Part of the Ataraxy Labs stack, agent-native infrastructure for software development. See also: sem (semantic version control) · inspect (semantic code review) · opensessions (tmux sidebar for coding agents).

Read the manifesto: https://ataraxy-labs.com/#thesis · Essays: https://ataraxy-labs.com/blogs · LLMs: https://ataraxy-labs.com/llms.txt

Entity-level semantic merge for Git.
Resolves merge conflicts that Git can't by parsing code into functions, classes, and keys with tree-sitter, then merging those entities instead of lines.

Install · Quickstart · How It Works · MCP Server · CLI · Releases

Quickstart

weave setup                 # this repo now merges through weave; git merge/rebase/cherry-pick unchanged
git merge <branch>          # real conflicts land as markers with a `refused_by:` line stating why
weave explain <file>        # per-hunk detail for one conflicted file, read off the actual git stages
#  ...edit to resolve...
weave check                 # verify the working tree against the three merge stages; exits 1 on findings

See Setup for --global/--local variants, CLI Commands for the rest of the weave binary, and MCP Server for agent-framework integration.

The Problem

Git merges by comparing lines. When two branches both add code to the same file, even to completely different functions, Git sees overlapping line ranges and declares a conflict:

<<<<<<< HEAD
export function validateToken(token: string): boolean {
    return token.length > 0 && token.startsWith("sk-");
}
=======
export function formatDate(date: Date): string {
    return date.toISOString().split('T')[0];
}
>>>>>>> feature-branch

These are completely independent changes. There's no real conflict. But someone has to manually resolve it anyway.

This happens constantly when multiple AI agents work on the same codebase. Agent A adds a function, Agent B adds a different function to the same file, and Git halts everything for a human to intervene.

How Weave Fixes This

Weave replaces Git's line-based merge with entity-level merge, a 3-way merge that compares base, ours, and theirs at the level of individual functions, classes, and keys instead of individual lines. That lets it tell where the two branches actually drifted apart, rather than just where their edits happen to land on the same line numbers. It works like this:

1. Parses all three versions (base, ours, theirs) into semantic entities: functions, classes, JSON keys, etc., using tree-sitter
2. Matches entities across versions by identity (name + type + scope), including renames
3. Merges at the entity level:
   • Different entities changed → auto-resolved, no conflict
   • Same entity changed by both → attempts intra-entity merge, conflicts only if truly incompatible
   • One side modifies, other deletes → flags a meaningful conflict

Run the same scenario above through weave, and it merges cleanly with zero conflicts: both functions end up in the output.

This merge algorithm is deterministic and stateless: it reads three file revisions and writes one result, the same way git merge-file does. It is not a CRDT. (Weave separately ships a CRDT-backed coordination layer, weave-crdt, for tracking live multi-agent edits before they hit Git; see Architecture.)

Weave vs Git Merge

Scenario	Git (line-based)	Weave (entity-level)
Two agents add different functions to same file	CONFLICT	Auto-resolved
Agent A modifies foo(), Agent B adds bar()	CONFLICT (adjacent lines)	Auto-resolved
Both agents modify the same function differently	CONFLICT	CONFLICT (with entity-level context)
One agent modifies, other deletes same function	CONFLICT (cryptic diff)	CONFLICT: function 'validateToken' (modified in ours, deleted in theirs)
Both agents add identical function	CONFLICT	Auto-resolved (identical content detected)
Both agents add different properties to same object	CONFLICT	Auto-resolved
Different JSON keys modified	CONFLICT	Auto-resolved

The key difference: Git produces false conflicts on independent changes because they happen to be in the same file. Weave only conflicts on actual semantic collisions when two branches change the same entity incompatibly.

Weave vs Mergiraf

31 hand-crafted merge scenarios across 7 languages, comparable to mergiraf's own test corpus. Run weave bench to reproduce:

Two of the 31 must not merge. When both sides add different decorators to the same Python or TypeScript function, decorator application is function composition, so the stack order is a semantic decision neither side made — @cache outside @auth serves cached responses without ever running the auth check. Weave refuses rather than fabricate an order, and a tool that merges those cleanly is wrong, not better. Annotations in Java, C# and Kotlin are unordered metadata, so weave still set-unions those.

Tool	Clean merges (of 29 mergeable)	Correct outcomes (of 31)
Weave	29/29 (100%)	31/31 (100%)
Mergiraf (v0.16.3)	26/29 (90%)	28/31 (90%)
Git	15/29 (52%)	17/31 (55%)

All three tools correctly refuse the two decorator scenarios. Mergiraf fails on both-add-at-end-of-file and insert-between-existing; weave resolves those because it operates at entity granularity (functions, classes, methods) rather than AST node level. Full breakdown at ataraxy-labs.github.io/weave.

