feat(sdk): reduce round trips when offloading large tool results with `sandbox.execute` by ccurme · Pull Request #4230 · langchain-ai/deepagents

ccurme (ccurme) · 2026-06-24T19:59:51Z

When the execute tool receives a large tool result and you're using a sandbox:

The result is returned to the agent server
The agent server offloads it back to the sandbox (assuming no re-routing of large tool results using a CompositeBackend)

Here we remove the round trip by wrapping all execute commands with additional shell commands inside the execute tool. The additional shell commands:

Redirect the command's combined stdout/stderr to a file in the sandbox (> "$__da_f" 2>&1)
Measures the captured size (wc -c)
If under eviction threshold, prints the full output inline and deletes the file (no offloading)
If above eviction threshold, leaves the file at artifacts_root/large_tool_results/tool_call_id and prints the content preview
Emits a one-line machine-readable header that the agent server parses to rebuilt the ExecuteResponse.

This only engages when the backend is a BaseSandbox and the offload path routes to the same backend (so no changes if we're routing large tool results away with CompositeBackend).

We also add a flag enable_capture_offload to BaseSandbox as an escape hatch.

Also have a standard integration test here. Passes with LangsmithSandbox although I haven't got creds yet for any others: https://github.com/langchain-ai/langchain/compare/cc/standard-tests-sandboxes

ccurme (ccurme) · 2026-06-24T20:00:41Z

+# command `exit` cannot abort the wrapper, and `eval` preserves the backend's own
+# shell/env. The command is embedded via a quoted heredoc with a random delimiter
+# to avoid shell-quoting issues; the (internal, sanitized) path is shell-quoted.
+_EXECUTE_CAPTURE_CMD_TEMPLATE = """# ===== deepagents capture-at-source offload (auto-generated wrapper) =====


Breakdown of what's happening here:

Redirect the command's combined stdout/stderr to a file in the sandbox (> "$__da_f" 2>&1) instead of streaming it back, so the full payload never leaves the sandbox.

Embed the requested command verbatim via a quoted heredoc (cat <<'DELIM' … DELIM) and run it in a subshell via eval (( eval "$__da_cmd" )). The heredoc avoids any shell-quoting pitfalls; the subshell+eval runs the command in the sandbox's own shell/env and ensures a stray exit in the user's command can't abort the wrapper.

Pipe the captured stream through head -c (default 10 MiB) so the on-disk file can never exceed a hard cap — this bounds sandbox disk usage and restores backpressure (the writer gets SIGPIPE) for runaway output like yes.

Recover the command's real exit code from a sidecar file (echo "$?" > "$__da_f.ec"). The capture pipe would otherwise leave $? as head's exit code, not the command's.

Measure the captured size (wc -c) and branch:

Under the eviction threshold → print the full output inline and delete the file. Indistinguishable from today's small-result behavior.

Over the threshold → leave the file at <artifacts_root>/large_tool_results/<tool_call_id> and print only a head/tail preview (whole lines, byte-bounded) with a [N lines truncated] marker.

Emit a one-line machine-readable header — <exit_code> — that the agent server parses to rebuild the ExecuteResponse and decide between returning the output inline vs. a preview + a read_file pointer.

open-swe

Open SWE Review found 1 potential issue.

Open in Web • View Open SWE trace

Eugene Yurtsev (eyurtsev)

lgtm, but maybe we could push more code into sandbox and out of file system (perhaps a method on base sandbox that accepts max output size?)

Mason Daugherty (mdrxy)

