|>' in first['text']}")
print(f"Bad think close absent: {'</{' not in first['text']}")