← Troubleshooting

Logs & diagnostics

Antigravity Manager: Logs, Token Usage & Bug Reports

Open Traffic Logs, turn on Recording, and reproduce the problem once. Use that request’s details to find the failed step, check its token counts, or send a focused bug report.

Updated · Settings and behavior below refer to Antigravity Tools v5.0.0.

01 / Start here

Where are the Antigravity Manager logs?

Open Traffic Logs in the navigation for individual API requests. For startup failures and internal errors, open the desktop app’s Settings → Advanced → Data Directory → Open and look in logs/.

LocationUse it for
Traffic Logs
proxy_logs.db
Request status, models, account, timing and saved payloads. Open a row to see its details.
logs/error.logInternal failures around the time of the problem. Older error.log* files may also be present.
Token Stats
token_stats.db
Recorded usage totals and breakdowns by account or model.
debug_logs/Extra request and response diagnostics when Debug Logging is enabled. A configured output directory can override this location.

The usual data directory is ~/.antigravity_tools on macOS/Linux or %USERPROFILE%\.antigravity_tools on Windows. Use the path shown by your running instance if you moved it. In Docker or headless mode, inspect the server’s data directory or mounted volume; the browser’s own computer holds a different filesystem. See data locations and backups.

Use Open to inspect the folder. The adjacent folder-selection action starts a data migration. Keep the database files intact and inspect request records through Traffic Logs.

02 / Find one operation

How do I find the request that failed?

Record the failure time with its time zone, then match the request by endpoint, model and account in Traffic Logs.

  1. If the recording control says Paused, click it so it says Recording. Send one small request from the affected client and wait for it to finish or fail.
  2. Use Error to find failing HTTP statuses, or enter a model, path or status in the search field. Use Filter by account when you know which account handled it. Click Reset to clear filters if the row is missing.
  3. If the client received HTTP 200 but showed an error message, empty answer or missing output, switch back to All. Read the response body and check the requested result.
  4. Open the row’s Request Details. Keep the timestamp, status, Model (Req / Real), account label and Duration / TTFT. Copy the exact error and the relevant timing breakdown.

No row appeared? Confirm that the client points to this Manager instance. Keep Recording enabled for the next reproduction. With Capture Health off, successful GET requests—including model-list requests—and recognized health probes are skipped. Enable it temporarily when diagnosing those calls. Administrative /api/ traffic is excluded from this request monitor.

A streaming row may arrive after the stream closes. For a request that is still hanging, save the client’s last output and inspect the application log at the same time. Use connection troubleshooting for failures before a response, or streaming stages to locate a later stall.

Which request ID should I include?

The Request ID displayed in Monitor identifies the local log record. Label it “Local Monitor record ID” in your report. Add a client request ID or upstream request ID separately when that layer exposes one.

Keep one consistent account label, such as Account A, across excerpts. If an identifier is missing, write “unavailable.” Session IDs group a conversation; retain them privately when matching requests, then replace them consistently in a public report. Use the timestamp and model alongside IDs to follow retries or fallback attempts.

03 / Inspect the saved content

Why is a request payload missing or shortened?

First open the request details and switch the viewer to Full Mode. If the saved record is still incomplete, check its storage mode and capture settings before repeating a small request.

The viewer’s Concise Mode changes how saved content is displayed. Separately, Log Settings → Monitor payload storage controls what gets saved: Simple (default) simplifies payloads, while Full original retains more detail for new captures. Select Full original, click Save, reproduce once, and restore the previous setting afterward. Changing the mode cannot recover content already discarded.

What you seeWhat to check next
Only the list summaryOpen the row. Inspect Request Payload, Forwarded Request Payload and Response Payload separately.
No forwarded payloadCheck the protocol and forwarding path. Some paths forward directly or leave this field unrecorded. Compare the client request with its response.
Shortened text or media placeholdersUse a small text-only reproduction or a small, shareable media file. Multipart uploads are summarized, and some image data is removed before storage.
Old details disappearedCheck the storage budget. Space cleanup can remove older bodies or whole records. Save the relevant excerpt promptly.
A combined streaming responseUse client-side stream diagnostics when you need event order or chunk boundaries. Monitor consolidates streaming content into a readable response.

In v5.0.0, the monitor middleware reads request bodies up to 100 MiB. Keep diagnostic uploads well below that limit. Full original still passes through protocol-specific summaries and storage limits, so choose the smallest input that triggers the failure.

Capture extra detail with Debug Logging

Open Settings → Advanced → Debug Logging. Note its previous state and Debug Log Output Directory, enable it, save the settings, and reproduce once using harmless input. The default output is debug_logs/ under the active data directory.

Copy only the files for that time interval to a private working folder, then restore the previous logging settings. Large debug streams are bounded: v5.0.0 keeps the first and last 256 KiB with a truncation marker when content between them is omitted.

Review the files before sharing. Prompts, file contents, signatures, account references and session identifiers can remain in diagnostic output. Prepare a separate redacted copy using the report checklist.

04 / Make sense of the numbers

Why does Token Stats show zero or an unexpected cost?

Compare one completed request with Token Stats using the same model, account and time range. Start with the numeric Input, Output and Cache values, then investigate the chart or cost estimate.

  1. Confirm that the client received the requested answer. If output is empty or interrupted, follow response completion checks.
  2. Inspect the request’s response payload for usage fields and compare them with the row’s counts. An interrupted stream can lose final usage data. The v5.0.0 Claude non-streaming collector can also leave output at zero when an update contains only output_tokens.
  3. Open Token Stats, refresh, and widen the range to All Time if needed. Compare the numeric model and account breakdowns. Record the server’s time zone as well as the client’s when comparing daily or hourly totals.
  4. Check account attribution. Token Stats records a request when an account and both input/output counts are available. A failure before account selection can appear in Traffic Logs without adding a usage entry.