doctring

addressed comment

> [!CAUTION] > Merging this PR will automatically publish to **PyPI** and create a **GitHub release**. For the full release process, see [`.github/RELEASING.md`](https://github.com/langchain-ai/deepagents/blob/main/.github/RELEASING.md). --- _Release notes preview: keep this section in sync with the package `CHANGELOG.md`. Publish reads the merged CHANGELOG via `release.yml`, not this PR description — keep them aligned anyway so the PR stays an accurate historical record for reviewers and anyone returning later._ --- ## [0.7.0](deepagents==0.6.12...deepagents==0.7.0) (2026-07-29) See [the docs](https://docs.langchain.com/oss/python/releases/changelog#deepagents-v0-7-0) for curated release notes. ### ⚠ BREAKING CHANGES * `create_deep_agent` no longer includes `TodoListMiddleware` by default, the `write_todos` tool, `todos` state channel, and todo-planning prompt are now absent. Pass `middleware=[TodoListMiddleware()]` to restore them on the main agent; add it to each `SubAgent`'s middleware to restore them there. ([#4929](#4929)) ([9340518](9340518)) * Default agent prompts are now lean: the authored base prompt is empty, and tool-usage prose that duplicates tool schemas is trimmed. `BASE_AGENT_PROMPT` is deprecated (removal in `deepagents==0.9.0`) but remains importable and still returns the previous authored prompt verbatim; pass it as `create_deep_agent(system_prompt=BASE_AGENT_PROMPT)` to restore the old behavior. ([#4859](#4859)) ([#4979](#4979)) ([a8d1b32](a8d1b32)) ([d9f54fc](d9f54fc)) * The built-in tool-usage prompt constants `TASK_SYSTEM_PROMPT`, `ASYNC_TASK_SYSTEM_PROMPT`, `SUMMARIZATION_SYSTEM_PROMPT`, `FILESYSTEM_SYSTEM_PROMPT`, and `EXECUTION_SYSTEM_PROMPT` are removed, and the `system_prompt` default on `SubAgentMiddleware`, `AsyncSubAgentMiddleware`, `SummarizationToolMiddleware`, and `create_summarization_tool_middleware` is now `None`, which injects no prose. Pass your own string to restore prompt text. ([#4859](#4859)) ([a8d1b32](a8d1b32)) * `FilesystemBackend` and `LocalShellBackend` now default to `virtual_mode=True`. Filesystem paths are anchored under `root_dir`, `..` traversal is rejected, and paths resolving outside `root_dir` raise `ValueError`. Previously an unspecified `virtual_mode` emitted a deprecation warning and fell back to `False`, where absolute host paths were used as-is and `..` could escape `root_dir`. Pass `virtual_mode=False` explicitly to restore the old filesystem behavior. ([#4541](#4541)) ([540a0fa](540a0fa)) * Agents now see a destructive, recursive `delete` filesystem tool whenever the backend supports it, and filesystem permissions classify `delete` as a write operation — so an existing rule allowing writes to a path also authorizes recursively deleting that subtree unless a narrower deny or interrupt rule covers the target. Because recursive deletes affect descendants, deny and interrupt checks use bulk path overlap instead of exact-path matching. To keep the previous behavior, add a deny or interrupt rule, or omit `delete` from `FilesystemMiddleware(tools=...)`. Missing paths return a not-found error, `CompositeBackend` reports an unsupported-operation error when a routed sub-backend cannot delete, and the tool is hidden from the model entirely when the backend itself does not implement it. ([#3659](#3659)) ([#3691](#3691)) ([#3765](#3765)) ([#3851](#3851)) ([f2a21ec](f2a21ec)) * `write_file` can now create a file if it is missing and replaces it entirely if it already exists, instead of returning a file-exists error. The `write_file` tool description no longer requires reading the file first. There is no "create-only" compatibility mode. Workflows, prompts, tests, or guardrails that relied on the file-exists error to force `edit_file` usage or to protect existing content must omit `write_file`, add explicit permission or interrupt rules, or use `edit_file` where preserving existing content matters. ([#4109](#4109)) ([2506fcc](2506fcc)) * Removed deprecated backend compatibility shims. Callers must pass concrete `BackendProtocol` instances (not factories), configure `StoreBackend` with an explicit `namespace`, and use the current `ls` / `glob` / `grep` / `ReadResult` APIs. ([#4541](#4541)) ([540a0fa](540a0fa)) * The deprecated `files_update` attribute and constructor keyword are removed from `WriteResult` and `EditResult`. Custom backends must stop passing `files_update=`, and callers must stop reading `result.files_update`; state writes are emitted directly by `StateBackend`. ([#4541](#4541)) ([540a0fa](540a0fa)) * Removed the deprecated `BackendProtocol` methods `ls_info`, `als_info`, `glob_info`, `aglob_info`, `grep_raw`, and `agrep_raw`. Use `ls` / `glob` / `grep` and their async counterparts. ([#4541](#4541)) ([540a0fa](540a0fa)) * `SummarizationMiddleware(history_path_prefix=...)` was removed and now raises `TypeError`. Configure `CompositeBackend(artifacts_root=...)` instead. ([#4541](#4541)) ([540a0fa](540a0fa)) * Agent-facing `ls` and `glob` tool output now renders empty results as `No files found` instead of `[]`; direct backend APIs continue to return structured empty `LsResult` and `GlobResult` values. Callers that parse tool output should update those checks. ([#3709](#3709)) ([efafd1e](efafd1e)) * `read_file` no longer renders raw text with a fixed-width `cat -n`-style line-number gutter and tab separator. Line and continuation markers are dynamically aligned and separated from source content by two spaces, and the `LINE_NUMBER_WIDTH` constant is removed from `deepagents.backends.utils` and `deepagents.middleware.filesystem`. Callers that parse raw tool output should update those parsers. ([#4561](#4561)) ([cf057b4](cf057b4)) ### Features * Custom middleware passed to `create_deep_agent(..., middleware=[...])` can replace a default middleware instance when `.name` matches, so defaults such as `SummarizationMiddleware` can be overridden without also excluding the built-in instance. ([#4251](#4251)) ([90c8472](90c8472)) * `FilesystemMiddleware(tools=[...])