Changelog
Changelog
Notable changes per release. Each entry links the commit that carries the full reasoning — the commit bodies are the long form, this is the index.
Unreleased
Added
- A new bot gets the TeleCoder icon.
telecoder initoffers it to a bot that has no profile picture yet, and a bot made with/newbotstarts with it. A plain letter on a coloured disc is hard to tell from the next bot’s.--iconreplaces a picture the bot has;--no-iconleaves it alone without asking. A picture that’s already there is never replaced unasked.
Changed
- The status line is on by default, with its topic line. Project, effort,
model, context and cost show nowhere else in Telegram, and a new user
doesn’t know
/statusline. It’s sent silently now: the reply above it already rang the phone.STATUS_LINE_DEFAULT=falserestores the old default; the last-prompts andclaude --resumelines stay opt-in.
Fixed
- A background task claude stopped no longer shows as running forever. Claude writes no task-notification for a task it ends with TaskStop, and that notification was the only thing that retired a task, so ”🔄 Running in background” stayed under every later reply until the bot restarted, and the idle restart waited for it. A Bash call’s response now records claude’s own id for the task, and TaskStop on that id retires it as stopped.
TTS_ENABLED=falseturns spoken replies off. The setting was read but never checked, so a chat with/ttson kept getting voice replies. Now it stops them for every chat,/ttssays they’re off, and the command leaves the/menu.TRANSCRIBE_ENABLED=falseturns off voice notes too, not only/transcribe. A voice note was still downloaded and transcribed.telecoder help <command>lists the command’s options. It showed only the one-line summary, though the install docs pointed to it for the flags.<command> --helpshows the same.
[1.2.0] — 2026-11-14
Install
npm install -g telecoder— TeleCoder installs as an npm package with atelecodercommand:telecoder startruns one bot, or every bot ininstances.jsonwhen there is one;telecoder doctorchecks Node, theclaudeCLI and its login (without spending anything), node-pty and the toolchain it needs, and the config, with--liveasking Telegram about the token;telecoder --versionandtelecoder helpdo what they say.telecoder service installruns the bot in the background: a systemd user unit on Linux (enabled, with lingering after asking, so it survives logout and reboot) or a LaunchAgent on macOS, pointing at thenodeandtelecoderit was installed with and aPATHthat findsclaudeandgit./restartbotrestarts it from chat; a configuration error stops it instead of crash-looping (exit 78,RestartPreventExitStatus=; exit 0 under launchd, which can’t single out one code). A changed unit is shown as a diff before it is replaced;status,logs,restartanduninstallgo with it, and--nameallows several installs. Undersudo -u(noXDG_RUNTIME_DIR) it pointssystemctl --userat the user’s running manager itself, found by its bus socket in/run/user/<uid>; when no manager runs (no login, no lingering) it leads with the fix,sudo loginctl enable-linger <user>. Withoutsystemd --userit prints the unit and a manual fallback; npx installs are pointed atnpm install -gfirst.telecoder initsets a bot up in a few questions — it runs doctor’s checks (offeringclaude auth loginwhen Claude Code isn’t logged in), takes the bot token masked and checks it with Telegram, and finds your user id by having you send the bot/start, so there is no @userinfobot step (the bot answers that/startat once, pointing you back to the terminal). Then the workspace, SDK or PTY, and voice transcription and Telegraph if you want them. It writes a.envwith only those settings, readable by you only, and the bot messages you to say it’s set up. An existing.envcan be kept, edited with its values as defaults, or started over (backed up first).--token-env VAR,--owner-id,--workspace,--transportand--yesrun it from a script. When the Claude login is a variable in the shell’s environment (CLAUDE_CODE_OAUTH_TOKEN,ANTHROPIC_API_KEY,ANTHROPIC_AUTH_TOKEN) rather than stored credentials, it offers to save it in the.envtoo (--yesdoes), since a background service never sees that environment and would start logged out;telecoder doctorwarns about such a login as well.- Config lives in
~/.telecoderfor an npm install (TELECODER_HOMEmoves it):.env,instances.jsonand the Telegraph account file sit next to the existing state files. A source checkout with a.envorinstances.jsonnext topackage.jsonkeeps using those, exactly as before. - An npm install updates and restarts without a checkout —
/updatealso offersnpm install -g telecoder@latest,/versioncompares the installed version with the latest release, and/restartbotexits for the service manager to start it again (telecoder service installsets one up; pm2 works too)./rebuildbotis for source checkouts only and says so elsewhere. - A shorter README — the README is now the pitch, how TeleCoder compares,
a three-command quick start and links; the reference moved to
docs/(install, features, commands, multiple bots, providers, sharing, configuration, security, architecture). Corrected on the way: theCLAUDE_SDK_LOG_LEVELand GroqTTS_VOICEdefaults,HOME_PROJECT,/spinoffand/autoapprovein the command list, and which commands are admin-only.
Security
-
/fileand/telegraph <path>can’t hand out credentials — both accepted any path underWORKSPACE_DIR, which defaults to your home directory, so any guest on a shared bot could fetch~/.claude/.credentials.jsonwith/file, or publish~/.ssh/id_rsaas a public page with/telegraph. Guests are now confined to the chat’s project directory, admins keep the whole workspace, and the scope guard’s list of credential and bot-state paths (.envfiles,~/.ssh,~/.aws,~/.claude,~/.telecoder, cloud CLI configs and so on) is refused for everyone, admins included. Symlinks are followed before anything is decided./telegraph <path>is admin-only;/telegraphon its own still shows the per-chat toggle to everyone. -
Telegraph is off unless you turn it on —
TELEGRAPH_ENABLEDdefaulted totrue, so every reply over 2500 characters, every/handoffand every verbose action log became a public telegra.ph page without anyone choosing that. It now defaults tofalse, in the code and in.env.example; set it totrueto get Instant View links back. Off, long replies arrive in full as chunked messages, as they already did underneath the link, and/handoffsends just the file. Turning it off globally now also stops chats that had switched it on with/telegraph, and/handofffollows the chat’s setting. -
ID lists are checked at startup, and strangers are told their ID —
ALLOWED_USER_IDS=@aliceused to parse toNaNand start a bot nobody could use (“Allowed users: NaN”), and the.env.exampleplaceholder123456789was taken as a real admin. A non-numeric entry inALLOWED_USER_IDS,ADMIN_USER_IDS,APPROVER_USER_IDSorALLOWED_GROUP_IDS, or the placeholder in a user list, now stops startup with a message saying what belongs there. Someone not on the list who messages the bot privately is told their numeric ID to pass on, and it is logged; a group that isn’t allow-listed gets no reply at all, where it used to get “not authorized” for every message when privacy mode was off. -
The bot token stays out of
psand out of error messages — photos, documents and voice notes were fetched withcurl, whose command line carriedhttps://api.telegram.org/file/bot<token>/…for any local user to read, and a failed download sent that command line, token and all, back to the chat. Downloads now use Node’s ownfetch, with the same timeouts and retries.sanitizeErroralso redacts Telegram bot tokens, API keys, bearer headers and credentials in URLs from anything it is given.
