fix: TaskOutput-timeout guard, actor-permission preflight, report deps

action.yml

@@ -91,6 +91,26 @@ runs:
           exit 1
         fi
 
+    - name: Verify triggering actor has write access
+      shell: bash
+      run: |
+        ACTOR="${{ github.event.sender.login }}"
+        if [ -z "$ACTOR" ]; then
+          echo "::warning::Could not determine the triggering actor (github.event.sender.login was empty) — skipping the write-access preflight. The underlying claude-code-action will still enforce this itself."
+          exit 0
+        fi
+        PERM=$(gh api "repos/${{ github.repository }}/collaborators/${ACTOR}/permission" --jq '.permission' 2>/dev/null || echo "unknown")
+        echo "Triggering actor: $ACTOR (permission: $PERM)"
+        case "$PERM" in
+          admin|write|maintain) ;;
+          *)
+            echo "::error::@${ACTOR} triggered this review but does not have write access to this repository (permission: ${PERM}). claude-code-action requires write access to run safely and will reject this with an opaque error after ~40s of setup — failing fast instead. Ask someone with write access to (re-)apply the '${{ inputs.trigger_label }}' label."
+            exit 1
+            ;;
+        esac
+      env:
+        GH_TOKEN: ${{ inputs.github_token }}
+
     - name: Pre-flight checks
       shell: bash
       run: |
@@ -132,6 +152,19 @@ runs:
       shell: bash
       run: git config --global url."https://github.com/".insteadOf "git@github.com:"
 
+    - name: Install report-rendering dependencies
+      shell: bash
+      run: |
+        # claudius's report scripts (consolidate_reports.py, validate_report.py,
+        # generate_review_report.py) expect these as system packages — see the
+        # claudius CLAUDE.md "Python deps come from apt" convention. Runner images
+        # don't have them; installing here avoids a mid-review crash on `import markdown`
+        # (or jsonschema/jinja2/etc.) inside the sandboxed Claude session.
+        sudo apt-get update -qq
+        sudo apt-get install -y -qq \
+          python3-markdown python3-jsonschema python3-yaml python3-jinja2 \
+          python3-bs4 python3-nh3 python3-markupsafe python3-reportlab python3-matplotlib
+
     - name: Create report directory
       shell: bash
       run: |
@@ -172,6 +205,15 @@ runs:
           comments) will silently never happen. Do not end your turn until report.json has
           actually been written and findings have been posted.
 
+          TaskOutput can time out and return before an agent is actually done (you'll see
+          something like status:running in its result instead of a final answer) — that is NOT
+          completion. On a timeout, call TaskOutput again for the same task_id and keep re-blocking
+          until you get the agent's real result. Never treat a timed-out TaskOutput call as
+          "collected" and move on. This rule applies transitively: any specialist agent you spawn
+          (e.g. via claudius:grumpy-review) that itself spawns further sub-agents is bound by the
+          same rule — tell it so explicitly in its own spawn prompt, since it has no way to know
+          it's running headless otherwise.
+
           Do NOT create temporary directories (mktemp). Use REPORT_DIR env var for report output.
 
           Write all PR comments in Claudius persona — witty, confident, subtly snarky,
@@ -258,6 +300,7 @@ runs:
       shell: bash
       run: |
         gh pr edit "${{ github.event.pull_request.number }}" \
-          --remove-label "${{ inputs.trigger_label }}" 2>/dev/null || true
+          --remove-label "${{ inputs.trigger_label }}" \
+          || echo "::warning::Failed to remove the '${{ inputs.trigger_label }}' label. It will stay on the PR, which will re-trigger a full review on the next push — check that the token has label-write access."
       env:
         GH_TOKEN: ${{ inputs.github_token }}