ticker-tape — Financial Data Terminal

v4.2.0

Real-time quotes, thesis-driven portfolio views, technical analysis, and AI chat — all in a TUI that fits in a tmux pane.

Architecture

Built on Textual (Python TUI framework) with Rich markup rendering. Data layer uses yfinance with a TTL-cached info pipeline and parallel batch fetching via ThreadPoolExecutor. IBKR integration through MCP streamable HTTP with multi-account support.

Performance

• Refresh — 15-second quote cycle, parallel sparkline fetches (6 workers); quotes display before charts load.
• Cache — 10s TTL on .info calls, 120s on technicals, 60s on benchmark history. Eliminates redundant yfinance requests across sidebar, thesis, and lookup views.

Display

• Compact mode — Toggle with c for a two-line-per-symbol view with sparklines, earnings, technicals, and sidebar watchlist. Auto-enables on narrow terminals.
• Watchlist groups — Named symbol groups with sidebar headers, synced to thesis buckets at runtime.
• i18n — Full English/Chinese with 500+ translation keys, CJK-aware column padding via pad().

Monitoring

• Feed health — The watchlist title carries a FRESH / RECOVERING 45s / DELAYED 6m chip from the age of the last good fetch; a cold start reads as recovering, never delayed. Every quote is stamped when its number actually landed, and a row whose print is late wears a dim age while a cache fallback wears a red STALE 7m, so a snapshot is never painted as current. The 5-minute ⚠ STALE banner still fires above the list.
• Alerts — Smart alerts on price levels, RSI thresholds, SMA crossovers, volume spikes, and margin cushion. Fire-once trigger with auto-removal; technical alerts on a 60s eval cycle.
• NLV History — SQLite-backed NLV snapshots every 60s via peewee ORM (WAL mode). timeline shows a 90-day ASCII chart with drawdown and leverage trend.
• P/L conventions — All change/P&L/drawdown math goes through a single pricing module. Daily change % baselines on the previous regular-session close (bulk history); extended-hours % baselines on the last regular-session close; daily P&L % is recomputed from raw P&L over NLV. Zero/negative/missing baselines yield no figure — never a fake 0%.

Trading workflows

• Morning Briefing — brief assembles portfolio health, macro context (10 indicators including DXY, 10Y, BTC), watchlist movers, sector snapshot, news headlines per top mover, and upcoming earnings with EPS estimates. brief ai adds AI synthesis.
• Position Sizing — whatif [buy/sell] SYM QTY runs an IBKR what-if with concentration and cushion analysis. Omit the direction to see the buy and sell legs side by side.
• Earnings Tracker — surprises shows watchlist-wide EPS beat/miss history with persistence to SQLite.
• Backtest / Thesis Replay — bt [SYM] [ccy CAD|USD] replays the fills ledger against a benchmark: equity curve with entry/exit marks, book vs buy-&-hold vs alpha, max drawdown. Mixed CAD/USD books normalize to one report currency at each date's FX rate. Seed the ledger from an IBKR Flex Query with python flex_import.py <export.csv> (the live execution API only reaches ~7 days back; Flex reaches years).
• Shadow books — shadows replays standing counterfactual ledgers (data/shadows/*.csv — "never trimmed", "kept the hedge") through the same engine over the same bars, so the running dollar cost of every major decision stays visible.
• Time travel — tt [date] freezes the terminal AS OF a past date: that day's closes, and positions/cost basis/P&L reconstructed from the fills ledger. ←/→ scrub by day, Home returns to live.
• Thesis breakers — breakers / tw renders the external watcher's sell-discipline checklist: health headline, fired/holding grouping, catalyst calendar, rotation estimate. Read-only glass — the watcher owns the discipline.
• Decision cards — a fired alert now attaches a pre-computed playbook (e.g. a trim ladder sized to restore a target margin cushion) so the next step is a decision, not a research task. note annotates fills with one-line thesis notes stored in a sidecar (survives ledger re-imports).
• Volume sparks — spark vol swaps every row's price sparkline for a volume histogram coloured by each bar's close direction; spark price restores it. Persisted.
• Sidebar layout — layout pulse calendar pnl risk sets which sections show and in what order (layout add earnings, layout rm risk, layout reset); persisted. The calendar section is an ambient countdown of the next six merged catalyst/macro rows; earnings lists the soonest prints in the watchlist with the EPS estimate.
• Wire order — wire top ranks the fragwire feed by priority (event type × on-book × thesis × source credibility × age decay, live rows first); wire latest is newest-first. The choice persists across sessions.
• Event tape — bottom strip of recent trading events (alerts, movers, calendar); toggle with Ctrl+E or events on/off. Deduped ring buffer, pure data structure.

Research & memos

• Deep-Dive Analysis — analyze <target> runs a Gemini Pro memo with full tool access and Google grounding. Memos archive to data/analyses/{slug}/ with YAML front-matter (target, date, conviction, key_claim) and load prior memos into context for "since last memo" continuity.
• Archive Navigation — memos / memo commands browse and reopen past memos inside the TUI: per-slug listings with conviction-color badges, banner view with front-matter, body rendered through the same markdown pipeline as a live analyze.