First run
-
A missing or wrong setting is explained, not dumped — with no
.env, a first start printed Zod’s raw JSON (invalid_type, “expected string, received undefined”) instead of saying a token was needed. Startup now prints one plain report: which.envwas read, or that none was found; each missing or invalid setting with where to get it (@BotFatherfor the token,@userinfobotfor your ID); andtelecoder initor copying.env.examplewhen there is no file. The ID-list and admin checks report the same way. It exits 1, with no stack trace. -
A bad bot token is caught before anything starts — a placeholder or revoked token ended in
Fatal error: GrammyError: Call to 'getMe' failed! (401)and a stack trace, after the IPC server had started. A token that isn’t shaped like one now fails the config check, and a well-formed one is checked withgetMebefore any other service starts: a rejected token says to get a fresh one from @BotFather, and a network failure says api.telegram.org couldn’t be reached and points atTELEGRAM_PROXY_URL. The token is never printed. Under the multi-bot launcher, a bot that stops on its configuration says so under its own name and isn’t respawned into the same error; fixing itsinstances.jsonentry starts it again. -
“Claude Code isn’t logged in” says how to log in — on a machine where nobody had signed in, every message came back as
❌ Error: Claude error: Claude Code returned an error result: Not logged in · Please run /login, and in PTY mode the CLI’s first-run theme and login screens were relayed to the chat as dialogs that led nowhere. Both transports now reply that Claude Code on the bot’s machine isn’t logged in, and to runclaudeand /login there (orclaude setup-tokenfor a long-lived token); an invalid API key or a 401 from the API gets the same reply. Startup also warns when there is no~/.claude/.credentials.jsonand noANTHROPIC_API_KEYorCLAUDE_CODE_OAUTH_TOKEN, without refusing to start (macOS keeps the login in the keychain, where the check can’t see it). -
doctorsees a login kept in the.env— aCLAUDE_CODE_OAUTH_TOKENorANTHROPIC_API_KEYset in the.envworks for the bot, butdoctorranclaude auth statuswithout it and reported “not logged in”. It now hands claude the login variables from the.env(and nothing else from it), its hint mentionsclaude setup-token, and its install hint is the native installer:npm install -g @anthropic-ai/claude-codefails with EACCES for a normal user on a distro’s Node. -
A fresh install keeps the name you gave the bot — the startup name sync renamed every bot to “TeleCoder”, BOT_NAME’s default, so the name picked in BotFather was gone on the first start. It only renames now when BOT_NAME is set, or when dynamic names are on (they’re built from it).
-
Node.js 22 or newer is required — Node 20 reached end of life in April 2026.
engines.nodeis now>=22, CI runs on Node 22 and 24, and the bot and the launcher check the version before loading anything else, saying which version they found and what to install instead of failing somewhere further in.
Multiple bots
-
instances.jsonis read while the launcher runs — adding a bot used to mean restarting the launcher, and with it every bot it ran. It now watches the file: a bot you add starts, one you take out stops, and one whose token, name or settings change restarts on its own and picks its sessions back up. The others carry on. Bots are matched by bot id, so a replaced token is the same bot restarting rather than one leaving and another arriving, and when a bot is taken out of the middle of atokenslist, the next one takes over its name once it has stopped. A file that doesn’t parse, lists no bots, or gives two bots one name is ignored until it’s fixed. -
/newbotcreates a bot for a project — a bot per project meant a BotFather chat each time: pick a name, copy the token, paste it intoinstances.json. Now/newbot sjsoffers a button that opens Telegram’s own create-a-bot screen with the name and username filled in, and once you confirm, the bot that sent it fetches the new token, restricts the new bot to you, and adds it toinstances.json, which the launcher picks up. It needs Bot Management Mode on in BotFather’s Mini App for the bot you send it to, and Threaded Mode, which Telegram offers no API for, is still switched on there by hand. Only a bot an admin created is adopted: anyone can create one managed by ours through a link. Asking Claude for a bot works too: itstelecoder_create_bottool sends the same button. The new bot sends the steps for turning threads on with the first message you send it, then checks once a minute and says so when they’re on, taking the setting in without a restart. -
A bot can have a home project —
HOME_PROJECT, which/newbotsets. A conversation with nothing to restore starts there instead of asking for/project, and the project is left out of the thread names and the bot’s display name while you’re in it, since the bot’s own name already says it. That leaves the thread names their full width for what they’re about. A thread that moves to another project gets that one in front again.
Threads
-
Each thread keeps its own model, effort, backend, transport and permission mode — in a bot with threads, a
/modelin one thread used to switch every thread in the chat, because these were stored per chat. Now they’re stored per thread. A thread that never picked one follows the chat’s setting, so a new thread starts out the way the chat was before threads, and resetting one in a thread goes back to the default, not the chat’s choice. Verbosity and the status line toggles stay per chat. Chats without threads are unchanged, and “apply to all bots” still applies to the other bots’ chats as a whole. -
Threads are named after what they’re about — a message sent from a bot’s “All” view starts a thread Telegram calls “New chat”, and every one of them looked the same in the list. The bot now renames a private chat’s thread to its session topic, and follows the topic as the conversation moves on, at most once a minute so the thread isn’t filled with rename notices. A name you gave the thread yourself, when creating it or by renaming it later, is never replaced. Group forum topics keep their names. The name is the topic’s first half, the broad area, which holds still while the second half follows each message. The project’s folder name goes in front,
telecoder · Thread naming, since one bot’s threads can each be in a different project; it’s left out when the topic already names it, and after a project switch it’s the whole name until the new topic comes in. Haiku is told the project, so the topic’s area names a part of it rather than repeating it, and a topic Claude or/topicsets while Haiku is still summarizing the message is no longer overwritten by Haiku’s late answer. -
A fork opens a thread of its own — on a bot with threads, the 🍴 Fork button and
/forkgo straight to a new thread, with no picker and no confirmation: someone who turned threads on wants them, and a thread you didn’t want is deleted as easily as a confirmation is cancelled. It opens a thread named after the topic with 🍴 in front, loads the conversation up to that point into it with the settings the thread you forked from chose, and replies with a link to it. The thread you forked from carries on unchanged, so both stay open side by side instead of one going to/resume. Bots without threads keep the picker.
Renamed
- The last Claudegram names — the agent’s tools are now
telecoder_*(telecoder_ask_user,telecoder_send_fileand so on) on a server calledtelecoder-tools. The state directory is~/.telecoder: the launcher moves~/.claudegramacross on its first start, before any bot runs, and leaves a link at the old path for anything still holding it. Files sent in the chat and/handoffnotes go to.telecoder/in the project, which gets its own.gitignore, so they stay out of git even in a project that only ignores.claudegram/. Existing.claudegram/folders in projects are left alone.
Removed
/medium,/reddit,/vredditand/extract, and the agent tools behind them — these content fetchers came from Claudegram: Medium articles through the Freedium mirror, which exists to get around Medium’s paywall, Reddit threads through a bot account’s OAuth app, Reddit videos, and YouTube, Instagram and TikTok downloads throughyt-dlp. None of them has to do with driving Claude Code, which can read a link itself. Theclaudegram_fetch_reddit,claudegram_fetch_mediumandclaudegram_extract_mediatools went with them, so a bare YouTube, TikTok, Instagram or Reddit video link is now an ordinary message to Claude instead of opening a download menu. Their settings (REDDIT_*,REDDITFETCH_*,VREDDIT_ENABLED,MEDIUM_*,FREEDIUM_*,EXTRACT_*,YTDLP_*andALLOW_PRIVATE_NETWORK_URLS) are ignored if they’re still in your.env, andyt-dlpandffprobeare no longer needed.
