Repository navigation
Add opt-in compact prompts mode to reduce LLM token usage #456
Description
LLM analyzer prompts carry redundant content that inflates token consumption without improving detection quality:
- Line numbers are zero-padded (L001: instead of L1:), adding ~3 wasted characters per line across every file sent to the LLM.
- The meta-analyzer embeds each finding's context snippet in the prompt alongside the full line-numbered file content, duplicating the same code.
- The OverallAssessment field in MetaAnalyzerResult is generated by the LLM but never consumed downstream (the report node computes its own risk score).
- Several analyzer prompts (SSD, SDI, SQP, meta-analyzer, TP4) contain repeated phrasing, verbose formatting (horizontal rules, per-section rule-ID reminders already present in the
summary table), and filler sentences.
Proposal: Introduce a SKILLSPECTOR_COMPACT_PROMPTS=true environment variable that activates condensed prompt variants. Default behavior stays identical for backward compatibility.
When enabled:
- Line numbers drop zero-padding (L1: instead of L001:)
- Meta-analyzer omits redundant finding context from the prompt
- Meta-analyzer uses a slimmer structured output schema (no OverallAssessment)
- All five LLM analyzer prompts use condensed variants (same semantic content, less formatting overhead)
- BASE_ANALYSIS_PROMPT output guidelines shortened from 5 lines to 2
Estimated savings per scan: 400-800 input tokens from prompt condensation, ~375 input tokens per call from line numbering (on a 500-line file), 30-50 output tokens per meta-analyzer call from schema reduction, plus variable savings from context omission.
Activity
rng1995 commentedon Sep 16, 2026
Two compact-prompt implementations are under review: PR #457 proposes the broader opt-in compact mode, while PR #542 implements the first line-number compaction slice. Keeping this feature open until maintainers select and merge the intended scope.
rng1995 commentedon Sep 21, 2026
Status correction: the broader compact-prompt implementation, PR #457, is now a draft. The smaller line-number compaction slice, PR #542, remains open and ready for review. Neither has merged; keeping this feature open pending acceptance and merge of the intended scope.
rng1995 commentedon Sep 22, 2026
PR #542 has now merged. It implements item 1:
SKILLSPECTOR_COMPACT_PROMPTS=trueremoves zero-padding from prompt line numbers while preserving the default behavior. The relevant default/compact prompt regressions pass on currentmainat844ac30.This remains a partial implementation of the issue. Items 2-5 (omitting duplicate meta-analyzer context, slimming its output schema, condensed analyzer prompts, and shorter base guidelines) have not landed. The broader PR #457 remains a draft. Keeping the issue open for that remaining scope; the merged line-number slice is not yet in the latest published release, v2.11.2.
rng1995 commentedon Sep 23, 2026
Publication update: the line-number compaction from PR #542 is now included in v2.12.0. This supersedes the earlier “not yet released” status. Items 2-5 remain outstanding, with the broader PR #457 still draft, so this issue remains open.
rng1995 commentedon Oct 4, 2026
PR-state snapshot checked on 2026-10-04:
Merged #542 delivered line-number compaction only. The remaining compact-prompt items are still in draft #457.
Keeping this issue open: the relevant implementation is not merged and the remaining scope still needs verification. Check/review status is a point-in-time snapshot, not a claim of merge readiness.
rng1995 commentedon Oct 6, 2026
Merge-state update checked on 2026-10-06: PR #457 is closed without merging, so it is no longer an active implementation.
Merged PR #542 delivered compact line numbering only. Context deduplication, the slimmer meta-analysis schema, condensed analyzer prompts, and shorter base output guidance remain outside that merged change. Keeping this issue open; no merged PR completes the remaining compact-prompt scope.