Problem
In the batch of 21 issues (20260521-214143), the coordinator manually overrode auto-detected pipeline modes 5 times:
All 5 were small bug fixes where "full" mode (8 agents including both researchers) was wasteful. The auto-detector defaulted to "full" because the issue body signals didn't match any lighter-mode heuristics.
Impact
Each unnecessary "full" run versus "fix" run costs 2 extra agents (researcher-local + researcher) per issue. At 4 false-full detections → 8 unnecessary agent invocations per batch. At 1,000 tokens/agent-minute, this is measurable overhead.
Root Cause
The mode detector likely uses presence-of-signals to select lighter modes (e.g., detect "fix" keywords → use --fix) but falls back to "full" on ambiguity. Small bug fixes that don't explicitly say "fix" in the issue body get classified as full.
Suggested Fix
Strengthen the --fix heuristics in the mode detector:
- Issue title starts with
fix: or fix( or contains "bug fix", "hardening", "guard" → --fix is a strong prior
- Issue body is under N words with no architecture changes → --fix
- Issue references a specific failing test → --fix
- Issue references a specific line/file → --fix
Alternatively, lower the default from "full" to "fix" when no architecture signals are present, requiring explicit signals for "full" (new feature, new agent, new hook, etc.).
Relevant file: plugins/autonomous-dev/lib/ — mode detector / issue classifier (likely pipeline_mode_detector.py or similar)
Plugin Version: 3.50.0 (f6f264c)
Filed automatically by continuous-improvement-analyst
Problem
In the batch of 21 issues (20260521-214143), the coordinator manually overrode auto-detected pipeline modes 5 times:
All 5 were small bug fixes where "full" mode (8 agents including both researchers) was wasteful. The auto-detector defaulted to "full" because the issue body signals didn't match any lighter-mode heuristics.
Impact
Each unnecessary "full" run versus "fix" run costs 2 extra agents (researcher-local + researcher) per issue. At 4 false-full detections → 8 unnecessary agent invocations per batch. At 1,000 tokens/agent-minute, this is measurable overhead.
Root Cause
The mode detector likely uses presence-of-signals to select lighter modes (e.g., detect "fix" keywords → use --fix) but falls back to "full" on ambiguity. Small bug fixes that don't explicitly say "fix" in the issue body get classified as full.
Suggested Fix
Strengthen the --fix heuristics in the mode detector:
fix:orfix(or contains "bug fix", "hardening", "guard" → --fix is a strong priorAlternatively, lower the default from "full" to "fix" when no architecture signals are present, requiring explicit signals for "full" (new feature, new agent, new hook, etc.).
Relevant file:
plugins/autonomous-dev/lib/— mode detector / issue classifier (likelypipeline_mode_detector.pyor similar)Plugin Version: 3.50.0 (f6f264c)
Filed automatically by continuous-improvement-analyst