redhat-et/ripwire

For a single developer, this tool is a good lookup accelerator. For an orchestrated fleet, it's load-bearing: it halved the research spend, twice redirected tasks before wasted work, prevented at least one silent-divergence shipped bug, and turned code-quality hygiene from a hope into a per-task mechanical gate. Whole-workflow ~2×; per-lookup 10–25×; and two moments where one call was worth more than the rest of the session's tooling combined.

— the report's bottom line

The model's own report of one engagement, on a version before 0.5; not a controlled measurement. Controlled measurements are in docs/EVALS.md.

Beside those sits a labelled survey of 237 tools that contributed nothing and says so. The two sets are disjoint by construction, so they add rather than nest — a tool that gave a lesson is never counted twice.

Both halves are load-bearing, and they are doing different jobs. The settled results are what make the quality lens trustworthy: McCabe on complexity (1976), Halstead on volume (1977), Spärck Jones on term specificity (1972), Nagappan & Ball on churn. Fifty years of replication means those are not opinions, and a tool that measures your code should be built on the ones that survived.

The recent work is what makes it current: seventeen of the folded papers are from 2026, seven published in the last two months and three in the last thirty days (dates as of 2026-09-08; every row carries its arXiv id, so the claim is checkable rather than atmospheric). Retrieval for coding agents, context-compression cost, placebo-controlled localization — that literature is months old, not decades, and several rows were folded within weeks of the paper appearing.

Neither half alone would be enough. A tool built only on the classics would not know what an agent needs; one built only on last month's preprints would have nothing underneath it. And the newest row is a result that failed when it was tested here — which is the point of writing them down. All three counts are re-derived from that document's own tables by test/readmedriftcheck.sh on every run, which fails if this page and those tables disagree, so the claim cannot quietly drift. The row-by-row ledger is docs/LINEAGE.md.

Measured on 48 matched questions across django, webpack and this repository. Across all three, ripwire indexes in 0.25–0.45 s and 6.6–16.5 MB against that server's 23–52 s and 391–623 MB. The full method, the wins named one by one and the losses included, is in Against the leading graph-database code-context MCP server and docs/EVALS.md.

One self-contained binary on your own machine, offline, installed in one line — and the same line installs and activates the task-shaped skills that teach your agent when to reach for it, not just how, for every agent it finds on the machine. If your agent can run shell commands — Claude Code, Codex, Cursor, Windsurf, Gemini, opencode, aider — it is set up the moment the install finishes; the MCP server is the optional second interface. Install it and ask it something before you finish reading this page:

Two things make that reachable in practice, and neither is the destination. Answers are honest about their own limits — a count that cannot be a total is labelled a floor, a zero means "none found" and never "none exists", every truncation is disclosed — so an answer never looks more complete than it is, and the map never degrades the code by guessing. And an answer can be given a token budget, so what one costs is something you ask for rather than discover; where a complete answer will not fit, it says it went over rather than silently dropping the row you needed.

Those two are the stair-steps: honest about what is missing, priced in what it spends. The step they climb toward is a question fully answered in one call, which is not always trivial to reach — and where it is not, the output says so rather than pretending otherwise.

How we measure the climb. A complete answer to every code question is genuinely hard, and it is reached over many releases rather than in one. Until an answer can be complete, it owes you three things: it stays bounded in size, it is fully honest about what it does not know, and it gives you a clue where to look next. So every change is judged on more than "was it right":

• Complete answers: the share of questions answered so you can act without re-checking.
• Honest-partial rate: how often an incomplete answer says what it is missing, instead of sounding confident.
• False-confidence rate: wrong answers that claim to be complete — the worst case, held near zero.
• Next-clue usefulness: for partial and wrong answers, whether following the answer's first suggested next step reaches what the question needed, within one hop. Graded blind.
• Recoverability: for every cut answer, whether following its next= returns everything that was cut.
• Tokens per complete answer: total tokens an agent spends across a scenario, divided by the questions it got fully answered.
• Cut-row recoveries: how often a needed item arrived only through a continuation.

Answering comes first; efficiency is how we get there. We aim for answers about as lean as the leanest tools, but a byte budget is never allowed to cost an answer. We have gone too far toward short answers before and cut information that was needed, so now every size limit has one of two jobs: a runaway guard far above typical answers (it catches an output bug, and says so when it trips), or a stair-step target where anything over it must be explained. A capped answer is always measured against the same answer uncapped, and the one that answers better wins.

A cut is only honest if you can get the rest back. Disclosing a cut is the floor, not the bar: a capped answer plus the next= it prints must return everything the uncapped answer had. A test that checks this for every budget option on every command is being built, so that "the answer was cut" always comes with "and here is the rest".

What a budget may spend first. When a budget is tight, ripwire drops bytes that carry no answer before it drops answers: repeated legend text, duplicated paths, bodies whose signature already says what you need. Only then the lowest-ranked rows, each counted. A whole section never disappears without a marker.

Budgets are yours to set, never ours to impose. The defaults and the setups ripwire wrap writes carry no budget that cuts a typical answer; a budget is a choice the caller makes, with its cost disclosed.

Efficiency is counted per complete answer, across the whole task. Bytes per answer can fall while the task gets more expensive, if the agent then has to ask again. So the efficiency figure is tokens per complete answer over a whole scenario — orient, then the questions that follow — not bytes per call.

Ranking is the tool's best guess; the next= is the safety net — and we measure how often the net is needed. ripwire cannot always know which row answers the question; it goes by ranking. When a needed item appears only after following an answer's next=, the honesty layer saved the answer but the ranking put that item below the cut. We count those per round:

• Recovered through next=: the item was cut but recoverable — one extra call. Driving this down is ranking work.
• Lost: the item was cut with no marker or no working next= — a bug, fixed first.

Ten everyday moments, re-measured on this repository, 2026-08-08. Figures are ~tokens (≈ bytes/4); every ratio comes from a real run, reproduced by the command in its row — raw byte counts and exact commands in docs/EVALS.md §5.

These aren't summaries that gamble with information. Each row is scored same-correct-answer-or-it-doesn't-count, and both sides were checked, not assumed: orient surfaces this repo's own pipeline files (ingest.cpp, graph.h, serialize.h) in the first screen, the same three docs/ARCHITECTURE.md names as central; the --for row lands mcpStale (src/mcpindex.h:633), the actual staleness check, 5th-ranked; --recall lands the container-rule doc (AGENTS.md) that states, verbatim, the same "no std::map" rule CONTRIBUTING.md explains in full; --pack-task names the same three touch points a human would — cachelint.h, mergeCachePack (src/main.cpp:1787), the lintrules.h helpers it reuses; --expand hands back the requested function's complete, unmodified body — the ranked-neighborhood addition costs the same ~22.6 KB regardless of which function you ask for, confirmed on two (a fixed floor, not per-function variance); --callers on langOfPath names its 2 real callers, the same ones a grep hit-list buries under 5 files of comment-only mentions; --impact+--uses on coversOrEquals names the same 2 direct call sites --uses alone would, plus (disclosed) a transitive reach --uses doesn't cover at all; --from-trace resolves all 7 frames of a real call chain by name to the same definitions a per-frame grep would eventually find, mixed with call sites and comments; --situ on a 2-file diff names the same 6 real test harnesses, 2 of which a filename grep across test/ cannot find even after opening every one of its 41 candidates — a completeness gap, not just a byte one; --pr-context surfaces co-change partners (test/regression.sh, src/main.cpp) a raw git diff has no way to know were usually touched and weren't this time — Fowler's Shotgun Surgery checked rather than merely named (the backtest is in docs/EVALS.md). The map ranks and discloses — it never paraphrases your code — and every truncation is disclosed in the header.

The honesty line, made concrete: the same auto-selection behind the --expand row also runs the other way. On a small file (pageRankDouble in src/pagerank.cpp, 5,559 B) the ranked bundle would cost 27,916 B — nearly 5× more than the file — so ripwire serves the file itself instead, disclosed as mode="whole-file" on the response, not silently. docs/EVALS.md §7 lists that and the other counterexamples this project publishes against itself.

The shape: plan the work, then run a loop that matches each task to the model that fits it. Every thread it spawns opens with an empty context on a repository it has never seen. Orienting an empty context to a large repository is the most repeated cost in the whole system, and the one this tool was built for — it is also the cost that grows with the size of the tree, which is why the pattern matters more the bigger the repository gets.

--for and --pack-task answer it in a single call, at the per-call rates in the table above, instead of a grep-and-read tour that every lane pays over again from scratch.

The return path matters as much. --quality-delta, --test-gate and --edit-check answer what did I make worse, which tests must run, did I change a contract — quantitative answers a lane can hand back as a verdict, rather than a transcript the orchestrator has to read to find out what happened.

What is and is not claimed here. Every figure on this page is a single-agent measurement. That the saving compounds with the number of cold orientations follows from the fixed-cost mechanism, but it is pre-registered and unrun — the reason the pattern is worth trying, not a result this project has published.

One self-contained HTML file (--html[=FILE]), no server, no CDN, no external asset. --color-by=lang|community|cx|churn|tested sets the initial colour; the page embeds all five and keeps a live selector, so switching lens costs no second run.

Read from the figures above, which state their own rules in a sidecar saved beside each image:

• arrow points caller → callee — the graph is directed, and the page draws it that way.
• 31 of 183 shafts dashed in this view = the resolver could not choose between same-name definitions and split the call over all of them — per edge, not per symbol. A symbol-level "this function makes some ambiguous calls" would mark every one of its edges, which would be a lie about most of them.
• labels: top 24 by in-view degree, one per name — one label per distinct name, so a picture of a container class stops crowding out the functions you asked about.
• shapes: ● fn ■ cls ✚ var — kind is nominal data on a nominal channel; complexity never uses shape.
• module outlines: 7 of 12 modules with 3+ nodes in view (cap 12; 3 dropped as too thin to read as a region; 2 dropped as enclosing mostly other modules) — three separate truncations, each with its own count and its own reason.

The cx and churn ramps share one five-stop scale, ordered so lightness rises with the value — it survives greyscale printing, and every adjacent pair stays separable under protanopia, deuteranopia and tritanopia. Thresholds are fixed rather than per-corpus quantiles, so a hot node cannot be manufactured by a cold repository.

churn needs real git history: a shallow clone reports every file as one commit, and a directory with no repository says churn unavailable rather than drawing zeros.

• The ranked symbols, in rank order — the cache-header constant kCacheMagic first (with its doc comment quoted in place and the one next= call that opens it), then spanTierMemoPath (the cache-path composer), ingestCommitTree, … ingest — each row with its file, line, and signature.
• Risk, annotated in place — complexity, git churn (ingest shows 128 recent edits), change amplification (touch ingest and 266 graph nodes feel it), purity and test coverage. The fragile spots are visible before anything touches them.
• One-hop call context — spanTierMemoPath calls shaKeyedCachePath, headSnapRepoHex, exclConfigHex; no second query needed to see the neighbourhood.
• Its own confidence — this answer says confidence="high" with the score margin attached; a flat ranking says low, so it reads as a starting point instead of masquerading as an answer. confidence= measures how clearly the ranking separates its head from the rest, not whether the head is what you meant: ask a repository about a concept it does not contain and the best lexical matches still rank, confidently. Phrase the task in your code's own words.

