Skip to content

copilot-centralization-drilldown.md

Latest commit

 

History

History
205 lines (183 loc) · 8.06 KB

copilot-centralization-drilldown.md

File metadata and controls

205 lines (183 loc) · 8.06 KB
name Copilot Centralization Drilldown
description Expands one mined centralization candidate into a concrete draft workflow or reusable prompt template.
true
workflow_dispatch
inputs
candidate_title target_kind recommendation_kind sample_prompt evidence_summary candidate_json
description required
Short name for the candidate skill or repeated task
true
description required default
workflow, shared_prompt, playbook, or auto
false
auto
description required default
Recommendation from the optimizer
false
shared_prompt_or_chatops
description required
Representative prompt text
false
description required
Compact summary of repeated behavior and expected value
false
description required
Optional compact JSON object from the optimizer output
false
permissions
contents copilot-requests
read
write
strict true
max-ai-credits 120
max-daily-ai-credits 500
safe-outputs
mentions allowed-github-references max-bot-mentions create-issue
false
1
title-prefix labels expires
[copilot-centralization-draft]
report
ai-optimization
workflow-design
30d
steps
name env run
Normalize candidate input
CANDIDATE_TITLE TARGET_KIND RECOMMENDATION_KIND SAMPLE_PROMPT EVIDENCE_SUMMARY CANDIDATE_JSON
${{ inputs.candidate_title }}
${{ inputs.target_kind }}
${{ inputs.recommendation_kind }}
${{ inputs.sample_prompt }}
${{ inputs.evidence_summary }}
${{ inputs.candidate_json }}
set -euo pipefail GH_AW_SAFE_OUTPUTS="${GH_AW_SAFE_OUTPUTS:-${RUNNER_TEMP:-/tmp}/gh-aw/safeoutputs/outputs.jsonl}" mkdir -p /tmp/gh-aw/data if [ -n "${CANDIDATE_JSON}" ]; then printf '%s' "$CANDIDATE_JSON" | jq '.' > /tmp/gh-aw/data/candidate-raw.json else jq -n \ --arg title "$CANDIDATE_TITLE" \ --arg recommendation_kind "$RECOMMENDATION_KIND" \ --arg sample_prompt "$SAMPLE_PROMPT" \ --arg evidence_summary "$EVIDENCE_SUMMARY" \ '{ title: $title, recommendation_kind: $recommendation_kind, sample_prompt: (if $sample_prompt == "" then null else $sample_prompt end), evidence_summary: (if $evidence_summary == "" then null else $evidence_summary end) }' > /tmp/gh-aw/data/candidate-raw.json fi candidate_title="$(jq -r '.title // empty' /tmp/gh-aw/data/candidate-raw.json)" sample_prompt="$(jq -r '.sample_prompt // empty' /tmp/gh-aw/data/candidate-raw.json)" evidence_summary="$(jq -r '.evidence_summary // empty' /tmp/gh-aw/data/candidate-raw.json)" if [ -z "$candidate_title" ]; then printf '%s\n' '{"type":"noop","message":"No candidate title was provided for drilldown."}' >> "$GH_AW_SAFE_OUTPUTS" exit 0 fi slug="$(printf '%s' "$candidate_title" \ | tr '[:upper:]' '[:lower:]' \ | sed 's/[^a-z0-9]/-/g' \ | sed 's/-\{2,\}/-/g' \ | sed 's/^-//' \ | sed 's/-$//' \ | cut -c1-48)" if [ -z "$slug" ]; then slug="candidate" fi resolved_target_kind="$TARGET_KIND" if [ "$resolved_target_kind" = "" ] || [ "$resolved_target_kind" = "auto" ]; then case "$RECOMMENDATION_KIND" in continuous_workflow) resolved_target_kind="workflow" ;; shared_prompt_or_chatops) resolved_target_kind="shared_prompt" ;; keep_ad_hoc_but_standardize) resolved_target_kind="playbook" ;; *) resolved_target_kind="workflow" ;; esac fi sample_len=${#sample_prompt} evidence_len=${#evidence_summary} candidate_strength="weak" if [ "$sample_len" -ge 24 ] || [ "$evidence_len" -ge 40 ] || [ -n "$CANDIDATE_JSON" ]; then candidate_strength="strong" fi jq -n \ --arg title "$candidate_title" \ --arg slug "$slug" \ --arg recommendation_kind "$RECOMMENDATION_KIND" \ --arg resolved_target_kind "$resolved_target_kind" \ --arg sample_prompt "$sample_prompt" \ --arg evidence_summary "$evidence_summary" \ --arg candidate_strength "$candidate_strength" \ --slurpfile raw /tmp/gh-aw/data/candidate-raw.json ' { title: $title, slug: $slug, recommendation_kind: $recommendation_kind, resolved_target_kind: $resolved_target_kind, candidate_strength: $candidate_strength, sample_prompt: (if $sample_prompt == "" then null else $sample_prompt end), evidence_summary: (if $evidence_summary == "" then null else $evidence_summary end), raw_candidate: $raw[0] } ' > /tmp/gh-aw/data/candidate.json jq -n \ --arg resolved_target_kind "$resolved_target_kind" \ --arg slug "$slug" ' { target_path: ( if $resolved_target_kind == "workflow" then ".github/workflows/" + $slug + ".md" elif $resolved_target_kind == "shared_prompt" then ".github/workflows/shared/" + $slug + ".md" else ".github/workflows/shared/" + $slug + "-playbook.md" end ), trigger_hint: ( if $resolved_target_kind == "workflow" then "Prefer workflow_dispatch first; choose schedule or event triggers only when repeated evidence justifies automation." elif $resolved_target_kind == "shared_prompt" then "Prefer a reusable prompt or shared workflow component before full automation." else "Keep this human-in-the-loop and produce a reusable playbook or prompt template." end ), implementation_bias: "Prefer the smallest durable artifact that reduces repeated prompting without widening scope.", report_style: "Use issue sections with visible summary and one fenced draft block." } ' > /tmp/gh-aw/data/derived-plan.json

