Changelog

0.1.9 — June 14, 2026

Scheduled background research

• watch can run in the background — scira watch "<goal>" -b saves a topic and re-runs it on a schedule, even after you close the terminal. Until now watch only looped in the foreground for as long as the command stayed open. Set the cadence with --every day|weekday|week and --at HH:MM (your local time), and label it with --name.
• A scheduler runs the saved watches — scira watch start spawns a background process that wakes once a minute and runs whatever is due; scira watch stop and scira watch status control and inspect it (status shows uptime, the recent log, and each watch's next run time). Saving a background watch doesn't start the scheduler for you — Scira tells you whether it's running and how to start it.
• You hear about changes — every background run diffs the new report against the previous one, sends a desktop notification with the added/removed section counts, and the next scira command prints a short banner of results you haven't seen yet. Results are marked read only once that banner has actually shown, so a crash mid-command won't make you miss one.
• Managing watches — scira watch list shows each saved watch and when it next runs; scira watch remove <id> deletes one. Saving the same goal, schedule, and project twice reuses the existing watch instead of stacking up duplicates.
• Foreground watch is unchanged — scira watch "<goal>" --daily|--hourly|--weekly|--interval <ms> [--runs N] still works exactly as before for monitoring you want to sit and watch live.

0.1.8 — June 13, 2026

TUI fixes

• Loading phrase now rotates — the animated loading line under a running turn was permanently stuck on "Thinking it through…": the frame counter wrapped modulo the spinner length, so it never grew enough to advance the phrase. The counter now increments monotonically (the spinner still takes it modulo its frame count), so the phrase cycles roughly every two seconds as intended.

0.1.7 — June 13, 2026

Updates & harness polish

• Update notifications — Scira now checks npm for a newer release (at most once a day, cached). The TUI shows it as an inline notice on open; CLI commands print a one-line reminder when they finish (and never both at once). The check is skipped for --version/--help so those stay instant, and a transient network failure no longer discards an update it had already found.
• Friendlier missing-file handling — scira export <run> --format md and the shell /plan command now degrade gracefully when a run hasn't produced report.md / plan.md yet (common on fresh local-harness runs), instead of erroring out.
• Codex tool rendering — Codex's bash output (returned as {exitCode, output}) and its fileChange events now render cleanly in the timeline instead of as raw JSON, matching the Claude Code treatment.
• Harness writes land in the run directory — research runs on the local Claude Code / Codex providers now write report.md, sources.jsonl, etc. into the correct run folder (the run's artifacts are no longer pre-seeded, which had blocked the CLI's "read before write" rule).
• Models get the run path — agent instructions now tell the model the run's absolute directory, and the file tools accept that path (or any path inside the run) in addition to bare filenames. Models that aren't happy with bare names — e.g. gateway-routed ones — no longer fail by guessing a .scira/runs/latest/… path.
• Skills in the harness — Scira's built-in research skills are now passed to the Claude Code / Codex runtimes natively.
• /help redesigned — leads with keyboard shortcuts (scroll, quit, autocomplete, group toggles) and a grouped command map, sized to fit; it now also opens from the home screen.
• xAI model list — the /model picker for xAI now lists only text models (via /v1/language-models), dropping image/video/embedding models.
• Faster quit — Ctrl+C-twice / Ctrl+D no longer leave a timer holding the process open; the TUI exits promptly.

0.1.6 — June 13, 2026

Local agent runtimes — Claude Code & Codex

• Run Claude Code or Codex as a provider — pick claude-code or codex from /llm (or the provider picker) and Scira drives the AI SDK harness instead of a plain language model. The CLI's own agent loop and tools run inside the TUI, streamed into the timeline like any other turn.
• Runs on your machine, on your real files — a local-filesystem sandbox runs the harness bridge as you, in your project directory, with full local filesystem and process access. No cloud sandbox.
• Uses your existing CLI login — no API key — authentication comes from your claude login / codex login session (~/.claude / ~/.codex). Scira never reads, passes, or inherits an API key for these providers — credential env vars are stripped from the bridge so your Pro/Max/ChatGPT subscription session is used as-is.
• Model & reasoning controls — /model lists Claude (Sonnet 4.6 default) and Codex (GPT-5.5 default) models; /thinking sets Claude Code extended thinking (off/on/adaptive) and /reasoning sets Codex effort (low/medium/high).
• Grounded web search — the harness's built-in web search is disabled; instead the CLI is given Scira's own multi-query webSearch and readUrl tools, so web answers run through your configured Exa/Firecrawl/Parallel pipeline and cite real sources.
• Same instructions as Scira — the harness receives Scira's agent instructions, so it follows the same workflow and persona as the built-in agent.
• Session continuity & resume — each run keeps one live harness session across turns, and the resume state is persisted per run so reopening a run (even after restarting the TUI) continues the same CLI conversation. Sessions are cleanly torn down on quit, and any resume failure falls back to a fresh session.
• scira doctor — now reports each local runtime: whether the claude / codex binary is on PATH and whether you're logged in (reading the account from ~/.claude.json / ~/.codex/auth.json).

