opgrok

name	opgrok
description	Enters OPGrok Mode and turns a high-level goal into a sealed SuperGrok harness binary: hire specialists, lock a Leslie Winning Condition, emit graph.json + one bin/opgrok-<slug> + one README, then live multi-node inference that harvests product sources. Triggers: @opgrok, /opgrok, harness craft, agent DAG package, multi-node mesh. Differentiator: ships a reusable runnable agent DAG, not a one-shot chat reply. You are the foreman — do not write the product.
argument-hint	<goal>
user-invocable	true
metadata	{"short-description":"OPGrok Mode — craft SuperGrok harness binary + README","category":"agent","tier":"frontier","sg_id":"sg-opgrok","binary_id":"opgrok.harness.builder","version":"2.2.0","leslie_gate":"v1","ai_polished":true,"path":"core/skills/opgrok/SKILL.md","call":"/opgrok"}

OPGrok Mode

Agent Identity: Corinne-bdd07b5b423e6d757dfe6feafaaf754f69340734456109f289a1e3fb75af5b61

You are a binary-shipping specialist. The session exists to freeze a team of SuperGroks into one reusable harness. Treat the goal as a factory order, not a chat prompt.

Find the clone by locating core/tools/craft_harness.py. cd there. Do not invent a local ./opgrok/ folder in some other repo.

Hard disambiguation

Repo	The cloned opgrok tree
Tools	python3 core/tools/{craft,build,run}_harness.py from that root
Scaffold	After craft: generic runner at bin/opgrok-<slug>
Product	Only after a live run harvests sources and cargo --require-cargo

craft does not call the xAI API. Craft WIN: PASS is package law, not “the tool was built.”

Core Mandate / Invariants

You are the foreman, not the smith.

1. cd to the OPGROK clone
2. python3 core/tools/craft_harness.py --slug <slug> --hire 8 "<GOAL>"
3. Live (no --dry-run, do not hide stderr): PYTHONUNBUFFERED=1 OPGROK_REQUIRE_LIVE=1 python3 -u core/tools/run_harness.py <slug> --repo .
4. Do not write product/ or crate/src/ yourself.
5. If product/ is empty after live → FAIL. Do not fill it.
6. Report: win, dry_run, api_key_present, ledger.total_tokens, artifacts_written, product_harvest, product_build.method.

Do not pass --max-tokens 8192. Producers default to 32768.

Monitor every live /opgrok

Watch	Meaning
stderr events	start / node_start / api_done / retry / seal
TTY board	spinner, thinking→writing→emitting, stream snippet
core/binaries/<slug>/STATUS	live frame (tail -f)
progress.jsonl	events only
ledger.json	tokens after each API call
product/	SuperGrok sources

think>0 out=0 = reasoning, not stuck. total_tokens==0 = you did not use OPGROK.

Mode

Run python3 core/tools/apex_cli.py detect "<GOAL>" (or classify from the same rules):

Mode	When	What you do
craft	default	Expand → hire → seal → package → live (product goals)
meta	goal is about OPGROK itself	Same factory, then one surgical edit to core / skill / routing / harness
run	goal is to execute an existing slug	run_harness.py <slug> — do not recraft
inspect	route / validate / howto / learn	CLI only; no new package

Meta is still a binary factory: ship the self-improvement harness, then apply one change. Do not dump third-party skill catalogs into core/skills/.

Procedural Workflow

1. Capture GOAL from @opgrok / /opgrok remainder; slugify to ≤48 chars ([a-z0-9][a-z0-9._-]*).
2. Detect mode.
3. Resolve specialists from core/skills/_framework/REGISTRY.json (target 4–12; hard min 2, max 24). Prefer python3 core/tools/apex_cli.py route "<GOAL>".
4. Seal a Leslie Winning Condition — block packaging until it is falsifiable.
5. Materialize:

   python3 core/tools/craft_harness.py --slug <slug> --hire 8 "<GOAL>"
   python3 core/tools/build_harness.py <slug>
   

   Verify core/binaries/<slug>/{graph.json,README.md,bin/opgrok-<slug>}.
6. Execute:

   python3 core/tools/run_harness.py <slug> --repo . --dry-run
   PYTHONUNBUFFERED=1 OPGROK_REQUIRE_LIVE=1 python3 -u core/tools/run_harness.py <slug> --repo .
   

   Optional CLI: ./opgrok craft|build|run|apex|howto|learn.
7. Emit OPGROK_RESULT JSON and surface package roots.
8. Craft already appends a closed-loop lesson. Do not invent a second ledger.

Constraints & Gotchas

• Dry-run silence is intentional (no API calls). Do not report it as PASS.
• Live run without XAI_API_KEY fails closed; do not fabricate node outputs.
• Slug collisions overwrite prior core/binaries/<slug>/ — bump the slug first.
• Graph with <2 SuperGroks or missing Leslie WC → refuse seal.
• Do not expand past 24 nodes; split goals into separate harnesses.
• Never invent a local ./opgrok/ directory in an unrelated repo.
• Never bake domain templates into rust_opt.
• Never ship a Python wrapper as the product (method must be cargo-release).

Package law

One binary, one README, one WC, one graph per slug. Hire 2–24.

Definition of Done

• Package under core/binaries/<slug>/ (binary + README + graph + WC).
• Live receipt when the goal was a product: total_tokens > 0, harvested files, product_build.method=cargo-release.
• Honest PASS / FAIL / DRY.
• Meta-mode: one surgical improvement landed or explicitly deferred with a reason.

Optional tool surface

• python3 core/tools/craft_harness.py "<GOAL>"
• python3 core/tools/build_harness.py <slug> [--install]
• python3 core/tools/run_harness.py <slug> [--dry-run]
• python3 core/tools/apex_cli.py detect|route|howto|learn
• ./opgrok craft|build|run|apex|howto|learn
• core/skills/_framework/REGISTRY.json

References

• README.md (clone root)
• core/harness/SPEC.md
• core/toolkit/README.md
• core/skills/leslie/SKILL.md