Implement the backend for the desktop feature for agents.
- Adds a new `/api/experimental/chats/$id/desktop` endpoint to coderd
which exposes a VNC stream from a
[portabledesktop](https://github.com/coder/portabledesktop) process
running inside the workspace
- Adds a new `spawn_computer_use_agent` tool to chatd, which spawns a
subagent that has access to the `computer` tool which lets it interact
with the `portabledesktop` process running inside the workspace
- Adds the plumbing to make the above possible
There's a follow up frontend PR here:
https://github.com/coder/coder/pull/23006
Adds real-time git status watching for workspace agents, so the frontend
can subscribe over WebSocket and show
git file changes in near real-time.
1. Subscription is scoped to a **chat** via `GET
/api/experimental/chats/{chat}/git/watch`.
2. The workspace agent automatically determines which paths to watch
based on tool calls made by the chat (and its ancestor chats).
3. Workspace agent polls subscribed repo working trees on a 30s
interval, on tools calls, and on explicit `refresh` from the client.
4. Scans are rate-limited to at most once per second.
5. Edited paths are tracked **in-memory** inside the workspace agent.
There is no database persistence — state is lost on agent restart. This
will be addresses in a future PR.
6. Messages sent over WebSocket include a full-repo snapshot (unified
diff, branch, origin). A new message is emitted only when the snapshot
changes.
This PR was implemented with AI with me closely controlling what it's
doing. The code follows a plan file that was updated continuously during
implementation. Here's the file if you'd like to see it:
[project.md](https://gist.github.com/hugodutka/8722cf80c92f8a56555f7bc595b770e2).
It reflects the current state of the PR.
## Summary
Adds a new agent-side process management HTTP API and rewrites the chat
execute tool to use it instead of SSH sessions.
## What changed
### New agent/agentproc/ package
- **headtail.go** — Thread-safe io.Writer with bounded memory (16KB head
+ 16KB tail ring buffer). Provides LLM-ready output with truncation
metadata and long-line truncation at 2048 bytes.
- **headtail_test.go** — 16 tests including race detector coverage for
concurrent writes.
- **process.go** — Manager + Process types for lifecycle management
using agentexec.Execer for proper OOM/nice scores.
- **api.go** — HTTP API following the agentfiles chi router pattern. 4
endpoints: start, list, output, signal.
### Agent wiring (agent/agent.go, agent/api.go)
Mounts the process API at /api/v0/processes, mirroring how agentfiles is
mounted.
### SDK (codersdk/workspacesdk/agentconn.go)
4 new AgentConn interface methods + 7 request/response types:
- StartProcess, ListProcesses, ProcessOutput, SignalProcess
### Execute tool rewrite (coderd/chatd/chattool/execute.go)
- SSH to Agent API: conn.StartProcess() + conn.ProcessOutput() polling
- New parameters: workdir, run_in_background
- Structured response: success, exit_code, wall_duration_ms, error,
truncated, note, background_process_id
- Non-interactive env vars: GIT_EDITOR=true, TERM=dumb, NO_COLOR=1,
PAGER=cat, etc.
- Output truncation: HeadTailBuffer caps at 32KB for LLM consumption
- File-dump detection with advisory notes suggesting read_file
- Default timeout: 60s to 10s
- Foreground polling: 200ms intervals until exit or timeout
## Architecture
State lives on the agent, surviving coderd failover and instance
changes. Any coderd replica can query any agent via HTTP over tailnet.
## Summary
Adds a new line-based file reading endpoint to the workspace agent,
replacing the unbounded byte-based approach for the `read_file` chat
tool and `coder_workspace_read_file` MCP tool.
**Problem**: The current `read_file` tool returns the entire file
contents with no limits, which can blow up LLM context windows and cause
OOM issues with large files.
**Solution**: Inspired by [`coder/mux`](https://github.com/coder/mux)
and [`openai/codex`](https://github.com/openai/codex), implement a
line-based reader with safety limits.