Screens

View	What it shows
Thesis	Two-line per symbol: price, change, ext hours, sparkline, RSI, 52-week range bar, earnings countdown, SMA signals, volume ratio, ATH%, relative strength. Portfolio header with breadth metrics and 10-indicator market context (incl. gold).
Heatmap	Color-coded performance grid sorted by daily change
Technicals	SMA 20/50/200, RSI, MACD with crossover detection, Bollinger Bands, ATR, relative strength vs benchmark
Intraday	5-minute bars with VWAP overlay, multi-row tall sparkline charts
Lookup	Full stock profile: valuation (P/E, P/S, PEG, EV), margins, financials, ownership breakdown, analyst consensus
Earnings	Calendar with countdown, EPS estimates from yfinance calendar dict
Economic	FOMC, CPI, NFP, GDP, PCE dates with urgency coloring
Event	event fomc / event 3 / event SYM opens a calendar row as a workspace: what it is in plain language and why it matters, the sectors and public proxies it touches with their live move, prior / consensus / actual with the surprise once printed, and a countdown with a pre / live / post phase
Insider	Recent insider transactions with type/value/shares
Options	Options chain with IV, ATM/ITM/OTM tagging, moneyness filtering, expiration picker. A priced strip above the chain: ATM-straddle expected move (±$ / ±%) against the typical realised print with a rich/cheap verdict, ATM IV, 25-delta put/call skew, IV term structure across the nearest expiries, and volume/OI outliers
Correlation	NxN correlation matrix across watchlist, color-coded by strength, avg pairwise metric
Comparison	Side-by-side multi-symbol performance comparison
Screening	Quick multi-symbol comparison table for filtering ideas
Signal boards	Saved rule sets over the technicals the sidebar already computes (sig add momo rsi>60 vol>1.5 above:sma200 rank vol), evaluated over the watchlist into match / pending / miss bands with the reason on every row; save matches as a group, or arm the closest single alert for a name
Timeline	90-day NLV ASCII chart with drawdown from peak, leverage trend, cushion
Surprises	Watchlist-wide earnings surprise history: EPS beat/miss, price reaction, streaks
Impact	Per-symbol historical earnings reactions: EPS surprise, next-day price move, beat streak, beat rate
Sizing	Pre-trade what-if: margin impact, concentration weight, cushion before/after
Briefing	Morning briefing: portfolio health, macro (10 indicators incl. DXY/10Y/BTC), movers, sector snapshot, news headlines per top mover, earnings with EPS estimates
Archive	Browse written analyze memos: slug list with counts/dates, per-slug numbered listing with conviction color + key claim, reopen view with front-matter banner
Cockpit	Risk cockpit: traffic-lit verdict, per-position margin treatment (from SELL-all what-ifs), VaR/vol/beta/drawdown, survivability
Carry	Cost of carry: margin interest vs dividend income, per-day $, % of NLV, breakeven return on the borrowed slice
Catalyst	Forward per-symbol catalyst calendar (conferences, product/policy dates) merged with the macro econ calendar
Backtest	Fills-ledger replay vs benchmark: equity curve with entry/exit marks, alpha, max DD, FX-normalized
Shadows	Counterfactual ledgers vs the real book — the running cost of past decisions
Breakers	Thesis-breaker checklist from the external watcher: health headline, grouping, catalyst calendar, rotation estimate
Time travel	The whole terminal AS OF a past date, ←/→ day scrubbing, ledger-reconstructed book

AI Chat

Nine models across three providers — switch mid-conversation with model.

Model	Provider	Thinking	Context	Notes
Flash 3.8	Google	1,024	900K	Fast answers, cheapest
Gemini Pro	Google	2,048	900K	Deep analysis
Sonnet 5	Anthropic	adaptive	1M	Balanced
Opus 5	Anthropic	adaptive	1M	Strongest Opus reasoning
Fable 5.1	Anthropic	adaptive	1M	Most capable — top reasoning
GPT-5.6 Sol	OpenAI	low	120K	Flagship GPT — low effort (already plenty for chat)
GPT-5.6 Terra	OpenAI	high	120K	Balanced GPT, cranked effort
GPT-5.6 Luna	OpenAI	high	120K	Fast, cheapest GPT, cranked effort
GPT-6 Astra	OpenAI	medium	120K	Most capable GPT

ticker> model

═══ MODELS ═══
Type 'model' to list, 'model <name>' to switch.

  ── GEMINI ──
  ◆ flash        Flash 3.8                 gemini-3.8-flash               ✓
    pro          Gemini Pro                gemini-3.1-pro-preview         ✓
  ── CLAUDE ──
    sonnet       Sonnet 5                  claude-sonnet-5                ✓
    opus         Opus 5                    claude-opus-5                  ✓
    fable        Fable 5.1                 claude-fable-5-1               ✓
  ── GPT ──
    sol          GPT-5.6 Sol               gpt-5.6-sol                    ✓
    terra        GPT-5.6 Terra             gpt-5.6-terra                  ✓
    luna         GPT-5.6 Luna              gpt-5.6-luna                   ✓
    astra        GPT-6 Astra               gpt-6-astra                    ✓

Context System

The AI assistant has a layered context system — it knows who you are, what the market is doing, and what you've told it before.

Live market context — The system prompt is rebuilt on every chat session with current data: real-time quotes for all watched symbols (price, change, extended hours), technical signals (RSI, SMA position, distance from 52-week high), and portfolio thesis bucket groupings. If IBKR is connected, condensed account summaries (NLV, leverage, margin cushion, top holdings) from all configured accounts are fetched in parallel and injected. The AI can answer "what's my cushion?" or "which positions are above their 50-day?" without running any commands first. Anthropic calls split the system prompt into a stable persona block (cached, ~90% read discount on subsequent turns) and a volatile live-data block (fresh every turn). Gemini calls similarly cache the stable prefix when above the per-model token threshold.

Slash commands — Feed specific data to the AI mid-conversation: /ta AAPL runs technicals and injects the output, /pos pulls IBKR positions, /acct gets account summary. The AI analyzes the data in context.

Web search — When the AI needs current information beyond what's in context, it searches the web automatically: Anthropic models use their native web_search tool (capped at 2 searches per turn to bound cost), Gemini grounds via Google Search, and OpenAI/other paths fall back to Tavily (optimized for LLM consumption) with DuckDuckGo as a free fallback. No manual trigger required.

