Conversation
Claude Code 2.1.x shows a "Bypass Permissions mode" acceptance screen on
launch with --dangerously-skip-permissions. Its cursor DEFAULTS to
"1. No, exit", so the trust-prompt auto-accept handler's bare Enter selected
"No, exit" and quit Claude — crash-looping every agent on first boot. The
option text "2. Yes, I accept" also made the existing recent.includes('Yes')
check fire that same fatal Enter.
Detect the screen and arrow-down to "2. Yes, I accept" before confirming,
exactly once per session (guarded so later redundant timers don't inject
stray keys into the live REPL).
Match on single tokens ('Bypass', 'responsibility'): Claude's TUI positions
each word with ANSI cursor-move codes, so there are no literal spaces in the
PTY buffer and a multi-word substring like 'Bypass Permissions mode' never
matches.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ername
detect-chat-id derived ALLOWED_USER as `user.username || String(user.fromId)`,
so any operator who has a Telegram @username got a non-numeric value written
to the agent's .env. The daemon then rejects it ("ALLOWED_USER must be a
comma-separated list of numeric Telegram user IDs") and SILENTLY refuses to
enable Telegram — the agent can send but never receives, with no obvious cause.
Always write the numeric fromId; keep @username only for the display label.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
HubKey76
changed the title
* fix(pty): harden bypass-permissions startup recovery Port the member-reported fix from #11 using the canonical bounded, stateful handler. Retry only while the gate remains in the final 4KB, cap navigation at three answers, strip ANSI, and keep bare Enter unreachable while an exit-default marker is visible. * fix(detect-chat-id): persist numeric Telegram user IDs Port the diagnosis from #28 and verify the real command JSON path. ALLOWED_USER now always uses the numeric sender ID even when Telegram supplies an @username. --------- Co-authored-by: NancyBraun <nancy@carolinaspropertymanagement.com> Co-authored-by: HubKey76 <mike@hubkey.co>
|
@codex review\n\nPlease review exact head |
|
Codex Review: Didn't find any major issues. You're on a roll. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
|
Closing this one as superseded - but with credit where it's due: you diagnosed this 11 days before the fix that landed, and the shape that shipped matches your diagnosis (default-to-"No, exit" handling, single-token matching, numeric user IDs). Where it landed on main: Thanks for being first to the diagnosis - that's what got it fixed. Hope to see more from you. |
2 participants
Two bugs that prevent a fresh agent from onboarding on Claude Code 2.1.x. Each is an independent commit.
1.
fix(pty)— auto-accept the "Bypass Permissions mode" promptEvery agent crash-loops on first boot. Claude Code 2.1.x shows a "Bypass Permissions mode" acceptance screen when launched with
--dangerously-skip-permissions(which the Anthropic PTY adapter does). Its cursor defaults to1. No, exit:The trust-prompt auto-accept in
AgentPTY.spawn()dismisses prompts with a bare Enter — which here selects "No, exit" and quits Claude. The watchdog marks the agentcrashedand restarts it into the same wall forever. (bypassPermissionsModeAccepted=truein~/.claude.jsondoes not suppress it.) Worsened because2. Yes, I acceptcontainsYes, so the existingrecent.includes('Yes')check also fired the fatal Enter.Fix: detect the screen and arrow-down to "2. Yes, I accept" before confirming, once per session (guarded so redundant timers don't inject stray keys into the live REPL). Match on single tokens (
'Bypass','responsibility') — Claude's TUI positions each word with ANSI cursor-move codes, so there are no literal spaces in the PTY buffer and a multi-word substring like'Bypass Permissions mode'never matches (same reason the existing single-token'trust'/'Yes'checks work).2.
fix(detect-chat-id)— write the numeric user id toALLOWED_USERdetect-chat-idderivedALLOWED_USERasuser.username || String(user.fromId), so any operator who has a Telegram @username got a non-numeric value written to the agent.env. The daemon then rejects it — "ALLOWED_USER must be a comma-separated list of numeric Telegram user IDs" — and silently refuses to enable Telegram. The agent can send messages but never receives them, with no obvious cause.Fix: always write the numeric
fromId; keep@usernameonly for the human-readable display label.Testing
Both reproduced and fixed live on Claude Code 2.1.176 (macOS, Node v24):
Exited with code 1, repeated watchdog restarts), cursor stuck on1. No, exit. After: cursor moves to❯ Yes, I accept, Claude boots, agent reachesBootstrap complete. Beginning poll loop.Refusing to enable Telegram; agent sent onboarding messages but never received replies. After (numericALLOWED_USER):Telegram configured … allowed_user: enabled,Telegram poller started, two-way messaging works.🤖 Generated with Claude Code