## Changes
### Agent (`agent/agentfiles/`)
- New `/read-file-lines` endpoint with `HandleReadFileLines` handler
- Line-based `offset` (1-based line number, default: 1) and `limit`
(line count, default: 2000)
- Safety constants:
| Constant | Value | Purpose |
|---|---|---|
| `MaxFileSize` | 1 MB | Reject files larger than this at stat |
| `MaxLineBytes` | 1,024 | Per-line truncation with `... [truncated]`
marker |
| `MaxResponseLines` | 2,000 | Max lines per response |
| `MaxResponseBytes` | 32 KB | Max total response size |
| `DefaultLineLimit` | 2,000 | Default when no limit specified |
- Line numbering format: `1\tcontent` (tab-separated)
- Structured JSON response: `{ success, file_size, total_lines,
lines_read, content, error }`
- Hard errors when limits exceeded — tells the LLM to use
`offset`/`limit`
- Existing byte-based `/read-file` endpoint preserved (used by
`instruction.go`)
### SDK (`codersdk/workspacesdk/`)
- `ReadFileLinesResponse` type added
- `ReadFileLines` method added to `AgentConn` interface
- Mock regenerated
### Chat tool (`coderd/chatd/chattool/`)
- `read_file` tool now uses `conn.ReadFileLines()` instead of
`conn.ReadFile()`
- Updated tool description to document line-based parameters
- Response includes `file_size`, `total_lines`, `lines_read` metadata
### MCP tool (`codersdk/toolsdk/`)
- `coder_workspace_read_file` updated to use line-based reading
- Schema descriptions updated for line-based offset/limit
- Removed `maxFileLimit` constant (agent handles limits now)
### Tests
- 13 new test cases for `TestReadFileLines`:
- Path validation (empty, relative, non-existent, directory, no
permissions)
- Empty file handling
- Basic read, offset, limit, offset+limit combinations
- Offset beyond file length
- Long line truncation (>1024 bytes)
- Large file rejection (>1MB)
- All existing tests pass unchanged
## Design decisions
| Decision | Rationale |
|---|---|
| Line-based, not byte-based | Both coder/mux and openai/codex use
line-based — matches how LLMs reason about code |
| Default limit of 2000 | Matches codex; prevents accidental full-file
dumps while being generous |
| 32 KB response cap | Compromise between mux (16 KB) and codex (no cap)
|
| 1024 byte/line truncation with marker | More generous than codex
(500), marker helps LLM know data is missing |
| Hard errors on overflow | Matches mux; forces LLM to paginate rather
than getting partial data |
| Preserve byte-based endpoint | `instruction.go` needs raw byte access
for AGENTS.md |
Fixes all our Go file imports to match the preferred spec that we've _mostly_ been using. For example:
```
import (
"context"
"time"
"github.com/prometheus/client_golang/prometheus"
"golang.org/x/xerrors"
"gopkg.in/natefinch/lumberjack.v2"
"cdr.dev/slog/v3"
"github.com/coder/coder/v2/codersdk/agentsdk"
"github.com/coder/serpent"
)
```
3 groups: standard library, 3rd partly libs, Coder libs.
This PR makes the change across the codebase. The PR in the stack above modifies our formatting to maintain this state of affairs, and is a separate PR so it's possible to review that one in detail.
Upgrades to slog v3 which includes a small, but backward incompatible API change to the acceptible call arguments when logging. This change allows us to verify via compile time type checking that arguments are correct and won't cause a panic, as was possible in slog v1, which this replaces (v2 was tagged but never used in coder/coder).
It also updates dependencies that also use slog and were updated.
I've left the `aibridge` dependency as a commit SHA, under the assumption that the team there (cc @pawbana @dannykopping ) will tag and update the dependency soon and on their own schedule.
Other dependencies, I pushed new tags.
Follows similarly to the bash tool (and some code to connect to an agent
was extracted from it).
There are two main parts: a new agent endpoint, and then a new MCP tool
that consumes that endpoint.
Fixes https://github.com/coder/internal/issues/907
We convert `workspacesdk.AgentConn` to an interface and generate a mock
for it. This allows writing `coderd` tests that rely on the agent's HTTP
api to not have to set up an entire tailnet networking stack.
Fixes https://github.com/coder/coder/issues/19372
We increase the read limit to 4MiB (we use this limit elsewhere). We
also make sure to stop sending messages when `containersCh` becomes
closed.