Agent
-
/steerand⚡ Interrupt & send— a correction no longer has to wait for the turn it corrects./steer <text>stops the running turn and sends the text in its place; Claude keeps the stopped prompt and what it did so far, so the text can refer to it. The same button sits on a⏳ Queuednotice, for a message already waiting (7cb7197). -
The status line turns amber, then red, as the context grows — the
📊 34%stat looked the same at 15% as at 80%, and with the 1M window even 80% is far past the point where turns get slower, costlier and less sharp. The chart now becomes🟡from 150k tokens or 60% of the window, and🔴from 300k or 80%, whichever comes first, a cue to/clear, fork or/compact. The usage footer and/contextbars follow the same rule. -
Fable gets a 💎 while it’s on — the model showed as one more
🤖stat in the status line, easy to read past, and not at all with the status line off, so a Fable picked for one hard task went on spending the plan’s limits on the easy ones after it. The live status now ends its first line with· 💎 fableevery turn, the status line shows💎 FABLE-5-1in place of the robot, and picking it in/modelsays what the badge means. Under CCR the alias says nothing about the backend, so it’s left unbadged. -
The activity card says at its bottom whether anything is still running — the running tasks were listed at the top of the card and the events grew below them, so on a long card only the events were on screen, and a card whose tasks had all finished looked the same as a busy one. It now ends on
⏳ 2 tasks runningor✅ All done. A card left behind when a turn starts says its tasks run on below it, since their end lands on the next card. -
The reply shows while it is being written — a reply used to reach the chat only once the turn had ended. It now appears in the status bubble as Claude writes it and is edited every second or two (every few seconds in a group, and less often after Telegram asks the bot to slow down), with the status line, or
✍️ Writing…without the terminal UI, underneath as the sign that it isn’t final yet. A reply that outgrows one message carries on in a silent message below. At the end the final text takes its place; a cancelled turn keeps what was written, with the banner under it. The SDK streams it as it is typed. PTY mode takes finished blocks from the session log and the one being written from the screen, a line at a time, and leaves out tool rows, the spinner and Claude Code’s own notices.CLAUDE_SDK_INCLUDE_PARTIALis gone: streaming needs it on. -
Formatting that can’t break a message — every message used to go out as MarkdownV2, which needs escaping, so an underscore in a path or a backtick in a command could get a reply refused and resent unformatted, and a long reply cut in the middle of bold text or a link broke both halves. Messages now go out as plain text with the formatting alongside as entities, from replies to menus, edits,
/status,/project,/recapand the background-task and monitor messages. Nothing needs escaping, and a long reply can be split anywhere: bold, links and code blocks carry on in the next part. A link Telegram won’t take, such ashttp://localhost:3000, is written out as text instead of failing the whole message. Code blocks now show their language. -
The task list is back — Claude Code stopped using
TodoWriteand keeps its list withTaskCreate,TaskUpdate,TaskListandTaskGetinstead, so the live checklist never appeared on a current CLI. The card now reads the list from where Claude Code keeps it (~/.claude/tasks/<session id>/), in both transports, and draws it the way the terminal does: a3 tasks (1 done, 1 in progress, 1 open)header,✔struck through for done,◼in bold for the one in progress,◻for the rest, and› blocked by #2while a blocker is open. It is posted on a turn’s first change and edited in place after that. A restored session (/resume,/continue, a restart, going back to a branch in/tree) shows its whole list again.TodoWritesupport is gone. -
Secrets that stay out of the chat — a task that needed a key or a token used to end in instructions to open a terminal on the machine over SSH and type it there, since anything sent in the chat, or run there with
!, lands in the session log, in the model’s context and in the Telegram history. The newclaudegram_request_secrettool has Claude name the secret and the command that consumes it, and shows you both. You send the secret as an ordinary message: the bot takes it out of the stream before anything can queue or forward it, deletes it, and runs the command with it in$SECRETand on stdin. Claude gets the exit code and the output with the secret blanked out. For a long-lived Claude token the bot runsclaude setup-tokenitself: it sends you the sign-in link, you approve on your phone and send back the code the page shows, and the token goes from the CLI straight into the command without passing through Telegram. It works in private chats with an admin, one request at a time, and/cancelor the Cancel button withdraws it.
Sessions
-
Every reply ends with the status line — it only came after a message you typed. A reply to a voice note, a photo or a file, to
/plan,/exploreor/loopwith the task on the same line, and the one Claude writes on its own when a background agent or task reports back, all left you on the tail of the answer with nothing to say where things stood. That last one is often the answer you were waiting for, and it tends to arrive while you’re in another chat, with no prompt of yours above it. Each now closes the way a typed turn does: your last prompts, the project, the model and the context, which under a background reply is the figure that reply left and not the one from the turn before. They get the compaction and new-session notices a typed turn gets as well, and a/loopshows the model and the context, which the line under one never had. A scheduled fire closes with a short line, the stats and the context bar alone: a schedule can fire every minute, and each fire counts as a prompt, so the list would soon show nothing else. A background turn that says nothing still posts nothing. -
The action log follows you past a question — when Claude asked you something mid-turn and you answered, the turn carried on in the action log posted before the question, so its next actions landed above the question and your answer, where you’d scrolled past them, and the bottom of the chat looked like nothing was happening. Answering (or the question timing out, or a secret request settling) now collapses that log and the next action opens a new one under your answer, with the status line and the prompt.
-
The status line names the project — a bot moves between projects and its display name only says where it is now, so scrolling back left you guessing which project an older answer belonged to.
📁 telecodernow leads the stats under each answer. After a switch it already names the new project, where your next message goes. -
A turn ends in one message — with the status line on under
/verbosity verbose, a turn closed with two messages that both gave the context and the cost, and each one notified you. The context bar now sits at the bottom of the status line, whose own 📊 and 💰 make way for it. With the status line off, the bar still comes on its own. -
/recapends like a turn — after a long reply,/recapleft you on the tail of an answer with the prompt it answers scrolled away, and it said nothing about the context, which is what has changed when you’ve just gone back to a fork or an earlier point. It now closes with the same status message as a normal turn: your last prompts, then the figure for the point the chat is at. In PTY mode that figure comes from the session log when no turn has cached one, so/statusnow shows its context line straight after a jump or a restart too, and/contextanswers from the log at those moments instead of starting a separateclaudeto ask. -
A restart carries on instead of asking — a chat quiet for between an hour and a week came back from a restart with “Continue where you left off, or start fresh?”, and a
/rebuildbot allleft that choice waiting in every bot you hadn’t used in the last hour. Every chat used in the past week now carries straight on in its conversation, as it would have with no restart at all;/clearis there when you’d rather start fresh. A project you’d picked but not talked in yet comes back with your next message too, where an hour on it used to answer “No project set”. A “Start fresh” still standing on an old prompt now runs/clear. -
A fork stands out where it starts — loading a fork left one plain line, ”✅ Fork loaded from …” or ”✅ Branched into a new conversation”, between two conversations that otherwise read as one thread when you scroll back. A branch taken in the same chat was the worst of it: the messages just above carry on past the point the branch starts from. The confirmation is now a banner — 🍴 Forked from … or 🌱 New branch, between heavy rules — that says outright a new conversation starts there and says the conversation above is in
/resume. It shows the exchange the fork picks up after: your prompt on one line, and the reply whole in a collapsed quote that opens on a tap — so a fork to another bot doesn’t leave that reply behind in the other chat, and the last reply above your next message is one the fork actually has. No summary: that would be a model call, and a retelling of the one thing you need verbatim. A reply too long to fit is cut at a word, pointing to/recapfor the rest. The ”✅ Forked to …” message in the sending chat now says that conversation carries on unchanged. -
A fork can be forked again from either chat — both messages confirming a fork now carry a 🍴 Fork button of their own: ”✅ Fork loaded” or ”✅ Branched” where the branch continues (after Accept,
/acceptor 🌱 New branch here), and ”✅ Forked to …” in the chat it was sent from. Forking the same point a second time — to a third bot, or into another branch — used to mean scrolling back through the source bot’s history to the message the fork was first taken from. Each button forks that same point, whatever either conversation has done since, so it hands over exactly what that message’s own button would. -
A fork offer previews the reply it was taken from — the “Fork received” message quoted the source chat’s latest reply as the last assistant message, so a fork taken from any earlier point arrived introduced by text it doesn’t contain. The preview is now read out of the cut transcript itself.
