Invest Skill

<skill-dir> is this skill's own directory — the folder that holds this SKILL.md (Claude Code shows it as the skill's base directory; Hermes / Codex / OpenClaw install it inside the plugin). run.sh below always means <skill-dir>/scripts/run.sh.

First-time fork users: run the invest-setup skill first to initialize. The main flow lives in this skill (the AI agent uses the CLI/MCP to view the portfolio / run the committee / replay decision history). The Web GUI has been retired (2026-07) — every capability is exposed via CLI subcommands / MCP tools. The backend is distributed from PyPI (pulled on demand via uvx) at the version pinned in run.sh, which moves with each plugin release: updating the plugin updates the backend. run.sh update re-downloads the pinned version if the uv cache breaks.

openInvest multi-asset AI investment committee. This skill is not Claude-exclusive — any agent that can run shell commands can use it; see "Choosing a path" below.

Reply in the language the user is currently speaking unless they explicitly ask to switch languages.

Choosing a path

The first question is not "which brand of agent are you" — it is "is a human present for this invocation":

Is a user asking in chat right now ("should I buy X / 该不该买 X") where you can
react in real time (wrong tool picked, blocked by a safety gate — the user sees it
on the spot and corrects it on the spot)?
  → Check whether you have an isolated sub-task delegation capability (Agent({...}) / delegate_task)
    Yes → Coordinator (table below)
    No  → Direct

Is this a cron / scheduled run with nobody watching?
  → Always go Direct, no matter which agent you are or whether you can delegate

Why unattended runs always go Direct, even with delegation capability: the Coordinator protocol depends on you improvising "which tool to call and how to assemble the prompt". That is fine normally, but on 2026-07-14 a real Hermes cron ran Coordinator unattended — it did not faithfully call the delegation tool per the protocol, picked its own route instead, then hit the "unattended cron cannot approve dangerous commands" safety gate and stalled. Direct is pure deterministic Python code and never improvises. Saving the small hassle of configuring a key is not worth a probabilistic hang / derailment in unattended scenarios — a bad trade.

Scenario	Who you are	Path	What to run	Credentials
Interactive (user present)	Claude Code (has the Agent({...}) tool)	Coordinator	prepare_committee → spawn 4 subagents → save_committee	No key needed
Interactive (user present)	Hermes (has the delegate_task tool)	Coordinator (Hermes variant)	Same as above; spawn syntax becomes delegate_task(tasks=[...])	No key needed
Interactive, no delegation capability	Codex / OpenClaw / Cursor / Cline / plain script	Direct	run_committee <SYMBOL> one-shot	Needs LLM_API_KEY
cron / unattended, any agent	Anyone, with or without delegation capability	Direct	run_committee / daily_report	Needs LLM_API_KEY

All three paths share the same foundation — the same prompts, the same data preparation (regime classification + probability framing + deterministic fact block), and the same on-disk format (memory/.committee/<date>/<asset>.md; both Coordinator variants must call save_committee with --provider <your-brand> so the transcript is not mislabeled as claude). The only difference is "who plays the 4 LLM roles": on Coordinator you play them yourself (the model the user already subscribes to); on Direct the configured LLM (billed per token) plays them. Verdicts may differ (different models — useful for cross-validation).

LLM_API_KEY is not DeepSeek-only, and the cost is near zero: utils/llm.py works with any OpenAI-compatible endpoint (three envs: LLM_API_KEY/LLM_BASE_URL/LLM_MODEL; DEEPSEEK_* remains supported for compatibility). Want zero cost → providers with free tiers right now (Qwen / Zhipu / MiMo etc.) all work (free-tier terms change — verify before use); don't want the hassle → just use DeepSeek: at daily-report volume (a few assets/day) roughly ¥0.01-0.03 per run, under ¥2 a month. Do not bet the reliability of unattended runs on Coordinator just to save that much.

No sub-task delegation capability and no key configured: do not force your way through Coordinator, do not fabricate a verdict, and do not write code to brute-force around it — just call run_committee (you will get a clear error) and honestly tell the user "LLM_API_KEY needs to be configured".

