default.exp-tas.com — Microsoft AID/Experimentation rollout
marketplace.visualstudio.com + vscode.blob.core.windows.net — VS Code Marketplace and VSIX downloads
*.githubassets.com, *.vscode-unpkg.net — asset/CDN hosts used by extension UI
.github.com, *.githubusercontent.com, and the rest of stt_pkg_domains were already allowed, so no overlap.
Validation on prod will be a simple login + completion request from an IDE through http://squid:3128.
Allow Hermes to reach GitHub Copilot network endpoints through the corporate squid proxy.
Adds a `copilot_domains` ACL covering the domains required by GitHub Copilot (IDE integration and ACP):
- `api.github.com` — user/auth and copilot REST endpoints
- `copilot-proxy.githubusercontent.com` — completion and chat streaming backend
- `*.githubcopilot.com` — newer Copilot control plane
- `copilot-telemetry.githubusercontent.com` — older Copilot telemetry
- `default.exp-tas.com` — Microsoft AID/Experimentation rollout
- `marketplace.visualstudio.com` + `vscode.blob.core.windows.net` — VS Code Marketplace and VSIX downloads
- `*.githubassets.com`, `*.vscode-unpkg.net` — asset/CDN hosts used by extension UI
`.github.com`, `*.githubusercontent.com`, and the rest of `stt_pkg_domains` were already allowed, so no overlap.
Validation on prod will be a simple login + completion request from an IDE through `http://squid:3128`.
## What
Wires up local STT (faster-whisper) into the hermes-gateway container so voice messages on Telegram transcribe without depending on external services.
Closes#19. Supersedes the closed#20 (which used a custom Dockerfile-hermes; this version uses a bind-mounted entrypoint script instead since pypi.org is now whitelisted).
## Changes
1. **`hermes_squid/squid.conf`** — expands the egress whitelist:
- new `stt_pkg_domains` for `pypi.org`, `files.pythonhosted.org`, `huggingface.co`, `github.com`, `models.dev`, `registry.npmjs.org`, `cdn-lfs.huggingface.co` — needed for runtime `pip install faster-whisper` and HF model fetch.
- new `telegram_ips` covering AS62041 (91.108.4.0/22, 91.108.8.0/22, 91.108.12.0/22, 91.108.16.0/22, 91.108.20.0/22, 91.108.56.0/22, 149.154.160.0/20, 185.76.151.0/24) as a backstop for the `dstdomain` ACL. squid's `dstdomain` does not match HTTPS CONNECT requests without `ssl_bump`, so the recent log shows 24 of 9083 `api.telegram.org` requests getting 403 in a 5-minute window.
- renamed `hermes_network` -> `localnet` (squid default name, no behavior change).
2. **`hermes_squid/hermes-gateway-entrypoint.sh`** (new) — bind-mounted into the container at `/opt/data/hermes-gateway-entrypoint.sh` and set as the service entrypoint. The script:
- is idempotent: on first start, runs `apt-get install -y python3-pip ffmpeg` and `pip3 install --break-system-packages faster-whisper`. Subsequent restarts skip install if `faster_whisper` is already importable.
- exports `STT_PROVIDER=local`, `STT_LOCAL_MODEL=base`, `STT_LOCAL_DEVICE=cpu`, `STT_LOCAL_COMPUTE_TYPE=int8`.
- `exec`s the original gateway command (`gateway run`).
3. **`hermes_squid/stack-compose.yml`** — `hermes-gateway` gets:
- `entrypoint: ["/opt/data/hermes-gateway-entrypoint.sh"]`
- `command: ["gateway", "run"]`
- bind-mount `./hermes-gateway-entrypoint.sh:/opt/data/hermes-gateway-entrypoint.sh:ro`
- new env: `STT_PROVIDER`, `STT_LOCAL_MODEL`, `STT_LOCAL_DEVICE`, `STT_LOCAL_COMPUTE_TYPE`
No other service in the compose file is touched.
## Why runtime install, not a custom image
- `pypi.org` is now whitelisted, so `pip3 install` works at runtime through the same squid.
- Avoids maintaining a custom hermes-agent image and a local registry tag.
- The upstream `nousresearch/hermes-agent:v2026.8.3` image is kept as-is.
## Deploy notes
- brencher apply will trigger `docker service update` for hermes-gateway; the new entrypoint runs the apt+pip install on first start (~1-2 min).
- First start also downloads the `base` faster-whisper model (~150 MB) from `huggingface.co` — covered by `stt_pkg_domains`.
- squid config is reread on `docker service update` for the squid service; new ACLs apply without a full stack restart.
- Subsequent restarts of hermes-gateway skip the install path entirely (idempotent guard).
## Verified locally
- `git diff --stat upstream/master` shows only the three files above (+25 entrypoint, +14 squid, +8 compose).
- The new worktree is `/opt/data/work/thread-114-Alex-InfraAlexBot/ansible` (per-session working dir).
## Risks
- `apt-get update` inside the container needs `deb.debian.org` reachable — already in `allowed_domains`.
- `STT_LOCAL_MODEL=base` is ~150 MB and runs on CPU. If latency on Raspberry-class nodes is bad, switch to `tiny`.
- `--break-system-packages` is required for PEP 668 (Debian trixie + Python 3.11+). If the upstream image switches to PEP 668 strictly, this flag is the right escape hatch.
