v0.2.0-beta.1

Changelog

Features

• b92ee60 feat(acp): serve Canopy to editors over the Agent Client Protocol
• b2ba9cf feat(agent): notice when an agent is going in circles
• 98f20ac feat(agents): agent definitions as files, dispatched by name
• dc19d92 feat(agents): dispatched agents sit under their orchestrator, with their branch
• 1c1c63b feat(agents): tool calls in panes carry their kind labels
• be65cf4 feat(anthropic): adaptive thinking, one cache breakpoint at the newest block, and paused turns
• 159178f feat(anthropic): hold many MCP tools back behind a tool search
• f411acd feat(attach): the interface opens on a conversation canopy serve runs
• fff5578 feat(bench): canopy bench measures pass rate and cost on real tasks
• 6ef07d5 feat(budget): spending caps hold inside a turn, and can be set
• a1cca37 feat(chat): !command runs in the project and its output goes with the next message
• db855d2 feat(chat): /grants lists what a conversation may do without asking, and takes it back
• 9e9ad88 feat(chat): @ file mentions, ctrl+x ctrl+e for your editor, # notes
• 5bc77b2 feat(chat): a command palette on ctrl+p
• 26528ea feat(chat): agents and conversations in the command palette
• 988a483 feat(chat): enter carries out a finished plan in build mode
• 5c609ed feat(chat): every conversation of this project is reachable from the palette
• 56596b9 feat(chat): find in the conversation on ctrl+f, copy the last reply on ctrl+y
• 306fed9 feat(chat): spending caps in the header, and a pause that says how to go on
• 2706170 feat(chat): up and down move through a draft, alt+k kills to the end, a long paste shows its size
• 83b48f5 feat(cli): canopy run, the full agent without the interface
• 7fa0793 feat(dispatch): agents share their orchestrator's cached prefix and report back
• cddad60 feat(dispatch): an agent's report carries the verifier's verdict on its work
• 1f932ae feat(engine): replay the conversation exactly and keep its prefix frozen
• 76c9b2c feat(exec): bound how many test runs execute at once
• 3688dc5 feat(hooks): hooks run in the sandbox
• a1d6e31 feat(hooks): pre-tool, post-tool and turn-end hooks, given JSON on stdin
• 475296c feat(instructions): send AGENTS.md, CLAUDE.md and project instructions
• 6f8732e feat(keys): a stored key is checked at once, and a price can be entered
• 839e9df feat(land): merge an agent's branch only if the merged result passes
• 52ed4e9 feat(lsp): a language server checks every file an agent edits
• 71a4c21 feat(lsp): find_definition and find_references from the language server
• 1fa9aa1 feat(mcp): remote servers over Streamable HTTP
• d4390a8 feat(openai): the Responses API for OpenAI's own endpoint, opt-in
• 1057ee7 feat(permission): outside content taints a conversation (D-57)
• 772aa4e feat(review): ask a model for an opinion of the ranked attempts
• b818d76 feat(run): verify with the project's tests and escalate effort on red
• f0f1cd7 feat(sandbox): Seatbelt and Landlock confinement policies
• 8ca2d9d feat(sandbox): a new worktree's setup and local MCP servers run confined
• 9af8241 feat(sandbox): confine agent shell commands with Seatbelt and Landlock
• 7a59b31 feat(sandbox): narrow a command's network to package registries (D-60)
• 8657ac8 feat(sandbox): the project's tests run in the sandbox too (D-61)
• c84ffdd feat(search): Anthropic's web search on Anthropic keys, and notices reach the screen
• 84253b9 feat(serve): keep agents working after their client leaves
• 6bc3c89 feat(session): compact when a conversation outgrows the window
• 34d3ba1 feat(session): measure the context by the last request and compact past a budget
• 8cf9a6f feat(session): what a compaction replaces is carried again
• 7678b3f feat(skills): load Agent Skills with their bodies on demand
• 7ab9da6 feat(theme): ship six palettes as data, load your own, gate raw colours
• be39824 feat(tools): keep long command output out of the conversation
• 952999b feat(tools): read less at the source
• 42b0f38 feat(tools): repo_map, an outline of the repository ranked by use
• 249bc41 feat(tui): /context shows where the next request's tokens go
• 28aa575 feat(tui): /mouse hands the mouse back to the terminal, and takes it again
• f0b1127 feat(tui): a footer under each finished turn with its tokens, cache share and cost
• 8cee462 feat(tui): demo images drawn from the golden screen fixtures
• aa04c95 feat(tui): desktop notifications, a live window title and tab progress
• c0c008b feat(tui): draw a finished turn in the order it happened
• 2faa0ed feat(tui): each agent pane carries its receipt on its border
• 86d7697 feat(tui): highlight code blocks with a real lexer
• bd291e9 feat(tui): move to Bubble Tea v2 and Lip Gloss v2 (V-01)
• 70f923b feat(tui): take steering back, stop and remove agents where they are seen
• 28e8566 feat(undo): show what an undo would change before doing it
• a019c57 feat(windows): it builds, and stopping a command ends everything it started
• 591e380 feat(worktrees): keep agent worktrees in Canopy's data directory and reclaim them
• 60dd6a7 feat: a failed turn is tried again with enter, and says what to fix when it cannot be
• da5f2c8 feat: canopy doctor, canopy init, and keys import for the first run
• 680ece2 feat: pictures in a message, scaled and sent to the model
• 6dc9a6f feat: what Canopy says while starting is shown inside the interface

