v0.0.12

What's Changed

• docs: Explain using dwim from Claude Code or Codex by @penberg in #143
• server: Run the model in a server that the sessions share by @penberg in #142

Full Changelog: v0.0.11...v0.0.12

v0.0.11

What's Changed

• cli: Show the model's memory apart from the context's by @penberg in #135
• tools: Replace a part of one line within it, and say when lines were copied in part by @penberg in #140
• docs: Point CONTRIBUTING.md at AGENTS.md, not CLAUDE.md by @penberg in #141

Full Changelog: v0.0.10...v0.0.11

v0.0.10

What's Changed

• Move compaction into harness/compaction.rs by @penberg in #130
• harness: Write the compaction note heading by heading by @penberg in #92
• cli: Sort the options under headings in --help and the man page by @penberg in #132
• Add a research-issue skill for writing up papers linked from issues by @penberg in #133
• tools: Keep glob within the working directory and cap its walk by @penberg in #119
• models: Remove the thinking from the context after each reply by @penberg in #95
• README: Say what dwim stands for under the title by @penberg in #137
• gpu: Allow more memory for context on a Mac by @penberg in #136
• tools: Run commands with no terminal, editor, pager or prompt by @penberg in #139
• cli: Drop "tok/s" from the status line before any tokens by @penberg in #138

Full Changelog: v0.0.9...v0.0.10

v0.0.9

What's Changed

• scripts: Score SWE-bench on arm64, and remove each image once scored by @penberg in #91
• cli: Add --bench, timing the model at depths of context by @penberg in #86
• cli: Add --temperature to sample at a temperature of one's own by @penberg in #75
• Run SWE-bench's agent in the instance's container, with a new --shell option by @penberg in #94
• scripts: Pull the next SWE-bench image and score the last patch while the agent works by @penberg in #96
• harness: Let the model think, and drop the plan and done prompts by @penberg in #100
• tools: Improve agent tools by @penberg in #101
• harness: Tell the model to fix by changing, testing, and working from the code by @penberg in #102
• harness: Keep more of AGENTS.md and of the compaction note by @penberg in #103
• gpu: Decode a token on the matrix accelerators past 2k of context by @penberg in #105
• tools: Read a file a page at a time, and tell the model to read more at once by @penberg in #106
• harness: Remind the model to make its change after many calls without one by @penberg in #107
• gpu: Fit the context on a Mac to the memory available by @penberg in #108
• harness: Limit a thought to 2,048 tokens by default by @penberg in #110
• cli: Choose the reasoning effort, medium by default by @penberg in #111
• cli: Show the status above the input box, the directory below it by @penberg in #112
• harness: Add auto mode and enable it for non-interactive sessions by default by @penberg in #113
• harness: Refuse commands that would discard the model's changes by @penberg in #114
• harness: Tell the model to check what it recalls, rather than think it over by @penberg in #115
• harness: Tell the model its commands run in the working directory by @penberg in #98
• cli: Hide what the model reads unless ctrl+o shows it by @penberg in #118
• harness: End a cut thought with a decision once the turn has called a tool by @penberg in #120
• Left-align the status line's elapsed seconds by @penberg in #121
• Random code cleanups by @penberg in #123
• Simplify code to follow Rust idioms by @penberg in #125
• Merge crates into one by @penberg in #126
• Keep the model in the Hugging Face cache by @penberg in #128
• swebench: Tell the model to fix the issue from the code here, not from upstream by @penberg in #129
• Replace Bonsai 2 27B with Qwen3.8-27B by @penberg in #124

Full Changelog: v0.0.8...v0.0.9

v0.0.8

Full Changelog: v0.0.7...v0.0.8

v0.0.6

What's Changed

• models: Count each sampled token in the status line by @penberg in #85
• CONTRIBUTING: Mark commits whose change dwim made with Generated-by by @penberg in #87
• gpu: Batched attention on the matrix accelerators by @penberg in #88
• tools: Refuse a long old_string, and read as many lines as asked for by @penberg in #89

