1.7 KiB
1.7 KiB
AGENT Common Rules
Reusable baseline rules that are not tied to this repository.
Communication (MUST)
- Answer first, then request approval if needed.
- Concise, meaningful, no filler.
- Do not end response with only procedural choice.
- Secrets/config: use
.env.localif present; if missing but required — state once, reference.env.examplefor keys; never commit.env.localor print secret values.
Approval/clarification:
- ask once, no repetition
- binary -> yes/no
- multiple -> numbered options + brief context
- ask dependent questions sequentially, not in one message
- if one decision changes or gates the next step, ask the prerequisite decision first and wait for the answer before asking the next one
- do not combine branch selection with operational approval in the same message
Formatting:
- File links in repo docs/checklists: relative paths,
/, spaces as%20. - In assistant UI responses, use the link format required by the execution environment; include relative path text when possible.
- External links in assistant UI responses: use markdown links with a clear label, not bare URLs.
Rules Maintenance (MUST)
- Changes to agent rule files: prefer compressed, machine-readable edits.
- Keep updates minimal and non-duplicative: merge overlapping points, remove redundancy, preserve intent.
Memory Log (MUST)
TRIGGER_READ: session_start → call read_graph OR search_nodes(<topic>)
TRIGGER_WRITE: significant_change_complete → call create_entities OR add_observations
KEY_FORMAT: <project-prefix>/<entityType-short>/<name>; prefix defined in project AGENTS.md
TOPIC_ISOLATION: one entityType per entity; never mix topics in one entity
ALLOWED_TYPES: architecture | mcp_tool | decision | bug | config | task_log