Paused recording with rising totals is expected: v5.0.0 records eligible token usage before it checks whether request logging is enabled. Pausing Traffic Logs stops new request records, while usage aggregation can continue. Some input counts also use a fallback estimate when the response lacks usable input usage.

Check the price rule behind the estimate

Open the pricing control in Token Stats and check Model Pricing ($/1M Tokens). Verify the model identifier and input, output and cache prices. Usage is grouped under the requested model name in this version; compare it with the mapped and upstream model in request details if you use aliases.

The displayed estimate uses the current price rules stored in that desktop view or browser. Changing a rule recalculates the estimate for the selected historical usage. Two browsers can therefore show different costs for the same recorded tokens. Check those rules even when the interface uses the label “Snapshot Accumulated.”

How the local cost estimate is calculated

The estimate adds uncached input, cached input and output at their respective rates per million tokens:

uncached input = max(0, input tokens − cached tokens)
estimated cost = (uncached input × input price
                + cached tokens × cache price
                + output tokens × output price) / 1,000,000

Cache tokens are included within the recorded input total; adding them again would double-count that portion. Use your provider’s billing records for actual charges and Accounts quota indicators for current model availability.

If a chart seems active while its numbers are zero, compare the breakdown values and request details before changing accounts or clearing data. Include the exact view and Manager version in the report so a display issue can be separated from missing usage records.

05 / Reduce the failing case

How do I make a useful reproduction?

Repeat the smallest task that reaches the failing step, then change one condition at a time. Keep the model, client and endpoint fixed for the first comparison.

  1. Start with a short answer. In a fresh session, ask for one sentence. Save the actual response and error text, including an “unsupported version” message even if the HTTP status is 200.
  2. Add the failing feature. For a tool problem, use the file-read and follow-up test in a disposable folder. For media, use one small file and check that the requested image, audio or video is returned.
  3. Compare session history. Repeat in the affected conversation if safe. Record whether failure starts after a tool result or only in a long session. Check files for partial changes before retrying a write.
  4. Compare one setting. For a stream failure, compare streaming and non-streaming if the client supports both. For a connection problem, compare the local route with the failing network route.
  5. Restart only when relevant. Save work first. State exactly which process you restarted and whether the same task then succeeded. Restore temporary settings when finished.

Stop at the first failing step and record passed, failed or not tested for each comparison. Use 429, 503 and account-pool troubleshooting when the failure happens before generation. A concise reproduction with one clear difference is easier to investigate than repeated retries with several settings changed.

How do I control log storage and clean up?

Save the relevant request details before clearing logs, then open Traffic Logs → Log Settings to adjust storage limits.

Review Request Logs (Sliding Window) and Internal Error Logs (Sliding Window) separately, then click Save. In v5.0.0, request history is managed by capacity and row limits; internal error logs have their own capacity budget. Older entries can be evicted as new entries arrive.

Clear All Historical Requests removes request history. Token statistics, debug files and thinking/session stores have separate storage and cleanup paths. Choose the action for the data you intend to remove, and keep account files untouched. Use the data and backup guide before broader cleanup or migration.

06 / Send a focused report

What should I include in a bug report?

Include the exact versions, a short reproduction, expected and actual results, and the smallest redacted log excerpt that shows the failure. Search existing issues for the exact error first; compare the affected version and steps before adding your case.

Copy this template and fill only the applicable fields. Use “not tested” for comparisons you have yet to run and “unavailable” for identifiers you could not obtain.

Title: [client] — [specific failure] — Manager [version/channel]

Manager version and release channel:
Operating system and CPU architecture:
Client name and exact version:
Deployment: desktop / Docker / headless
Protocol and endpoint (remove credentials and private hostnames):
Requested model / mapped model / upstream model, if shown:
Time of failure and time zone:

Steps to reproduce:
1.
2.
Expected result:
Actual result and exact error:
HTTP status, if available:
Last completed stage:

Local Monitor record ID, if available:
Client request ID, if available:
Upstream request ID, if available:
Account label: Account A / Account B / unknown

Comparisons actually run and their results:
Fresh session:
Same task after changing one setting:
Client or Manager restart, if relevant:

Relevant log excerpt (redacted):
Capture settings: Recording / storage mode / Debug Logging
Anything omitted: media / long fields / private content
Changes still active:

Remove private data from the copy you share

Remove API keys, access and refresh tokens, cookies, OAuth codes and authorization URLs, signed download URLs, email addresses, private prompts, file contents, personal paths and real thinking signatures. Inspect screenshots, headers and URL query strings as well as the main error text.

Automatic header masking covers common credentials but keeps session markers and can leave account emails in other fields. Use consistent labels such as Account A, Request A and Session A. Keep the unedited original privately; attach a short redacted excerpt rather than an account export, credential database, full HAR or entire data folder.

How should a redacted report read?

Keep the sequence and exact error visible while replacing private values. For a failure after a tool result, fill in this structure with your own observations:

Time: [timestamp with UTC offset]
Client: [name and exact version]
Request: POST /v1/messages
Account: Account A
Local Monitor record: Request A
Upstream request ID: unavailable

Expected: answer the follow-up using the file-read result.
Actual: [exact error text] after the tool returned its result.
Fresh-session comparison: [result, or "not tested"]
Attachments: redacted error excerpt and timing details.
Removed: account email, private file path and file contents.

The timestamp and request path locate the operation; Account A and Request A connect your excerpts. The expected and actual results identify the failed step. Listing what you removed helps the maintainer request a smaller follow-up sample if needed.

Version-specific implementation: Traffic Logs controls, request capture, usage recording, and local cost calculation.

WINDOWS DOWNLOAD

Archive password

Keep this password to open your Windows download if prompted.