Bug fixes

• d532c34 fix(acp): questions follow their conversation, and serve stops cleanly
• ab6f159 fix(acp): stop on a signal with stdin open, withdraw moved questions
• 4abd6d9 fix(agent): a changing workspace resets the no-progress count
• dc576c6 fix(agents): a definition's tools list caps its agents' trust
• af8a233 fix(agents): s stops an agent waiting on a question, and taken-back guidance keeps its lines
• 3479ffa fix(anthropic): never replay a thinking block without a signature
• 9b13aaf fix(attach): a conversation is sent from its first unfinished turn, trimmed
• 6386c88 fix(attach): a record that changed under the kept turns is sent whole
• 68a878d fix(bench): Python's bytecode is not a test the attempt added
• 457369c fix(bench): a pass needs the task's own tests seen passing
• b96e290 fix(bench): a test file the attempt added fails it too
• 5fec76e fix(bench): attempt in build mode and refuse a pass made by editing tests
• a21573b fix(budget): the overall cap counts turns still running
• 5a3ee82 fix(chat): a ! command is typed and confirmed, and its output cannot forge its frame
• 2790e1a fix(chat): a note is one typed line, confirmed, and never blesses another edit
• 5c6e1cd fix(chat): a question takes the keyboard back from the palette
• 3b9e643 fix(chat): a search answer is matched to its words, not only its generation
• 81b2e34 fix(chat): anything that happens between the two enters asks for the second again
• bb681fe fix(chat): find gives way to a question and matches without changing case
• d45086d fix(chat): only a plan with steps is offered, and carrying it out takes two enters
• 394e22c fix(chat): say when pictu...

v0.1.0-beta.2

Changelog

Features

• 01bc86a feat(acp): reach Claude by driving the Claude Code you already signed in to
• d99d0df feat(cli): sign a subscription in, sign it out, and say what a test can check
• b87753a feat(codex): reach ChatGPT by driving OpenAI's own Codex app server
• 64f04c6 feat(copilot): run a turn on the user's own GitHub Copilot subscription
• 7b959c2 feat(keys): a Claude credential that is a delegation rather than a secret
• dc9a266 feat(keys): a credential can be one somebody signed in to
• 26c2070 feat(keys): renew a sign-in before the request, not after the 401
• 9ce03b6 feat(trust): run nothing a repository configures until it is trusted
• 33ba0f0 feat(tui): a wizard branch that reaches a credential with nothing to paste

