cmd/

Command askctx runs one find_context against a chosen binary and prints the response an agent host would receive.

Command chainprobe measures whether the call graph makes a chain followable, without an agent in the loop.

Command confcal calibrates the confidence gate.

Command corebench evaluates an embedding model on a downloaded subset of CORE-Bench (arXiv:2606.11864, HF: zhangfw123/CORE-Bench) in BEIR format: each repo dir holds corpus.jsonl, queries.jsonl and qrels/test.tsv.

Command deepprobe measures what ranking metrics structurally cannot see: whether the answer is inside the excerpt the response actually shows.

Diagnostic tool: for selected (project, query_id) pairs, runs retrieval in both vector-only and hybrid modes with top-30 + features, and prints a compact side-by-side comparison highlighting the expected symbol's rank.

Command embprobe is a 30-second semantic sanity check for an embedding model in OUR runtime (tokenizer + pooling + prefixes + ORT): it embeds a few fixed code snippets and two NL queries and prints the cosine matrix.

Command exploreprobe scores the served retrieval against SWE-Explore, the first external benchmark that grades the WHOLE pipeline rather than its entrance.

Command ftdata turns CORE-Bench-style BEIR data into embedder fine-tuning triplets: {"query", "pos": [...], "neg": [...]} JSONL, one line per query.

Command giteval scores the served stack against labels nobody on this project authored: each case is a real commit — query = the commit subject, gold = the symbols whose enclosing-function hunk headers appear in that commit's diff, resolved against the index.

Command goldengate is THE regression gate: one command, one report, always the same definition of "our numbers", measured through the shipped binary.

Command idxstats prints symbol and edge counts for one or more index DBs.

Command lateexp is an offline experiment: does late-interaction (ColBERT-style MaxSim over per-token vectors) improve paraphrastic SEED recall over the production single-vector mean-pool? It scores the WHOLE corpus by both methods (not just a re-rank of the seed set) because the hypothesis is that MaxSim catches paraphrases the single vector drops at the seed stage — a re-rank of an already-missed seed could never show that.

Command mcpeval drives the gen-eval corpus through the REAL served stack — it spawns the shipped binary's `mcp` subcommand per project and talks JSON-RPC over stdio, exactly like an agent host does.

Command negprobe measures the false-confidence rate: queries about plausible concepts that do NOT exist in the indexed repo.

Command reachprobe asks why a gold file is never retrieved.

Command regionprobe asks where the gold sits relative to what we returned.

Command rspbreak prices a find_context response by section.

Command selfsweep is a label-free ranking radar for ANY indexed repo: every documented symbol must be findable by its own docstring.

Command soak asks one long-lived server the same questions over and over and checks that the answers do not drift.