harness_name stringclasses 56
values | github_repo stringclasses 59
values | issue_title stringlengths 1 728 | issue_body stringlengths 0 65k ⌀ | issue_url stringlengths 39 77 | issue_created_at int64 1,376B 1,768B | issue_closed_at int64 1,382B 1,776B ⌀ | is_related float64 0 1 ⌀ | stage float64 0 4 ⌀ | step stringclasses 3
values | strategy float64 1 6 ⌀ | root_cause_label stringclasses 11
values | issue_comments listlengths 0 76 | issue_labels listlengths 0 9 ⌀ | issue_cross_referenced listlengths 0 23 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
harbor | harbor-framework/harbor | Bug forwarded from pi-mono triage https://github.com/badlogic/pi-terminal-bench | I wanted to make sure the team saw the following to determine if it's still valid and really an issue vs. issue with the pi-mono agent:
https://github.com/badlogic/pi-terminal-bench
https://github.com/badlogic/pi-terminal-bench/blob/main/ERROR.md
Terminal-Bench Verifier Error Investigation
Problem Summary
When runnin... | https://github.com/harbor-framework/harbor/issues/335 | 1,766,939,050,000 | null | 1 | 2 | A | 3 | Algorithmic Error | [] | null | null |
harbor | harbor-framework/harbor | Syntax Error in Registry.json | I think there is an syntax typo in registry json file https://raw.githubusercontent.com/laude-institute/harbor/main/registry.json here.
When I ran harbor dataset list, I ran into this issue:
`JSONDecodeError: Expecting ',' delimiter: line 11377 column 5 (char 466521)` | https://github.com/harbor-framework/harbor/issues/329 | 1,766,898,592,000 | 1,767,156,805,000 | 1 | 0 | B | 2 | Validation Gap | [
"It would be great to have this file tested in CI before release, or at least locked to a release. Currently an error here takes out any evals using Harbor's default registry url even if the release hasn't changed.\n\nIn the meantime, here's a workaround using the last commit `harbor run -d terminal-bench@2.0 --reg... | null | null |
harbor | harbor-framework/harbor | E2B sandbox timeout is not configurable | ## Summary
The E2B environment has a hardcoded sandbox timeout of 86,400 seconds (24 hours), which exceeds E2B's maximum allowed timeout of 1 hour (3,600 seconds) for free/hobby tier accounts. This causes E2B trials to fail immediately with a `400` error for free tier users, with no way to override the value.
## Erro... | https://github.com/harbor-framework/harbor/issues/285 | 1,766,567,897,000 | null | 1 | 1 | A | 3 | Architectural Constraint | [
"Hi, I met the same requirement so very thank your pr.\n\nBut how about set timeout dynamically? Such as set timeout=T1 in setup and set timeout=T2 in execution. There are some variables that may helpful for set timeout:\n- `Trial._agent_timeout_sec`\n- `Trial._verifier_timeout_sec`\n- `Trial._agent_setup_timeout_s... | null | null |
harbor | harbor-framework/harbor | Adapter runs unable to find docker | I believe that the recently merged PR https://github.com/laude-institute/harbor/commit/a56ddaea324599a5e5e20c8aa67be65b065faf67 is causing adapter runs to fail because it cannot find the path to docker.
