Open
Conversation
`burn waste` now hard-filters its input slice against the coverage flags
each detector requires, refuses with a non-zero exit when no turn meets
the prereqs, and annotates partial exclusions both in text and JSON
output.
- `attributeWaste` / `aggregateBy*` need `hasToolCalls` +
`hasToolResultEvents`. The orchestrator filters the slice before
calling them and reports an "analyzed N of M turns; K excluded ..."
notice when some turns survive.
- `--patterns retries|failures` need the same flags; `reverts` needs
`hasToolCalls` + `hasRawContent` (the editPreHash / editPostHash
fields the parser computes from raw content); `compaction` is
unchanged because the compaction sidecar is independent of
`TurnRecord.fidelity`. Per-detector notices name the missing prereq
and the source kinds responsible.
- Refusal: when every input turn fails the prereq, exit 2 with
`burn waste: 142/142 turns lack tool-call/tool-result coverage
required for waste attribution. Sources: codex
(per-session-aggregate, missing tool-call records, tool-result
events). No waste analysis was performed.`
- `--json` carries a `fidelity` block (`{ analyzed, excluded, summary,
refused }`) mirroring `summary --json`. `--patterns` JSON additionally
exposes a `perDetector` array with each detector's `required` flags
and `excludedBySource` breakdown.
- Tests cover full refusal, partial exclusion, per-detector messaging,
and the JSON contract for both the attribution and patterns paths.
Closes #100. Refs #41, #76.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
Comment on lines
+677
to
+697
| const analyzedUnion = new Set<string>(); | ||
| for (const d of perDetectorCoverage) { | ||
| if (d.kind === 'compaction') continue; | ||
| const slice = perDetector.get(d.kind)!; | ||
| for (const t of slice) analyzedUnion.add(`${t.sessionId}|${t.messageId}`); | ||
| } | ||
| const analyzedCount = analyzedUnion.size; | ||
|
|
||
| if (args.flags['json'] === true) { | ||
| process.stdout.write( | ||
| JSON.stringify( | ||
| { | ||
| turnsAnalyzed, | ||
| turnsAnalyzed: analyzedCount, | ||
| retryLoops, | ||
| failureRuns, | ||
| compactions, | ||
| compactions: compactionLosses, | ||
| editReverts, | ||
| sessionSummaries: patterns.sessionSummaries, | ||
| sessionSummaries, | ||
| fidelity: { | ||
| analyzed: analyzedCount, | ||
| excluded: total - analyzedCount, |
Contributor
There was a problem hiding this comment.
🟡 Top-level turnsAnalyzed and fidelity.excluded are wrong when compaction is the only (or one of the) selected pattern detector(s)
When --patterns compaction is used, analyzedUnion at line 677-682 skips every compaction entry via if (d.kind === 'compaction') continue, so analyzedCount ends up 0 even though compaction successfully analyzed all total turns. This flows into:
turnsAnalyzed: 0in JSON (line 689) and text output (line 715)fidelity.excluded: total - 0 = total(line 697)
These contradict the per-detector breakdown which correctly reports compaction: { analyzed: total, excluded: 0 } (waste.ts:560-563). A JSON consumer seeing fidelity.excluded: N alongside fidelity.refused: false with an empty turnsAnalyzed: 0 would conclude nothing was analyzed. In the mixed case (--patterns retries,compaction), excluded also over-counts by not crediting turns that were only analyzed by compaction.
Prompt for agents
In runPatternsMode, the analyzedUnion set (line 677-682) deliberately skips compaction because compaction doesn't use fidelity-based coverage filtering. However, the resulting analyzedCount (which is 0 in the compaction-only case) is used for both turnsAnalyzed in the output and fidelity.excluded (total - analyzedCount) in the JSON fidelity block. This creates a contradiction: the top-level says excluded=N but the per-detector detail for compaction says excluded=0.
To fix this, the top-level fidelity counts need to account for compaction's contribution. One approach: when computing fidelity.analyzed and fidelity.excluded, take the max of analyzedCount and the compaction analyzed count (which is always total when compaction is selected). Or better: define the top-level fidelity.analyzed/excluded as a summary of the per-detector entries rather than deriving it from a union that ignores compaction.
Similarly, turnsAnalyzed in the headline should probably include turns that were analyzed by compaction when that's the only detector. A simple fix: if compaction is in selected, include all turns in the union.
Was this helpful? React with 👍 or 👎 to provide feedback.
This was referenced Apr 26, 2026
Open
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
burn wastenow hard-filters its input slice against the coverage flags each detector requires, refuses with a non-zero exit when no turn meets the prereqs, and annotates partial exclusions in both text and JSON output.attributeWaste/aggregateBy*needhasToolCalls+hasToolResultEvents.--patterns retries|failuresneed the same;revertsneedshasToolCalls+hasRawContent(the source ofeditPreHash/editPostHash);compactionis unchanged because its sidecar is independent ofTurnRecord.fidelity.burn waste: 142/142 turns lack tool-call/tool-result coverage required for waste attribution. Sources: codex (per-session-aggregate, missing tool-call records, tool-result events). No waste analysis was performed.and exits 2.--jsonnow carries afidelityblock{ analyzed, excluded, summary, refused }mirroringsummary --json.--patternsJSON additionally surfaces aperDetectorarray with each detector'srequiredflags andexcludedBySourcebreakdown.Test plan
pnpm run buildpnpm run test:ts— 533 tests pass (20 new inpackages/cli/src/commands/waste.test.ts)--patterns retries,failures, JSONfidelityshape (analyzed/excluded/summary/refused), per-detectorrequired+excludedBySourceJSON, and the orthogonalcompaction-only path that must not refuse on aggregate-only input.Refs
summarizeFidelity/hasMinimumFidelityand called out this wiring as deferred work)Out of scope per the issue: detecting waste patterns on aggregate-only data; modifying
attributeWasteitself (all gating happens at the CLI layer so analyze stays pure).Generated with Claude Code