code-hardener

name	code-hardener
description	Mandatory second-stage logic filter and architectural hardening skill that runs after NEURO. It validates NEURO recommendations against repo truth, calls NEURO a second time with filtered critique, then emits the only implementation plan Gemini may follow.

Purpose

code-hardener is the second skill in the mandatory hardening chain.

It consumes first-pass NEURO review output, filters it against repository truth and project context, calls NEURO again with the filtered critique, and produces the final implementation plan.

The chain is:

User prompt + repo context → neuro → code-hardener logic filter → NEURO second pass → final Code Hardener plan → Gemini implementation → validation → PROJECT_CONTEXT.md update

Activation

Activate this skill automatically:

• After any neuro skill usage
• When .neuro/reviews/*.json exists for the current task
• When the user asks for code hardening
• When the user asks for architectural improvement
• When the user asks to apply NEURO recommendations
• During non-trivial code review or implementation
• Whenever Gemini is about to edit production code after a NEURO review

Logic filter

Before asking NEURO for the second pass, check:

• Do the files NEURO mentions actually exist?
• Do the functions/classes/endpoints actually exist?
• Does the advice match current repo state?
• Does the advice conflict with PROJECT_CONTEXT.md?
• Would the advice delete or break working behavior?
• Is the patch too broad?
• Are there missing tests?
• Are there hidden data-contract changes?
• Are there security/secrets risks?
• Are there performance regressions?
• Are there UI/API compatibility risks?
• Are there generated files or caches that must not be edited?

Reject ungrounded recommendations.

Second NEURO call

The second NEURO call must include:

• Original user prompt
• Current project context
• First NEURO output
• Logic-filter critique
• Exact files under consideration
• Existing tests/validation commands
• Request for architectural improvement
• Request for minimal safe patch plan

Final implementation gate

Gemini may only edit source code after these files exist:

• .neuro/chains/latest/01_neuro_initial.json
• .neuro/chains/latest/02_logic_filter.md
• .neuro/chains/latest/03_neuro_architectural_second_pass.json
• .neuro/chains/latest/04_final_code_hardener_plan.md
• .neuro/chains/latest/05_implementation_manifest.md

The final plan must include:

• Files allowed to edit
• Files forbidden to edit
• Exact intended changes
• Changes explicitly rejected
• Required tests/lint/build commands
• Rollback risks
• Post-change verification steps
• PROJECT_CONTEXT.md update requirements

If the final plan is missing, stale, empty, or unrelated to the current task, editing is forbidden.

Implementation discipline

Gemini must:

• Implement only the final plan
• Preserve existing behavior unless the plan explicitly changes it
• Prefer small patches over rewrites
• Run validation
• Fix validation failures
• Never claim tests pass unless they pass
• Update PROJECT_CONTEXT.md

Progressive disclosure

Keep this SKILL.md concise. Detailed workflow is in:

• workflows/logic-filter.md
• workflows/two-pass-neuro-chain.md
• workflows/implementation-gate.md

Executable orchestrator is in:

• scripts/neuro_code_hardener_chain.py