rune/oai-swift
rune/oai-swift
Watch
1
0
Fork
You've already forked oai-swift
 0
Code Issues Pull requests Releases 17 Activity

2.5.2 #12

Merged
rune merged 23 commits from 2.5.2 into main 2026-08-17 08:37:04 +02:00
Conversation 0 Commits 23 Files changed 39 +12683 -3834
rune commented 2026-08-17 08:36:56 +02:00
Owner
Copy link
No description provided.
rune added 23 commits 2026-08-17 08:36:56 +02:00 
Each round of a multi-tool-call chain used to append a new "Calling: X"
message, so a long tool chain stacked up a growing list of rows in the
transcript. Replaced with a transient status line under the thinking
indicator that updates in place each round; once the response
completes, the whole chain collapses into one expandable summary
message ("Used N tool calls") instead of N separate ones.
- ThinkingVerbs now picks from a hand-written verb list per active
  display language (en/nb/sv/da/de/fr) instead of always English,
  fixed to be nonisolated at the type level (this project defaults
  to MainActor isolation, which was breaking the static verb arrays).
- SyncStatusIndicator.tooltipText was typed String instead of
  LocalizedStringKey, silently bypassing localization for .help() —
  fixed.
- Translated ~150 strings (750 individual translations) into
  nb/sv/da/de/fr that had never been localized: the entire slash-
  command dropdown, sync status labels, reasoning-effort
  descriptions, model sort options, and settings rows added across
  recent features (Git Sync conflict recovery, crash recovery,
  notes, backup, external MCP servers). Most were invisible to
  xcodebuild -exportLocalizations because they're LocalizedStringKey-
  typed properties/helpers rather than literal Text() calls, per the
  same gap documented from the original Phase 7 audit.
- AI-translated per the README's existing disclosure; verified live
  in nb and fr builds, and diffed the catalog to confirm zero
  pre-existing translations were altered, only additions.
New CLIServerService listens on a Unix domain socket
(~/Library/Application Support/oAI/cli.sock) speaking a minimal
HTTP/1.1 subset, for one-shot non-streaming shell access to a single
fixed model — e.g. an `ai "prompt"` zsh function — without opening
the app window and without going through the tool-calling loop.

Configured in Settings > MCP > CLI Access (toggle, provider, model —
deliberately independent of the chat UI's active model). JSON
request/response envelope rather than raw text so new fields (model
override, streaming, tool support) can be added later without a
breaking wire-format change.

Verified live end-to-end against a real OpenRouter request, error
paths, and clean-shutdown socket cleanup. 10 new unit tests cover
the HTTP framing/parsing logic.
Six strings from the new CLI Access section (897bfdc) had never
been added to the localization catalog: the section title and
description, the enable toggle, and the shell-function instructions.
Verified live in a Norwegian build.
The description Text sits inside ModelInfoView's ScrollView, where
Text with a lineLimit doesn't reliably compute wrapping/truncation
(a documented SwiftUI/AppKit quirk). Without a fixedSize hint it
hard-clipped mid-word with no ellipsis; adding one fixed that but
silently broke the "More…" button below it in the same VStack.

Removed the whole toggle instead of chasing further edge cases — the
modal already scrolls, so a long description just means more
scrolling. Separately confirmed via OpenRouter's public API that at
least one model's description is truncated server-side with no fuller
version available through any endpoint, so this was never going to
fully solve "show the complete description" for every model regardless.
.onExitCommand alone wasn't enough: clicking into the multi-line
.textSelection(.enabled) description handed it real AppKit
first-responder status, and its own cancelOperation: handling for
Escape consumed the key event before it ever reached the modal's
exit-command handler.

Replaced text-selection on the description with an explicit Copy
button (same pattern as the chat message copy button in
MessageRow.swift) so it can no longer grab keyboard focus at all.
infoRow's single-line values keep .textSelection(.enabled) — only
the multi-line description reproduced the bug.
Each row in the Run History list now opens a Run Details sheet showing
the full, untruncated output/error (the old inline chevron-expand
capped output at 20 lines) plus duration, trigger, tokens, and cost.
Output/error use an explicit Copy button rather than
.textSelection(.enabled), to avoid the same Escape-beeps-instead-of-
dismissing bug just fixed in ModelInfoView.

