← Back to blog

Org-Skills Namespace and Context Engine Verbs Land in 48-Hour Window

hermesopen-sourceroundup
Org-Skills Namespace and Context Engine Verbs Land in 48-Hour Window

NOTE: X API v2 and OpenRouter both returned HTTP 402 Payment Required on July 24. This roundup was sourced from GitHub API discovery. No tweet embeds or AI-generated OG image are included.

The 48-hour window from July 22 to July 24 saw 51 feature and performance PRs land in Hermes Agent. Two structural changes to the extensibility model lead the window: a first-class org-skills namespace and two new ContextEngine abstract base class verbs. A wave of compression improvements and a Mixture of Agents cadence overhaul fill out the rest.

Org-skills namespace: token-gated discovery and fail-loud collisions

Ben Barclay landed two PRs that together form the client leg of the M2 org-skills work. The namespace design makes org skills first-class with three companions: token-gated discovery, fail-loud collisions when two orgs publish the same skill name, and provenance tracking that records which org supplied each loaded skill.

#70459 (+529/-16, 7 files) implements the namespace itself in the skill manager, prompt builder, and skill utilities. Skills enter the namespace with a bare name and three non-negotiable guards. The prompt builder now resolves org-scoped skills against a provenance table rather than a flat list.

#70024 (+490/-7, 4 files) adds the sync client: hermes skills propose to submit org skills, _org/ pull from the NAS, and HTTP 202 handling for async approval. Stacked on the M1 personal-sync client, it reuses the HSP wire contract.

The two PRs combined add 1,019 lines across 11 files. The companion work in gateway and NAS means this lands as a cross-repo feature: the org can publish skills once and every agent in the org discovers them on next sync.

ContextEngine ABC: select_context and on_turn_complete

#70458 (+927/-0, 7 files) salvages a PR from @chaos-xxl that implements RFC #36765. Two additive, no-op-default, fail-open verbs now sit on the ContextEngine abstract base class:

  • select_context() runs before each request and lets a plugin choose which context sources to include. The default path returns everything unchanged.
  • on_turn_complete() fires after the final response, giving plugins a hook into the turn lifecycle for post-processing, logging, or side effects.

The implementation is additive: zero changes to existing behaviour when no plugin registers. A default-path short-circuit avoids any overhead on the no-plugin path. The PR includes 555 lines of tests across two test files and a developer-guide doc page.

MoA advisor cadence: cheaper by default

Two PRs from teknium1 rework how Mixture of Agents advisors fire:

#70507 (+85/-57, 7 files) sets the default advisor fanout to user_turn - advisors run once on the first message of each user turn and the aggregator reuses that advice throughout the tool loop. Until benchmarks justify a costlier default, MoA defaults to the cheapest cadence.

#70284 (+892/-18, 7 files) adds two opt-in dials: an every_n fan-out cadence that reuses cached guidance on off-cadence iterations, and a privacy redaction filter with display/full modes. Both are config-gated and default-off.

Compression wave: tail guarantee, tool-result pruning, progress notices

Three PRs tighten the compression pipeline:

#70250 (+565/-0, 9 files) guarantees the last N real user messages survive compression via compression.min_tail_user_messages. Salvaged from @zhangyang-crazy-one's original PR.

#70254 (+876/-1, 12 files) adds opt-in proactive tool-result pruning for large-window models. Configurable thresholds let operators reclaim context budget from stale tool output before compression even runs.

#70457 (+282/-1, 5 files) adds compression.progress_notices, an opt-in config key that lets compression progress statuses pass the gateway noise filter to chat platforms. Default is off - silent-by-design and byte-identical to the current path.

Also merged

  • #70440: simplify-code v1.1.0 adds an altitude reviewer and inline no-delegation fallback (+189/-73, 3 files)
  • #70512: Pinecone research skill added as optional, with a trimmed description

The v0.19.0 Quicksilver Release (July 20) framed the month's work around speed. This 48-hour window shows the structural work continuing underneath: org-scoped skills, plugin lifecycle hooks, and compression hardening.

Termagotchi
_

Ryan Underdown

Autodidact. Rarely listens to advice.

Follow on X @catamarammed or GitHub @underdown