fix(usage): keep the richest copy of a deduped usage record by xkelxmc · Pull Request #5700 · pingdotgg/t3code

xkelxmc · 2026-08-08T13:27:23Z

What changed

Usage de-duplication (within-file and across the scan) now keeps the richest copy of a dedupe key instead of the first one encountered, replacing a poorer copy's contribution exactly (tokens, cost, cache savings, record counts). Identical copies still keep the first, so the content-block-repeat case this rule was written for behaves as before.

usageScanCache.ts — dedupeWithinFile keeps the copy with the most tokens, in its original position; USAGE_SCAN_CACHE_VERSION bumped to 2 so entries holding first-seen copies are re-parsed rather than reused.
usageAggregation.ts — #seen grows from a Set of keys to a Map of contributions, so a richer copy arriving later can retract the poorer copy from its bucket and take its place.

Why it should exist

Copies sharing a message.id:requestId key are not always identical. Claude Code writes intermediate snapshots of a streamed message, and the earlier copies carry a partial output_tokens — the copy with the most tokens is the message as billed. First-wins keeps whichever copy the filesystem walk met first, which is frequently a partial one.

On our transcripts this is not a rounding error: comparing both strategies over the same one-week window, first-wins lost roughly half of all output tokens (input, cache-read and cache-write totals were identical), which understates Claude cost by a few percent — output is the most expensive channel.

Notes

duplicatesDropped keeps its meaning: the superseded poorer copy counts as the dropped one.
Sessions are deliberately not retracted; a superseded copy is the same message from the same session.
Without the cache version bump the fix appears to do nothing on warm installs — the scan cache stores records already deduped within their file.

Tests cover both arrival orders (partial→finished and finished→partial), the identical-copies case, and the cost retraction.

Note

Medium Risk
Changes how usage and cost are aggregated for billing dashboards; scope is limited to usage parsing/cache with targeted tests and a forced cache refresh.

Overview
Usage de-duplication now keeps the highest-token copy per dedupe key instead of the first seen, so Claude Code streaming snapshots (same key, growing output_tokens) bill against the finished message rather than a partial write.

dedupeWithinFile in usageScanCache.ts compares totalTokens and replaces a held row when a later duplicate is richer; identical duplicates still keep the first. USAGE_SCAN_CACHE_VERSION is bumped to 2 so v1 cache entries (first-seen partials) are discarded and files are re-parsed.

Tests cover within-file and aggregator paths for partial→finished and finished→partial order, identical copies, and cost retraction when a partial is superseded.

^{Reviewed by Cursor Bugbot for commit 052e18e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix usage deduplication to keep the richest copy of a duplicate record

Rewrites dedupeWithinFile in usageScanCache.ts to track the position of the first occurrence of each dedupe key, replacing it only if a later record has more total tokens.
Updates aggregation logic in usageAggregation.ts to apply the same richest-copy preference when folding duplicates across buckets.
Increments USAGE_SCAN_CACHE_VERSION from 1 to 2, invalidating existing cache entries so they are re-parsed with the new deduplication semantics.
Behavioral Change: previously the first-seen duplicate was always kept; now a later record with more tokens replaces it at its original position.

^{Macroscope summarized 052e18e.}

coderabbitai · 2026-08-08T13:27:33Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 8ae59077-93ba-4cf6-9e8b-842d36c72a11

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

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.}

macroscopeapp · 2026-08-08T13:33:47Z

Approvability

Verdict: Needs human review

This PR modifies usage/billing metering logic (deduplication to keep the richest token count). Changes affecting billing or metering require human review regardless of apparent simplicity.