Image Input

Paste images into chat with Ctrl+P (macOS clipboard) or drop file paths directly into the input. Works across all nine models — Anthropic, Gemini, and OpenAI all receive the image as base64 content blocks in their native format. Use it for chart analysis, screenshot questions, or anything visual. Hard cap at 2 MB per image (override with TICKERTAPE_MAX_IMAGE_BYTES); oversize images get rejected with a notify, since image tokens add up fast — a 4K screenshot can cost ~16K input tokens per turn the conversation references it.

Memory

Memories are persistent facts that survive across sessions, model switches, and history compaction. Stored as JSON on disk and injected into every model's system prompt, so all nine models share the same knowledge base.

Three ways to save: memory add <text> from the command bar, remember <text> while in chat mode (direct, no API call), or just tell the AI conversationally — "remember that AAPL reports Jan 30" — and it saves automatically.

Deleting works the same way — memory delete 5 from the command bar, or tell the AI "forget memory 5" in conversation.

Compaction — memory compact sends all stored memories through Flash 3.8, which distills them into a smaller set of consolidated facts. Useful when you've built up 20+ memories and want to keep the signal without the noise.

ticker> memory

═══ MEMORIES ═══

  #1  AAPL reports Jan 30, after hours                   2026-03-28
  #2  Sold 200 MSFT at $420 for rebalancing              2026-03-26
  #3  Earnings week — no new positions until Friday      2026-03-24
  #4  TSLA support at $180, resistance at $210           2026-03-22

memory <ID> (full text) · add <text> · delete <ID> · compact

History

Conversation history persists to disk across sessions and model switches. history shows a paginated view of past exchanges with per-model cost breakdown.

ticker> history

═══ HISTORY (12 turns, page 1/2) ═══
  Fable $0.04  Opus $0.03  Sonnet $0.01  GPT $0.02  Gemini $0.02
  ~8.2K tokens  ~$0.12

  1  Q: what do you think about AMZN's setup going into earnings?
     sonnet ▸ AMZN looks solid — RSI 52 is neutral, sitting right on the 50d...

  2  Q: compare it to GOOGL
     sonnet ▸ Both are mega-cap tech but with different growth drivers...

history 2 · peek N · search <term> · delete N[-M] · compact · clear

history search <term> finds old conversations. history delete 3 removes a single exchange. history delete 6-10 removes a range.

Compaction — history compact sends the full history through Flash 3.8, which produces a concise summary. The original messages are replaced with a single summary entry. The AI retains the knowledge without burning tokens re-reading 50 old exchanges.

Chain-of-Thought, Search & Token Usage

Models with thinking budgets or reasoning effort (Flash 3.8, Gemini Pro, Claude Sonnet, Claude Opus, and the GPT-5.6 tiers + GPT-5.5) stream their internal reasoning before responding — buffered by paragraph for readability. Toggle visibility with Ctrl+O. When hidden, thinking still runs and improves the answer.

Native web search fires automatically when a model needs current information. Each provider uses its own search: Claude uses web_search, Gemini uses Google Search with grounding metadata, GPT uses the Responses API web_search. Search indicators appear inline during thinking. Citation markers are stripped from final output.

Token usage displays input/output token counts, elapsed time, and per-model cost breakdown. Toggle with Ctrl+T. Anthropic calls use prompt caching on the system block — the static persona/context block is written once and read at 90% discount on subsequent turns. Cache hits are tracked per call.

Gemini context caching (opt-in) — set TICKERTAPE_GEMINI_CACHE=1 to cache the stable system-prompt prefix (user profile + communication rules + output format + memory-tool instructions + tool schemas) server-side via google-genai's caches API. Per-call, only the volatile context (date, live quotes, IBKR snapshot, technicals, recallable memories) flows through contents. Cached portion bills at ~25% of normal input rate, so a chat session with the cache active sees 70% input-cost reduction on flash-preview. Disabled by default because the per-hour storage fee ($1/M tokens/hour) is only worth paying for active sessions; break-even is ~1 call/hour. The token-usage footer shows (N,NNNt cached) alongside the prompt total when a cache hit occurred. Cache fail-opens: any error during create/lookup falls back to the uncached path, so caching can never silently break a chat turn.

A random thinking placeholder (from a pool of ~80 words including "Catfishing", "Moonwalking", "Glazing") shows while waiting for the first token.

Q: what's happening with the TSMC arizona fab timeline?
Moonwalking...
Analyzing the Request
    The user is asking about TSMC's Arizona fab construction
    timeline. I should search for the latest updates on the
    N4P process node ramp and any delays or policy changes.
Searching the Web
    🔍 Searching: "TSMC Arizona fab timeline 2026"
    🔍 Searching: "TSMC N4P production ramp update"
Synthesizing
    The latest reports confirm first silicon on the N4P node
    ahead of schedule. This is significant for the domestic
    semiconductor supply chain buildout...

A:
TSMC's Arizona Fab 1 hit first silicon on the N4P node two weeks
ahead of the revised schedule. Yield rates are tracking at 92%,
which is competitive with Tainan. Fab 2 (N3E) broke ground last
month with a 2028 target. The CHIPS Act second tranche disbursed
$2.1B in March, removing the funding overhang.

┌───────────────────────────┐
│ ↑ 3,156t  ↓ 809t  ◷ 10.4s │
└───────────────────────────┘

Journal & Resume

journal save [N] captures the last N chat exchanges losslessly to a persistent journal — useful for archiving analysis or trade rationale. journal lists saved entries with timestamps.

resume replays the tail of chat history (~3000 chars) with full markdown rendering, so you can pick up where you left off after restarting.

Auto News Context

Mentioning a watchlist ticker in chat automatically fetches its recent headlines and injects them into the AI's context. No manual /news command needed — just say "what do you think about AAPL" and the model sees the latest headlines alongside its market data.

