@plainconceptsplatform/workflows 1.14.0 → 1.14.1

package/loops/actions/validate-refine-output/validate-refine-output.sh

@@ -29,19 +29,31 @@ jq -r --arg marker "$marker" --arg prefix "$comment_prefix" --arg issue "$issue_
     any(.items[]; .type == "update_issue");
 
   # A temporal draft is a replacement body the worker still expects to grow: the draft
-  # marker distinguishes it from a finished body.
+  # marker distinguishes it from a finished body. The gh-aw process_safe_outputs strips
+  # HTML comments from the body, so the draft marker (<!-- agent-refine-draft -->) can
+  # be gone by the time the validator reads it (Bots #1336, run 42214146974). When the
+  # draft marker is absent but the pending mark (_pending see questions below_) is
+  # present, the body is still a draft: the pending mark only appears in drafts, and a
+  # finished body has none.
   def has_draft_body:
     any(.items[]; .type == "update_issue" and
       (.item_number == null or (.item_number | tostring) == $issue) and
       (.body | type == "string") and
-      (.body | contains($draft_marker)));
+      ((.body | contains($draft_marker)) or (.body | contains($pending_mark))));
 
   # The batched questions, by the sentence that opens them. It decides the outcome on its own:
   # Bots #561, #871 and Numa #821 asked, left a draft without the draft marker, were read as
   # complete, and implement started two seconds later on a story nobody had answered.
+  # The prompt asks the model to open with the exact questions_line, but models paraphrase
+  # it: GLM-5-2 wrote "I have some questions that the code cannot settle on its own" and the
+  # exact-line match failed, dropping a valid questions outcome (Bots #1336). Match the exact
+  # line, or its opening "I have some questions" prefix, which is specific enough in a refine
+  # comment to never match the complete or split messages.
   def asks_questions:
     any(.items[]; .type == "add_comment" and (.item_number | tostring) == $issue and
-      (.body | type == "string") and (.body | contains($questions_line)));
+      (.body | type == "string") and
+      ((.body | contains($questions_line)) or
+       (.body | test("I have some questions"; "i"))));
 
   def has_pending_marks:
     any(.items[]; .type == "update_issue" and

package/loops/actions/validate-review-output/validate-review-output.sh

@@ -24,7 +24,10 @@ jq -r --arg pr "$pr_number" --slurpfile threads "$threads_file" '
   .items as $items
   | ($threads[0] | map(select(.isResolved == false and .isOutdated == false) | .id) | sort) as $expected
   | ($items | [.[] | select(.type == "add_comment" and (.item_number | tostring) == $pr and (.body | type == "string"))]) as $comments
-  | ($comments | map(.body | if test("\\*\\*Review outcome:\\*\\*\\s*(implemented|already-satisfied|needs-human)"; "i") then capture("\\*\\*Review outcome:\\*\\*\\s*(?<v>implemented|already-satisfied|needs-human)"; "i").v | ascii_downcase else empty end) | .[0] // "invalid") as $outcome
+  # The line is matched by shape, not exact markdown: a model that writes
+  # **Review outcome** : implemented or **Review outcome - implemented still means
+  # the same. See the triage validator for the same pattern.
+  | ($comments | map(.body | if test("Review outcome[^\\w\\n]{0,6}(?<v>implemented|already-satisfied|needs-human)\\b"; "i") then capture("Review outcome[^\\w\\n]{0,6}(?<v>implemented|already-satisfied|needs-human)\\b"; "i").v | ascii_downcase else empty end) | .[0] // "invalid") as $outcome
   # `add` on an empty array is null, not [], and `unique` then dies with "Cannot iterate over
   # null". That is not an edge case: it is every pull-request-level review. Requesting changes
   # from the Files tab, or a plain review body with no inline comment, creates a review and no

package/loops/actions/validate-triage-output/validate-triage-output.sh

@@ -20,12 +20,15 @@ fi
 # must contain a verdict line matching **Verdict:** pass|needs-info|needs-maintainer|block.
 # capture() returns an object use a named group (?<v>...) to extract the value.
 # test() before capture() avoids jq errors on non-matching bodies.
+# The line is matched by shape, not exact markdown: a model that writes
+# **Verdict** : pass or **Verdict - pass still means the same (Bots #1336 showed
+# GLM-5-2 paraphrasing the expected sentence; the same happens with verdict lines).
 jq -r --arg issue "$issue_number" '
   def extract_verdict:
     [.items[] | select(.type == "add_comment" and (.item_number | tostring) == $issue and (.body | type == "string"))]
     | map(.body |
-      if test("\\*\\*Verdict:\\*\\*\\s*(pass|needs-info|needs-maintainer|block)"; "i") then
-        capture("\\*\\*Verdict:\\*\\*\\s*(?<v>pass|needs-info|needs-maintainer|block)"; "i").v | ascii_downcase
+      if test("Verdict[^\\w\\n]{0,6}(?<v>pass|needs-info|needs-maintainer|block)\\b"; "i") then
+        capture("Verdict[^\\w\\n]{0,6}(?<v>pass|needs-info|needs-maintainer|block)\\b"; "i").v | ascii_downcase
       else empty end
     )
     | .[0] // "none";

