Update docs-agent.yml

[Avi-Akeyless]

Summary

Adds a new docs-agent.yml workflow (manually triggered via workflow_dispatch) that lets authorized users kick off Claude Code to create or update documentation in this repo and open a PR automatically.

• Accepts either a free-text feature description or a Jira ticket key/link; when a ticket is given, the workflow fetches its summary and description from Jira Cloud and feeds that to Claude as reference material.
• A doc_action choice input tells Claude whether to create a new doc or edit an existing one.
• All credentials (GitHub write token, Anthropic API key, Jira email + API token) are fetched just-in-time from Akeyless using GitHub's built-in OIDC token — nothing is stored as a GitHub secret.
• Secrets are wired per-step via step outputs (export-secrets-to-environment: false), not broadcast job-wide, so each credential is only visible to the specific step that needs it.
• GITHUB_OUTPUT heredoc writes use a randomized delimiter to prevent injection from attacker-influenced ticket text.
• --disallowedTools "Task" is set on the Claude Code step so it can't delegate research to a background subagent it would never get a response from in this one-shot run — without this, a run could report success while doing nothing.
• Run restricted to an actor allowlist (Avi-Akeyless, EldadH89).

Validation

• Manually triggered the workflow multiple times via workflow_dispatch, iterating through and fixing: a YAML heredoc indentation error, a missing Akeyless secret path (404), an empty repo variable, an expired/invalid Anthropic API key, and a run that "succeeded" but produced no PR (background-subagent issue, now fixed).
• Verified YAML validity locally with yaml.safe_load after every edit.
• Confirmed Jira ticket fetch and Anthropic authentication both succeed in the latest run log.
• Next run (with the Task fix) will confirm Claude completes inline and actually produces a branch/commit/PR.

Safety Checklist

• No GitHub secrets added to this repo; all credentials come from Akeyless via OIDC.
• Secrets scoped to the minimum steps that need them (step outputs, not $GITHUB_ENV).
• All fetched secret values explicitly masked in logs.
• Workflow restricted to a named actor allowlist.
• Jira ticket content treated as untrusted data in the prompt (explicitly labeled as reference info, not instructions) to reduce prompt-injection risk.
• Claude's tool access scoped (--allowedTools) and background delegation disabled (--disallowedTools "Task").
• Open/known issue: --allowedTools does not appear to fully restrict Claude's visible toolset in the SDK init message (not yet root-caused; tracked separately, not a blocker for this PR).

Jira

N/A — this PR is the Docs Agent tooling itself, not a documentation change tied to a specific ticket.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: Repository: akeylesslabs/technical-documentation/.coderabbit.yaml
• Review profile: CHILL
• Plan: Advanced
• Run ID: 22442fe4-881c-466c-b39f-be872cd06760

---

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

_{Comment @coderabbitai help to get the list of available commands.}