<ctx task="incremental cache invalidation" confidence="high" margin_pct="20"
     bundle="compact" bodies="0" reason="compact-route" est_tokens="3995">
  <sigs shown="23" total="40" capped="1">
    <d l="106" n="kCacheMagic" p="src/ingest_cache.h" cx="0" in="0" churn="11" amp="71" pure="1" r="1"
       next="--expand=src/ingest_cache.h:kCacheMagic">
      <doc>incremental cache (--cache): per-file content hash + raw facts so a
           re-run re-parses ONLY …</doc>constexpr std::uint32_t kCacheMagic = …</d>
    <d l="1307" n="spanTierMemoPath" p="src/ingest_astquery.h" cx="1" in="2" churn="5" amp="44" r="2"> … </d>
    <d l="247" n="ingestCommitTree" p="src/dmm.h" cx="6" in="1" churn="6" amp="27" r="3"> … </d>
    …
    <d l="191" n="ingest" p="src/ingest.cpp" cx="4" in="14" churn="128" amp="266" tested="1" r="13"> … </d>
    … </sigs>
  <hops shown="2" total="6" capped="1" noedge="2">
    <h l="1307" p="src/ingest_astquery.h" n="spanTierMemoPath">
      <calls total="3"><c n="shaKeyedCachePath" l="1621"/> … </calls></h> … </hops>
</ctx>

cx= complexity, churn= git edit frequency, amp= change amplification, r= rank; <hops> rows carry the one-hop call context, caps disclosed. The one legend at the top of the real output defines them — tersely by default, every reading in full with --legend=full — and the root self-reports the bundle's cost — est_tokens="3995" here.

And against five retrieval competitors on a held-out LocBench slice, it finds all gold files in the top 10 on 58.3% of instances — the best alternative lands 40.0% — while indexing in 0.31 s. The full leaderboard, losses included ↓

The 48 questions span symbol lookup, conceptual search, blast radius, and one-call task orientation. ripwire indexes the same three repositories in 0.25–0.45 s and 6.6–16.5 MB, against that server's 23–52 s and 391–623 MB, and answers a warm query in a median 197 ms against its 1,082 ms. Its seven wins are real and named one by one in the method.

Both arms warm with a pre-built index, median of 3 timed calls, stdout to a file rather than a pipe. The competitor ran in its stronger retrieval configuration; its numbers were recorded once and then frozen, and ripwire's side was re-run after the fixes the first pass produced. Per class, as a share of the competitor's bytes on totals: symbol lookup 1.35×, conceptual search 1.23×, blast radius 0.39×, task orientation 0.06×.

Where it loses: on a plain one-symbol lookup the competitor answers in about a kilobyte carrying callers and callees, and ripwire spends roughly three times that to also hand back the body. It ranks a chunk-id plugin first on one webpack query where ripwire never surfaces the directory at all — a ranking miss, and a fix for it was built, met its pre-registered band, and was reverted anyway for failing a separate standing requirement. Its depth-labelled blast radius and its import edges are both better presentations than ripwire's flat reaching-set.

Full method, pins, per-class tables, the carried-versus-re-judged ledger, and the complete list of what the competitor does better: docs/EVALS.md §2.

Nothing it is unsure about reaches your agent unlabelled — and nothing it could not see goes unnamed. Every guess is marked in the output, and every mark has a next step — up to handing it a compiler-grade index. Point it at a repository whose main language it has no grammar for and the map's first line says so (unindexed="ml:793,mli:607,…" on a facebook/infer clone); a file it indexed but cannot vouch for carries a parse-health row; every file the crawl passed over is itemized with its reason. A confident-looking map that lies by omission is the failure mode this tool refuses.

58.3% against 40.0% for the best tool tested — and it answers before they finish indexing. Every arm below was re-run in full on 2026-08-08 — one ripwire binary (the profile-guided release build that now ships), one evaluator, one 60-instance held-out LocBench slice: paired, zero exclusions, same gold set, and the metric code imported unmodified into every arm. Strict file@10 = all gold files inside the top 10, which is whether your agent starts in the right place at all.

Ripwire leads every arm on both accuracy metrics and in both strata. Paired, the losses are small and they are published: 2 instances to codebase-memory-mcp, 2 to repowise, 1 each to graphify and aider. Cold from nothing to an answer — parse, rank, reply, no cache — ripwire takes 0.213 s, against a ~35 s index-then-query for repowise; its worst single index in this run was 352 s.

Three things this table costs us, said plainly. codebase-memory-mcp is the real runner-up at 40.0%, not repowise — an earlier round credited it with 26.7%, and re-running it fairly raised it. The margin over the best competitor is therefore 1.46×, not the 1.75× two separately-dated tables used to imply. And multi-file gold is hard for everyone: ripwire leads the stratum at 21.4% strict, but its own any@10 there is 78.6% — it finds a gold file and misses the siblings, and no arm in this table solves that.

Held out wider — 243 instances across 78 repositories — ripwire lands 60.9% against 27.6% for its own pre-routing baseline: +33.3pp paired, clustered-bootstrap 95% lower bound +25.0pp, bought for +3.4% warm latency and −39.4% tokens. Full provenance, the losing instances one by one, and a third round against a compression-layer competitor: Measured and bench/headtohead/r4-2026-08-06/, whose harness is committed so anyone can re-run the whole comparison.

Test scaffolding does not pollute the ranking from inside source files either: #[cfg(test)] mod tests, describe() blocks, Test* classes and [Fact] attributes are detected syntactically, wherever they live — not by file path alone. On an astral-sh/ruff clone (5,945 files), --ignore-tests removes 23,907 test symbols where path rules alone caught 18,532 (ripwire <ruff> --ignore-tests, 2026-08-14; the per-language fixtures are pinned by test/testscopecheck.sh).

A Survey of the Neighbourhood's Eligible Instruments — being an account of five gentlemen of retrieval, and one lady of no pretension whatsoever.

It is a truth universally acknowledged, that an engineer in possession of a large repository must be in want of a map.

Mrs. Codebase-Memory must be named first, for she has risen a great deal in the estimation of the neighbourhood — two-fifths of her answers entirely correct, which is more than any other caller can say, and she is ready in a second and a quarter. It must nevertheless be recorded that her card announces accomplishments in the semantic line; that upon enquiry the semantic line is not at home; and that the household denies all knowledge of it. One is left with the impression of a capable woman ill-served by whoever prints her cards.

Mr. Repowise is by common consent the most substantial of the party, and no one who has waited upon him would dispute it. He is possessed of a handsome index and a manner of great thoroughness; but he must be seen to. One does not simply address Mr. Repowise. One sends word, and dresses, and waits — three-and-thirty seconds on an ordinary morning, and upon one memorable occasion in the country, seven minutes and four seconds — during which interval a less consequential neighbour has answered the question, taken her leave, and thought no more about it. He answers creditably when at last he arrives, one time in three; whether that is worth the toilette, each family must determine for itself.

Mr. Graphify enjoys a great many admirers. He does not rank his acquaintances; he calls upon them in whatever order his walk happens to take him, and reports the order of the walk as though it were an opinion. He has been known to arrive carrying a hundred and thirty megabytes of correspondence. Pressed once for any answer at all, he replied that no matching nodes were found, and considered the matter closed.

Mr. Aider is the most gentlemanly of the company and by far the most difficult to consult. He cannot be asked a question — the thing is simply not done. One may mention names in his hearing and hope he takes the hint; he does take it, and is fully ten points the better for it, which says more about the hint than about Mr. Aider. But he forms his view of the neighbourhood before you speak and retains it after, and one cannot escape the feeling that the conversation was never truly with you.

Mr. Codeseek is a young gentleman of quick habits who suffers from an affliction of address. Speak to him plainly, in the language of ordinary complaint, and he will regard you with perfect composure and say nothing whatever — nothing, upon sixty occasions out of sixty. Name a person precisely as that person is named, and he grows animated directly. It is not stupidity; it is a want of imagination in the matter of introductions.

And there is ripwire, of whom nothing is said in the drawing rooms, because she has already gone home. She was asked; she answered, in thirteen hundredths of a second; every gold file within the first ten, in eight-and-fifty cases of the hundred. She keeps no establishment, corresponds with no distant authority, and has never once been indexed at a party. Mr. Repowise finds her abrupt.

She is.

_{Every figure above is a measured number from the round-4 table on this page: the index medians (1.24 s, 34.0 s, 3.37 s) and repowise's 352 s worst case, the 40.0% and 33.3% strict file@10, the 0-results-on-60/60 fallback arm, the absent semantic_query tool, graphify's 129 MB largest graph and its 1-of-60 empty ranking, aider's +10 pp personalization delta, and ripwire's 0.108 s / 58.3%. Provenance in bench/headtohead/r4-2026-08-06/ and docs/EVALS.md; only the manners are editorial. These are other people's real work, and the joke is aimed at the trade-offs, never at the authors.}

The Long Index — in which a man asks six informants one simple question, and only one of them has the decency to answer it.

It was a million lines of somebody else's mistakes, and I needed one file out of it before the coffee went cold. So I did what you do. I went and talked to the people who say they know the neighborhood.

Codebase-Memory had the best record in the room and she knew it — two answers right out of every five, handed over in a second and a quarter, which in this business is practically a kindness. Trouble was the card. Right under her name it said semantic query, real classy, real expensive-looking. I asked to see it. She said it wasn't in. I asked the house. The house had never heard of it. I've known a lot of good people ruined by whoever printed their cards.

Repowise was the heavyweight — everybody told me so before I got through the door. Big index, good tailoring, thorough as a tax man. Only you don't just ask Repowise a question. You send word. You wait. Thirty-three seconds on a good day, and one bad morning out in the country, seven minutes and four seconds — long enough to get the same answer somewhere else, drive home, and forget his name. He came through one time in three. For some outfits that's worth the wait. I've got a metabolism.

Graphify never met a fact he wouldn't hand you. Ask him one thing and he turns up with a hundred and twenty-nine megabytes of everything, unsorted, in whatever order he tripped over it — and he'll report that order like it's a considered opinion. It isn't. Leaned on him once for anything at all; he looked me dead in the eye, said no matching nodes, and figured we were square.

Aider was a gentleman, which is another way of saying you couldn't file a straight question into him in triplicate. Wouldn't be asked. You mention things, loud, and hope — and sure enough, drop the right names and he's ten points sharper, which tells you everything about the names and nothing about Aider. He'd made up his mind about the place before I opened mine, and kept it after. You never did feel the conversation was with you.

Codeseek was young and had a condition. Talk to him like a human being — plain, tired, the way a man actually asks for help — and he'll look clean through you and say nothing. Sixty times out of sixty, nothing. But name the thing exactly, badge number and all, and the kid lights right up. It isn't that he's slow. He just never learned how people knock on a door.

And ripwire. Nobody at the table brought her up, on account of she'd already left. Took the question, answered it in thirteen hundredths of a second — every file I needed inside the first ten, fifty-eight times out of a hundred — keeps no office, wires no head branch, never once got herself indexed at a party. Repowise says she's abrupt.

She is. That's why I hired her.

_{Every figure above is a measured number from the round-4 table on this page: the index medians (1.24 s, 3.37 s, 34.0 s), repowise's 352 s / six-minute worst case, the 40.0% and 33.3% strict file@10, codeseek's 0-of-60 plain-language arm, codebase-memory's advertised-but-absent semantic tool, graphify's 129 MB largest graph and its 1-in-60 empty return, aider's +10 pp name-drop delta, and ripwire's 0.108 s / 58.3%. Provenance in bench/headtohead/r4-2026-08-06/ and docs/EVALS.md; only the cynicism is editorial. These are other people's real work, and the joke is aimed at the trade-offs, never the authors.}

Name a symbol and it is the first hit — and it is never a mystery which ranker answered. Every --for query is served by one of three lanes; a confidence-gated router picks by reading the query's shape, discloses its choice on the output (route=), and --no-route overrides it. Routing lifts recall@1 on name-shaped queries 61.1% → 91.3% in src/, 59.2% → 85.5% at the repository root — and the gate is load-bearing in both directions: route everything to the name lane and prose queries collapse from 0.967 MRR to 0.016. Both numbers ship together. Reproduce with ripwire <dir> --eval-retrieval, which grades EVERY doc-commented symbol in the corpus and prints its own population=/scored=/rule=; the full per-ranker tables are in bench/ANSWERQUALITY.md.

