agentty: oudere versies downloaden

Added
- Credit the model in commits (#59). Settings › General, off by
default. On, commits the agent makes end with
Co-authored-by: agentty (<model>) <noreply@agentty.org> (aider's
shape), naming the model that wrote that turn. Works in agentty run
and subagents too.
- The ^U reader shows what's known about the thinking on its header
row: model, effort, full text or summary, time and tokens (the billed
count when the provider reports one). Turns now record their effort.
Fixed
- The ^U reader follows streaming thinking while you're scrolled to
the bottom.
Full history: CHANGELOG.md

Added
- ^U reads thinking in full (#123). On an empty composer it opens the
newest thinking block over the screen; ←/→ go to older ones, Esc lists
them all, ^U closes. Collapsed blocks say ⋯ N lines above · ^U to read,
and the palette has "Read thinking". A live retrieved-context card still
takes ^U first.
Changed
- The ^O and ^U viewers don't move under you. On the newest entry while
it runs, scrolled to the bottom, you follow new tools or thinking blocks.
Anywhere else you stay put and new ones show as ↑ N new; t goes to
the newest.
- Text arriving in those viewers types in with the transcript's
typewriter and pacing. The Motion setting turns it off the same way.
Fixed
- Collapsed thinking counts its hidden lines right. Long blocks said
"24 lines above" whatever their length.
- -m picks the model the TUI opens on. It was saved but the
provider's saved model won at startup.
- Reasoning settings take effect on the turn being written, the
"Reasoned · Ns" timer starts when thinking starts and stops on errors,
Hidden no longer fetches a summary it throws away, and Haiku/Sonnet 4.5
can think again.
- Alt+←/→ in threads includes threads made this session.
Full history: CHANGELOG.md

Fixed
- Reasoning off really is off on Claude 5 models. Sonnet 5 kept
thinking with effort off because the request left thinking out, and
Claude 5 models think by default. Off now follows what Anthropic's
/v1/models says each model accepts (disabled, between_tools on
Sonnet 5.5), with rules matching Anthropic's per-model table as the
fallback. A rejected thinking mode is retried and remembered. Models that
always think (Opus 5.5, Fable, Mythos) no longer offer off at all. Smart
Mode's routed off is sent explicitly too.
- Switching providers keeps the right model. The saved model per
provider was looked up under the name you typed (codex) but saved under
the canonical one (openai), so switching back fell to the default. A
local server (Ollama, LM Studio) that doesn't have the saved model now
picks one it has and says so.
- Effort survives switching models. It used to be rewritten on every
switch, so off became low on a model that always thinks and stayed low
after switching back.
- agentty run -m no longer changes the TUI's saved model, and neither
does a headless --provider.
Changed
- The tool output pane (^O) is readable and searchable. It uses the
terminal's height and never runs past it, wraps long lines and the full
command instead of cutting them off, and shows the output as you'd want
to read it: no description echo, ``` fences or notes meant for the model,
in normal text with tokens coloured. A cut is shown as a ⋯ line. `/`
finds as you type, `n`/`N` step through matches, PgUp/PgDn page, and `y`
copies the clean text. The list never cuts off a row's status.
- The diff review panel was redone. It says up front that the edits
are saved and you keep or undo each one, shows files as tabs, lists each
change with real old and new line numbers, and counts kept and undone.
- Tool output reads in the normal text colour with paths, numbers, strings
and flags coloured, instead of a dull grey. Cards drop notes the tool
wrote for the model, read opens at the top of the file, and git status shows just the changed files.
- The status bar shows the whole thread title when there's room.
Full history: CHANGELOG.md

Changed
- Tool cards got a full visual pass. The header is one two-tone pill:
the tool name on a solid chip, the rest on a soft band of the same
colour, and the output below sits on a faint tint of it. Write, edit and
diff output is syntax coloured on the red/green bands, and paired lines
light up the words that changed. Grep rows are syntax coloured too, and
every match sits on an amber band. File paths get a language dot and a
dim directory with a bold file name, and edit headers end with +3 −5.
- Grep highlights whatever the pattern matched, regex included. It used
to give up on any pattern with an escape like resolve\(.
- Running tools keep their spinner whenever the whole card fits on
screen, instead of a still dot, and their header has a soft shimmer.
- Permission prompts show what bash will run. bash was missing from the
shell case, so approving it showed the whole args object as json.
- Every theme reads. A contrast floor lifts colours that were too dim on
their own background (links, headings, hints), keeping the hue. Across
the 615 bundled themes that took 870 dim colours to none.
- The native theme gets the same look on terminals that report their
colours: maya asks once at startup, in the round-trip it already makes.
- Reasoning uses a quiet neutral ink instead of the theme's secondary hue,
says how many lines are hidden when collapsed, and its rail is coloured
again (a renderer bug left left-only borders uncoloured).
- Smaller: failed cards drop repeated [kind] tags, the composer colours
@file and /command, nested markdown lists indent, picker rows and the
diff review panel use the same tinted bands, MCP tools read as
Search Issues · github, and the ── branch before each tool is gone.
Added
- tool_gallery renders every surface from real recorded tool calls
(--running, --ask, --chrome, --reasoning, --prose, --welcome,
--panels), and a ui_snapshots test diffs its output so a change that
alters how something looks shows up in review.
Full history: CHANGELOG.md

