Commits

List commit history with landing attribution, filter it by file or directory, read per-commit diffs, and inspect workspace and repository graphs from the CLI or over HTTP.

Repository history is queryable without a clone. Two surfaces cover it:

• CLI: tl git log and tl git smartlog for humans and agents working at a mount or a terminal.
• HTTP API: commit listing, per-commit diffs, workspace logs, and repository graphs for control planes and dashboards. This is the same API the Tensorlake dashboard uses.

Commits that landed through Tensorlake carry landing attribution: how the commit reached the branch (push, promote, merge) and which actor did it, classified as a human or an agent.

In a clone, git log, git diff, and git show work as usual; Tensorlake is a normal remote. This page is for when there is no checkout: mounts, control planes, CI, and dashboards.

History from the CLI

Workspace history: tl git log

tl git log shows the active snapshot chain of a workspace, plus any retained recovery chains created by rebases:

$ tl git log /work
workspace 3f9a2b7e1c4d
  8b21f6a9 add notes
  5d0c7e2a resolve parser conflict
  4f8c2a17 initial app

• With a mount path, it logs that mount’s workspace.
• With a repository name, it logs every workspace in the repository.
• With no argument, it uses the mount containing the current directory.

A read-only view has no workspace and no snapshot chain; use tl git smartlog there. Add --json for the raw pages.

Repository graph: tl git smartlog

tl git smartlog shows branch, tag, workspace, snapshot, and mount positions in one view — where everything is, and how far each workspace is ahead of its base:

$ tl git smartlog agent-outputs
repository project_9f3c2a1b/agent-outputs
main             1c4d9a2f  2h ago
├─ 3f9a2b7e1c4d  8b21f6a9  add notes                      ahead 2
└─ 7c2e9b4d1a6f  (no snapshots)                           at base

Unlike tl git log, smartlog works on read-only views: it reads repository state, not a workspace. Add --project for the bounded project-wide graph across every repository, and --json for the raw node/edge pages.

List Commits over HTTP

GET /project/{project}/repos/{repo}/commits

The endpoint walks a branch’s first-parent history and returns one page of commits, each with landing attribution when available. Authenticate with a short-lived Git credential over HTTP Basic auth (a read-scoped credential is enough):

TOKEN=$(tl git token agent-outputs --json | jq -r .token)

curl -u t:$TOKEN \
  "https://git.tensorlake.ai/project/project_9f3c2a1b/repos/agent-outputs/commits?ref=main&limit=2"

{
  "repo": "project_9f3c2a1b/agent-outputs",
  "ref_name": "refs/heads/main",
  "items": [
    {
      "oid": "1c4d9a2f7e5b...",
      "subject": "add notes",
      "author_name": "Agent User",
      "author_email": "[e-mail]",
      "author_at_secs": 1756000000,
      "committed_at_secs": 1756000000,
      "parents": ["4f8c2a17b9d3..."],
      "landing": {
        "via": "promote",
        "actor": { "name": "api-key:key_7d2f9a1c", "kind": "agent" },
        "source_workspace_id": "3f9a2b7e1c4d",
        "source_snapshot": "8b21f6a9c3d5..."
      }
    },
    {
      "oid": "4f8c2a17b9d3...",
      "subject": "initial app",
      "author_name": "Agent User",
      "author_email": "[e-mail]",
      "author_at_secs": 1755990000,
      "committed_at_secs": 1755990000,
      "parents": [],
      "landing": {
        "via": "push",
        "actor": { "name": "user:diptanu", "kind": "human" }
      }
    }
  ],
  "truncated": false
}

Query Parameters

Parameter	Meaning
ref	Branch name, tag name, full refs/... name, or a full 40-hex commit. Defaults to the default branch.
via	Only commits that landed through this operation: push, promote, or merge.
actor	Only commits landed by this actor, matched against landing.actor.name exactly.
path	Only commits that changed this file, or anything under this directory.
after	Opaque resume cursor from a previous page’s next_after.
limit	Page size, clamped to 1–200. Default 50.

Pagination is cursor-based: while truncated is true, pass next_after back as after. Filters apply server-side, and a filter that matches nothing still terminates promptly with a resumable cursor instead of walking unbounded history. An unknown ref returns 404.

Landing Attribution

The landing object records how a commit reached the branch:

Field	Meaning
via	The operation that landed it: push, promote, or merge.
actor.name	The authenticated principal that ran the operation.
actor.kind	human (a user or personal access token), agent (a project API key), or unknown.
source_workspace_id	For promotions: the workspace the commit came from.
source_snapshot	For promotions: the workspace snapshot that was published.