While wiring this up, found the JarvisAgentRun model didn't match the
real oAI-Web API response shape: output decoded from a nonexistent
"output" key instead of "result" (always nil, hence "No output for
this run" even on successful runs with real content), finishedAt
decoded from "finished_at" instead of "ended_at" (Duration silently
never showed), and the status icon only recognized "completed"/
"failed" instead of the API's actual "success"/"error" values (plain
gray circle instead of a green checkmark). Fixed all three, verified
against a real API response, added 4 decoding tests.
detailRow's label and outputBlock's title were typed String instead
of LocalizedStringKey, so Text(label)/Text(title) inside them could
never localize regardless of catalog content — the exact anti-pattern
CLAUDE.md's i18n rules warn against for helper functions. Fixed the
types, then translated the 6 genuinely-new strings (Run Details, No
output for this run., Started, Duration, Trigger, Tokens) into
nb/sv/da/de/fr. A few others (Cost, Error, Output, Copied!) were
already covered from being reused elsewhere in the catalog.
ErrorDetail.code was typed String? but OpenRouter returns it as a JSON
number, so decoding the whole error body silently failed and every
non-200 response fell back to "Unknown error: HTTP <code>" instead of
showing OpenRouter's actual message (e.g. why image generation was
rejected). Dropped the unused code field, and made the streaming path
attempt to decode the error body too instead of skipping it outright.
sendAndClose is main-actor-isolated (implicit under -default-isolation
MainActor), but the Task in readRequest's Network.framework callback
runs nonisolated (DisableOutwardActorInference means it no longer
inherits the class's actor from lexical context). Calling sendAndClose
without hopping onto the actor is a hard error in Swift 6 mode.
Image-gen models are billed per-image/per-request, not at the plain
per-token prompt/completion rates ModelInfo.Pricing captures — so
token-based cost calculation was always near-zero for them regardless
of actual spend. Now requests OpenRouter's usage.include=true, decodes
the actual billed usage.cost, and prefers it over calculated cost via
a new resolveCost() helper (covers the dedicated Images API, regular
chat completions, streaming, and the tool-calling loop alike).
New large modal off the Stats sheet with a 6-way timeframe picker
(Today/7 Days/Week/Month/Year/Total), three tappable summary tiles,
and native Swift Charts (bar chart over time, pie chart by model).
Backed by date-range-filtered DatabaseService queries plus a new
daily-bucketed query — no schema changes needed since messages
already carry timestamp/tokens/cost/role.
Conversation text is still only persisted when explicitly saved (⌘S),
but tokens/cost/model/provider are now logged to a new usage_events
table for every completed AI response regardless — the Analytics
view now reads from this table instead of messages, so it reflects
real usage even for conversations that were never saved. Adds a
By Provider chart mode alongside Over Time/By Model.

Also fixes the Analytics entry point: ToolbarItem(placement: .navigation)
silently doesn't render in a plain .sheet-presented NavigationStack on
macOS. Moved the button inline next to the segmented picker, matching
the codebase's existing convention (e.g. the model-favorites star filter).
External MCP Servers previously only spoke stdio (spawn a local
command + args). Adds:
- env vars for stdio servers (merged into the subprocess environment,
  not embedded in the args string), with a masked key-value editor
- a native Streamable HTTP transport (URL + Bearer token + custom
  headers), so HTTP-based MCP servers like Obsidian's Local REST API
  plugin connect directly without needing npx/Node.js as a bridge

Introduces an MCPTransport abstraction (stdio/HTTP) so ExternalMCPClient
stays transport-agnostic — mirrors how Provider.swift already abstracts
AI backends in this codebase.