` accepts a keyword-only allowlist of built-in filesystem tools, typed by the newly exported `FsToolName` literal (`"ls"`, `"read_file"`, `"write_file"`, `"edit_file"`, `"delete"`, `"glob"`, `"grep"`, `"execute"`); pass `"all"` or omit the argument to keep every tool. A list must include `"read_file"` or the constructor raises `ValueError`. Omitted built-in tools are non-executable, and custom user tools are unaffected. ([#4325](#4325)) ([#4698](#4698)) ([704a70d](704a70d)) ([9709525](9709525)) * Shorten LLM-facing descriptions for the `task` tool and filesystem tools (`read_file`, `grep`, `edit_file`, `glob`, `execute`). ([#5009](#5009)) ([761f5f0](761f5f0)) * `GrepResult` and `GlobResult` now carry a `truncated` flag so supporting backends can return valid partial results when a match cap or backend deadline is reached; agent-facing tool output adds a note telling the model to narrow the search. `FilesystemBackend` returns partial `grep` and `glob` results on its backend timeout rather than erroring, while other backend or middleware timeouts may still return errors. Its `glob` also gains brace expansion such as `*.{py,md}` (already supported by the state and store backends). ([#4063](#4063)) ([ef591e7](ef591e7)) * The agent-facing `grep` match cap is configurable: `FilesystemMiddleware(grep_max_count=...)` sets the default (`1000`; `None` disables it) and the model can override it per call through the tool's new `max_count` argument. `grep` / `agrep` on `BackendProtocol` and all built-in backends accept a keyword-only `max_count`. Local ripgrep output is streamed and terminated once the cap is reached. Direct `FilesystemBackend.grep()` callers can request surrounding lines with keyword-only `context_lines`. ([#4570](#4570)) ([#4706](#4706)) ([8e86f5e](8e86f5e)) ([65230df](65230df)) * Paginated built-in `read_file` responses report the returned source-line range and next `offset`; total and remaining line counts are included when the backend knows the file length. Resume offsets remain safe when sandbox or middleware limits shorten the visible page. ([#4540](#4540)) ([8321194](8321194)) * Optional video frame extraction for `read_file`, enabled by the new `deepagents[video]` extra. Video files are sampled into JPEG frames, with `offset` and `limit` interpreted as seconds. Without the extra, existing generic video/file content-block behavior remains. ([#4094](#4094)) ([b927147](b927147)) * `FilesystemMiddleware` can capture oversized `execute` tool output directly inside the sandbox artifact path on compatible, opted-in `BaseSandbox` implementations to reduce round trips; `LangSmithSandbox` opts in by default. ([#4230](#4230)) ([02f5bd7](02f5bd7)) * Automatically enable Fireworks prompt-cache session affinity when a compatible `langchain-fireworks` installation is available. ([#4598](#4598)) ([5d878bf](5d878bf)) * Add a built-in NVIDIA Nemotron 3 Ultra harness profile and NVIDIA NIM app-origin attribution. ([#4192](#4192)) ([#4455](#4455)) ([d5a60ec](d5a60ec)) ([4cb4749](4cb4749)) * `RubricMiddleware` now accepts any positive `max_iterations` cap instead of enforcing a hard upper bound. ([#4405](#4405)) ([d6692a7](d6692a7)) ### Bug Fixes * Keep fields marked with `PrivateStateAttr`, including fields declared through `create_deep_agent(state_schema=...)`, out of subagent inputs and returned parent-state updates. ([#4587](#4587)) ([a4662c0](a4662c0)) * Preserve `ContextT` through the `create_deep_agent(..., middleware=[...])` type annotation so type checkers accept context-aware middleware when a matching `context_schema` is passed. ([#4055](#4055)) ([7be76c7](7be76c7)) * Accept YAML list values as well as comma-separated strings for skill `allowed-tools` frontmatter, and make skill truncation warnings actionable with field name, path, length, configured limit, and impact. ([#4140](#4140)) ([#4141](#4141)) ([d62534c](d62534c)) ([2f5f5b8](2f5f5b8)) * Align filesystem instructions with the tools that remain after allowlist and backend-capability filtering, so agents no longer reference hidden `grep`/`glob` tools or prohibit equivalent shell search when dedicated search tools are unavailable. ([#4920](#4920)) ([#4921](#4921)) ([d3650c7](d3650c7)) ([b65cc00](b65cc00)) * Propagate default-backend failures from `CompositeBackend.ls("/")` and `CompositeBackend.als("/")` instead of returning successful route-only listings. ([#4925](#4925)) ([4c3b166](4c3b166)) * Correct `CompositeBackend.glob` / `CompositeBackend.aglob` routing so explicit default-backend paths such as `/tools` do not also return files from routed backends such as `/memories`. ([#4531](#4531)) ([cbdb0a7](cbdb0a7)) * Propagate default- and routed-backend failures from root `CompositeBackend.glob(..., path=None)` / `aglob(..., path=None)` and `path="/"` searches instead of returning incomplete successful results. ([#4063](#4063)) ([ef591e7](ef591e7)) * Constrain sandbox `glob` and slash-pattern `grep` searches to their declared search root by treating leading `/` as search-root-relative, rejecting `..` traversal segments, and filtering symlink-resolved matches outside the root. ([#4588](#4588)) ([c6c7213](c6c7213)) * Unify `grep(..., glob=...)` include-glob semantics across filesystem and in-memory backends: basename patterns like `*.py` match at any depth, and slash-containing patterns like `src/**/*.py` match relative paths consistently. ([#3936](#3936)) ([feab6e0](feab6e0)) * Improve agent-facing `grep` descriptions and no-match hints to steer regex-looking patterns toward literal searches, route slash-containing sandbox include-globs correctly, and shorten default search timeouts so bad patterns and huge trees return guidance faster. ([#4168](#4168)) ([b1dbf5e](b1dbf5e)) * Align sandbox delete behavior with other backends by returning not-found errors for missing paths, and avoid over-blocking unrelated sibling deletes when deny rules use glob patterns. ([#4321](#4321)) ([d77496b](d77496b)) * Improve rubric grader failure diagnostics with configured model, structured-output strategy, and integer HTTP status when available. ([#4938](#4938)) ([#4967](#4967)) ([f51d3a0](f51d3a0)) ([bca70aa](bca70aa)) * Emit `max_iterations_reached` as the terminal `RubricMiddleware` status when the iteration cap is exhausted, instead of a final `needs_revision` event that will not loop. ([#4406](#4406)) ([a51c8d2](a51c8d2)) * Handle missing async subagent URLs consistently in `check_async_task` and `cancel_async_task`. ([#3967](#3967)) ([b0d92c0](b0d92c0)) ### Performance Improvements * Run LangSmith sandbox commands over the async client. ([#5061](#5061)) ([0d08747](0d08747)) --- _End release notes preview._ --- > [!NOTE] > A **New Contributors** section is appended to the GitHub release notes automatically at publish time (see [Release Pipeline](https://github.com/langchain-ai/deepagents/blob/main/.github/RELEASING.md#release-pipeline), step 2). --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Mason Daugherty <github@mdrxy.com>

… `sandbox.execute` (langchain-ai#4230)

> [!CAUTION] > Merging this PR will automatically publish to **PyPI** and create a **GitHub release**. For the full release process, see [`.github/RELEASING.md`](https://github.com/langchain-ai/deepagents/blob/main/.github/RELEASING.md). --- _Release notes preview: keep this section in sync with the package `CHANGELOG.md`. Publish reads the merged CHANGELOG via `release.yml`, not this PR description — keep them aligned anyway so the PR stays an accurate historical record for reviewers and anyone returning later._ --- ## [0.7.0](langchain-ai/deepagents@deepagents==0.6.12...deepagents==0.7.0) (2026-07-29) See [the docs](https://docs.langchain.com/oss/python/releases/changelog#deepagents-v0-7-0) for curated release notes. ### ⚠ BREAKING CHANGES * `create_deep_agent` no longer includes `TodoListMiddleware` by default, the `write_todos` tool, `todos` state channel, and todo-planning prompt are now absent. Pass `middleware=[TodoListMiddleware()]` to restore them on the main agent; add it to each `SubAgent`'s middleware to restore them there. ([langchain-ai#4929](langchain-ai#4929)) ([9340518](langchain-ai@9340518)) * Default agent prompts are now lean: the authored base prompt is empty, and tool-usage prose that duplicates tool schemas is trimmed. `BASE_AGENT_PROMPT` is deprecated (removal in `deepagents==0.9.0`) but remains importable and still returns the previous authored prompt verbatim; pass it as `create_deep_agent(system_prompt=BASE_AGENT_PROMPT)` to restore the old behavior. ([langchain-ai#4859](langchain-ai#4859)) ([langchain-ai#4979](langchain-ai#4979)) ([a8d1b32](langchain-ai@a8d1b32)) ([d9f54fc](langchain-ai@d9f54fc)) * The built-in tool-usage prompt constants `TASK_SYSTEM_PROMPT`, `ASYNC_TASK_SYSTEM_PROMPT`, `SUMMARIZATION_SYSTEM_PROMPT`, `FILESYSTEM_SYSTEM_PROMPT`, and `EXECUTION_SYSTEM_PROMPT` are removed, and the `system_prompt` default on `SubAgentMiddleware`, `AsyncSubAgentMiddleware`, `SummarizationToolMiddleware`, and `create_summarization_tool_middleware` is now `None`, which injects no prose. Pass your own string to restore prompt text. ([langchain-ai#4859](langchain-ai#4859)) ([a8d1b32](langchain-ai@a8d1b32)) * `FilesystemBackend` and `LocalShellBackend` now default to `virtual_mode=True`. Filesystem paths are anchored under `root_dir`, `..` traversal is rejected, and paths resolving outside `root_dir` raise `ValueError`. Previously an unspecified `virtual_mode` emitted a deprecation warning and fell back to `False`, where absolute host paths were used as-is and `..` could escape `root_dir`. Pass `virtual_mode=False` explicitly to restore the old filesystem behavior. ([langchain-ai#4541](langchain-ai#4541)) ([540a0fa](langchain-ai@540a0fa)) * Agents now see a destructive, recursive `delete` filesystem tool whenever the backend supports it, and filesystem permissions classify `delete` as a write operation — so an existing rule allowing writes to a path also authorizes recursively deleting that subtree unless a narrower deny or interrupt rule covers the target. Because recursive deletes affect descendants, deny and interrupt checks use bulk path overlap instead of exact-path matching. To keep the previous behavior, add a deny or interrupt rule, or omit `delete` from `FilesystemMiddleware(tools=...)`. Missing paths return a not-found error, `CompositeBackend` reports an unsupported-operation error when a routed sub-backend cannot delete, and the tool is hidden from the model entirely when the backend itself does not implement it. ([langchain-ai#3659](langchain-ai#3659)) ([langchain-ai#3691](langchain-ai#3691)) ([langchain-ai#3765](langchain-ai#3765)) ([langchain-ai#3851](langchain-ai#3851)) ([f2a21ec](langchain-ai@f2a21ec)) * `write_file` can now create a file if it is missing and replaces it entirely if it already exists, instead of returning a file-exists error. The `write_file` tool description no longer requires reading the file first. There is no "create-only" compatibility mode. Workflows, prompts, tests, or guardrails that relied on the file-exists error to force `edit_file` usage or to protect existing content must omit `write_file`, add explicit permission or interrupt rules, or use `edit_file` where preserving existing content matters. ([langchain-ai#4109](langchain-ai#4109)) ([2506fcc](langchain-ai@2506fcc)) * Removed deprecated backend compatibility shims. Callers must pass concrete `BackendProtocol` instances (not factories), configure `StoreBackend` with an explicit `namespace`, and use the current `ls` / `glob` / `grep` / `ReadResult` APIs. ([langchain-ai#4541](langchain-ai#4541)) ([540a0fa](langchain-ai@540a0fa)) * The deprecated `files_update` attribute and constructor keyword are removed from `WriteResult` and `EditResult`. Custom backends must stop passing `files_update=`, and callers must stop reading `result.files_update`; state writes are emitted directly by `StateBackend`. ([langchain-ai#4541](langchain-ai#4541)) ([540a0fa](langchain-ai@540a0fa)) * Removed the deprecated `BackendProtocol` methods `ls_info`, `als_info`, `glob_info`, `aglob_info`, `grep_raw`, and `agrep_raw`. Use `ls` / `glob` / `grep` and their async counterparts. ([langchain-ai#4541](langchain-ai#4541)) ([540a0fa](langchain-ai@540a0fa)) * `SummarizationMiddleware(history_path_prefix=...)` was removed and now raises `TypeError`. Configure `CompositeBackend(artifacts_root=...)` instead. ([langchain-ai#4541](langchain-ai#4541)) ([540a0fa](langchain-ai@540a0fa)) * Agent-facing `ls` and `glob` tool output now renders empty results as `No files found` instead of `[]`; direct backend APIs continue to return structured empty `LsResult` and `GlobResult` values. Callers that parse tool output should update those checks. ([langchain-ai#3709](langchain-ai#3709)) ([efafd1e](langchain-ai@efafd1e)) * `read_file` no longer renders raw text with a fixed-width `cat -n`-style line-number gutter and tab separator. Line and continuation markers are dynamically aligned and separated from source content by two spaces, and the `LINE_NUMBER_WIDTH` constant is removed from `deepagents.backends.utils` and `deepagents.middleware.filesystem`. Callers that parse raw tool output should update those parsers. ([langchain-ai#4561](langchain-ai#4561)) ([cf057b4](langchain-ai@cf057b4)) ### Features * Custom middleware passed to `create_deep_agent(..., middleware=[...])` can replace a default middleware instance when `.name` matches, so defaults such as `SummarizationMiddleware` can be overridden without also excluding the built-in instance. ([langchain-ai#4251](langchain-ai#4251)) ([90c8472](langchain-ai@90c8472)) * `FilesystemMiddleware(tools=[...])` accepts a keyword-only allowlist of built-in filesystem tools, typed by the newly exported `FsToolName` literal (`"ls"`, `"read_file"`, `"write_file"`, `"edit_file"`, `"delete"`, `"glob"`, `"grep"`, `"execute"`); pass `"all"` or omit the argument to keep every tool. A list must include `"read_file"` or the constructor raises `ValueError`. Omitted built-in tools are non-executable, and custom user tools are unaffected. ([langchain-ai#4325](langchain-ai#4325)) ([langchain-ai#4698](langchain-ai#4698)) ([704a70d](langchain-ai@704a70d)) ([9709525](langchain-ai@9709525)) * Shorten LLM-facing descriptions for the `task` tool and filesystem tools (`read_file`, `grep`, `edit_file`, `glob`, `execute`). ([langchain-ai#5009](langchain-ai#5009)) ([761f5f0](langchain-ai@761f5f0)) * `GrepResult` and `GlobResult` now carry a `truncated` flag so supporting backends can return valid partial results when a match cap or backend deadline is reached; agent-facing tool output adds a note telling the model to narrow the search. `FilesystemBackend` returns partial `grep` and `glob` results on its backend timeout rather than erroring, while other backend or middleware timeouts may still return errors. Its `glob` also gains brace expansion such as `*.{py,md}` (already supported by the state and store backends). ([langchain-ai#4063](langchain-ai#4063)) ([ef591e7](langchain-ai@ef591e7)) * The agent-facing `grep` match cap is configurable: `FilesystemMiddleware(grep_max_count=...)` sets the default (`1000`; `None` disables it) and the model can override it per call through the tool's new `max_count` argument. `grep` / `agrep` on `BackendProtocol` and all built-in backends accept a keyword-only `max_count`. Local ripgrep output is streamed and terminated once the cap is reached. Direct `FilesystemBackend.grep()` callers can request surrounding lines with keyword-only `context_lines`. ([langchain-ai#4570](langchain-ai#4570)) ([langchain-ai#4706](langchain-ai#4706)) ([8e86f5e](langchain-ai@8e86f5e)) ([65230df](langchain-ai@65230df)) * Paginated built-in `read_file` responses report the returned source-line range and next `offset`; total and remaining line counts are included when the backend knows the file length. Resume offsets remain safe when sandbox or middleware limits shorten the visible page. ([langchain-ai#4540](langchain-ai#4540)) ([8321194](langchain-ai@8321194)) * Optional video frame extraction for `read_file`, enabled by the new `deepagents[video]` extra. Video files are sampled into JPEG frames, with `offset` and `limit` interpreted as seconds. Without the extra, existing generic video/file content-block behavior remains. ([langchain-ai#4094](langchain-ai#4094)) ([b927147](langchain-ai@b927147)) * `FilesystemMiddleware` can capture oversized `execute` tool output directly inside the sandbox artifact path on compatible, opted-in `BaseSandbox` implementations to reduce round trips; `LangSmithSandbox` opts in by default. ([langchain-ai#4230](langchain-ai#4230)) ([02f5bd7](langchain-ai@02f5bd7)) * Automatically enable Fireworks prompt-cache session affinity when a compatible `langchain-fireworks` installation is available. ([langchain-ai#4598](langchain-ai#4598)) ([5d878bf](langchain-ai@5d878bf)) * Add a built-in NVIDIA Nemotron 3 Ultra harness profile and NVIDIA NIM app-origin attribution. ([langchain-ai#4192](langchain-ai#4192)) ([langchain-ai#4455](langchain-ai#4455)) ([d5a60ec](langchain-ai@d5a60ec)) ([4cb4749](langchain-ai@4cb4749)) * `RubricMiddleware` now accepts any positive `max_iterations` cap instead of enforcing a hard upper bound. ([langchain-ai#4405](langchain-ai#4405)) ([d6692a7](langchain-ai@d6692a7)) ### Bug Fixes * Keep fields marked with `PrivateStateAttr`, including fields declared through `create_deep_agent(state_schema=...)`, out of subagent inputs and returned parent-state updates. ([langchain-ai#4587](langchain-ai#4587)) ([a4662c0](langchain-ai@a4662c0)) * Preserve `ContextT` through the `create_deep_agent(..., middleware=[...])` type annotation so type checkers accept context-aware middleware when a matching `context_schema` is passed. ([langchain-ai#4055](langchain-ai#4055)) ([7be76c7](langchain-ai@7be76c7)) * Accept YAML list values as well as comma-separated strings for skill `allowed-tools` frontmatter, and make skill truncation warnings actionable with field name, path, length, configured limit, and impact. ([langchain-ai#4140](langchain-ai#4140)) ([langchain-ai#4141](langchain-ai#4141)) ([d62534c](langchain-ai@d62534c)) ([2f5f5b8](langchain-ai@2f5f5b8)) * Align filesystem instructions with the tools that remain after allowlist and backend-capability filtering, so agents no longer reference hidden `grep`/`glob` tools or prohibit equivalent shell search when dedicated search tools are unavailable. ([langchain-ai#4920](langchain-ai#4920)) ([langchain-ai#4921](langchain-ai#4921)) ([d3650c7](langchain-ai@d3650c7)) ([b65cc00](langchain-ai@b65cc00)) * Propagate default-backend failures from `CompositeBackend.ls("/")` and `CompositeBackend.als("/")` instead of returning successful route-only listings. ([langchain-ai#4925](langchain-ai#4925)) ([4c3b166](langchain-ai@4c3b166)) * Correct `CompositeBackend.glob` / `CompositeBackend.aglob` routing so explicit default-backend paths such as `/tools` do not also return files from routed backends such as `/memories`. ([langchain-ai#4531](langchain-ai#4531)) ([cbdb0a7](langchain-ai@cbdb0a7)) * Propagate default- and routed-backend failures from root `CompositeBackend.glob(..., path=None)` / `aglob(..., path=None)` and `path="/"` searches instead of returning incomplete successful results. ([langchain-ai#4063](langchain-ai#4063)) ([ef591e7](langchain-ai@ef591e7)) * Constrain sandbox `glob` and slash-pattern `grep` searches to their declared search root by treating leading `/` as search-root-relative, rejecting `..` traversal segments, and filtering symlink-resolved matches outside the root. ([langchain-ai#4588](langchain-ai#4588)) ([c6c7213](langchain-ai@c6c7213)) * Unify `grep(..., glob=...)` include-glob semantics across filesystem and in-memory backends: basename patterns like `*.py` match at any depth, and slash-containing patterns like `src/**/*.py` match relative paths consistently. ([langchain-ai#3936](langchain-ai#3936)) ([feab6e0](langchain-ai@feab6e0)) * Improve agent-facing `grep` descriptions and no-match hints to steer regex-looking patterns toward literal searches, route slash-containing sandbox include-globs correctly, and shorten default search timeouts so bad patterns and huge trees return guidance faster. ([langchain-ai#4168](langchain-ai#4168)) ([b1dbf5e](langchain-ai@b1dbf5e)) * Align sandbox delete behavior with other backends by returning not-found errors for missing paths, and avoid over-blocking unrelated sibling deletes when deny rules use glob patterns. ([langchain-ai#4321](langchain-ai#4321)) ([d77496b](langchain-ai@d77496b)) * Improve rubric grader failure diagnostics with configured model, structured-output strategy, and integer HTTP status when available. ([langchain-ai#4938](langchain-ai#4938)) ([langchain-ai#4967](langchain-ai#4967)) ([f51d3a0](langchain-ai@f51d3a0)) ([bca70aa](langchain-ai@bca70aa)) * Emit `max_iterations_reached` as the terminal `RubricMiddleware` status when the iteration cap is exhausted, instead of a final `needs_revision` event that will not loop. ([langchain-ai#4406](langchain-ai#4406)) ([a51c8d2](langchain-ai@a51c8d2)) * Handle missing async subagent URLs consistently in `check_async_task` and `cancel_async_task`. ([langchain-ai#3967](langchain-ai#3967)) ([b0d92c0](langchain-ai@b0d92c0)) ### Performance Improvements * Run LangSmith sandbox commands over the async client. ([langchain-ai#5061](langchain-ai#5061)) ([0d08747](langchain-ai@0d08747)) --- _End release notes preview._ --- > [!NOTE] > A **New Contributors** section is appended to the GitHub release notes automatically at publish time (see [Release Pipeline](https://github.com/langchain-ai/deepagents/blob/main/.github/RELEASING.md#release-pipeline), step 2). --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Mason Daugherty <github@mdrxy.com>

> [!CAUTION] > Merging this PR will automatically publish to **PyPI** and create a **GitHub release**. For the full release process, see [`.github/RELEASING.md`](https://github.com/langchain-ai/deepagents/blob/main/.github/RELEASING.md). --- _Release notes preview: keep this section in sync with the package `CHANGELOG.md`. Publish reads the merged CHANGELOG via `release.yml`, not this PR description — keep them aligned anyway so the PR stays an accurate historical record for reviewers and anyone returning later._ --- ## [0.7.0](deepagents==0.6.12...deepagents==0.7.0) (2026-07-29) See [the docs](https://docs.langchain.com/oss/python/releases/changelog#deepagents-v0-7-0) for curated release notes. ### ⚠ BREAKING CHANGES * `create_deep_agent` no longer includes `TodoListMiddleware` by default, the `write_todos` tool, `todos` state channel, and todo-planning prompt are now absent. Pass `middleware=[TodoListMiddleware()]` to restore them on the main agent; add it to each `SubAgent`'s middleware to restore them there. ([#4929](#4929)) ([9340518](9340518)) * Default agent prompts are now lean: the authored base prompt is empty, and tool-usage prose that duplicates tool schemas is trimmed. `BASE_AGENT_PROMPT` is deprecated (removal in `deepagents==0.9.0`) but remains importable and still returns the previous authored prompt verbatim; pass it as `create_deep_agent(system_prompt=BASE_AGENT_PROMPT)` to restore the old behavior. ([#4859](#4859)) ([#4979](#4979)) ([a8d1b32](a8d1b32)) ([d9f54fc](d9f54fc)) * The built-in tool-usage prompt constants `TASK_SYSTEM_PROMPT`, `ASYNC_TASK_SYSTEM_PROMPT`, `SUMMARIZATION_SYSTEM_PROMPT`, `FILESYSTEM_SYSTEM_PROMPT`, and `EXECUTION_SYSTEM_PROMPT` are removed, and the `system_prompt` default on `SubAgentMiddleware`, `AsyncSubAgentMiddleware`, `SummarizationToolMiddleware`, and `create_summarization_tool_middleware` is now `None`, which injects no prose. Pass your own string to restore prompt text. ([#4859](#4859)) ([a8d1b32](a8d1b32)) * `FilesystemBackend` and `LocalShellBackend` now default to `virtual_mode=True`. Filesystem paths are anchored under `root_dir`, `..` traversal is rejected, and paths resolving outside `root_dir` raise `ValueError`. Previously an unspecified `virtual_mode` emitted a deprecation warning and fell back to `False`, where absolute host paths were used as-is and `..` could escape `root_dir`. Pass `virtual_mode=False` explicitly to restore the old filesystem behavior. ([#4541](#4541)) ([540a0fa](540a0fa)) * Agents now see a destructive, recursive `delete` filesystem tool whenever the backend supports it, and filesystem permissions classify `delete` as a write operation — so an existing rule allowing writes to a path also authorizes recursively deleting that subtree unless a narrower deny or interrupt rule covers the target. Because recursive deletes affect descendants, deny and interrupt checks use bulk path overlap instead of exact-path matching. To keep the previous behavior, add a deny or interrupt rule, or omit `delete` from `FilesystemMiddleware(tools=...)`. Missing paths return a not-found error, `CompositeBackend` reports an unsupported-operation error when a routed sub-backend cannot delete, and the tool is hidden from the model entirely when the backend itself does not implement it. ([#3659](#3659)) ([#3691](#3691)) ([#3765](#3765)) ([#3851](#3851)) ([f2a21ec](f2a21ec)) * `write_file` can now create a file if it is missing and replaces it entirely if it already exists, instead of returning a file-exists error. The `write_file` tool description no longer requires reading the file first. There is no "create-only" compatibility mode. Workflows, prompts, tests, or guardrails that relied on the file-exists error to force `edit_file` usage or to protect existing content must omit `write_file`, add explicit permission or interrupt rules, or use `edit_file` where preserving existing content matters. ([#4109](#4109)) ([2506fcc](2506fcc)) * Removed deprecated backend compatibility shims. Callers must pass concrete `BackendProtocol` instances (not factories), configure `StoreBackend` with an explicit `namespace`, and use the current `ls` / `glob` / `grep` / `ReadResult` APIs. ([#4541](#4541)) ([540a0fa](540a0fa)) * The deprecated `files_update` attribute and constructor keyword are removed from `WriteResult` and `EditResult`. Custom backends must stop passing `files_update=`, and callers must stop reading `result.files_update`; state writes are emitted directly by `StateBackend`. ([#4541](#4541)) ([540a0fa](540a0fa)) * Removed the deprecated `BackendProtocol` methods `ls_info`, `als_info`, `glob_info`, `aglob_info`, `grep_raw`, and `agrep_raw`. Use `ls` / `glob` / `grep` and their async counterparts. ([#4541](#4541)) ([540a0fa](540a0fa)) * `SummarizationMiddleware(history_path_prefix=...)` was removed and now raises `TypeError`. Configure `CompositeBackend(artifacts_root=...)` instead. ([#4541](#4541)) ([540a0fa](540a0fa)) * Agent-facing `ls` and `glob` tool output now renders empty results as `No files found` instead of `[]`; direct backend APIs continue to return structured empty `LsResult` and `GlobResult` values. Callers that parse tool output should update those checks. ([#3709](#3709)) ([efafd1e](efafd1e)) * `read_file` no longer renders raw text with a fixed-width `cat -n`-style line-number gutter and tab separator. Line and continuation markers are dynamically aligned and separated from source content by two spaces, and the `LINE_NUMBER_WIDTH` constant is removed from `deepagents.backends.utils` and `deepagents.middleware.filesystem`. Callers that parse raw tool output should update those parsers. ([#4561](#4561)) ([cf057b4](cf057b4)) ### Features * Custom middleware passed to `create_deep_agent(..., middleware=[...])` can replace a default middleware instance when `.name` matches, so defaults such as `SummarizationMiddleware` can be overridden without also excluding the built-in instance. ([#4251](#4251)) ([90c8472](90c8472)) * `FilesystemMiddleware(tools=[...])