^{You can customize Macroscope's approvability policy. Learn more.}

caezium · 2026-08-09T08:20:23Z

Could duplicate usage be merged by per-field maxima instead of selecting the record with the largest overall token total? Streaming snapshots can advance input, output, cache-read, 5m cache-write, and 1h cache-write independently, so a single “richest” record can discard a maximum from another snapshot. A crossing-fields regression test would capture this.

xkelxmc · 2026-08-08T13:27:23Z

What changed

Usage de-duplication (within-file and across the scan) now keeps the richest copy of a dedupe key instead of the first one encountered, replacing a poorer copy's contribution exactly (tokens, cost, cache savings, record counts). Identical copies still keep the first, so the content-block-repeat case this rule was written for behaves as before.

usageScanCache.ts — dedupeWithinFile keeps the copy with the most tokens, in its original position; USAGE_SCAN_CACHE_VERSION bumped to 2 so entries holding first-seen copies are re-parsed rather than reused.
usageAggregation.ts — #seen grows from a Set of keys to a Map of contributions, so a richer copy arriving later can retract the poorer copy from its bucket and take its place.

Why it should exist

Copies sharing a message.id:requestId key are not always identical. Claude Code writes intermediate snapshots of a streamed message, and the earlier copies carry a partial output_tokens — the copy with the most tokens is the message as billed. First-wins keeps whichever copy the filesystem walk met first, which is frequently a partial one.

On our transcripts this is not a rounding error: comparing both strategies over the same one-week window, first-wins lost roughly half of all output tokens (input, cache-read and cache-write totals were identical), which understates Claude cost by a few percent — output is the most expensive channel.

Notes

duplicatesDropped keeps its meaning: the superseded poorer copy counts as the dropped one.
Sessions are deliberately not retracted; a superseded copy is the same message from the same session.
Without the cache version bump the fix appears to do nothing on warm installs — the scan cache stores records already deduped within their file.

Tests cover both arrival orders (partial→finished and finished→partial), the identical-copies case, and the cost retraction.

Note

Medium Risk
Changes how usage and cost are aggregated for billing dashboards; scope is limited to usage parsing/cache with targeted tests and a forced cache refresh.

Overview
Usage de-duplication now keeps the highest-token copy per dedupe key instead of the first seen, so Claude Code streaming snapshots (same key, growing output_tokens) bill against the finished message rather than a partial write.

dedupeWithinFile in usageScanCache.ts compares totalTokens and replaces a held row when a later duplicate is richer; identical duplicates still keep the first. USAGE_SCAN_CACHE_VERSION is bumped to 2 so v1 cache entries (first-seen partials) are discarded and files are re-parsed.

Tests cover within-file and aggregator paths for partial→finished and finished→partial order, identical copies, and cost retraction when a partial is superseded.

^{Reviewed by Cursor Bugbot for commit 052e18e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix usage deduplication to keep the richest copy of a duplicate record

Rewrites dedupeWithinFile in usageScanCache.ts to track the position of the first occurrence of each dedupe key, replacing it only if a later record has more total tokens.
Updates aggregation logic in usageAggregation.ts to apply the same richest-copy preference when folding duplicates across buckets.
Increments USAGE_SCAN_CACHE_VERSION from 1 to 2, invalidating existing cache entries so they are re-parsed with the new deduplication semantics.
Behavioral Change: previously the first-seen duplicate was always kept; now a later record with more tokens replaces it at its original position.

^{Macroscope summarized 052e18e.}

coderabbitai · 2026-08-08T13:27:33Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 8ae59077-93ba-4cf6-9e8b-842d36c72a11

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

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.}

macroscopeapp · 2026-08-08T13:33:47Z

Approvability

Verdict: Needs human review

This PR modifies usage/billing metering logic (deduplication to keep the richest token count). Changes affecting billing or metering require human review regardless of apparent simplicity.

