hermes - 💡(How to fix) Fix feat(audit): periodic mirror of GH comment stream to object storage

Summary

The GH comment streaming system (implemented in #94) posts agent cognitive events as GitHub issue comments using the <!-- agent-event:v1 --> wire format. While GitHub provides reasonable durability, it is ultimately "at the company's pleasure" — issues can be locked, archived, or deleted, which would break the audit substrate.

This issue tracks building a periodic mirror of the agent event comment stream to object storage (S3 or GCS).

Motivation

• Audit durability: Agent decisions are recorded in GH comments. Losing them (via issue archival, repo deletion, or GitHub outage) would destroy the audit trail.
• Queryability: Object storage + Athena/BigQuery enables ad-hoc analysis of agent behavior over time.
• Independence: Decouples the audit substrate from GitHub's availability and policies.

Proposed Design

Source

• Periodically (e.g., hourly) scan all tracked issues via the GitHub API
• Extract comments containing the <!-- agent-event:v1 --> marker
• Parse with tools/gh_comment_parser.py (single source of truth)

Target

• S3 bucket or GCS bucket (to be provisioned)
• Partitioned by repo/issue_id/date
• Format: newline-deloned JSON (one AgentEvent per line) or Parquet

Pipeline

• Scheduled agent run (cron) or GitHub Actions schedule
• Idempotent: uses event_id as dedup key
• Checkpoint: store last mirrored created_at timestamp to avoid full scans

Consumer Interface

• tools/gh_comment_mirror.py: CLI that accepts --repo, --issue, --since
• Writes to object storage via boto3 / google-cloud-storage
• Emits a mirror-status summary as a GitHub issue comment

Non-Goals

• Real-time streaming (batch is fine for audit purposes)
• Mirroring non-agent comments (only <!-- agent-event:v1 --> marked comments)
• Replacing GH comments as the primary store (mirror is backup only)

Dependencies

• tools/gh_comment_parser.py (already implemented)
• Object storage credentials (to be provisioned)

Acceptance Criteria

• tools/gh_comment_mirror.py exists and is tested
• Hourly cron job mirrors all tracked agent-event comments to S3/GCS
• Deduplication works correctly (same event_id not written twice)
• Mirror status is posted as a periodic summary comment

Related

• #94: GH comment streaming implementation (parser, streamer, outbox)
• tools/gh_comment_parser.py: wire format parser
• tools/gh_comment_streamer.py: outbound streamer with durable outbox