This PR replaces the use of the **container** ID with the
**devcontainer** ID. This is a breaking change. This allows rebuilding a
devcontainer when there is no valid container ID.
This change introduces a refactor of the devcontainers recreation logic
which is now handled asynchronously rather than being request scoped.
The response was consequently changed from "No Content" to "Accepted" to
reflect this.
A new `Status` field was introduced to the devcontainer struct which
replaces `Running` (bool). This reflects that the devcontainer can now
be in various states (starting, running, stopped or errored).
The status field also protects against multiple concurrent recrations,
as long as they are initiated via the API.
Updates #16424
Closes https://github.com/coder/vscode-coder/issues/447
Closes https://github.com/coder/jetbrains-coder/issues/543
Closes https://github.com/coder/coder-jetbrains-toolbox/issues/21
This PR adds Coder Connect support to `coder ssh --stdio`.
When connecting to a workspace, if `--force-new-tunnel` is not passed, the CLI will first do a DNS lookup for `<agent>.<workspace>.<owner>.<hostname-suffix>`. If an IP address is returned, and it's within the Coder service prefix, the CLI will not create a new tailnet connection to the workspace, and instead dial the SSH server running on port 22 on the workspace directly over TCP.
This allows IDE extensions to use the Coder Connect tunnel, without requiring any modifications to the extensions themselves.
Additionally, `using_coder_connect` is added to the `sshNetworkStats` file, which the VS Code extension (and maybe Jetbrains?) will be able to read, and indicate to the user that they are using Coder Connect.
One advantage of this approach is that running `coder ssh --stdio` on an offline workspace with Coder Connect enabled will have the CLI wait for the workspace to build, the agent to connect (and optionally, for the startup scripts to finish), before finally connecting using the Coder Connect tunnel.
As a result, `coder ssh --stdio` has the overhead of looking up the workspace and agent, and checking if they are running. On my device, this meant `coder ssh --stdio <workspace>` was approximately a second slower than just connecting to the workspace directly using `ssh <workspace>.coder` (I would assume anyone serious about their Coder Connect usage would know to just do the latter anyway).
To ensure this doesn't come at a significant performance cost, I've also benchmarked this PR.
<details>
<summary>Benchmark</summary>
## Methodology
All tests were completed on `dev.coder.com`, where a Linux workspace running in AWS `us-west1` was created.
The machine running Coder Desktop (the 'client') was a Windows VM running in the same AWS region and VPC as the workspace.
To test the performance of specifically the SSH connection, a port was forwarded between the client and workspace using:
```
ssh -p 22 -L7001:localhost:7001 <host>
```
where `host` was either an alias for an SSH ProxyCommand that called `coder ssh`, or a Coder Connect hostname.
For latency, [`tcping`](https://www.elifulkerson.com/projects/tcping.php) was used against the forwarded port:
```
tcping -n 100 localhost 7001
```
For throughput, [`iperf3`](https://iperf.fr/iperf-download.php) was used:
```
iperf3 -c localhost -p 7001
```
where an `iperf3` server was running on the workspace on port 7001.