^{You can customize Macroscope's approvability policy. Learn more.}

caezium · 2026-08-09T08:20:23Z

Could duplicate usage be merged by per-field maxima instead of selecting the record with the largest overall token total? Streaming snapshots can advance input, output, cache-read, 5m cache-write, and 1h cache-write independently, so a single “richest” record can discard a maximum from another snapshot. A crossing-fields regression test would capture this.

xkelxmc · 2026-08-08T13:27:23Z

What changed

Usage de-duplication (within-file and across the scan) now keeps the richest copy of a dedupe key instead of the first one encountered, replacing a poorer copy's contribution exactly (tokens, cost, cache savings, record counts). Identical copies still keep the first, so the content-block-repeat case this rule was written for behaves as before.

usageScanCache.ts — dedupeWithinFile keeps the copy with the most tokens, in its original position; USAGE_SCAN_CACHE_VERSION bumped to 2 so entries holding first-seen copies are re-parsed rather than reused.
usageAggregation.ts — #seen grows from a Set of keys to a Map of contributions, so a richer copy arriving later can retract the poorer copy from its bucket and take its place.

Why it should exist

Copies sharing a message.id:requestId key are not always identical. Claude Code writes intermediate snapshots of a streamed message, and the earlier copies carry a partial output_tokens — the copy with the most tokens is the message as billed. First-wins keeps whichever copy the filesystem walk met first, which is frequently a partial one.

On our transcripts this is not a rounding error: comparing both strategies over the same one-week window, first-wins lost roughly half of all output tokens (input, cache-read and cache-write totals were identical), which understates Claude cost by a few percent — output is the most expensive channel.

Notes

duplicatesDropped keeps its meaning: the superseded poorer copy counts as the dropped one.
Sessions are deliberately not retracted; a superseded copy is the same message from the same session.
Without the cache version bump the fix appears to do nothing on warm installs — the scan cache stores records already deduped within their file.

Tests cover both arrival orders (partial→finished and finished→partial), the identical-copies case, and the cost retraction.

Note

Medium Risk
Changes how usage and cost are aggregated for billing dashboards; scope is limited to usage parsing/cache with targeted tests and a forced cache refresh.

Overview
Usage de-duplication now keeps the highest-token copy per dedupe key instead of the first seen, so Claude Code streaming snapshots (same key, growing output_tokens) bill against the finished message rather than a partial write.

dedupeWithinFile in usageScanCache.ts compares totalTokens and replaces a held row when a later duplicate is richer; identical duplicates still keep the first. USAGE_SCAN_CACHE_VERSION is bumped to 2 so v1 cache entries (first-seen partials) are discarded and files are re-parsed.

Tests cover within-file and aggregator paths for partial→finished and finished→partial order, identical copies, and cost retraction when a partial is superseded.

^{Reviewed by Cursor Bugbot for commit 052e18e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix usage deduplication to keep the richest copy of a duplicate record

Rewrites dedupeWithinFile in usageScanCache.ts to track the position of the first occurrence of each dedupe key, replacing it only if a later record has more total tokens.
Updates aggregation logic in usageAggregation.ts to apply the same richest-copy preference when folding duplicates across buckets.
Increments USAGE_SCAN_CACHE_VERSION from 1 to 2, invalidating existing cache entries so they are re-parsed with the new deduplication semantics.
Behavioral Change: previously the first-seen duplicate was always kept; now a later record with more tokens replaces it at its original position.

^{Macroscope summarized 052e18e.}

coderabbitai · 2026-08-08T13:27:33Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 8ae59077-93ba-4cf6-9e8b-842d36c72a11

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

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.}

macroscopeapp · 2026-08-08T13:33:47Z

Approvability

Verdict: Needs human review

This PR modifies usage/billing metering logic (deduplication to keep the richest token count). Changes affecting billing or metering require human review regardless of apparent simplicity.

