GRASP.
Grasp · Comprehension Index Scan Report

Who still understands crewAIInc/crewAI?

Computed from git metadata only; no source code contents were read.
touch-weighted · full history, 3-year analysis window · last commit 14d ago · scores are comparable only against same-mode scans

Grasp Score
85
touch-weighted living knowledge across the repo
Active humans
28 / 350
contributors active in last 90 days vs. all-time
AI-attributed minimum
≥ 6.2%
hard evidence only. Just 19% of commits here carry any attribution signal, so the real share is higher
Inactive share
7.1%
work by authors with no commit in the window; silence, not proof of departure
Composition
15/6/78/1%
source / tests / docs / config share of scored work
Living, source only
74%
the headline recomputed over hand-written source alone
Score stability
moderate
90-day-cliff variant reads 82%–94% across 60/90/180-day activity windows; read the width before the middle value
Squash-merged history: 57% of the commits analysed here collapsed a whole pull request into one commit by one author. Squashing credits a single person for everyone's work and can strip the co-author trailers that name the rest (measured across heavy squashers, roughly a quarter to a third of squash commits still carry them, and those are credited), so for this repository the concentration figures are an upper bound, the bus factors are a lower bound, and the AI-attribution minimum is lower than it would be with the same work merged conventionally. That information is destroyed at merge time and cannot be recovered from the repository, so it is reported rather than corrected. Comparisons against a project that does not squash are comparisons across two different record-keeping systems.

Dark-code ranking

Modules ranked by comprehension risk: the unknown share, amplified by ongoing change and thin ownership. The green meter is living knowledge.

ModuleRiskLiving knowledgeAI-attr. InactiveBusHeat/180d
docs/imagesLOW 42
52%
5%59%244
docs/conceptsLOW 42
44%
5%67%10
docsLOW 37
63%
3%42%187
docs/toolsLOW 37
51%
0%62%10
docs/guidesLOW 34
55%
7%58%10
docs/how-toLOW 32
57%
1%48%10
docs/enterpriseLOW 31
58%
0%52%10
docs/enLOW 30
65%
2%36%3203
docs/pt-BRLOW 30
66%
2%36%3177
docs/koLOW 29
66%
2%34%3176
.githubLOW 27
71%
5%25%284
docs/arLOW 24
76%
3%4%1394
docs/core-conceptsLOW 24
69%
0%30%10
(root)LOW 23
75%
4%20%2160
srcLOW 20
67%
3%28%30
libLOW 19
78%
8%8%34225
docs/v1.15.3LOW 17
83%
0%0%11107
docs/v1.15.4LOW 17
83%
0%0%11111
docs/v1.15.5LOW 17
83%
0%0%11111
testsLOW 15
75%
5%20%40

One commit touched 42% of every file changed in this window. Wide, shallow commits, a formatter run or a license-header pass, add weight for their author across the whole surface while changing little. Worth looking at what that commit was before crediting the score to its author.

One commit accounts for 32% of the analysed lines. In a repository this size a single large change can move the score on its own, and git cannot tell a thoughtfully written thousand-line change from a thousand lines a formatter rewrote. Worth looking at what that commit actually was before reading much into the number; the touch-weighted view, which counts each file-touching commit once, is the cross-check.

Public-repo caveat: "departed" includes drive-by contributors, overstating attrition relative to a company repo with a real roster. Engagement is a proxy for comprehension. Treat it as a potential indicator rather than a diagnosis.