Skip to content

ci: harden AI triage (idempotent, non-clobbering, tighter labels) #85

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
duzos merged 1 commit into from
Jun 7, 2026
Merged

ci: harden AI triage (idempotent, non-clobbering, tighter labels) #85

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
186 changes: 107 additions & 79 deletions .github/workflows/issue-triage.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,11 +1,18 @@
name: "Triage: AI"

# AI-assisted issue triage using GitHub Models (actions/ai-inference).
# On a new issue it asks a model to classify the issue against the AmbleKit
# label taxonomy, then applies labels, sets the issue type, sets the org
# Priority/Effort fields (best effort), and leaves a triage comment that
# includes a likely-cause (bugs) / recommended-implementation (features) note.
# If the model output can't be parsed it falls back to "S: Untriaged".
# On a new/reopened issue (or manual dispatch) it asks a model to classify the
# issue against the AmbleKit label taxonomy, applies labels, sets the org
# Priority/Effort fields (best effort), and posts a triage comment with a
# likely-cause (bugs) / recommended-implementation (features) note.
#
# Idempotent: the comment carries a hidden <!-- auto-triage --> marker, so a
# re-run edits the existing comment instead of stacking a new one. On a re-run
# of an already-triaged issue it refreshes only the comment and leaves labels
# untouched, so it never clobbers a maintainer's later label edits.
#
# If the model output can't be parsed it falls back to "S: Untriaged" (only on
# the first triage).

on:
issues:
Expand Down Expand Up @@ -102,7 +109,7 @@ jobs:
"effort": "High" | "Medium" | "Low",
"difficulty": "DB" | "D3" | "D2" | "D1" | "D0",
"areas": [ subset of: "A: Suits","A: Powers","A: Entities","A: Datagen","A: Admin Tooling","A: Core Tech","A: Build","A: Docs" ],
"changes": [ subset of: "C: Render","C: UI","C: Networking","C: Textures","C: Audio","C: Translations","C: Structures","C: No Java" ],
"changes": [ subset of: "C: Render","C: UI","C: Networking","C: Textures","C: Audio","C: Translations" ],
"summary": "one short lowercase sentence",
"insight": "for a Bug: 2-4 sentences on the most likely cause, naming the suspect subsystem/class/package; for a Feature: 2-4 sentences on a recommended implementation approach, pointing at where in the codebase it would go; for a Task: empty string",
"needs_info": true | false
Expand All @@ -128,6 +135,14 @@ jobs:
run: |
set -uo pipefail

MARKER="<!-- auto-triage -->"
# find our previous triage comment (if any). its presence means this issue
# was already triaged, so a re-run refreshes the comment but leaves labels
# and fields alone - never clobbering a maintainer's later edits.
CID=$(gh api "repos/$REPO/issues/$NUM/comments" --paginate \
--jq "[.[] | select(.body | contains(\"$MARKER\")) | .id] | last // empty" 2>/dev/null || true)
FIRST_TRIAGE=true; [ -n "$CID" ] && FIRST_TRIAGE=false

Comment thread
duzos marked this conversation as resolved.
# prefer the whole response if it already parses as json (expected case:
# the model is asked for one line of minified json). otherwise strip any
# code fences and fall back to the first {...} block.
Expand All @@ -139,15 +154,15 @@ jobs:
echo "model json: $JSON"

if [ -z "$JSON" ] || ! printf '%s' "$JSON" | jq -e . >/dev/null 2>&1; then
echo "no valid json -> S: Untriaged"
gh issue edit "$NUM" --repo "$REPO" --add-label "S: Untriaged" || true
if [ "$FIRST_TRIAGE" = "true" ]; then
echo "no valid json -> S: Untriaged"
gh issue edit "$NUM" --repo "$REPO" --add-label "S: Untriaged" || true
else
echo "no valid json on re-triage -> leaving already-triaged issue untouched"
fi
exit 0
fi

# a previous run may have left S: Untriaged - clear it now that we have a
# valid classification so the issue isn't both triaged and untriaged.
gh issue edit "$NUM" --repo "$REPO" --remove-label "S: Untriaged" >/dev/null 2>&1 || true

TYPE=$(echo "$JSON" | jq -r '.type // empty')
PRIORITY=$(echo "$JSON" | jq -r '.priority // empty')
EFFORT=$(echo "$JSON" | jq -r '.effort // empty')
Expand All @@ -156,75 +171,79 @@ jobs:
INSIGHT=$(echo "$JSON" | jq -r '.insight // empty' | tr '\r\n' ' ' | cut -c1-1200)
NEEDS=$(echo "$JSON" | jq -r '.needs_info // false')

