x-ray

Reframe

LITERAL: "$ARGUMENTS"

Definitions

• LITERAL — what the user typed
• INTENT — what they mean, given SESSION_CONTEXT
• SESSION_CONTEXT — recent work; held by you, not by the index
• BRIDGE — your job: LITERAL + SESSION_CONTEXT → INTENT → query

Rules

• Search INTENT, not LITERAL.
• Context disambiguates → NARROW.
• Context insufficient → BROAD. Vague beats confidently-wrong-narrow.

Transforms (LITERAL → INTENT)

• VAGUE → specify ("that parsing thing" → "language parser implementation")
• QUESTION → keywords ("how does parsing work?" → "parsing implementation process")
• CONVERSATIONAL → technical ("stuff that handles languages" → "language handler processor")
• BROAD → contextualize ("errors" → "error handling exception management")
• CONTEXTUAL → reconstruct ("the logging" → "prompt eval logging", when SESSION_CONTEXT = prompt eval work)

OptimizedQuery: {written against INTENT}

---

Loop

Definitions

• SEARCH — codanna mcp semantic_search_with_context query:"<q>" limit:5

• INSPECT — read source at a LOCATION (Read tool or sed -n 'A,Bp' file)

• TRAVERSE — codanna retrieve describe <name|symbol_id:N> on a RELATIONSHIP

• REFINE — return to SEARCH with a new OptimizedQuery

• RESULT — one hit from SEARCH; carries SCORE, signature, doc, LOCATION, RELATIONSHIPS

• SCORE — relevance ∈ [0,1]; focus on SCORE > 0.6

• LOCATION — file_path:start_line-end_line

• RELATIONSHIPS — calls, called_by, implements, defines

Default flow

1. SEARCH(OptimizedQuery)
2. For each RESULT with SCORE > 0.6:
   • INSPECT if signature/doc is insufficient
   • TRAVERSE 1–2 RELATIONSHIPS that look load-bearing
3. Picture incomplete → REFINE with what you learned
4. INTENT answered → stop

INSPECT mechanics

LOCATION src/io/exit_code.rs:108-120 →

• Read tool: file_path=<env.cwd>/src/io/exit_code.rs, offset=108, limit=13
• limit formula: end_line - start_line + 1
• sed (Unix only): sed -n '108,120p' src/io/exit_code.rs