* feat: add macro scanner feedback loop and lessons memory
- Implements `LessonStore` to persist JSON screening lessons
- Adds `selection_reflector.py` to fetch performance and news, and generate LLM lessons
- Adds `memory_loader.py` to filter negative lessons into `FinancialSituationMemory`
- Integrates a rejection gate in `candidate_prioritizer.py` based on past negative lessons
- Adds `reflect` command to CLI
Co-authored-by: aguzererler <6199053+aguzererler@users.noreply.github.com>
* feat: update macro scanner feedback loop for dual-lesson output (Trend DNA)
- Update `selection_reflector.py` to calculate exact terminal returns, `mfe_pct`, `mae_pct`, and `days_to_peak`.
- Update LLM prompt to generate distinct `screening_advice` and `exit_advice`.
- Update `lesson_store` tests to reflect new schema.
- Update `memory_loader.py` to use `screening_advice` for negative selection filtering.
- Update `micro_summary_agent.py` to inject `exit_advice` into PM context for current holdings.
- Update `cli/main.py` default horizons to `30,90` and print dual-advice columns.
Co-authored-by: aguzererler <6199053+aguzererler@users.noreply.github.com>
* fix: resolve bugs in macro scanner feedback loop
- Address Key mismatch (stock_return_pct vs terminal_return_pct)
- Fix missing persistence of mfe_pct and mae_pct
- Use create_report_store() instead of raw ReportStore() in load_scan_candidates
- Clean up unused imports in fetch_news_summary
- Ensure default horizons match code in cli description
- Create isolated `_local_safe_pct` to remove cross-module dependency
Co-authored-by: aguzererler <6199053+aguzererler@users.noreply.github.com>
* fix: address PR 124 feedback on macro scanner memory feedback loop
- Use `l.get('screening_advice')` gracefully in `memory_loader` to prevent KeyErrors.
- Properly instantiate `selection_memory` inside the graph in `portfolio_setup.py` and pass it to the prioriziation rejection gate.
Co-authored-by: aguzererler <6199053+aguzererler@users.noreply.github.com>
---------
Co-authored-by: google-labs-jules[bot] <161369871+google-labs-jules[bot]@users.noreply.github.com>
Co-authored-by: aguzererler <6199053+aguzererler@users.noreply.github.com>
Replace run_id with flow_id as the primary grouping concept (one flow =
one user analysis intent spanning scan + pipeline + portfolio). Reports
are now written as {timestamp}_{name}.json so load methods always return
the latest version by lexicographic sort, eliminating the latest.json
pointer pattern for new flows.
Key changes:
- report_paths.py: add generate_flow_id(), ts_now() (ms precision),
flow_id kwarg on all path helpers; keep run_id / pointer helpers for
backward compatibility
- ReportStore: dual-mode save/load — flow_id uses timestamped layout,
run_id uses legacy runs/{id}/ layout with latest.json
- MongoReportStore: add flow_id field and index; run_id stays for compat
- DualReportStore: expose flow_id property
- store_factory: accept flow_id as primary param, run_id as alias
- runs.py / langgraph_engine.py: generate and thread flow_id through all
trigger endpoints and run methods
- Tests: add flow_id coverage for all layers; 905 tests pass
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: introduce flow_id with timestamp-based report versioning
Replace run_id with flow_id as the primary grouping concept (one flow =
one user analysis intent spanning scan + pipeline + portfolio). Reports
are now written as {timestamp}_{name}.json so load methods always return
the latest version by lexicographic sort, eliminating the latest.json
pointer pattern for new flows.
Key changes:
- report_paths.py: add generate_flow_id(), ts_now() (ms precision),
flow_id kwarg on all path helpers; keep run_id / pointer helpers for
backward compatibility
- ReportStore: dual-mode save/load — flow_id uses timestamped layout,
run_id uses legacy runs/{id}/ layout with latest.json
- MongoReportStore: add flow_id field and index; run_id stays for compat
- DualReportStore: expose flow_id property
- store_factory: accept flow_id as primary param, run_id as alias
- runs.py / langgraph_engine.py: generate and thread flow_id through all
trigger endpoints and run methods
- Tests: add flow_id coverage for all layers; 905 tests pass
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: PM brain upgrade — macro/micro summary agents, memory split, forensic dashboard
Replaces the PM's raw-JSON context (~6,800 tokens on deep_think) with a
MAP-REDUCE compression layer using two parallel mid_think summary agents,
achieving ~70% cost reduction at the PM tier.