Decision tree (identical up front, whichever path you take)

1. Run `run.sh doctor`                               ← mandatory first step
   ├─ status: "ready"        → go to step 2 (keep using this skill)
   └─ status: "needs_setup"  → **switch to the `invest-setup` skill** (not handled here)

2. Pick the subcommand by user intent:

   "show portfolio / how much do I have / 看持仓"     → run.sh status
   "assess the situation / risk / concentration
    / 分析战况"                                       → run.sh status (see Constraints before
                                                        calling anything "overweight")
   "what is my strategy / 我的策略是什么"             → run.sh strategy
   "recent trades / transaction log / 最近交易"       → run.sh history
   "where is the market / VIX now / 现在大盘"         → run.sh live_prices
   "if X drops 5% how much do I lose / 如果X跌5%"     → run.sh what_if --symbol X --pct -5
                                                        (X is a yfinance symbol in the user's
                                                        portfolio; --gold-pct / --ndq-pct kept
                                                        for legacy usage)
   "should I buy/sell X / analyze X / 该不该买卖X"    → committee protocol ↓
   "track AAPL / I want to watch TSLA / 跟踪苹果"     → see references/adding-assets.md
   "I already held 3000 X before using this /
    补录我原来就有的持仓"                              → run.sh status first — never re-add a
                                                        symbol it lists (double count); else
                                                        run.sh buy ... --existing-position
                                                        (MCP tool: record_existing_position;
                                                        `Unknown tool` = old server, do NOT
                                                        fall back to buy) — records it WITHOUT
                                                        deducting cash; a plain `buy` is a new
                                                        purchase paid from ledger cash

3. Run the committee per your path:
   - Coordinator → read references/committee-protocol.md (spawn 4 subagents)
   - Direct      → just `run.sh run_committee <SYMBOL>` for a JSON verdict

4. After getting the verdict / cio_memo:
   - **`cio_memo` is a Markdown string** (with `# title ## verdict` structure etc.).
     Render it **as Markdown for the user** — do not print raw JSON and make the
     user parse it themselves
   - Execution step: check the next_step field and guide the user as it says.
     **Never** write memory/ directly (see Constraints)

Coordinator path details (Claude Code only)

Read references/committee-protocol.md and follow it strictly. Full 6 stages:

• Stage 0: same-day check (if memory/.committee/<today>/<asset>.md exists, reuse it directly)
• Stage 1: prepare_committee to get the brief
• Stage 2: Round 1 — 3 Agent({...}) in parallel (Macro + Quant + Risk)
• Stage 3: Round 2 — Cross-challenge (2 Agents)
• Stage 4 (optional): run Round 3+ if not converged
• Stage 5: CIO synthesis (you write it yourself, do not delegate)
• Stage 6: save_committee to persist

Critical warning: the regime_brief / sentiment_brief / valuation_brief / reentry_reference / event_brief emitted by prepare_committee must be pasted verbatim into the corresponding workers' prompts per the instructions: event_brief (news/event layer, may be empty) goes into Macro; regime + valuation + sentiment go into Quant; all three blocks + the path reference go into the CIO. Consequences of omission: Quant loses the probability framing and the defensive-sentinel background; the CIO's EXPECTED_PATH is made up out of thin air; INDEP_DEFENSE_FLAG never reaches the transcript → save_committee's deterministic defense downgrade (crash sentinel) breaks entirely.

Direct path details (any agent)

# One command does it all
<skill-dir>/scripts/run.sh run_committee NDQ.AX

JSON output:

{
  "status": "ok",
  "asset": {...},
  "verdict": {"verdict": "ACCUMULATE", "confidence": 0.72, ...},
  "confidence_lookup": "样本不足（n=12）",
  "cio_memo": "<full CIO memo markdown>",
  "transcript_path": "memory/.committee/2026-05-09/NDQ.AX.md",
  "next_step": "..."
}

Options:

• --force: rerun even if already run today (default reads the cache to save tokens)
• --max-rounds N: cap on cross-challenge rounds (default 1)

Daily report (cron on the host-agent side)

<skill-dir>/scripts/run.sh daily_report   # = uvx openinvest daily_report

Runs the full daily-report pipeline (multi-asset committee + Gemini second opinion + translator + discipline ledger); stdout emits markdown identical to the email body, does not send email — delivery belongs to the host agent (a Hermes cron can forward --no-agent --script output as-is; report formatting is guaranteed uniformly by the backend). On a circuit-breaker trip / unconfigured target_assets, stdout is a structured JSON error. Rerunning reruns the entire committee (burns tokens).

Prerequisite: .env has DEEPSEEK_API_KEY. If the calling agent runs on the user's machine but has no key, tell the user to run run.sh init to configure it first. (Remote-mode exception: the key lives on the hub, not needed locally — see next section.)

Remote mode (hub-and-spoke, multiple devices sharing one dataset)

Recommended new path (2026-07, BETA — not yet field-tested by the author in a real multi-device setup): when the hub runs openinvest-mcp --http (remote MCP), spokes register the HTTP MCP directly — all tools fully available, no CLI forwarding: claude mcp add --transport http openinvest https://<hub>/mcp --header "Authorization: Bearer $INVEST_API_TOKEN" The CLI→REST forwarding below is still supported (maintenance mode); the Coordinator protocol (prepare/save) and doctor/event_check still go through it. For deployment see backend wiki 08 §9.

