fix(actions): report an outage when the run fails ahead of the agent

claude/action.yaml

@@ -115,6 +115,7 @@ runs:
     # and runs the script (reached via ../shared/steps, since this action is
     # one level deep).
     - name: Security preflight
+      id: security
       shell: bash
       run: bash "${{ github.action_path }}/../shared/steps/security-preflight.sh"
       env:
@@ -607,8 +608,27 @@ runs:
           "*Cost at API list prices — a large multiple of the effective rate on Claude Code subscriptions.*"
         ' <<< "$USAGE" >> "$GITHUB_STEP_SUMMARY"
 
+    # Gate on the job being red, not on the agent step specifically. A dozen
+    # steps run before `claude` — the rate-limit preflight, auth validation,
+    # the proxy and sandbox build, the adopter's `sandbox_setup:` — and any of
+    # them failing strands the same work, so each needs to reach the outage
+    # tracker. Keying on `steps.claude.outcome` meant a pre-agent failure went
+    # out as a red run and nothing else; a rate-limit preflight abort held
+    # every workflow on this repo for a full UTC day without filing a thing.
+    #
+    # The security preflight is the one exclusion, and what carves it out is
+    # that reporting would have the bot write to the repo (`gh issue create`,
+    # and `gh issue close` on the reconcile path) with its PAT immediately
+    # after the security gate refused to let it operate there. Its failure is
+    # also persistent — red until a human fixes branch protection, so an issue
+    # titled "Bot temporarily unavailable" would collect a row per trigger —
+    # but that argument doesn't single it out: `Validate auth configured` and
+    # the adopter's `sandbox_setup:` fail the same way and stay in. Bounding
+    # that repeated append belongs in `report-failure.sh`, where it covers
+    # every such step at once (#859); enumerating more step ids here would go
+    # stale as steps move.
     - name: Report failure
-      if: failure() && steps.claude.outcome == 'failure'
+      if: failure() && steps.security.outcome != 'failure'
       shell: bash
       run: bash "${{ github.action_path }}/../shared/steps/report-failure.sh"
       env:

codex/action.yaml

@@ -83,6 +83,7 @@ runs:
     # Preflight/teardown step bodies are shared with the Claude actions via
     # ../shared/steps; codex's auth, prompt staging, and run path are its own.
     - name: Security preflight
+      id: security
       shell: bash
       run: bash "${{ github.action_path }}/../shared/steps/security-preflight.sh"
       env:
@@ -329,8 +330,15 @@ runs:
       env:
         GITHUB_TOKEN: ${{ inputs.github_token }}
 
+    # Gate on the job being red, not on the agent step specifically — see the
+    # matching note in claude/action.yaml, including why the security preflight
+    # is excluded. Every step ahead of `codex` (the rate-limit preflight, bot-ID
+    # resolution, auth validation, the CLI and plugin installs, the AGENTS.md
+    # staging, the fork-PR pinning) strands the same work when it fails, and
+    # keying on `steps.codex.outcome` kept all of them out of the outage
+    # tracker.
     - name: Report failure
-      if: failure() && steps.codex.outcome == 'failure'
+      if: failure() && steps.security.outcome != 'failure'
       shell: bash
       run: bash "${{ github.action_path }}/../shared/steps/report-failure.sh"
       env:

shared/steps/report-failure.sh

@@ -1,7 +1,12 @@
 #!/usr/bin/env bash
 # File or append to a `tend-outage` issue when a run fails, so outages are
-# tracked until resolved. Shared verbatim by both harness actions; the
-# caller gates it on the agent step having failed.
+# tracked until resolved. Shared verbatim by both harness actions; the caller
+# gates it on the job being red, so a failure anywhere in the action — the
+# rate-limit preflight and sandbox build ahead of the agent as much as the
+# agent itself — lands in the tracker. The one exclusion is the security
+# preflight: that failure is a persistent config refusal rather than an
+# outage, and the issue this files ("temporarily unavailable", closed once
+# resolved) would never resolve.
 #
 # Just records the run link. Error annotations and logs are not reliably
 # available while the job is in_progress, so the nightly skill enriches these