fix(bot): scan precision - quoted secrets only, low-risk paths to nit

[samson910022]

First live ai-scan dry_run produced 3 false positives (bare code assignment, unquoted yml env line, test fixture). Two precision tunings: (1) secret patterns require quoted assignments (bare code and unquoted env lines no longer match; committed .env still caught by forbidden-file check); (2) secret-pattern hits under tests/docs/yml/md downgrade to nit so real findings sort first (fingerprints unchanged, dedupe unaffected). 38 python tests green.

[github-actions]

Verdict matrix

• Roles: gateway_safety, api_compat, docs_config
• FINAL_VERDICT: APPROVE

Sensitive paths

• github_bot/src/repo_scan.py

Deterministic pre-checks

• possible-password: github_bot/src/repo_scan.py,tests/test_ai_bot.py:62
• private-key-material: github_bot/src/repo_scan.py,tests/test_ai_bot.py:64
• possible-password: github_bot/src/repo_scan.py,tests/test_ai_bot.py:71

gateway_safety

VERDICT: APPROVE

No BLOCKING findings (no secret leaks, no auth bypasses, no broken wire shapes, no regressions to [provider] invariants).
No SHOULD_FIX findings.

NITS:

• github_bot/src/repo_scan.py:19: Original unquoted secret patterns are commented rather than deleted; minor audit-trail vs. cleanliness tradeoff.
• tests/test_ai_bot.py:327: import re + from repo_scan import SECRET_PATTERNS moved inside test method (acceptable but non-idiomatic; other tests import at top).
• tests/test_ai_bot.py:340: test_prior_dry_run_fps_resolved calls live collect_findings() (full worktree scan); acceptable for regression verification but could be mocked for pure unit-test purity.

All changes preserve recall for real leaks (quoted pasted secrets + separate forbidden-file check for .env etc.), downgrade only noise under tests/ / *.yml/.md/.github/*yml, keep private-key pattern untouched, and pass the stated 38 Python tests. No [provider] safety invariants violated.

api_compat

API Compatibility Review

This PR modifies only the GitHub bot's repository scanner (github_bot/src/repo_scan.py) and its tests (tests/test_ai_bot.py). It does not touch any [provider] runtime code (src/proxy.ts, src/converters/*, src/config/*, route handlers, collectors, retry logic, session management, or documentation).

---

Checklist evaluation

#	Area	Impact
1	Route error wire shapes	Not touched — no changes to chat/responses/messages/interactions handlers
2	Converter thinking/signature preservation	Not touched
3	previous_response_id / session chaining	Not touched
4	Reasoning mapping / budget thresholds	Not touched
5	Model alias rules	Not touched
6	Collector finish==='tool' / flushHeaders	Not touched
7	resolveMaxRetries / retry backoff	Not touched
8	New endpoints / docs	Not applicable

No [provider] code is affected; all eight compatibility invariants are preserved.

---

Deterministic pre-findings triage

The three pre-findings (possible-password ×2, private-key-material ×1) reference lines inside the scanner's SECRET_PATTERNS list and test assertions. These are regex pattern strings and test fixtures, not actual secrets. The diff evidence is [REDACTED], and the visible diff confirms the lines are pattern definitions and re.search calls against synthetic test strings. No real credential material is present.

---

Findings

BLOCKING: (none)

SHOULD_FIX:

1. github_bot/src/repo_scan.py:38 — is_low_risk_path uses str.startswith(tuple) correctly, but the suffix check rel.endswith(LOW_RISK_SUFFIXES) will also match paths like src/config/proxy.yaml or a hypothetical src/secrets.yml. If the intent is strictly fixture/doc noise, consider combining the directory and suffix checks (e.g., only downgrade suffix matches when also under a low-risk directory). Current behavior is defensible but slightly broader than described.

2. tests/test_ai_bot.py:361–377 — test_prior_dry_run_fps_resolved calls collect_findings() which internally runs scan_worktree against the real working tree. This makes the test environment-dependent (it will pass only when the repo is checked out and the working tree matches expectations). If CI runs from a shallow clone or a different worktree state, the assertions on specific files may become flaky. Consider mocking scan_worktree or using a temp directory fixture.

NITS:

1. github_bot/src/repo_scan.py:28–29 — The {3,} minimum in the quoted-value capture ([^'\"]{3,}) means a 1- or 2-character quoted secret like password: "ab" won't match. This is likely intentional to reduce noise, but worth a brief inline comment noting the minimum-length choice.

2. tests/test_ai_bot.py:335 — Test strings use [REDACTED] as placeholder secret values. These are fine for pattern matching, but using a more obviously synthetic value (e.g., sk-test1234abcd) would make the test intent clearer at a glance.

---

Test gap note

No cross-protocol regression risk — this PR is entirely within the bot/scanner subsystem. The 38 Python tests cover the changed logic. No TypeScript [provider] tests are affected or need updating.

---

VERDICT: APPROVE

The changes are scoped entirely to the bot scanner's precision tuning and do not touch any [provider] runtime, wire shapes, or API compatibility surfaces. The should-fix items are minor robustness concerns, not blockers.

docs_config

VERDICT: APPROVE

BLOCKING: None

• No new/changed [provider] env: six-way matrix untouched.

SHOULD_FIX: None

NITS: None

Matrix gap closure — files that must change: none.

Checked:

• github_bot/src/repo_scan.py:22-35 — quoted-only secret regex + LOW_RISK_DIRS/SUFFIXES + is_low_risk_path is bot-internal, not [provider] config. No sync required in .env.example, config.json.example, Dockerfile, docker-compose.yml, index.ts, docs/configuration.md, AGENTS.md.
• tests/test_ai_bot.py:330-369 — bot unit tests only. No npm run lint reference, no test:unit rename, no README/CHANGELOG drift.
• No README.md / README.zh-CN.md / docs/README.md changes needed; no link to git-ignored docs/host-deployments.md added.

FINAL_VERDICT: APPROVE

Session: response_id=1528ca6b-1bef-9426-a795-8dc61486fb5d,resp_req_vrtx_011CetUCLcG41LjCJxJawAtv,chatcmpl-f5bf5235-cf69-4aaf-bd96-f3b5904cbd7a (chain via previous_response_id, 30min TTL). model=[model] provider=[provider].