package/loops/actions/verify-refine-output/verify-refine-output.sh

@@ -75,6 +75,16 @@ assert_output 'complete output cannot update another issue' invalid \
   '{"items":[{"type":"update_issue","item_number":42,"body":"# User story"},{"type":"update_issue","item_number":7,"body":"# Other story"}]}'
 assert_output 'draft update with questions comment is questions' questions \
   '{"items":[{"type":"update_issue","item_number":42,"body":"<!-- agent-refine-draft -->\n### Proposal\n_pending see questions below_"},{"type":"add_comment","item_number":42,"body":"<!-- agent-refine -->\nRefinement update\nI have some questions about this issue. Please reply in one comment and I''ll process your answers.\nWhich area owns this behavior?"}]}'
+
+# From Bots #1336, run 42214146974. gh-aw process_safe_outputs strips HTML comments
+# from the body, so the draft marker (<!-- agent-refine-draft -->) is gone. The
+# model also paraphrased the questions line as "I have some questions that the code
+# cannot settle on its own" instead of the exact sentence. The pending mark
+# (_pending see questions below_) survives in the body and the comment is still on
+# the source issue, so the outcome is questions, not invalid.
+assert_output 'stripped draft marker with paraphrased questions is questions' questions \
+  '{"items":[{"type":"update_issue","item_number":42,"body":"**Complexity:** standard path\n\n> Refinement draft. Three open questions are in a separate comment. The parts they leave open are marked `_pending see questions below_`."},{"type":"add_comment","item_number":42,"body":"## Refinement update\n\nI have some questions that the code cannot settle on its own. Reply in one comment to change any of them.\n1. Does the conductor bot stay local only too?"}]}'
+
 assert_output 'draft update alone is invalid' invalid \
   '{"items":[{"type":"update_issue","item_number":42,"body":"<!-- agent-refine-draft -->\n### Proposal\n_pending see questions below_"}]}'
 

package/loops/workflows/agent-refine.md

@@ -738,7 +738,20 @@ jobs:
             output_tokens=$(jq -r '.output_tokens // empty' usage-read/agent_usage.json 2>/dev/null || echo "")
           fi
           if [ -n "$output_tokens" ] && [ "$output_tokens" -gt 20000 ]; then
-            echo "truncated=This attempt consumed ${output_tokens} output tokens and still produced no outcome. That is what provider output truncation looks like, not an outage: the model hit the single-response cap mid-generation and every call it was writing was lost. If this repeats, split the issue or tighten its body." >> "$GITHUB_OUTPUT"
+            # A run that burned tokens and produced items the validator refused is not
+            # truncation: the agent finished, and the validator did not recognise the
+            # outcome shape (Bots #1336, run 42214146974: 50926 output tokens, 2 items,
+            # draft marker stripped by process_safe_outputs, questions line paraphrased).
+            # Only call it truncation when the agent spent tokens and produced nothing.
+            item_count=0
+            if [ -f usage-read/agent_output.json ]; then
+              item_count=$(jq '.items | length' usage-read/agent_output.json 2>/dev/null || echo 0)
+            fi
+            if [ "$item_count" -gt 0 ]; then
+              echo "truncated=This attempt consumed ${output_tokens} output tokens and produced ${item_count} item(s), but the validator could not recognise the outcome. The agent finished its work: check the run artifacts (agent_output.json, safeoutputs.jsonl) against the validator's expected shape before retrying." >> "$GITHUB_OUTPUT"
+            else
+              echo "truncated=This attempt consumed ${output_tokens} output tokens and still produced no outcome. That is what provider output truncation looks like, not an outage: the model hit the single-response cap mid-generation and every call it was writing was lost. If this repeats, split the issue or tighten its body." >> "$GITHUB_OUTPUT"
+            fi
           else
             echo "truncated=" >> "$GITHUB_OUTPUT"
           fi

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@plainconceptsplatform/workflows",
-  "version": "1.14.0",
+  "version": "1.14.1",
   "description": "Install and update Platform GitHub agentic workflows.",
   "keywords": [
     "github-actions",