pi-srcwalk

A Pi extension wrapper around srcwalk — makes structural code-intelligence easy for AI coding agents. srcwalk is a CLI for symbol search, callers/callees, deps, and overviews; this package wraps it into five agent-safe semantic tools with bounded output, error handling, and registry-based candidate handoff.

Five agent-facing tools:

• semantic_query — discover ranked code evidence from natural-language questions, symbols, files, overviews, deps, and tests.
• semantic_inspect — deep-inspect known symbol(s): context, callers, callees, and references in one shot.
• semantic_show — read exact source target(s) via srcwalk show with a fixed -C 12 surrounding-line window.
• semantic_review — review staged or working-tree changes with diff evidence and risk hints.
• semantic_grep — search raw text or regex with trigram-indexed candidate pruning and full-scan fallback.

No Python runtime. Pure TypeScript. Ships as a Pi extension package.

---

Prerequisites

• srcwalk CLI — the structural code-intelligence engine.

  Install via:

  npm install -g srcwalk
  # or npx srcwalk for on-demand use
  

  Verify:

  npx srcwalk --version
  # or srcwalk --version if installed globally
  

---

Quick start

# Install from npm
pi install npm:@sting8k/pi-srcwalk

# Or load the local checkout during development
pi -e ./extensions/pi-srcwalk/index.ts

After /reload in Pi, all five tools become available.

---

Which tool should I use?

unknown target / broad question
  -> semantic_query

known exact symbol
  -> semantic_inspect

raw text / regex
  -> semantic_grep

exact path:line or candidate id
  -> semantic_show

changed code
  -> semantic_review

Need	Use
Find likely files, symbols, deps, tests, or overview	semantic_query
Understand one known symbol deeply	semantic_inspect
Search raw text or regex	semantic_grep
Read exact source from a candidate or path:line	semantic_show
Review current diff	semantic_review

---

Tools

semantic_query

Use this for broad code discovery: questions, fuzzy symbols, file targets, overviews, deps, and tests.

What it does:

• Accepts a natural-language question, a fuzzy symbol, a file path, path:line, or a request like "overview of Z", "deps of Y", or "tests for W".
• Looks up a TypeScript-native process-local BM25/PRF memory cache for broad queries.
• Calls srcwalk for structural evidence (discover, context, trace, deps, overview, show).
• Prioritizes exact symbol anchors when a natural-language query names CamelCase or method-like symbols.
• Uses RRF fusion when both BM25 and srcwalk rank lists are available, then gates confidence.
• Returns a compact evidence packet.

You ask	It returns
"how does ranking work?"	ranked file/function candidates + code context
"overview of src/search"	structural overview with function list + relations
"deps of rank.rs"	dependency list
"tests for bm25"	test files matching the topic
execute_search (fuzzy symbol)	definition candidates + evidence
src/index/bm25.ts:17 (exact)	evidence at that location

Every result includes:

• Retrieval confidence — high when candidates cluster tightly; medium when results spread across modules or query is broad.
• Bounded evidence — compact srcwalk evidence by default, with full blocks available in deep detail. Open returned candidates with semantic_show, or inspect exact known symbols with semantic_inspect.

When semantic_query abstains (abstained: true), it means no strong match was found — do not fabricate evidence from thin air.

semantic_grep

Use this for regex search by default. Set literal: true for exact strings, and add scopes, glob, ignoreCase, or context only when needed.

What it does:

• scopes accepts one or more dir/file paths (relative, ../, or absolute). Overlapping scopes are canonicalized before indexing.
• Searches regex patterns by default and exact strings when literal: true is set.
• Uses a trigram-indexed pipeline: literal anchor extraction, candidate pruning, then exact line verification.
• Falls back to full scanning when regex anchors are too weak or too complex.
• Returns raw file:line matches with optional context lines.
• Replaces built-in grep in the pi-srcwalk injected tool contract.

You ask	It returns
execute.*Search	regex matches with pruned candidates
literal: true, "foo.bar"	exact literal file:line matches
--glob "**/*.ts"	same search, but limited to matching files
-i / ignoreCase: true	case-insensitive matches

semantic_inspect

Use this when you already know the exact symbol name. Pass up to three symbols with commas when needed.

What it does:

• Deep-inspects known symbol(s). Use symbol: "A,B,C" for up to 3 symbols.
• Runs srcwalk context first to show local structure and call neighborhood.
• Adds explicit relation evidence:
  • callers — upstream call sites
  • callees — downstream detailed call sites
  • references — symbol matches / reference candidates
• Default relation: "all" returns Context, Callers, Callees, and References.
• Results include inspect_id so targets can be opened with semantic_show.

// Full inspect for one known symbol
semantic_inspect({ symbol: "executeSearch" })
// -> Context, Callers, Callees, References, inspect_id

// Narrow output to callers
semantic_inspect({ symbol: "buildOrLoadIndex", relation: "callers" })
// -> Context, Callers

// Bounded multi-symbol inspect
semantic_inspect({ symbol: "executeSearch,readArg,formatInspectPacket" })
// -> one section per symbol; max 3

semantic_show

Use this to open exact source from a previous search_id / inspect_id, or from a direct path:line target.

