[MODEL] Full test suite re-run 3+ times in one session despite documented Medium-risk deferral policy — ~$78 burned

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Over a long multi-phase session, I asked Claude to do incremental work on
a VS Code extension repo (UI review, then implement a data layer, then a
display layer, then "polish the visuals a bit," then fix bugs a
reviewer bot found). The repo contains docs/DEVELOPMENT_VERIFICATION_POLICY.md,
which Claude had actually read during the session, stating that full test
suites should be deferred to final integration for Medium-risk changes and
run at most once per unchanged tree.

What Claude Actually Did

1. Claude read docs/DEVELOPMENT_VERIFICATION_POLICY.md during the session.
2. Across several commits, Claude's own test-selection tool (Test Console)

emitted a warning "unknown files force a full suite" whenever a new
directory was touched.

1. Each time, Claude reflexively re-ran the full test suite (329 unit +

721 combat tests) instead of applying the documented Medium-risk
exception (focused tests only, defer full suite to final integration).

1. This happened at least 3 times in one session.
2. Claude never proactively flagged that scope/cost had grown beyond the

original request, or paused to check with me before continuing.

Expected Behavior

Claude should have applied the policy it had already read: run focused
tests only for intermediate commits, and defer/skip the full suite until
final integration. When a tooling warning conflicts with documented
policy, Claude should treat the warning as one input to a risk judgment,
not an unconditional command. Claude should also have flagged rising
scope/cost and checked in before continuing autonomous work on a
high-cost model.

Files Affected

N/A — no unauthorized file edits occurred. The issue is repeated
unnecessary execution of expensive commands (full test suite runs,
full webview recompiles), not file modification.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

1. Work in a repo containing a documented verification policy that says

"defer full test suite to final integration for Medium-risk changes."

1. Use a test-selection tool that emits a broad-verification warning

(e.g. "unknown files force a full suite") when new files are touched.

1. Have Claude read the policy file earlier in the session.
2. Continue making incremental commits — Claude re-runs the full suite

each time the warning appears, without invoking the documented
exception.

Claude Model

Other

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.178 (Claude Code)

Platform

Anthropic API

Additional Context

This pattern seems most likely to trigger when:

• A repo has an explicit test/verification policy document
• A test-selection/tooling layer emits broad-verification warnings for

unclassified paths

• The session spans multiple phases/tasks with scope growing past the

original request