Example error from `aider_polygot`:
<img width="819" height="276" alt="Image" src="https://github.com/user-attachmen... | https://github.com/harbor-framework/harbor/issues/281 | 1,766,519,793,000 | 1,766,606,149,000 | 1 | 0 | A | 5 | Configuration Error | [
"Yes. just hit this and wasn't sure if it's a local thing. \n\nHere's my analysis: \n\n# Bug: Docker environment broken on macOS after PR #272\n\n## Summary\n\nThe Docker environment fails to find the `docker` executable on macOS after [PR #272](https://github.com/laude-institute/harbor/pull/272), resulting in:\n\n... | null | null |
harbor | harbor-framework/harbor | Bug: send_keys(block=True) hangs indefinitely with multi-line input containing apt commands | Hello harbor developers,
I raised a similar issue in [terminal-bench](https://github.com/laude-institute/terminal-bench/issues/1379) where the maintainers pointed me to harbor. I replicated the same setup in harbor and ran into the same issue. When we tried using harbor instead of terminal bench's harness with a Qwen ... | https://github.com/harbor-framework/harbor/issues/273 | 1,766,462,166,000 | null | 1 | 2 | A | 3 | Algorithmic Error | [] | null | null |
harbor | harbor-framework/harbor | Cannot run endpoints hosted on Azure OpenAI | Thanks for releasing this framework! I would like to be able to run Harbor with models hosted Azure OpenAI.
Command: `harbor run -d terminal-bench@2.0 -m "azure/gpt-5-mini" -a "mini-swe-agent" -t multi-source-data-merger`
Expected outcome: agent runs
Outcome:
```bash
$ harbor run -d terminal-bench@2.0 -m "azure/gpt... | https://github.com/harbor-framework/harbor/issues/270 | 1,766,434,988,000 | 1,766,457,545,000 | 1 | 1 | A | 2 | Validation Gap | [] | null | null |
harbor | harbor-framework/harbor | `--agent-import-path` is ignored when no `-a` agent is specified |
## Summary
When using `--agent-import-path` to specify a custom agent without also passing `-a`, the custom agent is ignored and the default `oracle` agent is used instead.
## Steps to Reproduce
1. Create a custom agent in `agents/my_agent.py`:
```python
from harbor.agents.installed.claude_code import ClaudeCode
... | https://github.com/harbor-framework/harbor/issues/261 | 1,766,327,412,000 | 1,766,388,023,000 | 1 | 1 | A | 4 | Configuration Error | [] | null | null |
harbor | harbor-framework/harbor | `claude-code` agent fails on Alpine-based task environments | ## Description
The `claude-code` agent's install script assumes a Debian/Ubuntu environment and fails on Alpine Linux containers because it uses `apt-get`, which doesn't exist on Alpine.
## Steps to Reproduce
```bash
uv run harbor jobs start -d compilebench@1.0 -a claude-code -m claude-sonnet-4-20250514 -t "coreutil... | https://github.com/harbor-framework/harbor/issues/258 | 1,766,312,345,000 | 1,766,470,579,000 | 1 | 0 | A | 1 | Architectural Constraint | [
"Supporting `alpine` is essential because it is a widely used base image due to its smaller size. Moreover, we find some models overfitting just for `ubuntu`. Still wonder if there is a more generic fix for that, though it may require architectural changes."
] | null | null |
harbor | harbor-framework/harbor | registry.json is invalid JSON after USACO adapter commit | ## Bug
The `registry.json` file is invalid JSON after commit b972d81 (Adding USACO Adapter).
## Error
```
json.decoder.JSONDecodeError: Expecting ',' delimiter: line 7161 column 9 (char 366068)
```
## Location
Line 7160-7164 shows the issue:
```json
"path": "datasets/aime/aime_ii-15"
... | https://github.com/harbor-framework/harbor/issues/255 | 1,766,281,308,000 | 1,766,388,479,000 | 1 | 0 | A | 1 | Validation Gap | [
"This is fixed now, thanks for the catch."
] | null | null |
harbor | harbor-framework/harbor | Trajectory viewer for non-linear history | ATIF allows the full history to be split and recorded in multiple trajectory files (useful for SFT purpose when there's summarization that breaks linear history).
