feat: add hedge-fund-hermes template with gbrain-backed output

win4r · claude · win4r · commit 9fbc74e74888 · 2026-04-11T21:10:00.000-07:00
The original hedge-fund template tells each analyst to publish their
report via `clawteam inbox send ...`. OpenClaw workers reliably execute
that. Hermes workers often don't — they treat -q as "answer this" and
return analysis to stdout without running the send command.

Result: kanban shows COMPLETED, but inboxes are empty. Users fall back
to parsing tmux scrollback which is fragile.

Fix: a Hermes-native template variant that uses gbrain as the output
channel instead of clawteam inbox. Hermes workers reliably call
mcp_gbrain_put_page because gbrain is their native MCP surface.

Changes:
- clawteam/templates/hedge-fund-hermes.toml (new)
  - command = ["hermes"] by default (no --command flag needed)
  - Each analyst stores findings via mcp_gbrain_put_page with slug
    {team_name}-&lt;analyst&gt; (buffett/growth/technical/fundamentals/
    sentiment/risk/final-report)
  - Portfolio-manager and risk-manager read from gbrain via
    mcp_gbrain_query + mcp_gbrain_get_page instead of clawteam inbox
  - 100% parity with hedge-fund.toml analysts, prompts rewritten for
    Hermes idioms

- skills/hermes/SKILL.md
  - Added "Reading Output from hedge-fund-hermes" section explaining
    the gbrain query pattern
  - Recommends hedge-fund-hermes over hedge-fund for Hermes users

- README.md
  - Added hedge-fund-hermes example to the AI Hedge Fund section

Benefits vs hedge-fund (OpenClaw-style):
- Findings persist in Postgres across sessions
- Cross-session memory: next week's analysis can reference prior reports
- No tmux scrollback fragility
- Works with unmodified Hermes installations (no coordination-prompt
  tuning needed)

Requires: gbrain MCP server configured in ~/.hermes/config.yaml.
Without gbrain, this template degrades to the same tmux-scrollback-only
behavior as hedge-fund.toml.

474 tests pass.

Co-Authored-By: Claude Opus 4.6 (1M context) &lt;noreply@anthropic.com&gt;
diff --git a/README.md b/README.md
@@ -370,11 +370,17 @@ Leader agent:
 A TOML template spawns a complete 7-agent investment team with one command:
 
 ```bash
+# OpenClaw (default)
 clawteam launch hedge-fund --team fund1 --goal "Analyze AAPL, MSFT, NVDA for Q2 2026"
+
+# Hermes Agent — use the Hermes-native variant that stores findings in gbrain
+clawteam launch hedge-fund-hermes --team-name fund1 --goal "Analyze AAPL" --force
 ```
 
 5 analyst agents (value, growth, technical, fundamentals, sentiment) work in parallel. Risk manager synthesizes all signals. Portfolio manager makes final decisions.
 
+Hermes users: the `hedge-fund-hermes` variant rewrites each analyst's task to use `mcp_gbrain_put_page` for output persistence instead of `clawteam inbox send`. Findings land in gbrain pages with predictable slugs (`<team-name>-buffett`, `<team-name>-growth`, etc.), persist across sessions, and can be queried later.
+
 Templates are TOML files — **create your own** for any domain.
 