How the conceptual lane finds what you didn't name. The subtoken+body lane is why a query with no symbol name in it still lands:

• Both sides are split into subtokens. SplitChunksPlugin becomes split+chunks+plugin, and so does your query — so words match pieces of names you never typed.
• Three evidence fields, not one. A symbol is scored on its name subtokens, its doc comment, and its body — vocabulary that only exists in a comment or an implementation still finds its symbol.
• BM25 with per-query IDF. Rare, discriminating words dominate the score; words the whole corpus shares contribute almost nothing. Type the three words only the right function uses and they carry the query.
• Lookalikes are down-weighted, not hidden. Fixture, test-data, and generated paths score at a fraction, so a test vocabulary-twin cannot outrank the real source (adversarial-class pollution@5: 28% → 0%, docs/EVALS.md §4) — but they stay in the index and are still found when asked for.
• The list ends at a cliff, not a quota. The cut is adaptive: output stops where the scores drop off, so a sharp answer is a short list and a diffuse one is disclosed as such, instead of a fixed top-k padding both.

How the router picks. The gate is built on cheap, corpus-derived evidence, and its bias reflects an asymmetry the table above makes plain: a missed name-route costs a few ranks; a false one is catastrophic (0.016 MRR).

1. Identifier shape is trusted outright. A camelCase/snake token — or a short query carrying one — routes name-exact. Someone who types chooseForRanker is naming, not describing.
2. The all-words test. If every content word equals some symbol's whole name (pack task where both pack and task are real symbols), that is strong evidence of a name query.
3. The plausibility test. Present is not enough — each matched name must be specific: few definitions, and not a subtoken carried by half the corpus's symbol names (thresholds derived from the index itself, not a hardcoded stdlib list). This is what catches split chunks: every word names a symbol, but split names a String method defined everywhere — so the route is declined and the conceptual lane runs, which finds SplitChunksPlugin easily.
4. Every decision is disclosed. route= states the lane that ran; a decline names the anchor that failed and why. --no-route forces the conceptual lane when you disagree.

The proof the gate earns its keep: the routed lane matches the best single lane on both query modes simultaneously — 0.960 MRR on names (equal to forced name-exact) and 0.967 on prose (within noise of pure subtoken+body, src/). No single lane does that.

The honest boundary: the router classifies the query's shape — it cannot rescue vocabulary that is not in the index. A prose query whose concept lives only in a compound class name (SplitChunksPlugin contains no splits subtoken) is correctly sent to the conceptual lane, which then has little to grab; that gap is measured and recorded in docs/EVALS.md §7, not hidden. Numbers re-derived 2026-08-08 on this tree; per-lane table and history in docs/EVALS.md §4.

On mid-task questions it had never seen, ripwire answers at 5.0% of what a grep-and-read pass spends — 5.2% on the questions both arms fully answered. --pack-signatures returns 74.7% fewer bytes than full bodies at top-50 (re-derived on this tree, 2026-09-06). The output is already dense enough that running a dedicated context compressor over it saved exactly 0 tokens.

Both of those first two figures moved when they were re-derived on 2026-08-23, and they moved in opposite directions — 7.3% → 5.0% overall, but 1.7% → 5.2% on the both-answered subset. Same frozen questions, same frozen verb ladders, same corpus pin, same tokenizer; the naive arm reproduced to the token. What changed is where ripwire spends: the compact conceptual route made its misses much cheaper, while richer default bundles made the questions it answers dearer. Both numbers are printed because printing only the one that improved would be the failure this project exists to not commit. The full per-question re-derivation is in the Round 3 note under Measured.

It is also cheap enough to call on reflex: this repository parses in ~0.15 s cold and ~0.10 s warm (time ./build/ripwire . --no-cache), so the agent asks instead of guessing.

A map is a fixed cost paid once per context, so it pays in proportion to what that context goes on to do with it. Two shapes get the most out of it, and one gets nothing.

• Orienting a context that does not know the repository. This is the strongest case, and it is strongest of all under an orchestrator that matches tasks to models: every lane it spawns starts cold on the same tree, and each one would otherwise re-derive the same structure from scratch. The map is the one artifact that does not have to be rediscovered per agent.
• The checking pass at the end. --quality-delta, --test-gate and --edit-check answer questions — what did I make worse, which tests must run, did I change a contract — that an agent cannot answer by reading more source, at any budget.
• It does not pay on a question one grep already answers. The map is charged whether or not it was needed, so a narrow lookup in a file you can already name is cheaper without it. --help-task exists to make that call, and a flat ranking says confidence="low" rather than pretending.

1. A budget. --token-budget=N caps the bundle; --top-k=N caps the rows of the default map, --query, --format=candidates, --recall and --graph-query — it does not shape plain --for. On --for, a positive, explicit --top-k is read by nothing: the run prints a stderr note (--top-k is not read by --for) and still emits the full bundle. Two neighbours of that shape are different and neither warns: --for --format=candidates --top-k=N does consume the flag (the candidate export composes with it and caps the rows), and --for --top-k=0 is refused outright by the payload-only guard (--top-k=0 needs a payload verb), not warned-and-emitted. Narrow plain --for with its own arguments instead — --signatures-only (drop the auto-bodies), --token-budget=N (shapes the bundle to fit), --detail=N (full bodies for just the top N). Unbudgeted --for returns a rich terminal bundle by design — right when it ends the question, wasteful when it does not.
2. Routing. ripwire . --help-task="<task>" names the ONE command the task actually wants, and abstains when the evidence is thin. It is advice, it never runs anything. An answer of "just grep for it" is a correct answer and the tool will give it.
3. The skills. skills/install.sh teaches an agent when to reach for which verb. Without them an agent has 175 flags and no map of which moment each is for, and it will reach for the map every time — including the times it should not.

What ripwire does not replace — reach for grep/read here even when a verb looks close:

What is measured and what is not. The per-question figures above and in Measured are single-agent measurements. That the saving grows with the number of cold orientations follows from the fixed-cost mechanism but is not a published result — it is pre-registered and unrun. Treat the single-agent numbers as the measured ones.

--quality-panel runs the calls a good reviewer makes by hand — is this function too tangled, is it named badly, does it hide control flow inside an idiom, does its history say it keeps breaking, must you read five other files to follow it, does it mutate state three hops away — as six independent evidence families, and ranks by how many of them agree, never as one blended score. Each family implements published work — McCabe on shape, Butler on naming, Gopstein's atoms of confusion on idiom, Nagappan & Ball on churn, Beck & Diehl on colocation, Henry & Kafura on state — with the lesson taken from each paper, and the rules measured and withdrawn, in docs/LINEAGE.md. Pooled over five corpora (n = 27,889) the largest correlation between any two families is +0.168: they really are measuring different things, so two families firing on the same function is corroboration rather than one metric counted twice.

That matters most for code an agent wrote. Empty-catch error masking is +47% more common in AI-authored commits, a function rewritten again inside two weeks +15% more likely, and reuse is declining as AI's share of commits grows (GitClear, AI Copilot Code Quality, 2026). Ten of --quality-delta's 12 kinds each target one measured mode like those (the eleventh, placeholder, lists the stubs and TODOs a change adds, and never gates; the twelfth, defect-shape, catches four defect shapes review kept finding by hand — new-symbol rows never gate, except defect-shape format-arity, because a format/argument mismatch is a defect, not debt), and it reports only what your change made worse — then --exemplar shows the pattern in your own repo to copy, and --test-gate names the tests that must run before "done."

Tree-local numbers above are reproducible with the commands shown; the rest are dated, sourced measurements in docs/EVALS.md — each with its instrument, its corpus, and its counterexamples, because the losses ship beside the wins. Zero runtime dependencies, C++23, builds with the network off.

<ctx task="incremental cache invalidation" route="subtoken+body" confidence="high" margin_pct="20"
     bundle="compact" bodies="0" reason="compact-route" est_tokens="3995">
  <sigs shown="23" total="40" capped="1">
    <d l="106" n="kCacheMagic" p="src/ingest_cache.h" cx="0" ccx="0" in="0" churn="11" amp="71" pure="1" r="1"
       next="--expand=src/ingest_cache.h:kCacheMagic"><doc>incremental cache (--cache): per-file content hash + raw facts so a re-run re-parses ONLY c…</doc>constexpr std::uint32_t kCacheMagic = 0x4b505443</d>
    <d l="1307" n="spanTierMemoPath" sc="rw" p="src/ingest_astquery.h" cx="1" ccx="0" in="2" churn="5" amp="44" r="2"><doc>Composed exactly the way every OTHER blob family is (quality.h): one fixed-width identity hex pe…</doc>inline std::string spanTierMemoPath( const std::string&amp; diskPath )</d>
    <d l="247" n="ingestCommitTree" sc="rw::dmm" p="src/dmm.h" cx="6" ccx="5" in="1" churn="6" amp="27" r="3"><doc>Ingest the tree at `sha`, materialized out of `root`&apos;s object store. …</doc>inline bool ingestCommitTree( const std::string&amp; root, const std::string&amp; sha, … )</d>
    <d l="841" n="mcpRefreshedThisRequest" sc="rw" p="src/mcpindex.h" cx="1" ccx="0" in="2" churn="20" amp="43" r="4"><doc>P1-15 — the `_reingest` envelope field for a response whose handling ran an INCREMENTAL pass, …</doc>inline bool mcpRefreshedThisRequest( std::uint64_t passesAtEntry )</d>
    …
  </sigs>
  <hops shown="2" total="6" capped="1" noedge="2">
    <h l="1307" p="src/ingest_astquery.h" n="spanTierMemoPath">
      <calls total="3"><c n="shaKeyedCachePath" l="1621"/><c n="headSnapRepoHex" l="1359"/><c n="exclConfigHex" l="1554"/></calls>
    </h>
    <h l="247" p="src/dmm.h" n="ingestCommitTree">
      <calls total="10" shown="7" capped="1">…</calls>
    </h>
  </hops>
</ctx>

The cache cluster, ranked and annotated in place: cx/ccx complexity, in reuse count, churn recent commits, amp change amplification, tested coverage — the fragile spots are visible before the agent touches them, in a few thousand tokens (est_tokens="3995", self-reported in the header) instead of five whole files. This is a conceptual query, so the bundle is the compact shape: the ranked map plus one-hop callee edges, no inline bodies, and the root says so rather than leaving you to notice. Read the map, then --expand=SYM the one you want — or pass --auto-bodies to get bodies inline as before.

$ ripwire . --quality-panel --limit=1
<quality_panel preset="default" families="6" enabled_n="6" cut="2" eligible="6497" ranked="524" …>
<s p="src/graph.h:723" n="buildGraph" fam="4" of="6" fired="structural,confusion,historical,colocation">
<e f="structural" counted="1" why="ccx=764 loc=1368 nest=8 humps=34 deep=315 ev=98 rrank=1"/>
<e f="confusion" counted="1" why="atom-embedded-crement*4"/>
<e f="historical" counted="1" why="hrank=12 churn=36"/>
<e f="colocation" counted="1" why="crank=32"/>
</s>
…
</quality_panel>

Four of six independent evidence families corroborate on buildGraph, each with its own reason shown inline — never a single blended score. eligible="6497" narrows to ranked="524" (2-of-6 agreement): an 8.1% shortlist of this repository's own functions, not a guess (re-derived 2026-08-23 — the corpus grew, the shortlist share did not move). Full six-family breakdown, real numbers per family → The quality panel below.

Around the core sit 191 long flags advertised in --help, across seven families — plus an MCP server, so a coding agent can call any of them mid-task instead of grepping and reading whole files. --help prints one line per flag (~4.5K tokens); --help=--FLAG prints that flag's full entry with every caveat, --help=SECTION one family, and --help=all the whole catalog. Not sure which of them fits the task in front of you? ripwire . --help-task="<task in words>" recommends ONE executable command with the evidence behind the pick — advice only, it never runs the recommendation — and abstains honestly when the evidence is too thin to name a winner.

./build/ripwire --help is generated from the binary's own flag table and is always the authority; docs/COMMANDS.md documents every one of the 176 documented flags — 162 of them with a real invocation and its recorded output (both counts are reported by the generator that writes the document, docs/docs_commands_build.py, and were re-derived from it on 2026-09-14; test/docscommandscheck.sh fails if that documented set and the binary's own flag table ever disagree). Each family below links there.