Agent Tools

The AI can call ticker-tape functions directly when you ask about specific data — no slash commands needed. Ask "what's MSFT's RSI?" and the model calls get_technicals(MSFT) to fetch real-time indicators, then analyzes the result. Ask "set an alert when NVDA crosses 200" and it calls set_alert — the alert lands in the same store the alert command uses. "What did my last AAPL memo say" searches the analyze archive. 19 tools across all nine models.

Tool	What It Does
get_quotes	Fresh real-time quotes for any symbols
get_technicals	RSI, SMA, MACD, Bollinger, ATR, volume ratio
get_news	Recent headlines with age
get_fundamentals	P/E, margins, growth, market cap
get_chart	Price history (configurable period/interval)
get_intraday	Today's 5-min bars with VWAP
ibkr_get_positions	Portfolio positions with P&L
ibkr_get_account_summary	Account NLV, margin, buying power
ibkr_briefing	Risk dashboard with margin metrics
ibkr_stress_test	Stress test under configurable scenarios
ibkr_what_if	Simulate trade: margin impact, cushion change
get_earnings_surprises	Historical EPS beat/miss across watchlist
get_briefing	Morning briefing: portfolio, macro, movers, earnings
search_memos	Search the analyze-memo archive (claims + bodies)
set_alert / delete_alert / list_alerts	Manage price/RSI/SMA/volume/cushion alerts
memory_add / memory_delete	Save/delete persistent memories

Tools are defined once in a provider-agnostic registry and translated to each provider's native format (Anthropic tool_use, Gemini function_declarations, OpenAI function tools). Execution is local — the model requests a tool, ticker-tape runs it in Python, feeds the result back, and the model continues with analysis. Tool rounds are capped per turn, so a confused model can't loop forever.

Write tools (set_alert, memory_add, deletes) validate and normalize arguments before persisting — junk operators or unknown alert types are refused with an error the model can correct, never silently stored as dead alerts.

Broker access is strictly read-only. The tool layer has no order-placement capability at all — IBKR tool names are checked against a hard whitelist (positions, account summary, briefing, stress test, what-if) before any network call. No model output can place, modify, or cancel an order through this code path.

Clipboard

copy captures the current screen content to your system clipboard (macOS pbcopy, Linux xclip, WSL clip.exe). copy 20 for just the last 20 lines. Falls back to file export if no clipboard tool is available.

Keyboard Shortcuts

Key	Context	Action
Ctrl+O	Chat	Toggle chain-of-thought display
Ctrl+T	Chat	Toggle token usage box
Ctrl+P	Chat	Paste clipboard image (macOS)
Ctrl+N	Chat	New line in input
c	Ticker	Toggle compact mode
l	Ticker	Switch language (en/zh)
Ctrl+E	Ticker	Toggle bottom event tape
← / →	Time travel	Scrub one day back / forward
Home	Time travel	Return to live

Deep-Dive Analysis

analyze <target> (alias: dive) runs a doc-grade memo via Gemini Pro with full tool access and Google grounding. Memos stream to the terminal in real time and are archived to data/analyses/{slug}/{YYYY-MM-DD-HHMMSS}.md with YAML front-matter.

analyze TSLA                    # symbol dive
analyze rotation                # thesis dive (requires thesis key in config)
analyze MSFT earnings             # symbol with angle hint
analyze "why is XLU up"         # freeform — nests under _freeform/<hash>/

Each memo follows a fixed six-section structure: Context, What Changed Since Last Memo, Current Read, Risks / Disconfirming Evidence, Suggested Actions, Sources. The system prompt forces a High/Medium/Low conviction phrasing so conviction level and key claim can be extracted into front-matter — this lets the index (_index.json) surface a timeline view without re-reading every memo.

Prior memos for the same target are loaded into the system prompt (newest 5), so re-running analyze TSLA a month later produces a "since last memo" update rather than a restart. Backend errors are caught and written as error memos (conviction=unknown) so failures stay auditable in the archive.

Archive Navigation

Read what you've written. All commands are disk-only — no AI calls, no tokens spent.

memos                           # list all slugs (memo count + newest date)
memos TSLA                      # list TSLA memos, newest first, with index + conviction
memo 1                          # reopen memo #1 from the last listing
memo latest                     # reopen the newest memo across the archive
memo latest TSLA                # reopen the newest memo for a slug

Slug resolution: uppercase → symbol (memos tsla → TSLA), lowercase → thesis, _freeform/<hash> → freeform. Unknown slug shows a dim not-found message; out-of-range memo N reports the last listing size so you can retry.

The reopen view starts with a banner (target · date · conviction-colored badge · key claim · angle · model · prior-memo count), then streams the body through the same markdown renderer used for a live analyze. Error memos surface with a UNKNOWN dim badge so failed runs remain visible in the timeline.

Status Bar

Scrolling top bar with global indices (S&P, Nasdaq, HSI, VIX, WTI, Brent, Gold, Silver, Natgas) and local ET clock. Character-level scroll using Rich Text object slicing to preserve per-segment coloring. Off-hours swaps to futures tickers (ES=F, NQ=F). VIX and natgas color-coded by absolute level (green/yellow/red thresholds). NYSE holiday detection with orange "HOLIDAY" tag.

Sidebar

Real-time watchlist quotes with refresh flash indicator. IBKR P&L section (daily/unrealized/realized aggregated across accounts, 30s refresh). Risk section (cushion, leverage, NLV from primary account, 60s refresh) with color-coded thresholds. Earnings calendar in the Pulse section.

