Fix context window displaying as 0 when resuming conversation #1590

xingyaoww · 2026-01-05T16:17:36Z

Summary

This fixes issue #1228 where the context_window (and other stats) were being reset to zero when sending ConversationStatusUpdateEvent (full_state) after resuming a conversation.

Root cause: In ConversationState.create(), when resuming a conversation, the stats were correctly deserialized from base_state.json but then immediately overwritten with an empty ConversationStats() object at line 216.

Changes:

Remove the stats reset line in state.py resume path (the stats are already correctly deserialized from base_state.json)
Add context_window field to TokenUsageData in integration test schemas
Add warning log when context_window is 0 despite LLM usage (to catch similar issues in the future)
Add regression test test_conversation_state_stats_preserved_on_resume for stats preservation on resume
Update existing test test_conversation_state_flags_persistence to reflect correct behavior

Fixes #1228

Checklist

If the PR is changing/adding functionality, are there tests to reflect this?
If there is an example, have you run the example to make sure that it works?
If there are instructions on how to run the code, have you followed the instructions and made sure that it works?
If the feature is significant enough to require documentation, is there a PR open on the OpenHands/docs repository with the same branch name?
Is the github CI passing?

@xingyaoww can click here to continue refining the PR

Agent Server images for this PR

• GHCR package: https://github.com/OpenHands/agent-sdk/pkgs/container/agent-server

Variants & Base Images

Variant	Architectures	Base Image	Docs / Tags
java	amd64, arm64	`eclipse-temurin:17-jdk`	Link
python	amd64, arm64	`nikolaik/python-nodejs:python3.12-nodejs22`	Link
golang	amd64, arm64	`golang:1.21-bookworm`	Link

Pull (multi-arch manifest)

# Each variant is a multi-arch manifest supporting both amd64 and arm64
docker pull ghcr.io/openhands/agent-server:70c1405-python

Run

docker run -it --rm \
  -p 8000:8000 \
  --name agent-server-70c1405-python \
  ghcr.io/openhands/agent-server:70c1405-python

All tags pushed for this build

ghcr.io/openhands/agent-server:70c1405-golang-amd64
ghcr.io/openhands/agent-server:70c1405-golang_tag_1.21-bookworm-amd64
ghcr.io/openhands/agent-server:70c1405-golang-arm64
ghcr.io/openhands/agent-server:70c1405-golang_tag_1.21-bookworm-arm64
ghcr.io/openhands/agent-server:70c1405-java-amd64
ghcr.io/openhands/agent-server:70c1405-eclipse-temurin_tag_17-jdk-amd64
ghcr.io/openhands/agent-server:70c1405-java-arm64
ghcr.io/openhands/agent-server:70c1405-eclipse-temurin_tag_17-jdk-arm64
ghcr.io/openhands/agent-server:70c1405-python-amd64
ghcr.io/openhands/agent-server:70c1405-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-amd64
ghcr.io/openhands/agent-server:70c1405-python-arm64
ghcr.io/openhands/agent-server:70c1405-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-arm64
ghcr.io/openhands/agent-server:70c1405-golang
ghcr.io/openhands/agent-server:70c1405-java
ghcr.io/openhands/agent-server:70c1405-python

About Multi-Architecture Support

Each variant tag (e.g., 70c1405-python) is a multi-arch manifest supporting both amd64 and arm64
Docker automatically pulls the correct architecture for your platform
Individual architecture tags (e.g., 70c1405-python-amd64) are also available if needed

github-actions · 2026-01-05T16:20:52Z

Coverage Report •

File	Stmts	Miss	Cover	Missing
openhands-sdk/openhands/sdk/llm
llm.py	420	157	62%	359, 364, 368, 372–373, 376, 380–381, 392–393, 395–396, 400, 417, 435–438, 485, 515–517, 538, 542, 558, 565–566, 590–591, 601, 626–631, 652–653, 656, 660, 672, 677, 679–681, 691, 699–706, 710–713, 715, 728, 732–733, 735–736, 741–742, 744, 751, 754–759, 816–821, 878–879, 882–885, 927, 944, 998, 1001, 1004–1012, 1016–1018, 1021, 1024–1026, 1033–1034, 1043, 1050–1052, 1056, 1058–1063, 1065–1082, 1085–1089, 1091–1092, 1098–1107, 1120, 1134, 1139
TOTAL	14547	6913	52%

github-actions · 2026-01-05T16:36:40Z

Hi! I started running the integration tests on your PR. You will receive a comment with the results shortly.