What it does:

• Reads exact source code via srcwalk show <target> -C 12 --budget 5000.
• Opens a candidate from semantic_query using search_id + candidate_id.
• Opens a candidate from semantic_inspect using inspect_id + candidate_id.
• Also accepts direct targets, including multi-target strings like a.ts:10,b.ts:20-30.
• Does not run srcwalk context or relation traces. For structural context around a known symbol, use semantic_inspect.

// Open candidate #1 from a previous search
semantic_query({ query: "buildOrLoadIndex" })
// -> returns search_id: "r595b4-s1", candidates: [...]
semantic_show({ search_id: "r595b4-s1", candidate_id: 1 })

// Open candidate #1 from a previous inspect
semantic_inspect({ symbol: "buildOrLoadIndex" })
// -> returns inspect_id: "r595b4-u1", candidates: [...]
semantic_show({ inspect_id: "r595b4-u1", candidate_id: 1 })

// Stateless direct source read
semantic_show({ target: "src/index/cache.ts:154-259" })

// Multi-target direct source read
semantic_show({ target: "src/engine.ts:45-55,src/router/intent.ts:1-5" })

semantic_review

Use this to review staged or working-tree changes, optionally narrowed to one scope.

What it does:

• Runs srcwalk review on staged changes by default.
• With target: "working-tree", reviews unstaged diffs.
• If scope points to a nested git repo, runs review inside that repo.
• Returns changed files, affected symbols, and risk hints.

Use semantic_review when the user asks to review, check, summarize, or assess current changes.

---

Common workflows

Understand and edit existing code

semantic_query({ query: "where is cache eviction handled?" })
  -> semantic_inspect({ symbol: "evictOldest" })
  -> semantic_show({ inspect_id, candidate_id: 1 })
  -> edit
  -> semantic_review({ target: "working-tree" })

Review current changes

semantic_review({ target: "working-tree" })
  -> semantic_show({ target: "file.ts:line-range" }) when exact source is needed

---

How it works

Agent (Pi)
  ├─ semantic_query(query, scope?)
  │    discovery / NL routing
  │    -> BM25/PRF + srcwalk evidence
  │    -> ranked candidates + confidence
  │
  ├─ semantic_grep({ pattern, scopes?, glob?, literal?, regex? })
  │    deterministic text/regex search
  │    -> literal anchors / trigram prune
  │    -> verify exact line matches
  │
  ├─ semantic_inspect({ symbol, relation?, scope?, limit? })
  │    known symbol(s), max 3
  │    -> discover target + refs
  │    -> srcwalk context
  │    -> callers / detailed callees
  │    -> inspect_id candidates
  │
  ├─ semantic_show({ target } | { search_id/inspect_id, candidate_id })
  │    exact source read
  │    -> srcwalk show <target> -C 12 --budget 5000
  │
  └─ semantic_review({ target?, scope? })
       changed-code evidence
       -> srcwalk review --staged | working-tree

Each lane returns bounded srcwalk evidence and truncates tool output to 50KB. semantic_query is broad discovery, semantic_grep is raw text/regex search, semantic_inspect is targeted symbol understanding, semantic_show is exact source reading, and semantic_review is changed-code review.

---

Cache

Broad-query BM25/PRF uses a process-local compact memory cache:

semantic_query process
└─ memory:<repo+scope+version hash>
   ├─ path table + chunk line ranges + short previews
   ├─ vocab table + token→termId map
   ├─ typed-array BM25 postings
   └─ typed-array doc terms for PRF

The cache does not write chunk/index files to /tmp. It rebuilds when file fingerprints change and is bounded with LRU eviction:

Env	Default	Purpose
PI_SRCWALK_MEMORY_CACHE_ENTRIES	4	maximum cached repo/scope indexes per process
PI_SRCWALK_MEMORY_CACHE_MAX_MB	512	approximate memory budget before LRU eviction

The retained memory index avoids full chunk text and duplicated token strings; it keeps only chunk metadata, short previews, vocabulary strings, and typed arrays. No database, no native dependencies.

---

Python lab (research phase)

router_lab_v9.py on the lab branch prototyped the router, BM25/PRF, RRF, and confidence architecture. Embedding was left out of TS v1 to keep the extension dependency-free.

---

Project structure

pi-srcwalk/
├── package.json                     # Pi package manifest
├── extensions/pi-srcwalk/index.ts   # Pi extension entrypoint (5 tools)
├── src/                             # TS-native runtime engine
│   ├── engine.ts                    # semantic_query orchestration
│   ├── cli.ts                       # dev smoke-test CLI
│   ├── index/                       # compact memory cache + BM25/PRF
│   ├── router/                      # intent detection + command planning
│   ├── srcwalk/                     # CLI runner + output parser
│   ├── ranking/                     # RRF fusion + confidence
│   └── output/                      # evidence formatting + truncation

---

Design principles

1. Router, not pattern-match — classify intent, generate strategies, early-stop on success.
2. Evidence, not summary — return raw srcwalk output as bounded evidence. Don't paraphrase.
3. Abstain, don't hallucinate — if no strong match, say so clearly with abstained: true.
4. Minimal agent surface — keep tool inputs small; advanced srcwalk knobs stay internal.
5. Fallback broadly — fall through to text search → symbol glob → overview, then abstain when evidence is still weak.

---

References

• srcwalk — structural code-intelligence CLI
• Pi coding agent — agent platform, extension spec
• Semble — RRF fusion inspiration
• potion-code-16M — embedding model (lab only)