 ---
diff --git a/clawteam/templates/hedge-fund-hermes.toml b/clawteam/templates/hedge-fund-hermes.toml
@@ -0,0 +1,237 @@
+[template]
+name = "hedge-fund-hermes"
+description = "AI Hedge Fund - Hermes-native variant using gbrain for persistence (not clawteam inbox)"
+command = ["hermes"]
+backend = "tmux"
+
+# DESIGN NOTE: This template is a Hermes-native rewrite of hedge-fund.toml.
+# Key differences from the OpenClaw variant:
+#   1. Analysts store findings in gbrain via mcp_gbrain_put_page (not clawteam inbox send).
+#      Hermes workers treat -q as "answer this prompt" and reliably return analysis
+#      to stdout, but they often skip the post-analysis `clawteam inbox send` step.
+#      Storing to gbrain uses Hermes's native MCP muscle memory.
+#   2. Output persists in Postgres. Works across sessions. Next week's analysis can
+#      reference prior findings.
+#   3. Portfolio-manager and risk-manager read findings via mcp_gbrain_query
+#      instead of clawteam inbox receive.
+#
+# Requires: gbrain MCP server configured in ~/.hermes/config.yaml.
+# Without gbrain, this template degrades to the same tmux-scrollback-only behavior
+# as hedge-fund.toml under Hermes.
+
+[template.leader]
+name = "portfolio-manager"
+type = "portfolio-manager"
+task = """You are the Portfolio Manager and team leader of an AI hedge fund.
+Your goal: {goal}
+
+Workflow (read-from-gbrain, not from inbox):
+1. Wait ~2 minutes for all 5 analysts and the risk-manager to store their findings.
+2. Query the knowledge brain for analyst reports:
+   mcp_gbrain_query("{team_name} analyst")
+   Expected pages (slug pattern): {team_name}-buffett, {team_name}-growth, {team_name}-technical, {team_name}-fundamentals, {team_name}-sentiment, {team_name}-risk
+3. Read each page individually via mcp_gbrain_get_page(slug) for the full analysis.
+4. Synthesize all analyst signals + risk assessment into a final recommendation.
+5. Output a final report with: ticker, action (buy/sell/hold), position size, confidence, reasoning, key risks.
+6. Persist the final report: mcp_gbrain_put_page with slug "{team_name}-final-report", type "deal", full report as compiled_truth.
+7. Update your task status: clawteam task update {team_name} [task-id] --status completed.
+
+If any analyst's page is missing after 3 minutes, proceed with partial data and note the missing analysts in the final report."""
+
+[[template.agents]]
+name = "buffett-analyst"
+type = "value-analyst"
+task = """You are Warren Buffett. Analyze stocks using value investing principles.
+The team goal is: {goal}
+
+Your analysis framework:
+- Moat analysis: durable competitive advantages, pricing power
+- Financial health: ROE >15%, debt-to-equity <0.5, operating margins >15%
+- Management quality: share buybacks, dividend history
+- Intrinsic value: owner earnings DCF with 15% margin of safety
+- Book value growth: CAGR and consistency
+
+OUTPUT INSTRUCTIONS (Hermes-native — do NOT use clawteam inbox send):
+After completing your analysis, store findings via mcp_gbrain_put_page with:
+  slug: {team_name}-buffett
+  type: deal
+  title: "{team_name} Buffett-style Value Analysis"
+  compiled_truth: Full analysis including:
+    - SIGNAL: bullish | bearish | neutral
+    - CONFIDENCE: 0-100
+    - MOAT ANALYSIS
+    - FINANCIAL METRICS (ROE, D/E, margins)
+    - INTRINSIC VALUE vs current price
+    - REASONING (2-3 paragraphs)
+
+Then update task status: clawteam task update {team_name} [task-id] --status completed"""
+
+[[template.agents]]
+name = "growth-analyst"
+type = "growth-analyst"
+task = """You are a growth/disruption investor (Cathie Wood style). Analyze stocks for growth potential.
+The team goal is: {goal}
+
+Your analysis framework:
+- Revenue growth trajectory (>20% preferred)
+- Total addressable market (TAM) expansion
+- Disruptive innovation potential
+- Network effects and platform dynamics
+- Management vision and execution
+
+OUTPUT INSTRUCTIONS (Hermes-native — do NOT use clawteam inbox send):
+After completing your analysis, store findings via mcp_gbrain_put_page with:
+  slug: {team_name}-growth
+  type: deal
+  title: "{team_name} Growth/Disruption Analysis"
+  compiled_truth: Full analysis including:
+    - SIGNAL: bullish | bearish | neutral
+    - CONFIDENCE: 0-100
+    - REVENUE GROWTH TRAJECTORY
+    - TAM & DISRUPTION POTENTIAL
+    - NETWORK EFFECTS
+    - REASONING (2-3 paragraphs)
+
+Then update task status: clawteam task update {team_name} [task-id] --status completed"""
+
+[[template.agents]]
+name = "technical-analyst"
+type = "technical-analyst"
+task = """You are a Technical Analyst. Analyze stocks using technical indicators.
+The team goal is: {goal}
+
+Your analysis framework (weighted ensemble):
+- Trend following (25%): EMA-8/21/55, ADX
+- Momentum (25%): 1/3/6-month returns, volume momentum
+- Mean reversion (20%): Z-score, Bollinger Bands, RSI-14/28
+- Volatility (15%): Historical vol, ATR ratio, vol regime
+- Statistical arbitrage (15%): Hurst exponent, skewness, kurtosis
+
+OUTPUT INSTRUCTIONS (Hermes-native — do NOT use clawteam inbox send):
+After completing your analysis, store findings via mcp_gbrain_put_page with:
+  slug: {team_name}-technical
+  type: deal
+  title: "{team_name} Technical Analysis"
+  compiled_truth: Full analysis including:
+    - SIGNAL: bullish | bearish | neutral
+    - CONFIDENCE: 0-100
+    - TREND (EMA crossovers, ADX)
+    - MOMENTUM (returns, volume)
+    - MEAN REVERSION (RSI, Bollinger, Z-score)
+    - VOLATILITY REGIME