With INVEST_API_BASE set in .env, this machine is a client: every subcommand is auto-forwarded to the remote hub (another machine running the invest web_api), and this machine has no memory/ and should not have one. All data (portfolio/strategy/verdicts/prompts) lives on the hub; change it once and every device sees it.

Minimal client .env (no DeepSeek key / Gmail / memory needed):

INVEST_API_BASE=https://your-hub.example.com   # or http://10.0.0.x:8765
INVEST_API_TOKEN=<the hub's token of the same name>   # only needed if the hub has auth enabled
# If going through Cloudflare Tunnel + Access, use this pair instead:
# CF_ACCESS_CLIENT_ID=...  CF_ACCESS_CLIENT_SECRET=...

Behavior differences (all other commands are forwarded transparently, output has the same shape as local, decision tree applies as usual):

Command	Remote-mode behavior
doctor	Returns hub-perspective checks + an extra remote section (api_base / auth mode / connectivity)
init	Disabled (data lives on the hub; connecting to the hub only needs the two .env lines above). Error includes a hint
live_prices / correlate	Still run locally (pure yfinance, touches no data)
run_committee	Runs on the hub (DeepSeek key is on the hub); the CLI polls automatically until done; same-day cache uses the hub's date semantics
prepare/save_committee	Via hub RPC — the Coordinator protocol (spawn 4 subagents) is completely unchanged
buy/sell/deposit/... write ops	Land in the hub ledger (history source: skill_remote)
buy --existing-position	Refused (the REST endpoint can't carry it, and forwarding would deduct hub cash) — run it on the hub, or use the remote MCP tool record_existing_position
event_check	Forwards the hub's manual scan; --live / --recall disabled (hub cron already covers them)

Discipline: in remote mode this machine has no memory/, so "reading/writing memory files directly" does not even exist — everything goes through run.sh or the hub API. In the Web API table below, replace :8765 with $INVEST_API_BASE in remote mode, and add Authorization: Bearer $INVEST_API_TOKEN when curling.

Tool lookup (MCP first; long tail in references/tools.md)

MCP users (auto-registered once the plugin is installed; same for Claude Code / Codex): status / strategy / history / live_prices / what_if / discipline / decisions / explain_decision / record_execution / ingest_event / buy / record_existing_position / sell / deposit / withdraw / set_allocations / track_asset / untrack_asset / news_sources / add_news_source / remove_news_source / run_committee — schemas auto-discovered; call them directly, no table lookup needed.

CLI/REST agents, or long-tail operations MCP does not cover (trades intent flow / config whitelist / events / holdings import / gold-specific endpoints) → read references/tools.md (full subcommand table + endpoint table). Full OpenAPI: http://127.0.0.1:8765/openapi.json.

Subcommand names are a closed set — commands outside the table do not exist. When tempted to call names like get_committee_context / analyze_asset, stop and check the table — you are most likely hallucinating; you probably want prepare_committee or run_committee. All output is JSON; always quote numbers from the JSON, never from memory/*.md.

Importing holdings from a broker-app screenshot (you do the OCR; backend has zero dependencies)

When the user sends a broker holdings screenshot: read the image yourself (you have vision), convert each row into text as symbol/quantity/cost/currency/channel, then use import --text "..." (or POST /api/holdings/import) — the backend LLM only parses text, it never receives images. Preview first without --commit; after the user verifies, run with --commit for a non-destructive write.

Feeding news (your search beats any crawler)

You have far stronger search capability than the backend crawler (including Chinese-language sources). When you come across finance news relevant to the user's holdings while browsing/searching, proactively call ingest_event to feed it into the event ledger (MCP tool, or CLI ingest_event --title --url [--snippet --source]) — the backend handles normalization / severity grading / dedup / RAG recall, and resending the same item never double-books. A new item that grades severity ≥ mid, non-neutral, and hits a holding/target automatically re-runs the committee for that symbol (rate-limited and shared with the crawler: 12h per-symbol cooldown that a strictly higher-severity item can bypass, ≤4 per rolling 24h; the result's committee_task_id says whether one fired) — so feed material news only. Ingestion is disabled in advisory mode. A-share / regional-market news especially: that is the crawler's blind spot and you are the only source. If the host has a quotes/news skill installed (e.g. Longbridge), its news is worth feeding too — the ledger cares about the information, not where it came from.

Decision-loop workflow (Decision Review + Reflection)

openInvest does the bookkeeping; you do the collection — that is the host agent's core duty (issue #133 Decision 2).