Attribution comes from the repository’s bounded operation log, so one landing entry attributes an entire pushed range — a large push costs one entry, not one per commit. Commits older than the attribution window come back without landing rather than erroring.

Filtering makes activity questions one call. Every commit an agent promoted to main:

curl -u t:$TOKEN \
  "https://git.tensorlake.ai/project/project_9f3c2a1b/repos/agent-outputs/commits?ref=main&via=promote"

File and Directory History

path= narrows the walk to commits that actually changed one file:

curl -u t:$TOKEN \
  "https://git.tensorlake.ai/project/project_9f3c2a1b/repos/agent-outputs/commits?ref=main&path=src/app.rs"

A directory path returns subtree history — every commit that changed anything under it:

curl -u t:$TOKEN \
  "https://git.tensorlake.ai/project/project_9f3c2a1b/repos/agent-outputs/commits?ref=main&path=src"

The response shape is unchanged: the same items, landing attribution, and after/limit cursor scheme, restricted to commits that changed the path. A deletion counts as a change to the path, and a rename counts as a change to both the old and the new path. A commit that only changed other files is not returned. path can be combined with via and actor.

Two implementation details affect how to use this:

• A commit is reported only after its tree was compared against its first parent at the given path. The changed-path index makes the walk faster, but it does not by itself decide whether a commit is included.
• A path filter over deep history scans at most 10,000 commits per request. If nothing matched in that window, the page is empty with truncated: true and a cursor; passing the cursor back continues the walk. Queries and pushes index older history as they run, so repeated queries over the same history get faster.

For line-level attribution of one file’s current content, see Blame.

Commit Diff over HTTP

GET /project/{project}/repos/{repo}/commits/{hash}/diff

Returns one page of a commit’s line-level diff against its first parent (a root commit diffs against the empty tree), paged by file path with the same after/limit cursor scheme:

curl -u t:$TOKEN \
  "https://git.tensorlake.ai/project/project_9f3c2a1b/repos/agent-outputs/commits/1c4d9a2f7e5b.../diff"

{
  "repo": "project_9f3c2a1b/agent-outputs",
  "commit": "1c4d9a2f7e5b...",
  "parent": "4f8c2a17b9d3...",
  "files": [
    {
      "path": "README.md",
      "change": "modified",
      "mode": 33188,
      "old_oid": "a94c8fe5...",
      "new_oid": "d3b07384...",
      "old_size": 16,
      "new_size": 31,
      "binary": false,
      "added_lines": 2,
      "removed_lines": 0,
      "patch_omitted": false,
      "hunks": [
        {
          "old_start": 1,
          "old_lines": 1,
          "new_start": 1,
          "new_lines": 3,
          "lines": [
            { "op": " ", "text": "# Agent Outputs" },
            { "op": "+", "text": "" },
            { "op": "+", "text": "## Parser Notes" }
          ]
        }
      ]
    }
  ],
  "truncated": false
}

• change is added, modified, or removed.
• Hunk line op is "+", "-", or " ", like a unified diff.
• Binary files are reported honestly: binary: true with sizes and oids, and no fabricated line hunks.
• A text file over the patch size caps comes back with patch_omitted: true and its line counts, so listings stay bounded on any commit.

Workspace Log and Smartlog over HTTP

The same data behind the CLI views is available directly.

Workspace log — the active snapshot chain and retained recovery chains of one workspace:

GET /project/{project}/repos/{repo}/workspaces/{id}/log

Each active-chain entry carries oid, subject, at_ms, actor, operation, and conflicted. Retained chains (created when a rebase replaces a snapshot chain) carry the recovery ref, head, base, and retention reason — nothing is silently discarded by a rebase.

Smartlog — a bounded node/edge graph of branches, tags, workspaces, snapshots, and live mounts:

GET /project/{project}/repos/{repo}/smartlog
GET /project/{project}/smartlog

Nodes carry a kind (branch, tag, workspace, snapshot, mount), position (oid), actor, and timestamp; edges connect them. The project-wide form covers every repository in the project and supports repo= and workspace= filters — this is the fleet view for a control plane watching many agents at once. Both endpoints page with after/limit.

Operation History

Separately from commit history, every repository keeps a durable operation log: pushes, promotions, merges, snapshots, branch deletions, and lifecycle events, each attributed to its actor. It is available from the SDKs as operations(repo) — see Repository SDKs.

Next Steps

Blame

Attribute every line of a file to the commit that introduced it.

Repository Mounts

The snapshot and promote workflow that produces attributed history.

Merging Changes

Land workspaces on moved branches and query structured conflict records.

Repository SDKs

Repository operations and the operation log from Python or TypeScript.

Authentication

Short-lived Git credentials, scopes, and token lifetime.