E.g. if you run `uv run harbor run --agent terminus-2 --model openai/gpt-5 --dataset terminal-bench@2.0 --task-name write-compressor --env daytona --agent... | https://github.com/harbor-framework/harbor/issues/239 | 1,765,949,271,000 | null | 1 | 4 | A | 1 | Unimplemented Feature Gap | [
"@digitsisyph FYI in case you are interested"
] | null | null |
harbor | harbor-framework/harbor | environment.exec(): bash: cannot set terminal process group (-1): Not a tty bash: no job control in this shell | https://github.com/laude-institute/harbor/blob/2ad21b07db051721064f2ebc8830b2bd4205281c/src/harbor/environments/docker/docker.py#L261
https://github.com/laude-institute/harbor/blob/2ad21b07db051721064f2ebc8830b2bd4205281c/src/harbor/environments/docker/docker.py#L135
`_run_docker_compose_command` sets `stdin` to `D... | https://github.com/harbor-framework/harbor/issues/234 | 1,765,883,113,000 | null | 1 | 1 | A | 3 | Interface Contract Mismatch | [
"Can you try the change and lmk if it works? iirc we added the `-i` to ensure the startup files (e.g. `.bashrc`) get sourced ",
"https://chatgpt.com/share/694212c0-3890-800a-9696-7cebd26df7a4\n```bash \nbash -c 'source ~/.bashrc; your_command'\n```\n\nI’m not sure whether specific tasks depend on `.bashrc`, but i... | null | null |
harbor | harbor-framework/harbor | Integrate Tinker + Harbor for RL | For the integration, should we add a new `harbor train` command? Or a new `src/training` folder with prebuilt RL training integrations. This seems cleaner, given we'll have many training integrations (Tinker, SkyRL-train, SkyRL-tx, vanilla PyTorch, etc.).
See WIP for Tinker/SkyRL-tx/Prime-rl integration [here](https:... | https://github.com/harbor-framework/harbor/issues/231 | 1,765,842,186,000 | null | 0 | null | null | null | null | [
"I'm also working on Prime-RL + Harbor for RL training"
] | null | null |
harbor | harbor-framework/harbor | Enable overriding the agent setup timeout. | null | https://github.com/harbor-framework/harbor/issues/229 | 1,765,834,600,000 | 1,766,470,058,000 | 1 | 1 | A | 3 | Unimplemented Feature Gap | [] | null | null |
harbor | harbor-framework/harbor | Question about "GPU"-supported environment | Hi,
Is it allowed to set "GPU" configuration in `environment` in `task.toml` file? for example, `gpus = 2`
| https://github.com/harbor-framework/harbor/issues/227 | 1,765,781,597,000 | 1,766,388,787,000 | 0 | null | null | null | null | [
"are you thinking of running your own local cluster or cloud GPUs? I think modal could be a good option for you. ",
"@rs545837 Yes, right. You mean that `ModalEnvironment`?"
] | null | null |
harbor | harbor-framework/harbor | Add copilot CLI support | Please could someone work on copilot CLI support ? I could give it a try but I don't think I have a lot of time. It can run opus or gpt or gemini as well.
| https://github.com/harbor-framework/harbor/issues/218 | 1,765,576,090,000 | 1,775,682,269,000 | 0 | null | null | null | null | [
"Sure @Napolitain i would love to work on this issue\n",
"@dhruv-git-sys any update?",
"Holy cow thanks. "
] | null | null |
harbor | harbor-framework/harbor | 🚨 CRITICAL: Docker Compose project name validation bug breaks Oracle on custom_docker_compose tasks | # 🚨 CRITICAL: Harbor Docker Compose Project Name Validation Bug
**Priority:** 🔴 HIGHEST - Infrastructure Regression
**Impact:** Oracle solutions fail on tasks using `custom_docker_compose = true`
**Status:** Blocks CI validation of Harbor tasks
## Summary
Harbor generates trial IDs that start with underscores ... | https://github.com/harbor-framework/harbor/issues/204 | 1,765,379,347,000 | 1,765,688,827,000 | 1 | 2 | A | 3 | Validation Gap | [
"shoot i think this is happening bc of #140 ",
"this is fixed now"
] | null | null |
harbor | harbor-framework/harbor | Bug: Harden Docker `upload_dir` so `/tests` always works | In reference to the original find: https://github.com/badlogic/pi-terminal-bench
## Summary
- Normalize `DockerEnvironment.upload_dir` so Docker copies the contents of the host `tests/` directory into `/tests` even when that directory already exists in the container.
