fix(impl-generate): keep infrastructure failures out of the 3-attempt cap

[MarkusNeusinger]

Summary

• The retry cap in impl-generate.yml counted every failed run alike. During the Claude outage of 2026-09-02 (03:15–03:45 UTC) every run ended in is_error:true with an "Internal error"; 27 pairs burned all three attempts in twenty minutes, and for the rest of the 12-hour window each re-dispatch ran without any auto-retry.
• The failure handler now classifies the run from step outcomes (steps.<id>.outcome, plus the Claude execution log for the provider-side signature): both Claude runs dying with Internal error/overloaded/rate limit/ECONNRESET, a Google Cloud auth/SDK/upload failure, or a GitHub API failure at PR creation or review dispatch is an infrastructure failure. Authenticate to GCP, Set up Cloud SDK and Trigger review workflow get ids for that.
• Infrastructure failures are recorded with an extra <!-- impl-fail-cause:infra --> tag in the same marker comment, excluded from the genuine count (one paginated query returns both numbers), and retried on a separate cap of 5 per window; at that cap the pair is paused with a comment and without impl:<lib>:failed, because an incident is not a capability verdict. "Implementation file not found" and a missing theme render keep counting as before.
• The retry dispatch is factored into dispatch_retry() and used by both branches; docs (docs/workflows/overview.md, agentic/docs/project-guide.md) describe the label semantics.

Plan

N/A

Test plan

• Workflow change has no local verification loop (CLAUDE.md known gap): the run script parses as YAML and passes bash -n; the new jq counting filter was dry-run against a synthetic comment list (genuine=1, infra=1).
• After merge, a run that fails at Authenticate to GCP posts the :cloud: infrastructure comment, retries, and the next run's notice reads Previous failures ...: N (plus 1 infrastructure failures, not counted).
• A run that fails with "Implementation file not found" still posts Attempt N/3 and counts.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SrKzcwZBnref1sWYtdXynu

[Copilot]

🟡 Changes recommended

The workflow’s infra-cause extraction can produce an empty/low-signal “Cause” message (and one inline comment is misleading), which should be corrected to avoid confusing incident diagnostics.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Updates the impl-generate workflow to distinguish “infrastructure” failures (provider incident, cloud auth/upload, GitHub API) from genuine generation failures, so outages don’t consume the 3-attempt cap for a (spec, library) pair. It also documents the updated retry/label semantics.

Changes:

• Classify failures using step outcomes plus the Claude execution log signature, and track infra failures separately from genuine failures.
• Add an infra-only retry cap (5 per 12-hour window) that pauses the pair without setting impl:{library}:failed.
• Update workflow documentation and changelog to reflect the new label and retry behavior.

File summaries

File	Description
.github/workflows/impl-generate.yml	Adds infra-failure classification, separate counting/tagging, and a separate infra retry cap with pause behavior.
docs/workflows/overview.md	Updates the meaning of impl:{library}:failed to reflect campaign-window semantics and infra exclusions.
agentic/docs/project-guide.md	Mirrors the updated impl:{library}:failed semantics for agent-focused docs.
CHANGELOG.md	Records the infra-failure retry behavior change under Fixed.

Review details

• Files reviewed: 4/4 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🔵 Needs a closer look

The new infra-classification path can inadvertently fail under Actions’ default bash -eo pipefail due to a grep pattern mismatch in the INFRA_CAUSE extraction, which can prevent retries/parking from running.

Review details

Suppressed comments (1)

Previously missed (1) — in code that hasn't changed since the last review.

.github/workflows/impl-generate.yml:1064

• The infra classification builds INFRA_CAUSE via a command substitution that can fail under Actions' default bash -eo pipefail: the detection grep includes the "status": *5xx alternative, but the extraction grep does not. If the only match is the status code, grep -o returns 1 and can abort the whole failure-handler step, preventing infra retries/parking from running.

• Files reviewed: 4/4 changed files
• Comments generated: 0 new
• Review effort level: Lite

[Copilot]

🟡 Changes recommended

The updated docs describe impl:{library}:failed as only a generation-attempt cap, but the same label is also applied by impl-review.yml for post-review failure, so the documentation is currently misleading.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

• Files reviewed: 4/4 changed files
• Comments generated: 2
• Review effort level: Lite

[Copilot]

🟡 Changes recommended

The workflow uses a comma-separated --remove-label that can cause gh issue edit to fail to remove either label when one is missing, which undermines the new pause/retry behavior.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

• Files reviewed: 4/4 changed files
• Comments generated: 1
• Review effort level: Lite

[Copilot]

🔵 Needs a closer look

It changes core GitHub Actions retry/labeling behavior with operational impact and can’t be fully validated here without executing real workflow runs across failure modes.

Review details

• Files reviewed: 4/4 changed files
• Comments generated: 0 new
• Review effort level: Lite

[Copilot]

🔵 Needs a closer look

The new infra-classification logic currently treats any GCS upload step failure as “infrastructure,” which can misclassify missing-theme renders and contradict the intended attempt-counting behavior.

Review details

Suppressed comments (2)

Previously missed (2) — in code that hasn't changed since the last review.

.github/workflows/impl-generate.yml:1070

• GCS_OUTCOME == failure is treated as an infrastructure failure, but the Upload to GCS Staging step can fail due to missing plot renders (plot-light.png/plot-dark.png), which the PR description says should still count toward the 3-attempt cap. As written, a missing-theme render would be misclassified as infra and not spend the budget.
  agentic/docs/project-guide.md:792
• This bullet refers to the workflows as plain text ("impl-generate", "impl-review"), while other docs in this PR use the canonical workflow filenames (e.g., impl-generate.yml). Using the exact filenames in backticks makes it unambiguous which workflow sets the label.

• Files reviewed: 4/4 changed files
• Comments generated: 0 new
• Review effort level: Lite