SymDex Code Search

Use SymDex before broad file browsing. Use it to save tokens by retrieving the exact code the agent needs instead of scanning whole files. SymDex currently covers 21 language surfaces, including Python, Go, Kotlin, Dart, Swift, HTML, CSS, Shell, Vue/Svelte script blocks, and Markdown headings plus supported fenced code blocks.

Current SymDex Snapshot

• Package version: 0.1.26
• Latest public tag: v0.1.26
• MCP tool surface: 21 tools
• Language coverage: 21 language surfaces, including Android/Kotlin, Flutter/Dart, iOS/Swift, HTML, CSS-family stylesheets, Shell, Vue/Svelte script blocks, Markdown headings, and supported fenced code blocks
• Route extraction: Python, JavaScript/TypeScript, Spring/Kotlin, Laravel, Gin-style Go, ASP.NET, Rails/Sinatra, Phoenix, and Actix
• Search outputs: one-line CLI token-savings footers plus MCP roi, roisummary, and roiagent_hint
• Quality outputs: MCP results and CLI JSON include quality with confidence, freshness, parser mode, language surface, generated-file hints, and embedding availability
• Context packs: CLI symdex pack and MCP buildcontextpack assemble token-budgeted evidence bundles for broader feature questions
• Semantic backends: local sentence-transformers, Voyage, OpenAI-compatible /embeddings, Gemini, and compatible proxies
• Hosted embedding support: SYMDEXEMBEDRPM plus symdex index --lazy for foreground structural indexing with background embedding fill
• No-embedding mode: symdex index --no-embed skips semantic embedding work entirely
• Watch behavior: low-memory structural refresh by default; use symdex watch --embed only when semantic embeddings must refresh continuously
• State model: global ~/.symdex by default, optional workspace-local ./.symdex with registry.json
• Markdown support: .md, .markdown, and .mdx headings plus supported fenced code blocks are indexed alongside source files

Start Here

1. If the SymDex CLI reports a newer release, prefer upgrading before long sessions.
2. Confirm the repo id.
3. If the repo id is already known, pass repo on every scoped tool call.
4. If the repo id is unknown, call list_repos and match the current worktree.
5. Check freshness with getindexstatus(repo).
6. If the current worktree is not indexed, call index_folder(path=".").
7. If the workspace already has .symdex, treat it as the intended local SymDex state and reuse it.
8. Reuse the returned repo id for the rest of the task.

If SymDex is unavailable or indexing fails, say so clearly and fall back to normal file reads only as needed.

Core Rules

• Search first.
• Pass repo whenever you know it.
• Use buildcontextpack for broad "how does this feature work?", docs, API, or bug-investigation questions before issuing many separate narrow searches.
• Prefer getsymbol or getfile_outline over full-file reads.
• Use call graph and route tools before manual tracing.
• Re-check getindexstatus after major edits or worktree switches.
• Read full files only when editing, reviewing unsupported or generated content, or when SymDex cannot answer.
• Optimize for lower-token retrieval, not broad context loading.
• If a search tool returns roi, roisummary, or roiagent_hint, mention the approximate token savings briefly in your response.
• Inspect quality before reasoning from a result. Prefer fresh, parser-backed, high-confidence evidence; warn the user when evidence is stale, generated, fallback text, or missing embeddings.
• If the repo uses workspace-local SymDex state (./.symdex), stay inside that workspace so the same index is auto-discovered.
• Treat symdex watch as low-memory by default; only request --embed when semantic embeddings must refresh on file changes.
• For remote embedding providers with strict request limits, prefer symdex index --lazy and set SYMDEXEMBEDRPM instead of blocking an agent session on a long foreground embedding run.

Tool Selection

Typical Flow

1. Confirm the repo id and freshness.
2. Index with index_folder if needed.
3. For broad feature, API, docs, or bug-investigation questions, start with buildcontextpack.
4. For narrow lookups, start with searchsymbols, semanticsearch, or search_text.
5. Narrow to getsymbol or getfile_outline.
6. Check quality fields on returned items before making code-understanding claims.
7. Use getcallers, getcallees, searchroutes, or getrepo_stats for deeper analysis.
8. Fall back to direct file reads only when SymDex cannot answer precisely enough.

Decision Guide

• "Where is X defined?" -> search_symbols
• "What does this do?" -> semanticsearch, then getsymbol
• "How does this feature work?" -> buildcontextpack
• "Gather evidence before documenting this API" -> buildcontextpack with include=["routes", "docs", "tests"]
• "Who uses this?" -> get_callers
• "What does this call?" -> get_callees
• "Where is the endpoint?" -> search_routes
• "Show me the file structure first" -> getfileoutline
• "Give me a repo-level picture" -> getrepooutline or getrepostats
• "Is the index current?" -> getindexstatus

Good Trigger Phrases

• "Find the function that validates JWTs"
• "Who calls this route handler?"
• "Show me the outline of this file"
• "Search for the code that parses webhook payloads"
• "Find the HTTP route for /api/checkout"
• "Give me the repo summary before I edit anything"
• "Find the code path that might explain this bug"
• "Check whether this logic is coherent across callers and callees"
• "Build a context pack for this feature"

Bug And Logic Investigation

Use SymDex as evidence retrieval, not as proof by itself.

1. Start with the symptom, route, function, or text clue.
2. Use buildcontextpack when the bug spans multiple symbols, routes, docs, or tests.
3. Retrieve the exact symbol, route, caller chain, callee chain, and relevant docs.
4. Inspect quality.indexfresh, quality.confidence, quality.parsermode, quality.isgenerated, and quality.hasembeddings.
5. If evidence quality is weak, say what is missing before claiming a bug.
6. Cite the concrete symbol, file, route, or call edge behind any bug hypothesis.

Editing

When you need to edit code:

1. Use SymDex to find the exact symbol or file location.
2. Read only that file or symbol slice.
3. Make the smallest change needed.

Watch And Semantic Search

• symdex watch refreshes structural indexes by default without loading local embedding models.
• Use symdex watch --embed only when the task needs semantic search to stay fresh continuously.
• If semanticsearch has no embeddings, fall back to searchsymbols or search_text, or re-index after enabling symdex[local], Voyage, OpenAI-compatible, Gemini, or another hosted embedding backend.
• Use symdex index --lazy when embeddings may be slow because of hosted-model latency or RPM limits.
• Workspace-local state keeps watcher metadata under ./.symdex, so commands should run from that workspace or pass the matching --state-dir.

Use Normal Browsing Only When Needed

• SymDex is unavailable.
• The repo is not indexed and cannot be indexed in the current environment.
• The target file type is unsupported or generated.
• You need surrounding context that the symbol-level response does not provide.

Output Checklist

• Repo id confirmed or derived
• Index freshness checked
• SymDex tool chosen before broad file reads
• Context pack used for broad multi-file questions
• Exact symbol or file outline used before whole-file reads when possible
• Quality metadata inspected before logic or bug claims
• Direct file reads used only when SymDex could not answer cleanly