LABELS=()
case "$TYPE" in
Bug) LABELS+=("T: Bugfix");;
Feature) LABELS+=("T: New Feature");;
# Task has no T: label in the taxonomy - it's conveyed by the GitHub
# issue type set further down, so no label is added here on purpose.
esac
case "$PRIORITY" in
Urgent) LABELS+=("P0: Critical");;
High) LABELS+=("P1: High");;
Medium) LABELS+=("P2: Raised");;
Low) LABELS+=("P3: Standard");;
esac
case "$DIFF" in
DB) LABELS+=("DB: Beginner Friendly");;
D3) LABELS+=("D3: Low");;
D2) LABELS+=("D2: Medium");;
D1) LABELS+=("D1: High");;
D0) LABELS+=("D0: Very High");;
esac
# allowlist the model's area/change labels to the documented taxonomy so a
# stray (but repo-existing) label can't sneak in if the model goes off-schema.
declare -A ALLOWED_AC=(
["A: Suits"]=1 ["A: Powers"]=1 ["A: Entities"]=1 ["A: Datagen"]=1
["A: Admin Tooling"]=1 ["A: Core Tech"]=1 ["A: Build"]=1 ["A: Docs"]=1
["C: Render"]=1 ["C: UI"]=1 ["C: Networking"]=1 ["C: Textures"]=1
["C: Audio"]=1 ["C: Translations"]=1 ["C: Structures"]=1 ["C: No Java"]=1
)
while IFS= read -r l; do
[ -n "$l" ] && [ -n "${ALLOWED_AC[$l]:-}" ] && LABELS+=("$l")
done < <(echo "$JSON" | jq -r '((.areas // []) + (.changes // []))[]')
[ "$NEEDS" = "true" ] && LABELS+=("Issue: Awaiting Response")

# only apply labels that actually exist in the repo - page through all of
# them into a set for O(1) membership instead of a nested scan.
declare -A HAVE
while IFS= read -r e; do [ -n "$e" ] && HAVE["$e"]=1; done \
< <(gh api "repos/$REPO/labels" --paginate --jq '.[].name')
ADD=()
for l in "${LABELS[@]:-}"; do
[ -n "${HAVE[$l]:-}" ] && ADD+=("--add-label" "$l")
done
if [ "${#ADD[@]}" -gt 0 ]; then
gh issue edit "$NUM" --repo "$REPO" "${ADD[@]}" || true
fi
# labels + fields are only touched on the FIRST triage. on a re-run we keep
# whatever the maintainers have set and just refresh the comment below.
if [ "$FIRST_TRIAGE" = "true" ]; then
# clear S: Untriaged now that we have a valid classification.
gh issue edit "$NUM" --repo "$REPO" --remove-label "S: Untriaged" >/dev/null 2>&1 || true

LABELS=()
case "$TYPE" in
Bug) LABELS+=("T: Bugfix");;
Feature) LABELS+=("T: New Feature");;
# Task has no T: label in the taxonomy on purpose.
esac
case "$PRIORITY" in
Urgent) LABELS+=("P0: Critical");;
High) LABELS+=("P1: High");;
Medium) LABELS+=("P2: Raised");;
Low) LABELS+=("P3: Standard");;
esac
case "$DIFF" in
DB) LABELS+=("DB: Beginner Friendly");;
D3) LABELS+=("D3: Low");;
D2) LABELS+=("D2: Medium");;
D1) LABELS+=("D1: High");;
D0) LABELS+=("D0: Very High");;
esac
# allowlist the model's area/change labels. note: diff-derived changes
# (C: No Java, C: Structures) are intentionally NOT here - those are decided
# from a PR's changed files by .github/labeler.yml, not guessable from an issue.
declare -A ALLOWED_AC=(
["A: Suits"]=1 ["A: Powers"]=1 ["A: Entities"]=1 ["A: Datagen"]=1
["A: Admin Tooling"]=1 ["A: Core Tech"]=1 ["A: Build"]=1 ["A: Docs"]=1
["C: Render"]=1 ["C: UI"]=1 ["C: Networking"]=1 ["C: Textures"]=1
["C: Audio"]=1 ["C: Translations"]=1
)
while IFS= read -r l; do
[ -n "$l" ] && [ -n "${ALLOWED_AC[$l]:-}" ] && LABELS+=("$l")
done < <(echo "$JSON" | jq -r '((.areas // []) + (.changes // []))[]')
[ "$NEEDS" = "true" ] && LABELS+=("Issue: Awaiting Response")