IBKR Integration

Multi-account MCP client over streamable HTTP. Two accounts on the same or separate IB Gateways, configured via environment variables. Per-account labels with gateway-down detection. Compact column formatting with currency filtering and FX conversion (USD/CAD, USD/JPY, USD/EUR, USD/CNY, USD/TWD, CAD/HKD, CAD/CNY). Consolidated view renders three blocks — per-account summaries plus a "Combined" cross-account roll-up with shared NLV/leverage/margin metrics.

Command	What it shows
/pos	Positions table with cost basis, P&L, weights, daily P&L summary
/acct	Account health: NLV, margin, buying power, cushion, leverage
/pnl	Daily/unrealized/realized P&L with daily return %
/risk	Full risk dashboard: health, concentration, alerts, VaR
/trades	Today's executions grouped by symbol, VWAP, commissions
/orders	Open/pending orders
/detail SYM	Single position deep dive: cost, weight, P&L, margin, performance
/margin	Margin summary, /margin SYM headroom, /margin SYM QTY what-if
/stress	Stress test: preflight, drawdown curve, overnight gap risk
/beta	Portfolio beta vs SPY with per-position breakdown
/corr	Pairwise correlation matrix
/sector	Sector exposure breakdown with HHI
/ibkr	Consolidated cross-account view with FX conversion

═══ POSITIONS (U12345678) ═══  14:32:05 ET

         Shares      Cost     Price         Value             P&L    Wt%
──────────────────────────────────────────────────────────────────────
 AAPL       500 $  185.20 $  198.45    $99,225.00    +   6,625.00  28.3%
 MSFT       300 $  378.50 $  412.30   $123,690.00    +  10,140.00  25.1%
 GOOGL      400 $  155.80 $  168.92    $67,568.00    +   5,248.00  19.4%
 AMZN       250 $  178.40 $  192.15    $48,037.50    +   3,437.50  13.7%

 Total Market Value              $350,482.50 USD
 Daily P&L                       +$2,841.20
 Daily P&L %                     +0.82%

Stack

• textual — TUI framework, reactive properties, CSS styling
• rich — Markup rendering, Text objects for scrolling tape
• yfinance — Market data, technicals, earnings, insider transactions
• google-genai — Gemini chat with streaming, chain-of-thought, Google Search, code execution
• anthropic — Claude chat with extended thinking, web search, code execution, prompt caching
• openai — GPT chat with Responses API web search
• tavily — LLM-optimized web search fallback (DDG as secondary fallback)
• mcp — IBKR MCP client (streamable HTTP, multi-account)
• httpx — Async HTTP transport
• peewee — SQLite ORM for NLV history and earnings persistence (WAL mode)
• pyyaml — Analyze memo front-matter serialization
• pytest — run python -m pytest --collect-only -q for the current count; coverage includes data layer, formatters, screens, chat, tool registry + agent write tools, demo mode, pricing conventions, journal, memory tags, MCP pipeline, smart alerts, db persistence, analyze orchestration, archive I/O, archive views, risk cockpit, cost of carry, semantic recall, catalyst calendar, backtest engine + FX layer + Flex import, time travel, event feed, decision cards, fill notes, shadow books, breakers

Demo

Multi-model AI chat with web search, chain-of-thought, and model switching mid-conversation.

Fully integrated Chinese language support with CJK-aware column alignment.

Changelog

v4.2.0 (2026-09-22) — The July roadmap. layout decides which sidebar sections show and in what order (layout pulse calendar pnl risk, layout add earnings, layout rm risk, layout reset), persisted, with two new ambient sections: a calendar countdown of the next six merged catalyst/macro rows and earnings, the soonest prints in the watchlist with the EPS estimate. spark vol swaps every row's price sparkline for a volume histogram coloured by each bar's close direction, in the sidebar and both thesis views; spark price restores it. The earnings-reaction parity check needed nothing: impact already carried beat rate, streak, average surprise and the average close-to-close move.

v4.1.0 (2026-09-22) — Feed honesty. Every quote is stamped when its number actually landed and the stamp survives the last-known-good fallback, so a row whose print is late wears a dim age and a cache fallback wears a red STALE 7m; the sidebar title carries a FRESH / RECOVERING 45s / DELAYED 6m chip that never calls a cold start delayed. wire top ranks the fragwire feed by event type, watchlist relevance, thesis weight, source credibility and age decay, live rows first with transcript chatter collapsed to the newest chunk per session; wire latest restores newest-first and the choice persists.

v4.0.0 (2026-09-22) — Web parity. The analyst reads built and proven on ticker-tape-web come home, all as pure functions over data the app already fetches. opt SYM now leads with what the market is pricing: the ATM-straddle expected move in dollars and percent, measured against the name's typical realised earnings reaction for a rich/fair/cheap verdict, plus ATM IV, 25-delta skew from Black-Scholes deltas on each contract's own IV, the IV term structure across expiries spread along the curve, and volume-over-open-interest outliers with the number that made them unusual; the chain fetch also stops raising on unquoted contracts and drops Yahoo's sub-1% placeholder vol. Signal boards (sig) are saved screens: field<value, field:low..high, above:sma200, AND-combined with a rank field, run over the watchlist into match / pending / miss bands where a symbol whose inputs have not landed is reported as pending rather than dropped; matches save as a group, and entry arms the closest single alert. The event workspace (event) opens any calendar row, macro or catalyst, into a plain-language read in both locales, the public instruments it touches with their session move, and prior / consensus / actual with the surprise once the print is in.

v3.9.0 (2026-08-11) — The wire release. The fragwire panel catches up to the server: cross-outlet story clustering with ×N badges and wire story N to unpack members, source-credibility pips, plain-English publication bylines, wire read N on-demand article extraction, long-form morning-brief rendering with CJK-aware wrapping, and Chinese output across the whole wire surface (headlines, statuses, health — the last locale leak). Alerts gain in-place editing (alert edit) with a persisted FIRED state; the watchlist drops its forced alphabetical sort in favour of user order with wl up/down/move and a one-line breadth summary. Startup quotes now come from one batched request instead of a per-symbol trickle. The economic calendar carries the full 2027 schedule plus derived quad-witching dates, guarded by a test that fails 60 days before the calendar runs dry. Lookup adds a company Profile block and a stale-quote badge.