^{You can customize Macroscope's approvability policy. Learn more.}

caezium · 2026-08-09T08:20:23Z

Could duplicate usage be merged by per-field maxima instead of selecting the record with the largest overall token total? Streaming snapshots can advance input, output, cache-read, 5m cache-write, and 1h cache-write independently, so a single “richest” record can discard a maximum from another snapshot. A crossing-fields regression test would capture this.

xkelxmc · 2026-08-08T13:27:23Z

What changed

Usage de-duplication (within-file and across the scan) now keeps the richest copy of a dedupe key instead of the first one encountered, replacing a poorer copy's contribution exactly (tokens, cost, cache savings, record counts). Identical copies still keep the first, so the content-block-repeat case this rule was written for behaves as before.

usageScanCache.ts — dedupeWithinFile keeps the copy with the most tokens, in its original position; USAGE_SCAN_CACHE_VERSION bumped to 2 so entries holding first-seen copies are re-parsed rather than reused.
usageAggregation.ts — #seen grows from a Set of keys to a Map of contributions, so a richer copy arriving later can retract the poorer copy from its bucket and take its place.

Why it should exist

Copies sharing a message.id:requestId key are not always identical. Claude Code writes intermediate snapshots of a streamed message, and the earlier copies carry a partial output_tokens — the copy with the most tokens is the message as billed. First-wins keeps whichever copy the filesystem walk met first, which is frequently a partial one.

On our transcripts this is not a rounding error: comparing both strategies over the same one-week window, first-wins lost roughly half of all output tokens (input, cache-read and cache-write totals were identical), which understates Claude cost by a few percent — output is the most expensive channel.

Notes

duplicatesDropped keeps its meaning: the superseded poorer copy counts as the dropped one.
Sessions are deliberately not retracted; a superseded copy is the same message from the same session.
Without the cache version bump the fix appears to do nothing on warm installs — the scan cache stores records already deduped within their file.

Tests cover both arrival orders (partial→finished and finished→partial), the identical-copies case, and the cost retraction.

Note

Medium Risk
Changes how usage and cost are aggregated for billing dashboards; scope is limited to usage parsing/cache with targeted tests and a forced cache refresh.

Overview
Usage de-duplication now keeps the highest-token copy per dedupe key instead of the first seen, so Claude Code streaming snapshots (same key, growing output_tokens) bill against the finished message rather than a partial write.

dedupeWithinFile in usageScanCache.ts compares totalTokens and replaces a held row when a later duplicate is richer; identical duplicates still keep the first. USAGE_SCAN_CACHE_VERSION is bumped to 2 so v1 cache entries (first-seen partials) are discarded and files are re-parsed.

Tests cover within-file and aggregator paths for partial→finished and finished→partial order, identical copies, and cost retraction when a partial is superseded.

^{Reviewed by Cursor Bugbot for commit 052e18e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix usage deduplication to keep the richest copy of a duplicate record

Rewrites dedupeWithinFile in usageScanCache.ts to track the position of the first occurrence of each dedupe key, replacing it only if a later record has more total tokens.
Updates aggregation logic in usageAggregation.ts to apply the same richest-copy preference when folding duplicates across buckets.
Increments USAGE_SCAN_CACHE_VERSION from 1 to 2, invalidating existing cache entries so they are re-parsed with the new deduplication semantics.
Behavioral Change: previously the first-seen duplicate was always kept; now a later record with more tokens replaces it at its original position.

^{Macroscope summarized 052e18e.}

coderabbitai · 2026-08-08T13:27:33Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 8ae59077-93ba-4cf6-9e8b-842d36c72a11

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

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.}

macroscopeapp · 2026-08-08T13:33:47Z

Approvability

Verdict: Needs human review

This PR modifies usage/billing metering logic (deduplication to keep the richest token count). Changes affecting billing or metering require human review regardless of apparent simplicity.

