trade-debrief

Purpose

Synthesize session data into a coaching-grade debrief with the bundled prompt library. Descriptive TCA lives in trade-journal. This skill is prescriptive. It surfaces mistakes and patterns, and proposes rules and measurable improvements.

Environment routing

Demo (simulation) and live are two separate MCP servers. Account names are unique to one server. Once a workflow resolves an account on a server, every downstream call must go through that same server. This includes my_portfolio, market_snapshot, place_order, create_alert, history tools, etc. Cross-routing fails or hits the wrong environment.

MCP tools used

Tool names below are bare. The NinjaTrader MCP server provides them. Your client adds its own prefix. See AGENTS.md at the repo root.

• timeline_report — per-session event timeline (single tool post-consolidation: account, date, optional endDate for a range)
• performance_summary — arrays of {name, value} stat pairs under extra.allTradeStats[] / extra.profitTradeStats[] / extra.lossTradeStats[] (e.g. {name: "Total P/L", value: "$765.00"}), plus extra.tradeLog[] and extra.hasTrades
• fill_history — fill-level timestamps
• position_history — closed-position summaries
• market_history — bar OHLC for MFE/MAE
• search_contracts — contract metadata (productName, tickSize, valuePerPoint)
• economic_calendar — event context (via event-watch skill, not this one directly)
• user_profile — discover available account names when the account is unknown

Skills consumed

• trade-journal's streaks.py — FIFO round-trip pairing over fill_history, to produce the trades[] list. Reuse it. Do not duplicate it.
• market-context — session VWAP, ATR, and regime narrative per symbol, for the market_context block of each trade.

Workflow

1. Classify the intent → pick the prompt

references/prompt-selection.md has the full routing table. Rough mapping:

2. Gather MCP inputs (in parallel)

Account resolution. All MCP calls below require account=<name>. If the account name is unknown, call user_profile() first. Its accounts[].name field lists every account on this MCP server.

timeline_report(account=..., date=<YYYY-MM-DD>, endDate=<YYYY-MM-DD>)
performance_summary(account=..., startDate=<YYYY-MM-DD>, endDate=<YYYY-MM-DD>)
fill_history(account=..., startDate=<YYYY-MM-DD>, endDate=<YYYY-MM-DD>)
position_history(account=..., startDate=<YYYY-MM-DD>, endDate=<YYYY-MM-DD>)
search_contracts(text=<each symbol>, includeFamilySiblings=false)

Date params are date-level. They use YYYY-MM-DD or a natural term like "last week", never a timestamp.

Then for each trade, fetch a bar window around its lifetime:

market_history(symbol=<sym>, barType="Minute",
  barSize=1, from=<entry_ts - 1min>, to=<exit_ts + 1min>)

(from and to are full ISO-8601 timestamps. from requires to.)

Keep per-trade bar fetches bounded. Unless the user asks for per-minute pattern analysis, do not fetch 1-min bars for a full day (expensive).

3. Pair fills into trades

Run trade-journal's streaks.py over the fill list. Use its output round-trips as the trade list. Attach each trade's corresponding bar window to it.

4. Compute derived metrics

Save the trades-with-bars payload to a file. Pass its path with --file. Never re-type or inline a large JSON payload in the command.

python3 scripts/compute_derived.py --file trades_with_bars.json

Per trade, it produces realized_points/dollars/R, mfe_points/dollars/R, mae_points/dollars/R, giveback_from_mfe_pct, and realized_vs_mfe_pct.

Some fields are always null: scale_out_count, stop_modified_count, and minute-by-minute timeline volume features. The MCP cannot source them. See references/report-schema.md for the full provenance map.

5. Assemble the report YAML

python3 scripts/assemble_report.py --file combined_input.json

It takes account_id, trade_date, timeline_report, performance_summary, trades (post-compute_derived), a contracts map, and an optional market_context map. It emits the YAML shape that the prompts expect. See references/report-schema.md for details.

