You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The Prompt Clustering Analysis workflow's clean_prompt() helper (around line 380 of the workflow .md) strips markdown code blocks, START COPILOT preambles, URLs, and special characters — but doesn't strip recurring bot-footer signatures embedded in PR bodies before TF-IDF vectorization. This cycle's Prompt Clustering Analysis — 2026-08-29 run flagged its own Cluster 1 as an artifact: 63 PRs (5.5% of the 1,140 analyzed) clustered together purely because of a repeated "PR Sous Chef ... Comment /souschef to run again" bot footer, not a coherent task theme. The same class of noise likely affects other bot-authored footers (e.g. <!-- gh-aw-agentic-workflow: ... --> HTML comments).
Expected Impact
Strips noise before vectorization so cluster assignments reflect actual task content rather than bot signatures — removing at least one full noise cluster (currently ~5.5% of all analyzed PRs) and improving keyword/theme accuracy for the remaining clusters.
Suggested Fix
In clean_prompt(), add a stripping step (before or alongside the existing markdown-code-block removal) for:
<!--\s*gh-aw-agentic-workflow:.*?--> style HTML comment footers
Known recurring bot-footer text patterns (e.g. "PR Sous Chef", "Comment /souschef to run again") — ideally via a small, extensible list/regex rather than one-off string matches, since new bot footers will appear over time.
Suggested Agent
prompt-clustering-analysis workflow author / New Agent
Description
The Prompt Clustering Analysis workflow's
clean_prompt()helper (around line 380 of the workflow.md) strips markdown code blocks,START COPILOTpreambles, URLs, and special characters — but doesn't strip recurring bot-footer signatures embedded in PR bodies before TF-IDF vectorization. This cycle's Prompt Clustering Analysis — 2026-08-29 run flagged its own Cluster 1 as an artifact: 63 PRs (5.5% of the 1,140 analyzed) clustered together purely because of a repeated "PR Sous Chef ... Comment /souschef to run again" bot footer, not a coherent task theme. The same class of noise likely affects other bot-authored footers (e.g.<!-- gh-aw-agentic-workflow: ... -->HTML comments).Expected Impact
Strips noise before vectorization so cluster assignments reflect actual task content rather than bot signatures — removing at least one full noise cluster (currently ~5.5% of all analyzed PRs) and improving keyword/theme accuracy for the remaining clusters.
Suggested Fix
In
clean_prompt(), add a stripping step (before or alongside the existing markdown-code-block removal) for:<!--\s*gh-aw-agentic-workflow:.*?-->style HTML comment footersSuggested Agent
prompt-clustering-analysis workflow author / New Agent
Estimated Effort
Quick (< 1 hour)
Data Source
DeepReport Intelligence Briefing analysis, 2026-08-29 cycle (window: discussions #56860–#56877, since prior baseline #56856). Source discussion: "Prompt Clustering Analysis - 2026-08-29" (#56873).
Warning
Firewall blocked 1 domain
The following domain was blocked by the firewall during workflow execution:
api.anthropic.comTo allow these domains, add them to the
network.allowedlist in your workflow frontmatter:See Network Configuration for more information.