fix(backend): prevent graph execution stuck + steer SDK away from bash_exec by majdyz · Pull Request #12548 · Significant-Gravitas/AutoGPT

majdyz · 2026-03-25T05:52:31Z

Summary

Two backend fixes for CoPilot stability:

Steer model away from bash_exec for SDK tool-result files — When the SDK returns tool results as file paths, the copilot model was attempting to use bash_exec to read them instead of treating the content directly. Added system prompt guidance to prevent this.
Guard against missing 'name' in execution input_data — GraphExecution.from_db() assumed all INPUT/OUTPUT block node executions have a name field in input_data. This crashes with KeyError: 'name' when non-standard blocks (e.g., OrchestratorBlock) produce node executions without this field. Added "name" in exec.input_data guards.

Why

The bash_exec issue causes copilot to fail when processing SDK tool outputs
The KeyError crashes the update_graph_execution_stats endpoint, causing graph executions to appear stuck (retries 35+ times, never completes)

How

Added system prompt instruction to treat tool result file contents directly
Added "name" in exec.input_data guard in both input extraction (line 340) and output extraction (line 365) in execution.py

Changes

backend/copilot/sdk/service.py — system prompt guidance
backend/data/execution.py — KeyError guard for missing name field

Checklist 📋

I have clearly listed my changes in the PR description
I have made a test plan
I have tested my changes according to the test plan

Test plan:

OrchestratorBlock graph execution no longer gets stuck
Standard Agent Input/Output blocks still work correctly
Copilot SDK tool results are processed without bash_exec

…sult files The model was using bash_exec (cd + cat) to read SDK tool-result files under ~/.claude/projects/.../tool-results/, which fails in E2B because the sandbox runs as /home/user and cannot access /root/. The Read tool handles these correctly by reading from the host filesystem. Changes: - Add "CANNOT read SDK tool-result files" warning to bash_exec description - Add tool-results mention to the Read tool description - Fix system prompt: replace stale "read_file" reference with "Read", explicitly warn that bash_exec cannot access host-side tool-results

coderabbitai · 2026-03-25T05:52:49Z

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

@coderabbitai resume to resume automatic reviews.
@coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

▶️ Resume reviews
🔍 Trigger review

Walkthrough

Updated copilot prompting and tool descriptions to require using the Read tool (not bash_exec or read_workspace_file) for SDK tool-result files under ~/.claude/projects/.../tool-results/, clarified those files are on the host filesystem (inaccessible to sandboxed bash_exec), and made a defensive fix in GraphExecution.from_db to only use exec.input_data["name"] when "name" exists.

Changes

Cohort / File(s)	Summary
Prompt & storage guidance `autogpt_platform/backend/backend/copilot/prompting.py`	Narrowed and corrected `_build_storage_supplement`: mandates using `Read` only for SDK tool-result files, disallows `bash_exec`/`read_workspace_file`, and clarifies files live on host FS under `~/.claude/projects/.../tool-results/` (not accessible from sandboxed `bash_exec`).
Read tool description `autogpt_platform/backend/backend/copilot/sdk/tool_adapter.py`	Appended explicit instruction to `_READ_TOOL_DESCRIPTION` to use `Read` (not `bash_exec`) for SDK tool-result files under `~/.claude/projects/.../tool-results/`.
Bash exec description `autogpt_platform/backend/backend/copilot/tools/bash_exec.py`	Updated `BashExecTool.description` to state it CANNOT read SDK tool-result files under `~/.claude/projects/...` and to direct use of `Read`; no execution logic changed.
Execution data defensive fix `autogpt_platform/backend/backend/data/execution.py`	Made `GraphExecution.from_db` defensive: when building `inputs` and `outputs` from completed `INPUT`/`OUTPUT` block executions, only access `exec.input_data["name"]` if `"name"` exists, avoiding unsafe indexing.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related PRs

refactor(backend/copilot): centralize prompt building in prompting.py #12324: Edits the same prompting/tool-supplement areas and tool descriptions for SDK tool-result handling.
fix(backend/copilot): fix tool-result file read failures across turns #12399: Similar changes to copilot tool descriptions enforcing Read for SDK tool-result files.
feat(copilot): persist large tool outputs to workspace with retrieval instructions #12279: Adds workspace persistence and read_workspace_file support; related because it touches which tools read persisted outputs.

Suggested labels

size/s, Review effort 3/5

Suggested reviewers

0ubbe
Swiftyos
Bentlybro

Poem

🐰 I hopped through prompts and filesystem trails,
Told bash to step back from host-side tales.
"Use Read," I hummed, nose twitching with glee,
Guarding tool-results where humans can't see —
A little rabbit's helpful decree.

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 33.33% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and specifically describes the two main fixes: preventing graph execution from getting stuck and steering the model away from bash_exec for SDK tool-result files.
Description check	✅ Passed	The description comprehensively details both fixes with clear sections on summary, rationale, implementation, and test plans that directly relate to the changeset.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Commit unit tests in branch fix/copilot-tool-results-guidance

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

majdyz

🤖 Self-review Summary

All three changes are consistent and reinforce each other — system prompt, Read tool description, and bash_exec description all point the model in the right direction.

No blockers or should-fix issues found.

