Silent truncation of legal documents is unacceptable - Merlyn was reading
partial files without knowing it.
New behavior:
- Files under AGENT_MAX_FILE_TOKENS (default 500,000): return full content
- Files over limit: return explicit message with token count and options
- Paralegal controls any summarization decision
- Merlyn never silently truncates legal documents
Both static and instance fallbacks increased from 4096 to 2,000,000 tokens.
When model isn't in models.json cache (e.g. Grok), it was falling back to
4096 tokens causing severe truncation of file reads.
Reuses AGENT_CONTEXT_WINDOW_FALLBACK env var from Patch 2.
Default reduced from 0.25 to 0.05.
Reserving 25% of context for response was excessive, leaving too little
for file content in legal document analysis.
Configure via AGENT_CONTEXT_RESERVE_RATIO in .env to override.
Default fallback increased from 8,000 to 2,000,000 tokens.
When provider doesn't declare promptWindowLimit (e.g. Grok via OpenRouter),
the fallback was far too small causing severe truncation.
Configure via AGENT_CONTEXT_WINDOW_FALLBACK in .env to override.
* fix: white flash when switching between threads
* nest thread route under workspace
* fix: white flash when switching between workspaces
* simplify Link usage in workspace/thread sidebar items
* smooth workspace and thread switching
* fix race condition on send during thread/workspace switch
---------
Co-authored-by: Timothy Carambat <rambat1010@gmail.com>
* fix(tts): strip Markdown syntax before sending text to TTS engines
Chat responses are rendered as Markdown but the TTS components piped the
raw response into Piper / the browser's `SpeechSynthesis` API. The
synthesizer reads every special character literally — `**bold**` becomes
"asterisk asterisk bold asterisk asterisk", `# Heading` becomes "pound
heading", code fences are read backtick-by-backtick, and bullet lists
become "hyphen item". The result is unintelligible whenever the assistant
includes any formatting, which is most of the time.
This commit adds a small `messageToSpeech` helper that converts a
Markdown chat message into plain text suitable for TTS:
- fenced code blocks and images are dropped (nothing useful to read)
- inline code and link labels keep their text content
- emphasis markers, headings, blockquote markers, list markers, and
horizontal rules are stripped while preserving the underlying words
- HTML tags are removed but their text content kept
- table pipes become commas so rows read naturally
The helper is regex-based — no new dependency — and is wired into both
the native (`SpeechSynthesis`) and Piper TTS components in
`WorkspaceChat/ChatContainer/ChatHistory/HistoricalMessage/Actions/TTSButton`.
Closes#5557.
---
Co-authored-by: Timothy Carambat <rambat1010@gmail.com>
* initialize
* expand tool result text limit | add syntax highlighting and json formatting to tool result rendering
* fix onError jsdoc
* lint
* fix unread icon
* route protection
* improve form handling for NewJobModal
* safeJsonParse
* remove unneeded comments
* remove trycatch
* add truncateText helper
* add explicit fallback value tos safeJsonParse
* add shared cron constant and helpers
* reduce frontend indirection
* use isLight to compute syntax highlighting theme
* remove dead code
* remove forJob and make job limit to 50
* create recomputeNextRunAt helper method
* add comment about nextRunAt recomputation
* add job queue and concurrency control to scheduled jobs
* use p-queue
* change default max concurrent value to 1
* add comment explaining internal scheduling system
* add recomputeNextRunAt on boot
* add generated documents to run details
* Modify toolsOverride functionality where no tools selected means no tools are given to the agent
add a select all/deselect all toggle button for easily selecting all
tools in the cerate job form
* create usePolling hook
* add polling to scheduled jobs and scheduled job runs pages
* add cron generation feature in job form
* remove cron generation feature | add cron builder feature | add max active scheduled jobs limit
* set MAX_ACTIVE to null
* replace hour and minute input fields with input with type time
* simplify
* organize components
* move components to bottom of page component
* change Generated Documents to Generated Files
* add i18n to cronstrue
* add i18n
* add type="button" to button elements
* refactor fileSource retrieval logic
* one scheduled job run can have status "running"
* add protection of file retrieveal from scheduled job in multiuser mode
* fix comments
* make job status default to queued
* add queued status
* fix bug with result trace rendering
* store timeout ref and clearTimeout once race settles
* remove unneeded handlerPromise tracking
* move imports to top level
* refactor hardcoded paths to path resolve functions
* implement new job form design
* simplify
* fix button styles
* fix runJob bug
* implement styles for scheduled jobs page
* apply dark mode figma styles
* delete unused translation key
* implement light mode for new new job modal, run history, and run details
* lint
* fix light mode scroll bar in tool call card
* adjust table header contrast
* fix type in subtitle
* kill workers when job is in-flight before deleting job
* add border-none to buttons
* change locale time to iso string
* import BackgroundService module level | instatiate backgroundService singltone once and reuse across handlers
* add p-queue, @breejs/later and cron-validate as core deps
* parse cron expression to a builder state once
* add theme to day buttons in cron builder
* fix stale tools selection caption
* flip popover when popover clips screen height
* make ScheduleJob.trigger() await the run insertion | disable run now button if job is in flight
* regen table
* refactor generated file card
* refactor frontend
* remove logs
* major refactor for tool picking, fix bree/later bug
* combine action endpoints, move contine to method
* fix unoptimized query with include + take + order
* fix dangerous use, refactor job to utils
* add copy content to text response
* improve notification system subscription for browser
* remove unused translations
* prevent gen-file cleanup job from deleting active job file generated references
* rich text copy
* Scheduled Jobs: Translations (#5482)
* add locales for scheduled jobs
* i18n
---------
Co-authored-by: Timothy Carambat <rambat1010@gmail.com>
* add config flag with UI notice
* update README
* telemetry datapoints
* Always use UTC on backend, convert to local in frontend
* fix tz render
* Add job killing
* cleanup thinking text in job notifications and break out reasoning in response text.
Also hide zero metrics since that is useless
* Port generatedFile schema to the normalized workspace chat `outputs` file format so porting to thread is simple and implem between chats <> jobs is 1:1
* what the fuck
* compiled bug
* fixed thinking oddity in complied frontend
* supress multi-toast
* fix duration call
* Revert "fix duration call"
This reverts commit 0491bc71f4223e65ea4046561b15b268fefb8da2.
* revert and reapply fix
---------
Co-authored-by: Timothy Carambat <rambat1010@gmail.com>
* add image lightbox for chat attachments
* wrap lightbox image triggers in button elements
* add images to dependency array
* add jsdoc to ChatAttachments and remove filter
