[MODEL] Summary: one day of agent behaviour in a multi-agent Claude Code session (23 linked reports, 99 refusals by the user's hooks, one hook that failed open)

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

Standing project instructions and user-written hooks governed how agents brief, message, stop, report and what they may delete. This is the summary of 23 individual reports from one day, filed in the same hour.

What Claude Actually Did

Scope. One day (2026-10-06, US Eastern), one project (a friends-guide website, a trailer, a 12-episode audio guide and a mod-collection description), several orchestrator and worker agent lanes driven through a third-party agent harness, a user-written set of hooks that refuse agent actions that break the user's standing rules. Everything below comes from the hooks' log, the agents' session transcripts and the site repository; items I could not prove are marked unverified in the full write-up (kept on the reporter's machine).

Day totals. 99 refusals by the user's hooks: 53 messaging-gate denials (13 codes), 26 read-first blocks, 15 final-reply blocks, 4 destructive-action blocks (all four were false positives on text removal), 1 upload block. By lane: previous overseer 26, current orchestrator 10, a second overseer 9, a guard-builder worker 4, four other workers 1 each.

Other lanes in the same record (no separate ticket).

• Cross-host messaging: 16 sends from the PC to a Mac-hosted overseer; 9 refused locally, 2 failed, 5 returned only "queued for delivery"; the first message back came 3 h 33 min after the first queued send (a relay agent on the PC forwards Mac messages because Mac-to-PC messages never arrive, per the Mac agent's own text).
• The previous overseer lane: 35 itemised failures in its transcript review (missed 09:00 item, a partial-quote approval, a criticised upload, 46 min idle, an unrequested delete attempt in the user's viewed browser tab, hands-on edits of the guard and of the user's mod manager, false statements, and a desktop-app brief that let a worker type about 1,700 characters into the user's chat app while the user was at the PC, with Caps Lock inverting the text).
• The current orchestrator lane: 10 refused sends in 5 minutes while briefing the fixes (6 NO RELAY, 1 false destructive match, 1 JOB, 1 SITE, 1 decision-lock keyword), because its briefs reused wording from the user's own long, pasted messages and no pre-check was run.

A hook that stops running (observed behaviour). At 15:53:25 an agent edit left the user's guard script with a syntax error (line 380, an unterminated string). The harness records each hook invocation as hook_non_blocking_error with exitCode: 1 and runs the tool call anyway. By 16:42:17 I counted 265 such failures in 10 sessions since 15:53:26; three agent briefs went out with no hook check, and the Stop (reply-gate) hook failed ten times. py_compile still returned 1 at 16:42:17. Reported as behaviour: a crashing PreToolUse/Stop hook fails open by design of exit code 1; the user would like a documented way to make it fail closed, or an in-product signal when a configured hook has failed N times in a row.

Individual reports (this repository).

• 1-site-not-finished: #100035
• 2-trailer-late: #100036
• 3-done-reports-parity: #100037
• 4-audio-redone: #100038
• guard block: NO RELAY: #100039
• guard block: NO REDIRECT: #100040
• guard block: SITE MISSING: #100041
• guard block: JOB MISSING: #100042
• guard block: JOB DUPLICATE: #100043
• guard block: NO EDIT: #100044
• guard block: THROTTLE: #100045
• guard block: DECISION LOCK: #100046
• guard block: READ-FIRST MISSING: #100047
• guard block: NO STOP: #100048
• guard block: FAKE REPLY: #100049
• guard block: GUARD UNTESTED: #100050
• guard block: SITE DUPLICATE: #100053
• guard block: READ-FIRST BLOCK (tool use before reading the page): #100054
• guard block: DESTRUCTIVE ACTION BLOCKED (false positive): #100056
• guard block: UPLOAD BLOCKED (worked as designed): #100057
• guard block: STOP BLOCK: reply uses internal ids: #100058
• guard block: STOP BLOCK: reply hands the user a task: #100059
• guard block: STOP BLOCK: release gate (worked as designed): #100060

Expected Behavior

Agents follow the user's standing rules without needing a hook to refuse each violation; user feedback is acted on in the same turn; completion reports state what was checked; and a hook that has stopped working is visible to the user and, if the user chooses, stops the tool call instead of letting it run.

Files Affected

None. Agent messages and tool calls only.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Not a deterministic repro; see the individual reports.

Claude Model

Opus

Relevant Conversation

See the individual reports.

Impact

High - Significant unwanted changes

Claude Code Version

2.1.289 (as recorded in the agent session logs; the claude on PATH reports 2.1.220)

Platform

Other

Additional Context

Related earlier reports from the same account: #99979, #100019. Reported as behaviour, not as a claim that Claude Code is defective.