v3.7.3 (2026-07-22) — Compact banner persistence. The final compact-thesis refresh now restores the startup banner after clearing the interim render, preserving the MacBook regression fix across the app-mixin refactor. The headless CLI banner and command headings now resolve and display the same release version as the TUI.

v3.7.2 (2026-07-22) — Correctness hardening. Backtest return, alpha, and drawdown calculations now remove later capital contributions instead of reporting deposits as gains; the benchmark receives the same dated cash flows. Shadow books retain ledger currency metadata and normalize every fill and bar into the configured report currency before comparison. Earnings reactions now distinguish before-market from after-market reports, include already-reported same-day BMO events, and refresh existing SQLite rows when corrected reaction data arrives.

v3.7.1 (2026-07-10) — GPT-5.6 tiers. Swapped the two GPT slots for the GPT-5.6 lineup: Sol (flagship), Terra (balanced), Luna (fast/cheapest), plus the prior flagship 5.5 kept. Removed GPT-5.4 mini. The registry keys are the switch names — model sol|terra|luna|5.5. Effort is inverted vs tier: Sol runs low (the flagship is already overkill for a chatbot — save the tokens), Terra/Luna run high to make the cheaper tiers punch up; 5.5 stays medium. Ten models now; the cost breakdown folds all four GPT tiers into one entry by provider, so it needed no change. Pricing per 1M tokens: Sol $5/$30, Terra $2.50/$15, Luna $1/$6, 5.5 $5/$30.

Older releases

v3.7.0 (2026-07-07) — Decision surfaces: breakers, time travel, event tape, decision cards, shadow books. Five features that turn the terminal from a data display into a decision instrument. Breakers (tw): read-only glass over the external breaker watcher — health headline (INTACT/BREACHED), fired/holding grouping, catalyst calendar, rotation estimate, per-breaker qtr coverage badges; the watcher owns the discipline, the terminal renders it. Time travel (tt [date]): the whole terminal AS OF a past date — that day's closes, book reconstructed from the fills ledger via the backtest engine's PositionBook (same matched-sell/average-cost semantics), ←/→ day scrubbing, Home = live. Event tape (Ctrl+E): a bottom strip of recent trading events (alerts, movers, calendar) on a deduped ring buffer. Decision cards: a fired alert attaches a pre-computed playbook — e.g. a trim ladder sized from live IBKR state to restore a target cushion — plus note, one-line thesis annotations on fills stored in a sidecar that survives ledger re-imports. Shadow books (shadows): standing counterfactual ledgers (data/shadows/*.csv) replayed through the same engine over the same bars, making the running dollar cost of past decisions a first-class view. Also: backtest data-quality hardening (split detection in the ledger window, raw unadjusted closes, deduped FX caveats).

v3.6.0 (2026-07-07) — Backtest / thesis replay (bt). Replays the book's realized fills against a benchmark: ASCII equity curve with entry/exit marks, book vs buy-&-hold return vs alpha, max drawdown. Fills come from a local CSV ledger (data/fills.csv) because IBKR's execution API only reaches ~7 days back; flex_import.py converts a multi-year IBKR Flex-Query export into the ledger in one command (column aliases, signed-qty side inference, CAD-listing suffixes). Mixed CAD/USD books are normalized to one report currency (bt [SYM] ccy CAD|USD, default config.BACKTEST_CCY) with each amount converted at its own date's USDCAD rate — never today's, never a fabricated parity rate. Pure engine (backtest.py), data layer (backtest_data.py), FX layer (backtest_fx.py), all unit-tested; demo-seeded.

v3.5.0 (2026-07-02) — Catalyst calendar (catalyst / cx). A forward calendar of per-symbol events beyond earnings — conferences (GTC, OFC), product launches, policy/export-control dates, supply cycles, capex days — the catalysts a thesis turns on that no earnings feed carries. Catalysts persist in a JSON store and the view merges them with the macro econ calendar (FOMC/CPI/NFP/GDP/PCE) into one date-sorted, countdown-colored list; catalyst SYM filters to one name. AI chat maintains it via two new write tools (add_catalyst, list_catalysts), so you can say "add NVDA's GTC keynote on March 17" mid-conversation. Pure formatter (catalyst.py), bilingual, demo-seeded.

v3.4.0 (2026-07-02) — Semantic recall (recall chat tool). Chat could only keyword-match memos; now it can find them by meaning. Memos and the trade journal are indexed into a dedicated vecgrep corpus (the local-first Qdrant + bge-m3 engine), and a new recall agent tool queries it — "what was my take on X, and did it play out" pulls the semantically-nearest past memos/journal entries regardless of wording. The journal (JSON) is rendered to a markdown shadow so it indexes alongside the memo files; reindex runs before each recall and is incremental on vecgrep's side. Degrades to the keyword memo search if vecgrep is unreachable, so a chat turn never hard-fails. Demo mode never touches the real corpus (keyword-only). Config: VECGREP_BIN, VECGREP_CORPUS.

v3.3.0 (2026-07-02) — Cost-of-carry dashboard. New carry command: what the book costs to hold per year. Margin interest (config rate × the debit balance) netted against forward dividend income and any cash-sweep credit, shown as annual $, per-day $, and % of NLV, plus the breakeven return on the borrowed slice. The gateway exposes the debit balance and the actual accrued interest but not the forward rate, so the modeled interest is rendered beside IBKR's real accrual as a reality check — if they diverge, the config MARGIN_RATE_PCT is stale. Handles both leveraged (debit) and cash-positive books. Rates live in config.py (MARGIN_RATE_PCT, CASH_YIELD_PCT, FX_CARRY_PCT), env-overridable. Pure formatter (screens/carry.py), Decimal money math, parallel fetch, bilingual, canned demo data.

