No existing tool proves full junk-level repo outlines plus before/after change semantics; a local code graph plus an agent workflow is the honest answer
Asked (summary):Find — do not create — an existing DeepWiki-like skill for a private repo, for two tasks: (1) a full repo outline covering every component, method, class, config, artifact, log and probable junk; (2) a change outline covering the task, abstractions touched and used, new symbols, roles, call chain, data flow before/after, what looks abandoned or excess, and what could or should have been touched but was not. Score discovered skills 1–100 with backing evidence, and output ranked feature matrices plus vanity metrics for each goal.
The research merged thousands of web-search results across GitHub, tool directories, vendor pages and independent reviews on 2026-09-13, then kept a 15-candidate shortlist per goal — not an exhaustive census of every tool. Scores are documented-feature coverage against a 100-point rubric: 100 means the searched source made claims in every rubric bucket, not that accuracy was audited. All rows below are visible in the heatmap and tables.
Rubric coverage, ranked: subscores per tool for each goal
Cell shade = share of the bucket's points claimed in the source; the number is the points. Amber left border = the row's evidence for that goal is mostly generic rubric wording, not a concrete product statement — treat the score as weak. Hover a tool name or cell for the actual evidence text.
Eleven of 15 Goal 1 tools tie at 100/100, yet only RustCodeGraph's source literally claims configs, artifacts, logs, binaries and probable junk in its inventory wording — and even that is an untested claim (github.com).
CodeGraph carries the most concrete Goal 2 evidence: git diff scope, symbol/call/import/route graph, named flow-edge types (flows_to, mutates, def_use), live dead-code and cycle detection, fully local — with an honest caveat on metaprogramming and dynamic dispatch (github.com).
GitNexus is the strongest blast-radius specialist: consumers before a return-type change, depth-grouped upstream/downstream dependencies with confidence scores, PDG source-to-sink data flow, local indexing — under a PolyForm Noncommercial license (github.com).
Hosted PR reviewers are better evidenced for automated review but not for the ask: Qodo, CodeRabbit and Greptile show concrete cross-service impact tracing, yet require hosted or enterprise deployment and none proves a complete before/after semantic model (qodo.ai).
Only docmap is a literal SKILL.md agent skill in either shortlist — and it outlines markdown, PDF and YAML documents, not source code; every other candidate is an MCP server, CLI, plugin or hosted service an agent skill would invoke (pkg.go.dev).
Recommendation by use case
Most complete local/private outline
CodeGraph or Codanna
Local-only, symbol-to-route graphs, broad agent integrations; pair with GitNexus for blast radius. No tool independently proves logs/junk coverage — have the agent walk ignored and untracked files itself.
Broadest language support
Serena
Over 40 languages via real language servers (LSP), 16.1K+ stars; limited to what each language server exposes and weak on external-dependency declarations.
Literal installable skill/plugin
docmap
The only SKILL.md in the set — but for docs (Markdown/PDF/YAML), not code. For code, install a graph tool's MCP server and wrap it in your own thin skill file.
Strongest change-impact graph
CodeGraph + GitNexus
CodeGraph for flow-edge semantics and dead code; GitNexus for confidence-ranked, depth-grouped dependency tracing and PDG source-to-sink findings. Both static: reflection and dynamic dispatch stay partly invisible.
Strongest turnkey PR review
Qodo / CodeRabbit / Greptile
Concrete cross-repo impact evidence and enterprise self-host or air-gap options, at per-seat pricing; they review diffs well but do not deliver a full before/after data-flow model.
What not to expect
DeepWiki alone
No discovered product credibly enumerates every symbol plus configs, artifacts, logs, binaries and junk while also reasoning about missed changes. The workable strategy is a code-graph tool plus an agent workflow.
Evidence table — both goals (30 rows)
Click a header to sort. Goal 1 rows list outline evidence; Goal 2 rows list change-analysis evidence, with local graph tools before hosted PR reviewers. Vanity metrics are popularity, not quality — captured from pages with differing dates. Rows flagged generic evidence mostly repeat rubric wording; their 100s are coverage ties, not proof of equal quality.
Method: two 15-row shortlists (30 rows total) compiled 2026-09-13 from thousands of merged web-search results across GitHub, tool directories, vendor pages and independent reviews. Each score out of 100 sums rubric subscores measuring whether the cited source documents a claim in that bucket; it is not an audited accuracy benchmark, and logs/junk coverage was not independently tested. Vanity metrics are popularity signals from pages of differing dates. Long evidence and language lists are truncated in tables for space; hover the heatmap for full evidence text. Sources span multiple hosts; no single URL backs more than half the rows.