6. Run the prompt (internally — Claude IS the LLM)

1. Read the chosen prompt file (e.g., references/prompts/07_daily_debrief.md).
2. Append the assembled YAML below the prompt's final ## Trade Timeline Report header.
3. Follow the prompt's STRICT RULES. These require fact-only content, YAML-path citations, and the mode-specific output structure.
4. When a rule says "cite exact YAML path", cite only fields that have a value, not fields that are null. references/report-schema.md lists which fields are null today.

Do not route this through an external LLM API. Claude Code IS the LLM. Just read the prompt and apply its constraints.

7. Echo a summary + point at the raw report

End with a 1-line pointer: "Full assembled report YAML available on request; source tools: timeline_report, performance_summary, fill_history, market_history, search_contracts." This gives the user transparency about where the numbers came from.

Output constraints (from the prompt library)

• Fact-only with citations. Every claim references a YAML path. No speculation.
• Dollar-first, R-second. Lead with "$875 left on the table", then "0.7R giveback" for context.
• Hypotheses, labeled explicitly. If a claim needs data the report does not have yet, mark it "hypothesis" and state the evidence it needs.
• No invented indicators. Only use fields present in the assembled report. Volume and POC features are null today. Do not cite them.

Output idioms

Prompt 07 example (daily_debrief), roughly 200 words:

Performance Summary Net +$765 / +1.2R across 2 trades on the E-Mini S&P 500 and E-Mini NASDAQ-100 (summary.total_pnl_dollars = 765.00). Win rate 100% (2 of 2), but MFE capture lagged on both.

Primary Pattern Systematic MFE giveback. ES gave back 47.5% from peak (trades[0].derived.giveback_from_mfe_pct = 47.5); NQ gave back 40% (trades[1].derived.giveback_from_mfe_pct = 40.0). Combined $600 left on the table from peak.

Top 3 Improvements

1. Rule: IF mfe_R >= 2.0 THEN scale_out_50pct
   • Evidence: trades[0].mfe_R = 2.5; exit at 1.3R (realized = $525, peak = $1,000, giveback = $475)
   • Expected impact: +$475 / +1.2R improvement on ES
2. Rule: IF mfe_R >= 0.5 AND time_in_trade > 10min THEN move_stop_to_breakeven
   • Trades affected: both
3. (Hypothesis; needs scale-out action data not yet in report.)

Disambiguation

• vs trade-journal: journal reports descriptive TCA and streaks — what happened. Debrief mines prescriptive rules — what to change.
• vs trade-replay: replay does one trade's deep dive, with counterfactual exits. Debrief synthesizes cross-trade patterns at the session level.
• vs risk-coach: coach handles live behavior, such as tilt, revenge trading, and overtrading. Debrief handles retrospective mechanics, such as giveback, scale discipline, and stop hygiene.
• vs market-context: debrief consumes it as input for the market_context block of each trade. It is not a substitute for debrief.

Explicit non-goals

• No external LLM round-trip. Claude Code IS the LLM.
• No modification of the bundled prompts. This skill uses them verbatim.
• No fabricated fields. timeline[n].participation_rate_vs_session, poc_shift, scale_out_count, and stop_modified_count are null today. Do not guess their values.
• No daily auto-push. This is a reactive skill. It runs only when the user asks.

Resource layout

• scripts/compute_derived.py — per-trade MFE/MAE + giveback metrics from bar scans. Shared pattern with trade-replay.
• scripts/assemble_report.py — combines compute-derived trades + contracts + performance summary + market context → the assembled report YAML.
• scripts/fixtures/trades_with_bars_synthetic.json — 2-trade multi-symbol fixture (ES long winner, NQ short winner with significant MFE giveback) for smoke tests.
• references/prompts/ — 7 prompt template files: 01_mistake_identification.md, 02_pattern_extraction.md, 03_trade_scoring.md, 06_correction_rules.md, 07_daily_debrief.md, 08_pattern_clustering.md, 09_expectancy_analysis.md. Load only the one file that matches the chosen prompt.
• references/prompt-selection.md — intent-to-prompt routing + combinations + how-to-run inside Claude Code. Load it when you classify the intent and choose a prompt.
• references/report-schema.md — field provenance map (MCP-sourced vs derived vs null-today), per-trade block reference, summary block reference. Load it when you check which fields the chosen prompt needs.