-
/resumesays which session each button is — a session is namedproject · topic, which is routinely wider than the single line Telegram renders a button on, and the client ellipsises the rest away with no way to expand it. Two conversations in the same project therefore arrived as the same half-line and you picked between them blind. The picker now writes every session out in the message body keyed A/B/C — name and age — and the same letter keys its button, the ruleclaudegram_ask_userhas followed since options moved into the body. Short names still ride on their own buttons (A · telecoder); the moment one is too wide the whole keyboard drops to bare letters, packed four to a row. Both features build that keyboard through one function now, so a question and a session list clip at the same place instead of each keeping its own threshold. -
The other two pickers that hid text on a button — the same rule now covers the “back to previous” button under
/clearand/project, and the/forktarget list. The back button named its session inside its own label and trimmed it at 45 characters, well past what a button renders, so the moment a session had a topic you were offered “back to” something you could not read. It keeps the name on the button while it fits —↩️ Back to telecoder (2h ago), the common case, adding nothing to the chat — and past that width puts the name in the message and a fixed label on the button. The fork picker listed sibling bots by name only on their buttons, so two bots sharing a prefix arrived as the same clipped line; targets are now written out A/B/C in the picker message. “New branch here” moved the bot’s own name off its label, which was single-handedly costing that keyboard its labels. Six commands offered the back button with four lines of identical keyboard-spreading each; that is one call now. -
/promptsreaches further back — it lists the last 15 prompts by default instead of 5, and/prompts Ngoes up to 25 instead of 20. Most prompts are a short line, so five used a fraction of the room one message has. Across 40 recent sessions, 15 always fit in one message, at about two phone screens. 25 fit in all but 5 of them, and in those the oldest are dropped to fit, as before, and the header says how many. -
The status line shows your last five prompts — with “Show last prompt” on in
/statusline, the line after each answer showed only the prompt that turn answered, and a turn often starts from an “ok” or a “yeah” that says nothing on its own. It now lists the last five, oldest first. They are read from the session log, so they reach back past a restart. A/compactand the notes Claude Code writes in your place, like “[Request interrupted by user]”, don’t take one of the five; a paste shows its text and an upload its caption, the way/treelabels a branch. -
The live action log says what the turn is for — in verbose mode the action log grows below your prompt until it pushes it off screen, so remembering what a long turn was working on meant scrolling back up to find it. The log now ends with that prompt, under the status line, as the same italic ▸ line the status line uses. A paste shows its text and an upload its caption; a skill run with a slash command shows as typed. It goes when the turn ends, with the answer posted below the log by then.
-
A prompt shows whole, and a long one its start and its end — the status line and the bottom of the live action log cut every prompt at 100 characters. Across 100 recent sessions that caught a third of the prompts written as one paragraph, and a cut that keeps only the start drops what a long prompt usually ends on: the question. A prompt now shows whole up to 300 characters, enough for 93% of those; past that it shows its start and its end with … between, cut between words. Five prompts under an answer take a median of 15 rows on a phone instead of 11.
/promptsand/recapcut the same way, at the 300 and 500 characters they already allowed. -
/treeshows each point’s prompt, not its first few words — a line of the tree had room for about 23 characters of the prompt a point starts from, so on the forked conversations on disk 38 of 44 points showed cut, most of them to a few words. The prompt now wraps onto lines of its own inside the drawing, with the branches carried on down beside it, and shows whole up to 300 characters; past that, its start and its end, the way/promptsshows one. Those trees take a median of 17 lines now, at most 47, where they took a line per point. A tree too big for one message gives every prompt the same shorter cut, so the longest give way first. Opening a point shows the prompt it ends on the same way, up to 300 characters instead of 160, and the list of forked conversations its prompts up to 300 instead of 80. -
A fork’s banner shows your prompt the way
/treedoes — the banner where a fork or a branch starts showed the prompt it picks up after cut at 120 characters, keeping only its start, and as the session log has it: a paste in its<pasted_content>tags, an upload as the note it reached Claude in, a reply under the quote it answers. It now shows whole up to 300 characters, past that by its start and its end, a paste by its text, an upload by its caption and a reply by the words written under the quote, the same as the point’s view in/tree. With the longest reply a banner holds, it still fits in one message. -
/promptsgoes back to any prompt — the list was something to read and nothing more: going back to an earlier point meant scrolling up past every reply since to that one’s 🍴 Fork button. Each prompt is now keyed A/B/C the way/treekeys its points, with a button per letter. A letter opens the prompt — its reply in a collapsed quote — with 🌱 New branch from here, which carries on just after that reply in a new conversation and leaves this one in/resume: the same banner and the same branch the reply’s own 🍴 makes. Anything sent after that reply stays out of the branch, a/compactincluded. A prompt where a branch already splits off is marked 🌿, and 🌳 shows the tree, where that branch can be carried on. Prompts read the way/treeand the status line show them, a reply under a quote by the words written under it./promptsalso answers while a turn runs, where it used to wait for the turn to finish, and after a restart lists where the chat left off, as/treedoes. Built from the transcripts on disk, 180 lists and 494 prompt views fit in a message, the longest at 3,325 characters. -
/promptsgives the prompt after a/compactits own line — a/compactleaves three records in the session log: the command, Claude Code’s record of it, and what it printed./promptsread them as part of the next prompt, so that prompt’s line opened with some 200 characters of<command-name>…</local-command-stdout>, terminal escape codes included, and only about 85 of its 300 were yours. Each prompt now gets a line of its own. Commands and the notes Claude Code writes in your place, like “[Request interrupted by user]”, are left out and don’t count toward N. A prompt interrupted before any reply was folded into the next one; it now gets its own line, marked ⏳. A paste shows without the<pasted_content>tags around it, in/recaptoo. -
/recapand/handoffshow only what you sent — the exchange after a/compactopened with the command’s three records, terminal escape codes and all, and an interrupted prompt carried “[Request interrupted by user]” into the next one. Both are left out now, there and wherever else exchanges are read off the log: a fork’s banner, a/treepoint, and the conversation carried over when you switch providers. A command that got a reply, like a skill run with a slash command, keeps its exchange./handoffalso writes a paste without the<pasted_content>tags around it. Across 312 recent sessions, 26 of 1,430 exchanges had such records in the prompt; the replies are unchanged.
