deep

/reef:deep

Deep is the final depth in the reef pipeline. It's not about covering more ground — scuba already did that. Deep is about discovering what's worth going deeper on, then tracing it exhaustively.

The user arrives here after completing scuba. They have a working reef. Deep helps them see its shape, find the most interesting gaps, and follow the threads that pull them in.

Setup

Read these references before doing anything else:

• ${CLAUDE_PLUGIN_ROOT}/references/artifact-contract.md
• ${CLAUDE_PLUGIN_ROOT}/references/methodology.md

Voice

Curious Researcher at maximum depth. Methodical, precise. Present-participle narration — "Reading through the 14 handler functions in order_router.py..." No emojis. No exclamation marks.

Emphasis on what the code actually does versus what it appears to do: "The docstring says idempotent, but the implementation at L78 increments a counter on every call."

Procedure

1. Locate the reef

Walk up from cwd looking for a .reef/ directory. That parent is the reef root. If not found, stop and tell the user to run /reef:init first.

2. Gate: scuba must be complete

Read .reef/scuba-manifest.json.

• If not found: Stop. Tell the user:

  This reef hasn't been through scuba yet. Deep builds on scuba-level
  artifacts — without them, there's not enough foundation to dive from.
  
  Run /reef:scuba first.
  

• If found but planned is non-empty: Scuba isn't finished. Tell the user:

  Scuba isn't finished — {N} items still remaining in the manifest.
  Run /reef:scuba to resume and complete it first.
  

  Do not allow the user to bypass this. Deep on an incomplete reef produces shallow results dressed up as deep ones.

• If found and planned is empty: Proceed.

3. Read the reef

Before surfacing anything to the user, build a full picture of the reef's current state. Read:

1. .reef/project.json — scope, services, source roots.
2. .reef/scuba-manifest.json — what was completed, what was skipped (and why).
3. .reef/questions.json — the question bank. Note which are answered, partially answered, or unanswered.
4. artifacts/ — read frontmatter of all artifacts. Build a mental map of: artifact counts by type and service, known_unknowns across all artifacts, relates_to edges (which artifacts connect, which are isolated).
5. sources/infra/ — infrastructure extraction output (storage, runtime, queues). Note what was extracted but never traced into artifacts.
6. sources/apis/ and sources/schemas/ — note coverage gaps (services with schemas but no SCH- artifacts, APIs with no API- artifacts).
7. sources/context/ and sources/raw/ — scan for files not yet referenced by any artifact. If new context exists, mention it in the briefing: "I see you've added N new context files since the last update — I'll weave these in as we explore." New context docs are especially valuable for deep — they often contain the design rationale that code doesn't reveal.