User asks "why HOLD today / why did it tell me to sell" (Decision Review):

1. explain_decision <decision_id> for the full 4-role debate transcript + CIO memo + path snapshot (+ confidence_lookup, the number to quote next to the verdict — see Constraints)
2. Combine with status (current portfolio) if needed
3. Answer using evidence from the transcript — do not invent reasons yourself

User reacts to a recommendation with "I didn't buy / I bought / I disagree" (Reflection):

1. First ask one question about the reason (valuation too high? insufficient funds? disagrees with the committee? forgot?) — do not skip this; the reason is the one piece of information the system cannot obtain on its own
2. record_execution <decision_id> [--rejected] --reason "..." to write it back (idempotent; the user may change their answer anytime)
3. If the user actually traded → guide them to log the trade (a same-symbol, same-direction fill within 7 days is also auto-matched as a fallback)

User asks "how often did I follow the advice / is the committee reliable": decisions --days 90 → adoption rate + the full verdict↔intervention↔execution↔outcome chain per decision (lists the newest 20 — narrow with --symbol / --verdict, --limit 0 for all; count and summary always cover every match); pair with discipline for the counterfactual P&L of rule-blocked actions (its inaction rate counts live verdicts only). For hit rates use GET /api/verdict_review/summary: it returns separate live / backtest / contaminated buckets — quote only live as performance, never add the buckets together, and say "sample too small" when a rate is null (n<30). Weekend decisions on weekend-closed assets reuse Friday's close, so they are left out of every bucket and counted in weekend_dup_excluded — that is why live.n is below the number of live decisions. When the user rejects the same class of recommendation several times in a row, proactively point out "the divergence pattern between you and the committee" — that is not a bad thing; it is a signal worth recording.

Constraints (guard these, do not break them)

• Don't call the portfolio "overweight → TRIM" from status alone (the 2026-05-12 mistake): the portfolio is often only part of the user's money, and the committee's concentration lens is off by default (ADR-020). Raise concentration only if the user asks about it. The cap on any add-on buy is always portfolio cash — this never changes
• Do not run daily_report proactively in conversation — unless the user explicitly says "run the deep analysis" / "run full report" or you are setting up the daily cron. That path burns DeepSeek tokens. In conversation, single-asset run_committee is enough.
• Never fabricate live prices. Always go through run.sh status or live_prices. yfinance may return stale data; watch the is_stale flag.
• Never write memory/ directly. All state changes go through CLI subcommands / MCP tools (atomic write + fcntl lock + audit trail). Direct editing causes schema drift + concurrent-write corruption.
• Do not rerun the committee for the same asset on the same day — run_committee reads the cache by default; on the Coordinator path, check with ls memory/.committee/<today>/<SYM>.md first.
• Do not fabricate CIO confidence. When workers disagree sharply, honestly write confidence: 0.4-0.5.
• Next to a verdict, quote confidence_lookup, not confidence. run_committee (CLI + MCP), explain_decision, /api/committee_sessions and /api/committee/{task_id} return it: how similar verdicts turned out 30 days later (deterministic table refreshed by the daily verdict_review job from live verdicts plus the prospective paper fleet; when this install has n<30 for that verdict it falls back to a fleet-only default table shipped with the package). The text ends with its source — 含纸面舰队样本 / 默认表 / 本机样本 — keep it when you quote the number: paper-fleet runs use a neutral portfolio and no event/valuation/sentiment inputs. For HOLD it is the share whose 30-day move stayed inside the normal volatility band, shown next to the market's own sideways base rate — that measures how sideways the market was, not accuracy, so never call it a hit rate. n<30 reads "样本不足（n=…）"; rule-forced HOLDs / self-reported ≤0.4 read "输入缺失/强制 HOLD". confidence is the CIO's self-reported number, kept for the record only (historically no better than a constant); if you mention it, label it self-reported (自报).
• Do not leak the user's email or other personally identifying information — never hard-code it in output.

Where to look when something breaks

Read the doctor JSON output carefully. Every check has a hint field telling you how to fix it. If doctor is all green but a subcommand still fails, read references/troubleshooting.md.

Common Direct-path errors:

• error: DEEPSEEK_API_KEY 未设 (backend's verbatim "key not set" message) → no key in .env; run run.sh init to configure it
• error: asset X not in strategy.target_assets → add X to the strategy first, see references/adding-assets.md

References index

File	When to read
references/tools.md	Full subcommand table + Web API endpoint table (long-tail operations MCP does not cover)
references/committee-protocol.md	Running the committee on the Coordinator path (Claude Code only)
references/two-paths.md	Understanding Coordinator vs Direct / DeepSeek cron triggering
references/adding-assets.md	User wants to track a new symbol
references/troubleshooting.md	doctor all green but still failing
references/onboarding.md	First-time install goes to the invest-setup skill; this file is kept as a detailed reference

For deeper architectural context see the project wiki: github.com/longsizhuo/openInvest/tree/main/docs/wiki