Sharing a bot
-
The permission gate runs on the SDK transport and CCR — it was a hook on the PTY CLI only, while the default transport is SDK and CCR always takes it. There the curated tool list pre-approves Bash, Write and Edit, so a guest let in on a default install ran shell commands with no prompt at all. The SDK transport now registers the same check as an in-process PreToolUse hook, in every
/modeand for subagents’ tool calls, and a gate that throws denies the call. One test table runs through both transports, and/permissionsno longer needs its “this chat’s transport has no gate” warning. -
Auto-approve windows —
/autoapprove 30mlets the permission gate approve instead of asking, for a stated length of time on one chat. A job made of guarded commands cost a tap each, and a guardrail people tap without reading has stopped being one. Admins and approvers can open a window; it posts every command it waves through rather than going quiet, announces itself when it lapses, and lives only in memory so a restart closes it.mkfs,dd of=/dev/,DROP DATABASE,terraform destroyand credential reads keep their prompt — the line is whether a wrong call can be walked back, not whether it is dangerous./autoapprove offand/autoapprove statusdo what they say. -
An approver tier —
APPROVER_USER_IDSnames guests who may answer someone else’s permission prompt without owning the bot. Being able to tap “Allow once” and being able to run/restartbotor/methodwere one flag, which forced a choice between one admin being the bottleneck for every prompt and handing guests the switch that turns the gate off. An approver stays a guest in every other respect — the charter judge still reads their messages, the gate still stops their tool calls — and cannot approve their own request, so the prompt still means something./userslists them as their own tier,/denynames both lists so removing one doesn’t strand an id the orphan check refuses to start on, and the toast tells an approver why their own request is the one they can’t answer. -
Admitting people from chat —
/allow,/denyand/users. A stranger who posts in the shared group gets an approve card naming them, and someone added to the group is asked about the moment they join; one tap makes them a guest, with no.envedit and no restart. Telegram has no API that resolves a@username, so the bot matches only handles it has seen go by and/allowas a reply is the exact form. Ids admitted this way persist in~/.claudegram/user-roster.json; ids fromALLOWED_USER_IDSstay the operator’s, and/denysays so rather than appearing to remove one.hasGuestUsers()now reads the effective list, so admitting the first guest switches on the permission gate, the scope guard and the charter judge.
Sharing a bot (continued)
-
Contributors and spectators in a group — a group can now hold people who read along without being able to prompt the agent, which is the difference between a seat in the room and a shell on the host.
GROUP_MEMBERS_DEFAULTdecides what membership alone is worth,/allowand/denymove individuals, and/membersreads back one group the way/usersreads back the bot.GROUP_REQUIRE_MENTIONkeeps human conversation in the group from becoming prompts.This landed as a merge of two access models that had been built in parallel, and the reconciliation is the substance of it. The roster stayed the door — who may use the bot at all — and the group roles became a second gate behind it, so admitting someone in a DM makes them a contributor everywhere while a grant in a group stays in that group.
/allowtherefore means the nearer layer: the room when typed in the room, the bot when typed in a DM.Three things were dropped rather than merged. The branch’s
OWNER_USER_IDScomputed exactly whatisAdmin()already did, so it went andownerbecameadminthroughout. Its owner-only command list duplicated theadminOnlywrapper the handlers already carry — but its judgement was better than main’s coverage, so/project,/newproject,/permissionsand/teleportare now wrapped too. And its username→id cache was a second answer to a question the roster already answers, free to drift from the first.GROUP_MEMBERS_DEFAULTships asspectator, not thecontributorthe branch defaulted to. That default was right when Telegram membership was the whole gate; layered on top of the roster it would have widened who can prompt the agent, so the merged default matches what the roster already enforced.
Reliability
-
A long paste reaches Claude as one message again — Telegram splits a paste over 4096 characters into several messages. Since messages started queueing behind a busy turn, each part waited for the one before it to be answered, so Claude saw the paste one part per turn. The batcher that joins them was off unless
MESSAGE_BATCH_TIMEOUT_MSwas set. It’s now on by default (1500ms), and only a message of 2000 characters or more opens a batch, so ordinary messages aren’t held up waiting for parts that won’t come. The parts after the first join it whatever they look like, including a short last part or one that starts with a slash because the split landed just before a path. -
No more catch-up repeating a reply — the reply goes out without its reasoning summary, but the session log keeps it, and the check that recovers prose the relay dropped counted the summary as missing. With summaries on, every reply got a ”📨 Catch-up — recovered from session log”, and one with a tool call in it was posted a second time in full. The check now leaves the summary out of the log’s text as well, and still recovers prose that really went missing.
-
The live action log gives its time and its count at both ends — the log is usually taller than a screen, the count of actions sat only at its top and the time it was last drawn only at its bottom, and both change all the time, so reading either meant scrolling to the other end. From the top the log also looked frozen through a long step, since nothing up there moved until the next action came in. A running log now opens and closes with the same line,
🕐 Updated 2:11 PM · last 9 of 20 actions, with the count left off while every action is showing. -
The live action log says when it last moved — Telegram shows when a message was sent but not when it was last edited, so a live log that had stopped moving looked just like one that was still going, and a long quiet step left you wondering whether the bot was still alive. The log now ends with the time it was last drawn, under the prompt. With the status line on it’s redrawn every ten seconds, so a time that keeps up with your clock means the bot is alive however long the step takes. That time and the one on each action now read in your own timezone; they were the server’s.
-
Commands that only read answer during a turn — every command behind the queue waited for the running turn, so
/commandssat behind a long turn, acked as queued, to print a fixed list once Claude was done./commands,/sessions,/botstatus,/recap,/projectcommandsand/contextnow answer at once. Commands that change the session or start a turn still wait their turn, as a message does. -
/latersets an idea aside before it’s sent — Ask me after is only on a message already waiting behind a turn. An idea you had while nothing was running had nowhere to go, and one sent as a turn was ending could go out before you’d tapped the button./later <text>sets it aside straight away, with the same question, Send and Drop afterwards. Sent while Claude is idle, it waits until the next turn has run, and/lateralone lists what’s set aside in the chat. A question left unanswered while the conversation moves on is now asked again below each new reply, without a second notification, instead of staying above them all. -
Decide on a follow-up after you’ve read the answer — a message sent during a turn went out as soon as that turn ended unless you took it back, so you had to decide before you’d seen the answer it depended on. The
⏳ Queuednotice now also has Ask me after. The message won’t go out by itself, and nothing behind it waits for it. Once nothing is running or waiting, the bot asks below the reply whether you still want to send it, with Send and Drop, and that question notifies you, unlike the notice. It stays until you answer it, through/canceland restarts. Send puts the message through as if it had just arrived, so a voice note is only transcribed then, and it waits its turn if another one has started. An edit made while a message waits, set aside or not, is now what goes out; Claude used to get the text as first sent. -
The subagent panel can’t go out as a reply — when the session log had no reply text for a turn, the bot read one off the screen, and with a subagent running it could send the panel under the input box instead. A log that holds your message but no reply now gives the usual “Done”. The screen is only read when the log has no record of the message, and it skips the panel, tool calls and older replies. Each time the log has no reply text, the bot logs enough to show why.
-
Messages get through while a subagent is running — the panel under the input box shows each running subagent with a live timer and token count, and the bot took those updates to mean the screen hadn’t settled. Every message then failed after a minute with “Claude Code’s input box never appeared”, for as long as the subagent ran, though the box was open the whole time. The bot now ignores those rows when it checks whether the screen has settled.
-
/allowand/denyapply to the bot you send them to — all the bots kept their guests in one file. A guest let into one bot reached the others at their next restart, along with the guardrails a first guest switches on, whatever those bots’.envsaid about who could use them. A/denycould also come undone: the other bots kept allowing the guest until they restarted, and the next save from any of them wrote the guest back for everyone. Each bot now has its own roster. On its first start it copies the shared one, so nobody loses access, which means a guest admitted before this stays on every bot until you/denythem on each. -
Favorites stay in step across your bots — each bot read the shared list at startup and saved from that copy, so a favorite added on one bot didn’t show on the others until they restarted, and could vanish, or a removed one come back, when another bot saved. Each bot now reads the list whenever it needs it.
-
A setting stays with the bot you set it on —
/tts,/terminalui,/telegraph,/suggestionsand/botnamekept their settings in files all the bots shared, and each bot rewrote a file from the copy it loaded at startup. A change on one bot could be undone by a save on another, and one that survived reached the other bots at their next restart. Each bot now has its own files, so changing a setting everywhere means setting it on each bot. On its first start a bot copies what the shared files hold, so nothing resets. -
One bot can’t undo what another remembers about its name — the bots shared a file of the display name each last sent, and each rewrote it from the copy it loaded at startup, putting back the others’ old entries. After a restart, a bot could take a real rename for no change and keep showing the old project, with no notice, until the next switch. The file of Telegram’s rate-limit blocks had the same flaw, so a bot could forget a block and call again. Each bot now keeps its own files. The old shared ones are no longer read, which costs at most one extra call per bot the first time it renames.