Real-World Benchmarks

Replayed against real merge commits from five long-lived open-source repositories. For each of the first 500 merge commits per repo, weave re-runs the merge (base/ours/theirs from the actual git history) and compares its output to both Git's line merge and the human-authored merge commit. Reproduce with weave bench-repo <path-to-clone> --limit 500; full per-repo breakdown, including which files disagree and why, is at ataraxy-labs.github.io/weave/benchmarks.html.

• Win: the line-based 3-way merge conflicted, weave resolved cleanly
• Regression: the line-based 3-way merge resolved cleanly, weave conflicted
• Human match: of weave's wins, how many are byte-identical (whitespace-normalized) to what the developer actually wrote

Note (0.5.3): regenerated on the 0.5.3 engine on 2026-09-01 (weave bench-repo <clone> --limit 500, fresh full clones). Read against the previous table with three caveats, stated rather than smoothed over. First, 0.5.3 conflicts on purpose where 0.5.2 sometimes resolved silently (divergent same-name additions, tightened same-entity and gap verdicts) — most of the regression increase is that tightening doing its job; on CPython, whose tested window is identical between runs, all of it is. Second, the earlier run's exact commit window was not recorded, and bench-repo walks the most recent N merges — for git, Go, and TypeScript the two runs replay substantially different commit sets, so cross-run rate comparisons there are indicative, not exact; this run's windows are current as of the date above. Third, audited details: the "clean line merge" baseline is a diff3 implementation (diffy), which disagrees with git merge-file on a small number of cases; 3 of the 86 regressions are a known guard false-positive (files whose source code contains conflict-marker string literals — e.g. TypeScript's own scanner); and in 2 regressions the line merge's "clean" result differs from what the human actually committed, i.e. weave's refusal was arguably the safer verdict.

Repository	Language	File merges tested	Wins	Regressions	Human match
git/git	C	1,701	183	23	72%
Flask	Python	67	15	1	33%
CPython	C / Python	256	11	10	45%
Go	Go	1,667	120	37	33%
TypeScript	TypeScript	1,280	15	15	53%

Across all five repos: 344 wins on 4,971 file merges (0.5.2 measured 83 wins on 4,517), with 86 total regressions spread across every repo (0.5.2 measured 3, all on TypeScript). The jump in regressions is expected and by design, not a quality drop we're hiding: 0.5.3 now conflicts on genuinely divergent concurrent additions that 0.5.2 silently merged, and no longer resolves a case that could resurrect a deleted JSON key — both changes move cases from "weave resolves" into "weave conflicts" under this benchmark's own regression definition (git resolves cleanly, weave doesn't). Wins also rose substantially on every repo. File-merge counts and human-match rates shifted too, partly because "first 500 merge commits" is a moving window and all five repos have advanced since the 0.5.2 run. See the per-repo breakdown on the benchmarks page before relying on weave for large merges in any of these languages.

Testing

The open test suite in this repository, 441 unit and integration tests plus a five-scenario sweep per supported language in crates/weave-core/tests/language_coverage.rs, covers the documented merge properties and runs in CI (cargo fmt --check, cargo clippy -D warnings, cargo test --workspace) on Linux and Windows on every push and PR.

Conflict Markers

When a real conflict occurs, weave gives you context that Git doesn't: which entity, what type, and, on the line inside the box, which internal guard declined to auto-merge and exactly which lines both sides disagree about.

<<<<<<< ours — function `process` (T, confidence: high)
// refused_by: statement_fold · collision: `    return data.upper()` +1 more
export function process(data: any) {
    return JSON.stringify(data);
}
=======
export function process(data: any) {
    return data.toUpperCase();
}
>>>>>>> theirs — function `process` (T, confidence: high)

Run weave explain <file> for more detail on every conflicted entity in the file, and weave check after editing to verify your resolution against the three merge stages; see Quickstart.

Supported Languages

TypeScript, TSX, JavaScript, Python, Go, Rust, Java, C, C++, Ruby, C#, PHP, Swift, Kotlin, Scala, Dart, Elixir, Bash, Fish, Fortran, Perl, OCaml, Zig, Elm, Clojure, EDN, D, Lua, Nix, SQL, HCL/Terraform, LaTeX, XML, JSON, YAML, TOML, CSV, Markdown. Falls back to standard line-level merge for everything else.

weave setup derives its .gitattributes rules directly from the parser registry: every extension the tree-sitter grammars recognize gets a merge=weave line automatically, with no hand-maintained list to fall behind. Each language on it passes a five-scenario merge sweep (two sides adding different definitions merges clean, two sides rewriting the same definition conflicts, and nothing is dropped) in crates/weave-core/tests/language_coverage.rs, and a separate parity test (crates/weave-core/tests/setup_extension_coverage.rs) fails the build if a newly added grammar is ever left unclaimed and undeclined.