# only apply labels that actually exist in the repo - page through all of
# them into a set for O(1) membership instead of a nested scan.
declare -A HAVE
while IFS= read -r e; do [ -n "$e" ] && HAVE["$e"]=1; done \
< <(gh api "repos/$REPO/labels" --paginate --jq '.[].name')
ADD=()
for l in "${LABELS[@]:-}"; do
[ -n "${HAVE[$l]:-}" ] && ADD+=("--add-label" "$l")
done
if [ "${#ADD[@]}" -gt 0 ]; then
gh issue edit "$NUM" --repo "$REPO" "${ADD[@]}" || true
fi

# GitHub issue type (best effort - org-level)
if [ -n "$TYPE" ]; then
gh api --method PATCH "repos/$REPO/issues/$NUM" -f type="$TYPE" >/dev/null 2>&1 \
|| echo "issue type not set (token may lack org perms)"
# org Priority/Effort issue fields (best effort)
RID=$(gh api "repos/$REPO" --jq '.id')
set_field() {
[ -z "$2" ] && return 0
# build the payload with jq so the value is properly json-escaped
jq -n --argjson fid "$1" --arg val "$2" \
'{issue_field_values:[{field_id:$fid,value:$val}]}' \
| gh api --method POST "repositories/$RID/issues/$NUM/issue-field-values" --input - >/dev/null 2>&1 \
|| echo "field $1 not set"
}
# ids default to this org's Priority/Effort fields; override via the
# PRIORITY_FIELD_ID / EFFORT_FIELD_ID Actions variables if they differ.
set_field "${PRIORITY_FIELD_ID:-30052706}" "$PRIORITY"
set_field "${EFFORT_FIELD_ID:-30052709}" "$EFFORT"
else
echo "re-triage: refreshing comment only, leaving labels/fields as-is"
fi

# org Priority/Effort issue fields (best effort)
RID=$(gh api "repos/$REPO" --jq '.id')
set_field() {
[ -z "$2" ] && return 0
# build the payload with jq so the value is properly json-escaped
jq -n --argjson fid "$1" --arg val "$2" \
'{issue_field_values:[{field_id:$fid,value:$val}]}' \
| gh api --method POST "repositories/$RID/issues/$NUM/issue-field-values" --input - >/dev/null 2>&1 \
|| echo "field $1 not set"
}
# ids default to this org's Priority/Effort fields; override via the
# PRIORITY_FIELD_ID / EFFORT_FIELD_ID Actions variables if they differ.
set_field "${PRIORITY_FIELD_ID:-30052706}" "$PRIORITY"
set_field "${EFFORT_FIELD_ID:-30052709}" "$EFFORT"

# triage comment
# triage comment (hidden marker first so re-runs can find + edit it)
{
echo "$MARKER"
echo "🤖 **auto-triage** (github models)"
echo
echo "| type | priority | effort | difficulty |"
Expand All @@ -243,8 +262,17 @@ jobs:
echo "$INSIGHT"
echo
fi
[ "$NEEDS" = "true" ] && echo "looks like it's missing repro steps / logs - added \`Issue: Awaiting Response\`."
[ "$NEEDS" = "true" ] && [ "$FIRST_TRIAGE" = "true" ] && echo "looks like it's missing repro steps / logs - added \`Issue: Awaiting Response\`."
[ "$FIRST_TRIAGE" = "false" ] && echo "_(re-triage: labels left as set by maintainers)_"
echo
echo "_labels and insight are a best guess, a maintainer will confirm._"
} > "$RUNNER_TEMP/comment.md"
gh issue comment "$NUM" --repo "$REPO" --body-file "$RUNNER_TEMP/comment.md" || true

if [ -n "$CID" ]; then
jq -Rs '{body: .}' "$RUNNER_TEMP/comment.md" \
| gh api --method PATCH "repos/$REPO/issues/comments/$CID" --input - >/dev/null \
&& echo "updated existing triage comment $CID" || echo "comment update failed"
else
gh issue comment "$NUM" --repo "$REPO" --body-file "$RUNNER_TEMP/comment.md" >/dev/null \
&& echo "posted triage comment" || true
fi
Loading