Architecture:
- MacroMemory: new regime-level memory class (MongoDB/JSON, separate from
per-ticker reflexion memory) with record_macro_state/build_macro_context
- ReflexionMemory: extended with collection_name param to isolate
micro_reflexion from the pipeline reflexion collection (with distinct
local JSON fallback path to prevent file collision)
- Macro_Summary_Agent (mid_think): compresses scan_summary into a 1-page
regime brief with memory injection; sentinel guard prevents LLM call on
empty/error scan data ("NO DATA AVAILABLE - ABORT MACRO")
- Micro_Summary_Agent (mid_think): compresses holding_reviews + candidates
into a markdown table brief with per-ticker memory injection
- Portfolio graph: parallel fan-out (prioritize_candidates → macro_summary
‖ micro_summary → make_pm_decision) using _last_value reducers for safe
concurrent state writes (ADR-005 pattern)
- PM refactor: Pydantic PMDecisionSchema enforces Forensic Execution
Dashboard output (macro_regime, forensic_report, per-trade
macro_alignment/memory_note/position_sizing_logic); with_structured_output
as primary path, extract_json fallback for non-conforming providers
- PM sentinel handling: "NO DATA AVAILABLE" in macro_brief substituted
with actionable conservative guidance before LLM sees it
62 new unit tests across 4 test files covering all new components.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: address code review — relaxed error guard, ticker_analyses, PM memory wiring
1. macro_summary_agent: relaxed error guard to only abort when scan_summary's
sole key is "error" (partial failures with real data are now processed)
2. micro_summary_agent: now reads ticker_analyses from state and enriches
the per-ticker table with trading graph analysis data
3. portfolio_graph: wires macro_memory and micro_memory to PM factory call
4. test_empty_state: updated test for new partial-failure behavior
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: introduce flow_id with timestamp-based report versioning
Replace run_id with flow_id as the primary grouping concept (one flow =
one user analysis intent spanning scan + pipeline + portfolio). Reports
are now written as {timestamp}_{name}.json so load methods always return
the latest version by lexicographic sort, eliminating the latest.json
pointer pattern for new flows.
Key changes:
- report_paths.py: add generate_flow_id(), ts_now() (ms precision),
flow_id kwarg on all path helpers; keep run_id / pointer helpers for
backward compatibility
- ReportStore: dual-mode save/load — flow_id uses timestamped layout,
run_id uses legacy runs/{id}/ layout with latest.json
- MongoReportStore: add flow_id field and index; run_id stays for compat
- DualReportStore: expose flow_id property
- store_factory: accept flow_id as primary param, run_id as alias
- runs.py / langgraph_engine.py: generate and thread flow_id through all
trigger endpoints and run methods
- Tests: add flow_id coverage for all layers; 905 tests pass
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: finalise storage layout, run history loading & phase-level re-run
Storage / persistence
- flow_id (8-char hex) replaces run_id as the disk storage key; all
sub-phases of one auto run share the same flow_id directory
- Startup hydration: hydrate_runs_from_disk() rebuilds in-memory run
store from run_meta.json on server restart (events lazy-loaded)
WebSocket / run history fixes
- Lazy-load events from run_events.jsonl on first WS connect; fixes
blank terminal when clicking a historical run after restart
- Orphaned "running" runs (server restarted mid-run) auto-detected and
marked "failed" with partial events replayed correctly
Phase re-run fixes
- Analysts checkpoint: use any() instead of all() — Social Analyst is
optional; all() silently blocked checkpoint saves in typical runs
- Checkpoint lookup: pass original flow_id through rerun_params so
_date_root() resolves to the correct flow_id subdirectory
- Selective event filtering on re-run: preserves scan nodes and other
tickers; only removes stale events for the re-run phase+ticker
- Frontend graph now shows full auto-flow context during phase re-runs
Documentation
- ADR 018: canonical reference for storage layout, event schema,
WebSocket streaming flows, checkpoint structure, MongoDB vs local
- ADR 013 updated: reflects background-task + lazy-loading evolution
- ADR 015 marked superseded by ADR 018
- CLAUDE.md: AgentOS storage section + 4 new critical patterns
- CURRENT_STATE.md updated
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: introduce flow_id with timestamp-based report versioning
Replace run_id with flow_id as the primary grouping concept (one flow =
one user analysis intent spanning scan + pipeline + portfolio). Reports
are now written as {timestamp}_{name}.json so load methods always return
the latest version by lexicographic sort, eliminating the latest.json
pointer pattern for new flows.
