Repository navigation
Minor. sigmap advise: what SigMap can tell you offline about how to give an agent SigMap in your project, a guide to which setup saves what, and --emit-root so the paired harness can be run on your own project (#1067, #1068). Nothing in extraction, ranking or the index changed; advise reads the project and its git history and writes nothing (--emit-root writes only under the directory it is given). The agent study's Stage I, run before any of it was built, found that no offline number predicts which repository saves, so advise reports facts and makes no call.
Added
sigmap advise: offline facts for choosing how an agent gets SigMap, and a guide that says which setup saves what. Read-only, no model call, nothing written, no index generated (without one it says to runsigmapfirst). It prints what each setup adds to a turn in your project (thelocate,coreandalltool definitions and the hook's per-prompt block, chars/4), then mines questions from your git history by the rule that built the agent study's held-out set (the title of a first-parent commit that touched one to three source files, the title naming none of them;--since,--max) and reports how often SigMap's first five files hold the changed file against the honest benchmark's single-shot grep (a Wilson 95% interval on each, and the headroom between them), and whether the prefetch hook would fire on those titles as prompts. It makes no call and prints no forecast, and--jsoncarriescall: nullandforecast: null: the study's Stage I tested, with its bar written before any feature was computed, whether any of these numbers predicts a repository's saving, and it did not (Spearman ρ between headroom and the change in real cost ofprefetchover twenty repositories −0.13 [−0.57, +0.36], bar ≤ −0.50 with the interval below zero; nothing else tracked it either).advisethen quotes what the study measured by goal. The mining rule and the grep scan are ports ofscripts/build-heldout-set.mjsandscripts/lib/grep-baseline.mjs, which are not in the package; a test pins that the scan ranks identically to the benchmark's. 18 tests (advise.test.js); 37 mutants of the rule, the scan, the cut at five, the headroom sign, the interval and the quoted study numbers were each applied to a temporary copy and each fails a test.sigmap advise --emit-root <dir>: the paired harness's inputs for your own project, private or not. The guide's "measure it paired" route built its questions withscripts/build-heldout-set.mjs, which clones public GitHub repositories only.--emit-rootwrites the manifest (the project pinned toHEAD), one task per mined question (with the commit that made the change) and a link to the project under the directory you name, then prints the dry-run and run commands with the path shell-quoted. It writes nothing in the project, refuses a directory inside it, one it did not write and one whose link slot holds a real directory, and refuses a project whose.gitis a file (a linked worktree: the harness resets a copy of the project, and git would write into the repository the file points at); uncommitted changes are reported (every arm is reset to the pinned commit). The real harness's dry run accepts the output on both agents, and the study'sretrieval-rank.mjsprepares real arms from it and leaves the project untouched. 13 tests (advise-emit.test.js) and 26 mutants of the writer, the printed commands and the CLI.- Choosing a setup: which way to give an agent SigMap, by goal. On the twenty held-out repositories (Messages-API agent, Haiku 5.5) the prefetch route was the only setup that cut time (19 of 20 repositories, −18% [−21, −14]) and tokens (−26% [−39, −14], none used more) and saved cost on 7 and cost more on 5;
strategy: "index"with thelocateprofile was the only one that cost more on none (−7% [−13, −2]) but saves no time and raises total tokens;indexalone raised tokens on 17 of 20. What not to use, where it saves more and less, and three ways to find out for your project, fromadviseto the paired harness (with the commands, checked with--dry-run). The installed hook is about a third of the benchmark'sprefetch(time −11%, tokens −23%, cost −4% in Claude Code on eight repositories). - Stage I of the agent study, in the protocol: plan and bar committed before any feature was computed (
1c508cc), the test script before it was run (ab240eb), the result after (553eb93;benchmarks/agent-study/setup-signal-eval.mjs, output indata/setup-signal-eval.txt). The primary test does not pass. The by-goal table (saves / no clear change / costs more per setup and measure, and the pooled change) was added after the test and is marked as descriptive; its cost rows reproducesummarize.mjs --by-repo.
Changed
- Published metrics: none moves. Hit@5 (80.4%), the task-success proxy (61.9%), prompt reduction (44.4%), token reduction (95.2%), test-discovery F1 (98.0%), the third-party gate (118 of 208, no task changed hit status) and grounding coverage (38,117 of 38,122 references, 43 repositories) hold, and every benchmark report reproduces v8.86.0 leaf for leaf apart from its version stamp, except the honest comparison. Its headline is the same (grep 49 of 125, SigMap 116 of 125, 2.37×); the count of top-5 places taken by SigMap's own generated files reads 133 of 625 where 134 was published, and an untouched merged
developreads 133 where thev8.86.0tag reads 135, so the movement belongs to the merged code and documents, which the scan reads.hardreads 62 of 90 (68.9%) against a floor of 68.0%, 0 tasks of headroom. The npm package goes from 230 to 234 files (4026 to 4095 KB unpacked).