Also fixes a real crash found via live testing against Obsidian:
convertInputSchema force-unwrapped a tool parameter's `type`, which
isn't required by JSON Schema — Obsidian's plugin was the first real
server to send a parameter without one. Live-verified end to end
(vault search/read/write/edit) before this commit, per the project's
standing rule to hold external-service-dependent changes until they're
actually confirmed working, not just compiling and passing tests.
An unquoted prompt containing ?/*/[...] (e.g. `ai Who are you?`) was
getting glob-expanded by zsh before the ai() function ever ran,
aborting with "no matches found" and never reaching the socket server.
A plain function can't protect its own call site from this - filename
generation happens during command-line parsing, before the shell
dispatches to a function. Switched the recommended snippet to
`alias ai='noglob _ai_impl'`, which suppresses globbing for the whole
command line via noglob as a precommand modifier. Quoting remains the
fully robust habit for other shell metacharacters, but this covers the
specific mistake a user is most likely to make by accident.
The multi-line backslash-continued curl command was fragile to
copy/paste: a trailing space after \ or a dropped backslash (both
common when pasting a multi-line snippet out of a chat UI or browser)
silently breaks the continuation, so each following line gets parsed
as its own bogus command instead of a curl flag - exactly what Rune
hit ("command not found: -H", "-d", "no such file or directory:
http://localhost/") after pasting the previous version. A long single
line has no continuation character left to mangle.
Prompted by a real incident: Confab's process silently wedged, and
was indistinguishable from "just idle" after the fact — lsof showed
the CLI socket held open, every connection was refused, and nothing
had been logged for the rest of that session. No crash, no error.

- CLIServerService: NWListener's state transitions are normally
  near-instant, but had no bound on that assumption. Now times out
  after 8s if .ready is never reached, logging clearly and clearing
  the listener slot instead of silently pretending to work forever.
- oAIApp: a 5-minute heartbeat log line, otherwise meaningless on its
  own, turns "the log's gone quiet" from an ambiguous signal into a
  plain read of roughly when a future freeze started.

Neither fixes a known root cause (none was found - no crash report,
no error, just silence), so this is diagnostics and a narrow
failsafe, not a claim the underlying freeze is resolved.
Anthropic doesn't allow this kind of third-party OAuth usage, so it
was never wired into any UI - no button or menu item anywhere ever
triggered it. Confirmed unreachable (only construction site for
AnthropicProvider used the apiKey initializer) before deleting:

- AnthropicOAuthService.swift (314-line PKCE flow) removed entirely
- AnthropicProvider: dropped the AuthMode enum/oauth case, now just
  holds a plain apiKey string
- ProviderRegistry.hasValidAPIKey(.anthropic) no longer checks
  OAuth auth state
- README: removed the "...or use OAuth" mention
It was only shown when Search Provider was set to Google, but the
same key is also needed for Google embeddings in Semantic Search
(Advanced tab) — a completely separate feature from web search. A
user who only wanted Google for embeddings had no way to find where
to enter the key. Now always visible under General → Web Search, with
a caption clarifying both uses; Search Engine ID stays conditional
since it's web-search-specific. Also pointed the Semantic Search tab's
"no providers available" hint at the same location.
ExternalMCPManager.convertToolDefinition/convertInputSchema were made
nonisolated static func in an earlier session for direct unit testing,
but this project's -default-isolation=MainActor makes every type
implicitly main-actor-isolated unless marked nonisolated — so those
functions referencing Tool.Function.Parameters.Property's init and
ExternalMCPServer.slug (both plain data types with no explicit
isolation) triggered "main actor-isolated ... can not be referenced
from a nonisolated context" warnings, surfacing as Xcode's opaque
"exit code 0 but produced no further output" compile failure in a
Release build.

Marked Tool (and all its nested types) in AIProvider.swift, and
ExternalMCPServer/MCPTransportKind/MCPToolDefinition/MCPInputSchema/
MCPPropertySchema in ExternalMCPModels.swift, nonisolated at the type
level - they're plain DTOs for JSON request/response mapping with no
reason to be actor-isolated at all. Verified with a clean Release
build (matching how the warnings were originally surfaced).
rune merged commit 46ca15f352 into main 2026-08-17 08:37:04 +02:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Reference
rune/oai-swift!12
No description provided.