v3.2.0 (2026-07-02) — Risk cockpit. New cockpit command (alias hf) composes every margin and risk number onto one screen: a traffic-lit verdict header (cushion, distance-to-forced-liquidation), account margin state, a per-position margin-treatment table, VaR / vol / beta / drawdown, and a survivability section. The margin table derives each position's maintenance requirement from a SELL-all what-if (account maint before minus after) rather than assuming a flat rate — so a position carrying elevated house/concentration margin is measured from the broker's own engine and flagged when its maint/value ratio runs above the book. Every section degrades independently when a tool is unavailable; canned demo data included. Position-table parsing is now header-driven (locates the Value/Weight columns by name) so it reads both the live and demo table layouts. Pure formatter in screens/cockpit.py, parallel fetch in the app worker, bilingual.

v3.1.1 (2026-06-11) — Lineup trim + cost-breakdown fix + ticker scroll repaint. Dropped Haiku 4.5 — too far behind Sonnet 4.6 / Opus 4.8 / Fable 5 to earn a slot (eight models now). Set Fable 5 to effort: medium (adaptive thinking is steered by effort, not a token budget — medium trims spend on the 2x-cost model). Rewrote the history cost breakdown: it was a hardcoded list that silently dropped Fable's cost and showed $0.00 for unused models — now it shows only models with spend this session, folds the GPT models into one entry (like Gemini), and lists Anthropic models individually. Fixed the scrolling ticker tape freezing under tmux — the marquee advanced its offset internally but the screen never repainted, because Textual coalesces the tiny per-tick (1-cell) label updates and tmux/SSH drop them; the tick now forces an explicit refresh() each frame.

v3.1.0 (2026-06-10) — Fable 5 + adaptive thinking. Added Claude Fable 5 (claude-fable-5) as the top model — 1M context, $10/$50 per MTok, the most capable of the lineup (now nine models across three providers). Fable 5 and Opus 4.8 use adaptive thinking (thinking: {type: "adaptive"}); the Claude call path now routes by a thinking_mode flag instead of always sending the legacy budget_tokens form, which 400s on adaptive-only models. Fixed Opus 4.8 along the way — it carried the same latent budget_tokens config and would 400 on thinking turns; switched it to adaptive and corrected its context window to 1M. Also surfaced the previously-undocumented Flash 3.5 (gemini-3.5-flash) in the model list.

v3.0.0 (2026-06-09) — Agent tools + demo mode. Tool layer extracted to chat_tools.py and extended to 19 tools across all seven models / three providers. New tools: get_quotes (fresh quotes on demand), search_memos (analyze-archive search), set_alert/delete_alert/list_alerts, memory_add/memory_delete — write tools validate and normalize args before persisting (GPT's "above"/"below" operators normalized, junk alert types refused instead of stored dead). Hard read-only IBKR whitelist: no tool name outside the read set is ever forwarded to the broker gateway — refusal happens before any network call; the layer has no order-placement capability. OpenAI streaming tool-call fixes (args accumulated by item_id, conversation preserved across tool rounds); Gemini translator type fixes. New demo mode (ticker-tape --demo / TICKERTAPE_DEMO=1): deterministic fake market + fake $500K account on generic tickers, all data under data/demo/, zero network, system prompt sanitized of all real context, seeded NLV history/alerts/memo/chat — 51 screens verified.

v2.7.0 (2026-06-09) — P/L correctness. New pricing module is the single source of truth for change/P&L/drawdown math; quotes, briefing, sidebar, lookup, status bar, market pulse, sectors, commodities, valuation, timeline, and IBKR P&L all route through it. Daily change % baselines on previous regular-session close, extended-hours % on last regular-session close, daily P&L % recomputed from raw P&L over NLV, NLV drawdown clamped at 0. Zero/negative/missing baselines now skip or mark the row stale instead of rendering a fake 0% or sign-flipped figure.

v2.6.1 (2026-04-19) — Archive navigation. memos lists archived slugs (count + newest date); memos <slug> shows each memo's date, conviction level, and key claim with an index column. memo <N> reopens memo #N from the last listing; memo latest [slug] jumps to the newest. Banner shows date, conviction, angle, model; body rendered through _md_to_rich so it flows like a fresh analyze. Read-only — no AI calls, no archive modification.

v2.6 (2026-04-19) — Deep-dive analysis. New analyze <target> command (alias: dive) runs doc-grade memos via Gemini Pro with full tool access and Google grounding. Memos archive to data/analyses/{slug}/{YYYY-MM-DD-HHMMSS}.md with YAML front-matter; prior memos load into the system prompt for "since last memo" continuity. Routes symbol / thesis / freeform targets under their own slug. Conviction level extracted into front-matter for timeline views. Error memos written on backend failure so failures stay auditable. Chat rendering: negative lookahead added to cross-line-bold fix so adjacent bold-prefixed list items stop collapsing onto one line (Gemini Pro dense-list regression).

v2.5.5 (2026-04-16) — View polish sweep. Thesis compact: %H / %R pulled left 1 cell each. Commodities: name column tightened 3 cells. Correlation matrix: proper heatmap — amber band for 0.6-0.8 ("hidden beta" zone), dim amber for 0.4-0.6, white 0.2-0.4; legend updated. Earnings surprises: full rewrite — $ sign in its own column with -$ leading negatives, Beat %/detail treated as fixed-width pair so subsequent columns don't shift, all percentages right-justified on visible width not markup-wrapped length. Valuation screen: column spacing tightened so the row no longer spills off ~80-char panes (inter-column gap 2→1, Chg% built from raw pct not color_pct helper). IBKR consolidated view: third purple "Combined" section for cross-account NLV/Leverage/Margin Util, positions table P&L sign aligned in its own column (all +/- glyphs in one vertical line regardless of magnitude), column headers shifted left to align with data, USD/TWD added as relevant FX pair. P&L view: stripped redundant "of NLV" suffix from daily P&L parenthetical. Briefing AI: brief ai now routes directly through Gemini Flash backend instead of the user's currently-selected chat model (was broken if user was on non-Gemini model).

