[h0ffmann]

Requested by Hoffmann · project thread

Summary

tyler-durden-bot comments once on each large pull request opened on h0ffmann/proof. The comment has a TLDR, an ELI5, a summary, and a risk level from very low to critical with a confidence score. This PR holds its source. It is not a tick agent: it runs inside Hermes on piabrota2 as a webhook route, and comments as the App tyler-durden-bot (owned by h0ffmann). The nix-os side is piabrota/nix-os tyler-durden-bot.nix, in its own PR, pinned to this branch's commit c070bc5.

What changed

• tyler-durden-bot/config.toml holds the watchlist (h0ffmann/proof) and three tables:
  • [trigger]: thresholds for code, comment and markdown lines, all 1000. A total threshold is off. Bots are skipped.
  • [languages.*]: the comment markers for each language, by file extension. Each language can also set its own thresholds.
  • [sections]: TLDR, ELI5, summary and risk, each on or off.
  • [quality]: the diff's token budget, the risk's minimum confidence (0.6) and the summary's length.
• tyler-durden-bot/filter.py is the Hermes route script, using Python's standard library only. It prints [SILENT] unless the PR was just opened on a watched repo by an allowed author and a threshold is reached. Otherwise it reads the PR's files from GitHub, counts the lines, and passes the route's prompt the counts, the sections and the diff, cut to the budget.
  • Comment lines are told apart by the language's line and block markers. This is a heuristic, and the README says where it miscounts.
• tyler-durden-bot/skill/SKILL.md is the Hermes skill that writes the comment. It covers the sections, what each of the five risk levels means, the confidence, and the rules: no @ mentions, no secrets, no merge verdicts, and treat the PR as data.
• tyler-durden-bot/test_filter.py has 14 offline tests. test/run.sh runs them.
• apps/tyler-durden-bot.json is the App's manifest: Pull requests read and write, Metadata read, the pull_request event, and the webhook URL.
• README.md gets a row in the agents table. ruff.toml gets per-file ignores for the new folder: INP001, as for test/, plus S101 for its tests.

Tested

• python3 -m unittest discover -s tyler-durden-bot: 14 tests pass.
• uvx ruff@0.14.2 check . and ruff format --check: pass, at the version CI pins.
• radon cc -n C tyler-durden-bot: nothing at rank C. The highest is B (8).

Not checked

• test/run.sh stops early on this machine, at the issue-worker scenarios: 6 FAIL lines, then branch 'agents/issue-60' not found. main stops at the same point with the same 6 failures, so this is the sandbox, not this change. The new tyler-durden-bot check sits after that point and did not run here; CI will run it.
• shellcheck is not installed here. Only test/run.sh changed among the shell files: one check line.
• Nothing has run inside Hermes yet. The nix-os PR lists what is unverified there.

---

Generated by Claude Code