|
Download docs/agent-trace-draft.md from build-small-hackathon/Mystery_Mail_Guardian: direct link, hf CLI and curl.
- Browser
- Download file 4.04 kB
-
https://huggingface.co/spaces/build-small-hackathon/Mystery_Mail_Guardian/resolve/main/docs/agent-trace-draft.md
- Command line
-
hf download hf://spaces/build-small-hackathon/Mystery_Mail_Guardian/docs/agent-trace-draft.md
-
curl -L -o agent-trace-draft.md https://huggingface.co/spaces/build-small-hackathon/Mystery_Mail_Guardian/resolve/main/docs/agent-trace-draft.md
4.04 kB
| # Agent trace β Mystery-Mail Guardian (PUBLISHED 2026-06-12, owner-approved) | |
| > **Status: PUBLISHED** with owner consent to | |
| > [build-small-hackathon/mystery-mail-guardian-agent-trace](https://huggingface.co/datasets/build-small-hackathon/mystery-mail-guardian-agent-trace). | |
| > `achievement:sharing` tag added to README. The published dataset card lives | |
| > in `dist/agent-trace/README.md`; this file is the working draft it was built | |
| > from. Sanitization was re-verified by automated secret scan before upload. | |
| ## What this is | |
| Mystery-Mail Guardian was built end-to-end with Claude Code (Anthropic) as the | |
| hands and a human owner as the director. This trace is the honest, sanitized | |
| record of how that collaboration actually worked β the prompts that steered it, | |
| the loop that ran it, and the evidence trail it left behind. | |
| ## How the build was driven | |
| - **Planning first:** every major phase started as a written plan | |
| (`docs/superpowers/plans/`) with TDD steps, verified against live docs | |
| rather than model memory (Gradio 6 API drift caught this way twice). | |
| - **The owner steered with short prompts**, e.g.: | |
| - "the loading icon is broken β¦ I provided this image and it gave this | |
| output which I think is wrong" β led to the prompt-echo discovery (the | |
| 1.3B model parroting our JSON schema back) and the single-overlay loader. | |
| - "make an Easter egg when the user switches to Japanese" β the Gen X Soft | |
| Club palette, held to the same WCAG-AA tests as the main themes. | |
| - "train the model more using our Modal credits" β answered with evidence | |
| instead of obedience: a labeled synthetic-letter dataset + eval matrix | |
| that showed safety behavior was already at ceiling, so no fine-tune | |
| shipped β the data made the call. | |
| - "enter a state of loop-craft β¦ keep improving until June 14" β a | |
| self-rescheduling build loop with a committed worklist | |
| (`docs/loop-worklist.md`), one verified iteration per wake. | |
| - **Every iteration had the same gate:** offline tests green β Modal GPU | |
| validation where behavior changed β deploy β live probes against the | |
| running Space β commit with evidence. | |
| ## What the trace shows (highlights) | |
| 1. **Live fire beats simulation.** Three JSON-repair bugs and one genuinely | |
| unsafe model suggestion ("check with the tax bureau and ensure you pay | |
| soon") were caught on the deployed Space; the safety layer discarded the | |
| bad advice by construction (BUILD_LOG 06-11). | |
| 2. **The QA harness caught real bugs on its first runs.** Run #1 of the Modal | |
| matrix found a scammer's reply address leaking through key facts; the | |
| multilingual eval found Devanagari/Japanese scam letters under-flagged. | |
| Both fixed, gated, and redeployed the same day (BUILD_LOG 06-12). | |
| 3. **A NO-GO gate did its job.** A prompt improvement regressed one Spanish | |
| case ("within a set time" paraphrase); the deploy was blocked until the | |
| urgency heuristic was generalized with tests, then re-gated to GO. | |
| 4. **Tooling archaeology, documented:** `launch(js=)` dead in Gradio 6.17.3, | |
| js-only `demo.load` registered-but-never-run, `head=` dropped by Space | |
| SSR β resolved with `head=` + `ssr_mode=False` (BUILD_LOG 06-12). | |
| ## Companion artifacts (already public in the repo) | |
| - `BUILD_LOG.md` β the full field notes, including dead ends | |
| - `docs/modal-validation-report.json`, `docs/modal-eval-report.json` β GPU | |
| evidence as committed JSON | |
| - `docs/loop-worklist.md` β the autonomous loop's rules and backlog, as run | |
| - Git history β every step, with co-authorship trailers | |
| ## Sanitization checklist (verify before publishing) | |
| - [x] No API keys, tokens, or secret values (HF token lives only in the local | |
| cache + a Modal secret; never in the repo) | |
| - [x] No real personal data β every letter in datasets/examples is synthetic; | |
| phone numbers use the reserved 555-01XX range | |
| - [x] No private URLs or credentials in logs quoted here | |
| - [x] Owner approved publication (chat, 2026-06-12) | |
| - [x] `achievement:sharing` tag added to README after upload | |