Four reflexes worth wiring into muscle memory: --from-trace=FILE for an error you have in hand, --edit-check=SYM right after an edit (did the contract change, and which callers are now provably incompatible), --merge-scout=REF1,REF2 before landing parallel branches, and --pack-task="…" for ranking, bodies, callers and tests in one budgeted bundle.

Prebuilt binary — macOS (arm64; 0.6.1 is the last release with an Intel macOS binary, and an Intel Mac builds later releases from source) and Linux (arm64 / x86-64, built for RHEL 8+; every release is smoke-tested on a RHEL 9 userland before it publishes). Downloads the latest GitHub Release, verifies its SHA-256, and installs to ~/.local/bin. From v0.2.2 the release tarball also ships the seventeen agent skills, and the installer stages them under ~/.local/share/ripwire/skills and activates them for every agent it detects (Claude Code, Codex), printing one line per agent saying what it did. Sixteen of the seventeen are for using the tool; the one about compiling ripwire itself (ripwire-opt-remarks, audience: contributor in its front matter) stays staged unless you pass --contributor to skills/install.sh. An agent that is not installed is never given a skills directory, hooks are never registered without an explicit --hook, and RIPWIRE_NO_ACTIVATE=1 stages without activating for image builds. When no agent is detected the activation one-liner is printed instead:

RIPWIRE_REPO=redhat-et/ripwire bash -c "$(curl -fsSL https://raw.githubusercontent.com/redhat-et/ripwire/main/scripts/install.sh)"
export PATH="$HOME/.local/bin:$PATH"      # not on PATH by default on macOS or most Linux shells; add it to your rc file

Windows x64 — preview. From 0.6.3 each release carries ripwire-<version>-windows-x64.zip and its .zip.sha256. It is built with clang-cl against the static C runtime, so it needs no Visual C++ Redistributable, and like the Linux x64 binary it needs an x86-64-v3 (AVX2) CPU. CI unzips and exercises it on every train, including a byte-for-byte comparison of its output with Linux's, but no maintainer runs Windows, so treat it as a preview until Windows users report back. The exe is not code-signed. Expand-Archive does not pass the download's Mark-of-the-Web on to the files it extracts, so SmartScreen does not prompt for an exe unpacked this way; no prompt is not a verdict. In PowerShell:

$v = "0.6.3"; $a = "ripwire-$v-windows-x64"; $u = "https://github.com/redhat-et/ripwire/releases/download/v$v"
Invoke-WebRequest "$u/$a.zip" -OutFile "$a.zip"; Invoke-WebRequest "$u/$a.zip.sha256" -OutFile "$a.zip.sha256"
# extracts only when the SHA-256 matches; on a mismatch it throws and nothing is unpacked
if ((Get-FileHash "$a.zip" -Algorithm SHA256).Hash -eq (Get-Content "$a.zip.sha256").Split(" ")[0]) { Expand-Archive "$a.zip" -DestinationPath "$env:LOCALAPPDATA\Programs" } else { throw "SHA-256 mismatch: do not run $a.zip" }
$bin = "$env:LOCALAPPDATA\Programs\$a"
[Environment]::SetEnvironmentVariable("Path", "$bin;" + [Environment]::GetEnvironmentVariable("Path", "User"), "User")
$env:Path = "$bin;$env:Path"   # this window too; new windows read the user Path
ripwire --version

• The hash check passes because -eq ignores case: Get-FileHash prints upper-case hex and the .sha256 file is lower-case. For a case-sensitive compare, use .Hash.ToLower() -ceq.
• Comparing two outputs in Git Bash: use cmp. There fc is a shell builtin (it replays history) and compares nothing; the Windows tool is fc.exe, run as MSYS_NO_PATHCONV=1 fc.exe /b a b so /b is not rewritten as a path.
• Git for Windows is needed for the git-history features (churn, --situ, the git row of --doctor) and for the skills installer. The map itself runs without it.
• ripwire . --doctor: every row should read ok="1". binary-path finds ripwire the way PowerShell does (Path in order, then PATHEXT). When it fails, which= names the ripwire that Path finds first, and which_version= is what that one prints for --version. same_bytes="unknown" means a copy could not be read; the row fails as unverified, so compare the two with Get-FileHash. The cache lives in %LOCALAPPDATA%\Temp\ripwire-<uid> (your TEMP). With TMPDIR, TEMP and TMP all unset, Windows' own temp-directory rule falls back to your profile folder, so the cache is %USERPROFILE%\ripwire-<uid>; the cache-dir row names the directory either way.
• Checking cache reuse: in PowerShell $env:RIPWIRE_CACHE_STATS=1; ripwire . > $null (in Git Bash RIPWIRE_CACHE_STATS=1 ripwire . > /dev/null) prints one cache-stats line on stderr. Run it twice with no other ripwire build in between: reparsed=0 and reused= equal to files= mean every file came from the cache. warm_growths= is a performance counter, not a reuse fact: it counts how often a worker thread's buffer had to grow, which depends on how the threads split the files, so it varies from run to run by design.
• Agent skills (Claude Code, Codex): from Git Bash, in the unzipped folder, bash skills/install.sh (Claude Code) or bash skills/install.sh --codex. From PowerShell, name Git Bash by full path, & "C:\Program Files\Git\bin\bash.exe" skills/install.sh: a bare bash there is often WSL's (C:\Windows\System32\bash.exe), which installs into the WSL home, where Windows agents never look. Or copy them by hand in PowerShell: New-Item -ItemType Directory -Force "$HOME\.claude\skills" | Out-Null; Copy-Item -Recurse -Force "$bin\skills\ripwire-*" "$HOME\.claude\skills\" (Codex reads $HOME\.agents\skills). --hook is untested on Windows, and scripts/install.sh (the curl installer) does not run there.

Or build from source. Requirements: CMake 3.24+ and a C++23 compiler — that means clang 16+ / AppleClang 15+ (Xcode 15) / gcc 13+, and if your distro's CMake is older than 3.24, pip install cmake or brew install cmake gets a current one everywhere. Nothing else — tree-sitter's core, the grammars listed in THIRD_PARTY.md and the test framework are vendored under third_party/deps, so there is no download step and no package manager to satisfy. Prove that with the network off: add -DFETCHCONTENT_FULLY_DISCONNECTED=ON and the build still completes.

Two builds, two jobs — pick by what you are doing:

# building to USE it — the fast binary (Release implies LTO; scripts/pgobuild.sh adds PGO, what CI ships)
cmake -S . -B build-release -DCMAKE_BUILD_TYPE=Release && cmake --build build-release -j

# building to WORK ON it — plain configure, no build type (why that matters: the trap, under the fold below)
cmake -S . -B build && cmake --build build -j

See Languages for supported source and document formats.

To put it on PATH, ./install.sh builds and atomically installs the binary plus the matching skills/ and hooks/ assets into a detected prefix (Homebrew's if present, ~/.local otherwise; override with RIPWIRE_INSTALL_PREFIX). Set RIPWIRE_ACTIVATE_CODEX=1 to also refresh Codex's skill links and advisory hooks from that same staged version; activation is otherwise explicit.

The CLI is the recommended baseline because it works in every shell-capable agent; the MCP server is optional, for agents whose workflow benefits from persistent tool registration. Full walkthrough, all six clients: Agent setup.

The trap, spelled out: never configure the dev tree (build/) with -DCMAKE_BUILD_TYPE=Release. Release defines NDEBUG, which compiles the degrade-path alerts out and blinds the gates that assert them — and every gate and bench number is measured against build/, so changing that tree's flavour silently moves all of them at once. Release belongs in its own tree (build-release/ above; ./install.sh builds its own build-install/ the same way). CI builds both flavours on purpose — see CONTRIBUTING.md.

The honesty contract, in one line: every count ripwire cannot prove is a total ships labelled a floor, every truncation is disclosed where it happens, and a zero means none found — never none exists. Two runs over the same tree are byte-identical, and a warm run equals a cold one. That is a contract, gated on every pull request and every push to main, not a tendency. The full discipline — and the losses published next to the wins →

Six independent evidence families, ranked by how many of them agree — never one blended score. Pointed at this repository's 4,956 eligible functions, 2-of-6 agreement leaves 401 worth a second look: an 8.1% shortlist. Pooled over five corpora (n = 27,889) no two families correlate above +0.168, which is what makes agreement corroboration rather than one metric counted twice.

--quality-panel joins six independent evidence families (structural shape, lexical naming, syntactic-confusion idioms, git churn history, cross-file colocation, and non-local mutable state) and ranks by the count of families agreeing, never a weighted composite — averaging correlated metrics and calling it several is the Maintainability Index's well-known failure mode. On this repository the panel measures 4,956 eligible functions and narrows them to 401 worth a second look at 2-of-6 agreement — an 8.1% shortlist, not a guess — and the largest correlation between any two families, pooled across five independent corpora (n = 27,889), is +0.168: the families really are measuring different things. The six-family table above shows what each finds on this repository's own source; the sections below are the parts that need more than a row.

The structural family's nest= reports the single deepest line in a function. One line at depth 9 and a thousand lines at depth 9 produce the same number — which means a long blocked-sequential body (a run of shallow scoped steps, its max set by one inner loop nobody has to hold in their head) is indistinguishable from a tangled one that sustains depth for hundreds of lines. Every consumer of nest= inherited that blindness: the panel's structural family, --biggest-first's rank, the ensemble join.

--metrics now emits the profile beside the max — humps= (how many maximal regions reach the nesting bar, CodeScene's "bumpy road": a rise above the threshold then a fall, so repeated missing abstractions read differently from one deep tangle) and deep= (how many lines lie inside them, against the loc= already on the row). Both come from the same fused walk that already computes ccx/nest, at zero extra parsing cost; deep= is a disclosed floor (deep_floor="1"). Both are absent exactly when nest < the bar — not-deep, never a hidden 0. deep counts lines and humps counts regions, and two regions can share a line — a one-line if(c){x;}else{y;} at the bar is two regions on one line — so deep below humps is legal output rather than a defect. On this repository's own source:

loc and nest alone rank main and dispatchMcpLine beside ingest and buildGraph; the profile separates them, and it promotes ur_walkTree — 87 lines, so no size bar fires, yet proportionally the densest thing in the table. This changes no ranking: humps > 0 is exactly nest >= bar, which is precisely when the nest bar already fired, so the family count and the panel's shortlist are untouched. It is strictly more evidence on rows that already appear — a reader can tell the two shapes apart without opening the file.

The panel also carries one join, and it is deliberately not a seventh family: a row whose structural evidence includes deep= (a body that sustains depth at the nesting bar) and that no indexed test reaches is annotated join="deep+untested" — the pair where a refactor is most wanted and least safe, put side by side because both facts are already on the row. It changes nothing: not fam=, not of=, not the order, not which rows appear (counting it would be the structural family wearing a second hat). The root reports tested_scope= (symbols any indexed test reaches — the join's honest denominator) and deep_untested= (rows carrying the annotation across the whole row set); at tested_scope="0" no indexed test reaches anything here, so "untested" would be a fact about what was crawled rather than about the code, and the annotation is emitted on no row.