Bug fixes

• a00ab33 fix(acp): a delegated turn names no model, because it is not Canopy's to name
• 14ebe4f fix(codex): make the route id the route, and the account an identity
• b165186 fix(codex): say CODEX_HOME in the caveat, which is what the code reads
• f64311a fix(delegated): reap the vendor's process through the thing that guards its pid
• 7669535 fix(delegated): refuse untrusted vendor settings and stop orphaning vendor agents
• f9c99eb fix(exec): keep credentials out of every child process environment
• f8804fb fix(gitsafe): neutralise only the filters a repository defines
• f7658b2 fix(keys): SetModel wrote back a copy of the file from before the last write
• f237c51 fix(keys): keep the Claude route building on its own commit
• e0bbb78 fix(keys): let a sign-in route answer only about its own credentials
• 3d76bd5 fix(pricing): price Opus 5.5 and Fable 5.1, and stop billing OpenAI cache twice
• a099595 fix(release): let the current release be the current release
• 0a4528a fix(session): make the history database owner-only
• bc0fdc7 fix(subscription): make a vendor client something the cleanup can always reach
• 4c8869d fix(tools): keep file tools out of .git and stop git running repo config
• 07a082c fix(tools): refuse fetches that resolve to this machine or a private network
• 948feb1 fix(tui): escape model text and vendor sign-in links before they reach the terminal
• b8bff55 fix(tui): hold one event subscription instead of one per event
• a5264fe fix: close the bypasses found in review of the hotfixes
• c39a81b fix: harden delegated subscription routes
• a09fa73 fix: make credential renames recoverable
• 92d9ccd fix: report stale inline approval outcomes
• 2980074 fix: report stale inline approval outcomes

Other changes

• 61a29ad ci(release): warn rather than refuse when the Copilot app id is unset
• 96d88fa docs(install): stop calling signout the end of a grant it does not touch
• 70c623d docs(limitations): fix the claims that were true of one route and stated of three
• b74a174 docs(limitations): say that eviction seeds a conversation too, not only a restart
• f856f35 docs(readme): name the ground the ChatGPT route actually rests on
• 3dc7441 docs(security): describe the two signout messages the build actually prints
• c0ad49c docs(security): scope the expiry and the read-only sandbox to what has them
• ead6bf3 docs(tasks): S-01 to review, with the storage argument written down
• 722ac0e docs(tasks): S-02 to review, with the margin and the two failures argued
• 210a30d docs(tasks): S-03 back to review, with the resource story argued from the tests
• 3f6ccd9 docs(tasks): S-03 to review, with the session-versus-request argument written down
• feaaf17 docs(tasks): S-04 to review, and say plainly whose gate a delegated turn runs under
• 813b754 docs(tasks): S-05 to review, with the fallback argued rather than built
• 497d245 docs(tasks): S-06 and S-07 to review, with the boundary caveat written down
• 535cb46 docs(tasks): note where the second provider fork actually lives now
• 920e324 docs(tasks): record what the live run against a real Claude Code caught
• 3acc991 docs(tasks): record what verifying phase S found, and block the two it could not support
• 1ea28f0 docs(tasks): reflow the S-03 notes to the width the file uses
• d512327 docs(tasks): return S-08 to review with the second pass written down
• 634d5cd keys: a credential can be renamed, and everything on it follows
• fb1698d keys: keep the rename note inside sixty columns
• 59fb540 ledger: U-26 moved to its own branch
• c26f726 merge: integrate credential and picker UX
• ec1c65f merge: stack header layout after inline approvals
• 03bbe28 merge: stack model and credential UX after header layout
• 27d77bc refactor(codex): drop the login mode accessor too, which was the fourth
• 1b47a1f refactor(codex): drop the three exported methods nothing reaches
• ffaa610 style: check writes in the trust prompt
• 9b4307a test(codex): hold the handshake, the refusals and the machines with nothing on them
• 11fc874 test(copilot): hold the sign-in, the session and the two switches that reach them
• 8aec0e8 test(subscription): hold four guarantees that were only held by their comments
• 6f44b3b tui: a question is answered where it is seen
• 7086b2d tui: a question is answered where it is seen
• 18f6a90 tui: the model picker stands where the message box stands
• 63f7084 tui: the name lives in the header corner, not in the middle of a new chat

v0.1.0-beta.1

Changelog

Features

• 99c55ce feat(chat): let navigation read a permission prompt without answering it
• b19d696 feat(chat): make ctrl+r offer a compaction before it pays for one
• b1899ec feat(tui): carry who needs you into every header and stop locking screens
• 3d93cf8 feat: /model, for the conversation rather than the key
• 46fbb60 feat: a key is one credential and many models
• 87358d9 feat: a question asked on the side is worth keeping
• 80009db feat: a reply that reads like a document, not like its source
• 3cf2f2d feat: an agent stuck on a question finds you where you are
• 5a90d76 feat: an edit you can read, not a filename and a duration
• fa204df feat: the corner says who you are with, not what we are called
• e8444dd feat: the task list becomes a block you can read across the room
• 6b09d4a feat: two sonnet agents, with no key called sonnet