v2.5.4 (2026-04-16) — i18n polish: briefing portfolio/macro/movers/earnings labels, IBKR account summary labels, and chat help menu all now translate (30+ new keys). CJK-aware column padding via pad() so Chinese labels preserve visual alignment. README changelog caught up.

v2.5.3 (2026-04-15) — IBKR view polish (P&L formatting consistent across all views, +1,234.56 no $, 2-space left pad on consolidated). Account summary adds excess liquidity + buying power; CAD/CNY FX pair. Morning briefing expanded to 40-60 lines with news headlines per top mover, sector snapshot, 10-indicator macro (DXY, 10Y, BTC added), EPS estimates on earnings. Status bar CJK fix — cell-aware slicing via unicodedata.east_asian_width() so Chinese chars (金 etc.) don't get cut off in scrolling tape. 30+ new i18n keys covering thesis indicators, impact screen, sizing, memories, history, chat shortcuts, journal. SOX dropped from compact bar to keep gold visible in Chinese mode.

v2.5.2 (2026-04-15) — System prompt overhaul with precise thesis context, geopolitical standing context, tool-use synthesis instruction. Fixed token counting in Gemini/Claude re-call loops (was showing initial turn tokens, not final response). Fixed Gemini thinking leak in tool re-call. Column spacing tuned for thesis view.

v2.5.1 (2026-04-15) — Smart alerts (price, RSI, SMA crossover, volume spike, margin cushion). NLV timeline with 90-day ASCII chart. Earnings surprise tracker. Morning briefing with AI synthesis. Position sizing what-if. Market context (3 rows: indices, commodities, yields). JsonStore memoization (~25x faster reads). 3 new AI tools. FX pairs in consolidated view.

v2.4.2 (2026-04-13) — Grouped tool indicators, compact tool results, dead code removal.

v2.4.1 (2026-04-13) — Pricing display, cost basis columns, tool UX improvements, SIGINT fix.

v2.4 (2026-04-13) — AI tool use (12 functions across 3 providers), consolidated IBKR view, /detail command, historical trades.

v2.3 (2026-04-09) — IBKR refactor, NaN handling, consistent box widths, screening view.

v2.2 (2026-04-07) — IBKR overhaul with risk sidebar, slash commands for portfolio data, screening view, README refresh.

v2.1 (2026-04-02) — Compact mode with sparklines/earnings/gold/breadth. Watchlist groups. Memory edit via AI. Margin dashboard. Trade grouping. Multi-account IBKR. Commodities screen. NYSE holiday detection. Chat help menu. Resume command. Image input (Ctrl+P). Prompt caching on Anthropic. GPT native search. Gemini code execution. History/memory compaction.

v2.0.5 (2026-03-29) — Chat memory tools, multi-line paste, history ranges, system prompt update.

v2.0.4 (2026-03-29) — Memory/history consolidation, history compact, Haiku/Opus pricing fix.

v2.0.3 (2026-03-27) — IBKR formatting, compact mode, AI screen context injection.

v2.0.2 (2026-03-26) — IBKR formatting fixes, what-if sell, history display improvements.

v2.0.1 (2026-03-26) — Options chain, correlation matrix, watchlist groups, live alerts, earnings upgrade. Tavily search, GPT search, Gemini code execution. Token/cost tracker.

v2.0 (2026-03-25) — Textual rewrite from scratch. 368 tests. Thesis-driven portfolio view, heatmap, technicals, intraday, lookup, earnings, economic calendar screens. Status bar with scrolling indices. Full i18n (EN/ZH, ~250 keys). Shared memory system. Multi-model AI chat (7 models, 3 providers). Streaming with paragraph-level flush. IBKR multi-account integration via MCP.

v1.5 (2026-03-17) — Multi-model AI chat with 7 providers. Streaming paragraph-level flush. Shared memory system across all models. i18n with ~250 translation keys (EN/ZH).

v1.4 (2026-03-15) — IBKR multi-account integration via MCP streamable HTTP. Economic calendar (FOMC, CPI, NFP, GDP, PCE). Price alerts and watchlist persistence.

v1.3 (2026-03-11) — Earnings impact analysis and valuation screen. Sector heatmap and insider tracking. Intraday price action with VWAP.

v1.2 (2026-03-07) — Technicals (MACD, RSI, Bollinger Bands, ATR). Charts and side-by-side comparison. Command bar with ~40 aliases.

v1.1 (2026-03-03) — Scrolling ticker tape widget. Status bar with market state and indices. Sidebar with sparklines and pulse metrics. Sectors, earnings, news screens.

v1.0 (2026-02-25) — Thesis dashboard with 52w range bars. Stock lookup with key stats. Market overview screen.

v0.1 (2026-02-19) — Project scaffolding. Textual app shell. yfinance data layer with TTL caching. Rich markup formatters.

Local validation

Run validation from an unlocked checkout before merging:

venv/bin/python -m pytest tests -q
venv/bin/python -m pip wheel --no-deps --wheel-dir dist .

Verify the built wheel in a separate virtual environment, including runtime imports and the packaged stylesheet. Source decryption stays on trusted local machines; this repository does not require a GitHub Actions decryption secret. Existing local PII and secret guards remain mandatory.