-
/versionsays whether you’re on the latest — the version a restart names tells you what came back, but not whether anything newer has landed since. A fix you’d committed and never built looked just like one that was live./versionnames the running build, compares it with the code in the checkout and the branch it tracks, and says what catches it up:/restartbotfor a build that’s done but not loaded,/rebuildbotfor code changed since,git pullfor commits that aren’t here yet. It compares the contents of what a build compiles rather than commits, so a docs commit, committing what was already built, or a history rewrite doesn’t count as falling behind. A check that couldn’t run says so instead of passing, and it answers during a running turn. -
A message you sent too soon can be taken back — a message sent while the bot was busy got no answer at all, so a bot that was only still working looked idle, and you found out when the reply to the last message landed. Deleting it didn’t help, because Telegram doesn’t tell a bot about a delete: Claude got it anyway. A message that has to wait now says so within half a second, as
⏳ Queued behind the running turn, with a Take back button that drops it before it reaches Claude and leaves the turn running. Text, voice notes, photos and files are all covered, and a bot command gets the same notice while it waits. That includes the turn you can’t see, one Claude started itself on a background task’s result: your message waits for it without a “thinking” bubble going up, and that turn’s reply still arrives whole./canceland/stopnow clear the queue. They used to stop the running turn and thereby start the message waiting behind it, and now they also reach a turn Claude started itself.CANCEL_ON_NEW_MESSAGE, which never took effect behind a normal turn, now does what.env.examplesays. -
A cancelled message stays cancelled — a message you stopped with
/cancelbefore Claude had written anything came back. Claude Code put it back in its input box, and it went out stuck to the front of your next message as one prompt, while a/costor/taskswould have submitted it by itself. The box is now emptied before anything else is typed into it, and a/cancelthat lands while a message is still being typed in stops it from being submitted at all. -
A restart says which version came back —
/rebuildbotreported a build and a restart but never what the bot came back on, so there was no telling from the chat whether a fix had gone live. The message a restart ends with now names it, as “Version: 63dcf45 feat: end /recap the way a turn ends (built 2m ago)”, with “plus uncommitted changes” whensrc/had some at build time, and/botstatussays the same whenever you ask. It comes from a stampnpm run buildleaves indist/rather than from what’s checked out, so a commit you haven’t built yet isn’t claimed as running. WithAUTO_RESTORE_SESSIONoff a restart posts nothing afterwards, so/botstatusis the place to look there. -
The bypass warning stays accepted — since Claude Code 2.1.314, every new session (after
/clear,/model, a project switch or a restart) opened with “WARNING: Claude Code running in Bypass Permissions mode”, even though you had accepted it, and “Yes, I accept” didn’t last past that session. The bot’s launch leaves your user settings out, which is where that answer is kept, and that version stopped reading it there. The bot now passes the answer along when your settings already have it. An account that hasn’t accepted still gets asked, once. -
/rebuildbotcan wait for each bot to finish — rebuilding while bots were busy on other projects meant waiting for a moment when none of them was, or cutting a turn off halfway. The/rebuildbotmenu now has “All instances, each once it is idle” (/rebuildbot idlefor short): it builds once, then restarts each bot the moment it has nothing left to finish — no turn running or queued, no message still being handled, no background task still going, and no reply still on its way — instead of waiting for the busiest. One message in the chat you asked from keeps the tally, saying what each bot still waiting is busy with, with buttons to restart the rest now or call it off. The launcher has to be running this build before it can take the request; until then the message says so and nothing restarts. -
Claude’s answer after a background task arrives whole — when a background task finished between your messages, what Claude wrote about it went onto the background activity card, under a ”📡 Monitor” line, cut at 1,200 characters, and the rest was only in
/recap. That answer is often what the background work was for, and in two weeks of session logs 50 of the 119 Claude wrote ran past the cut. It now comes as a reply of its own once Claude finishes, sent the way any reply is: whole, split over messages only where Telegram needs it, and with a phone notification. The card keeps the tasks — what’s running, what finished. A monitor firing still lands on the card at once, and what Claude says about it follows as a reply. If a question or a file comes in the middle, what Claude wrote before it goes out first, above it. -
A restart stops its Claude Code sessions before the new bot takes over — a restarted bot left every Claude Code session it had open to carry on until that session noticed. On 10-27 one wrote another reply into its conversation’s log a second after the new bot had already resumed that conversation in a session of its own, so two processes had one log open. Shutting down now hangs each session up and waits for it to exit, which takes under a second when it’s idle, and kills one still running after four seconds. A reply the restart cuts short isn’t reported as an error on the way out, and a message that arrives mid-shutdown is told to come again.
-
A question asked after a background task reports back reaches you — when a background task finished between your messages, Claude Code started a turn of its own to deal with it, and TeleCoder had no turn on record to run its tools in. A question it asked with buttons failed within a second with “MCP error -32602: Invalid tools/call result” and a schema dump, and never reached the chat; sending a file, a poll or a schedule failed the same way. The two weeks of logs on hand hold seven such calls across two bots. Those tools now reach the chat the session belongs to. A question or a file posted there seals the background card above it first, so what Claude writes after your answer lands below it. A tool with no chat to reach now says so in words.
-
A question keeps its turn while it waits — Claude Code 2.1.311 moves an MCP call that is still running after two minutes to the background, and the turn carries on without it. A question you took longer than that to answer ended its turn, and your answer came back later as a background task instead of in the reply. TeleCoder’s Claude Code sessions now let a call run until it would time out, which is past the question’s own ten minutes.
-
The context reading after a
/compactshows what it left — under “🗜️ Context compacted — 536.4k → 13.7k tokens” the usage footer still read 54%, the context from before, and only the next reply brought it down. A compaction writes no usage of its own, so the newest figure in the session log was the one for the conversation it had just replaced. The reading now drops to what the compaction left, 1% there, in the footer, the status line and/contextalike. That figure counts the conversation only, so the next reply, which adds the system prompt and tools back, reads a few percent higher. -
/statusreads the context the footer does — its 📐 Context line added up the reply and the input sent fresh, and left out what came from the prompt cache, which once a conversation is under way is nearly all of it: 0% (322/1.0M) where the footer read 54%. It now counts cache reads too, the same sum as the footer. -
Claude Code stays on the screen layout TeleCoder can read — Claude Code 2.1.310 can start in its new fullscreen layout without being asked: a trial, a fresh install or a remote feature switch turns it on. Everything TeleCoder reads off the screen — the input box, the spinner, the dialogs — is built for the classic layout, and in a test run that came up fullscreen,
/mcp’s tool list offered “Enter — select” on a list whose cursor it couldn’t see. TeleCoder’s Claude Code now always starts in the classic layout. -
Closing a menu no longer sends your last answer again —
/mcpand the other Claude Code menus write no reply of their own. When one closed, TeleCoder went looking for the latest reply in the session log, found the one before the command, and posted it again as the menu’s answer. The chat now gets what the command printed — “Reconnected to sjs-dev.”, “MCP dialog dismissed” — or ✅ Done when it printed nothing. -
Claude Code’s menus can be walked from the chat — picking a server in
/mcpopens that server’s own screen, but TeleCoder expected every tap to close the dialog it answered. The tap went in, the menu stayed open on the server’s screen, and the turn ended calling the dialog “still open”: the menu could be opened from the chat but never used. Each screen a tap leads to now comes to the chat in turn — a server’s actions, its tool list, the list Esc steps back to — until Claude Code is back at its input box. A list too long for the screen, like a server’s 111 tools, offers every row it shows instead of its first seven, with buttons to scroll it, and the↓ 92 moreunder it and the group headings between/mcp’s servers are no longer offered as rows. -
A message that starts with a long link gets through — Claude Code wraps its input box at word boundaries. In a message like “I’m about to submit https://…long link…”, the link moved down a row, which split the opening words TeleCoder looks for to confirm a message landed. Every attempt was reported as “didn’t take your message”, however often you resent it. Worse, clearing the box between attempts removed only the bottom row. Copies of the opening words piled up three per attempt, and they were sent along with your next message. TeleCoder now finds the opening words across a row break, and clears every row it filled.
