fix(agentic-engineering): define classifier output contract

.claude-plugin/marketplace.json

@@ -27,7 +27,7 @@
     {
       "name": "agentic-engineering",
       "description": "The autonomous engineering system for repository portfolios — engineer, read-only surveyor, and meta-engineer agents; portfolio, product, spend, and improvement workflows; cross-tool instruction architecture and skill discovery; configured by the consumer AGENTS.md",
-      "version": "5.1.4",
+      "version": "5.1.5",
       "source": "./plugins/agentic-engineering"
     },
     {

.github/plugin/marketplace.json

@@ -27,7 +27,7 @@
     {
       "name": "agentic-engineering",
       "description": "The autonomous engineering system for repository portfolios — engineer, read-only surveyor, and meta-engineer agents; portfolio, product, spend, and improvement workflows; cross-tool instruction architecture and skill discovery; configured by the consumer AGENTS.md",
-      "version": "5.1.4",
+      "version": "5.1.5",
       "source": "./plugins/agentic-engineering"
     },
     {

plugins/agentic-engineering/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agentic-engineering",
   "description": "The autonomous engineering system for repository portfolios — engineer, read-only surveyor, and meta-engineer agents; portfolio, product, spend, and improvement workflows; cross-tool instruction architecture and skill discovery; configured by the consumer AGENTS.md",
-  "version": "5.1.4",
+  "version": "5.1.5",
   "author": {
     "name": "devantler-tech",
     "url": "https://github.com/devantler-tech"

plugins/agentic-engineering/agents/portfolio-surveyor.agent.md

@@ -502,6 +502,21 @@ by a successful consumer of partial output and the read-only role never writes a
 The bundled `forge-readonly-guard.sh` recognises only this exact installed sibling; offline `--input`
 remains denied. Exit 0 is a complete classification; exit 2 means `unknown`, never green.
 
+**Read the verdict from the helper's native tool result, never by appending the guard-denied `; echo
+"EXIT=$?"` idiom.** The guard rejects shell chaining before the helper runs, while the tool result
+already surfaces stdout, stderr, and the process status. Classify that complete result with this closed
+set:
+
+| Native process status and output | Meaning |
+|---|---|
+| observed native process status 0 and completely empty output | no red runs → that branch is **green** |
+| observed native process status 0 and **well-formed TSV rows** — exactly eight tab-separated fields in helper order: numeric `workflow_id`, red `conclusion` (`failure`, `timed_out`, or `startup_failure`), `html_url`, `name`, supported `event`, `path`, valid `created_at`, numeric `run_id` | those are the **red runs** |
+| any nonzero or unavailable native process status; or any other output, including mixed valid and malformed rows | the helper FAILED → **`QUERY-UNKNOWN`**; never `nothing_on_fire: true` |
+
+Every nonempty line must match the complete eight-field row shape. Do not accept a numeric first
+field, a diagnostic prefix, or any other partial predicate as sufficient, and never treat merely
+nonempty output as red runs. Status 0 plus empty output is the only green case.
+
 Invoke it from a process that already has `GH_TELEMETRY=0` (or `false`) in the environment,
 or rely on the helper's own export of `GH_TELEMETRY=0` before its remote `gh api` GET — GitHub CLI
 2.96.0 otherwise writes `gh/device-id` on a certified read. Do not prefix the helper with

plugins/agentic-engineering/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agentic-engineering",
   "description": "The autonomous engineering system for repository portfolios — engineer, read-only surveyor, and meta-engineer agents; portfolio, product, spend, and improvement workflows; cross-tool instruction architecture and skill discovery; configured by the consumer AGENTS.md",
-  "version": "5.1.4",
+  "version": "5.1.5",
   "author": {
     "name": "devantler-tech",
     "url": "https://github.com/devantler-tech"

plugins/agentic-engineering/resources/provider-neutral.desired-state.json

@@ -66,7 +66,7 @@
       "portfolio-surveyor": {
         "enabled": true,
         "mode": "delegated-read-only",
-        "definitionSha256": "53dfcda04e4e404b6d13fb465aa76c574a79401d1b2be1058beae1ccc9bc6425"
+        "definitionSha256": "8e31a79bb60e0c68f92f95c5f4fbf2abda78e719dd40c04392f9e2f74de91275"
       },
       "agent-improver": {
         "enabledWhen": "Both optional consumer contract sections are present",

plugins/agentic-engineering/scripts/portfolio-surveyor-agent.test.sh

@@ -23,6 +23,26 @@ if grep -Fq -- '--json issueType,blockedBy,assignees' "$SURVEYOR"; then
   fail 'dependency reads must not enumerate blocker nodes through gh issue view'
 fi
 
+CI_STEP=$(sed -n \
+  '/^### 4\. CI red on the default branch/,/^### 5\. Triage, stale, and advance signals/p' \
+  "$SURVEYOR" | tr '\n' ' ' | tr -s '[:space:]' ' ')
+[ -n "$CI_STEP" ] || fail 'could not extract the default-branch CI step'
+# shellcheck disable=SC2016 # The asserted agent text contains a literal shell idiom.
+grep -Fq 'Read the verdict from the helper'"'"'s native tool result, never by appending the guard-denied `; echo "EXIT=$?"` idiom.' \
+  <<<"$CI_STEP" ||
+  fail 'default-branch CI must prescribe the native tool result instead of denied shell exit capture'
+grep -Fq '| observed native process status 0 and completely empty output | no red runs → that branch is **green** |' \
+  <<<"$CI_STEP" ||
+  fail 'green must require both observed native process status 0 and completely empty classifier output'
+# shellcheck disable=SC2016 # Backticks belong to the asserted Markdown contract.
+grep -Fq '| observed native process status 0 and **well-formed TSV rows** — exactly eight tab-separated fields in helper order: numeric `workflow_id`, red `conclusion` (`failure`, `timed_out`, or `startup_failure`), `html_url`, `name`, supported `event`, `path`, valid `created_at`, numeric `run_id` | those are the **red runs** |' \
+  <<<"$CI_STEP" ||
+  fail 'the complete eight-field TSV predicate must map to the red verdict as one table row'
+# shellcheck disable=SC2016 # Backticks belong to the asserted Markdown contract.
+grep -Fq '| any nonzero or unavailable native process status; or any other output, including mixed valid and malformed rows | the helper FAILED → **`QUERY-UNKNOWN`**; never `nothing_on_fire: true` |' \
+  <<<"$CI_STEP" ||
+  fail 'nonzero or unavailable status and malformed or mixed output must fail closed as QUERY-UNKNOWN'
+
 JQ_FILTER=$(sed -n \
   "/issueDependenciesSummary{blockedBy totalBlockedBy}/{n;s/^[[:space:]]*--jq '\\(.*\\)'$/\\1/p;}" \
   "$SURVEYOR")