Key changes:
- report_paths.py: add generate_flow_id(), ts_now() (ms precision),
flow_id kwarg on all path helpers; keep run_id / pointer helpers for
backward compatibility
- ReportStore: dual-mode save/load — flow_id uses timestamped layout,
run_id uses legacy runs/{id}/ layout with latest.json
- MongoReportStore: add flow_id field and index; run_id stays for compat
- DualReportStore: expose flow_id property
- store_factory: accept flow_id as primary param, run_id as alias
- runs.py / langgraph_engine.py: generate and thread flow_id through all
trigger endpoints and run methods
- Tests: add flow_id coverage for all layers; 905 tests pass
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: load flow_id in FE to resume runs and fix max_tickers cap on continuation
- Add flow_id to RunParams interface and initial state
- loadRun() now restores flow_id + max_auto_tickers from history so the next
run continues in the same flow directory (Phase 1 scan skipped, already-done
tickers skipped via skip-if-exists logic)
- startRun() spreads flow_id into the request body when set, letting the backend
reuse the existing flow directory instead of generating a fresh flow_id
- After each run, params.flow_id is updated from the response so subsequent
runs automatically continue from the same flow
- max_auto_tickers restored from run.params.max_tickers ensures the ticker cap
matches the original run; scan_tickers[:max_t] on the backend then limits
the Phase 2 queue to the user's setting even when the existing scan has more
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(mongo): fast-fail timeout + lazy ensure_indexes to avoid 30s block on fallback
MongoClient previously used pymongo's 30-second serverSelectionTimeoutMS default,
causing store_factory to hang for 30s before falling back to the filesystem when
Atlas is unreachable. Also, ensure_indexes() was called eagerly in __init__,
making every store construction attempt block on a live network call.
- Set serverSelectionTimeoutMS=5_000 so fallback is triggered in ≤5s
- Move ensure_indexes() call out of __init__ — indexes are now created lazily
on the first _save() call via a guarded self._indexes_ensured flag
- ensure_indexes() is still idempotent and safe to call explicitly in tests
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(store): wrap all DualReportStore mongo calls in _try_mongo() for graceful degradation
Any MongoDB exception (SSL error, ServerSelectionTimeout, auth failure) was
propagating uncaught through DualReportStore and crashing the run. Reads
would return an error instead of falling back to local, and writes would
abort mid-run without saving anything.
Introduce a single _try_mongo(fn, default) helper that:
- Executes the Mongo callable
- Catches *any* exception, logs it as WARNING with type + message
- Returns the default value so the caller continues with local-only data
Pattern per method:
writes → try mongo (fire-and-forget); always return local result
reads → try mongo first; fall back to local on None or exception
lists → try mongo; fall back to local on empty/None
Runs now complete successfully even when Atlas is unreachable or returns SSL
errors. MongoDB sync resumes automatically once connectivity is restored.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(observability): non-blocking MongoDB inserts + 5s timeout in RunLogger
Every LLM and tool callback called _append() which synchronously called
insert_one() against MongoDB. When Atlas was unreachable this blocked the
entire LangGraph run for pymongo's 30-second default timeout per event,
effectively serializing all agent work behind MongoDB retries.
Two fixes:
1. serverSelectionTimeoutMS=5_000 on the RunLogger's MongoClient — consistent
with the same fix applied to MongoReportStore.
2. MongoDB inserts are now fire-and-forget via daemon threads — _append() spawns
a Thread(target=_insert, daemon=True) and returns immediately. LLM callbacks
and tool events are never delayed by MongoDB connectivity issues.
Failures are still reported via WARNING log from the background thread.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* revert(observability): restore synchronous MongoDB inserts in RunLogger
Root cause was an IP whitelist issue on Atlas causing SSL failures, not
insert volume. The background-thread approach added unnecessary complexity.
The 5s serverSelectionTimeoutMS is retained as a defensive safeguard.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Feature 1 - Configurable max_auto_tickers:
- Add max_auto_tickers config key (default 10) with TRADINGAGENTS_MAX_AUTO_TICKERS env override
- Macro synthesis agent accepts max_scan_tickers param, injects exact count into LLM prompt
- ScannerGraph passes config value to create_macro_synthesis()
- Backend engine applies safety cap on scan candidates (portfolio holdings always included)
- Frontend adds Max Tickers number input in params panel, sends max_tickers in auto run body
Feature 2 - Run persistence + phase-level node re-run:
- 2A: ReportStore + MongoReportStore gain save/load_run_meta, save/load_run_events,
list_run_metas methods; runs.py persists to disk in finally block; startup hydration
restores historical runs; lazy event loading on GET /{run_id}
- 2B: Analysts + trader checkpoint save/load methods in both stores; engine saves
checkpoints after pipeline completion alongside complete_report.json
- 2C: GraphSetup gains build_debate_subgraph() and build_risk_subgraph() for partial
re-runs; TradingAgentsGraph exposes debate_graph/risk_graph as lazy properties;
NODE_TO_PHASE mapping + run_pipeline_from_phase() engine method;
POST /api/run/rerun-node endpoint with _append_and_store helper
- 2D: Frontend history popover (loads GET /api/run/, sorts by created_at, click to load);
triggerNodeRerun() calls rerun-node endpoint; handleNodeRerun uses phase-level
re-run when active run is loaded
All 890 existing tests pass (10 skipped).