Vue, Svelte, ERB and Haskell are parsed but deliberately not claimed: weave declines exactly these four (weave_core::DECLINED_EXTENSIONS), and nothing else. Their entity model treats a whole <script> block, template, or type signature as a single unit, so two people adding two different definitions conflict where they should merge cleanly, and the conflict marker can land mid-definition. Those files get Git's line-level merge instead, which is the better answer until the parser gains a real per-definition model for them. They're excluded from weave setup automatically; nothing to opt out of by hand.

Install

brew install weave

Or build from source (requires Rust). Two binaries, both required: weave (the CLI you run: setup/explain/check/...) and weave-driver (the one git itself invokes on every merge; weave setup fails without it on PATH):

git clone https://github.com/Ataraxy-Labs/weave
cd weave
cargo install --path crates/weave-cli      # the `weave` binary
cargo install --path crates/weave-driver   # the `weave-driver` binary git calls

Upgrading an existing source install? cargo install refuses to overwrite a binary it didn't put there itself, so add --force to either command above.

Setup

In any Git repo:

weave setup

This configures Git to use weave for all supported file types. Then use git merge as normal.

To revert back to normal git merging:

weave unsetup

To set up for just yourself (without modifying .gitattributes), write the same supported file type rules to .git/info/attributes instead:

weave setup --local

Global (every repo)

To make weave the default merge driver for all your repos at once (no per-repo setup, like mergiraf):

weave setup --global

This writes the driver to your ~/.gitconfig and the supported file-type rules to git's global attributes file (~/.config/git/attributes, or your core.attributesfile if set). No git repo required. Make sure weave-driver is on your PATH (it ships next to the weave binary).

The equivalent manual config, if you prefer:

git config --global merge.weave.name "Entity-level semantic merge"
git config --global merge.weave.driver "weave-driver %O %A %B %L %P"
# then add `*.ts merge=weave` (etc.) to ~/.config/git/attributes

Jujutsu (jj)

Add to your jj config (jj config edit --user):

[merge-tools.weave]
program = "weave-driver"
merge-args = ["$base", "$left", "$right", "-o", "$output", "-l", "$marker_length", "-p", "$path"]
merge-conflict-exit-codes = [1]
merge-tool-edits-conflict-markers = true
conflict-marker-style = "git"

Resolve conflicts with jj resolve --tool weave, or set as default:

jj config set --user ui.merge-editor "weave"

Preview

Dry-run a merge to see what weave would do:

weave preview feature-branch

  src/utils.ts — auto-resolved
    unchanged: 2, added-ours: 1, added-theirs: 1
  src/api.ts — 1 conflict(s)
    ✗ function `process`: both modified

✓ Merge would be clean (1 file(s) auto-resolved by weave)

After a real conflict, weave explain <file> and weave check are the next two commands; see Quickstart.

CLI Commands

Beyond setup/explain/check/preview above, the weave binary has commands for the CRDT coordination layer and for typed entity patches. Run weave --help or weave <command> --help for the full flag list; the table below is what each one is for.

Command	What it does
weave status [--file] [--agent]	Entity and agent state from the CRDT: claims, last editor, merge state
weave claim <agent-id> <file> <entity>	Claim an entity before editing it (advisory: weave does not enforce it)
weave release <agent-id> <file> <entity>	Release a previously claimed entity
weave apply <file>...	Materialize entity edits held in the CRDT back onto the working files
weave patch extract <base-file> <changed-file>	Emit the typed ops that turn base-file into changed-file
weave patch apply <ops-file> <target-file>	Apply those ops to a target file, three-way against the ops' base, in case the target has drifted since the ops were extracted
weave summary <file>	Parse a file's weave conflict markers into a structured (optionally JSON) summary
weave stats	Lifetime merge counters, if you've opted in with WEAVE_STATS=1 (off by default)
weave bench	Run the 31-scenario synthetic benchmark against weave, Mergiraf, and git
weave bench-repo <path> [--limit N]	Replay real merge commits from a cloned repo; see Real-World Benchmarks

claim/release/status/apply all operate on the same .weave/state.automerge CRDT document as the MCP tools below: the CLI and MCP server are two front ends onto one coordination state. That document lives in the repo's working tree but is never repo content: the first time weave writes it, it adds .weave/ to the repo's local .git/info/exclude (never your own .gitignore), so it never shows up in git status or gets swept into git add -A.

