fix(security): v0.35.0 — escape Markdown report table cells; old-DB upgrade regression test

[hxddh]

What changed

A focused release from a report-generation audit.

🔴 Security — Markdown report table injection / stored XSS

The report renderer's _table (sidecar/app/runs/analysis_report.py) interpolated cell values straight between | delimiters with no escaping — and the cells carry the most attacker-influenceable data in the product: object keys and user-agents (an S3 key legally contains |, newlines, and angle brackets). Unescaped:

• a | in a key misaligns the row's columns (size/storage-class shift under the wrong headers);
• a newline splits one row into two, corrupting the rest of the table;
• a key like '<img src=x onerror=...>' lands verbatim in the saved .md → stored HTML/script in any viewer that renders inline HTML.

Credential redaction (applied to the whole document on write) only scrubs secret patterns — it does not touch markdown metacharacters, so it didn't cover this. Fixed with a _cell() escaper (| \r \n \ < >`) applied at the render boundary, covering all four report types (access-log, inventory, config-review, account-discovery) at once. Backward-compatible with app-generated content.

🟢 Regression coverage — old-DB upgrade survival

Added a test proving an old-schema DB upgrades cleanly: rows in the tables that later migrations rebuild (tool_calls@m002, datasets@m004) survive with data intact (kind→dataset_type, source_path→stored_path, counts preserved). The path was sound but had no coverage.

What checks were run

• cd sidecar && pytest -q — 596 passed (+3 in tests/test_v0350_fixes.py).
• ruff check app — clean.
• Existing report tests (test_config_review, test_runs, test_account_discovery) — green (escaping is backward-compatible).

Note

This round's other two mining angles (skills-content drift, migration deep-dive) were interrupted by a session limit. I completed the skills-drift audit inline — it came back clean (every tool/param a SKILL.md references still resolves to a real tool) — and verified the migration happy-path is sound (hence the regression test above). A fuller re-run of those angles is in progress separately.

Security invariants

No new writes, no shell, no destructive S3 op. The fix only escapes rendered output — it doesn't change what data reaches the report (redaction still runs). No behavior change for legitimate content.

🤖 Generated with Claude Code

https://claude.ai/code/session_01M3YojFPzjkfMj6YYSwvzdx

---

Generated by Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 240a77b1ac

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Teach the report parser to honor escaped pipes

When a key or user-agent contains |—the exact input this change targets—_cell emits \|, but the built-in report viewer's cells() in frontend/src/components/Markdown.tsx line 176 unconditionally calls .split("|") without recognizing escapes. Consequently, a row such as | a\|b | 5 GB | is still parsed as three cells in the report overlay and RunDetail, shifting columns and corrupting the displayed results despite the new backend test passing; update that parser or encode pipes in a form it does not treat as a delimiter.

Useful? React with 👍 / 👎.