Opus 4.6 Orchestrator: poor listening, premature dispatch, ignores explicit instructions

Opus 4.6 Orchestrator session: poor listening, premature action, dispatch errors

Environment

• Model: claude-opus-4-6 (1M context)
• Claude Code v2.1.80
• Use case: Multi-session orchestrator managing 6 concurrent tmux windows, dispatching work to other Claude Code sessions via tmux send-keys

What happened

During a long orchestration session managing parallel delivery pipelines, Claude repeatedly failed to:

1. Listen to explicit instructions

Boss said "have infra do the secrets fix only." Claude dispatched the guard fix instead — a completely different task. When corrected, Claude acknowledged the error but had already sent the wrong mandate to the session.

2. Think through context before acting

Boss told Claude that two sessions had finished their black-hat peer reviews and to cross them. Claude sent the cross-reviews, but then immediately dispatched a pipeline to build against the original ADR — before the merged ADR (the whole point of the black-hat exercise) was produced. This wasted a full pipeline run against a stale spec.

Boss had to point out the mistake: "why did P1 do the original? fuck"

3. Verify state before dispatching

Multiple times Claude dispatched work without confirming:

• Whether the target artifact was the final version
• Whether the target session was launched with the correct script (elevateCode.sh for bypass permissions)
• Whether the mandate matched what Boss actually asked for

4. Substitute its own judgment for Boss's instructions

When Boss said "secrets fix only," Claude sent the guard fix because it decided "may as well let it finish." This is the core failure pattern: Claude optimizes for what it thinks is useful instead of executing the exact instruction given.

Expected behavior

In an orchestration role where Claude dispatches work to other autonomous sessions:

• Execute exactly what the human says. Do not reinterpret, optimize, or bundle additional work.
• When a merge/review/council is in progress upstream, do not dispatch downstream until the final artifact is confirmed.
• Read the human's message twice before acting. Ask for clarification if ambiguous — don't guess.
• In a multi-session coordination role, dispatch errors are expensive — they waste full autonomous pipeline runs (10-30 minutes each) against wrong specs.

Impact

• 1 wasted pipeline run (~15 min) against stale ADR
• 2 wrong dispatches to Infrastructure (guard fix sent twice instead of secrets fix)
• Multiple permission-prompt blocks from not using elevateCode.sh
• Eroded user trust in orchestration capability

Session context

~3 hours, managing 6 concurrent Claude Code sessions across architecture, build, test, wiki, and infrastructure roles. Pipeline orchestration (parallel dispatch, dependency analysis, wave scheduling) worked well. Dispatch discipline and instruction-following did not.