skills/bounded-repo-executor/SKILL.md
Bounded execution skill for Work Order lanes requiring minimal diff, exact proof, and closeout or verification against repo/main truth without scope widening or design detours.
npx skillsauth add ic1558/0luka bounded-repo-executorInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Use this skill when the task is already bounded by a Work Order or explicit lane definition and the job is to execute, verify, close out, or report repo truth without widening scope.
This skill is for disciplined execution.
It is not for broad design or architecture invention.
Take a bounded repo task and drive it to one of these honest outcomes:
The result must be grounded in:
No inflated claim is allowed.
Before using this skill on a governed 0luka lane:
/Users/icmini/0luka/core_brain/governance/AGENTS.yaml/Users/icmini/0luka/CONTRACT.yaml/Users/icmini/0luka/skills/manifest.mdIf a lane-specific WO or protocol exists, read that before taking action.
Do not begin execution from memory alone.
Use this skill when:
Do not use this skill when:
stack-engineer or blueprint-specialistThink like a bounded engineer, not a strategist.
That means:
If all are true:
then default behavior is:
think -> decide -> execute
not:
think -> explain -> wait
Extract from the WO:
Rewrite scope mentally into a narrow execution boundary. Do not improvise beyond it.
Before mutation, verify:
If unsafe:
Perform only the smallest changes required to satisfy the WO.
Rules:
At minimum produce:
If proof is missing, the lane is not PASS.
Truth must distinguish:
Never claim:
If the WO includes closeout:
main if possibleAlways preserve these:
Not allowed:
A lane is PASS only if:
PARTIAL if:
FAIL if:
This skill must explicitly distinguish:
What exists only in the current working copy or branch
What exists in the PR branch but not on main
What exists on refs/heads/main or refs/remotes/origin/main
Never collapse these into one.
If main ref cannot be verified directly, say so.
Do not fake VERIFIED-ON-MAIN.
For closeout tasks, return exact disposition only:
If a prior claim was inflated, correct it directly. Do not preserve old wording for politeness.
A real blocker must be one of:
Not blockers:
Return only:
This skill is done only when:
One bug is fixed, one file changed, one proof artifact returned, PASS.
Only some paths needed updates, others were already satisfied on main, final report matches exact repo truth.
Feature is proven, but direct main-ref verification is missing, so verdict is PARTIAL not FAIL.
WO expected 5 files but branch touched 1; report is downgraded to FAIL until scope matches truth.
Do not ask:
What else can I improve while I’m here?
Ask:
What is the smallest truthful diff that satisfies this WO and how do I prove it?
End of Document
development
--- name: auto-workflow-creator description: Create a new skill from a workflow you just performed. Generates a valid 3-layer Anthropic-style skill — YAML frontmatter (name + description + triggers), imperative instruction body, and optional scripts/ subdirectory — and prints the registration commands per target runtime (Claude Code auto-discovers from ~/.claude/skills/; Codex needs an explicit [[skills.config]] entry in ~/.codex/config.toml; OpenCode needs a registry entry in ~/.opencode/openco
tools
# Cost-Control Skill Enforce token discipline and model routing optimization (Flash vs Pro) via the Gemini CLI 0.40 Gemma local router for governed-agent workflows. ## Guidelines - **Preference for Local/RTK/File Tools:** Always prioritize local tools (ls, read_file, grep_search) and RTK for initial data gathering before calling model logic. - **Flash-First Routing:** - Route all routine code analysis, file reads, summarization, search, and classification tasks to `gemini-3.5-flash-preview` o
development
--- name: verify-first description: Run verification gates before and after code changes, prioritize health and regressions, and fail closed on missing evidence. Mandatory Read: YES --- # Verify First Mandatory Read: YES ## Workflow 1. Capture baseline verify results. 2. Apply minimal changes. 3. Re-run targeted and system-level verify commands. 4. Report explicit pass/fail evidence. ## Caps - Run deterministic verification commands. - Compare before/after outcomes. - Block claims without pr
testing
Use this skill when you need to perform a READ-ONLY forensic validation of a completed task (evidence), verify compliance against a TaskSpec, or issue a formal verdict (PASS/FAIL) before GMX approval. Strictly forbids modification or execution.