Quandora
This package contains eight user-facing Skills and one authenticated production MCP connection. Use it for production research and Simulation workflows. Real Live evidence is read-only through MCP.
Skills and responsibility
| Skill | Owns | Routes elsewhere |
|---|---|---|
account-connection | Explicit login, reauthorization, account switching, current identity and problem reports to the Quandora team | Research and Trading operations |
factor-mining | Hypothesis, Factor source/validation, authorized backtests, history and export | Deep Factor diagnosis |
factor-analysis | Existing Factor evidence, validity and controlled experiment design | Authoring and submission |
strategy-building | Eligible Factors, immutable Strategy definitions and requested backtests | Deep diagnosis and multi-Strategy composition |
strategy-analysis | Existing Strategy results and reusable logic cards | Definition changes and backtest submission |
portfolio-building | Candidate selection, initial capital weights, immutable Portfolios, backtests, discovery/favorites and sharing | Deep Portfolio diagnosis and Trading management |
portfolio-analysis | Actual-member Portfolio backtest comparisons and incremental-value diagnosis | Selection/creation, sharing and forward evidence |
portfolio-trading-analysis | Simulation account/run management and Simulation/real Live forward analysis | Backtest authoring and all real Live writes |
Alongside account management, six Skills cover backtests and one covers forward tests. The research intent and evidence contract defines request modes, capability-aware analysis and handoffs. Design-only requests do not create resources. Exact supplied members/weights are not silently optimized. Every Strategy or Portfolio the agent designs or recommends uses equal Factor weights or equal initial capital; non-equal weights come only from the user and are compared with the equal-weight version. Analysis suggestions do not authorize experiments; already-authorized work continues through the owning builder without repeated confirmation.
The research notebook carries research across conversations:
each research task reads get_research_context first, each decision carries a research_note
(relation, parent, hypothesis, expected effect), each interpreted result is recorded with
record_research_conclusion, and wait_for_runs replaces polling. Factor Mining submits new
research with submit_factor_research and tests an existing Factor version at another horizon with
run_factor_version. Notes belong to versions; a note about one run records it in run_id.
Each Skill uses only relevant tools and preserves opaque identities and pagination. Portfolio research accepts one to five distinct immutable Strategies with equal or explicit positive initial capital weights. Capital allocation is not continuous rebalancing. Definitions and Runs remain separate, and actual source StrategyRuns and aligned windows govern comparisons. Missing advanced joint-risk evidence limits conclusions rather than generating invented correlations or contributions.
Portfolio backtests use submit_portfolio_backtest on the exact immutable definition, following
its current exchange/schema contract. FM prepares missing or stale sources; PB returns persisted
state and authoritative research. PB runs automatic Monitoring only for a shared
Portfolio; get_portfolio_backtest returns its association in run.automatic_paper, null when
the Portfolio has never been shared. A pending or absent association does not authorize an extra start. Portfolio Trading Analysis observes that exact association and owns explicitly
requested Simulation operations. Legacy simulated live aliases do not describe real Live accounts.
Real Live uses separate account-oriented read tools; its writes remain frontend-only.
Connection
Supported hosts load the plugin's quandora remote HTTP MCP declaration. Antigravity reads
it from mcp_config.json with plugin.json as its manifest; WorkBuddy reads the same
plugin.json as an Agent Plugins manifest and takes the server from the Agent Plugins mcp.json. Complete the
host-native browser authorization flow for initial setup; reload tools only when the host requires it.
Do not create a local MCP server or paste credentials.
For incidental failures, the Agent lets the host refresh the connection automatically first, then uses supported reconnect or tool discovery and resumes the task. User interaction is limited to unavoidable sign-in or consent after terminal authorization failure, a host consent requirement, or an explicit user request to connect, reauthorize or switch accounts. Use Quandora Account for these requests; verify the active identity after login. Do not reinstall a working plugin, display token lifetimes, or ask users to run refresh commands. Check the release once per session using actual installed package metadata, then follow the host update guide within existing update authority. See Connection and Security for bounded retries and host-native fallbacks, and Failed mutations for what to do after a failed or unanswered mutating call. Never request secrets or bypass MCP with internal HTTP.
Claude Desktop Code authorization bridge
The plugin intentionally ships these two host-support assets:
scripts/claude-mcp-login-macos.sh
scripts/claude-mcp-login-windows.ps1
They support the authorization step of the one-click Claude Desktop Code Agent installation flow,
where redirected command input/output cannot host the official interactive claude mcp login
prompt. Both launch only plugin:quandora:quandora in a native interactive
terminal, keep OAuth output out of status files, and require manual browser identity and consent.
They are not general shell utilities, local MCP servers, or credential helpers. Interactive Claude
Code users should use the direct host command instead.
Search and pagination
Factor discovery actions accept supported query and filters fields. Canonical Strategy and
PortfolioFamily/Run lists use parent identities and pagination, without query/filter arguments. Continuation requests
must preserve the same query, filters, archive mode, and page size while copying the returned
opaque page token byte-for-byte.
Result files
All complete Result Bundle exports use download_result_bundle(kind, id). Follow
Unified Result Bundle Download for the canonical
selectors, retryable HTTP links, resume and integrity verification. Large ZIPs do not use
MCP base64 chunks; the host needs an HTTP/file download capability.
When local file writes are supported, verified Factor and Strategy bundles use these default workspace-relative destinations:
Quandora result/factor/<factor_slug>.zip
Quandora result/strategy/<strategy_slug>.zip
The validated Factor source's exact non-empty lowercase snake_case FACTOR_TYPE determines the
factor slug; it never silently falls back to a generic factor slug. The user-facing submitted
Strategy name determines the Strategy slug. The verified archive remains intact, and returned
metadata and manifests are authoritative for readiness, contents, omissions, and integrity.
Safety boundary
- OAuth and credentials are host-managed and must never be requested, printed, copied, or stored.
- Choose research settings from context and returned candidates, record why in the decision note, and carry them through Skill handoffs. Simulation, sharing and account operations remain request-only.
- An ambiguous response is not proof of failure and does not authorize a changed replacement.
- Simulation lifecycle and real Live read-only inspection are separate; all Live changes are frontend-only.
- User-facing answers follow the user's language while preserving tool names and schema fields exactly.
Community discovery supports authorized contextual research and separate Portfolio favorites. Sharing supports public or specified-user grants with required UTC dates, active-grant extensions and concurrency-checked intro edits. Current grant/channel constraints are server-owned; shared child visibility does not create Mine ownership or execution eligibility. Factor/Strategy stars remain independent of LibraryFolder placement.
Simulation creation/start uses a recoverable operation with the same request identity on retries. PB owns capital, quotas, grants and lifecycle rules. Net-ledger evidence owns account finance; legacy member observations must not be summed into an authoritative net account. Real Live reads preserve observation freshness and the bounded recent scope of orders/fills.
Named-object discovery uses search_quandora_objects; exact identities, current/historical source
filters and authorization remain authoritative. Strategy backtest intent uses a fresh
client_request_id per requested Run and the same identity only for unchanged retries.
Production host installation support
Production also retains the reviewed WorkBuddy support scripts:
scripts/workbuddy-cn-bootstrap-macos.shscripts/workbuddy-cn-install-macos.jsscripts/workbuddy-mcp-login-macos.shscripts/workbuddy-mcp-login-windows.ps1
Use the production installation guides for their supported scope and integrity checks.