^{You can customize Macroscope's approvability policy. Learn more.}

caezium · 2026-08-09T08:20:23Z

Could duplicate usage be merged by per-field maxima instead of selecting the record with the largest overall token total? Streaming snapshots can advance input, output, cache-read, 5m cache-write, and 1h cache-write independently, so a single “richest” record can discard a maximum from another snapshot. A crossing-fields regression test would capture this.

xkelxmc · 2026-08-08T13:27:23Z

What changed

Usage de-duplication (within-file and across the scan) now keeps the richest copy of a dedupe key instead of the first one encountered, replacing a poorer copy's contribution exactly (tokens, cost, cache savings, record counts). Identical copies still keep the first, so the content-block-repeat case this rule was written for behaves as before.

usageScanCache.ts — dedupeWithinFile keeps the copy with the most tokens, in its original position; USAGE_SCAN_CACHE_VERSION bumped to 2 so entries holding first-seen copies are re-parsed rather than reused.
usageAggregation.ts — #seen grows from a Set of keys to a Map of contributions, so a richer copy arriving later can retract the poorer copy from its bucket and take its place.

Why it should exist

Copies sharing a message.id:requestId key are not always identical. Claude Code writes intermediate snapshots of a streamed message, and the earlier copies carry a partial output_tokens — the copy with the most tokens is the message as billed. First-wins keeps whichever copy the filesystem walk met first, which is frequently a partial one.

On our transcripts this is not a rounding error: comparing both strategies over the same one-week window, first-wins lost roughly half of all output tokens (input, cache-read and cache-write totals were identical), which understates Claude cost by a few percent — output is the most expensive channel.

Notes

duplicatesDropped keeps its meaning: the superseded poorer copy counts as the dropped one.
Sessions are deliberately not retracted; a superseded copy is the same message from the same session.
Without the cache version bump the fix appears to do nothing on warm installs — the scan cache stores records already deduped within their file.

Tests cover both arrival orders (partial→finished and finished→partial), the identical-copies case, and the cost retraction.

Note

Medium Risk
Changes how usage and cost are aggregated for billing dashboards; scope is limited to usage parsing/cache with targeted tests and a forced cache refresh.

Overview
Usage de-duplication now keeps the highest-token copy per dedupe key instead of the first seen, so Claude Code streaming snapshots (same key, growing output_tokens) bill against the finished message rather than a partial write.

dedupeWithinFile in usageScanCache.ts compares totalTokens and replaces a held row when a later duplicate is richer; identical duplicates still keep the first. USAGE_SCAN_CACHE_VERSION is bumped to 2 so v1 cache entries (first-seen partials) are discarded and files are re-parsed.

Tests cover within-file and aggregator paths for partial→finished and finished→partial order, identical copies, and cost retraction when a partial is superseded.

^{Reviewed by Cursor Bugbot for commit 052e18e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix usage deduplication to keep the richest copy of a duplicate record

Rewrites dedupeWithinFile in usageScanCache.ts to track the position of the first occurrence of each dedupe key, replacing it only if a later record has more total tokens.
Updates aggregation logic in usageAggregation.ts to apply the same richest-copy preference when folding duplicates across buckets.
Increments USAGE_SCAN_CACHE_VERSION from 1 to 2, invalidating existing cache entries so they are re-parsed with the new deduplication semantics.
Behavioral Change: previously the first-seen duplicate was always kept; now a later record with more tokens replaces it at its original position.

^{Macroscope summarized 052e18e.}

coderabbitai · 2026-08-08T13:27:33Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 8ae59077-93ba-4cf6-9e8b-842d36c72a11

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

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.}

macroscopeapp · 2026-08-08T13:33:47Z

Approvability

Verdict: Needs human review

This PR modifies usage/billing metering logic (deduplication to keep the richest token count). Changes affecting billing or metering require human review regardless of apparent simplicity.

^{You can customize Macroscope's approvability policy. Learn more.}

caezium · 2026-08-09T08:20:23Z

Could duplicate usage be merged by per-field maxima instead of selecting the record with the largest overall token total? Streaming snapshots can advance input, output, cache-read, 5m cache-write, and 1h cache-write independently, so a single “richest” record can discard a maximum from another snapshot. A crossing-fields regression test would capture this.

xkelxmc · 2026-08-08T13:27:23Z

What changed

Usage de-duplication (within-file and across the scan) now keeps the richest copy of a dedupe key instead of the first one encountered, replacing a poorer copy's contribution exactly (tokens, cost, cache savings, record counts). Identical copies still keep the first, so the content-block-repeat case this rule was written for behaves as before.

usageScanCache.ts — dedupeWithinFile keeps the copy with the most tokens, in its original position; USAGE_SCAN_CACHE_VERSION bumped to 2 so entries holding first-seen copies are re-parsed rather than reused.
usageAggregation.ts — #seen grows from a Set of keys to a Map of contributions, so a richer copy arriving later can retract the poorer copy from its bucket and take its place.