+    - REASONING (2-3 paragraphs)
+
+Then update task status: clawteam task update {team_name} [task-id] --status completed"""
+
+[[template.agents]]
+name = "fundamentals-analyst"
+type = "fundamentals-analyst"
+task = """You are a Fundamentals Analyst. Evaluate stocks using financial metrics.
+The team goal is: {goal}
+
+Your analysis framework (4 dimensions):
+- Profitability: ROE >15%, net margin >20%, operating margin >15%
+- Growth: revenue growth >10%, earnings growth >10%, book value growth >10%
+- Financial health: current ratio >1.5, D/E <0.5, FCF >80% of earnings
+- Valuation: P/E <25, P/B <3, P/S <5
+
+Signal = majority vote of 4 dimensions. Confidence = % of agreeing signals.
+
+OUTPUT INSTRUCTIONS (Hermes-native — do NOT use clawteam inbox send):
+After completing your analysis, store findings via mcp_gbrain_put_page with:
+  slug: {team_name}-fundamentals
+  type: deal
+  title: "{team_name} Fundamentals Analysis"
+  compiled_truth: Full analysis including:
+    - SIGNAL: bullish | bearish | neutral
+    - CONFIDENCE: 0-100
+    - PROFITABILITY (ROE, margins)
+    - GROWTH (revenue, earnings, book value)
+    - FINANCIAL HEALTH (liquidity, leverage, FCF)
+    - VALUATION (P/E, P/B, P/S)
+    - REASONING (2-3 paragraphs)
+
+Then update task status: clawteam task update {team_name} [task-id] --status completed"""
+
+[[template.agents]]
+name = "sentiment-analyst"
+type = "sentiment-analyst"
+task = """You are a Sentiment Analyst. Analyze market sentiment for stocks.
+The team goal is: {goal}
+
+Your analysis framework:
+- News sentiment analysis (70% weight): classify recent news as bullish/bearish/neutral
+- Insider trading activity (30% weight): net insider buying = bullish, selling = bearish
+- Combined weighted signal with confidence score
+
+OUTPUT INSTRUCTIONS (Hermes-native — do NOT use clawteam inbox send):
+After completing your analysis, store findings via mcp_gbrain_put_page with:
+  slug: {team_name}-sentiment
+  type: deal
+  title: "{team_name} Sentiment Analysis"
+  compiled_truth: Full analysis including:
+    - SIGNAL: bullish | bearish | neutral
+    - CONFIDENCE: 0-100
+    - NEWS SENTIMENT (recent headlines classified)
+    - INSIDER ACTIVITY (net buy/sell)
+    - REASONING (2-3 paragraphs)
+
+Then update task status: clawteam task update {team_name} [task-id] --status completed"""
+
+[[template.agents]]
+name = "risk-manager"
+type = "risk-manager"
+task = """You are the Risk Manager. Consolidate all analyst signals and assess portfolio risk.
+The team goal is: {goal}
+
+Workflow (read-from-gbrain, not from inbox):
+1. Wait ~90 seconds for all 5 analysts to finish.
+2. Query the knowledge brain for analyst reports:
+   mcp_gbrain_query("{team_name} analyst")
+   Expected pages: {team_name}-buffett, {team_name}-growth, {team_name}-technical, {team_name}-fundamentals, {team_name}-sentiment
+3. Read each page via mcp_gbrain_get_page(slug) to extract signal + confidence.
+4. Consolidate signals:
+   - For each ticker, tally bullish/bearish/neutral votes with confidence-weighted scores.
+   - Compute volatility-adjusted position limits (low vol: 25% max, high vol: 10% max).
+   - Adjust for correlation with existing positions.
+5. Persist your risk report via mcp_gbrain_put_page with:
+     slug: {team_name}-risk
+     type: deal
+     title: "{team_name} Risk Report"
+     compiled_truth:
+       - CONSOLIDATED SIGNAL (per ticker)
+       - POSITION LIMITS
+       - CORRELATION ADJUSTMENTS
+       - KEY RISKS (2-3 bullets)
+       - ALL ANALYST VOTES (table)
+6. Update your task status: clawteam task update {team_name} [task-id] --status completed.
+
+If fewer than 5 analyst pages exist after 3 minutes, proceed with partial data and note which analysts are missing in the risk report."""
+
+[[template.tasks]]
+subject = "Coordinate analysis, read gbrain, and make final investment decisions"
+owner = "portfolio-manager"
+
+[[template.tasks]]
+subject = "Value investing analysis (Buffett style) — store in gbrain"
+owner = "buffett-analyst"
+
+[[template.tasks]]
+subject = "Growth/disruption analysis — store in gbrain"
+owner = "growth-analyst"
+
+[[template.tasks]]
+subject = "Technical indicators analysis — store in gbrain"
+owner = "technical-analyst"
+
+[[template.tasks]]
+subject = "Fundamentals and financial metrics analysis — store in gbrain"
+owner = "fundamentals-analyst"
+
+[[template.tasks]]
+subject = "News and insider sentiment analysis — store in gbrain"
+owner = "sentiment-analyst"
+
+[[template.tasks]]
+subject = "Consolidate signals from gbrain and assess portfolio risk"
+owner = "risk-manager"
diff --git a/skills/hermes/SKILL.md b/skills/hermes/SKILL.md
@@ -40,7 +40,10 @@ For `clawteam spawn` (manual mode):
 ## Launch a Template (Recommended Path)
 
 ```bash
-# Stock analysis (7 specialist analysts: portfolio-manager, buffett, growth, technical, fundamentals, sentiment, risk-manager)
+# Stock analysis (Hermes-native template using gbrain for output — RECOMMENDED for Hermes users)
+clawteam launch hedge-fund-hermes --team-name tesla --goal "Analyze TSLA" --force
+
+# Stock analysis (original OpenClaw-style template — relies on clawteam inbox send which Hermes workers often skip)
 clawteam launch hedge-fund --team-name tesla --goal "Analyze TSLA" --command hermes --force
 
 # Research paper summary
@@ -53,6 +56,31 @@ clawteam launch code-review --team-name review1 --goal "Review PR #42" --command
 clawteam launch strategy-room --team-name strat --goal "Q2 planning" --command hermes --force
 ```
 