TUI fixes

• Dedicated rendering for built-in tools — Claude Code / Codex tools now render with purpose-built timeline UI instead of raw JSON: Edit as a colored diff, TodoWrite as a checklist, Write as a file preview, Read/Bash/Grep via the file/shell/grep views, plus clean summaries for Task, ToolSearch, and WebFetch.
• Mode indicator — a PLAN MODE / FULL RESEARCH chip now sits inline with the shortcut hints, and the input box tints when plan mode is active, so the current mode is always visible. /plan on the home screen arms plan mode for the next session you start.
• Solid overlays — the /model, /llm, and /provider selector now paints a solid panel instead of letting the conversation bleed through it, and clips long rows so its border stays intact on narrow terminals.
• Quit takes intent — pressing q no longer quits. Use Ctrl+D, or Ctrl+C twice (the first press shows a confirmation); Ctrl+C while a turn is running stops the turn instead.
• Short conversations anchor to the top — opening an older, short chat now shows messages from the top instead of jammed against the input box; the feed only bottom-anchors once it overflows the viewport.
• Faster timeline rendering — tool-result formatting is now cached per item, so the feed no longer re-parses every result on each animation frame (~20fps while busy). Large outputs (e.g. a megabyte page fetch) no longer stall the UI; the full result is still saved, only what's laid out on screen is bounded.
• Wording — "New research" → "New session" and related hints, reflecting that Scira now runs coding and local-agent sessions too.

Internal

• Bun-native runtime — the test runner moved from vitest to bun test, and core utilities now use Bun APIs instead of Node shims: Bun.spawn/Bun.spawnSync for subprocesses, Bun.file()/Bun.write() for file I/O, Bun.serve for the OAuth callback, Bun.which/Bun.version for the doctor checks, and Web Crypto (crypto.getRandomValues/crypto.randomUUID + Bun.CryptoHasher) for PKCE/IDs. scira doctor now also reports the Bun runtime version.

0.1.5 — June 13, 2026

TUI fixes

• Sent messages anchor to the top — when you send a message it now jumps to the top of the conversation with empty room below for the incoming reply, instead of staying pinned to the bottom of the screen. As the reply grows past the viewport the feed falls back to following the latest output. A blank line above keeps the message off the top edge.
• Loading indicator while the agent works — the gap between sending and the first streamed output is no longer blank. An animated spinner shows a slowly rotating status line ("Connecting the dots…", "Gathering context…", "Chasing down answers…", and more) and stays visible for as long as the agent is doing non-text work (reasoning, tool calls).
• Chrome DevTools MCP tools auto-collapse — devtools_* browser tools now start collapsed in the timeline like the other long-output tools, so a browser session no longer floods the feed. Click to expand any of them.

0.1.4 — June 13, 2026

TUI fixes

• Clickable links — markdown links in the timeline now render as native OSC 8 terminal hyperlinks via ink-link, so supporting terminals (iTerm2, WezTerm, kitty, VS Code, GNOME Terminal) make them Cmd/Ctrl-clickable. Terminals without hyperlink support fall back to the same styled text as before, with no change to layout.
• Slash menu no longer leaves a gap — the message feed is now sized to its content and anchored to the bottom, so opening the command menu (and the help, approval, and link overlays) no longer leaves blank padding between the conversation and the input box.
• Overlays align with messages — the command, help, approval, and link boxes dropped their stray horizontal margin, so they now share the same left edge and width as the input box and the message column.

0.1.3 — June 13, 2026

Latest published.

X search

• xSearch tool — search X (formerly Twitter) for real-time posts using the xAI Grok responses API with native xSearch tool support. Returns up to 30 posts per query with full tweet text and author handles hydrated via react-tweet. Use it for breaking news, public reactions, and social discussions alongside webSearch.
• Requires XAI_API_KEY. When the key is set the tool is automatically available to the agent; when absent it is omitted from the toolset entirely so the model is never offered a tool it cannot use.
• Supports date ranges, up to 5 parallel queries, and per-query handle allow/block lists.
• xSearch calls are logged to the run audit log alongside all other tool calls.

Code agent

• Plan mode — explore and plan before making changes. Read-only commands run freely; writes, edits, and other side effects wait for approval. Toggle it with /plan.
• Dedicated tool sets — research, one-shot, and coding tools are split out, so the agent gets the right toolbox for the task.
• Workspace tools — grep and list files across your project, with separate path handling for run artifacts and project files.
• Background tasks — long-running commands can run in the background instead of blocking the turn.
• Todos — a tool for tracking multi-step work within a run.

Setup & keys

• Guided key setup — scira keys now shows per-provider signup links and exactly where to save each key.

Internal

• Agent tools moved from src/agent/ into src/tools/ and split by capability.
• Added a bash approval policy and broader test coverage.

0.1.2 — June 11, 2026

• Improved CLI commands across init, doctor, and run handling.
• Enhanced TUI theme support: dark, light, and auto, following your terminal, with /theme to override.
• Dependency updates.

0.1.1 — June 11, 2026

• Renamed the package to @scira/cli.
• Migrated the build and CI from npm to Bun.

0.1.0 — June 11, 2026

Initial public release.

Research

• Grounded web search through Exa, Parallel, or Firecrawl, with full-page reading via readUrl.
• Claim extraction and verification — every factual statement is tied to a source URL and a status.
• Markdown reports with inline citations, written to .scira/runs/<id>/report.md.

Code

• Workspace tools: readFile, writeFile, editFile, and bash.
• Approval modes — manual, suggest, or auto — so nothing runs without your say-so.

Interface

• Interactive Ink TUI with session history, slash commands, @ file mentions, and collapsible tool output.
• Headless mode: scira "question" writes a report and exits.

Providers & extensions

• Model providers: Vercel AI Gateway, xAI, Cloudflare Workers AI, and Hugging Face.
• MCP support, including built-in Chrome DevTools for live page research.