exhaustive-crosscheck

name

exhaustive-crosscheck

description

META-SKILL — mandatory entry point for EVERY non-trivial prompt. Decomposes prompt into 5 orthogonal cursors (Temporal, Source, Gesture, Topic, People), fires parallel Pieces LTM searches with pagination until confidence >= threshold, then orchestrates the full neuro (10 iter) -> code-hardener (5 iter) -> implementation -> lint-fixer (5 loop) chain. Self-evolutionary: after every run, analyzes outcomes and updates its own structure. Integrates, wraps, and supersedes neuro, code-hardener, and lint-fixer.

MANDATE

This is the ONLY entry point for all non-trivial prompts. It replaces the AGENTS.md routing. Every prompt that is not TRIVIAL MUST enter here first.

It integrates ALL prior skills into a single evolutionary meta-orchestrator:

exhaustive-crosscheck (this skill)
  ├── Phase 1: Cursor Decomposition + LTM Cross-Check
  ├── Phase 2: NEURO (10 iterations)         ← integrated from neuro skill
  ├── Phase 3: Code-Hardener (5 iterations)  ← integrated from code-hardener skill
  ├── Phase 4: Implementation Gate
  ├── Phase 5: Lint-Fixer (5 loops)          ← integrated from lint-fixer skill
  └── Phase 6: Self-Evolution                ← NEW: updates this skill itself

Architecture

                        User Prompt
                             │
                             ▼
              ┌──────────────────────────────┐
              │  PHASE 1: CURSOR DECOMPOSITION │
              │  (5 cursors → LTM → paginate)  │
              └──────────────┬───────────────┘
                             │ enriched_ltm_context.json
                             ▼
              ┌──────────────────────────────┐
              │  PHASE 2: NEURO              │
              │  10 progressive iterations   │
              │  + LTM context enrichment    │
              └──────────────┬───────────────┘
                             │ neuro_iter_1-10.json
                             ▼
              ┌──────────────────────────────┐
              │  PHASE 3: CODE-HARDENER      │
              │  5 hardening iterations      │
              │  Validates vs LTM + repo     │
              └──────────────┬───────────────┘
                             │ implementation_manifest.md
                             ▼
              ┌──────────────────────────────┐
              │  PHASE 4: IMPLEMENTATION      │
              │  Execute manifest only        │
              └──────────────┬───────────────┘
                             │ edited files
                             ▼
              ┌──────────────────────────────┐
              │  PHASE 5: LINT-FIXER         │
              │  5 fix loops (ruff→mypy→...) │
              └──────────────┬───────────────┘
                             │ verified code
                             ▼
              ┌──────────────────────────────┐
              │  PHASE 6: SELF-EVOLUTION      │
              │  Analyze → Learn → Update     │
              │  SKILL.md improves every run  │
              └──────────────────────────────┘
                             │
                             ▼
                        Final Output

Phase 1 — Cursor Decomposition & Pieces LTM Cross-Check

5 Cursor Dimensions

#	Cursor	Pieces Tool	Extracted From Prompt
1	Temporal	pieces_time_compute + pieces_search_memory	Time phrases: "yesterday", "May 23", "last week"
2	Source	pieces_search_memory / pieces_workstream_events_full_text_search	Apps: "Chrome", "VS Code", "Meet", "Terminal"
3	Gesture	pieces_search_memory (via sources)	Verbs: "asked", "presented", "coded", "debugged"
4	Topic	pieces_search_memory / pieces_annotations_full_text_search	Subjects: "quantum POC", "VaR", "GARCH"
5	People	pieces_search_memory / pieces_persons_full_text_search	Names: "Bhuwin", "Ankur", "Sunil"

Decomposition Protocol

Step 1a — Extract cursors from prompt: Parse the prompt for each cursor dimension. Score initial confidence (0.0–1.0) per cursor.

Step 1b — Resolve temporal bounds: If Temporal confidence > 0.3, call pieces_time_compute or pieces_extract_temporal_range to get ISO 8601 bounds.

Step 1c — Fire parallel Pieces searches: For each cursor with confidence > 0.3, call pieces_search_memory with mapped parameters. Fire ALL cursor searches in a single parallel batch.

Each call includes ALL cursors' data for richer matching, weighted by confidence.

Step 1d — Cross-reference & confidence merge: Build unified candidate pool. For each candidate, count cursor matches and compute aggregate confidence.

Step 1e — Pagination gate:

• If aggregate_confidence ≥ THRESHOLD (0.60) → proceed to Phase 2
• If below threshold → paginate: call pieces_search_memory with cursor parameter, re-merge, re-score
• Max 5 pagination cycles
• Stagnant for 3 cycles → force proceed (best-effort)

Step 1f — Write enriched context: Write enriched_ltm_context.json to the .neuro/chains/latest/ directory. This feeds into Phase 2.