Bug fixes

• b0747ab fix(catalog): advertise only compatible OpenAI models
• 766b4fa fix(keys): a credential you just typed is the one that answers
• 1671279 fix(keys): open on a conversation when several credentials are stored
• 340adfb fix(keys): report credential selection truthfully
• eece6d3 fix(lint): drop the three helpers the inline rewrite orphaned
• 18eb361 fix(lint): highlight the fragment once rather than assigning it twice
• 58e7023 fix(tui): apply the mode the key stops on, not the ones it passes
• 7723752 fix(tui): keep markdown tables within terminal width
• 3d909ee fix(tui): key transcript cache by tool classification
• 0e44719 fix(tui): neutralize terminal controls in tool output
• 38ba5c5 fix(tui): preserve pending mode across every exit
• 03224fb fix(tui): scope approval input to its conversation
• 5d06887 fix(tui): scope approval input to its conversation
• 5f230b6 fix: a column is not a rune, and a frame is not a suggestion
• 87e9ded fix: a focus you cannot see cannot exist
• 9b0010e fix: a key can be asked for the model it already runs
• 3e7a588 fix: a refused model change moves nothing, and the picker takes a typed one
• f04a428 fix: a stale catalog says so, and two spellings are one model
• f260f24 fix: the renderer could stop returning, and did so on ordinary output

Performance

• 8a32c98 perf: render a finished turn once

Other changes

• f24aeb1 Merge remote-tracking branch 'origin/fix/cell-width-and-chrome' into repair/transcript-cache-classification
• e5e495c Merge remote-tracking branch 'origin/main' into feat/one-key-many-models
• 2161915 Merge remote-tracking branch 'origin/main' into release/beta-1
• 8e124df Merge remote-tracking branch 'origin/main' into repair/cell-width-integration
• 234775d Merge remote-tracking branch 'origin/main' into repair/tool-output-controls
• a86747e Merge remote-tracking branch 'origin/main' into tui/ambient-attention
• 96e2c62 Merge remote-tracking branch 'origin/perf/render-once-per-turn' into repair/credential-selection-truth
• 318f724 Merge remote-tracking branch 'origin/tui/what-the-agent-did' into repair/markdown-table-width
• bb99d94 merge: main's one exit path, with the picker on it
• 30c2ac9 merge: take the branch below, which resolved main the same way
• 4de4b9f merge: the same assertion, finished on both branches
• 734af64 test(tui): account for Update results

v0.1.0-alpha.3

Changelog

Features

• e98aae9 feat: a bigger fire, and the name unframed with its edges picked out
• df60c54 feat: a header bar that stays, with what you actually want at a glance
• 792b5fd feat: a jump-to-bottom marker sitting on top of the message box
• 632e7fc feat: a pane is the conversation screen in miniature
• 9d28d95 feat: actually connect the configured MCP servers
• 44fb0db feat: actually run the hooks a project configures
• fe573be feat: add project commands and cost outcomes
• f7bf0fc feat: add the canopy report command
• 308520a feat: btw gets a panel of its own
• f3f8210 feat: drag to select the conversation, and only the conversation
• ccaa5d9 feat: draw the name at the size it deserves, and keep it on screen
• 81273d5 feat: errors keep the provider's own words
• a81db87 feat: every agent on screen at once, each pane wearing its own fire
• 0422975 feat: label a tool call by what it is, and frame the question
• 4846596 feat: one drawing of the name at a time, and a fire that goes out
• 86e4b78 feat: queued steering shows itself
• df4c789 feat: the name casts a shadow, and the fire smokes
• 487b80b feat: the name edged in grey, a taller fire, and the rough edges of the shell filed off
• 54172f0 feat: the questions that create agents stop whispering

