vexp-cli 3.2.5 → 3.3.0

package/README.md

@@ -2,7 +2,7 @@
 
 **Give your AI coding agent the right context, then verify its work — 100% on your machine.**
 
-`vexp-cli` is a local-first [MCP server](https://modelcontextprotocol.io) that pre-indexes your codebase into a dependency graph and feeds any AI coding agent only the code that matters for the task — cutting **87% of the tokens** it would otherwise waste reading whole files. New in 2.5: it also checks the agent's work mechanically when it says it's done, and scans your code for PII and secrets before any model reads them. Your source never leaves your laptop.
+`vexp-cli` is a local-first [MCP server](https://modelcontextprotocol.io) that pre-indexes your codebase into a dependency graph and feeds any AI coding agent only the code that matters for the task, so it stops spending tokens rediscovering your repo before every change. New in 2.5: it also checks the agent's work mechanically when it says it's done, and scans your code for PII and secrets before any model reads them. Your source never leaves your laptop.
 
 ```bash
 npm install -g vexp-cli
@@ -20,8 +20,8 @@ vexp index      # builds the local graph — no login, no API key
 
 Your AI agent burns most of its budget exploring your repo before writing a line of code. It reads entire files, floods its context window, and still misses the one function that mattered. vexp fixes the input **and** the output:
 
-- **📉 The whole session costs less (new in 3.0).** Tuned on full agent sessions against the same agent without vexp, on tasks the tuning never saw: **more tasks resolved (24/25 vs 22/25), 13% less context replay, 8% lower session cost — 20% lower on long sessions.** The turn-zero orientation now carries up to 3× more evidence, and everything that could add a turn to your session is opt-in (`vexp setup --interventions`): by default vexp only subtracts.
-- **⚡ 87% fewer tokens per call.** vexp returns ranked pivot files with exact line ranges and blast radius — not whole files. Measured on real usage, not a synthetic benchmark.
+- **📉 Fable-5-level results from Opus 5. At half the token price.** On the SWE-bench Pro public set, Claude Code running Claude Opus 5 with vexp resolved **81.7% (597/731)**. Anthropic reports 80.3% for Claude Fable 5 and 79.2% for Opus 5. $1.52 per task, 95% CI 78.7-84.3, every session published at [vexp.dev/benchmark](https://vexp.dev/benchmark).
+- **⚡ One call, not a repo crawl.** vexp returns ranked pivot files with exact line ranges and blast radius instead of whole files, and since 3.0 the turn-zero orientation carries up to 3x more evidence, so your agent stops rediscovering the codebase before every change. Everything that could add a turn to your session is opt-in (`vexp setup --interventions`): by default vexp only subtracts.
 - **🔍 Verifies the work (new in 2.5).** When your agent declares a task done, `verify_done` reports what's provably broken or missing — broken imports, parse errors, dependents never updated — each with file and line. No model grades anything; it's git + tree-sitter + the code graph.
 - **🛡️ Shield scan (new in 2.5).** `vexp shield scan` shows the emails, API keys, tokens and private keys sitting in your comments and strings that every AI agent currently reads. Local, deterministic, masked.
 - **🔒 100% local.** Zero network calls, zero cloud indexing, no account, no API key. Your code stays yours.
@@ -32,7 +32,7 @@ Your AI agent burns most of its budget exploring your repo before writing a line
 vexp parses **34 languages** with tree-sitter into a local SQLite dependency graph, then serves your agent through the Model Context Protocol:
 
 - `run_pipeline` — ranked pivot files + line ranges + blast radius, in one call
-- `get_skeleton` — file structure at 70-90% token savings
+- `get_skeleton` — file structure (signatures, no bodies)
 - `verify_done` — mechanical completion check
 - `expand_vexp_ref` — full symbol bodies on demand