Changelog
What shipped, release by release.
Generated from the repository history: every commit is a feature or a fix, and a
version bump closes a release. Last regenerated 2026-09-06.
-
Features
- keep non-badge rendering a config-only option
- simplify patch calls with flat replacement arguments
- document Homebrew release workflows
- update the Rust toolchain to 1.98.1
- add GPT-6 Astra to OpenAI catalogs
Fixes
- guard header border alignment during model changes and resizing
- align footer and activity text with the conversation
- keep thought toggle hints on the heading to save vertical space
- keep prompt rules consistent and chat content aligned
- preserve scrollback while answers stream after exploration
- prevent session data loss and inconsistent access checks
- generate a canonical Cask homepage
-
Features
- make thinking headings quieter
- align both themes around the cyan accent
- keep narrow markdown tables readable
- separate analysis from implementation requests
Fixes
- preserve exploration boundaries when sessions resume
-
Features
- prepare Apple Silicon releases for Homebrew
- keep implementation changes deliberately small
- keep the published model catalog current
- substantiate the token savings claim
- keep the core tools survey plan in the repository
Fixes
- keep editor borders stable across thinking levels
- keep the release version visible at startup
- align subagent rows with footer content
- keep partial tool output inside terminal width
- require explicit consent before delegating work
- remove the stray sample script
- document the current runtime behavior
-
Features
- ship built-in skills and let plan mode ask for one
- make hooks reliable across agent lifecycles
- keep background work visible after session resume
- make startup context visible at a glance
- let the Rust workspace stand on its own
Fixes
- keep the BG- badge on detached subagents
- show each running work item in exactly one place
- keep local development artifacts out of releases
- keep local state and interface presentation consistent
- keep badge tool headers visually consistent
- preserve chat history through compaction
- keep escape cancellation alive for the whole turn
- align badge tool output columns
-
Features
- add GLM-5.3-Flash on z.ai
- let /btw open its panel before there is a question
Fixes
- drop the render oracle that pinned a reference we left behind
- carry two deliberate changes into the tests that pinned them
- make the gate green again
-
Fixes
- wire the keys the panel promises, and split stopping from closing
-
Features
- add /btw, a question the conversation does not have to carry
- plan the side-question panel
- put required ahead of properties, and tier reasoning within a turn
- state the licence and the port's attribution
- add scope, report requirements and the generated-expectation rule
- add CONTRIBUTING.md
- add AGENTS.md
-
Fixes
- drop the run's clock when the label stops being about the run
-
Fixes
- stop everything that is running on one key, not the first branch
-
Fixes
- bound the wait for a cancelled tool, and answer the key at once
-
Features
- forbid emoji in the system prompt
- let a local runtime take an API key
-
Fixes
- offer local models only the effort levels their runtime accepts
-
Fixes
- read the listings LM Studio and Ollama actually serve
-
Features
- add Ollama, LM Studio and a custom OpenAI-compatible endpoint
-
Features
- add /init, which delegates writing AGENTS.md to a child
-
Features
- hide the footer workspace line behind a setting, off by default
Fixes
- stop reporting a non-JSON catalog answer on every picker open
-
Fixes
- lower the MTPLX fallback context window to 32k
-
Features
- bundle queued messages and let the first request answer the prompt alone
- rewrite the README around the port's own features
- start badge rows at the margin as a trial
- add a README to every crate
- add crate overview and permissions note to the README
- add a README and mark the MTPLX login entry as experimental
- close the plan's last two tasks
- surface the server's silent MTP fallback as a diagnostic
- cross the retention coupling the header tests bypassed
- record the two defects the plan's implementation uncovered
- cover the picker path from activation to a served model
- ask the MTPLX server which model it serves
- add the MTPLX provider factory and its session header
- plan the MTPLX provider
Fixes
- point the pi credit at the right repository
- move the pi credit from the intro to an acknowledgements footer
- one accent purple everywhere, a touch more lilac
- leave the todo icons uncoloured
- colour the subagent markers by outcome, not by shell
- the shown and decided context window follows the catalog
- decide compaction against the catalog's current context window
- a thought closes the running exploration, and the clock stops shimmering
- two edge cases the self-review of the branch found
- actually start the model picker's catalog refresh
- keep MTPLX out of the shared catalog overlay
- let models.json actually reach the MTPLX session header
- correct the plan's session-header claim after measuring it
-
Fixes
- park the cursor when no component owns it
- indent the todo rows and drop the colour from their shapes
-
Features
- leave the status line standing with what the work took
-
Fixes
- pale down the accent colour, not only the muted one
-
Features
- replace the status spinner with a travelling shimmer, and pale down the muted accent
-
Features
- keep the user's messages through a compaction and write a handoff note
- plan the compaction rework
-
Features
- add the undo tool, ported from notagent-main-rust
- record why the undo plan ships without a remove tool
- plan the undo tool ported from notagent-main-rust
-
Fixes
- take the glare off inline code in the dark theme
-
Features
- paste an image as a marker instead of a temp path
- replace delegation modes with two unified subagent types
- plan the two unified subagent types with star aliases
- rename the edit tool to patch and retune the footer
- add the plan_create tool so a plan session can record its plan
Fixes
- report what background work did, not only that it is running
- show a subagent once, in the panel that counts its tokens
-
Features
- retune the dark theme colours
- launch the prebuilt release binary from notagent-test.sh
- keep user-level state in ~/.notagent-v2
- add a badge block style for tool rendering
- share rendered TUI lines and settle the diff by pointer identity
- close the gaps against the MCP references
- connect to MCP servers and use their tools
- goal mode, driven by /goal
- the bash filter, switched with /bash-filter
- atomic file leases, switched with /leases
- add the ClinePass provider and GLM-5.3 on z.ai
- badge block style as the default, with a settings entry
Fixes
- settle the explore block like the reference on aborted turns
- colour the explore block from its call results, not at the close
- cache the Codex WebSocket connection, not just the continuation
- a borrowed tool meets the permission chain
- an aborted exploration block settles red
- the session knows which plans authenticate with an API key
- a subscription reports no price in the footer
- keep one explore block per run, as the reference does
- render diffs the way the reference does
- give badges their own fills, name the block explore, group every read-only tool
- group reads with the searches, restore the thinking toggle, correct the badge colours
- apply the badge block style to every chat block
- hide the completed todo list after five seconds
- justify the held editor borrow for the autocomplete pump
- install the rustls default provider and pump editor autocomplete