Bug fixes

• abb4cda fix(verification): harden freshness attribution
• 3625398 fix: agree on what makes two workspace directories the same
• 0b75d0c fix: align confined mode with its trust ceiling
• ba643c2 fix: bound the tool list out loud, and stop what a server started
• f606962 fix: build the report from one consistent view of the worktree
• 86f46b3 fix: close MCP hardening review gaps
• c6bedd3 fix: close the hook in-flight loop
• d582257 fix: confirm direct agent workspace
• 7f92c11 fix: dispatch runs the profile the user meant, and a spawned agent cannot spawn
• c87ef93 fix: enforce permission and workspace contracts
• 1e0aef9 fix: hold a conversation until its checks have finished
• 985120a fix: implement D-05, so a missing program is not a failing test
• 64e26b8 fix: keep a conversation's mode across a fork and a restart
• 8f3e5f9 fix: keep a shared workspace out of the review queue too
• 81c5330 fix: make the green gate actually wait for the tests
• a6ae8d7 fix: never signal a process group after its leader is reaped
• 7d2b273 fix: preserve the test command contract
• a43a841 fix: prove the review queue guard, and correct two ledger claims
• 6fe310f fix: quiet the linter on the selection tests
• cc7a0de fix: read the gate's mode through the resolver
• be6ad47 fix: refuse an ambiguous external call, and stop overclaiming two tasks
• 4b0f1e5 fix: refuse to pick a conversation up in another project
• f6cd808 fix: scope a remote approval to the exact call, and show it
• 108307f fix: scope an approval for a remote tool to its arguments
• caaf251 fix: stop a committing hook from firing on its own commit
• 16c0147 fix: stop a report's own values writing into the report
• 1468b7b fix: stop idle TUI animation churn
• f826d43 fix: sweep the engine for the six robustness properties
• b95a0e0 fix: tell a server's own requests apart from its replies
• 03633ca fix: the configured trust level is a ceiling, not a default

Other changes

• ba721b9 Merge remote-tracking branch 'origin/main' into codex-fix-fix-mode-trust-ceiling
• 357074e Merge remote-tracking branch 'origin/main' into codex-fix-hooks-loop-guard
• 0530e17 Merge remote-tracking branch 'origin/main' into codex-fix-release-0.1
• b36f3bb Merge remote-tracking branch 'origin/main' into feat/commands-and-cost
• 2e49510 Merge remote-tracking branch 'origin/main' into feat/interface-polish
• 595a769 Merge remote-tracking branch 'origin/main' into feat/modes-and-commands
• 2e6a252 a run report that cannot overstate what was checked
• 2ac7c25 a tent instead of a mushroom cloud, and the brand colours
• 2a6f3c6 claim A8-06 and A8-05
• c1e2f80 close the agents footer conditional the merge left open
• cdbca5c compose the screen a conversation opens on
• 8c76435 declare the config fields both branches need, and clear three stale notes
• d55f6f5 do not offer runway where no test is configured
• 7c76878 drop the em dash from the command list
• 3c415f3 drop the two bare colours nothing needs any more
• b02df0a fifteen commands, and let the mode change mid reply
• ca32b9c fire a code hook once per revision, not once per state
• df9eaa9 forbid constructing a colour outside the theme package
• 5e9981b four modes on the key, and the smoke in the accent grey
• 67c618a give the campfire the secondary brand colour
• 0b38aac give the mark a campsite, and drop the symmetry rule it broke
• 26782c2 keep A8-05 out of this branch's claim
• f60bda1 make the theme actually reach the screens that ignored it
• 4fc273d merge: bring main into the mosaic view
• 04c0340 merge: combine freshness and review follow-ups
• 67eb8d3 never build a revision key out of half a git status
• 8faa9b1 open on a conversation, not on a launch screen
• 0cae049 open the conversation you were given, not session-1
• 624e4e8 pick a theme from the terminal, redraw the welcome, and build the frames
• bc65c52 pin the mode prompt to the wire, and record what is left of M-09
• 7fa18c5 print a code on the way out, and pick the conversation up with it
• 16a181b record A8-05 in the ledger
• 8d7038b record A8-08 in the ledger
• 5cc2568 record M-08 in the ledger
• bb49264 record the gate and the aside in M-09
• a514d14 record the interface work, and the part of it that has to wait
• c2f2bdf record the mode ladder as M-09
• f8d5b98 record the mode ladder as M-09
• a4a012b rerun the three permission findings, and fix what the fixes left behind
• da89f9a run something when a state actually changes
• b36bdc7 say which mode and which model in the message box
• 75ff09a ship commands by default, and a list that drops out of the box
• 4146315 speak MCP, and refuse to let a server grade its own homework
• 5008b11 style: drop an em dash from the cost outcome header
• 4e945b3 the credential screen was ignoring the theme too
• 4e3ecc3 the green gate, and a question that changes nothing