` accepts a keyword-only allowlist of built-in filesystem tools, typed by the newly exported `FsToolName` literal (`"ls"`, `"read_file"`, `"write_file"`, `"edit_file"`, `"delete"`, `"glob"`, `"grep"`, `"execute"`); pass `"all"` or omit the argument to keep every tool. A list must include `"read_file"` or the constructor raises `ValueError`. Omitted built-in tools are non-executable, and custom user tools are unaffected. ([#4325](#4325)) ([#4698](#4698)) ([704a70d](704a70d)) ([9709525](9709525)) * Shorten LLM-facing descriptions for the `task` tool and filesystem tools (`read_file`, `grep`, `edit_file`, `glob`, `execute`). ([#5009](#5009)) ([761f5f0](761f5f0)) * `GrepResult` and `GlobResult` now carry a `truncated` flag so supporting backends can return valid partial results when a match cap or backend deadline is reached; agent-facing tool output adds a note telling the model to narrow the search. `FilesystemBackend` returns partial `grep` and `glob` results on its backend timeout rather than erroring, while other backend or middleware timeouts may still return errors. Its `glob` also gains brace expansion such as `*.{py,md}` (already supported by the state and store backends). ([#4063](#4063)) ([ef591e7](ef591e7)) * The agent-facing `grep` match cap is configurable: `FilesystemMiddleware(grep_max_count=...)` sets the default (`1000`; `None` disables it) and the model can override it per call through the tool's new `max_count` argument. `grep` / `agrep` on `BackendProtocol` and all built-in backends accept a keyword-only `max_count`. Local ripgrep output is streamed and terminated once the cap is reached. Direct `FilesystemBackend.grep()` callers can request surrounding lines with keyword-only `context_lines`. ([#4570](#4570)) ([#4706](#4706)) ([8e86f5e](8e86f5e)) ([65230df](65230df)) * Paginated built-in `read_file` responses report the returned source-line range and next `offset`; total and remaining line counts are included when the backend knows the file length. Resume offsets remain safe when sandbox or middleware limits shorten the visible page. ([#4540](#4540)) ([8321194](8321194)) * Optional video frame extraction for `read_file`, enabled by the new `deepagents[video]` extra. Video files are sampled into JPEG frames, with `offset` and `limit` interpreted as seconds. Without the extra, existing generic video/file content-block behavior remains. ([#4094](#4094)) ([b927147](b927147)) * `FilesystemMiddleware` can capture oversized `execute` tool output directly inside the sandbox artifact path on compatible, opted-in `BaseSandbox` implementations to reduce round trips; `LangSmithSandbox` opts in by default. ([#4230](#4230)) ([02f5bd7](02f5bd7)) * Automatically enable Fireworks prompt-cache session affinity when a compatible `langchain-fireworks` installation is available. ([#4598](#4598)) ([5d878bf](5d878bf)) * Add a built-in NVIDIA Nemotron 3 Ultra harness profile and NVIDIA NIM app-origin attribution. ([#4192](#4192)) ([#4455](#4455)) ([d5a60ec](d5a60ec)) ([4cb4749](4cb4749)) * `RubricMiddleware` now accepts any positive `max_iterations` cap instead of enforcing a hard upper bound. ([#4405](#4405)) ([d6692a7](d6692a7)) ### Bug Fixes * Keep fields marked with `PrivateStateAttr`, including fields declared through `create_deep_agent(state_schema=...)`, out of subagent inputs and returned parent-state updates. ([#4587](#4587)) ([a4662c0](a4662c0)) * Preserve `ContextT` through the `create_deep_agent(..., middleware=[...])` type annotation so type checkers accept context-aware middleware when a matching `context_schema` is passed. ([#4055](#4055)) ([7be76c7](7be76c7)) * Accept YAML list values as well as comma-separated strings for skill `allowed-tools` frontmatter, and make skill truncation warnings actionable with field name, path, length, configured limit, and impact. ([#4140](#4140)) ([#4141](#4141)) ([d62534c](d62534c)) ([2f5f5b8](2f5f5b8)) * Align filesystem instructions with the tools that remain after allowlist and backend-capability filtering, so agents no longer reference hidden `grep`/`glob` tools or prohibit equivalent shell search when dedicated search tools are unavailable. ([#4920](#4920)) ([#4921](#4921)) ([d3650c7](d3650c7)) ([b65cc00](b65cc00)) * Propagate default-backend failures from `CompositeBackend.ls("/")` and `CompositeBackend.als("/")` instead of returning successful route-only listings. ([#4925](#4925)) ([4c3b166](4c3b166)) * Correct `CompositeBackend.glob` / `CompositeBackend.aglob` routing so explicit default-backend paths such as `/tools` do not also return files from routed backends such as `/memories`. ([#4531](#4531)) ([cbdb0a7](cbdb0a7)) * Propagate default- and routed-backend failures from root `CompositeBackend.glob(..., path=None)` / `aglob(..., path=None)` and `path="/"` searches instead of returning incomplete successful results. ([#4063](#4063)) ([ef591e7](ef591e7)) * Constrain sandbox `glob` and slash-pattern `grep` searches to their declared search root by treating leading `/` as search-root-relative, rejecting `..` traversal segments, and filtering symlink-resolved matches outside the root. ([#4588](#4588)) ([c6c7213](c6c7213)) * Unify `grep(..., glob=...)` include-glob semantics across filesystem and in-memory backends: basename patterns like `*.py` match at any depth, and slash-containing patterns like `src/**/*.py` match relative paths consistently. ([#3936](#3936)) ([feab6e0](feab6e0)) * Improve agent-facing `grep` descriptions and no-match hints to steer regex-looking patterns toward literal searches, route slash-containing sandbox include-globs correctly, and shorten default search timeouts so bad patterns and huge trees return guidance faster. ([#4168](#4168)) ([b1dbf5e](b1dbf5e)) * Align sandbox delete behavior with other backends by returning not-found errors for missing paths, and avoid over-blocking unrelated sibling deletes when deny rules use glob patterns. ([#4321](#4321)) ([d77496b](d77496b)) * Improve rubric grader failure diagnostics with configured model, structured-output strategy, and integer HTTP status when available. ([#4938](#4938)) ([#4967](#4967)) ([f51d3a0](f51d3a0)) ([bca70aa](bca70aa)) * Emit `max_iterations_reached` as the terminal `RubricMiddleware` status when the iteration cap is exhausted, instead of a final `needs_revision` event that will not loop. ([#4406](#4406)) ([a51c8d2](a51c8d2)) * Handle missing async subagent URLs consistently in `check_async_task` and `cancel_async_task`. ([#3967](#3967)) ([b0d92c0](b0d92c0)) ### Performance Improvements * Run LangSmith sandbox commands over the async client. ([#5061](#5061)) ([0d08747](0d08747)) --- _End release notes preview._ --- > [!NOTE] > A **New Contributors** section is appended to the GitHub release notes automatically at publish time (see [Release Pipeline](https://github.com/langchain-ai/deepagents/blob/main/.github/RELEASING.md#release-pipeline), step 2). --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Mason Daugherty <github@mdrxy.com>