MCP Server

For agent frameworks that speak MCP:

# Claude Code
claude mcp add --scope user weave -- weave-mcp

# Any MCP client, via stdio (~/.config/claude/claude_desktop_config.json etc.)
{ "mcpServers": { "weave": { "command": "weave-mcp" } } }

The server discovers the repo from the first tool call's file path, the WEAVE_REPO env var, or its working directory. It exposes 22 tools in two independent groups (each tool's own description states when to call it and what an empty result means):

• Merge analysis reads git refs or the working tree directly, no setup needed: weave_findings, weave_check, weave_preview_merge, weave_diff, weave_merge_audit, weave_validate_merge, weave_merge_summary.
• Entity and dependency inspection reads a file's or the repo's structure: weave_extract_entities, weave_get_dependencies, weave_get_dependents, weave_impact_analysis.
• Live coordination tracks edits in the shared CRDT (.weave/state.automerge) for agents editing the same repo at the same time, starting with weave_agent_register: weave_agent_register, weave_agent_heartbeat, weave_claim_entity, weave_release_entity, weave_status, weave_who_is_editing, weave_potential_conflicts, weave_update_entity_content, weave_get_entity_content, weave_merge_file, weave_resolve_conflict.

Start with weave_findings after (or before) a merge between two branches, or weave_check for the cross-file binding risk a per-file git merge driver can't see: a rename in a.py whose surviving caller lives in b.py merges both files cleanly on its own, and the break is only visible repo-wide.

Architecture

weave-core       # Library: entity extraction, entity-level 3-way merge, reconstruction
weave-driver     # Git merge driver binary (called by git via %O %A %B %L %P)
weave-cli        # CLI: `weave setup`, `weave explain`, `weave check`, `weave patch`, ...
weave-crdt       # Automerge-backed CRDT: live multi-agent coordination state only
weave-mcp        # MCP server exposing weave to agent frameworks (22 tools)
weave-github     # GitHub webhook service behind the hosted PR-comment integration
                 #   (publish = false, not a binary you install; runs weave's merge
                 #   analysis on pull_request events and posts the result as a comment)

Uses sem-core for entity extraction via tree-sitter grammars.

The merge algorithm (weave-core) and the coordination state (weave-crdt) are deliberately separate concerns with different data models: the merge is a pure function over three file revisions, run fresh on every git merge/weave preview/ weave check call. The CRDT is the thing that persists; it's what lets two live agents see each other's claims and in-flight edits before either one commits, via weave_claim_entity/weave_update_entity_content or weave claim/weave apply. Nothing in the merge path depends on the CRDT ever having run.

How It Works

         base
        /    \
     ours    theirs
        \    /
       weave merge

1. Parse all three versions into semantic entities via tree-sitter
2. Extract regions, alternating entity and interstitial (imports, whitespace) segments
3. Match entities across versions by ID (file:type:name:parent), detecting renames
4. Resolve each entity: one-side-only changes win, both-changed attempts intra-entity 3-way merge
5. Reconstruct file from merged regions, preserving ours-side ordering
6. Fallback to line-level merge for files >1MB, binary files, or unsupported types

Limitations

• Vue, Svelte, ERB, Haskell parse, but their per-file entity model is too coarse to merge well, so they take Git's line merge, always (see Supported Languages).
• Files over 1MB, binary files, and file types with no parser fall back to Git's line-level merge automatically.
• Entity claims are advisory, not enforced. weave_claim_entity and weave claim are cooperative locks inside the CRDT coordination layer; weave does not stop a second agent (or you) from editing a claimed entity anyway.
• Crashed agents aren't reaped. weave_agent_heartbeat's liveness timestamp is informational; weave does not currently expire or release a claim automatically when an agent stops heartbeating, so a crashed agent's claims stay visible until another call to weave_release_entity/weave release.
• weave stats is empty until you opt in. Lifetime merge counters are off by default; set WEAVE_STATS=1 in the environment your git/jj merges run in to start accumulating them.

Contributing

Bug reports and issues are welcome. This is a small team maintaining a merge engine that runs inside other people's git workflows: incoming PRs get read, but are reviewed and adapted before merge rather than merged as-is. Open an issue first for anything beyond a small, obviously-correct fix, so the approach can be agreed on before you write the code.

License

Dual-licensed under MIT or Apache 2.0, at your option.

Star History