## Test Cases
### Testcase 1: `coder ssh` `ProxyCommand` that bicopies from Coder Connect
This case tests the implementation in this PR, such that we can write a config like:
```
Host codercliconnect
ProxyCommand /path/to/coder ssh --stdio workspace
```
With Coder Connect enabled, `ssh -p 22 -L7001:localhost:7001 codercliconnect` will use the Coder Connect tunnel. The results were as follows:
**Throughput, 10 tests, back to back:**
- Average throughput across all tests: 788.20 Mbits/sec
- Minimum average throughput: 731 Mbits/sec
- Maximum average throughput: 871 Mbits/sec
- Standard Deviation: 38.88 Mbits/sec
**Latency, 100 RTTs:**
- Average: 0.369ms
- Minimum: 0.290ms
- Maximum: 0.473ms
### Testcase 2: `ssh` dialing Coder Connect directly without a `ProxyCommand`
This is what we assume to be the 'best' way to use Coder Connect
**Throughput, 10 tests, back to back:**
- Average throughput across all tests: 789.50 Mbits/sec
- Minimum average throughput: 708 Mbits/sec
- Maximum average throughput: 839 Mbits/sec
- Standard Deviation: 39.98 Mbits/sec
**Latency, 100 RTTs:**
- Average: 0.369ms
- Minimum: 0.267ms
- Maximum: 0.440ms
### Testcase 3: `coder ssh` `ProxyCommand` that creates its own Tailnet connection in-process
This is what normally happens when you run `coder ssh`:
**Throughput, 10 tests, back to back:**
- Average throughput across all tests: 610.20 Mbits/sec
- Minimum average throughput: 569 Mbits/sec
- Maximum average throughput: 664 Mbits/sec
- Standard Deviation: 27.29 Mbits/sec
**Latency, 100 RTTs:**
- Average: 0.335ms
- Minimum: 0.262ms
- Maximum: 0.452ms
## Analysis
Performing a two-tailed, unpaired t-test against the throughput of testcases 1 and 2, we find a P value of `0.9450`. This suggests the difference between the data sets is not statistically significant. In other words, there is a 94.5% chance that the difference between the data sets is due to chance.
## Conclusion
From the t-test, and by comparison to the status quo (regular `coder ssh`, which uses gvisor, and is noticeably slower), I think it's safe to say any impact on throughput or latency by the `ProxyCommand` performing a bicopy against Coder Connect is negligible. Users are very much unlikely to run into performance issues as a result of using Coder Connect via `coder ssh`, as implemented in this PR.
Less scientifically, I ran these same tests on my home network with my Sydney workspace, and both throughput and latency were consistent across testcases 1 and 2.
</details>
- Update go.mod to use Go 1.24.1
- Update GitHub Actions setup-go action to use Go 1.24.1
- Fix linting issues with golangci-lint by:
- Updating to golangci-lint v1.57.1 (more compatible with Go 1.24.1)
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <claude@anthropic.com>
Fixes: https://github.com/coder/internal/issues/377
Added an additional SSH listener on port 22, so the agent now listens on both, port one and port 22.
---
Change-Id: Ifd986b260f8ac317e37d65111cd4e0bd1dc38af8
Signed-off-by: Thomas Kosiewski <tk@coder.com>
Builds on top of https://github.com/coder/coder/pull/16623/ and wires up
the ReconnectingPTY server. This does nothing to wire up the web
terminal yet but the added test demonstrates the functionality working.
Other changes:
* Refactors and moves the `SystemEnvInfo` interface to the
`agent/usershell` package to address follow-up from
https://github.com/coder/coder/pull/16623#discussion_r1967580249
* Marks `usershellinfo.Get` as deprecated. Consumers should use the
`EnvInfoer` interface instead.
---------
Co-authored-by: Mathias Fredriksson <mafredri@gmail.com>
Co-authored-by: Danny Kopping <danny@coder.com>
Fixes https://github.com/coder/coder/issues/16268
- Adds `/api/v2/workspaceagents/:id/containers` coderd endpoint that allows listing containers
visible to the agent. Optional filtering by labels is supported.
- Adds go tools to the `coder-dylib` CI step so we can generate mocks if needed
re: #14715
This PR introduces the Coder service prefix: `fd60:627a:a42b::/48` and refactors our existing code as calling the Tailscale service prefix explicitly (rather than implicitly).
Removes the unused `Addresses` agent option. All clients today assume they can compute the Agent's IP address based on its UUID, so an agent started with a custom address would break things.
First PR to address #14244.
Adds common potential reasons as to why a direct connection to the workspace agent couldn't be established to `coder ping`:
- If the Coder deployment administrator has blocked direction connections (`CODER_BLOCK_DIRECT`).
- If the client has no STUN servers within it's DERP map.
- If the client or agent appears to be behind a hard NAT, as per Tailscale `netInfo.MappingVariesByDestIP`
Also adds a warning if the client or agent has a network interface below the 'safe' MTU for tailnet. This warning is always displayed at the end of a `coder ping`.
Currently, importing `codersdk` just to interact with the API requires
importing tailscale, which causes builds to fail unless manually using
our fork.
Hugo Dutka