Fixed
- Settings stopped reverting when you run several agentty sessions at
once. Each session reads settings.json at startup and keeps it in
memory, and the save-time merge treated the whole appearance block as one
value. So one session setting scrollback, and another (still holding the
startup copy) changing its theme, wrote the stale scrollback back over
it. The setting did persist — until any other session saved anything.
Appearance settings now merge field by field: a field a session did not
change keeps whatever is on disk.
- Setting up a provider no longer pulls you off the one you are using
in two more cases. Working on Ollama with no API keys counted as "nothing
configured", so adding any provider adopted it; the active local backend
now counts. And an empty Anthropic key read as signed in — every other
provider already required a non-empty key.
Changed
- Appearance settings are declared once. The settings writer, its
loader and the save-time merge each used to spell out every field by
hand, and nothing made those lists agree — a field missing from one was
silently never saved, never loaded, or clobbered by another session. They
now walk a single table, defaults included, so none of them names a
field. Two build-time checks were added alongside it: one that every
appearance field is in that table, and one that keeps UI drawing out of
agentty (maya owns every pixel decision).
- The release's channel audit reads AUR's version from the RPC as well as
the cgit mirror. v0.9.48 went red on a package that had published
correctly 90 seconds earlier, because the mirror lags.
Full history: CHANGELOG.md

Changed
- Setting up a provider no longer switches to it. Adding a custom host
or pasting an API key used to make that provider active — and because a
switch has to pick a model for the new backend, it also changed the model
you were working with. Registering a spare key while mid-task ended your
session. The credential is saved and the toast offers Ctrl+P to switch
when you actually want to. First run still adopts, since there is no
session to protect and leaving you unconfigured would be worse.
- The provider picker reads as columns. Descriptions start at a fixed
column so the provider names scan vertically, and the marker column now
carries meaning: ● for the active provider, › for "show more", +
for the row that adds a host.
Fixed
- Picking a model updates the badge immediately, instead of waiting for
the next message. The badge shows the model that actually served the live
turn (they differ under Smart Mode), and that marker was not cleared when
you chose a different model — so it kept naming the previous one until a
new turn overwrote it. The same staleness made a provider switch show an
impossible pair, like Ollama beside Opus 4.5.
- A backend still loading its catalogue says so. Adding a local host
left the badge showing a provider with a blank model, which reads as a
half-finished switch; it now says loading models… while the fetch is in
flight, or no model · ^/ to pick if it finished without one. The model
fills itself in when the catalogue lands — dismissing the picker is safe.
- Sign-in adopts only when nothing usable is configured. A keyless local
host is stored with an empty key, so the old check counted it as a working
setup; someone whose only entry was such a host got no adopt and a dead
session.
Full history: CHANGELOG.md

Changed
- The provider picker (Ctrl+P) shows what you have set up instead of
the whole catalogue. The rest collapse behind one "Show N more
providers" line; press Enter on it or Ctrl+E to reveal them. Typing
still searches every provider.
- A provider with more than one account says which one is serving, on
the model picker's section header and the status badge.
- RECENT in the model picker is ordered by recency. It used to be
alphabetical, so the model you picked moments ago could sit below one
you last used a week back.
Fixed
- Changing model no longer renames every turn already in the thread
(#120). A turn's header named whichever model was selected rather
than the one that produced it. Turns recorded before this release
still show the selection.
- Switching account on the active provider reloads its model list,
which matters when two accounts have different entitlements.
- Keys in the provider picker no longer appear to need two presses.
- The @ and # pickers show their keys; Enter, Esc and the arrows
always worked, nothing said so.
- Settings > Plugins shows edit and remove on an unapproved server.
- Footer hints describe the highlighted row: Ctrl+F reads unfavorite on
a model you have starred, and Ctrl+/ reads "all providers" when the
list is scoped to one.
Full history: CHANGELOG.md

Added
- Allowed hosts. Settings → Sandbox → Network → Allowed hosts limits
sandboxed commands to the sites you list: github.com exactly, or
*.github.com for its subdomains. Leave it empty for any public site.
Pasted URLs are trimmed to the host, typos and IP addresses are refused
with the reason, and the agent is told which hosts it may use.
Changed
- On macOS, sandboxed commands reach the network only through agentty's
filtering proxy, as on Linux. Your local network, services on
localhost and cloud metadata addresses are refused; the internet
still works for git, curl, npm, pip, Python and Node.
- The macOS sandbox no longer lets a command read your clipboard, list
your keychains, or check whether files exist outside what it was
granted.
Fixed
- Picking hardened or airgapped and restarting no longer shows
custom. The memory cap was lost on every reload, so those presets
came back without it. A memory cap or /tmp size you set yourself was
dropped the same way. If an older build saved one of these presets for
you, pick it once more.
- Picking permissive no longer shows custom straight away.
- Under the macOS sandbox, /usr/bin/git, cc and python3 run again
instead of failing at startup.
- A command refused by the sandbox proxy now gets the [sandbox] note,
so the agent stops retrying it.
Full history: CHANGELOG.md