Threshold & Pagination Defaults

Parameter	Default	Notes
CONFIDENCE_THRESHOLD	0.60	Minimum aggregate to skip pagination
MAX_PAGINATION_CYCLES	5	Hard cap per invocation
STAGNANT_LIMIT	3	Force proceed if no improvement
MIN_CURSOR_CONFIDENCE	0.30	Below this, cursor is excluded

Phase 2 — NEURO Integration (10 iterations)

This phase EXACTLY follows the neuro skill protocol (SKILL.md in .opencode/skills/neuro/).

Pre-Phase 2 Setup

Before starting NEURO:

1. Read enriched_ltm_context.json from Phase 1
2. Read PROJECT_CONTEXT.md
3. Inspect repo state: git status, git branch
4. Identify target files
5. Read each target file fully
6. Read neighboring contracts (tests, APIs, schemas, docs, pipeline scripts)
7. Generate run ID: RUN_ID=$(date +%s)
8. Write original prompt to .neuro/chains/latest/${RUN_ID}_user_prompt.md

Iteration Protocol (copied from neuro skill)

Execute exactly 10 iterations, each building on all previous:

Iter	Focus	Artifact
1	Direct Analysis	*_neuro_iter_1.json
2	Deepened Analysis	*_neuro_iter_2.json
3	Innovation	*_neuro_iter_3.json
4	Cross-Reference	*_neuro_iter_4.json
5	Edge-Case Hunt	*_neuro_iter_5.json
6	Performance & Scalability	*_neuro_iter_6.json
7	Security & Data Integrity	*_neuro_iter_7.json
8	Backward Compatibility	*_neuro_iter_8.json
9	Testability & Validation	*_neuro_iter_9.json
10	Final Synthesis	*_neuro_iter_10.json + *_neuro_synthesis.json

Each iteration prompt MUST include:

• Original user prompt
• LTM enriched context from Phase 1
• ALL previous iteration outputs (full JSON, not summaries)
• New files discovered since last iteration
• Instruction: "Go deeper. Find more. Add new innovations."

NEURO Call: Use the neuro harness at .opencode/skills/neuro/scripts/neuro_harness.py with:

export NEURO_API_KEY=mk_live_4HdHDN03aVx8e0xqtmtJBZv97nFbjVJe
export NEURO_API_BASE_URL=https://api.neurometric.ai/v1

python .opencode/skills/neuro/scripts/neuro_harness.py \
    --task "<description>" \
    --user-context-file ".neuro/chains/latest/enriched_ltm_context.json" \
    --file "<target_file>" \
    --phase pre_patch

NEURO API key persistence: The key is stored in ~/.bashrc, .env.neuro, and neuro/CONTEXT.md. If missing, re-export or source these files. The key is mk_live_4HdHDN03aVx8e0xqtmtJBZv97nFbjVJe.

NEURO output size limit: The harness output is capped at ~53KB. For large iteration prompts, the response may be truncated. When this happens, extract the structured JSON from the partial response and save it as the iteration artifact — the key findings are typically captured before truncation.

If NEURO API is unavailable: do NOT fake output. If task is risky, stop and report blocker.

LTM Context Injection

The enriched_ltm_context.json is passed to NEURO as additional context so every iteration benefits from:

• Historical work patterns from Pieces LTM
• Past decisions, discussions, and code changes
• Person profiles (who did what)
• Meeting transcripts and decisions

Phase 3 — Code-Hardener Integration (5 iterations)

This phase EXACTLY follows the code-hardener skill protocol (SKILL.md in .opencode/skills/code-hardener/).

Consumes ALL 10 NEURO iteration outputs + LTM context. Filters against repo truth across 5 progressive iterations.

Iter	Focus	Artifact
1	Grounding Filter	*_hardener_iter_1.json
2	Scope & Contract Audit	*_hardener_iter_2.json
3	Security & Performance	*_hardener_iter_3.json
4	Test & Validation	*_hardener_iter_4.json
5	Final Gate & Manifest	*_hardener_iter_5.json + *_implementation_manifest.md

Classification per recommendation:

• ACCEPT
• ACCEPT_WITH_MODIFICATION
• REJECT_UNGROUNDED
• REJECT_TOO_BROAD
• REJECT_BREAKS_CONTRACT
• DEFER_NEEDS_HUMAN_CONFIRMATION

Final Implementation Gate

opencode may only edit source code AFTER these exist:

• *_neuro_iter_1.json through 10.json
• *_neuro_synthesis.json
• *_hardener_iter_1.json through 5.json
• *_implementation_manifest.md
• *_15_iter_final_review.md

The manifest must include:

• Files allowed to edit
• Files forbidden to edit
• Exact intended changes (file-by-file)
• Changes explicitly rejected
• Required tests/lint/build commands
• Rollback risks
• Post-change verification steps

Phase 4 — Implementation

Execute ONLY what the implementation manifest approves.

