Privacy and data handling

This is the operational policy for repository workflows, not a promise that all inputs stay on your computer. Review the selected skill, tool permissions and provider terms before using personal trading data. Keep account identifiers, holdings, tax records, API keys and unredacted conversations out of public examples, Issues, PRs and model prompts unless the disclosure is intentional. Use fictional examples, remove direct identifiers and review derived text too. Truncation is not anonymization. There is no general automatic PII redactor. Before sharing reports, logs or postmortems, replace names, account numbers, order IDs and broker identifiers with fictional labels; remove free-text identifiers and check filenames and metadata too. Share only the fields needed for review.

Local storage and Git boundaries

Location Likely contents Handling
state/theses, state/journal Thesis rationale, positions, decisions, trade outcomes Private; back up before changes
reports/, reviews/ Portfolio analysis, generated reports and review findings Private by default; inspect before sharing
logs/, .cache/, .staging/ Diagnostics, provider responses, automation state and intermediate output May contain sensitive payloads or error excerpts
.env, .env.*, .envrc*, .mcp.json*, .aws/, .ssh/ Credentials and local service configuration OS access controls; never publish
.claude/, ~/.claude/projects/ Local agent configuration and session transcripts Session transcripts can contain user/tool data; see mining below
private-backups/ Recovery copies and hash manifests Restrict OS access; encrypt off-device copies

The root .gitignore excludes these repository-local canonical locations. The active-index guard check_private_data_paths.py rejects root state, reports, logs, reviews, .cache, .claude, .staging, private-backups, .aws, .ssh, dotenv/.envrc variants and .mcp.json variants, including forced adds and case variants. Only exact, regular, empty indexed state/journal/.gitkeep and state/theses/.gitkeep placeholders are allowed. It runs in pre-commit and in the existing root-script CI tests. Install the repository hooks before contributing; CI checks cannot retract an already pushed secret.

This guard checks names and Git object metadata, not file contents. Reviewed fictional fixtures under examples/, docs/ or tests are allowed; real private data placed there is not detected by this guard. Ignore rules do not encrypt, delete tracked history, prevent uploads, or protect external absolute paths. The separate detect-secrets hook excludes several extensions, including Markdown, JSON and CSV; it is not a complete secret or PII control. If a credential is exposed, revoke/rotate it with its provider and assess history/artifacts separately.

External data flow

Integration Data sent or exposed Data returned / boundary
FMP client Ticker, dates and filters; API key in header or query depending on endpoint Market data; cache/output can preserve raw responses
FINVIZ Screening filters and configured authentication key Screen results
Alpaca tools / full portfolio workflow Broker credentials authorize account/position requests; returned holdings may enter agent context Broker/account data; local CSV comparison/normalization commands are offline, but connecting other tools changes the boundary
WebSearch / browser tools Search terms, tickers and any user text included in queries External pages/results; do not include account details in queries
Claude / configured model provider Conversation, pasted files and tool results selected for context; mining/review prompts described below Model output; use provider retention/access settings appropriate to the data
GitHub automation Selected generated candidates, skill changes, PR text and committed files May become public; inspect derived output and artifacts before publication

Transport security depends on the configured endpoint/tool; use HTTPS provider endpoints and avoid logging keys or pasting them into prompts. This repository does not enforce redaction across every network path. Skill instructions about broker usage are not a universal technical order-execution barrier; inspect actual tool capabilities and permissions. Offline scripts do not authorize network tools.

Session mining is a separate disclosure decision

mine_session_logs.py reads immediate *.jsonl files under selected ~/.claude/projects/ project directories, normally modified within the last seven days. The default allowlist is claude-trading-skills, claude-market-agents, trade-edge-finder, trade-strategy-pipeline and weekly-trade-strategy. A --project override can change/broaden the selected projects. It does not scan all Codex history folders.

Parsing extracts external user messages, assistant tool names/inputs and tool results/errors to find signals; sidechains and malformed records are skipped. The model prompt can include project names, patterns, up to five selected user excerpts (200 characters each), and up to three samples per signal (100 characters each). It invokes claude -p through stdin, with up to three turns, a USD 1 budget and a 600-second timeout. These limits do not anonymize the text. Review consent and content before running mining over sessions containing personal information.

--dry-run skips model abstraction but still reads/parses local logs and writes derived candidates. Raw candidates/session details may preserve sensitive text. The weekly generation pipeline invokes the miner; later generation/backlog/PR steps may publish derived material. The improvement loop also submits skill/review prompts through claude -p. Inspect output before sharing; neither path guarantees PII removal. Disabling a schedule alone does not erase existing transcripts or generated files.

Retention policy and current automation

Operational target: retain routine diagnostics for no more than 30 days, minimize report/session extracts, and retain investigation evidence only until its review and recovery needs are resolved. The operator must review and remove unnecessary private files and copies; confirm backup needs before any deletion. Provider-side retention is separate and governed by the configured provider/account settings.

Current implementation is narrower: generation and improvement scripts rotate only immediate logs/*.log files older than 30 days and cap their own state histories at 60 entries. Nested logs, JSON state files, reports, provider caches, Claude transcripts and backup copies have no uniform automatic purge. This policy does not claim that such purging is implemented. Do not use broad cleanup commands against active or unreviewed evidence.

For a source-preserving recovery drill, follow the tested thesis backup/restore procedure. It requires stopped writers, private storage, fresh destinations, a full file-list and SHA-256 comparison, and the store’s read-only doctor. Tests use fictional state and restore after fixture source loss; this is not proof of your own backup. See SECURITY.md for reporting limitations and support policy. Capability statements here describe source reviewed on 2026-11-14; deployed documentation and provider settings may differ.