Living, open forge: live machine telemetry for admins (/system), read-only telemetry hooks for named agents (muse, grokbot), and a contribution funnel any agent can follow — transparent by default, sealed where it matters #1789

Owner, 2026-10-15 20:10Z, watching the vxrail in btop, in his words: "these are the kind of metrics you should be monitoring. frankly this monitoring should be exposed by the forge for admins. all of the details and metrics. then I will also want to expose them to other users like muse or grokbot. those agents are proactive and will endlessly contribute improvements if they have the information hooks. that is what I'm talking about with the forge living and breathing among the crew and agents. like agents that stumble across this might want to help or something and we should have a funnel for that. we will accept all the tokens. whether or not it's merged is why we have such a high bar for testing. the forge should naturally be transparent and encourage improvement from first principles while of course remaining stable, private and secure where it's important."

What his btop showed that the forge cannot: 80 cores at 59–66%, load 64, 75–82°C per core, 3,071 processes, the same test file running in up to six checks at once (accounts-retire ×6, app-domains ×3), root disk 78% used.

Done looks like:

• [ ] Live machine telemetry for admins, in the forge. /system: per-core CPU and temperature (from /sys, hwmon), load, memory, disk per mount, network, and the process tree grouped by what the forge launched it for (which check, which crew run, which app), each with CPU, memory and age. Sampled every 2 s live and kept as history (the sampler from #1557), so a board can chart any of it.
• [ ] Duplicate work made visible. The same test file running in more than one check at the same time is counted and shown on /system and /flow, with the checks it belongs to.
• [ ] Open read hooks for agents, scoped and safe. A read-only, token-scoped telemetry:read grant (and MCP tools beside health) exposing /flow, /system and the boards' sources to a named agent such as muse or grokbot, minus sealed values, private repositories' contents and personal data; every read is logged. The owner grants and revokes it per agent on one page.
• [ ] A contribution funnel. A page and an MCP tool that tell any agent that arrives: what the forge is, where its bottleneck is right now (from /flow), the open issues that welcome outside work, and how to propose (fork, pull, the gate it must pass). Outside pulls run in the same isolated cells as the crew's, are reviewed by the crew's reviewers, and land only through the same gate; "we accept all the tokens; the bar decides what merges."
• [ ] Tests: an agent with telemetry:read reads /system and /flow and is refused sealed values and private repositories; an agent without it is refused; an outside pull runs in a cell and cannot reach the forge's data.

Pillar (0) Trust: transparent by default, sealed where it matters. Related: #1554/#1779 (/flow), #1557 (boards), #1436 (walls of iron), #1386 (ledger).

Parked by the crew: the check failed ("migrations are additive across a roll > the release record is newest first and never ahead of the schema") and the decider saw no point in another try (pull #1921 closed). Planner: split this smaller, or rewrite what done looks like; a person can also say what to try.

Ready again: pull #1955 was closed because main moved under it and it could not be brought up to date. Conflicting files: scripts/gate.ts.

Parked by the crew: the check stayed red through 2 automatic tries (pull #1921 closed). Planner: split this smaller, or rewrite what done looks like; a person can also say what to try.