+## Reading Output from `hedge-fund-hermes` (Preferred Path)
+
+The Hermes-native template stores each analyst's findings as a gbrain page with a predictable slug. After launch + ~2 minutes wait:
+
+```bash
+# Query all findings at once
+mcp_gbrain_query("<team-name> analyst")
+
+# Or read each analyst's page directly
+for analyst in buffett growth technical fundamentals sentiment risk; do
+  mcp_gbrain_get_page("<team-name>-$analyst")
+done
+
+# Portfolio manager stores the final synthesis here
+mcp_gbrain_get_page("<team-name>-final-report")
+```
+
+**Why this works better than `hedge-fund`**: Hermes workers reliably call `mcp_gbrain_put_page` because gbrain is their native MCP surface. They often skip the `clawteam inbox send` step in the OpenClaw-style template.
+
+Benefits:
+- Findings persist in Postgres across sessions
+- No tmux scrollback fragility
+- Next week's analysis can query prior reports
+- Cross-agent memory works natively
+
 ## CRITICAL: Timing Expectations (DO NOT cleanup early)
 
 Spawned Hermes workers need substantial time to produce output:
@@ -207,11 +235,43 @@ clawteam spawn -t research -n analyst --task "Research topic X. Query gbrain for
 | `clawteam board live <team>` | Auto-refreshing board |
 | `clawteam board serve --port 8080` | Web dashboard |
 
