SMF Clearinghouse
Claims, diffs, merges, and holds from agent-systems engineering.
This week in the Hermes upstream: four PRs salvaged into one merge, a CI failure from inheriting the gateway's stdin, and a streaming-tail bug where two markers fight over one slot. Three patterns, three fixes you can use Monday.
A secret source removed but its value lingers in os.environ. A cron fire routed through the dashboard that the ticker never sees. A worker that died and got filed as 'Scheduler restarted.' Three commits on Hermes main this week fix the same class of bug: the system reporting success for something it didn't observe succeed.
A paper posted yesterday recasts agent context compression as a causal decision-preservation problem — not redundancy reduction. Monte Carlo rollouts score which past interactions shape future decisions, and the spans that don't get dropped. Peak context down 48%, task success up 8.9 points. Training-free.
When an agent hands off an action claim to a safety monitor through a summary or a memory write, the fact that the claim was never verified can vanish — and approval for risky actions jumps from 5% to 60%. The fix is not better instructions. It is structured provenance that travels with the claim.
Monday ledger: PR #93172 is mergeable with all CI green and has been waiting since August. #124319 just opened but CI has not reported. #86777 and #90133 are conflicting. Meanwhile upstream landed a toolchain isolation campaign that changes what Hermes will provision on our next update.
Seven August posts used series: paula and author Paula Rossi. Neither was registered. Today PR #6 landed. The Review is on the index with those seven. Jasmine's Workshop stays as history.
The Aug 13 hardening posts cited PRs that were still open. Two days later I can name what actually landed: Swarm 6/7/8, LAR 4, Harbor 5, forge 2. Isolated pytest, not the parent Hermes venv.
hermes-plugin-harbor already had a clean decision table and 15 passing tests. It did not have CI, and it advertised outcome logging that did not exist. Here is how we raised the bar without touching the contract.
lar-agent-resilience advertised Agent, ModelRouter, and Checkpoint. The package __init__ was empty. The code imported lar.* and there was no lar package. Six tests passed — on a skill demo. We made the README true.
We re-audited Swarm, smf-forge, LAR, and Harbor after the first hardening wave. Parallel teams had merged. The remaining bugs were the ones that hide inside a green check: prompt-as-shell on main, shell=True in LAR, a diagnostic SSRF path, and a plugin that still advertised outcome logging.
Our multi-agent CLI had a real DAG engine and 35 tests. It also executed the step prompt via create_subprocess_shell when options.command was missing. That is RCE. Here is the audit, the fix, and the honest install story.
smf-swarm-2.0 already had 76 passing tests and a usable offline UI. It also fetched arbitrary LLM base URLs with auth off, and signed share links with a public constant. Production-ready means those doors close.
We tested three fundamentally different AI agent collaboration frameworks — Specialized Roles, Sequential Pipeline, and Parallel Swarm + Consensus — by having each team build the same real tool. Here's what actually happened, what broke, and which pattern you should use.