github-actions · 2026-01-05T16:52:45Z

🧪 Integration Tests Results

Overall Success Rate: 96.0%
Total Cost: $1.95
Models Tested: 6
Timestamp: 2026-01-05 16:52:38 UTC

📁 Detailed Logs & Artifacts

Click the links below to access detailed agent/LLM logs showing the complete reasoning process for each model. On the GitHub Actions page, scroll down to the 'Artifacts' section to download the logs.

litellm_proxy_deepseek_deepseek_chat: 📥 View & Download Logs
litellm_proxy_claude_sonnet_4_5_20250929: 📥 View & Download Logs
litellm_proxy_vertex_ai_gemini_3_pro_preview: 📥 View & Download Logs
litellm_proxy_moonshot_kimi_k2_thinking: 📥 View & Download Logs
litellm_proxy_mistral_devstral_2512: 📥 View & Download Logs
litellm_proxy_gpt_5.1_codex_max: 📥 View & Download Logs

📊 Summary

Model	Overall	Integration (Required)	Behavior (Optional)	Tests Passed	Skipped	Total	Cost	Tokens
litellm_proxy_deepseek_deepseek_chat	100.0%	100.0%	N/A	8/8	1	9	$0.06	536,384
litellm_proxy_claude_sonnet_4_5_20250929	100.0%	100.0%	N/A	9/9	0	9	$0.67	579,762
litellm_proxy_vertex_ai_gemini_3_pro_preview	100.0%	100.0%	N/A	9/9	0	9	$0.54	363,395
litellm_proxy_moonshot_kimi_k2_thinking	100.0%	100.0%	N/A	8/8	1	9	$0.28	415,593
litellm_proxy_mistral_devstral_2512	87.5%	87.5%	N/A	7/8	1	9	$0.25	604,509
litellm_proxy_gpt_5.1_codex_max	87.5%	87.5%	N/A	7/8	1	9	$0.15	211,982

📋 Detailed Results

litellm_proxy_deepseek_deepseek_chat

Overall Success Rate: 100.0% (8/8)
Integration Tests (Required): 100.0% (8/9)
Total Cost: $0.06
Token Usage: prompt: 524,021, completion: 12,363, cache_read: 500,032
Run Suffix: litellm_proxy_deepseek_deepseek_chat_cc46f95_deepseek_run_N9_20260105_163707
Skipped Tests: 1

Skipped Tests:

t08_image_file_viewing: This test requires a vision-capable LLM model. Please use a model that supports image input.

litellm_proxy_claude_sonnet_4_5_20250929

Overall Success Rate: 100.0% (9/9)
Integration Tests (Required): 100.0% (9/9)
Total Cost: $0.67
Token Usage: prompt: 567,099, completion: 12,663, cache_read: 476,176, cache_write: 89,951, reasoning: 3,523
Run Suffix: litellm_proxy_claude_sonnet_4_5_20250929_cc46f95_sonnet_run_N9_20260105_163704

litellm_proxy_vertex_ai_gemini_3_pro_preview

Overall Success Rate: 100.0% (9/9)
Integration Tests (Required): 100.0% (9/9)
Total Cost: $0.54
Token Usage: prompt: 344,824, completion: 18,571, cache_read: 208,345, reasoning: 13,349
Run Suffix: litellm_proxy_vertex_ai_gemini_3_pro_preview_cc46f95_gemini_3_pro_run_N9_20260105_163704

litellm_proxy_moonshot_kimi_k2_thinking

Overall Success Rate: 100.0% (8/8)
Integration Tests (Required): 100.0% (8/9)
Total Cost: $0.28
Token Usage: prompt: 397,283, completion: 18,310, cache_read: 329,216
Run Suffix: litellm_proxy_moonshot_kimi_k2_thinking_cc46f95_kimi_k2_run_N9_20260105_163710
Skipped Tests: 1

