Repository navigation
ask without the repo-root positional silently wipes the graph: freshness check anchors on cwd and rebuilds 0 nodes (cards survive) #438
Description
Version: @nanonets/graft 0.18.0 (npm global, /opt/homebrew), Node 24, macOS arm64 (darwin 25.6.0)
Summary
Running graft --dir <context-dir> ask "<query>" without the repo-root positional silently wipes the graph: ask anchors the repo root on the current working directory, the pre-query freshness check concludes every file changed, and it rebuilds an empty graph (0 nodes) — while the per-file card markdown survives, masking the damage. Exit code is 0 throughout.
Reproduction (self-contained, ~10 s)
# sources: a small git repo outside the "hub"
mkdir -p /tmp/gresrc/repo/src && cd /tmp/gresrc && git init -q repo
printf 'export function utilFn(n: number): number {\n return n * 2;\n}\n' > repo/src/util.ts
printf 'import { utilFn } from "./util";\nexport function mainFn(): number {\n return utilFn(21);\n}\n' > repo/src/main.ts
# "hub": a separate git repo holding only the context dir + a symlink to the source repo
mkdir -p /tmp/grepro/graft/demo && cd /tmp/grepro && git init -q .
ln -s /tmp/gresrc/repo myentry
graft build --dir graft/demo myentry # ✓ wiring: 4 nodes, 4 edges, 2 cards — wiring.json 2440 bytes
# the buggy form: no repo-root positional
graft --dir graft/demo ask "who calls utilFn" # [graft] refreshed the graph (2 files changed) → empty answer, exit 0
wc -c graft/demo/.graph/wiring.json # 232 bytesObserved after the buggy invocation:
[graft] refreshed the graph (2 files changed) before answering
graft ask — "who calls utilFn" (empty)
⚠ structural index: no entries for 'utilFn' — showing lexical matches; ...
no matching nodes — try different words, or `graft build` if graft/ is empty
graft/demo/.graph/wiring.jsonshrinks 2440 → 232 bytes:nodeCount: 0, edgeCount: 0, languages: []graft/demo/src/util.md/main.md(the per-file cards) are untouched — a user who checks the context dir at a glance sees nothing wrong- exit code 0
- the only tell before the answer degrades is one info line: "refreshed the graph (2 files changed)" — phrased as if the sources changed, when in fact the refresh just re-rooted the graph on the wrong directory
Expected vs actual
- Expected:
askresolves the repo root from the graph's recorded source root (what the originalgraft build --dir graft/demo myentrywas given), answers structurally, and touches nothing. If it cannot reconcile cwd with the recorded root, it should warn/fail, not rebuild. - Actual: the freshness check treats cwd as the repo root. In the repro (and in our real layout) the cwd's file set is unrelated to the graph, so the refresh re-roots and rebuilds empty, destroying a valid graph in under a second.
Why this is easy to hit
The hub-of-graphs layout that triggered it for us: one git repo that is a graph hub, graft/<entry>/ context dirs, and <entry> symlinks to sibling source repos — one graph per repo. The natural muscle memory is graft --dir graft/<entry> ask "<query>"; the required positional is easy to omit. We hit this on a 9-repo hub (graphs of up to 1,651 nodes) — one ask emptied a graph while the surviving cards made it look intact.
Recovery
Both of these restore the graph in place:
graft --dir graft/<entry> ask "<query>" <entry> # corrected ask form rebuilds it before answering
graft build --no-gitignore --no-ignore --dir graft/<entry> <entry>(The 232-byte wiring.json is the emptied signature.)
Suggested fix
Anchor the pre-query freshness check on the graph's recorded source root rather than the cwd — e.g. persist the root (and/or its identity) that the last build/ask used in .graph/wiring.json meta or a sibling file, and compare against that. If cwd ≠ recorded root, either warn and re-anchor, or refuse with a clear message like "graph was built for ; pass it as an argument" — never rebuild from an unrelated cwd into the graph directory.
Bonus safety net: a rebuild that would take a non-empty graph to 0 nodes could require confirmation or at least print a loud warning — "silently emptied" is the worst failure mode here because the surviving cards mask it.
Activity
[-]ask` without the repo-root positional silently wipes the graph: freshness check anchors on cwd and rebuilds 0 nodes (cards survive)[/-][+]ask without the repo-root positional silently wipes the graph: freshness check anchors on cwd and rebuilds 0 nodes (cards survive)[/+]fix(graph): refuse to refresh a graph from a foreign root, and scream…
hr1juldey commentedon Nov 18, 2026
Fix up as a draft: the graph now records its source root, and the pre-query freshness check refuses to re-root it anywhere else — the issue's hub layout (symlinked entries) is handled by realpath, and old graphs keep the old behavior exactly. PR: #857.
fix(graph): a query rooted elsewhere never re-roots the graph (trailh…
fix(graph): a query rooted elsewhere never re-roots the graph (trailh…