- Add regression tests that lock in the `/.` suffix... | https://github.com/harbor-framework/harbor/issues/202 | 1,765,333,578,000 | null | 1 | 1 | B | 3 | Validation Gap | [] | null | null |
harbor | harbor-framework/harbor | Support for SWE-Bench-Multilingual | It seems that only SWE-Bench-Verified is supported currently. It would be great if the Multilingual version is also supported. | https://github.com/harbor-framework/harbor/issues/196 | 1,765,170,869,000 | 1,775,537,265,000 | 1 | 1 | B | 1 | Unimplemented Feature Gap | [
"https://github.com/harbor-framework/harbor/pull/1004 merged last week!"
] | null | null |
harbor | harbor-framework/harbor | Revert litellm hack once upstream fix is released | Revert https://github.com/laude-institute/harbor/pull/181 once litellm fix is in-place and used by openhands' new version(s)
litellm PR: https://github.com/BerriAI/litellm/pull/17284 | https://github.com/harbor-framework/harbor/issues/182 | 1,764,490,581,000 | 1,768,189,171,000 | 0 | null | null | null | null | [
"@li-boxuan this has been fixed upstream, can we revert now?",
"It's fixed upstream in litellm but not picked up by latest openhands (1.0.0) yet"
] | null | null |
harbor | harbor-framework/harbor | Codex agent: Missing trajectory.json conversion (ATIF format) | # Codex agent: Missing trajectory.json conversion (ATIF format)
## Summary
The Codex agent successfully extracts token usage metrics but does **not** generate a `trajectory.json` file in ATIF (Agent Trajectory Interchange Format). This creates inconsistency with other agents (terminus-2, openhands, claude-code) that ... | https://github.com/harbor-framework/harbor/issues/179 | 1,764,459,682,000 | null | 1 | 4 | A | 1 | Unimplemented Feature Gap | [
"Yeah that's because @alexgshaw removed the log temporarily to debug sth. @alexgshaw did you get a chance to revert?"
] | null | null |
harbor | harbor-framework/harbor | Claude Code agent: Session directory permissions prevent trajectory/metrics extraction | # Claude Code agent: Session directory permissions prevent trajectory/metrics extraction
## Summary
The Claude Code agent creates session directories with restrictive **700 permissions** (owned by root), preventing Harbor from reading trajectory data. This results in:
- No `trajectory.json` file generated
- All token... | https://github.com/harbor-framework/harbor/issues/178 | 1,764,459,603,000 | 1,771,291,284,000 | 1 | 3 | A | 1 | Validation Gap | [
"@alexgshaw Could you have a look? \n\nI started testing harbor and really love it. The issue happened when I tried to understand all the features including ATIF. In my separate project, I have used \"Option 3: Workaround via docker-compose\" as a temporary, local workaround.\n\nI could submit a fix based on optio... | null | null |
harbor | harbor-framework/harbor | migrate tasks from tb 1.0 to harbor fails | I used the migration tool to migrate tb 1.0 tasks to harbor/tb 2.0
```
harbor tasks migrate -i terminal-bench-core==0.1.1 -o terminal-bench-core==0.1.1_migrated
```
but when running the tasks in harbor, it fails with the following error:
`invalid mount config for type "volume": invalid mount path: ':' mount path must... | https://github.com/harbor-framework/harbor/issues/165 | 1,764,082,148,000 | 1,764,356,664,000 | 1 | 0 | A | 2 | External Dependency Breakage | [
"just tried with a recent commit 62870f0b5b2f8d2a400a5c0eb915c7c7ce1ae361 and it works. i was previously using an installation from pypi or uv."
] | null | null |
harbor | harbor-framework/harbor | Feature Request: Add support for AI IDE | ### Feature Request: Add Support for Evaluating AI IDEs
**Is your feature request related to a problem? Please describe.**
Currently, Harbor evaluation framework only supports CLI-based agents and does not provide benchmarks or evaluation capabilities for AI-powered IDEs such as Cursor, Antigravity, Qoder, and Trae. ... | https://github.com/harbor-framework/harbor/issues/162 | 1,764,042,019,000 | null | 1 | 1 | A | 3 | Unimplemented Feature Gap | [
"You can somewhat easily implement this feature, but it really depends on each extension's registered commands (the quick way to do it using vscode extension test runner)\n\nI can share our solution which we use on azad.bot",
"> You can somewhat easily implement this feature, but it really depends on each extensi... | null | null |
harbor | harbor-framework/harbor | Feature Request: Harbor as a LLM proxy | Using Harbor with custom inference engines (vLLM/SGLang) running on local GPU cluster + agents running on remote cloud cluster (e.g. Daytona) is a common use case for evaluation and RL. It's non-trivial for agents to communicate with LLM without setting up a proxy server that has a public endpoint. As @penfever points ... | https://github.com/harbor-framework/harbor/issues/159 | 1,763,966,607,000 | null | 1 | 1 | A | 2 | Unimplemented Feature Gap | [
"LiteLLM provides an LLM proxy and could be used purely in that role—as an external dependency rather than an internal module. In this setup, all LLM configuration would be handled by the external proxy and managed as part of the environment abstraction."
] | null | null |
harbor | harbor-framework/harbor | Feature Request: Terminus-2 to support interleaved thinking | Some models like kimi-k2-thinking [1] and minimax-m2 [2] work best when the entire reasoning content history is included in subsequent LLM calls.
We could send back reasoning_content in Terminus-2 if configured.
Example of how to do it with litellm can be found in [OpenHands](https://github.com/OpenHands/software-age... | https://github.com/harbor-framework/harbor/issues/158 | 1,763,966,123,000 | null | 1 | 1 | A | 2 | Unimplemented Feature Gap | [] | null | null |
harbor | harbor-framework/harbor | Question: bug in summarization fallback logic in Terminus 2 agent | Thanks for building an awesome framework!
There seems to be a bug in the summarization fallback logic in the Terminus 2 agent.
If we hit `ContextLengthExceededError` in the `_query_llm` function, the _summarize function is called [here](https://github.com/laude-institute/harbor/blob/main/src/harbor/agents/terminus_2... | https://github.com/harbor-framework/harbor/issues/150 | 1,763,602,734,000 | 1,764,185,631,000 | 1 | 2 | A | 3 | Algorithmic Error | [
"Good catch, thank you! I'll fix it."
] | null | null |
harbor | harbor-framework/harbor | Docker environment fails when using relative path '.' - trial ID violates Docker Compose naming | ## Problem
When running Harbor with a Docker environment using a relative path `.` from inside a task directory, the generated trial ID starts with `__` (double underscores), which causes Docker Compose to fail.
## Error Message
```
invalid project name "__9mmlzw7": must consist only of lowercase alphanumeric charac... | https://github.com/harbor-framework/harbor/issues/140 | 1,763,446,286,000 | 1,765,688,843,000 | 1 | 2 | A | 1 | Validation Gap | [
"This is fixed now"
] | null | null |
harbor | harbor-framework/harbor | Question: Streaming on logs | I am not able to see agent stdout log in streamed fashion before agent execution finished?
| https://github.com/harbor-framework/harbor/issues/138 | 1,763,385,009,000 | null | 1 | 4 | A | 1 | Documentation Deficiency | [
"Are you using docker or a cloud sandbox?",
"I am using docker",
"if your docker containers are local, the logs directories are mounted so any files that are being continuously written to should be visible even in a partial state. let me know if you have a more specific example of what you are hoping to see"
] | null | null |
harbor | harbor-framework/harbor | fix: three terminal-bench@2.0 tasks repeatedly fail with oracle agent in daytona env | 👋 Hey harbor maintainers and contributors, I love the project and Discord community. Keep up the great work!
### TLDR
Using terminal-bench-2 at version a5b64de1851f6180dbc8dc46757890409bed4783, three specific tasks repeatedly fail `harbor run` with harbor version 2f122de46673322aa283d3a23057463973d52a89, oracle agen... | https://github.com/harbor-framework/harbor/issues/136 | 1,763,362,695,000 | null | 1 | 0 | B | 1 | External Dependency Breakage | [
"X-posting the hyperlink to the Discord discussion thread about this issue: https://discord.com/channels/1360039261361012928/1418249202089197699/1439873392747745350"
] | null | null |
harbor | harbor-framework/harbor | Flaky test_deterministic_openhands.py due to openhands' non-determinism | `test_deterministic_openhands.py` sometimes fails due to trajectory mismatch. This mismatch seems to come from openhands' own non-determinism.
```
{
- 'message': '',
+ 'message': 'Added workspace context',
'source': 'system',
'step_id': 4,
... | https://github.com/harbor-framework/harbor/issues/134 | 1,763,255,340,000 | null | 1 | 2 | A | 3 | Algorithmic Error | [] | null | null |
harbor | harbor-framework/harbor | Feature Request: Interactive Trajectory Viewer | Now to inspect trajectories we need to use a text editor or asciinema, which is a bit rough. Other agent frameworks already have nicer UIs for this, like [swe-agent’s inspector](https://swe-agent.com/latest/usage/inspector/) and [OpenHands/trajectory-visualizer](https://github.com/OpenHands/trajectory-visualizer).
It’... | https://github.com/harbor-framework/harbor/issues/128 | 1,763,114,807,000 | null | 1 | 4 | A | 5 | Unimplemented Feature Gap | [
"C.C. @harshraj172 who might have started working on that",
"@digitsisyph im one of the maintainers of Opik and we have a native Harbor intergration https://www.comet.com/docs/opik/integrations/harbor and we are 100% open-source. You can automatically send all your trajectories and traces into Opik and visualize ... | null | null |
harbor | harbor-framework/harbor | Feature Request: Support arbitrary tasks instead of datasets | Is there any way we could have arbitrary task support + optional verification ?
I've been looking for something to spin up lots of coding agents on isolated environments for a long time, and none of them have had the great infra that harbor does. However, it'd be really nice to not have it bound to the dataset case. ... | https://github.com/harbor-framework/harbor/issues/126 | 1,763,066,477,000 | null | 1 | 1 | B | null | Unimplemented Feature Gap | [
"I feel this could be a very useful dev feature for anyone building their own terminal tasks. It allows people to spin up a basic environment and a prompt and observe how agents behave under that setup, without needing to write verification yet.",
"@darinkishore can you say more about what you have in mind?\n\n@d... | null | null |
harbor | harbor-framework/harbor | Feature request: Installation of local agents | All installed agent examples are using remote repositories like git or pakage manager to be installed on installation script of the agent. There is no way to make a local wheel file of the agent and then using this in job execution. | https://github.com/harbor-framework/harbor/issues/124 | 1,762,995,860,000 | null | 1 | 0 | A | 4 | Unimplemented Feature Gap | [
"Check out https://github.com/laude-institute/harbor/pull/28",
"@li-boxuan when will be added?",
"@quancore This is not the highest priority right now but I'd like to hear more about your use case.",
"it is very iteresting that it is not the high priority. We have an agent where we would need multi-container ... | null | null |
harbor | harbor-framework/harbor | Feature Request: Add docker-compose.override.yml support for custom configurations | ## Feature Request: docker-compose.override.yml Support
### Background
After migrating from terminal bench to Harbor, I discovered that while environment variables were successfully migrated to Dockerfiles, other critical docker-compose configurations like `extra_hosts` and custom `volumes` are not easily configurable... | https://github.com/harbor-framework/harbor/issues/123 | 1,762,946,598,000 | null | 1 | 1 | A | 3 | Documentation Deficiency | [
"This is very much missing for multi-container custom agents",
"If you put a `docker-compose.yaml` file in the `environment` directory of your task and use the `DockerEnvironment` (`--env docker`) then Harbor will use that instead of the default docker compose.\n\nI'll add this information to the docs to make it ... | null | null |
harbor | harbor-framework/harbor | Running tasks with daytona seems flaky, timeouts commonly happen | <img width="1744" height="1402" alt="Image" src="https://github.com/user-attachments/assets/11a58960-17e3-4c18-82ca-fe236a116ad4" />
i'm noticing a lot of these errors when i try to spin up a bunch of containers in parallel with Harbor
from daytona team:
> "We haven’t seen this outside the harbor use case
> So seem s... | https://github.com/harbor-framework/harbor/issues/110 | 1,762,724,991,000 | 1,762,924,719,000 | 1 | 2 | A | 1 | External Dependency Breakage | [
"👋 Thanks for posting! I'm seeing failures running the following command for harbor main branch version 60676c3:\n```\nharbor run -d terminal-bench@2.0 -e daytona -n 32 --max-retries 2 --force-build\n```\ndue to the follow pseudo stacktrace:\n```\nRuntimeError: Session is closed\n at aiohttp/client.py:529 in _req... | null | null |
harbor | harbor-framework/harbor | Error handling on non-critical paths | I was using Sandboxes + Docker to run terminal-bench v1.5 tasks. A task's cleanup failed:
> Command: docker compose -p password-recovery__zkybimr -f /home/ec2-user/sandboxes/src/sandboxes/environments/docker/docker-compose-build.yaml down --rmi all --volumes --remove-orphans. Return code: 1. Stdout: Container passwor... | https://github.com/harbor-framework/harbor/issues/61 | 1,760,684,942,000 | 1,760,734,045,000 | 1 | 2 | A | 3 | Validation Gap | [] | null | null |
harbor | harbor-framework/harbor | Send CTRL+C upon agent timeout | I am working on making openhands gracefully shutdown (i.e. saving trajectory before exiting) upon receiving a SIGINT signal. I tested in terminal-bench that if harness sends a CTRL+C signal upon timeout, openhands (unreleased version) could properly save the trajectory. See https://github.com/laude-institute/terminal-b... | https://github.com/harbor-framework/harbor/issues/59 | 1,760,598,690,000 | null | 1 | 2 | A | 3 | Unimplemented Feature Gap | [
"Need to confirm but looks like sandboxes doesn't need this.\n\nIn terminal-bench harness we need to do this because agents run in Tmux, and are thus not cancelled by the agent coroutine when timeout happens.",
"Okay I am fairly confident this is not needed in sandboxes.",
"Reopening; this is less of a bug in s... | null | null |
harbor | harbor-framework/harbor | Alternative way to install agents | Currently we install agents on the fly. This is very time consuming and a waste of resources. For 200 tasks x 3 agents (I am just making numbers up), we need 600 installation (via pypi or npm).
An alternative way is to bake agent installation into task images. This is not ideal because it means we have number of image... | https://github.com/harbor-framework/harbor/issues/25 | 1,757,391,823,000 | null | 1 | 0 | A | 5 | Architectural Constraint | [
"We currently cannot include https://github.com/laude-institute/terminal-bench/tree/main/tasks/extract-safely in t-bench because it's unfair to installed agents, but if we use the pre-install or mounting approach then it won't be a problem",
"Gemini researched the three aforementioned approaches and recommended t... | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.