---

name: buildbuddy-invocation-compare description: Compare two BuildBuddy invocation URLs or IDs to troubleshoot hermeticity/reproducibility and cache invalidation. Use when asked to diff canonical Bazel flags, find the first shared action-cache misses, inspect Action/ActionResult differences (inputs/command/env/platform/outputs), or analyze compact execution logs with bb explain. metadata: short-description: Compare two BuildBuddy invocations and pinpoint cache/action differences.

BuildBuddy Invocation Compare

Overview

Compare two invocations end-to-end to isolate the earliest shared action-cache miss and pinpoint why actions are rebuilt or produce divergent outputs.

Workflow

0) Preflight: auth + inputs

• Extract invocation IDs from the URLs.
• Confirm API base URL and API key presence without printing the key.
• Create a temp working directory and cache API responses there.
• Use the request templates in references/requests.md.

1) Diff canonical flags

• Call GetInvocation for both IDs.
• Extract the canonical command line (structuredCommandLine.commandLineLabel matching canonical).
• Sort flags and diff; highlight any differences in startup or command options.
• If canonical is missing, fall back to the only structured command line and note the limitation.
• Some entries may have commandLineLabel=null; filter them out to avoid jq errors.

2) Find the first shared AC miss

• Call GetCacheScoreCard for both invocations filtered to:
  • cache_type=AC, request_type=READ, response_type=NOT_FOUND
  • ordered by start_time ascending
• Use scripts/find_first_shared_ac_miss.py to identify the earliest shared {target, mnemonic} with AC misses on both sides.
• Record target_id, action_mnemonic, and action_id from each side for the next step.

3) Locate the rerun action(s)

• Call GetExecution with inline_execute_response=true for each invocation.
• Filter executions by target label, mnemonic, and primary output path (if known) to disambiguate actions.
• Capture:
  • action_digest (Action proto in CAS)
  • action_result_digest or execute_response.result (ActionResult)
  • execute_response.result (stdout/stderr/output digests)
  • primary_output_path to ensure you are comparing the correct action

4) Diff Action and ActionResult

Fetch protos via bazel run //tools/cas and compare with diff -u or jq.

Focus on these common divergence points:

• Input root tree: Compare Action.input_root_digest; if it differs, fetch both trees and identify differing file paths/digests. If only a few files differ, download and diff them.
• Command args / outputs: Compare Command.arguments, Command.output_files, and Command.output_directories. Expect platform or toolchain mismatches to surface here.
• Env / platform properties: Compare Command.environment_variables and Action.platform.properties. Treat differences as evidence of non-strict env or platform drift.
• Output digests: Compare ActionResult.output_files and output_directories; if digests differ with identical inputs/commands, suspect non-determinism.

Prefer bb explain when input roots differ and compact execution logs are available to pinpoint exact input changes.

5) Compact execution logs (optional)

• If execution_log.binpb.zst appears in the build tool logs, download both logs and diff with bb explain.
• Alternatively, pass invocation IDs directly to bb explain when the logs are available in BES.

Response requirements

• Summarize the root cause clearly and tie it to the earliest shared AC miss.
• Call out which of the divergence categories applies (inputs, command/outputs, env/platform, output digest).
• Include a pointer to the compare UI sources: app/compare/compare_actions.tsx or app/compare/compare_invocations.tsx for deeper inspection.

Resources

• references/requests.md for API + jq templates.
• scripts/find_first_shared_ac_miss.py to identify the earliest shared AC miss.