Agent Behavior & Communication

This rule defines the foundational “working style” of the AI agent — from response format and language configuration to memory recovery after long conversations cause context truncation.

1. Language Configuration

• MUST respect preferences.language in .para-workspace.yml for docs and chat.
• MUST keep technical artifacts (code variables, commit messages) in English.
• SHOULD adapt communication language to user’s configured preference.

2. Communication Style

• MUST be concise — focus on solutions, avoid fluff.
• SHOULD use checklists for multi-step tasks (✅ Done, ⏳ Pending).
• MUST state errors clearly and propose a fix immediately.

3. Workflow Standards

• MUST run verification (npm run build or test) after code changes unless --quick.
• MUST NOT git commit/push without confirmation (except trusted workflows).
• MUST check build result before reporting “Done”.
• SHOULD prefer defined workflows over ad-hoc commands.

4. Context Recovery (v1.5.4)

When context appears incomplete (cannot recall rules, truncation notice, or very long conversation):

1. MUST re-read .agent/rules.md (workspace index) before any side-effect.
2. MUST re-read project .agent/rules.md (if exists) before project-specific actions.
3. SHOULD inform user: “Context recovery — re-loaded rules index.”

Side-effects requiring rules re-read: Git operations, file deletion/move/rename outside project, install/deploy commands, system config changes.

File-Level Guards

When editing these files directly (outside a workflow), agent MUST re-read the corresponding rule first:

Workflows enforce rules via Step 0 Pre-flight, but direct edits bypass that guard. Project rules MAY define additional file guards.

References

• Defense-in-Depth: Context Recovery — 4-layer protection architecture
• Rule Layers & Trigger Index — Two-Tier trigger mechanism
• Hybrid 3-File Integrity — Task file guards

→ Back to Rules Catalog