Why it should exist

Copies sharing a message.id:requestId key are not always identical. Claude Code writes intermediate snapshots of a streamed message, and the earlier copies carry a partial output_tokens — the copy with the most tokens is the message as billed. First-wins keeps whichever copy the filesystem walk met first, which is frequently a partial one.

On our transcripts this is not a rounding error: comparing both strategies over the same one-week window, first-wins lost roughly half of all output tokens (input, cache-read and cache-write totals were identical), which understates Claude cost by a few percent — output is the most expensive channel.

Notes

duplicatesDropped keeps its meaning: the superseded poorer copy counts as the dropped one.
Sessions are deliberately not retracted; a superseded copy is the same message from the same session.
Without the cache version bump the fix appears to do nothing on warm installs — the scan cache stores records already deduped within their file.

Tests cover both arrival orders (partial→finished and finished→partial), the identical-copies case, and the cost retraction.

Note

Medium Risk
Changes how usage and cost are aggregated for billing dashboards; scope is limited to usage parsing/cache with targeted tests and a forced cache refresh.

Overview
Usage de-duplication now keeps the highest-token copy per dedupe key instead of the first seen, so Claude Code streaming snapshots (same key, growing output_tokens) bill against the finished message rather than a partial write.

dedupeWithinFile in usageScanCache.ts compares totalTokens and replaces a held row when a later duplicate is richer; identical duplicates still keep the first. USAGE_SCAN_CACHE_VERSION is bumped to 2 so v1 cache entries (first-seen partials) are discarded and files are re-parsed.

Tests cover within-file and aggregator paths for partial→finished and finished→partial order, identical copies, and cost retraction when a partial is superseded.

^{Reviewed by Cursor Bugbot for commit 052e18e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix usage deduplication to keep the richest copy of a duplicate record

Rewrites dedupeWithinFile in usageScanCache.ts to track the position of the first occurrence of each dedupe key, replacing it only if a later record has more total tokens.
Updates aggregation logic in usageAggregation.ts to apply the same richest-copy preference when folding duplicates across buckets.
Increments USAGE_SCAN_CACHE_VERSION from 1 to 2, invalidating existing cache entries so they are re-parsed with the new deduplication semantics.
Behavioral Change: previously the first-seen duplicate was always kept; now a later record with more tokens replaces it at its original position.

^{Macroscope summarized 052e18e.}

coderabbitai · 2026-08-08T13:27:33Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 8ae59077-93ba-4cf6-9e8b-842d36c72a11

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

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.}

macroscopeapp · 2026-08-08T13:33:47Z

Approvability

Verdict: Needs human review

This PR modifies usage/billing metering logic (deduplication to keep the richest token count). Changes affecting billing or metering require human review regardless of apparent simplicity.

^{You can customize Macroscope's approvability policy. Learn more.}

caezium · 2026-08-09T08:20:23Z

Could duplicate usage be merged by per-field maxima instead of selecting the record with the largest overall token total? Streaming snapshots can advance input, output, cache-read, 5m cache-write, and 1h cache-write independently, so a single “richest” record can discard a maximum from another snapshot. A crossing-fields regression test would capture this.

xkelxmc · 2026-08-08T13:27:23Z

What changed

Usage de-duplication (within-file and across the scan) now keeps the richest copy of a dedupe key instead of the first one encountered, replacing a poorer copy's contribution exactly (tokens, cost, cache savings, record counts). Identical copies still keep the first, so the content-block-repeat case this rule was written for behaves as before.

usageScanCache.ts — dedupeWithinFile keeps the copy with the most tokens, in its original position; USAGE_SCAN_CACHE_VERSION bumped to 2 so entries holding first-seen copies are re-parsed rather than reused.
usageAggregation.ts — #seen grows from a Set of keys to a Map of contributions, so a richer copy arriving later can retract the poorer copy from its bucket and take its place.