ccurme (ccurme) added 8 commits June 20, 2026 10:23

offload large execute output at the source

0219be0

bound capture size and signal truncation

243077b

cosmetic changes to content preview

20d1ad2

refactor offloaded bool

c18bde8

add tests

1562466

add escape hatch

34de240

add comments in capture wrapper

332902a

Merge branch 'main' into cc/sandbox_offload

41aa2a3

github-actions Bot changed the title ~~feat(deepagents): reduce round trips when offloading large tool results with sandbox.execute~~ feat(sdk): reduce round trips when offloading large tool results with sandbox.execute Jun 24, 2026

github-actions Bot added deepagents Related to the `deepagents` SDK / agent harness feature New feature/enhancement or request for one internal User is a member of the `langchain-ai` GitHub organization size: M 200-499 LOC labels Jun 24, 2026

ccurme (ccurme) commented Jun 24, 2026

View reviewed changes

open-swe Bot reviewed Jun 24, 2026

View reviewed changes

Comment thread libs/deepagents/deepagents/backends/sandbox.py Outdated

Mason Daugherty (mdrxy) changed the title ~~feat(sdk): reduce round trips when offloading large tool results with sandbox.execute~~ feat(sdk): reduce round trips when offloading large tool results with sandbox.execute Jun 25, 2026

ccurme (ccurme) added 2 commits June 25, 2026 15:53

preserve exit code when capture output hits the cap

df565b6

Merge branch 'main' into cc/sandbox_offload

6777f18

Hunter Lovell (hntrl) approved these changes Jun 25, 2026

View reviewed changes

Eugene Yurtsev (eyurtsev) approved these changes Jun 26, 2026

View reviewed changes

ccurme (ccurme) added 4 commits June 26, 2026 13:10

move offloading to execute_with_offload on sandbox

4de7ad1

make opt-in (on for langsmith)

1e9d3d6

Merge branch 'main' into cc/sandbox_offload

63813cf

cr

67f0605

Eugene Yurtsev (eyurtsev) reviewed Jun 26, 2026

View reviewed changes

Comment thread libs/deepagents/deepagents/backends/sandbox.py Outdated

Mason Daugherty (mdrxy) previously requested changes Jun 26, 2026

View reviewed changes

ccurme (ccurme) added 2 commits June 26, 2026 16:10

cr

cdc87ee

doctring

162cba5

ccurme (ccurme) merged commit 02f5bd7 into main Jun 26, 2026
62 checks passed

ccurme (ccurme) deleted the cc/sandbox_offload branch June 26, 2026 22:50

github-actions Bot mentioned this pull request Jun 26, 2026

release(deepagents): 0.7.0 #4297

Merged

ccurme (ccurme) mentioned this pull request Jun 29, 2026

chore(standard-tests): add sandbox integration test for offloading large execute results langchain-ai/langchain#38537

Merged

Marcelo5444 pushed a commit to Marcelo5444/deepagents that referenced this pull request Jul 30, 2026

feat(sdk): reduce round trips when offloading large tool results with…

1de5cf1

… `sandbox.execute` (langchain-ai#4230)

Navigation Menu

Uh oh!

feat(sdk): reduce round trips when offloading large tool results with `sandbox.execute` - #4230

feat(sdk): reduce round trips when offloading large tool results with `sandbox.execute`#4230
ccurme (ccurme) merged 16 commits into
mainfrom
cc/sandbox_offload

ccurme (ccurme) commented Jun 24, 2026 •

edited

Loading

Uh oh!

ccurme (ccurme) Jun 24, 2026

Uh oh!

open-swe Bot left a comment

Uh oh!

Uh oh!

Eugene Yurtsev (eyurtsev) left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Mason Daugherty (mdrxy) left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

ccurme (ccurme) commented Jun 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ccurme (ccurme) Jun 24, 2026

Choose a reason for hiding this comment

Uh oh!

open-swe Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Eugene Yurtsev (eyurtsev) left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Mason Daugherty (mdrxy) left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

ccurme (ccurme) commented Jun 24, 2026 •

edited

Loading