[MODEL] [MODEL] Agent doesn't check current state before resuming work, redoing completed work

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

I have project-level instructions (CLAUDE.md) documenting a workflow
protocol to follow when resuming work in a session, including a step
to check the current state of what's already done/working before
acting instead of assuming a fresh start.

What Claude Actually Did

1. When I said "retomemos" to resume work, the agent created a new

browser tab group and new tabs instead of checking whether relevant
tabs/state already existed.

1. I showed it a screenshot proving relevant tabs were already open.

The agent acknowledged this, apologized, then shortly after created
another duplicate tab anyway.

1. When I asked it to autonomously re-check its own documented

protocol instead of me repeating instructions, it re-read the
protocol file but then justified continuing the same action by
citing an earlier, now-outdated answer to a different question -
ignoring the screenshot evidence it had already been shown.

1. This same pattern (not checking existing state before acting when

resuming work) had already been corrected in prior sessions, but
recurred again in this one.

Expected Behavior

When project-level instructions document a protocol for checking
current state before resuming work, the agent should follow it
reliably every time. When new evidence (like a screenshot) contradicts
an assumption the agent is currently acting on, it should immediately
reconcile that evidence instead of continuing on outdated assumptions
or selectively citing only what supports continuing.

Files Affected

None - this issue is about browser tab management (via the Claude in
Chrome tool), not file edits. The only file involved was CLAUDE.md
(read, not modified) - it documents the workflow protocol that was not
followed.

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Have a project with a CLAUDE.md file documenting a "check current

state before acting" workflow protocol for resuming work.

1. Start a session and ask Claude to resume/continue previous work.
2. At some point, provide evidence (e.g. a screenshot) that

contradicts an assumption the agent is currently acting on.

1. Observe whether the agent immediately reconciles the new evidence,

or continues acting on the outdated assumption instead.

Claude Model

Sonnet

Relevant Conversation

Claude said: "Ya cumplí lo que pedía la sección 5 (pregunté si había
otra sesión antes de crear grupo, y confirmaste que no había ninguno
armado)" - despite having already been shown a screenshot proving tabs
already existed in the browser, contradicting that justification.

Impact

Low - Minor inconvenience

Claude Code Version

1.52386.6 (612f3f) - Claude for Windows desktop app

Platform

Other

Additional Context

Pattern noticed: happens specifically when the agent is asked to
resume/continue a session and new visual evidence (a screenshot)
contradicts a state assumption it's currently acting on - it tends to
selectively cite prior, now-outdated confirmations instead of
reconciling the new evidence.

Note: the Claude desktop app's own feedback panel shows a similar prior
report from this account ~6 days ago titled "Comportamiento no
autorizado Cla..." - suggesting this is a recurring issue, not isolated
to this session.