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

Fixed
- Headless runs now resolve the real context window, including on custom hosts. agentty run, agentty acp and agentty mcp-serve never build a Model, so they installed the subagent seam with an empty model catalog — leaving context_window at 0 on every turn (the Ollama transport reads 0 as "use my tiny default" and truncates exactly the long runs headless exists for), cheapest_capable_model() with nothing to choose from, and Smart Mode's role resolver falling back to the parent model for every pinned slot. They now fetch the live catalog through the same seam the stream path dispatches on, layered over the bundled floor so an id that outlived its catalog row (a -m pinned in settings) still resolves. Verified: a custom OpenAI-compatible host goes from 0 to its advertised 131072.

All about reaching any OpenAI-compatible endpoint, and telling the truth about what happened when you do.
Added
- Type an endpoint into ^P and it offers to connect. agentty has always spoken to any OpenAI-compatible server (--provider host:port, or the "Custom host…" row), but the picker never said so: typing a hostname made every familiar provider disappear and left one muted grey row behind, which reads as "no". Now a query that looks like an endpoint — a dot, a host:port, an https:// — promotes a concrete "Use <host> as a custom OpenAI-compatible host" row to the TOP, and Enter carries the typed text straight into the connect probe instead of reopening an empty field. Saved custom hosts are searchable too (they used to vanish the moment you typed, so the only way back to one was scrolling past every built-in), the escape hatch is no longer painted in the muted style this TUI reserves for unavailable, and the footer says what Enter will actually do.
- The connect probe says what went wrong and what to do about it. Three very different mistakes used to arrive as the same unhelpful string. A 200 with HTML — what you get from pasting a dashboard URL instead of the API base — now reads "that's a web page, not an API — use the API base URL (usually ends in /v1)" instead of "HTTP 200, no model list at any known path". A 401/403 means the address was right, so it routes you straight to the key prompt instead of dumping you back into a URL field you typed correctly. And the probe now ranks its attempts, so a 401 on the configured path is no longer buried under a 404 from the /v1 fallback.
Fixed
- A Retry-After measured in hours is a quota, not a burst limit. A 429 carrying retry-after: 29410 (8 hours, from a provider's daily free-tier cap) was clamped to 600s in the reducer and 30s in the agent loop, then retried three times — 95 seconds of waiting to print the exact message the first response already carried. agentty now honours a server hint up to 15 minutes and treats anything beyond it as terminal: 95s → 1.5s, with the provider's own actionable text. Burst limits are unaffected, and the error is still classified RateLimit so the loop-backoff schedule is unchanged.
- A context window a gateway made up no longer breaks compaction. The window drives the gauge and auto-compaction, and the extraction narrowed straight to int — so 18446744073709551615 became -1, 4000000000 became -294967296, and the string "1e6" became 1. A negative window disables compaction entirely (every used < window test is false) and a window of 1 compacts every turn; both read as "agentty is broken" rather than "this gateway reported nonsense". Values are now validated before narrowing and rejected outside a plausible 1024–100M band, and a partial string parse is treated as a misread rather than a lenient read.
- Headless runs had an empty model catalog. agentty run and agentty acp never build a Model, so the subagent config was installed without candidates — leaving context_window at 0 on every headless turn (the Ollama transport then fell back to the daemon's ~2k/4k default, truncating exactly the long runs headless is for), cheapest_capable_model() with nothing to choose from (read-only subagent roles never routed down), and Smart Mode's role resolver with no candidates (pinned slots silently fell back to the parent model). Seeded from the same bundled catalog the TUI uses.
- Reasoning text no longer disappears on gateways that send both spellings. Some OpenRouter passthroughs emit an empty reasoning_content alongside a populated reasoning; taking the first key that merely exists dropped every reasoning token, silently.
Changed
- The OpenAI-compatible dialect is now observed through one table, not scattered ifs. Every way this spec-less wire spells a field — the two reasoning spellings, three error envelopes, Mistral's structured content parts, optional tool-call indices — lives in one place with the endpoint each was observed on. Two sites were separately 

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See [CHANGELOG.md](https://github.com/1ay1/agentty/blob/master/CHANGELOG.md).

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See [CHANGELOG.md](https://github.com/1ay1/agentty/blob/master/CHANGELOG.md).

See [CHANGELOG.md](https://github.com/1ay1/agentty/blob/master/CHANGELOG.md).

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.

See CHANGELOG.md.