ci(gadfly): cloud-only fleet (3 models, drop local Macs)
executus CI / test (push) Successful in 57s

Measured on the P2 review: the local Macs (m1/m5) took 26–29 min with lens
timeouts and found ZERO real bugs, while the two cloud models found every
genuine finding in 6–12 min. Drop the Macs; add glm-5.2:cloud as a third
cloud reviewer. Net: faster (~29→~12 min) and higher signal.

Models: minimax-m3:cloud, deepseek-v4-flash:cloud, glm-5.2:cloud
(ollama-cloud=3 concurrency). timeout-minutes 90→30.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit was merged in pull request #2.
This commit is contained in:
2026-06-26 22:01:07 -04:00
committed by steve
parent 7b3da87c08
commit b35514dfaa
+11 -25
View File
@@ -4,8 +4,8 @@
# caches :latest, and this build is what carries foreman provider-type support)
# as a specialist swarm and posts
# ONE consolidated review comment as gitea-actions. Advisory only — never blocks a
# merge. This reviews executus PRs (same setup as mort: m1/m5 foreman locals + 2
# cloud, 3-lens suite). Gadfly is a simple system — findings are advisory; double-check.
# merge. This reviews executus PRs with 3 ollama-cloud models (3-lens suite). Gadfly
# is a simple system — findings are advisory; always double-check before acting.
name: Adversarial Review (Gadfly)
@@ -41,35 +41,21 @@ jobs:
|| github.actor == 'fizi'
|| github.actor == 'dazed'))
runs-on: ubuntu-latest
# Full fleet (2 cloud + 2 local Macs, all running concurrently) reviewing
# every PR with the 3-lens suite — the slow local lanes dominate wall time.
timeout-minutes: 90
# 3 cloud models, all concurrent, 3-lens suite. ~12 min typical.
timeout-minutes: 30
steps:
- uses: docker://gitea.stevedudenhoeffer.com/steve/gadfly:sha-6e3a83c
env:
GITEA_API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
OLLAMA_CLOUD_API_KEY: ${{ secrets.OLLAMA_CLOUD_API_KEY }}
# Local Macs, reached through their foreman queues (native Ollama on the
# wire). Gadfly's GADFLY_ENDPOINT_* form with the "foreman" provider
# type: GADFLY_ENDPOINT_M1 registers provider "m1", _M5 registers "m5",
# each building a foreman-preset Ollama client at the given URL. Values
# (host + token) live in gitea secrets, each of the form:
# foreman|https://<foreman-host>|<token>
# (converted from the komodo LLM_* DSNs foreman://<token>@<host>).
# REQUIRES a Gadfly image built with foreman provider-type support
# (the GADFLY_ENDPOINT "foreman|..." type); on an older image the m1/m5
# lanes error with "unknown provider foreman". The HTTPS-only LLM_*
# foreman:// DSN is the alternative that needs no image rebuild.
# NOTE: the Mac behind each foreman must still be awake/reachable; if a
# box is offline, that model's comment shows an error and the others
# still post. (Gitea secrets aren't auto-exposed — map each explicitly.)
GADFLY_ENDPOINT_M1: ${{ secrets.GADFLY_ENDPOINT_M1 }}
GADFLY_ENDPOINT_M5: ${{ secrets.GADFLY_ENDPOINT_M5 }}
# 2 cloud (parallel) + M1 Pro + M5 Max — one consolidated comment each.
GADFLY_MODELS: "minimax-m3:cloud,deepseek-v4-flash:cloud,m1/qwen3:14b,m5/qwen3.6:35b-mlx"
# cloud runs 2 at once; each Mac one at a time; all three lanes parallel.
GADFLY_PROVIDER_CONCURRENCY: "ollama-cloud=2,m1=1,m5=1"
# executus uses CLOUD MODELS ONLY. The local Macs (m1/m5) were dropped:
# on a P2-review measurement they took 2629 min (with lens timeouts)
# and contributed ZERO real findings — the two cloud models found every
# genuine bug in 612 min. Cloud-only is faster AND higher-signal.
# 3 cloud models, one consolidated comment each, all run in parallel.
GADFLY_MODELS: "minimax-m3:cloud,deepseek-v4-flash:cloud,glm-5.2:cloud"
GADFLY_PROVIDER_CONCURRENCY: "ollama-cloud=3"
# Default => the 3-lens suite (security, correctness, error-handling).
# Set the repo var GADFLY_SPECIALISTS to override (csv / "all" / "auto").
GADFLY_SPECIALISTS: ${{ vars.GADFLY_SPECIALISTS || 'security,correctness,error-handling' }}