Two verbs sit beside the panel, not inside its six-family join — worth knowing the boundary rather than blurring it:

• --field-affinity (cache-friendly co-access) is explicitly not a panel family, by unit: it measures which struct fields are read together but declared far apart, and its subject is a type, not a function — attributing a struct's finding to the functions that touch it would be a claim the lens itself never makes (docs/EVALS.md §9.9.2). MainDispatch (src/main.cpp:1335, the 144-byte struct threaded through nearly every verb) still carries 12 real findings against a separation cost of 92.88 — fields read together in the same call routinely cross cache-line boundaries. Chilimbi, Davidson & Larus's cache-conscious structure definition (PLDI 1999), validated on real hardware counters; the advice-not-transform posture (report a split, never auto-reorder a struct) follows Hundt, Mannarswamy & Chakrabarti (CGO 2006). This is a genuinely rare kind of tool: code review catches cache-unfriendly patterns constantly, but every adjacent tool that reasons about memory-access shape (Intel Advisor's pattern classifier, DMon, PerfLint) does it by running the program first — a two-round adversarial literature and patent search found no shipping tool and no published work that does this statically, before a line executes (tier: RARE BUT REAL, the full citation trail and hedged claim wording in docs/LINEAGE.md). The same verb now also classifies each loop's access shape — index/handle-based (predictable, the hardware prefetcher can hide the latency) vs. pointer-chase (data-dependent, no struct layout fixes an unhideable per-hop stall) — and, for genuine chases, checks whether the pointer you dereference to reach the next node sits next to the payload you're about to read, since that cache-line fetch is unavoidable and colocating there is a strictly higher-value fix than generic field reordering. On this repository: 1,374 loops classified, 5 genuine pointer-chases found in a codebase deliberately built handle-based rather than pointer-linked (guardrail G2) — the lens staying quiet on code written to avoid the problem is itself a check that it isn't firing at random. Ships entirely report-only: an A/B benchmark against a real 64 MB shuffled linked list measured a mostly-null result, so the ranking-affecting half of this feature is a provable no-op until a blind real-corpus validation session clears it — reported here at the same honesty level as everything else in this table, not oversold ahead of the evidence.
• The cache-* lint pack + --with-profile (cache-friendly access patterns — the other half of the locality story, shipped 2026-08-07/08) covers what --field-affinity deliberately does not: eight AST shapes practitioners agree hurt — node-based containers, vector<T*>/vector-of-indirect (the "matrix as vector of vectors"), heap allocation inside loops, p = p->next chase advances, a[b[i]] gather subscripts, by-value shared_ptr parameters, and existing manual prefetches flagged for re-measurement — loop-fenced by span algebra, C-family only, facts never verdicts. The honesty numbers, both directions: on this repository the aggressive rules fire only in benches and test fixtures, zero in shipping src/ (guardrail G2 holding is itself the check the rules aren't firing at random), and a 13-agent adversarial triage of all 327 findings confirmed zero as fix-worthy — every plausible refactor died on "win unmeasurable without a profile". That gap is exactly what --lint --with-profile=FILE closes: it joins a RIPWIRE_PROFILE build's own per-scope hardware counters (#PROF_TSV) onto findings, so a row carries heat_total_ms / heat_l1d_mpki from a real run — static shape × measured PMU weight, the two halves of SYZYGY's advice mode (Hundt, CGO 2006) finally in one command. Worked example: the one surviving refactor candidate (flattening the Louvain adjacency) was settled by its new PROFILE_SCOPE in a single measured run — 5.2 ms, 5.9% of the verb — a wasted afternoon prevented by a number (docs/CACHELINT.md holds the full catalog, the wave-2 specs, and the compiler-handled myths deliberately not checked).
• --biggest-first (renamed from --readability, which still works — a stderr note points scripts at the new spelling) is a sibling lens, not a panel family either — the one classic model in the tree with a published closed form: Halstead volume (Halstead, Elements of Software Science, 1977) and the Posnett/Hindle/Devanbu sigmoid fit (MSR 2011, doi:10.1145/1985441.1985454), fitted on snippets of 20 lines or fewer — past that the fitted score saturates and only the ordering stays meaningful, which is exactly how the verb is used: largest-first, never as a grade. Halstead's volume specifically (not the later, less-trusted difficulty/effort derivatives) is among the metrics shown to track measured cognitive load directly (Peitek, Apel, Parnin, Brechmann & Siegmund, ICSE 2021, doi:10.1109/ICSE43902.2021.00056) — --biggest-first emits volume and stops there; difficulty and effort are computed nowhere in this tree. A ranking lens, never a grade, and here is what it actually orders: on ripwire's own history at the pinned v0.6.2 tag (412 function pairs mined from 80 refactor/simplify/cleanup commits), the order between two versions of a function followed the sign of its token-count change in 96.0% of pairs — read a move as more or fewer tokens, not as more or less readable. Of the 412, 154 (37.4%) ran the commit's implied direction and 258 (62.6%) ran opposite it; Halstead volume drove 91.1% of those 258; a separate self-consistency check confirms the formula computes exactly what it says it computes, so this is a construct-validity finding about the ordering claim, not an arithmetic bug. (First recorded as 484 pairs / 30.2% from an unpinned git log --all walk that does not reproduce — full derivation, the instrument fix and the retracted figure: docs/EVALS.md §8.) WITHDRAWN: the ordering claim is withdrawn — stratified into narrow token-count bands, the later-fix association disappears in 8 of 10 deciles (CIs include 1), so the order this verb produces is better explained as a residual size proxy, measured in the lens's own units, than as an independent readability signal (derivation: docs/EVALS.md §8). The lens's computation is unchanged pending a proper human study, and --help=--biggest-first carries the same figures where a CLI reader meets them.
• --naming-consistency is the lexical family's one exception to "evidence, never advice": every other lens in this panel tells you WHAT is wrong, never a computed fix. Case-style consistency is the one property with a corpus-derivable answer — on this repository's src/, camelCase is the dominant convention at 1,677/1,803 (93.0%) agreement, and the verb flags 136 off-convention names with a mechanically recombined propose= value for each (no dictionary, no synonym judgment — see What it answers).
• --lint --naming-locals points those same naming rules at local variable names — the thing a human reviewer flags immediately in a sprawling function and no static tool measured until this session. Opt-in, off by default: on this repository, a plain --lint finds 2,225 findings; adding --naming-locals finds 3,198 — 973 findings that were structurally invisible a moment ago, scoped tightly (only inside functions already flagged large/complex, only locals nested two blocks deep for the short-name rule) so it doesn't just relabel every loop counter in the tree. Ships disabled by default on purpose: this repository's own history includes a naming rule that shipped on plausibility and was later measured to flag its best-named functions — see the withdrawn-rule note below — so a rule this new stays opt-in until a real-corpus audit clears it.

Full citation table, evidence tiers, and what got measured and withdrawn (a naming rule that flagged this repository's best-named functions, kept as the standing argument for measuring before shipping) → docs/LINEAGE.md.

Four real invocations against this repository, each printed as the binary actually prints it: --callers (and why its count is a floor), the default ranked map (and what amb= admits), --test-gate (exit 4 while obligations remain), and --from-trace (feed it the error itself, not a paraphrase of one).

Output is minified — one line, no whitespace between tags — so the excerpts below are wrapped for reading, and each one's leading legend comment is elided. Nothing else is edited, except that corpus-size numbers (file/symbol/edge counts, the ranked-map header's token/ambiguity tallies, PageRank k= values, and the test-gate example's script_gates_unmodelled= — a count of the script runners under test/, recursively) drift as this repository grows: the ranked map elides those specifically, and says so again at the point of use, and the test gate additionally trims its <u> rows down to 2 of the 25 the real run prints, behind a trailing ….

Ten seconds, no index server, no embeddings, no API key — a parse and a call graph, built on the spot. The rows below are a real capture: the callers and their files are gate-held current (test/readmeexamplecheck.sh), the :line suffixes were true when captured and drift as the files grow — nothing can keep a line number true in a document, so it is not claimed here. p= is where each caller is defined; sites_at= lists its call sites as pasteable file:line tokens, so the lines to edit come with the first answer.

$ ripwire . --callers=rankGraphTeleport
<callers of="rankGraphTeleport" defs="1" count="7" root="." hop_tested="0" hop_untested="7" counts_floor="1">
<s t="fn" n="getIndex" p="src/mcpindex.h:1165" sites_at="src/mcpindex.h:1268"/>
<s t="fn" n="rankGraph" p="src/graph.h:7965" sites_at="src/graph.h:7968"/>
<s t="fn" n="anchoredLexicalRank" p="src/graph.h:8611" sites_at="src/graph.h:8662"/>
<s t="fn" n="runEval" p="src/eval.h:171" sites_at="src/eval.h:325"/>
<s t="fn" n="churnDecayRanking" p="src/main.cpp:1383" sites_at="src/main.cpp:1397"/>
<s t="fn" n="churnRankedGraph" p="src/main.cpp:1422" sites_at="src/main.cpp:1436 src/main.cpp:1437 src/main.cpp:1451"/>
<s t="fn" n="runDefaultMap" p="src/main.cpp:1632" sites_at="src/main.cpp:1752"/>
</callers>

counts_floor="1" is the point. Call edges are extracted from source text by name, so dynamic dispatch contributes no edge (a call through a function pointer or callback is an edge only when ONE function is bound to that variable in scope and the variable never escapes — its address taken or reference-bound — and a macro-generated call site — tagged role="macro" — only when its function-like #define is indexed): count="7" is a floor, and the element says so before you read a single row.

The ranked map — the default run, capped to three symbols so it fits here:

$ ripwire . --top-k=3
<!-- files=… symbols=… edges=… shown=3 est_tokens=… ambiguous=… unresolved=…
     precise=… skipped_oversize=… order=important-first -->
<r est_tokens="435">
<f p="./src/infra/svector.h" layer="infra">
<s t="method" n="size" sc="svector" k="…"></s>
<s t="method" n="push_back" sc="svector" amb="2" k="…">
<c n="buf" l="…,…"/><c n="grow" l="…"/></s>
</f>
<f p="./src/scipoverlay.h">
<s t="method" n="empty" sc="ScipOverlay" k="…"></s>
</f>
</r>

files=/symbols=/edges= and the k= rank values are elided: this repository is the corpus here, so they move every time README.md itself gains or loses a line, which is not what the example demonstrates. The rest of the header measures the whole corpus, not the excerpt — the ambiguous= tally is the call-graph completeness gauge, and amb="2" on a row says two of that symbol's calls hit a name with several definitions and the resolver guessed. Read the source when which-target matters.

The test gate — --test-gate names the obligations and exits 4 while any remain. Captured with an uncommitted change in the tree: changed="1" and the rows below appear only because something was actually pending. A clean clone exits 0 with every changed/impacted/test count at zero. The structural counts describe the tree, not git status, so they stay nonzero even then: script_gates_unmodelled= (script-to-binary test runners the call graph cannot see), script_gates_registered=, script_gates_mapped= and script_gates_unresolved_dynamic= (the suite's registered shell gates, and how many of them map to their dependencies), and the resolver gauges graph_ambiguous=, graph_unresolved= and graph_unindexed=. The capture below predates the script_gates_* registry counts and those gauges:

$ ripwire . --test-gate          # exit code: 4
<test-gate changed="1" impacted="80" tests="2" untested="76" shown_tests="2" tests_capped="0"
           shown_untested="25" untested_capped="1" script_gates_unmodelled="332" at="9cf0b16f3+dirty">
<t p="./test/adaptivecutshapefix/adaptive_cut_shape_test.cpp" run="bash test/adaptivecutshapecheck.sh"/>
<t p="./test/verify_radix.cpp" run_unknown="1"/>
<u sym="buildGraph" p="./src/graph.h" ccx="712"/>
<u sym="dispatchMcpLine" p="./src/mcp.h" ccx="428"/>
…
</test-gate>

A run= attribute appears only when a runner is derivable from real evidence — a test-dir script whose stem matches the harness, or whose text names it; for TypeScript/JavaScript, a file already recognized as test code (test//tests//__tests__/, or a _test./.test./_spec./.spec. name — the __tests__/ directory convention is shared with every language, the rest are TS/JS's own) whose name ALSO matches vitest/jest's own test-name shape (.test./.spec., or living under __tests__/ at all, jest's own two default conventions — never a .d.ts declaration file) — whose nearest package.json names vitest, jest, or node's own test runner. Its own package.json is AUTHORITATIVE for the whole subtree below it: a non-empty, non-placeholder scripts.test entry decides the runner (or decides none — a same-named dependency never overrides it, whether that dependency sits in the same manifest or a parent one), and only a manifest with NEITHER a real scripts.test NOR a vitest/jest dependency is skipped in favour of one further up. A TS/JS file with no manifest evidence of its own still falls back to the SAME test-dir shell/Python driver search every other language uses — a named driver beats an unguessed default. A row with none says so — run_unknown="1", never a guessed suite command — and a <t> or <g> row carries one or the other, never neither. A <g hops="2" n="3" p="a,b,c" run_unknown="1"/> row is two or more contiguous runner-less rows whose attributes are byte-identical, served as one: n= is how many, p= is their paths verbatim in list order, and the disclosure is paid once per group rather than once per row. Everything else stays its own row — a row with a run=, a row whose attributes differ from its neighbour's, and a path containing a comma, which is never grouped at all, so p= splits on , into exactly n= paths. A shown=/total= over these rows counts test FILES: a <g> row is n= of them.

script_gates_unmodelled="332" is the same discipline: script-to-binary is not a call edge, so those gates are invisible to this walk, and the number says so rather than letting tests="2" read as complete. The <u> rows are the untested blast radius: impacted symbols that no test in the corpus reaches — never a file's own <file-scope> module scope (a top-level statement or anonymous-callback body), since nothing in any language can name it and no test could ever be written for it; it still counts toward impacted= when it is a real caller in the blast radius, just never as an obligation. Those excluded owners are counted, not dropped without a trace: untested_modscope="N" (always present, alongside untested=) says how many, so a change whose only reader is an untestable entrypoint discloses why untested= reads zero instead of looking like there was nothing to find. The capture above predates that attribute (and a few other root attributes added since), so its root lacks it; today's root carries untested_modscope="N" immediately after untested=.