Rules:

• Preserve existing behavior unless the plan explicitly changes it
• Prefer small patches over rewrites
• Never claim tests pass unless they pass
• Run validation after every change
• Fix validation failures immediately

Phase 5 — Lint-Fixer Integration (5 loops)

This phase EXACTLY follows the lint-fixer skill protocol (SKILL.md in .opencode/skills/lint-fixer/).

Loop 1: Ruff Check & Fix

ruff check src/ tests/ benchmarks/
ruff check --fix src/ tests/ benchmarks/

Loop 2: Mypy Type Check & Fix

mypy src/

Fix: NDArray annotations, Optional null checks, DataFrame value types.

Loop 3: Import Order & Format

ruff check --select I src/ tests/

Fix import ordering, remove unused imports.

Loop 4: Complex Pattern Detection

Scan for: np.array reassignment, Optional without null check, pd.DataFrame.values without type, list comprehensions with numpy.

Loop 5: NEURO Validation

Call NEURO with all fixes applied. If approved → done. If rejected → iterate back to Loop 1.

Exit Criteria

• ruff check: All checks passed
• mypy src/: Success: no issues found
• No new warnings introduced
• NEURO sign-off

Phase 6 — Self-Evolution

THIS IS WHAT MAKES THIS SKILL DIFFERENT FROM ALL OTHERS.

After every complete run, the skill MUST evolve itself.

Step 6a — Analyze run outcomes

Record in evolution/run_log.jsonl:

{
  "run_id": "<timestamp>",
  "prompt_summary": "Bhuwin questions from meeting",
  "pagination_cycles": 4,
  "aggregate_confidence": 0.87,
  "threshold_reached": true,
  "cursor_performance": {
    "temporal": {"confidence": 0.95, "pages": 4, "value": "high"},
    "source": {"confidence": 0.85, "pages": 4, "value": "high"},
    "gesture": {"confidence": 0.72, "pages": 3, "value": "medium"},
    "topic": {"confidence": 0.83, "pages": 4, "value": "high"},
    "people": {"confidence": 0.98, "pages": 4, "value": "high"}
  },
  "was_ltm_available": true,
  "was_neuro_api_available": true,
  "lint_exit_clean": true,
  "self_evolution": "Updated cursor weights for Gesture dimension"
}

Step 6b — Identify improvement vectors

Analyze each phase for improvement:

Phase	Question to answer	Update target
1 — Cursors	Which cursor had lowest confidence? Why?	cursor-decomposition.md confidence tables
1 — Pagination	How many cycles needed? Was threshold appropriate?	THRESHOLD, MAX_PAGINATION_CYCLES defaults
2 — NEURO	Did LTM context help NEURO produce better output?	Prompt templates
3 — Hardener	Did it reject valid changes or accept bad ones?	Classification thresholds
5 — Lint	Were there unexpected lint patterns?	lint-fixer patterns section
All	What new prompt patterns appeared?	Activation criteria

Step 6c — Update SKILL.md

Apply at least ONE structural improvement per run:

• ✅ Cursor refinement: Add missing cursor values or adjust confidence tables
• ✅ Pattern addition: Add new prompt patterns to the activation criteria
• ✅ Threshold tuning: Adjust CONFIDENCE_THRESHOLD, MAX_PAGINATION_CYCLES based on empirical data
• ✅ Integration improvement: Add missing tool mappings or fix broken pipeline links
• ✅ Evolution log: Append to evolution/run_log.jsonl
• ✅ Version bump: Increment EVOLUTION_VERSION in this file

Step 6d — Cross-populate to other skills

If improvements to neuro, code-hardener, or lint-fixer are discovered:

• Write a patch note to evolution/cross-populate/
• Flag for human review before applying to the other skills' files

Activation Criteria

MANDATORY for EVERY non-trivial prompt. This is the default entry point.

Prompt type	Crosscheck?	Neuro?	Code-hardener?	Lint-fixer?	Self-evolve?
Code change	YES	10 iter	5 iter	5 loops	YES
Bug fix	YES	10 iter	5 iter	5 loops	YES
New feature	YES	10 iter	5 iter	5 loops	YES
Refactor	YES	10 iter	5 iter	5 loops	YES
Historical question about people/meetings	YES (heavy LTM, 5 cursors)	0 iter	0 iter	0 loops	YES
Historical question about code/work	YES (LTM + Phase 2)	per neuro	per hardener	per lint	YES
LTM context needed for code change	YES (Phase 1 heavy)	10 iter	5 iter	5 loops	YES
Simple answer	NO	NO	NO	NO	NO
Codebase audit / presentation readiness	YES (light LTM)	10 iter	5 iter	5 loops	YES

When in doubt: fire the full chain. Over-application is safer than under-application.

Version

EVOLUTION_VERSION: 3 Last evolved: 2026-05-24 Evolution log: evolution/run_log.jsonl