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
- ReportStore.clear_portfolio_stage(date, portfolio_id): deletes pm_decision
(.json + .md) and execution_result files for a given date/portfolio
- DELETE /api/run/portfolio-stage endpoint: calls clear_portfolio_stage
and returns list of deleted files
- Dashboard: 'Reset Decision' button calls the endpoint, then user can
run Auto to re-run Phase 3 from scratch while skipping Phase 1 & 2
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
TRADINGAGENTS_REPORTS_DIR now controls where all reports land (scans,
analysis, portfolio artifacts). Both report_paths.REPORTS_ROOT and
ReportStore.data_dir read from the same env var so the entire
reports/daily/{date}/... tree is rooted at one configurable location.
PORTFOLIO_DATA_DIR still works as a portfolio-specific override.
Falls back to "reports" (relative to CWD) when neither is set.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Extend run_portfolio to save Holding Reviews, Risk Metrics, PM Decision,
and Execution Result from final state
- Add run_trade_execution method for resuming trade execution from a saved
PM decision without re-running the full portfolio graph
- Update run_auto skip logic to check for execution result (not just decision)
and resume from saved decision when available
- Gitignore uv.lock and untrack it from version control
Co-Authored-By: Oz <oz-agent@warp.dev>
Replaces the O(N) database operations in the `TradeExecutor`'s
`execute_decisions` SELL loop with a single `batch_remove_holdings`
call to the repository. The new repository method calculates updates
in memory, resolves duplicate operations on the same ticker, and issues
the updates via newly implemented `psycopg2.extras.execute_batch`
routines on the `SupabaseClient`.
Co-authored-by: aguzererler <6199053+aguzererler@users.noreply.github.com>
Optimize the _percentile calculation to use heapq.nsmallest or heapq.nlargest when requesting small extreme percentiles (like 5% VaR) from large lists, falling back to sorted() only when necessary. This avoids fully sorting the entire array.
Co-authored-by: aguzererler <6199053+aguzererler@users.noreply.github.com>
This commit optimizes `PortfolioSnapshot` instantiation when loaded from the DB or JSON dictionaries by removing the immediate `json.loads(holdings_snapshot)` parsing inside the `from_dict` constructor. Instead, it overrides `__getattribute__` to implement a lazy-loading pattern, where the `holdings_snapshot` string is only parsed into a Python dictionary the very first time the field is accessed.
This optimization ensures that parsing large JSON payloads doesn't block the instantiation of snapshots, which is particularly beneficial in workflows that load a large historical series of snapshots but never access their `holdings_snapshot` field.
Co-authored-by: aguzererler <6199053+aguzererler@users.noreply.github.com>
- tradingagents/portfolio/risk_metrics.py: pure-Python computation of
Sharpe, Sortino, VaR, max drawdown, beta, sector concentration from
PortfolioSnapshot NAV history — no LLM, no external dependencies
- tradingagents/portfolio/__init__.py: export compute_risk_metrics
- tradingagents/agents/utils/portfolio_tools.py: 4 LangChain tools
wrapping Holding.enrich, Portfolio.enrich, ReportStore APIs, and
compute_risk_metrics so agents can access portfolio data without
reimplementing computations
- tests/portfolio/test_risk_metrics.py: 48 tests for risk metrics
- tests/unit/test_portfolio_tools.py: 19 tests for portfolio tools
- tests/portfolio/test_repository.py: fix pre-existing import error
Co-authored-by: aguzererler <6199053+aguzererler@users.noreply.github.com>
Replace supabase-py stubs with working psycopg2 implementation using
Supabase pooler connection string. Implement full business logic in
repository (avg cost basis, cash accounting, trade recording, snapshots).
Add 12 unit tests + 4 integration tests (51 total portfolio tests pass).
Fix cash_pct bug in models.py, update docs for psycopg2 + pooler pattern.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ahmet guzererler
Copilot
google-labs-jules[bot]