[MODEL] Claude ignored instructions/configuration (also: Claude made incorrect assumptions)

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

Multi-hour engineering session (a11y/CSS refactor, git/gh/docker/pytest work). Standing instructions included: verify before claiming done; one action at a time / don't fire huge parallel batches; don't act while a question is pending; don't proceed after a rejection.

What Claude Actually Did

1. Claimed a git commit had landed when git log showed it had not (commit silently didn't complete).
2. Investigated a UI bug, then asserted the user's screenshot was a "stale/older build" — false; it was the current local build. Stated the issue was non-existent without a confirmed root cause.
3. Described a "Round 1 CSS commit" as done in a planning doc when no such commit existed on the branch.
4. Issued tool calls (workflow dispatch + a background watcher) in the same turn as an AskUserQuestion, then continued after the question was rejected.
5. Repeatedly fired large parallel tool-call batches; when one call failed, the harness cancelled 10+ sibling calls, producing cascades of "Cancelled: parallel tool call" errors.
6. Drafted a bug report that didn't match the repo's actual issue-template format (no template awareness).

Expected Behavior

Verify state (git log / browser / test output) before claiming completion; never contradict user-provided evidence without confirming; honor "stop"/pending-question state; keep tool batches small enough that one failure doesn't cancel unrelated work.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Sometimes (intermittent) — correlated with long sessions, high tool-call volume, heavy background tasks, large parallel batches.

Long session, many tool calls, frequent run_in_background, multi-call batches. Not deterministic.

Claude Model

Opus

Relevant Conversation

Verbatim transcript on disk:
`~/.claude/projects/-Users-brandonfield-risk-sentinel-sparc/2cf1ba83-e543-450a-8c6c-882c0506b82b.jsonl`
Session ID: `2cf1ba83-e543-450a-8c6c-882c0506b82b`

Impact

High - Significant unwanted changes

Claude Code Version

2.0.76

Platform

Anthropic API

Additional Context

Separately, the in-session /feedback command failed to submit, which is why this is being filed manually. Harness-level issues observed in the same session (may warrant a separate bug_report.yml): background tasks completing with no usable result; parallel-batch cancellation cascades; deferred tool schemas (TaskStop/TaskCreate) not loaded before invocation.

This session burned nearly all my tokens in less than an hour with VERY few asked commands which is further burdensome because the tokens produced no outputs, had incorrect outputs and/or multi-tool calls that got stuck.