v0.1.0-alpha.2

Changelog

Other changes

• 0d146c0 follow what an agent is doing, and stop showing worktrees that do not exist
• fb523fc give the wheel to the conversation and the arrow keys to the message box
• 34bd2ba input history, a new conversation key, and a logo worth keeping
• 5b619de mark prereleases as prereleases, and say how to install in the readme
• b099978 plan the phase that makes the pre-alpha usable
• edfc799 write down when the next release goes out

v0.1.0-alpha.1

Changelog

Features

• 8a12c02 feat(cmd): add canopy ask
• fd0d654 feat(cmd): add the headless harness
• 592d6e6 feat(core): add domain types and the closed state vocabulary
• 3ca218f feat(core): add provider, key and profile types
• 9ce309e feat(core): add the fake store with four scripted worktrees
• 702c893 feat(core): add the five engine interfaces
• a12c2e3 feat(core): add the provider contract
• 0284b6a feat(core): add the workspace roll-up
• 2bbe933 feat(core): derive visible test and service state from evidence
• ad5e555 feat(keys): store credentials in the OS keychain
• b154733 feat(provider): add the Anthropic client
• 618c86f feat(tui): add the dashboard
• 2155d28 feat(tui): full screen application shell
• 4131179 feat(tui): manage credentials from the interface

Bug fixes

• e56be96 fix(ci): pin golangci-lint and fix the findings it was hiding

Other changes

• 889e65a a second theme, a help screen, and a config for this project
• 78f0105 add a web fetch tool
• c31cb93 add openai compatible provider
• ee0c1ba add plan first mode
• eb08e1e add provider fallback chains
• fb7a1dc add session and turn types to core
• 6396ca2 add session forking
• 1240662 add structured git tools
• defbd68 add the agent registry
• c805abb add the agents view
• b1ed98c add the context meter and compaction
• 3c1e698 add the permission model and audit trail
• 41a57be add the session engine
• 573e894 add the shell tool and command runner
• 2a83944 add the tool contract and the file tools
• 1fc2132 add the tool use loop and give the app real tools
• 4da4243 add worktree discovery and lifecycle
• 9d14846 ask before running a command
• 01c9e92 bring a fresh worktree to a runnable state
• b1063cb cache the stable parts of a prompt
• ac08d1a checkpoint the worktree before every turn
• 3ed8fe8 commit an agent's work, and see where two of them collide
• d1ff486 finish the overnight write up
• f5dc3d7 fix a race between the stall guard test and its own callback
• c6112f3 fix two bugs that made the interface unusable
• aa47e48 give agents a task list, wire dispatch in, and record what got done
• 17290c6 give an agent a worktree of its own when it asks for one
• 6cc7767 give the live tests a realistic time budget
• cad743f give up on a provider that goes silent
• 6b9bc36 hash worktree content instead of guessing at it
• 987b648 let a rate be attached to a stored credential
• 4bd0af2 list the model flag in the keys usage text
• 1a0e103 make chat the home screen
• 7c347db note why the stall guard test went out red
• a53410b only count settled turns in the session cost
• ca4a595 open and create agents from the agents view
• 019f14a persist sessions to sqlite
• 621c537 price turns from a dated table
• d6ac0e2 put the verification engine on screen and give it something to run
• 41b91e3 rank agents on evidence, and refuse to when there is none
• 8a7b1dc record A5-04 and A5-11, and correct four stale statuses
• 1377d74 report a cancelled turn as stopped, not failed
• a84694b run the configured tests and measure what an agent changed
• cc8a18f say which model each credential talks to in the listing
• b1278ce set up the homebrew tap and write down how a release happens
• d58c7ad steer an agent without stopping it, dispatch from the chat, and cap the spend
• bec46ac use the stored model, and ask for a credential when none is chosen
• e83e332 write up the overnight work for review