A TS/JS file with NO manifest evidence of its own (no package.json anywhere in the crawl boundary, or the nearest one decides nothing) still gets one more chance before falling to the shell/Python driver search above: its own bytes are read for a node:test import or require — import test from "node:test", import { test, describe } from "node:test", require("node:test"), single or double quoted (#60). This is a real parse, not a substring scan: a "node:test" mention inside a comment or an unrelated string literal is not evidence. An explicit package.json scripts.test (or vitest/jest dependency) still wins over this — including an authoritative-but-unrecognized script, which is a decided "no" — so the import fallback only ever fires on what was previously run_unknown="1".

For a .ts/.mts/.cts file — whether the runner was named by scripts.test: "node --test" or inferred from the import above — a run= command is spelled only where the checks below find nothing that makes Node refuse to start it. They are read off the repository's own bytes, so they cannot see the Node that will run it, and a refusal is run_unknown="1", never a guess:

• .tsx and .jsx never get a command. Node's type stripping does not cover .tsx at all (ERR_UNKNOWN_FILE_EXTENSION), and plain node cannot load a .jsx file at all, on any Node version — both stay run_unknown="1" unconditionally.
• Every relative import/require must resolve exactly as written, in the test file and in every local TypeScript module it reaches. Node's module resolver, under type stripping, never probes an extension and never maps a .js specifier onto a .ts source — the exact shapes tsc-, tsx- and bundler-run TS code uses to import its own siblings. So a command is spelled only when every relative (.//../) static import/export … from specifier or require(...) argument names a file that exists on disk at that exact path. The walk follows each specifier that lands on a .ts/.mts/.cts file, reads at most 64 modules, and a walk cut at that bound stays run_unknown="1" too.
• No module on that walk may use syntax type stripping cannot erase. Node strips types and rewrites nothing, so an enum, a namespace with runtime code (or holding only declare statements), the legacy module M {} keyword, a constructor parameter property, an import alias (import A = B.C, import x = require(…)), export =, an angle-bracket assertion <T>x or a decorator stops it with ERR_UNSUPPORTED_TYPESCRIPT_SYNTAX (or a parse error) before any test runs; any of them outside a declare stays run_unknown="1". A type-only namespace and a declare enum are erased and do not count.
• The command is additionally Node-version-aware, from engines.node. --experimental-strip-types itself exists from Node 22.6 only (an older Node refuses to start at all with it); stripping is ON BY DEFAULT — the flag becomes a harmless no-op — from Node 22.18 and separately from Node 23.6 (two floors, not one continuous range: 23.6 turned it on first, the 22.x line got it later by backport, and a bare 23.0–23.5 does not have it). This tool cannot see which Node will run the emitted command, so it reads engines.node from the nearest manifest: the bare node --test <file> when that range proves every satisfying Node has stripping on by default; the flagged node --experimental-strip-types --test <file> when it proves >= 22.6 but not provably default-on, or when there is no manifest at all (a stated assumption of a Node that strips types with the flag, not a guess at an unseen runtime); and run_unknown="1" when the range admits ANY Node below 22.6 (a plain >=18/^20, or a compound range like >=24 || ^20, where every || alternative must pass on its own) or cannot be read with confidence at all. Each alternative is read for its upper bound too, because the default-on versions have a gap: ^22.18.0 gets the bare form, but >=22.18 also admits 23.0–23.5 and keeps the flag.

Every node:test file, .js or TypeScript, must also load as the module kind its own syntax needs. A file with a static ES import/export runs as an ES module only as .mjs/.mts, under "type": "module" in its nearest package.json, or on a Node with default module-syntax detection (22.7 and later, 20.19 on the 20.x line). Without one of those the command stays run_unknown="1": an explicit "type": "commonjs" (or a .cjs/.cts file) turns detection off, and with no "type" the engines.node range must prove detection — >=22.7 does, >=20.19 does not, since it admits 21.x. With no engines.node at all, a TypeScript file keeps the stated assumption above and a .js file stays run_unknown="1", since its command otherwise rests on no assumption at all. A CommonJS file (require, no ES import/export) runs either way.

.js/.mjs/.cjs never need the flag. Their one version question is the runner itself: the --test flag exists from Node 18.1 and, by backport, 16.17 — never on 17.x, and 18.0 has the node:test module but not the flag. So engines.node must not admit a Node below 18 without it: ^16.17.0, >=18 and >=18.1 get the bare form, while >=16, >=16.17 and 16.17 - 18 (they reach 17.x) stay run_unknown="1", and so does a range confined to 18.0.x such as 18.0.x. A hyphen range A - B is bounded by B. >=18 is accepted although it admits 18.0: that is one release from April 2022, the command fails loudly there, and refusing the most common spelling would make the answer useless on most projects. No engines.node at all keeps the bare form.

A TS/JS run_unknown="1" can still mean the manifest genuinely names nothing recognized (and the test file itself names no node:test import either), one of the refusals above, or a real runner this tool does not yet derive: node's own test runner invoked through tsx (when neither the manifest nor the test file's own bytes name node:test directly) and bun's test runner. These stay an honest unknown rather than a guess; pnpm/yarn-prefixed test scripts derive correctly today (spelled npx …, which finds a local binary first).

From an error, not a paraphrase of one — --from-trace takes a stack trace, sanitizer report or compiler error on stdin or from a file, maps its frames onto indexed symbols innermost-first, and returns the innermost in-corpus body with them:

./build/ripwire . --from-trace=asan_report.txt
cmake --build build 2>&1 | ./build/ripwire . --from-trace=-

Every published number lives in docs/EVALS.md with the instrument that produced it, the corpus it ran on, and the in-tree file that pins it — alongside a counterexample section and a list of the claims this project deliberately does not publish. Read those first if you are here to check whether the tool is oversold.

Round 4 (re-run in full 2026-08-08): 58.3% strict file@10 against 40.0% for the best competitor — a 1.46× margin, at a 0.31 s index. N = 60 paired instances, zero exclusions, one binary (the profile-guided release build that now ships), one evaluator, all arms re-run on the same day. Strict file@10 = all gold files inside the top 10.

Round 4 (first run 2026-08-06; re-run in full 2026-08-08 on the profile-guided release build) — N = 60 paired instances, zero exclusions, one binary, one evaluator. Every competitor was re-run on the same day against the same ripwire binary and the same evaluator, so the arms are directly comparable to each other. Earlier rounds could not say that: r1 (2026-07-13/14) and r2 (2026-08-03) each scored a different subset against a different binary, which is why this page used to print two tables and ask the reader not to compare them. Strict file@10 = all gold files inside the top 10. Harness, per-instance JSONL, and a reproduction recipe: bench/headtohead/r4-2026-08-06/.

Paired, ripwire's losses are 2 instances to codebase-memory-mcp, 2 to repowise, 1 each to graphify and aider; codeseek never beat it. Cold from nothing to an answer — parse, rank, reply, no cache — ripwire takes 0.213 s, against a ~34 s index-then-query for repowise, whose worst single index in this run was 352 s. That comparison is only possible because ripwire's own index was measured this round; r2 recorded competitor index walls and deliberately refused to tabulate them, since a one-sided cost table is not evidence.

What re-running cost us, stated because it is the reason to re-run at all. codebase-memory-mcp is the true runner-up at 40.0% — r1 credited it with 26.7%, and scoring it fairly against today's binary raised it. graphify rose 21.7% → 31.7% and aider 13.3% → 20.0% the same way. The margin over the best competitor is therefore 1.46×, not the 1.75× that two separately-dated tables implied; the old framing flattered us by comparing today's ripwire against year-old competitor runs.

Three limits travel with this table. ¹ repowise's walls include a fresh MCP-server spawn per query; resident-server usage is faster. ² codeseek's raw row returned 0 results on 60/60 queries — its keyless fallback matches function names only, so that row measures a query-protocol boundary, not its embedder-backed shipping mode (unbenchmarked here). And the slice is Python-dominant — 107 of 134 gold files are .py — so this measures Python localization, not every supported language. Vexp and CodeIndexer were excluded, not beaten: their free tiers (node/project/chunk caps) cannot run a fair 60-instance sweep. An independent adversarial pass attacked the r2 comparison's design, and its findings and dispositions ship with that report (VERIFIER.md); r1's and r2's tables remain in their own directories as the historical record.

Multi-file gold is hard for every arm, including ours, and four attempts have failed to fix it. ripwire leads the stratum at 21.4% strict — but its own any@10 there is 78.6%: it finds a gold file and misses the siblings. If each gold file were an independent draw at the single-file rate (90.6%), multi-file would score 71.2%; it scores 21.4%, so the failures are strongly correlated. The mechanism is that a sibling is a file which changed because the primary changed — the primary carries the issue's vocabulary, the siblings carry the consequences. Four pre-registered rounds (r1_anchorhop, r1cpp_anchorhop, r4_siblift, r5_pooling) have now been rejected at ±0.00pp against it, each archived with its grid and its verdict under bench/locbench/results/.

Round three (2026-08-03): headroom — the compression-layer competitor. headroom (headroom-ai==0.33.0, 64k★) compresses context an agent already fetched; it retrieves nothing — so this round's instrument is tokens-to-correct-answer on 12 pre-registered mid-task questions (django @ pinned commit, five arms, one tokenizer), not file@k. ripwire won every measure the two tools share. headroom's default config passed every code chunk through byte-identical — its own protective guards fired throughout, netting −410 tokens on a 685,682-token workload (its own limitations page says "Code — Passthrough"; this run confirms it live) — and stacking it on ripwire's output added exactly 0 tokens of savings: the map is already past the density compression targets. As measured then, ripwire answered at 7.3% of the naive grep-and-read baseline's tokens (1.7% on the subset both arms fully answered), warm in ~0.14 s per verb. Those two figures were re-derived on 2026-08-23 and now read 5.0% and 5.2% — the block below. The losses in this round are ripwire's own, and they are published first: under the frozen no-human verb ladders it strictly satisfied only 5/12 questions vs the naive baseline's 11/12 — four ranking defects, one missing symbol kind, two harness artifacts, each bucketed with its fix disposition in the report. What the round does not show: headroom's home turf (JSON/log tool-output compression, provider-cache economics) was deliberately not measured — ripwire does not compete there. Provenance: docs/EVALS.md §2, full record + adversarial verification (which materially corrected the draft's arithmetic in headroom's favor) in bench/headtohead/r3-headroom-2026-08-03/.