+## Known Failure Modes
+
+1. **risk-manager crash**: The risk-manager agent is prone to crashing mid-execution (`"Agent 'risk-manager' exited unexpectedly. Reset N task(s) to pending"`). This can leave the team's final synthesis incomplete. If this happens, read tmux scrollback for the portfolio-manager (leader) window first — it may have the final recommendation before the crash.
+
+2. **Inbox empty despite COMPLETED tasks**: Hermes workers write to tmux scrollback, not to the inbox system. The kanban board reflects task state but not content. Always capture tmux panes to read actual output.
+
+3. **Workers do not persist findings**: Unless explicitly instructed to use gbrain, agent outputs exist only in tmux scrollback and are lost when the session is cleaned up. Always include brain-instructions in analysis goals.
+
+## CRITICAL: Inbox Peek Is Unreliable — Use Tmux Capture
+
+`clawteam inbox peek` often returns EMPTY even when tasks show COMPLETED on the board. The actual agent outputs live in **tmux scrollback**, not in the inbox system.
+
+**Always use tmux pane capture to read agent reports:**
+
+```bash
+# List all tmux windows for the team
+tmux list-windows -t clawteam-<team-name> 2>/dev/null
+
+# Capture each window's scrollback (each window = one agent)
+for win in $(tmux list-windows -t clawteam-<team-name> -F '#{window_index}' 2>/dev/null); do
+  echo -e "\n\n=== Window $win ==="
+  tmux capture-pane -t clawteam-<team-name>:$win -p 2>&1 | tail -100
+done
+```
+
+**Also:** When launching a team for analysis work, include brain-instructions in the goal so agents store findings in gbrain — this provides a persistent fallback when tmux sessions are lost:
+
+```bash
+clawteam launch hedge-fund --team-name <name> --goal "Analyze AAPL. IMPORTANT: After completing your analysis, store key findings as brain pages using mcp_gbrain_put_page so findings persist even if the tmux session is lost." --command hermes --force
+```
+
 ## Anti-Patterns (Do Not Do)
 
 - Don't use `delegate_task` when the user asks for multi-agent/swarm/team analysis. Use clawteam.
 - Don't pass `--team` (the flag is `--team-name`).
 - Don't pass `--command hermes` to `spawn` (it's a positional arg at the end). That flag only works on `launch`.
 - Don't forget `--command hermes` on `launch` — templates default to openclaw.
 - Don't assume templates will use Hermes by default. They don't. Always pass `--command hermes`.
-- Don't clean up the team before reading final reports from inboxes.
+- Don't clean up the team before reading final reports — BUT `inbox peek` alone is unreliable; always ALSO capture tmux panes.
+- Don't rely on `clawteam inbox peek` alone — tasks can be COMPLETED on the board while inbox messages are empty. Tmux scrollback capture is the source of truth.