Why it should exist

Copies sharing a message.id:requestId key are not always identical. Claude Code writes intermediate snapshots of a streamed message, and the earlier copies carry a partial output_tokens — the copy with the most tokens is the message as billed. First-wins keeps whichever copy the filesystem walk met first, which is frequently a partial one.

On our transcripts this is not a rounding error: comparing both strategies over the same one-week window, first-wins lost roughly half of all output tokens (input, cache-read and cache-write totals were identical), which understates Claude cost by a few percent — output is the most expensive channel.

Notes

duplicatesDropped keeps its meaning: the superseded poorer copy counts as the dropped one.
Sessions are deliberately not retracted; a superseded copy is the same message from the same session.
Without the cache version bump the fix appears to do nothing on warm installs — the scan cache stores records already deduped within their file.

Tests cover both arrival orders (partial→finished and finished→partial), the identical-copies case, and the cost retraction.

Note

Medium Risk
Changes how usage and cost are aggregated for billing dashboards; scope is limited to usage parsing/cache with targeted tests and a forced cache refresh.

Overview
Usage de-duplication now keeps the highest-token copy per dedupe key instead of the first seen, so Claude Code streaming snapshots (same key, growing output_tokens) bill against the finished message rather than a partial write.

dedupeWithinFile in usageScanCache.ts compares totalTokens and replaces a held row when a later duplicate is richer; identical duplicates still keep the first. USAGE_SCAN_CACHE_VERSION is bumped to 2 so v1 cache entries (first-seen partials) are discarded and files are re-parsed.

Tests cover within-file and aggregator paths for partial→finished and finished→partial order, identical copies, and cost retraction when a partial is superseded.

^{Reviewed by Cursor Bugbot for commit 052e18e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix usage deduplication to keep the richest copy of a duplicate record

Rewrites dedupeWithinFile in usageScanCache.ts to track the position of the first occurrence of each dedupe key, replacing it only if a later record has more total tokens.
Updates aggregation logic in usageAggregation.ts to apply the same richest-copy preference when folding duplicates across buckets.
Increments USAGE_SCAN_CACHE_VERSION from 1 to 2, invalidating existing cache entries so they are re-parsed with the new deduplication semantics.
Behavioral Change: previously the first-seen duplicate was always kept; now a later record with more tokens replaces it at its original position.

^{Macroscope summarized 052e18e.}

coderabbitai · 2026-08-08T13:27:33Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 8ae59077-93ba-4cf6-9e8b-842d36c72a11

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

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.}

macroscopeapp · 2026-08-08T13:33:47Z

Approvability

Verdict: Needs human review

This PR modifies usage/billing metering logic (deduplication to keep the richest token count). Changes affecting billing or metering require human review regardless of apparent simplicity.

^{You can customize Macroscope's approvability policy. Learn more.}

caezium · 2026-08-09T08:20:23Z

Could duplicate usage be merged by per-field maxima instead of selecting the record with the largest overall token total? Streaming snapshots can advance input, output, cache-read, 5m cache-write, and 1h cache-write independently, so a single “richest” record can discard a maximum from another snapshot. A crossing-fields regression test would capture this.

xkelxmc · 2026-08-08T13:27:23Z

What changed

Usage de-duplication (within-file and across the scan) now keeps the richest copy of a dedupe key instead of the first one encountered, replacing a poorer copy's contribution exactly (tokens, cost, cache savings, record counts). Identical copies still keep the first, so the content-block-repeat case this rule was written for behaves as before.

usageScanCache.ts — dedupeWithinFile keeps the copy with the most tokens, in its original position; USAGE_SCAN_CACHE_VERSION bumped to 2 so entries holding first-seen copies are re-parsed rather than reused.
usageAggregation.ts — #seen grows from a Set of keys to a Map of contributions, so a richer copy arriving later can retract the poorer copy from its bucket and take its place.