The 2026-08-03 round is frozen; this is the same instrument pointed at today's binary. Arms A (idealized grep-and-read) and C (ripwire's pre-registered verb ladders) were re-run from the committed harness.py, importing its metric functions unmodified — same questions.json, same frozen arms_spec.json ladders, same tiktoken cl100k_base counter, same corpus pin (django/django @ 70f39e46). The headroom arms (B/B′/D) were not re-run; they contribute nothing to the ripwire-vs-naive ratio, and the round's finding about them is a passthrough result that does not depend on ripwire's side.

The naive arm reproduced exactly — all 12 questions token-for-token, 685,682 t in total — which is what makes the comparison a measurement of ripwire's change and not of drift somewhere else. Satisfaction did not move either: 5/12 for ripwire, 11/12 for naive, and the both-satisfied set is the identical {q02, q03, q04, q05, q11}.

Why they moved apart. The compact conceptual route made the seven questions ripwire misses dramatically cheaper — q07 13,198 → 4,416 t, q08 11,192 → 2,016 t, q01 11,974 → 3,045 t — which is what pulls the overall share down. On the five it answers, richer default bundles cost more: q05 522 → 6,736 t and q11 280 → 1,373 t dominate that column. The 58× figure was always the flattering framing of a 5-of-12 result, and the honest reading of the pair is that ripwire got much better at being wrong cheaply and somewhat more expensive at being right. Nothing about the four ranking defects, the missing symbol kind or the two harness artifacts in the original loss buckets has been re-adjudicated here — the satisfaction column is unchanged, so those losses stand.

LocBench held-out, N = 243 across 78 repositories. Strict file@10 60.9%, against 27.6% for the pre-routing baseline — a paired +33.33pp with a clustered-bootstrap 95% lower bound of +25.00pp, bought for +3.4% warm latency and −39.4% on the production token ceiling. More accurate and cheaper, which is why it shipped.

Read the first row's methodology before quoting it: element bytes are counted root-neutralised, with the corpus-root prefix subtracted from both sides, because the root repeats inside every element, is charged in both forms, and is not what this verb elides. Quote the top-50 figure — the signature payload is top-50 whatever --top-k says, and top-10 is a ten-symbol sample that one one-line accessor can move several points. The gate fails if the documentation drifts more than 1.5 points from the binary.

And the third row's caveat is not small: it was measured 2026-06-20 on a large private C++ corpus, it is not publicly reproducible from this tree, and it proves cheaper and faster, not better outcomes.

The losses ship next to the wins. --grep's 2026-06 anti-headline (+19.7% / −11.2% vs raw grep) was re-derived 2026-08-15 after the emission overhaul: the un-tiered view — --grep-in=any, which was the default when this was measured — lands at −1.9% / +4.3% of a clean grep dump while carrying enclosing symbols and in-band honesty grep cannot give — one term either side of parity, which is not a headline in either direction — but an uncapped dump of that same un-tiered view remains token-negative (+150% / +214%), and this project still says so. (Span tiers landed 2026-08-19 and made the default a filtered view of the same scan; every byte number on this page is the un-tiered one, deliberately — the instrument may not move under a display change — so none of them describes today's default. See docs/EVALS.md §grep.) Nor is that the worst regime: on small-hit uncapped answers the fixed per-answer legend is not amortized over anything, and a gated set of twelve such queries on this repository runs +65.1% median, larger than plain grep on 12 of 12. When the agent knows a second term, --grep=A --and=B flips hard positive (−77.7% measured, complete both sides). --pack-signatures inverts on a short symbol: 303 bytes of signature-plus-doc-comment against a 158-byte body. The headline is a property of large result sets, and the full counterexample list is part of the contract, not an appendix — the four-fact grep re-derivation lives in docs/EVALS.md §5.

Every pipeline phase is bracketed by two hardware-counter reads, and the numbers say what wall-clock cannot. Parse and query retire 3.2–3.5 instructions per cycle; the graph phases stream at 32–35 L1D MPKI and still hold IPC above 3.1 (guardrail G2 doing its visible job); and a warm run replaces the dominant phase's 8.74 B instructions with a 52.3 M-instruction cache load, ≈167× fewer. Two opt-in builds go faster still — PGO by 14–25% cold — with byte-identical output on every one.

The self-profiler (-DRIPWIRE_PROFILE=ON; src/infra/profileScope.h + profilePmc.h) brackets every pipeline phase with two hardware-counter reads — kperf on Apple Silicon, a pinned perf_event_open group on Linux. Below: ripwire mapping its own public tree on an Apple M5 Pro, cold (--no-cache) except the last row. Counters are raw integers, never scaled; reproduce steps in bench/PROFILE.md (arming needs root; unprivileged runs print the same table with timing columns only — the honest degrade). The report renders on stderr; stdout stays the deterministic XML map, so redirects and | xmllint pipes behave the same in the profile flavour.

Three things the counters say that wall-clock alone cannot. The parse/query phases are compute-dense, not stall-bound: 3.2–3.5 instructions retired per cycle at ~3–4 L1D misses per thousand. The graph phases stream hard — 32–35 L1D MPKI — yet hold IPC above 3.1, which is the DOD/SoA/CSR layout (G2) doing its visible job. And the auto-cache's whole story in two rows: a warm run replaces the dominant phase's 8.74 B instructions with a 52.3 M-instruction cache load, ≈167× fewer instructions. Caveats travel with the table: counters are per-thread and aggregated per scope, so rows must not be summed across scopes (a recursive site samples only its outermost frame); one machine, one corpus — re-run on yours. Backend contract is gated by test/pmccheck.sh; M5 Pro event names verified (the last-level alias resolves via PL2_CACHE_MISS_LD).

Every number on this page is the DEFAULT build — and there is a faster one you can opt into. A clang optimization-remarks pass over src/ (-DRIPWIRE_OPT_REMARKS=ON; the whole triage is in docs/OPTREMARKS.md) found that the phases above spend their time calling tree-sitter's C API across a translation-unit boundary — 831 of 1,437 distinct inline/NoDefinition sites in the hot TU name a ts_* accessor. Two build options answer that, both off by default:

Measured by interleaved A/B (A,B,A,B,…, median and min, 9–31 runs per arm, repeated) on this repository and on a ~2000-file C++ tree that appears in no training run — the held-out corpus shows the same or larger gain, which is what rules out training on the benchmark. Output is byte-identical on all three builds and on both corpora, and the determinism gate passes on every tree; these options change how fast the answer arrives, never what it is. Cold gains 2–3× more than warm because cold is a branchy walk over tree-sitter's parse tree, while warm already runs in the cache-tuned CSR/SoA/B+tree structures G2 exists to produce. LTO is on by default — it costs link time and nothing else, and link time is not what this tool is optimized for. PGO needs a training run, so it stays a driven build (scripts/pgobuild.sh) rather than something a bare cmake --build does behind your back.

And on machines with no PMU at all — most cloud VMs and CI boxes — the counter columns no longer vanish. A kernel that refuses every hardware event (ENOENT; no vPMU is the common cloud case) still offers software counters, so the Linux backend's per-event graceful skip now extends to the group leader and falls through to two PERF_TYPE_SOFTWARE rows — task-clock (on-CPU ns) and page-faults — under their own column names, never as a stand-in for the hardware counts. Below, ripwire mapping its own tree, cold, on the 2-vCPU vPMU-less Intel Xeon VM this was validated live on:

Four things this buys that wall-clock alone cannot say. The wall−task-clock gap is off-CPU time: the parse phase's 36% gap is 2-vCPU oversubscription made visible (a pool parsing 799 files on two cores). The CSR row is a live cross-check: task-clock is the kernel's clock, the wall column is the profiler's own — two independent clocks agreeing to 0.05% on a single-threaded scope. (Short hot scopes diverge by the documented read-bracket overhead instead: the counter bracket encloses the tick bracket.) The page-faults column is a G2 witness: PageRank's power iteration retires with zero page faults and the CSR build with 7, against ~3,100 in the allocation-heavy model-build scopes — the no-allocation rule inside the ranked loop, watchable on a box with no PMU. And the doc post-pass row caught something real: 2.17 s of wall on 0.5 ms of CPU is work happening outside the process — with markitdown installed, the showcase PDF and PPTX are re-extracted by subprocess on every run, cache or no cache, which on this box is ~97% of a warm run's wall (2.05 s of 2.11 s); child CPU is invisible to every per-thread counter, and the wall-vs-task-clock gap is precisely the signature that flags it. (The extraction is already documented in-tree as a pure function of the file bytes — a cache candidate, now measured.) Only a kernel that offers nothing at all — perf_event_paranoid>=3, seccomp — still degrades to timing-only, and pmccheck's inactive arm now proves that was truly the case.

Almost none of the ideas here are new; the combination and the constraints are. Lessons folded from 49 repositories and 71 papers into one deterministic executable, alongside a labelled survey of 237 tools that folded nothing and are catalogued separately — the two sets are disjoint, so they add rather than nest. The row-by-row ledger, each with the lesson taken and where it lives, is docs/LINEAGE.md. Those three counts are derived from that document's own tables by test/readmedriftcheck.sh, which fails if this sentence and those tables disagree.

One of those papers measures the alternative architecture directly: an independent study found a planner that delegates exploration to a sub-agent in an isolated context window losing to search over a pre-built index — 65.2% vs 46.2% accuracy, at less than half the cost — with 41.8% of the delegated arm's failures happening silently at the planner→sub-agent hand-off (arXiv:2608.01507). A single process answering in one call has no hand-off to fail at.

A name-based call graph cannot see dynamic dispatch, a callback routed through a table, or a symbol that exists only after macro expansion. Every static tool has that horizon; the one that costs you a bug is the one that hides it. So the contract is not it sees everything — it is nothing it is unsure about reaches your agent unlabelled.

Measured against a compiler-grade oracle, its silent-miss count is zero. Of 68 answers scored against a scip-clang index, six were imperfect and four flagged themselves; the other two were right, and the oracle was the one that could not see the files. No imperfect answer arrived unmarked.

When a mark says the cheap answer is not enough, escalate on purpose — never by guessing:

1. --expand=SYM — read the body it ranked, still a fraction of a whole-file read.
2. --uses=SYM / --impact=SYM — every read, write and import site, and the transitive blast radius. --callers alone under-counts, and says so.
3. --scip=index.scip — hand it a compiler-grade index and precise edges replace the name-based guesses, tagged prov="scip". A path that is missing, empty or not a regular file refuses (exit 1); a corrupt index warns on stderr and proceeds name-based.

The map for this repository's own src/ grades itself in the header before a single answer (2026-08-10, at 4e245d7): files=109 symbols=3233 edges=10132 ambiguous=4768 unresolved=1097.

Read ambiguous= as a property of the indexed tree, not a score for the resolver. It counts call names that hit more than one definition, so it moves when the corpus gains same-named symbols — and it just did, visibly. On 2026-08-09 this number rose from 2991 to 4507. The cause was attributed by swapping one file back and re-measuring: src/infra/svector.h alone accounts for 4507 → 2741. The small-vector there grew a full container interface, so the tree now defines its own size, begin, end, push_back and data — the most-called names in any C++ corpus. Adding a container is a name collision by construction. Expect the number to climb again as that container replaces std::vector across the tree; a rise there means the map is disclosing more overloading, not resolving worse.

A 34-query sweep of this repository, 68 answers across --uses and --callers, scored against a scip-clang index (2026-08-09, the shipping binary). Six answers were imperfect; four carried a discriminating self-flag (amb=, defs>1, external="1"). The two unflagged were not misses: the oracle could not see those files — a Python file under bench/, an uncompiled file under test/ — and ripwire's answer was the correct one.

It errs the other way instead. 14 of 62 correct answers carried a caution flag they did not need — 22.6% over-hedging, up from 18.6% before this round's resolver fixes, because closing loss buckets added marks faster than it removed them. That bias is deliberate: a flag you did not need costs you one source read; a miss you were never told about costs you a bug.