🟡 Nice to Have (pre-existing, not in diff): Line 187 of prompting.py says "SDK file tools (Read/Write/Edit/Glob/Grep) and bash_exec share one filesystem" — slightly misleading now since Read can also access host-side tool-result files that bash_exec cannot. Could add a qualifier in a follow-up.

Address self-review: the "Shell & filesystem" section listed Read alongside bash_exec as sharing one filesystem, which is misleading now that Read also accesses host-side SDK tool-result files.

Read shares the sandbox filesystem for regular files; it just additionally reads host-side tool-results. Removing it from the list was misleading. The tool-results section below already covers the distinction.

Prevents KeyError when node executions from non-standard blocks (e.g., OrchestratorBlock) don't include 'name' in input_data. The from_db fallback assumes all INPUT/OUTPUT block executions have a 'name' field, but this isn't guaranteed for all block types.

…m prompt only Remove redundant tool-results mentions from Read and bash_exec tool descriptions — the system prompt storage supplement already covers this clearly. Avoids wasting tokens on triple-repeated instructions.

Revert the early-continue refactor and logger.debug — debug logs aren't read in production. Keep the simple inline guard style from the original PR #12548.

majdyz requested a review from a team as a code owner March 25, 2026 05:52

majdyz requested review from 0ubbe and Swiftyos and removed request for a team March 25, 2026 05:52

github-project-automation Bot added this to AutoGPT development kanban Mar 25, 2026

github-project-automation Bot moved this to 🆕 Needs initial review in AutoGPT development kanban Mar 25, 2026

github-actions Bot added platform/backend AutoGPT Platform - Back end size/m labels Mar 25, 2026

majdyz commented Mar 25, 2026

View reviewed changes

majdyz added 3 commits March 25, 2026 12:53

fix(backend/copilot): clarify Read tool's dual access in filesystem note

d0f558d

Address self-review: the "Shell & filesystem" section listed Read alongside bash_exec as sharing one filesystem, which is misleading now that Read also accesses host-side SDK tool-result files.

fix(backend/copilot): revert filesystem note — Read shares sandbox too

3273b95

Read shares the sandbox filesystem for regular files; it just additionally reads host-side tool-results. Removing it from the list was misleading. The tool-results section below already covers the distinction.

majdyz changed the title ~~fix(backend/copilot): steer model away from bash_exec for SDK tool-result files~~ fix(backend): prevent graph execution stuck + steer SDK away from bash_exec Mar 25, 2026

majdyz added 2 commits March 25, 2026 13:31

style: inline 'name' guard into existing condition

9f41880

majdyz enabled auto-merge March 25, 2026 06:39

majdyz disabled auto-merge March 25, 2026 06:58

majdyz merged commit 336114f into dev Mar 25, 2026
25 checks passed

majdyz deleted the fix/copilot-tool-results-guidance branch March 25, 2026 06:58

github-project-automation Bot moved this from 🆕 Needs initial review to ✅ Done in AutoGPT development kanban Mar 25, 2026

majdyz removed request for 0ubbe and Swiftyos March 25, 2026 06:58

majdyz commented Mar 25, 2026

View reviewed changes

Comment thread autogpt_platform/backend/backend/data/execution.py

majdyz commented Mar 25, 2026

View reviewed changes

Comment thread autogpt_platform/backend/backend/data/execution.py

majdyz mentioned this pull request Mar 25, 2026

fix(backend): add logger.debug and test for nameless node exec skipping #12552

Closed

2 tasks

This was referenced Mar 31, 2026

fix(backend/copilot): prevent duplicate block execution from pre-launch arg mismatch #12632

Merged

fix(copilot): P0 guardrails, transient retry, and security hardening #12636

Merged

This was referenced Apr 11, 2026

fix(copilot): unified MCP file tools (Read/Write/Edit) to prevent truncation data loss #12750

Merged

fix(backend/copilot): make system prompt fully static for cross-user prompt caching #12790

Merged

coderabbitai Bot mentioned this pull request Apr 20, 2026

fix(copilot): dedupe tool row, lift bash_exec timeout, Stop+resend recovery #12862

Merged

8 tasks

coderabbitai Bot mentioned this pull request May 21, 2026

fix(backend/copilot): make SDK tool-result reads navigable + redirect wrong-tool access #13179

Merged

4 tasks

Navigation Menu

Uh oh!

fix(backend): prevent graph execution stuck + steer SDK away from bash_exec - #12548

fix(backend): prevent graph execution stuck + steer SDK away from bash_exec#12548
majdyz merged 6 commits into
devfrom
fix/copilot-tool-results-guidance

majdyz commented Mar 25, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Mar 25, 2026 •

edited

Loading

Reviews paused

❌ Failed checks (1 warning)

Uh oh!

majdyz left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

majdyz commented Mar 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Why

How

Changes

Checklist 📋

Test plan:

Uh oh!

coderabbitai Bot commented Mar 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Reviews paused

Walkthrough

Changes

Estimated code review effort

Possibly related PRs

Suggested labels

Suggested reviewers

Poem

❌ Failed checks (1 warning)

Uh oh!

majdyz left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

majdyz commented Mar 25, 2026 •

edited

Loading

coderabbitai Bot commented Mar 25, 2026 •

edited

Loading