Send Codeep an instruction from Telegram and it runs, with the answer coming back to the same chat — plus Claude Fable 5.1, a Sonnet 5 price that was 50% too high, and cache reads on Kimi and Qwen that were billed wrong in opposite directions.
Added
- Start a task from your phone. V3 let Telegram answer a question the agent
chose to ask. This lets it ask one: send the bot an instruction in plain words and it runs, shown in the terminal like any other, with the agent's answer coming back to the same chat.
Its own switch, off by default — /settings → Start tasks from Telegram. A chat that can only say Run, Skip or Cancel is bounded by what the agent already decided to do; a chat that can send a prompt is a keyboard attached to your machine. Both are gated on the same chat ID, but they are not the same permission, so turning one on does not turn on the other.
An instruction sent while a run is in flight waits, and the bot says so. One waits at a time and a newer one replaces it: two corrections typed during one long run mean the second, not both in order.
Four refusals, each with something to say back rather than silence. Non-text, slash commands, anything over 2000 characters, and anything older than five minutes — Telegram holds undelivered messages for 24 hours, so without that last rule an instruction typed to a machine that was switched off would run the moment the CLI next started, hours later, with nobody watching. A message from any other chat is ignored with no reply at all: answering confirms the bot is live and attached to something worth probing.
A task started from the phone is not exempt from confirmation. It stops at a dangerous tool exactly as it would have for someone at the keyboard, and that question arrives on the same phone.
- Telegram says when a run is over. Past a minute of run time, so a phone is
not buzzed about work you watched finish. It carries the task, the duration and the token or cost line. The agent's answer travels only for a run the phone started — at the terminal it is already on the screen you are sitting at, and sending it would put file contents or a command line into a chat that syncs to Telegram's servers for nothing.
- Claude Fable 5.1 (
claude-fable-5-1). Prices like Fable 5, but reads a
cached token at 0.025× the input rate where every other Anthropic model charges 0.1×, so the cache rate is resolved per model before per provider now. Fable 5 stays listed: dropping an id does not leave a pinned config on the older model, it drops the lookup and lands you on another provider's default.
Fixed
- Claude Sonnet 5 was priced 50% high. The $2/$10 launch rate was billed as
introductory through 2026-08-31 and the scheduled rise to $3/$15 was cancelled. Nothing failed when the table still said $3 — the number was simply wrong, with no test to notice.
- Cache reads on Kimi and Qwen were billed wrong, in opposite directions.
Two errors, and the first hid the second. Cache hits were read only from prompt_tokens_details.cached_tokens; Kimi reports them at the top level of usage, so every Kimi cache hit came back as zero and the whole prompt billed at the cache-miss rate. And where a cache read was seen it billed at 0.1× — Anthropic's multiplier, hardcoded for everyone — while Kimi lists $0.19 against $0.95 and Alibaba prices Qwen's implicit cache at 20% of input. Only the metered keys were affected; the subscription tiers quote no dollar figure at all.
npm run devhad been dead for two weeks. Two interfaces were re-exported
with value syntax. tsc erases the types and compiles happily, so the build, dist/ and the published CLI were all fine and CI had nothing to report — but tsx transpiles file by file, leaves the re-export in, and the loader then fails to find an export that never existed at runtime. isolatedModules is on now, which named all eleven sites of that shape in one pass.
- A run reported itself by session id, not by its task. The dashboard, the
stats report and the Telegram notice all read a name that was copied by value before it was set, so session-2026-09-02-ddc1f13c went out in place of the task. And the notice used the session's name, which is the first task's and stays put — every task after the first announced itself under the first one's name.
- The agent refused commands it is allowed to run.
execute_commandwas
described as "Use for npm, git, build tools, tests, etc.", and a few examples read as an exhaustive list: it declined sleep, which is on the allowlist beside echo, pwd, date, ls, cat and curl, explaining that its tool was limited to package managers and version control. It believed the description over its own capability.
- The "N new · PgDn" badge offered a key that does nothing mid-run. While
the agent runs, the timeline owns the whole screen and returns before the transcript is drawn, so scrolling moves an offset nothing reads. The badge also replaced the status bar's runtime and token counts to say it.
- A failing Telegram poll looked like an idle bot. Errors were caught and
turned into null with nothing said, so a webhook left configured (409) and a revoked token (401) were indistinguishable from a phone nobody had messaged. It now says so once when polling starts failing and once when it recovers.
Changed
- One Telegram poll, one cursor.
getUpdatesconfirms every update older
than its offset regardless of allowed_updates — that parameter only filters the response. A second poller would have acknowledged approval taps it never delivered, so approvals would have begun failing intermittently with nothing to say why. The loop is shared now, asks for both kinds, and dispatches locally where losing one is impossible.