Skipped Tests:

t08_image_file_viewing: This test requires a vision-capable LLM model. Please use a model that supports image input.

litellm_proxy_mistral_devstral_2512

Overall Success Rate: 87.5% (7/8)
Integration Tests (Required): 87.5% (7/9)
Total Cost: $0.25
Token Usage: prompt: 598,340, completion: 6,169
Run Suffix: litellm_proxy_mistral_devstral_2512_cc46f95_devstral_2512_run_N9_20260105_163706
Skipped Tests: 1

Skipped Tests:

t08_image_file_viewing: This test requires a vision-capable LLM model. Please use a model that supports image input.

Failed Tests:

t02_add_bash_hello ⚠️ REQUIRED: Shell script is not executable (Cost: $0.0085)

litellm_proxy_gpt_5.1_codex_max

Overall Success Rate: 87.5% (7/8)
Integration Tests (Required): 87.5% (7/9)
Total Cost: $0.15
Token Usage: prompt: 208,837, completion: 3,145, cache_read: 129,408, reasoning: 1,280
Run Suffix: litellm_proxy_gpt_5.1_codex_max_cc46f95_gpt51_codex_run_N9_20260105_163702
Skipped Tests: 1

Skipped Tests:

t09_token_condenser: This test stresses long repetitive tool loops to trigger token-based condensation. GPT-5.1 Codex Max often declines such requests for efficiency/safety reasons.

Failed Tests:

t06_github_pr_browsing ⚠️ REQUIRED: Agent's final answer does not contain the expected information about the PR content. Final answer preview: I don’t have access to browse external sites like GitHub from this environment, so I can’t open that pull request directly. If you can share the key details (e.g., the PR description, discussion, or s... (Cost: $0.007)

xingyaoww · 2026-01-05T17:07:02Z

@OpenHands #1590 (comment)

Help me download the artifact and can you make sure the context window dispalying 0 is fixed?

openhands-ai · 2026-01-05T17:07:10Z

I'm on it! xingyaoww can track my progress at all-hands.dev

openhands-ai · 2026-01-05T17:13:58Z

Summary

I downloaded the integration test artifacts from PR #1590 and identified the root cause of the context_window displaying as 0 issue.

Root Cause Analysis

The integration test results showed context_window: 0 for all tests. After analyzing the code, I found that the actual bug was different from what the original PR commit addressed:

Original PR fix: Addressed stats being reset when resuming a conversation (issue Context Window Displays as 0 When Sending ConversationStatusUpdateEvent (Full State) #1228)
Additional bug found: context_window was only being passed to telemetry when log_enabled was True. Since logging is typically disabled in production/integration tests, context_window was always 0.

Fix Applied

Modified openhands-sdk/openhands/sdk/llm/llm.py to always pass context_window to telemetry regardless of whether logging is enabled:

# Before: log_ctx was None when logging disabled
log_ctx = None
if self._telemetry.log_enabled:
    log_ctx = {..., "context_window": self.max_input_tokens or 0}

# After: Always include context_window for metrics tracking
log_ctx: dict[str, Any] = {"context_window": self.max_input_tokens or 0}
if self._telemetry.log_enabled:
    log_ctx.update({...})  # Add logging-specific fields

This change was applied to both the completion() and responses() methods.

Testing

Added regression test test_context_window_tracked_without_logging_enabled to verify context_window is tracked even when logging is disabled
All 482 LLM tests pass
Pre-commit hooks pass (ruff, pyright, etc.)

