Install and Setup
`entire graph` prints the main Entire help instead of running
`entire graph` prints the main Entire help instead of running
graph as a command, so it falls back to its own help output.Queries fail with a `git cat-file --batch-command` error
Queries fail with a `git cat-file --batch-command` error
PATH:`entire graph version` prints `dev`
`entire graph version` prints `dev`
scripts/install-local.sh or a manual go build. That is expected during development and not expected otherwise.Reinstall a released build:`plugin install` refuses the download
`plugin install` refuses the download
checksums.txt. A release that publishes no checksums is refused.If you trust the source and understand the tradeoff, pass --allow-unverified. Otherwise check that you have the right repository URL.Installing from a URL that is not in the plugin index also asks for confirmation. In a script or CI, add --yes.`init-agents` fails before writing anything
`init-agents` fails before writing anything
init-agents checks both instruction files before its first write, so a failure means nothing was changed.Two things cause it. First, AGENTS.md or CLAUDE.md is not a regular file. A directory, pipe, socket, or device is rejected, and the error names the type it found. Second, a file has a broken marker layout. Each file must contain either zero entire-graph marker tokens, or exactly one begin marker followed by one end marker.Marker tokens count even inside code fences, examples, or HTML comments, because they make the replacement range ambiguous. To recover:- Back up
AGENTS.mdandCLAUDE.md. - Replace any non-regular target with a regular file, or move it aside.
- Reword any stray marker strings that appear in examples or comments.
- Rerun
entire graph init-agents --repo ..
Your edits to `.entire/graph-agent.md` disappeared
Your edits to `.entire/graph-agent.md` disappeared
init-agents run. Manual edits there are not preserved by design.Put your own instructions in AGENTS.md or CLAUDE.md, outside the <!-- entire-graph:begin --> and <!-- entire-graph:end --> markers. Everything outside those lines is left alone.Remove entire-graph from a repository
Remove entire-graph from a repository
.entire/graph-agent.md, then remove the section entire-graph manages in AGENTS.md and CLAUDE.md. You can delete .entire/ or either instruction file if it is otherwise empty.To also remove entire-graph from your machine, run:Your Agent Is Not Using entire-graph
The agent greps instead of running `entire graph search`
The agent greps instead of running `entire graph search`
@ imports. Claude Code loads CLAUDE.md at session start and resolves the import. Other clients may treat the block as plain text, in which case they still get a written instruction to read .entire/graph-agent.md, but whether they follow it depends on the client and the model. Verify the behavior rather than assuming it.`CLAUDE.md` seems to have the wrong block
`CLAUDE.md` seems to have the wrong block
CLAUDE.md already contains a live standalone import that resolves to your root AGENTS.md, entire-graph deliberately puts only an inheritance notice in its CLAUDE.md block, so the guide is not imported twice:AGENTS.md import and rerun init-agents to switch between the two layouts. Import detection only recognizes standalone paths. A mention inside inline code, a code fence, or an HTML comment does not count, and anything ambiguous keeps the direct pointer instead of assuming inheritance.Queries Are Slow
Every query rebuilds the index
Every query rebuilds the index
go.mod:go.mod and package.json all count. Dirty files entire-graph cannot index, such as a .bin, do not.The most common version of this is forgetting to commit the three activation files after running init-agents.You ran `index` but queries still miss the cache
You ran `index` but queries still miss the cache
index only warms committed-tree queries, and the interactive commands read the working tree unless you pass --head. index also defaults to --profile full while plain search defaults to --profile fast, and the profile is part of the cache key.So a default index run does not warm a default search. Match both:.graphignore also selects a different cache entry.`def` and `explain` never seem to cache
`def` and `explain` never seem to cache
def and explain only cache when --cache-dir or ENTIRE_PLUGIN_DATA_DIR is set. Set one of them if you are calling those two repeatedly.Results Are Not What You Expected
A symbol name is ambiguous
A symbol name is ambiguous
--symbol FILE:LINE.No symbols matched
No symbols matched
search, which takes plain language rather than an exact symbol.`impact` reports fewer callers than really exist
`impact` reports fewer callers than really exist
resolution and confidence fields rather than a completeness promise.`search` says LOW CONFIDENCE
`search` says LOW CONFIDENCE
A language returns files but no relationships
A language returns files but no relationships
Files are missing, or coverage says degraded
Files are missing, or coverage says degraded
partial_failures and warnings fields in JSON output, or the coverage line in agent format.Vendored and generated sources that are tracked in Git are the usual cause, since they show up as E_FILE_TOO_LARGE or E_PARSE_ERROR. Exclude them with a .graphignore at the repository root, which uses gitignore syntax and applies on top of .gitignore.Checking the Environment
entire graph doctor reports the resolved repository, the Entire environment variables, whether the plugin data directory is writable, and confirms no network egress: