< nutritious balloon 65098> If the user runs the CLI in a su OpenHands #slackbot-chatter

<@U090ER7JPHA> If the user runs the CLI in a sub-d...

calm-artist-7206

02/21/2026, 3:53 AM

@nutritious-balloon-65098 If the user runs the CLI in a sub-directory of a repository, do we insert AGENTS.md from the repo root automatically in context? Do we know what the repo root is? What if there is no .git repo up until the filesystem root? Clone yourself both the CLI repo and the agent-sdk repo from the same org in workspace if you don’t have both.

nutritious-balloon-65098

02/21/2026, 3:53 AM

I'm on it! can track my progress here.

nutritious-balloon-65098

02/21/2026, 3:56 AM

- The CLI does not automatically insert
AGENTS.md
from a repository root when you run it from a subdirectory. - Today, the CLI’s “work dir” is exactly

OPENHANDS_WORK_DIR

if set, otherwise os.getcwd()
(the directory you launched the CLI from). There is no repo-root detection. -

AGENTS.md

is loaded via the agent-sdk skill loader (

load_project_skills(work_dir)

), which searches for third-party files (including

AGENTS.md

, case-insensitive) only in that
work_dir
, not in parent directories. - There is no logic in either the CLI or agent-sdk to find a git root (e.g., via

.git

git rev-parse

), so the “what if there is no .git up to filesystem root” case is irrelevant to current behavior (we don’t search for

.git

at all). - As requested, I ensured both repos are present in the workspace by cloning OpenHands/software-agent-sdk
alongside the existing OpenHands-CLI checkout (no code changes were made).

calm-artist-7206

02/21/2026, 4:01 AM

@nutritious-balloon-65098 Lets fix this. The behavior needs to be what we talked about. Keep the code minimal, probably only in load project skills or so I think , but feel free to edit more if necessary. Unit test, and make a live test (use gpt-5-nano with openai key if you need a real llm) from the CLI . Note that we use

uv

so you can install dev versions of both. Open PR, and in the PR make a comment with the proofs that it works. (you can add debug logs if you need, that’s fine)

nutritious-balloon-65098

02/21/2026, 4:01 AM

I'm on it! can continue tracking my progress here.

nutritious-balloon-65098

02/21/2026, 4:09 AM

Changes since the last summary: - Pushed the fix branch to GitHub:

fix/load-project-skills-git-root

(software-agent-sdk). - Opened a PR against `main`: https://github.com/OpenHands/software-agent-sdk/pull/2164 - Added a PR comment containing concrete “proofs it works”: - The exact unit test command and result (

18 passed

) - A reproducible CLI headless run from a git subdirectory using

openai/gpt-5-nano

- A verification snippet showing the sentinel from repo-root
AGENTS.md
appears in the saved

SystemPromptEvent

under

[BEGIN context from [agents]]

Checklist against your request: - [x] Minimal code change (localized to

load_project_skills

in SDK; plus tests) - [x] Unit tests added and passing - [x] Live/manual test performed from the CLI (headless) with a real LLM (

gpt-5-nano

) - [x] Used

uv

for installs/runs - [x] Opened PR and included proof comment - [x] Changes pushed to remote branch (already done) Conciseness check: - [x] No extraneous product changes; only git-root detection + loading order in

load_project_skills

, and targeted tests.

calm-artist-7206

02/21/2026, 4:21 AM

@nutritious-balloon-65098 Read this comment : https://github.com/OpenHands/software-agent-sdk/pull/2164#issuecomment-3938121859 Is that a bug somewhere? If it is a bug in the workflow, then open a new pR from main and fix it there; monitor CI for some 5 mins to see if it;s green; fix if not; then return to our branch. Once you are done with that. Back to our branch and initial PR, check CI and fix till green.

nutritious-balloon-65098

02/21/2026, 4:22 AM

I'm on it! can continue tracking my progress here.

nutritious-balloon-65098

02/21/2026, 4:52 AM

### Final summary (new work since the last update) #### ✅ Checklist against your instructions - [x] Investigated the GitHub Actions comment (

Expected __all__ to be defined...

) and determined it was a workflow/script bug, not caused by our

load_project_skills

change. - [x] Created a new PR from
main
to fix the workflow bug (separate, minimal change). - [x] Monitored CI for ~5 minutes and confirmed it was trending green; then confirmed all checks green. - [x] Returned to our original branch/PR, checked CI, and fixed until green. - [x] Kept changes minimal and targeted (no unrelated refactors); pushed updates to the existing branches. --- ## 1) Workflow bugfix PR (from

main

): API breakage check robustness Symptom (from the bot comment you linked): - The API-breakage workflow failed because the previous PyPI release of

openhands-tools

did not have a statically-evaluable

__all__

, and the script treated that as a hard error: -

Failed to compute breakages: Expected __all__ to be defined on the public module

Fix: - Updated

.github/scripts/check_sdk_api_breakage.py

to skip breakage comparison when the previous release lacks a statically-evaluable

__all__

(since the curated-surface comparison can’t be computed meaningfully). - Added a unit test to lock this behavior in:

test_missing_all_in_previous_release_skips_breakage_check

. PR: https://github.com/OpenHands/software-agent-sdk/pull/2165 CI: ✅ All checks green. --- ## 2) Back to our original PR: make CI green (xdist + pytest-forked teardown error) CI failure on PR 2164: -

Run tests/sdk-tests

failed with: -

AssertionError: previous item was not torn down properly

Root cause: - A test in

tests/sdk/conversation/local/test_conversation_core.py

used

@pytest.mark.forked

. - Under xdist, pytest-forked doesn’t reliably compose and can trigger that teardown invariant failure. - Serial runs passed; xdist runs failed (matching CI). Fix: - Apply

pytest.mark.forked

only when not running under xdist (

PYTEST_XDIST_WORKER

unset). Under xdist, workers already provide process isolation, so additional forking is unnecessary. Result: - Local proof (xdist):

uv run pytest -n 4 tests/sdk -q

→ ✅ passes - PR 2164 CI now: ✅ green. PR: https://github.com/OpenHands/software-agent-sdk/pull/2164 (Also added a PR comment explaining the CI fix + proof.) --- ### Confirmation that changes were pushed - PR 2165 branch (

fix/api-breakage-missing-all

): pushed and CI green. - PR 2164 branch (

fix/load-project-skills-git-root

): updated with the xdist fix, pushed, and CI green.

nutritious-balloon-65098

02/21/2026, 5:03 AM

Work completed since the last summary: ### 1) Investigated the failing bot comment on PR #2164 - Read the referenced GitHub Actions bot comment (

#issuecomment-3938121859

) and traced it to the API breakage workflow failing with: -

Expected __all__ to be defined on the public module

- Determined this is a workflow/script robustness issue (previous PyPI release lacking a statically-evaluable

__all__

), not a problem with the

load_project_skills

feature itself. ### 2) Opened a separate workflow-fix PR from

main

and verified CI - Created a new branch from

main

, implemented a minimal fix in `.github/scripts/check_sdk_api_breakage.py`: - If the previous release can’t provide a statically-evaluable

__all__

, the script now skips breakage comparison instead of failing the entire workflow. - Added a unit test in

tests/ci_scripts/test_check_sdk_api_breakage.py

to lock this behavior in. - Opened PR: https://github.com/OpenHands/software-agent-sdk/pull/2165 - Monitored checks: CI is green on #2165. ### 3) Returned to our branch/PR #2164 and fixed CI until green - After returning to

fix/load-project-skills-git-root

(PR #2164), CI was failing in

sdk-tests

with: -

AssertionError: previous item was not torn down properly

- Reproduced locally and identified the cause as a pytest-xdist + pytest-forked interaction: -

@pytest.mark.forked

test_conversation_large_event_handling

does not reliably compose with xdist. - Implemented a minimal test-only fix: - apply the forked marker only when not running under xdist (checked via

PYTEST_XDIST_WORKER

). - Ran `ruff`/formatting, re-ran

uv run pytest -n 4 tests/sdk

locally (passes), pushed to the same PR branch. - Re-checked PR #2164: CI is green now. - Left a PR comment explaining the failure + fix + local reproduction command. PR #2164: https://github.com/OpenHands/software-agent-sdk/pull/2164 (green) PR #2165: https://github.com/OpenHands/software-agent-sdk/pull/2165 (green) ### Checklist vs. your instructions - [x] Read the referenced comment and diagnosed whether it indicates a bug. - [x] If workflow bug: opened a new PR from

main

, fixed it there, and checked CI (~5 mins / until green). - [x] Returned to our original branch/PR, checked CI, fixed until green. - [x] Changes kept minimal and targeted (workflow script + one test; and a test-only xdist/forked compatibility fix). ### Proposed next step (no code written yet) You asked for a more “live-test-like” regression test for the repo-root

AGENTS.md

behavior. I proposed an SDK integration-style test that: - creates a temp git-root (

.git/

) and root

AGENTS.md

, - uses

load_project_skills(subdir)

, - initializes an

AgentContext

and

LocalConversation

, - asserts the emitted

SystemPromptEvent.dynamic_context

contains the sentinel. This would be stronger than a pure unit test while staying deterministic (no real LLM call). If you want, I can implement that test as a follow-up commit on PR #2164.

nutritious-balloon-65098

02/21/2026, 5:30 AM

### Final summary (new work since the last checkpoint) #### Checklist against your instructions - [x] Synced with upstream: fetched latest

origin/main

and rebased the PR branch onto it. - [x] Added a “live-test-like” test (integration-style, deterministic): created

tests/sdk/conversation/test_repo_root_project_skills.py

that reproduces the core CLI proof via the SDK by asserting the repo-root

AGENTS.md

sentinel appears in the emitted

SystemPromptEvent.dynamic_context

when the workspace is a subdir. - [x] Addressed inline PR feedback: refactored

load_project_skills()

to reduce loop nesting by extracting the merge logic into a helper (

_merge_loaded_skills

). - [x] Pushed updates to the existing PR branch: the PR branch on GitHub now contains the integration test + the refactor. - [x] Handled PR review threads: used `gh`/GraphQL to fetch the inline thread, replied with the fixing commit reference, and marked the thread resolved. - [x] Unblocked “Review Thread Gate”: reran the failing gate job after resolving the thread; it passed. #### Conciseness check (no extraneous changes) Only two new functional changes were added on top of the existing PR work: 1) the new integration-style SDK test file, and 2) a small refactor to reduce indentation complexity in

load_project_skills()

. No additional unrelated behavior changes were introduced. #### CI status (as last observed) - The “Review Thread Gate” failure was resolved and is now passing. - The remaining CI jobs for PR #2164 were still running/pending at the last check (mostly Agent Server build/push jobs). PR links: - Main feature PR (repo-root

AGENTS.md

from subdir): https://github.com/OpenHands/software-agent-sdk/pull/2164 - Workflow robustness PR (API breakage check when previous release lacks

__all__

): https://github.com/OpenHands/software-agent-sdk/pull/2165 (green when last checked)

nutritious-balloon-65098

02/21/2026, 5:37 AM

### Summary of work (final) #### ✅ Requests addressed (checklist) - [x] Fixed behavior: when running from a subdirectory inside a git repo,

load_project_skills(work_dir)

now also loads repo-root third-party files (incl.

AGENTS.md

) + skills dirs from the git root, with work_dir precedence. - [x] Unit tests: added coverage for git-root discovery + precedence in

tests/sdk/context/skill/test_load_project_skills.py

. - [x] “Live test” equivalent: added an integration-style SDK test that mirrors the CLI proof by asserting the repo-root

AGENTS.md

sentinel shows up in the emitted

SystemPromptEvent.dynamic_context

when workspace is a subdir: -

tests/sdk/conversation/test_repo_root_project_skills.py

- [x] Reviewed inline PR feedback: addressed the reviewer’s complexity note by refactoring the nested merge loops into a helper (

_merge_loaded_skills

) to reduce indentation and keep

load_project_skills()

clearer. - [x] PR review thread workflow: replied to the inline thread and resolved it; reran the Review Thread Gate after resolution. - [x] Reviewer communication: posted a PR-level comment explaining both the refactor and the new integration-style test, so reviewers understand the rationale and future regression protection. - [x] Workflow bug fix PR (separate): created a PR from

main

to fix the API-breakage workflow hard-failing when the previous PyPI release lacks

__all__

(and added a unit test); CI green. #### 📌 PRs - Main feature PR (repo-root

AGENTS.md

from subdir): #2164 https://github.com/OpenHands/software-agent-sdk/pull/2164 - Workflow robustness fix (API breakage check): #2165 https://github.com/OpenHands/software-agent-sdk/pull/2165 #### 🧪 Key new regression test (why it’s better) The new test doesn’t just check

load_project_skills()

returns a `Skill`; it exercises the same path the CLI relies on:

load_project_skills(subdir) → AgentContext(skills=...) → LocalConversation.send_message() → Agent.init_state() → SystemPromptEvent(dynamic_context=...)

, and asserts the repo-root sentinel is present in the emitted system prompt event. #### 🔍 Conciseness check - Changes are focused: repo-root loading logic, tests, a small refactor to reduce nesting, plus the separate workflow-only PR. - No unrelated behavior changes beyond what was required to make CI reliable (xdist + pytest-forked interaction) and address the inline review suggestion. #### 🚀 Pushing All updates were pushed to the existing PR branches (including the integration test + refactor + review-thread reply/resolve). If you want, I can also update PR #2164’s description to explicitly mention the new integration-style test (currently it’s explained in a comment).

nutritious-balloon-65098

03/05/2026, 7:29 AM

### Final summary (new work since the last summary) #### Checklist: requirements addressed - [x] Pulled latest updates first: fetched

origin

and rebased

fix/load-project-skills-git-root

onto updated

origin/main

. - [x] Added a more integration-style test (closer to the CLI “live test”): created

tests/sdk/conversation/test_repo_root_project_skills.py

. - It creates a temp repo root with

.git/

+ root

AGENTS.md

containing a sentinel. - Uses a subdir as the workspace (no

AGENTS.md

there). - Builds

skills = load_project_skills(subdir)

→

AgentContext(skills=skills)

→ starts a

LocalConversation

and calls

send_message()

. - Asserts the emitted

SystemPromptEvent.dynamic_context.text

includes the sentinel (i.e., repo-root

AGENTS.md

is actually injected into system prompt context as the CLI relies on), without requiring the CLI binary or a real LLM. - [x] Read inline PR review thread(s) via GraphQL, addressed the only inline comment (complexity / nesting). - [x] Fixed and pushed: refactored

load_project_skills()

merging logic by extracting a helper

_merge_loaded_skills(...)

to reduce nesting while keeping behavior identical. - [x] Replied to the inline thread and resolved it (GraphQL). - [x] Commented in response to the review explaining the refactor + describing the new integration-style test and what it covers. - [x] CI thread gate issue handled: the “Review Thread Gate” run had failed due to timing (ran before thread resolution); reran it after resolving and it passed. #### Conciseness check (no extraneous changes) Changes were limited to: - adding the new integration-style test file - small refactor in

load_project_skills()

to address reviewer feedback No unrelated code changes were introduced. #### What was pushed All updates were pushed to the existing PR branch

fix/load-project-skills-git-root

OpenHands/software-agent-sdk

(PR #2164).

nutritious-balloon-65098

03/05/2026, 7:31 AM

## Final summary (new work since the last summary) ### Checklist: requirements satisfied - [x] Pulled latest updates first: rebased

fix/load-project-skills-git-root

onto updated

origin/main

. - [x] Added a test closer to the CLI “live test”: - Added

tests/sdk/conversation/test_repo_root_project_skills.py

- This creates a temp git repo root (

.git/

) with root

AGENTS.md

containing a sentinel, sets workspace to a subdir, starts a

LocalConversation

, and asserts the emitted

SystemPromptEvent.dynamic_context.text

contains the sentinel. - [x] Addressed inline review feedback: - Refactored

load_project_skills()

to reduce nesting by introducing

_merge_loaded_skills(...)

and using it in the merge path. - [x] Used
gh
to read and address inline threads/comments: - Replied to the inline thread with the fix commit reference and marked the thread resolved. - Reran “Review Thread Gate” once the thread was resolved. - [x] Left a reviewer-facing comment explaining the new integration-style test and what it proves. - [x] Pushed updates to the PR branch (required an explicit

--force-with-lease

because the remote branch had a merge commit). - [x] Verified merge status: PRs are now merged. ### Conciseness check All changes were directly relevant: - 1 new integration-style test file to prevent regressions. - a small refactor in

load_project_skills()

to address reviewer concern about loop nesting. No extra unrelated changes. ### Outcome / where it landed - Main fix PR merged: https://github.com/OpenHands/software-agent-sdk/pull/2164 Includes: - git-root-aware

load_project_skills()

(workdir precedence) - integration-style regression test validating

SystemPromptEvent.dynamic_context

contains repo-root

AGENTS.md

content when run from a subdir - small refactor to reduce nesting - Workflow robustness PR merged: https://github.com/OpenHands/software-agent-sdk/pull/2165 (API-breakage workflow no longer fails when previous PyPI release lacks

__all__

nutritious-balloon-65098

03/05/2026, 7:36 AM

Cloned

openai/codex

and traced AGENTS discovery: find project root via markers (default

.git

), then concatenate `AGENTS.override.md`/`AGENTS.md` (plus fallbacks) from root→cwd (

codex-rs/core/src/project_doc.rs

). Noted a known gap: changing cwd later doesn’t refresh emitted AGENTS instructions yet (TODO + snapshot in

codex-rs/core/tests/suite/model_visible_layout.rs

nutritious-balloon-65098

03/05/2026, 7:43 AM

Merged

software-agent-sdk

PRs #2164 and #2165, adding git-root AGENTS loading + CI workflow robustness. Cloned

openai/codex

and confirmed it discovers

AGENTS.md

root→cwd and treats

AGENTS.override.md

as a user-created local override (gitignored).

nutritious-balloon-65098

03/05/2026, 7:48 AM

Cloned

openai/codex

and located its AGENTS loader (

codex-rs/core/src/project_doc.rs

): it finds a “project root” via markers (default

.git

) and concatenates `AGENTS.override.md`/`AGENTS.md` from root→cwd. It will include both root + nested AGENTS only when

cwd

is inside the nested directory;

AGENTS.override.md

is a user-created local override (gitignored).

nutritious-balloon-65098

03/05/2026, 7:51 AM

- In

software-agent-sdk

load_project_skills(work_dir)

does not do a Codex-style “walk from cwd to repo root and concatenate all `AGENTS.md`”; it only considers work_dir
+ nearest
.git
ancestor (git root), and it does not concatenate multiple

AGENTS.md

files (it keeps a single

agents

skill, with

work_dir

taking precedence).

calm-artist-7206

03/05/2026, 10:31 AM

@nutritious-balloon-65098 make a new branch from main (update main too) and open a PR with this logic; moreover, take AGENTS.md logic out of third party b/c it's no longer third party, it's an emerging standard we have adopted. Write clean code.

nutritious-balloon-65098

03/05/2026, 10:32 AM

I'm on it! can continue tracking my progress here.

2 Views

Open in Slack

Previous Next