## Related
- #19 — investigation that found missing faster-whisper + squid deny pattern
- #20 — superseded (used Dockerfile-hermes instead of runtime install)
---------
Co-authored-by: hermes <hermes@rudolf.keenetic.link>
Reviewed-on: http://192.168.1.96/admin/ansible/pulls/21
Reviewed-by: admin <rudolfss13@gmail.com>
Co-authored-by: hermes <rudolfss13+hermes@gmail.com>
Flip CONTAINERS/IMAGES/NETWORKS/VOLUMES/SERVICES/TASKS/NODES from 0 to 1 so hermes-agent can inspect running stacks (read container logs, enumerate swarm services) without going through each node's shell.
POST stays 0 — no container creation, no image builds, no exec into running containers. Anything that mutates state remains blocked at the proxy level.
---------
Co-authored-by: admin <rudolfss13@gmail.com>
Co-authored-by: hermes <hermes@rudolf.keenetic.link>
Reviewed-on: http://192.168.1.96/admin/ansible/pulls/4
Co-authored-by: hermes <rudolfss13+hermes@gmail.com>
## Problem
The built-in `act_runner` cache server binds to an interface unreachable from per-job container networks (`container.network: ""` is the default). `ACTIONS_CACHE_URL` ends up pointing at an unreachable IP, and every CI job burns ~9 minutes on:
```
::warning::Failed to restore: getCacheEntry failed: connect ETIMEDOUT 192.168.112.21:37051
::warning::Failed to save: reserveCache failed: connect ETIMEDOUT 192.168.112.21:37051
```
Measured on `admin/brencher#15` (https://gitea.rudolf.keenetic.link/admin/brencher/actions/runs/16):
| Job | wall-clock | of which dead air on cache |
|---------|------------|----------------------------|
| 3.12.3 | 12m 40s | ~9m (271s × 2 timeouts) |
| 3.13 | 11m 13s | ~9m |
| 3.14 | 11m 17s | ~9m |
So even though `python-package.yml` enables `cache: 'pip'`, the cache backend is unreachable, jobs hit `pip install` cold every time, and CI ends up ~4x slower than necessary.
## Fix
Run a dedicated `act_runner cache-server` container on the same overlay network as the runner, and point the runner at it via `external_server`. Shared secret comes from a Docker secret mounted into both containers and read via `ACT_RUNNER_EXTERNAL_SECRET_FILE`.
## Changes
- `poc/gitea/stack-compose.yml`:
- New `cache_server` service (`command: ["cache-server"]`) on the `gitea` overlay with a `cache_data` named volume.
- `runner` and `cache_server` both mount the external Docker secret `gitea_cache_secret` at `/run/secrets/gitea_cache_secret` and set `ACT_RUNNER_EXTERNAL_SECRET_FILE` accordingly.
- Top-level `secrets:` block declares `gitea_cache_secret` as `external: true`, with the create command commented inline next to it.
- `poc/gitea/config.yaml`: point `cache.{host,port,external_server}` at `gitea_cache_server:37051`; `external_secret` left empty because the secret is loaded from the file.
## Verification
- Both YAMLs parse (PyYAML).
- Service name `gitea_cache_server` resolves on the `gitea` overlay via Swarm service-name DNS.
- `docker secret inspect gitea_cache_secret` works after the one-time create step.
## Deploy notes
1. On the swarm manager, create the secret once:
```
echo -n "$(openssl rand -hex 32)" | docker secret create gitea_cache_secret -
```
2. `docker stack deploy -c stack-compose.yml gitea`.
3. Re-run a CI job and confirm cache `restore`/`save` log lines no longer ETIMEDOUT.
## Related
- Related: admin/brencher#15 — `python-package.yml` enables `cache: 'pip'` and was timing out for the reason this PR fixes.
- Refs: gitea/actions/runs/16 jobs #248/249/250.
## Supersedes
- Supersedes: admin/ansible#33 — first attempt was based on a stale `origin/master` and accidentally included `hermes_squid/` changes from the actual upstream merge. The current branch is rebased on `admin/master` (`c3b7128`) and only touches `poc/gitea/`.Reviewed-on: http://192.168.1.96/admin/ansible/pulls/34
Reviewed-by: admin <rudolfss13@gmail.com>
Co-authored-by: hermes <rudolfss13+hermes@gmail.com>
Allow Hermes to connect to api.github.com, copilot-proxy,
githubcopilot.com, vscode marketplace and Microsoft experimentation
endpoints required by GitHub Copilot.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Allow Hermes to reach GitHub Copilot network endpoints through the corporate squid proxy.
Adds a
copilot_domainsACL covering the domains required by GitHub Copilot (IDE integration and ACP):api.github.com— user/auth and copilot REST endpointscopilot-proxy.githubusercontent.com— completion and chat streaming backend*.githubcopilot.com— newer Copilot control planecopilot-telemetry.githubusercontent.com— older Copilot telemetrydefault.exp-tas.com— Microsoft AID/Experimentation rolloutmarketplace.visualstudio.com+vscode.blob.core.windows.net— VS Code Marketplace and VSIX downloads*.githubassets.com,*.vscode-unpkg.net— asset/CDN hosts used by extension UI.github.com,*.githubusercontent.com, and the rest ofstt_pkg_domainswere already allowed, so no overlap.Validation on prod will be a simple login + completion request from an IDE through
http://squid:3128.Pull request closed