-
A command moved to the background is counted while it runs — Claude Code doesn’t kill a Bash command that runs into its 10-minute limit: it moves it to the background and carries on. TeleCoder only counted commands started in the background, so one moved there was missing from the ”🔄 N tasks running in background” footer,
/tasksand the background card for as long as it ran. It now counts from the moment it moves, and its running time starts from when the command did. -
A crashed Claude Code no longer leaves its background tasks counted as running — when Claude Code exits on its own, killed for running out of memory say, the tasks it ran in the background end with it. TeleCoder only cleared them when it shut Claude Code down itself, so after a crash they stayed in the count indefinitely. They’re now cleared however Claude Code exits.
-
A background task that finished mid-turn no longer counts as running — if a background task finished while Claude was in the middle of a turn, the ”🔄 N tasks running in background” footer,
/tasksand the background card went on counting it as running until the session ended. TeleCoder 4 showed two tasks running with nothing running at all. Claude Code folds a completion like that into the turn in progress instead of logging it as a message of its own, and TeleCoder only looked for the message. It now reads the folded-in completion too. -
A running turn no longer keeps “typing…” up — every turn renewed Telegram’s “typing…” indicator every four seconds until it ended, and Telegram Desktop draws that animation on the CPU for as long as it shows, so a long turn kept the client busy the whole time. A turn with the live status line (
/terminalui, on by default) no longer sends it: the status line already shows the turn is alive. With the terminal UI off, “typing…” stays, since nothing else would. In verbose mode the status line now sits at the bottom of the action log, because the log is posted below the status bubble and is usually tall enough to push it off screen. That left “typing…” as the only live sign you could see. The bubble points down at the log, and the log’s header starts with ⏳ while the turn runs, so the chat list shows it’s still going. The waiting-for-your-answer line now points down at the question too; it said “above”, but the question is posted below it. -
The live action log catches up — the log was edited at most every ten seconds, and an action that arrived sooner was left off until another one came in. After a burst of quick reads followed by a long build, the log stayed a couple of actions behind for the whole build. It now draws the waiting actions once the ten seconds are up, and after a Telegram rate limit it draws again instead of dropping the change. Commands and output go out as entities rather than MarkdownV2, so they show verbatim and nothing in them can break an edit.
-
A running command reads as code in the status bubble — while a Bash command ran, the status bubble showed it as plain text under the timer, so a long one-liner, such as a loop polling
gh pr checks, sat there as a raw run of pipes, quotes and$(...)for as long as it ran. The command is now in abashcode block, the way the permission prompt shows one, and the block’s copy button hands back just the command. It goes out as Telegram entities rather than MarkdownV2, so nothing in the command can break the update. -
A message of several lines is acted on, not held for a confirmation tap — Claude Code now wraps a long paste in
<pasted_content>tags, and its system prompt has the model act on instructions inside one only where your own words outside the tags ask it to. TeleCoder delivers every message as a paste, so anything over two lines or 800 characters arrived as one tagged block with nothing outside it: the model took your request for text copied from somewhere else and asked for a button tap before doing what it said. The system prompt a PTY session starts with now says what the tags mean here — how the message was delivered, not where it came from. Text you quote or forward, and pages TeleCoder fetches for you, are still read as someone else’s. SDK mode never pasted, so it was never affected. -
Claude Code’s dialogs arrive laid out for a phone — a dialog relayed to the chat, most often a permission prompt (Claude Code raises some even in auto mode, such as a heredoc whose braces and quotes trip its “expansion obfuscation” check), came through as its raw terminal render. It was hard-wrapped at 120 columns and then rewrapped by the phone, the command sat in a
│gutter, an edit’s diff sat between two screen-wide rows of dashes, and the rows repeated the buttons behind a❯cursor. Blank lines were dropped, so title, command, reason and question ran together. Now the title and the question are bold, the command or diff is in a code block, prose Claude Code wrapped is rejoined, and each choice is written out once in full under its button’s number, since the button clips it. The last line says what a tap does, and a dialog that opens mid-turn no longer promises to “deliver your message”. Answering the dialog, or letting it time out, keeps the formatting. The answered edit also used to strip formatting from the bot’s own permission prompts; it keeps it now too. -
A relayed permission prompt no longer offers “Tab — amend” — the button pressed Tab, which doesn’t answer the prompt: it turns the highlighted row into a text field (“Yes, and tell Claude what to do next”) for a note a button can’t type. The dialog stayed open, so the turn failed with “Claude Code hasn’t returned to its input box”. Tab is left out of the buttons now, along with the other keys that change a dialog without answering it.
-
Claude Code asking to fetch a page reaches the chat — WebFetch’s permission prompt was never relayed. The dialog reader found a dialog by its footer, the “Esc to cancel · …” line under the choices, and this prompt has none: it ends on its last choice, with the one key it names written into that choice (“No, and tell Claude what to do differently (esc)”). A turn that wanted to fetch a page in a mode that asks first sat on the dialog, with nobody asked, until the turn’s hard ceiling. A numbered list with the cursor on it, under the dialog’s top rule and reaching the bottom of the screen, now counts as the end of a dialog. It gets its choices as buttons, and no key buttons, since it names none.
-
A dialog left open mid-turn no longer says your message wasn’t delivered — when a dialog that opened partway through a turn went unanswered, or couldn’t be answered from the chat, the turn ended with “your message wasn’t delivered. Send it again”. It had been delivered: the dialog came from the work it started, and resending it only queued a duplicate. Mid-turn it now says the bot stopped waiting on the turn, that Claude Code is still paused on the dialog, and that your next message brings its buttons back. Before a message goes in, the wording is unchanged, since there the message really wasn’t typed in.