Changes Pushed

The fix has been pushed to branch openhands/fix-context-window-zero-issue-1228 and will be included in PR #1590.

View full conversation

The context_window was only being passed to telemetry when log_enabled was True. This caused context_window to always be 0 in production when logging was disabled. Changes: - Always pass context_window to telemetry regardless of log_enabled - Add context_window field to integration test TokenUsageData schema Co-authored-by: openhands <openhands@all-hands.dev>

openhands-sdk/openhands/sdk/llm/llm.py

enyst · 2026-01-05T23:47:46Z

@OpenHands Do a /codereview-roasted on this PR diff. Understand the goal is actually:

Additional bug found: context_window was only being passed to telemetry when log_enabled was True. Since logging is typically disabled in production/integration tests, context_window was always 0.

Is this correct?

Then read the review threads (inline threads) and answer the question there.

openhands-ai · 2026-01-05T23:47:56Z

I'm on it! enyst can track my progress at all-hands.dev

openhands-ai · 2026-01-05T23:49:25Z

Final summary

What I did: I reviewed the PR diff for openhands/fix-context-window-zero-issue-1228, verified the stated “additional bug” against the actual code changes, pulled the inline review comment(s) via the GitHub API, and answered the inline question with a roasted-style code review + technical explanation.

Restated answer to the main question (“Is this correct?”)

Yes. The statement is correct:

context_window was only being passed to telemetry when log_enabled was True. Since logging is typically disabled in production/integration tests, context_window was always 0.

In the old code, log_ctx was None unless self._telemetry.log_enabled was true, and context_window lived inside log_ctx. So telemetry never saw it when logging was off, which led to 0 downstream. The PR fixes that by always including {"context_window": self.max_input_tokens or 0} in the telemetry context regardless of log_enabled, while still gating heavy debug payload (messages/tools/kwargs) behind log_enabled.

Answer to the inline thread question

The reviewer asked why we only pass context_window in log_ctx but not “completion_tokens etc”.

My answer: token counts aren’t sourced from log_ctx—they come from the LLM response usage / telemetry instrumentation around the response. log_ctx is request-side context (inputs/tools/kwargs). context_window is a model/config property that cannot be reliably inferred from the response usage, so it must be passed explicitly for metrics even when logging is disabled.

Checklist against your request

Did a /codereview-roasted of the PR diff (no code changes)
Confirmed whether the “additional bug” explanation is correct
Retrieved and read inline review threads
Answered the question in the inline thread (in this chat, per instructions: no PR comments)

Notes on pushing / changes

No code changes were made in this session, so:

no diff to double-check
nothing to commit
nothing to push

View full conversation

enyst · 2026-01-05T23:56:05Z

🟡 Acceptable taste — fixes the real issue with minimal churn, but the naming and data flow are still muddled.

[IMPROVEMENT OPPORTUNITIES]

1) Misleading variable name: log_ctx is no longer “log” context

You now send it even when logging is disabled.
So it’s not a “log context”, it’s a “telemetry context” or “request context”.
This is exactly how code rots: names lie, future devs “fix” the wrong thing.

Concrete suggestion: rename log_ctx → telemetry_ctx (or request_ctx) and keep the “logging extras” gated.

2) The dict update() pattern is fine, but it’s slightly backwards
Right now:

build minimal dict
mutate it conditionally

That’s okay, but it reads like “logging is special, everything else is baseline” — which is correct — but the naming (log_ctx) contradicts it.

3) The real architectural smell: telemetry depends on logging structure
The root bug exists because the system coupled:

“stuff we want for verbose logging”
with
“stuff we want for metrics”

This PR patches it by splitting the payload, which is good, but the fact that it’s still the same variable passed into telemetry suggests telemetry API itself might be too logging-shaped.

openhands-ai · 2026-01-06T16:15:21Z

Looks like there are a few issues preventing this PR from being merged!

