Live-Test a Daily EMA20 and EMA50 Crossover

You are building a deterministic daily EMA crossover research scanner for me.

Configuration:
- Project path: [PROJECT_PATH]
- Symbol: [SYMBOL]
- Timeframe: 1 day
- Timezone: [TIMEZONE]
- Strategy version: ema20_ema50_daily_v1
- Discord destination: the channel connected to the protected environment variable EMA_CROSS_DISCORD_WEBHOOK_URL

Strategy rules:
1. Use completed daily candles only. Never evaluate an unfinished current candle.
2. Calculate EMA20 and EMA50 from the same adjusted-close policy and document that policy.
3. Create a LONG signal when the previous completed candle has EMA20 less than or equal to EMA50 and the newest completed candle has EMA20 greater than EMA50.
4. Create a CLOSE signal when the previous completed candle has EMA20 greater than or equal to EMA50 and the newest completed candle has EMA20 less than EMA50.
5. This starter system is long or flat. Do not create short signals.
6. Do not place broker orders. This is a research and alerting workflow only.

Architecture and token rules:
1. AI tokens may be used while you design, write, test, review, or repair the system.
2. Normal scheduled runs must use deterministic Python only and must not call an LLM, agent, web-search tool, or AI API.
3. Build one Python entry point that fetches data, validates it, calculates the EMAs, checks the crossover, updates state, writes logs, creates the chart, and sends Discord alerts.
4. The Hermes cron must be script-only with no_agent set to true. Do not schedule a recurring agent prompt.
5. A healthy run with no new signal must print nothing and send nothing.
6. A hard failure must exit nonzero with a concise error so the scheduler can report it.

Data and reliability requirements:
1. Use a documented market-data provider supported by the symbol. Isolate the provider behind a small function so it can be replaced later.
2. Require enough valid history to calculate EMA50 with a safety buffer.
3. Reject empty data, duplicate timestamps, non-numeric closes, stale data, and an unfinished newest candle.
4. Use retry with bounded exponential backoff for transient network failures.
5. Save an append-only JSONL run log containing run time, symbol, latest completed candle date, close, EMA20, EMA50, signal decision, strategy version, data freshness, and any rejection reason.
6. Save deduplication state atomically. The key must include symbol, signal type, strategy version, and candle date.
7. Do not mark a signal as delivered until Discord confirms a successful response.
8. Never print or log the Discord webhook URL.

Discord and chart requirements:
1. Send every new LONG or CLOSE signal to the Discord webhook stored in EMA_CROSS_DISCORD_WEBHOOK_URL.
2. Attach a PNG chart to every signal. Do not send a text-only signal.
3. Plot at least the latest 180 completed daily candles, closing price, EMA20, EMA50, and a visible marker on the signal candle.
4. The alert text must include symbol, signal type, completed candle date, close, EMA20, EMA50, strategy version, and the statement: Research alert only, not financial advice.
5. If chart generation or attachment upload fails, treat the delivery as failed and do not mark the signal as sent.

Project requirements:
1. Create a local virtual environment and a pinned requirements file.
2. Keep configuration separate from code. Secrets belong in environment variables.
3. Add focused automated tests using synthetic candles for bullish cross, bearish cross, no cross, duplicate prevention, unfinished-candle rejection, stale data, Discord failure, and successful chart attachment.
4. Add a dry-run mode that creates the chart and alert payload without contacting Discord.
5. Add a test-alert mode that sends a clearly labeled TEST alert and chart without changing strategy deduplication state.
6. Write a README with setup, environment variables, commands, log locations, cron schedule, troubleshooting, and removal instructions.

Before finishing:
1. Run the automated tests and show the real result.
2. Run dry-run mode against current data.
3. Run test-alert mode only after confirming the webhook variable exists.
4. Inspect the generated PNG and verify it is nonempty.
5. Report exact created paths and the final script command.

Do not stop at a plan or pseudocode. Build and verify the working local project.

Create a Hermes cron job for the verified EMA crossover scanner at [PROJECT_PATH].

Requirements:
1. Schedule it at [RUN_TIME] in [TIMEZONE], after the selected market's daily candle is confirmed closed.
2. Use the project's virtual-environment Python and absolute script path.
3. Configure the job as script-only with no_agent=true. It must not call an LLM or start an agent session during normal runs.
4. The Python script must send each new signal and PNG chart directly to the Discord channel webhook in EMA_CROSS_DISCORD_WEBHOOK_URL.
5. Keep cron delivery local or otherwise disabled for ordinary output so Discord does not receive duplicate scheduler messages.
6. Empty stdout means a healthy no-signal run and must remain silent.
7. Nonzero exit means a real operational failure and should trigger the scheduler's normal error alert.
8. Do not put the webhook URL inside the cron definition, command line, logs, or prompt.
9. Run the cron manually once, inspect its actual status, and verify the run log changed.
10. Confirm that the recurring job has no model override, no skills, and no unnecessary toolsets.

Before creating anything, inspect the actual scanner command and Hermes cron capabilities on this installation. After creation, return the job name, job ID, schedule, script path, no_agent value, delivery mode, and the result of the manual verification run.

Audit the EMA20 and EMA50 daily crossover project at [PROJECT_PATH] as an operational research system.

Do not redesign the strategy. Verify the current implementation with real commands and test fixtures.

Confirm all of the following:
1. Only completed daily candles can generate signals.
2. Bullish and bearish crossover logic uses the previous and newest completed candles exactly as specified.
3. Re-running the same candle cannot send a duplicate alert.
4. A Discord signal is not marked delivered unless both the message and PNG attachment succeed.
5. The chart contains price, EMA20, EMA50, and the signal marker.
6. No webhook secret appears in source, logs, process arguments, or test output.
7. Normal scheduled runs use deterministic Python and no AI tokens.
8. The Hermes cron is script-only with no_agent=true and has no unnecessary model, skills, or toolsets.
9. Healthy no-signal runs are silent.
10. Data, Discord, or chart failures exit nonzero and leave useful bounded diagnostics.
11. The append-only run log and atomic deduplication state remain valid after an interrupted write.
12. The README matches the actual commands and file paths.

Run the focused tests, dry-run mode, and a simulated Discord failure. Inspect the newest log records and generated chart. Return a pass or fail checklist backed by the real command output. Fix any failure you find, rerun the checks, and report the final verified state.