Copilot Centralization Drilldown

Turn one mined centralization candidate into a concrete, reviewable draft.

Read these prepared files first:

  • /tmp/gh-aw/data/candidate.json
  • /tmp/gh-aw/data/derived-plan.json

Stay narrow. Do not perform broad repo exploration. Use the current repository's existing workflow style as the default convention.

If candidate_strength is weak, or the provided evidence is too vague to justify a concrete draft, call noop with a short explanation.

Task

  1. Decide the smallest useful deliverable:
    • a new gh-aw workflow draft
    • a reusable shared prompt/template draft
    • a reusable human playbook when automation is premature
  2. Produce exactly one primary draft aligned with resolved_target_kind unless the evidence clearly supports a better smaller artifact.
  3. Optimize for AI-credit savings by reducing repeated prompt construction, ambiguous setup work, or repeated context rebuilding.
  4. Keep the draft implementation-ready and short. Do not design a platform.

Output

Use create-issue.

Structure the issue with these ### sections:

  • ### Summary
  • ### Pattern Fit
  • ### Proposed Draft
  • ### AI Credit Savings Rationale
  • ### Inputs Still Needed
  • ### References

In ### Proposed Draft:

  • include Path: <target path> on its own line before the draft
  • include exactly one fenced md code block containing the full proposed file content
  • if the best artifact is a workflow, produce a full gh-aw workflow markdown file with frontmatter and prompt body
  • if the best artifact is a shared prompt or playbook, produce the full markdown file content for that path

Constraints

  • Prefer workflow_dispatch for a first workflow draft unless the evidence strongly supports a scheduled or event-driven trigger.
  • Keep permissions read-only and route visible writes through safe outputs.
  • Reuse deterministic preprocessing or explicit noop guidance when that materially lowers token use.
  • Do not create more than one draft artifact in the issue.