-
The last message of a long reply lost its formatting — when a reply was split across messages at a point inside a code block, the splitter closed the block at the end of one part and reopened it at the top of the next, but then counted that reopening fence as a closer as well. The next part ended with a stray fence Telegram couldn’t match (“Can’t find end of Pre entity”), so it was rejected and re-sent as plain text, with its bold and code showing as raw
*and backtick markers. Past two parts, every fence after the first cut was counted the wrong way round. Each part now starts outside a code block, with the reopened fence as part of its own text. The first line of a reopened block also keeps its indentation, which used to be trimmed off. -
A quote under a list item no longer costs the message its formatting — a
>quote on the line right under a list item came out indented, as that item’s continuation. Telegram only reads>as a quote at the very start of a line; anywhere else it is a reserved character, so the message was rejected and re-sent as plain text. The same went for a collapsible<expandable>quote written on an item’s line. Quotes now always start their line, which is the only way Telegram draws them anyway. A$in a collapsible quote also stays as written;$$used to collapse to$. -
Long replies stopped being replaced by a 200-character stub — with Instant View on, any reply past 2500 characters — or one merely containing a markdown table — was swapped in chat for
text.substring(0, 200) + '...'and a telegra.ph link. The answer never arrived, the stub was cut at a raw character offset mid-word with its markdown stripped out, and the only way back to what the model wrote was/recap. Telegraph is additive now: the reply goes out in full, chunked as usual, and the page URL hangs under the last chunk — or takes a chunk of its own when appending would overflow it. A telegra.ph outage costs the link, never the text. The table trigger is gone outright;convertToTelegramMarkdownhas rewritten pipe tables into bullet lists for a long time, so there was nothing left to compensate for. -
Turns started from a voice note, a photo, a text file, a Reddit summary,
/planor/explorehad no catch-up —relayCatchUpIfMissedis the safety net that compares what was relayed against the canonical assistant text in the session log and posts whatever is missing. Text messages, force-reply prompts and scheduled runs called it; these turns never did, in streaming or wait mode, so they also never updated the relayed tracker — leaving/synccomparing a fresh log against a stale baseline and offering to re-post an earlier turn’s reply. The/loopcommand is deliberately still out: it concatenates several turns, and the single-turn comparison would either no-op or duplicate the last iteration. -
/restartbotno longer kills a systemd-managed bot for good — the restart ranbotctl recoveras a detached child, butdetached: trueescapes the process group and not the cgroup. Under systemd the helper was reaped along with the bot the moment it exited, before it reached the start half of its job, and the clean exit toldRestart=on-failurethere was nothing to fix. The bot now recognises that systemd is holding it up and exits75for systemd to bring it back, instead of trying to start its own successor. It checks the unit’sRestart=first and refuses — while still running — rather than announcing a restart it cannot come back from. The shipped unit moves toRestart=always. -
A lost hook could outrank the Stop signal and hang a finished turn — the end-of-turn check refuses to finish while the hooks say a tool is still open, which is right:
claudegram_ask_userlong-polls for a button tap and the pty is silent for as long as it waits. But that count was checked before the Stop hook, on the reasoning that Stop only fires once every tool is done.It does — which is exactly why a count that survives Stop cannot be a live tool. The hooks are
curlcalls into a loopback server, fired and forgotten (>/dev/null 2>&1; exit 0); a POST that fails takes its decrement with it, nothing retries, and nothing else ever brings the count down. One dropped request then outranked the only authoritative end-of-turn signal there is, and the turn ran to the two-hour ceiling with claude sitting idle at its prompt and its finished answer already written to the session log. A chat waited an hour and seven minutes past “Done” for a reply that never came.Stop now wins, the stale count is logged and cleared rather than silently believed, and the count keeps guarding the path where it earns its keep.
-
Answering a dialog could end the turn and strand the session — the end-of-turn check infers “claude is finished” from a quiet pty, the input box being back, no spinner, and the session log having moved. Every one of those is true in the moment after a relayed dialog is answered, and none of them mean what they usually do: the box is back because the dialog went away, and the spinner is missing because claude has not resumed yet.
A turn resolved in that gap does not just finish early, it takes the session down with it. The bot hands back an answer and forgets the turn, while claude carries on working with nobody listening — hooks arriving with no active turn to attach to and being dropped, output going nowhere, and every later message held against a screen that now does say “generating” until the ceiling rejects it. One live session spent 37 minutes unreachable that way, and
/stoponly reaches it while a message happens to be in flight.The turn now waits for claude to show itself after a dialog is answered — a spinner, or a tool opening through the hooks — bounded at 15s so a dialog whose answer genuinely did end the turn still resolves. The Stop hook stays authoritative and is never waited past.
-
The PTY could not find the
claudebinary under systemd —CLAUDE_BINfell back to the bare nameclaude, which node-pty execs directly against the service’s PATH. A systemd user unit gets the bare default, without~/.local/bin, which is where the native installer puts the CLI. execvp wrote “No such file or directory” into the pty and exited, so the bot saw a session that produced nothing and reported that the input box never appeared.resolveBin()has searched~/.local/binfirst on Linux since it was written, for this exact reason — the pty provider just never called it. @code_share1_bot had not completed a single turn since the day it moved to a systemd unit; it went unnoticed because the failure names a symptom two layers above the cause. -
A dialog with a status bar under it is still a dialog —
parseModaltook the last non-empty line and required it to parse as key hints, so anything claude drew beneath the footer (its⏵⏵ bypass permissions onstatus bar, a transcript warning, a welcome notice) made the parser answer null. The readiness loop reads null as “a screen we cannot drive” and falls through to the timeout, so the dialog that should have arrived in the chat as two buttons instead spent its whole ceiling being silent and came back as “Claude Code’s input box never appeared”.This is how @code_share1_bot went quiet after a CLI update began asking about a folder it had been working in for weeks: the trust prompt was on screen, recognisable, and never relayed. The footer is now located as the lowest hint-bearing line rather than assumed to be last, bounded to three lines of chrome so a stale footer scrolled up in the transcript cannot pose as this screen’s.
-
A stuck TUI now says what it is stuck on — the warning named the symptom (“input box absent”) and nothing else, so diagnosing the above meant spawning ptys by hand outside the bot to see a screen the bot had already read. It now logs the last dozen screen lines alongside the verdict.
[1.1.0] — 2026-08-31
The release that makes a bot shareable. Everything before this assumed one person on one machine; this batch adds the access model, the guardrails and the visibility that a bot with other people in it needs.
Sharing a bot
- Admins and guests (
9470e1d) —ADMIN_USER_IDSnames the subset of allowed users who may approve permission prompts and run lifecycle and transport commands. Unset, every allowed user is an admin, so single-user installs are unchanged. Includes a scope guard, a charter read before anything runs, and prompt-hold so a guest’s request waits for an admin. - Its own Unix account (
e72ee84) —scripts/setup-shared-bot-user.shprovisions the account that decides what is actually reachable: no sudo, no docker, its own checkout, and a--verifythat becomes the account and tries to read the operator’s secrets rather than reasoning about file modes. - Resource ceilings (
d8efbd9) — MemoryHigh throttles before MemoryMax kills, CPUQuota leaves cores for everything else, IOWeight yields disk. Disk is reported rather than enforced, since only filesystem quotas would be honest. - Per-bot model and effort (
365d134) —/modeland/effortapply to the bot you ran them in, with an opt-in fan-out to the rest.
Added
/prompts(680c436) — lists just the prompts you sent, one line each, without the replies. The fastest way to remember what a conversation was about./cost(562c205) — usage limits on a subscription, dollar totals on API billing, plus a running per-conversation total that Claude Code cannot keep because its own counter dies with every turn.CLAUDE_PLUGINS(7cdbf09) — brings marketplace plugins into the agent. Enabling one in the terminal writes to a settings file neither transport reads, so its skills and commands were silently missing in Telegram.- Claude Code’s own slash commands (
86d9622) —/commandsreports which of them work here. EXTRA_MCP_CONFIG(11dae19) — load MCP servers you name yourself, without dragging in every server the machine has ever registered.
Fixed
- Refuse to start on an unreachable admin (
2f19ba1) — an id inADMIN_USER_IDSbut notALLOWED_USER_IDScan never act. This replaces a warning that went to a log nobody reads and surfaced days later as an approval sent to the wrong person. - Truncated replies after a background task (
6628922) — a task reporting in mid-turn wrote a user-role record that the transcript reader treated as a turn boundary, dropping everything above it. - MarkdownV2 code spans (
11ab6fc) — added the escaper they need, which has different rules from ordinary text.
1.0.0
Everything before this changelog existed. See the git history.