GitHub Actions are failing:
- Agent Server
- Run tests

If you'd like me to help, just leave a comment, like

@OpenHands please fix the failing actions on PR #1590 at branch `openhands/fix-context-window-zero-issue-1228`

Feel free to include any additional details that might help me get this PR into a better state.

_{^{You can manage your notification settings}}

openhands-ai bot mentioned this pull request Jan 5, 2026

Context Window Displays as 0 When Sending ConversationStatusUpdateEvent (Full State) #1228

Open

xingyaoww requested a review from hieptl January 5, 2026 16:36

xingyaoww added the integration-test Runs the integration tests and comments the results label Jan 5, 2026

xingyaoww force-pushed the openhands/fix-context-window-zero-issue-1228 branch from 208dd7a to 6ecbc5f Compare January 5, 2026 17:19

enyst reviewed Jan 5, 2026

View reviewed changes

openhands-sdk/openhands/sdk/llm/llm.py Show resolved Hide resolved

enyst reviewed Jan 5, 2026

View reviewed changes

openhands-sdk/openhands/sdk/llm/llm.py Show resolved Hide resolved

Merge branch 'main' into openhands/fix-context-window-zero-issue-1228

0331f23

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Fix context window displaying as 0 when resuming conversation #1590

Fix context window displaying as 0 when resuming conversation #1590

xingyaoww commented Jan 5, 2026 •

edited by github-actions bot

Loading

Uh oh!

github-actions bot commented Jan 5, 2026 •

edited

Loading

Uh oh!

github-actions bot commented Jan 5, 2026

Uh oh!

github-actions bot commented Jan 5, 2026

Uh oh!

xingyaoww commented Jan 5, 2026

Uh oh!

openhands-ai bot commented Jan 5, 2026

Uh oh!

openhands-ai bot commented Jan 5, 2026

Uh oh!

Uh oh!

Uh oh!

enyst commented Jan 5, 2026

Uh oh!

openhands-ai bot commented Jan 5, 2026

Uh oh!

openhands-ai bot commented Jan 5, 2026

Uh oh!

enyst commented Jan 5, 2026

Uh oh!

openhands-ai bot commented Jan 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Fix context window displaying as 0 when resuming conversation #1590

Are you sure you want to change the base?

Fix context window displaying as 0 when resuming conversation #1590

Conversation

xingyaoww commented Jan 5, 2026 • edited by github-actions bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Checklist

Uh oh!

github-actions bot commented Jan 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Jan 5, 2026

Uh oh!

github-actions bot commented Jan 5, 2026

🧪 Integration Tests Results

📁 Detailed Logs & Artifacts

📊 Summary

📋 Detailed Results

litellm_proxy_deepseek_deepseek_chat

litellm_proxy_claude_sonnet_4_5_20250929

litellm_proxy_vertex_ai_gemini_3_pro_preview

litellm_proxy_moonshot_kimi_k2_thinking

litellm_proxy_mistral_devstral_2512

litellm_proxy_gpt_5.1_codex_max

Uh oh!

xingyaoww commented Jan 5, 2026

Uh oh!

openhands-ai bot commented Jan 5, 2026

Uh oh!

openhands-ai bot commented Jan 5, 2026

Summary

Root Cause Analysis

Fix Applied

Testing

Changes Pushed

Uh oh!

Uh oh!

Uh oh!

enyst commented Jan 5, 2026

Uh oh!

openhands-ai bot commented Jan 5, 2026

Uh oh!

openhands-ai bot commented Jan 5, 2026

Final summary

Restated answer to the main question (“Is this correct?”)

Answer to the inline thread question

Checklist against your request

Notes on pushing / changes

Uh oh!

enyst commented Jan 5, 2026

[IMPROVEMENT OPPORTUNITIES]

Uh oh!

openhands-ai bot commented Jan 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

xingyaoww commented Jan 5, 2026 •

edited by github-actions bot

Loading

github-actions bot commented Jan 5, 2026 •

edited

Loading