0.5.1

The Phase 1.8 §2 self-dogfood cut. Every fix in this release came from
a single round of running getdebug against a real 1,217-file Python
repo (crewAIInc/crewAI) — plus a full MCP round-trip against the
debug org itself. Two demo killers, fifteen smaller cuts, zero new
features.

Top-line numbers

• Python false-positive rate cut roughly 70% on prompt-injection
  • unbounded-stream + secret-placeholder shapes. 14 of 22 HIGH FPs
    on crewAI traced back to message and query in the
    prompt-injection identifier set; both removed.
• --fail-on=verified-* is now actually a precision lane. Before
  this cut, verified-high and plain high produced the same exit
  code on every real repo — the verified suffix only skipped
  status=invalid secrets. It's now an inclusion gate: a finding
  counts only with affirmative evidence (provider 2xx for secrets).
• MCP list_findings → get_finding works end-to-end. The
  get_finding tool was hitting an API route that didn't exist — so
  any AI agent that listed findings could never drill into one. The
  route now exists and accepts both the legacy id form and the new
  stable form. Live on prod.

Demo killers (P0)

• --fail-on=verified-* is an inclusion gate, not a subtraction.
• TTY output surfaces [LIVE] / [REJECTED] / [UNVERIFIED] badges
  per row + a verification summary line. Pre-fix, verifier output
  was silently dropped on the way to the renderer.
• Python prompt-injection: message + query removed from the
  identifier set. self.message = f"…" in custom Exception
  subclasses and query = "SELECT …" + … SQL builders no longer
  fire.

Detector improvements

• New detectors for xAI, GitLab, and npm access tokens —
  verifiers shipped in 0.5.0 but no detector regex existed.
• Anthropic (sk-ant-…) now correctly classifies as Anthropic, not
  OpenAI. The scan loop tracks per-line consumed ranges so a token
  can produce only one finding.
• Placeholder values prefixed fake- / mock_ / stub- suppressed
  in .env-style files.
• PEM-block detector respects Python docstrings + doctest lines.
  Narrow — real sk-… shapes inside docstrings still fire.
• streamTruePyRe requires kwarg context — kills FPs on
  self.stream = True attribute assignments.

CLI ergonomics

• --local-llm files are ranked by security relevance before
  the file cap. Pre-fix, the default 50-file cap burned on
  __init__.py + constants before reaching auth/sql handlers.
• New per-file progress meter [N/M] path · elapsed Ns · ETA Ns
  during --local-llm.
• New --local-llm-per-file-timeout (default 3 min) — pre-fix
  ceiling was the localllm HTTP client's 10 min, so one stuck file
  could pin a whole run.
• getdebug fix . --local-only accepts . as the workdir alias
  (every other command did already).
• getdebug undo deletes the .getdebug-backup-<TS>/ directory by
  default. Use --keep-backup to retain it.
• getdebug login no longer sticky-caches localhost API URLs
  across sessions.
• The 20 MB truncation hint now names the file the walk stopped at
  • suggests concrete next steps.

MCP

• GET /v1/findings/:id exists. mcp__getdebug__get_finding works.
• Finding ids on the wire are now finding_<projectId>_<contentHash>
  — stable across rescans. Legacy ids still resolve on the input
  side for backwards compat. MCP server source unchanged; same npm
  artifact, but it gets stable ids back from the API now.

Docs

• /docs Local-mode section now explicitly walks the chat-model
  pull (ollama pull qwen2.5-coder:7b) needed for --local-llm,
  not just the embedding model used by index --local.
• Documents the per-model wall-clock tradeoff: deepseek-r1:7b ~75s
  per file needs --local-llm-max-files=10, not the default 50.

Upgrade

npm install -g @getdebug/cli@0.5.1
# or for MCP:
npm install -g @getdebug/mcp@0.5.1

No breaking changes. CI gates using --fail-on=verified-high will see
sharply lower fail counts — this is the intended behavior; use plain
--fail-on=high if you want the prior gate.

Full changelog: cli/CHANGELOG.md