TL;DR — Claude Haiku 5.5 arrives at $0.10/$0.50 (five times that past 100K prompt tokens), far fewer runs end "without a summary", and Always Allow now lasts the whole chat.
Added
- Claude Haiku 5.5 (
claude-haiku-5-5), Anthropic's new fast model for
high-volume work: 1M context, 128K output, and thinking on by default at medium effort. /thinking offers low to max for it, and Auto leaves it at medium. Codeep leaves it at least 32K reply tokens (64K at Max), as for Opus 5.5 and Sonnet 5.5, because its thinking shares that limit, and sends it no temperature, top_p or top_k, which it rejects. Through OpenRouter it is anthropic/claude-haiku-5.5. Claude Haiku 4.5 stays in the picker, now named "Claude Haiku 4.5", and a setting pinned to it keeps it; Opus 5.5 stays the Anthropic default. The same text counts as about 30% more tokens than on Haiku 4.5, so the token counts you see rise by about that much; at a tenth of Haiku 4.5's rates up to 100K prompt tokens, and half of them above, the bill still falls.
- A price that depends on the prompt. Unlike the other Claude models
Codeep offers, Haiku 5.5 is priced by prompt length: $0.10 in / $0.50 out per million tokens for a prompt of up to 100,000 tokens, and $0.50 / $2.50 for one over it — for every token of that request, with cache reads and writes counted in the prompt's length. /cost and /stats price each request on its own, so a long session of short prompts stays at the low rate. /cost says when a request was billed at the high rate, /stats lists both rates in its pricing table, and the cache savings use the rate each request paid. A cost that OpenRouter reports itself is used as it comes.
Fixed
- "Always Allow" is kept for the chat, and says what it covers. It was
forgotten when the run ended, so the next prompt asked about php again, and it covered the whole tool: answering it to one php artisan migrate let every other command through until the run ended. It now lasts until a new chat is started, another session is loaded or Codeep ends, and for a command it covers the program only: the button reads "Always Allow php (this session)", php runs with any arguments, and npm still asks. Any other tool is allowed by its name, as before. A Deny still lasts one run, so one misclick does not turn a tool off for the whole session; a write to a file that decides what runs later is still asked about every time; and a dry run keeps nothing it was told "always". Renaming a chat takes its answers along. In an editor over ACP the option reads the same and holds for the conversation: /session new, or loading another session, starts with nothing allowed, and a skill's shell line offers "Allow all of this skill's commands", which is what that answer does.
- **Far fewer runs end on "The model ended the run without writing a
summary".** Some models, GLM-5.3 at max effort on Z.AI among them, answer where they mean to call a tool with [tool call: read_file] — the line Codeep keeps for each of their earlier turns that only called tools, so there is nothing in it to run. In a test run on Z.AI it was about one reply in four that made no call, and asked for the call or a summary, the model sometimes wrote the line again until the two requests were spent. Codeep now asks again without those turns in what it sends (the tool results stay, joined into one message), so the model has nothing to copy, and goes on that way for the rest of the run; the OpenAI Responses API, which replays the messages as they were, is left as it is. The nudge no longer calls the reply a "placeholder", the second one is shorter and firmer, and the two are counted from the last tool call, not from the start of the run: three recovered replies spread over a long one used them up, and the third ended it. A model that still answers with neither a call nor a summary after two such requests in a row ends the run as before.
- Stopping
codeep acpstops the MCP servers it started. A SIGTERM,
SIGINT or SIGHUP ended it with every MCP server it had started still running: the handler meant to stop them never attached, because the library behind the settings file had already claimed those signals. It now stops them, then exits with 143, 130 or 129 as a shell would show. SIGHUP is new to it. The settings library still cleans up its temporary files first.
- A signal sent to the
codeepcommand reaches the program it started. The
command npm installs only starts the real process, so a kill, a service stopping it or an editor ending codeep acp with a signal ended the command and left the real process running; and if the real process was ended by a signal, the command exited 0. It now passes SIGINT, SIGTERM and SIGHUP on, waits for the real process to finish, and ends with its exit code or the signal that ended it. Where stdin is a terminal, Ctrl+C is not passed on again — the terminal already sent it to both — so a kill -INT aimed at the command alone is not passed on there; with stdin elsewhere (an editor, a service, > out.log) it is.
- **A provider named like an object member no longer finds a key that was never
stored.** Where the system keychain is not available, API keys are kept in a plain-text map in the config file, and a provider called constructor, toString or __proto__ was answered with a function or an object, handed back as its key. Only keys that were stored are found now, a key can be stored under any provider name, and storing one cannot change what every object inherits.