One corpus, 34 queries, and the oracle has blind spots of its own — the limits that travel with the number are listed with it. Receipt, per-query scores and reproduction: bench/headtohead/r9-2026-08-09/.

Three rules the output enforces on itself: a zero reads as none found, never none exists; every truncation says what it withheld and how to page it; and every count names its unit, because the units differ by verb.

• A zero is a measurement, not an absence. counts_floor="1" marks every count that name-based resolution cannot prove is a total. Read a zero as none found, never none exists.
• Truncation is disclosed where it happens. shown_*, *_capped= and the paging attributes say what was withheld and how to page it; the legend comment that leads each document defines the vocabulary in full, so the output explains itself without this README.
• Units are named, because they differ by verb. The callers count is distinct symbols, the impact count is a reach set, the uses count is call sites. The legend says which, so two numbers that look contradictory can be read as the different questions they answer.

The evaluation labels were authored by reading the source and deciding which symbol is the on-task answer — never by transcribing the ranker's own output — so the eval is allowed to say the ranker is wrong, and it has. These are the results that say so, all in-tree, all published on purpose:

• --grep costs more tokens than it saves, and --pack-signatures can make output bigger — both quantified next to the savings they qualify, because saying so is cheaper than being caught.
• The public C++ number is materially lower than the retired private one was. SFML: strict file@10 28.7%, any@10 41.7%, first-hit MRR 0.21 — measured at commit d411f3de4 (bench/cppbench/results/sfml_scoreboard.md). Until 2026-08-07 this bullet compared against a roughly-89%-any@10 private corpus that is no longer reproducible from this tree; that comparison is retired. The public number is the baseline going forward.
• PageRank is a bad co-change ranker — 3.8% recall@5 against 40.3% for plain lexical, and fusing the two made it worse. Relatedness is lexical; importance is structural; the tool uses different machinery for each because the measurement said so.
• Strict multi-file localization is hard and stays hard. Held-out LocBench: single-file gold 73.4%, multi-file 18.2%. Every corpus shows the same cliff.

test/regression.sh names 684 gate scripts and is the authoritative list; python3 test/pargates.py . ./build/ripwire -j 6 runs the same set in parallel. On top of them sit the contracts that do not fit a unit test: two runs byte-identical, warm output identical to cold, output that pipes clean through xmllint --noout, a sanitizer build with -fno-sanitize-recover=all, and a differential argv harness that runs a reference binary and the candidate over every argument vector and requires stdout, stderr and exit code to match on each.

The house rule behind all of it: write the gate before the code it measures. A ranking, a token estimate and a call graph all look plausible whether or not they are correct.

If your agent can run shell commands, it is already set up. The CLI is the primary interface: ripwire on PATH costs the agent nothing until the moment it runs a command — no server to register, no tool schemas riding in every request's context. The one step that matters is installing the skills (step 2 below), which teach the agent when to reach for which verb.

The MCP server is the optional second interface, for what a shell pipe can't give you: clients without shell access, lazy body handles (fetch a symbol's source only when actually needed), and the same span-addressed edit operations against an already-warm index. The CLI exposes those edits too, with the same staleness/ambiguity refusal and atomic-write contract:

ripwire . --replace-symbol-body=SYM --edit-payload=definition.txt
printf '%s' "$BLOCK" | ripwire . --insert-before-symbol=SYM --edit-payload=-

That MCP convenience has a cost the CLI doesn't carry — the verb schemas sit in the agent's context every session — so register it when you want those verbs, not as a default.

One stdio server, 33 verbs — 17 read, 13 flagship-reflex, 3 span-addressed edit — and a client that isn't one of the six above can be pointed at the same process by hand.

That registers one stdio server — ripwire --mcp — exposing 33 verbs: 17 read verbs, 13 flagship-reflex verbs, and 3 span-addressed edit verbs. Read verbs mirror the CLI (analyze, for, grep, cochange, fetch_body, lego, mentions, owners, memory_recall, situational_awareness, batch, …); find_symbol and find_referencing_symbols attach a stable handle instead of a body, so the agent fetches source only when it actually needs it. The edit verbs mirror the CLI flags and enforce the shared safety contract — staleness refusal, ambiguity refusal, mode preservation and atomic writes. Full reference: skills/ripwire-mcp/.

Before printing, wrap security-scans ./skills and .agents/skills with the same engine as --scan-skills: a CRITICAL finding blocks the recipe, warnings print and continue.

If your client is not one of the six, the server is a plain stdio MCP process and every client that speaks MCP can be pointed at it by hand. The whole configuration is:

{
  "mcpServers": {
    "ripwire": { "command": "ripwire", "args": ["--mcp"] }
  }
}

Use an absolute path in command if ripwire is not on the agent's PATH. For a client that wants a socket instead of stdio, ripwire --listen=HOST:PORT serves the same verbs.

skills/ ships seventeen task-shaped skills that tell an agent which verb answers the moment it is in — orienting cold, tracing a call, sizing a refactor, checking a diff, hunting a bug, writing tests, reviewing security. Without them an agent has 33 verbs and no map of when each applies; the skills name the moment each verb is for. Install as symlinks back into this repo, so edits here take effect immediately:

skills/install.sh                 # → ${CLAUDE_CONFIG_DIR:-~/.claude}/skills
skills/install.sh --codex         # → ${AGENTS_HOME:-~/.agents}/skills (canonical Codex/agent path)
skills/install.sh --codex --hook  # → also install Codex's task router, CLI nudge + session primer
skills/install.sh --codex-legacy  # → ${CODEX_HOME:-~/.codex}/skills (older Codex installs)
skills/install.sh /some/path      # → an explicit destination
skills/install.sh --contributor   # → also the contributor-facing skill (compiling ripwire itself)
ripwire --scan-skills=skills      # read the security scanner's verdict first, if you would rather

The script's own header documents its other modes, including the opt-in advisory PreToolUse hook. wrap also prints a pasteable use-when blurb for your client's rules file (CLAUDE.md, AGENTS.md, .cursor/rules, …) and works from a prebuilt install with no checkout — from v0.2.2 it points at the installer's staged copy of the skills when the cwd is not a checkout.

How to run one:

1. Build the tool first — most loops need a binary to measure against: cmake -S . -B build && cmake --build build -j
2. Open your coding agent at the root of a ripwire checkout.
3. Paste the contents of one prompt file. It needs nothing else from that directory.
4. It writes a plan and stops for your go-ahead. Nothing runs before you approve it — read the plan, cut what you disagree with, then say go.
5. The loop runs its own gates in the foreground and reports what it left green.

The other ten — a paired head-to-head against a competitor, a ranking-eval loop that mines real retrieval misses from your own sessions, a per-language improvement pass, a zero-context onboarding study, a sibling sweep, a live command tour, a showcase build, the path a new language's grammar actually took, a quality-panel calibration round, and a COBOL two-pass corpus measurement — are listed with their audiences in prompts/README.md. Each states its own scope and its honesty rules, and most name the gates they must leave green.

After a session on your own repository — your first one counts, and counts most — hand your agent prompts/improve-for-my-language.md. It harvests that session's own transcript — where ripwire answered, where it missed, where you fell back to grep — and every finding it produces has to cite the moment it came from: what you asked, which command ran, what came back. Open an issue with the result.

That is worth more to this project than a bug report, because it arrives in the form the project already runs on: evidence with its provenance attached, not an impression. Several languages here are one contributor's corpus away from being measurably better, and we cannot see your code.

C, C++, Objective-C / Objective-C++, Metal (Metal Shading Language, .metal — indexed with the C++ grammar, since MSL is a C++14 dialect, so a dual-compile header's symbols resolve from both the GPU and CPU halves), CUDA (.cu/.cuh — indexed with the vendored tree-sitter-cuda grammar, so kernel<<<grid, block>>>( … ) launch sites are real call edges and --callers of a kernel names its host-side launchers; __constant__ module tables index as symbols even uninitialized — the cudaMemcpyToSymbol idiom — and SCREAMING_SNAKE __device__/__managed__ globals join them; dual-compile .cuh headers resolve from both halves), Python, TypeScript, JavaScript, Java, Ruby, PHP (.php/.phtml — classes, interfaces, traits, enums, functions, methods and constants; $o?->m() null-safe and A::m() static calls are edges; use directives are imports. Dynamic dispatch — $fn(), call_user_func, __call — names its callee at run time and is a stated floor, not a silence), Lua (all five spellings that define a function, including the M.f = function and table-constructor forms; function M:f() is a method. Metatable inheritance is a runtime call with no syntax to read, so a Lua corpus reports no inheritance edges — stated, not implied), Dart (.dart — classes, mixins, extensions, enums, typedefs, functions, methods, getters/setters; recv.m(), recv?.m() and cascade ..m() invocations are edges. Two stated floors: named constructors and factories index under the CLASS name, so C(), C.seeded() and factory C.fromA() are overloads of C; and noSuchMethod dynamic dispatch names its callee at run time. The grammar makes a function body a SIBLING of its signature rather than a child, so the definition span is extended through it at capture time — without that, every call in a body attributes to the enclosing class), Elixir (.ex/.exs — nested modules, structs, protocols and implementations, functions, macros, guards, delegates, types, callbacks, attributes and literal ExUnit tests; module/name/arity resolution with lexical aliases, filtered imports, default arguments, captures and pipes; see the static-analysis limits), Kotlin (.kt — classes, objects, companion objects, interfaces, enum classes and functions, extension functions included; bare and navigation calls and constructor delegation are call-graph edges; an import is a dependency edge only (role="import" on --uses, never a --callers/--impact edge — T13/fix3, 2026-09-20). Kotlin and Java share one call graph, and a call reaches the other language only when its own defines no candidate of that name, so adding .kt files never moves a Java edge. Stated floors: an explicit receiver (A.f()) does not narrow candidates; a multiplatform expect/actual type pair is two candidates; .kts is not indexed; and a file nesting string templates past 128 levels is refused and listed by --skipped — see the Kotlin limits), GDScript (.gd — a Godot file is a class body: class_name names it and its file-scope func/var are its members, with inner classes, constants, enums and their members, signals and call edges; preload/load produce no dependency edge yet, and .tscn/.tres/.gdshader are not indexed — see the GDScript notes), Clojure (.clj/.cljs/.cljc/.bb — namespaces, vars, functions, macros, multimethods, protocols, records and types; direct symbolic list-head calls are edges. Quoted, syntax-quoted and reader-discarded forms are excluded. Macro expansion, dynamic invocation, namespace aliases/import resolution, protocol/type method symbols, and exact multi-arity signatures are stated floors — see the Clojure limits), Bash, Go, Rust, Swift, C#, JSON + TOML + YAML (config keys — a [tool.ruff.lint] table is one symbol under its full dotted name, and pyproject.toml/Cargo.toml/CI workflows become greppable), and Markdown (.md/.markdown — the DOC tier: every heading, ATX or setext, is a section symbol whose span runs to the next same-or-higher heading, so --for ranks the section, --expand serves the section body, --recall answers section-granular, and links/backtick mentions are doc→doc and doc→code edges). All grammars are vendored; THIRD_PARTY.md owns their inventory and provenance.

Want another language? The pipeline is language-agnostic past the parse: a new language is a vendored tree-sitter grammar, its query file, and entries in the declarative extension → { grammar, queries } table (the "declarative constexpr tables" rule in CONTRIBUTING.md). Some grammars also need capture filters: Elixir, for example, represents definitions as ordinary calls. Open an issue naming the grammar and the repo you'd run it on.

Notebooks, HTML and CSV are indexed as documents for --recall and the doc↔code edges behind --mentions; Office and PDF join them through an optional bridge. Markdown graduated from that tier: it parses with its own vendored grammar, so its headings are symbols, not just document text.

The counts come from docs/LINEAGE.md. Counts are current as of 2026-09-14. Seventeen of the folded papers are from 2026, seven published in the last two months, and three in the last thirty days. The survey describes 237 tools that contributed no lesson. The two sets are disjoint, so the counts add.