Full Changelog: v0.0.5...v0.0.6

v0.0.5

What's Changed

• tools: End a pipeline with the status of its last command to fail by @penberg in #81
• harness: Give the plan room, and ask for the changes alone by @penberg in #83
• tools: Add an edit tool that replaces a string and shows the lines by @penberg in #82
• harness: Ask whether it is done when acting changes nothing for a while by @penberg in #84

Full Changelog: v0.0.4...v0.0.5

v0.0.4

What's Changed

• cli: Do not draw a bullet for a reply with no text yet by @penberg in #41
• cli: Default to a context of 16384 tokens by @penberg in #42
• gpu/vulkan: Validate the kernels with every check naga has by @penberg in #43
• gpu/vulkan: Run the kernels on a compute-only queue when there is one by @penberg in #44
• Context window improvements by @penberg in #40
• gpu/vulkan: Wait for the end of the pipe before a matrix product by @penberg in #47
• README: Bring your own sandbox by @penberg in #51
• gpu/vulkan: Score keys in rounds when the head is wider than a subgroup takes by @penberg in #53
• gpu: Store the key/value caches as 8-bit integers by @penberg in #52
• gpu/vulkan: Count a subgroup's lanes when scoring keys in attention by @penberg in #55
• models/bonsai: Refuse files with parts the model does not implement by @penberg in #60
• testing/difftest: Check the GPU against the CPU with the real model by @penberg in #61
• cli: Fit the default context to the memory the device has free by @penberg in #56
• harness: End a thought at a thinking budget by @penberg in #64
• gpu: Optimize Metal kernels for M5 by @penberg in #62
• harness: Keep a transcript of every session by @penberg in #57
• tools: Fix parsing tool calls with tags in values by @penberg in #67
• build: Add a Makefile with an install target by @penberg in #73
• models: Add min-p sampling, 0.05 by default by @penberg in #76
• tools: Outline a long file read without a start line by @penberg in #74
• docs: Minimize AGENTS.md by @penberg in #77
• build: Print one line per step, and the commands with V=1 by @penberg in #79
• harness: Investigate with thinking, then plan and act without by @penberg in #80

Full Changelog: v0.0.3...v0.0.4

v0.0.3

What's Changed

• Upload through two staging buffers by @penberg in #28
• Sum the tile kernel's products in packed half precision by @penberg in #29
• Multiply a few tokens through ternary weights at once by @penberg in #30
• Batch the bf16 matmul and split small matrices in the tile kernel by @penberg in #31
• Take a batch of up to 31 tokens eight at a time through the Vulkan ternary matmul by @penberg in #32
• Name the GPU kernels by what they do by @penberg in #33
• gpu/vulkan: Look a single token's bytes up in tables of activation sums by @penberg in #34
• Format with cargo fmt and check it and clippy in CI by @penberg in #35
• cli: Show only the directory in the footer by @penberg in #36
• cli: Hide the model's thinking in the shell unless asked by @penberg in #37
• harness: Send the environment with the first message, not the system prompt by @penberg in #38
• cli: Refuse to run on a Mac with less than 16 GiB of memory by @penberg in #39

Full Changelog: v0.0.2...v0.0.3

v0.0.2

What's Changed

• Preserve earlier thoughts in the conversation by @penberg in #13
• Do not make tool calls written inside the thought by @penberg in #14
• Round the Vulkan key/value cache to the nearest half by @penberg in #18
• gpu/vulkan: Store ternary weights as block-major by @penberg in #16
• Multiply batches of tokens through ternary weights as tiles by @penberg in #19
• Look a single token's trits up in a table by @penberg in #20
• Speed up the norm, Hadamard and attention kernels of a decode token by @penberg in #22
• README: Add an Installation section and show the shell by @penberg in #21
• Report where a conversation's time went by @penberg in #24
• Multiply a batch's activations as half floats in the tile kernel by @penberg in #25
• Time the kernels for a batch of 64 tokens too by @penberg in #26
• Take the system prompt up from a saved state by @penberg in #27

Full Changelog: v0.0.1...v0.0.2