explore

Explore phase (Full tier)

Explore is phase 1 of the four-phase loop. Its output is understanding, not artifacts. You are reading, not writing to jobs/.

Steps

1. Read MEMORY.md. Read it once, in full. It's designed to be thin.
2. Select relevant memory files by descriptor. For this task, decide which memory/*.md files are actually relevant. Read only those. Do NOT read every file in memory/.
3. Pull external sources if needed. If the task references URLs, documents, or other external sources, fetch them into resources/fetched/<slug>.md with the required frontmatter (source, fetched_at, fetched_by — the job slug, or literal light if this turns out to be Light).
4. Test intent. Before writing a proposal, can you name at least one concrete acceptance criterion for the task? If not, the intent is still ambiguous. Ask the user ONE clarifying question (not a list).
5. STOP at the end of Explore. Do NOT write jobs/<slug>/plan.md yet.

Allowed writes (this phase)

• resources/fetched/<slug>.md (or binary + .meta.yaml sibling) for any source pulled during Explore.

Forbidden writes (this phase)

• Anything under jobs/.
• Anything under memory/.
• Anything appended to MEMORY.md.
• Anything appended to TODO.md.
• AGENTS.md itself.

Guardrails

• Lazy load. Do not read every memory file by default. The MEMORY.md descriptors are the index; use them.
• One question at a time. If intent is unclear, ask a single focused clarifying question. Don't stack three.
• No preset domain assumptions. Memory has no categories. Don't ask "should I check travel memory?" — check MEMORY.md, pick what's relevant by descriptor.
• Fetched is immutable. If you've already fetched a source in a prior session and want a fresh copy, save it under a new filename; do not overwrite.

Exit criterion

You exit Explore when you have enough clarity to write a plan.md with at least one concrete, testable acceptance criterion. Advance to Design.