Why it should exist

Copies sharing a message.id:requestId key are not always identical. Claude Code writes intermediate snapshots of a streamed message, and the earlier copies carry a partial output_tokens — the copy with the most tokens is the message as billed. First-wins keeps whichever copy the filesystem walk met first, which is frequently a partial one.

On our transcripts this is not a rounding error: comparing both strategies over the same one-week window, first-wins lost roughly half of all output tokens (input, cache-read and cache-write totals were identical), which understates Claude cost by a few percent — output is the most expensive channel.

Notes

duplicatesDropped keeps its meaning: the superseded poorer copy counts as the dropped one.
Sessions are deliberately not retracted; a superseded copy is the same message from the same session.
Without the cache version bump the fix appears to do nothing on warm installs — the scan cache stores records already deduped within their file.

Tests cover both arrival orders (partial→finished and finished→partial), the identical-copies case, and the cost retraction.

Note

Medium Risk
Changes how usage and cost are aggregated for billing dashboards; scope is limited to usage parsing/cache with targeted tests and a forced cache refresh.

Overview
Usage de-duplication now keeps the highest-token copy per dedupe key instead of the first seen, so Claude Code streaming snapshots (same key, growing output_tokens) bill against the finished message rather than a partial write.

dedupeWithinFile in usageScanCache.ts compares totalTokens and replaces a held row when a later duplicate is richer; identical duplicates still keep the first. USAGE_SCAN_CACHE_VERSION is bumped to 2 so v1 cache entries (first-seen partials) are discarded and files are re-parsed.

Tests cover within-file and aggregator paths for partial→finished and finished→partial order, identical copies, and cost retraction when a partial is superseded.

^{Reviewed by Cursor Bugbot for commit 052e18e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix usage deduplication to keep the richest copy of a duplicate record

Rewrites dedupeWithinFile in usageScanCache.ts to track the position of the first occurrence of each dedupe key, replacing it only if a later record has more total tokens.
Updates aggregation logic in usageAggregation.ts to apply the same richest-copy preference when folding duplicates across buckets.
Increments USAGE_SCAN_CACHE_VERSION from 1 to 2, invalidating existing cache entries so they are re-parsed with the new deduplication semantics.
Behavioral Change: previously the first-seen duplicate was always kept; now a later record with more tokens replaces it at its original position.

^{Macroscope summarized 052e18e.}

coderabbitai · 2026-08-08T13:27:33Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 8ae59077-93ba-4cf6-9e8b-842d36c72a11

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

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.}

macroscopeapp · 2026-08-08T13:33:47Z

Approvability

Verdict: Needs human review

This PR modifies usage/billing metering logic (deduplication to keep the richest token count). Changes affecting billing or metering require human review regardless of apparent simplicity.

^{You can customize Macroscope's approvability policy. Learn more.}

caezium · 2026-08-09T08:20:23Z

Could duplicate usage be merged by per-field maxima instead of selecting the record with the largest overall token total? Streaming snapshots can advance input, output, cache-read, 5m cache-write, and 1h cache-write independently, so a single “richest” record can discard a maximum from another snapshot. A crossing-fields regression test would capture this.

fix(usage): keep the richest copy of a deduped usage record
052e18e

github-actionsBot added vouch:unvouched PR author is not yet trusted in the VOUCHED list. size:M 30-99 changed lines (additions + deletions). labels Aug 8, 2026

This was referenced Aug 9, 2026
fix(server): reconcile Codex cumulative usage #5814
Open
fix(server): exclude inherited Codex fork usage #5811
Closed