v0.31.4 — agent-stack

• agent-interop and agent-harness name their subject within the first 57 characters (hub
  board row B-285; Hermes shows only that many). agent-interop was "Use when an agent talks to
  anything outside its process: MCP server or client, A2A " and is now "Use when an agent speaks
  MCP, A2A or a gateway beyond its"; agent-harness was "Use when the question is what the agent
  is told, not how its loop is wired: system" and is now "Use when deciding an agent's system
  prompt and tools, not". Lengths 283 and 296; the quoted triggers and the Not-for boundaries are
  unchanged.

v0.31.3 — agent-stack

• Every skill description now fits a budget of about 300 characters (hub board row B-255 (a)):
  Claude Code lists every installed skill's description each session against a shared budget of
  about 30,000 characters, and the operator's listing measured 51,178, so less-used skills lost
  their descriptions. Measured with whitespace collapsed: agent-evals 422 → 294, agent-harness
  445 → 291, agent-interop 368 → 299, agent-orchestrator 458 → 299; total 1693 → 1183.
• Each description keeps its router sentence, its boundary against sibling skills, the opt-out
  phrase and a short English and Russian trigger list. Every quoted trigger the description
  dropped is kept in the body's ## Triggers section, so the family hook's words stay advertised
  (test/advertised_check.js in the hub).

v0.31.2 — agent-stack

• test/family_gate.py adds # diverges: none under # shared-mechanism: family-gate; the hub's
  cross-repository check requires both lines. No behaviour change.

v0.31.1 — agent-stack

• test/family_gate.py opens with # shared-mechanism: family-gate, as the hub's
  cross-repository check requires of every copy of a shared mechanism. No behaviour change.

v0.31.0 — agent-stack

• test/family_gate.py, the family's canonical member gate (hub B-108, B-109): the board
  parses (every row as wide as its header, no id twice) and release.yml keeps the tag-ancestry
  check (fetch-depth: 0, git merge-base --is-ancestor), each watched failing on a plant by
  --self-test. Wired into the gate; a byte-for-byte copy of the hub's member-gate/family_gate.py.
• agent-evals references/support-quality.md: the support quality standard v1 for agents
  that answer people (issue #56). The per-turn record (A1); four metric families per channel
  per closed UTC day, a metric not computed is "not measured", never 0 (A2); first-reply p90
  targets per channel kind: private chat under 30 s, mail under 4 h, store reviews under 24 h,
  group chats within the agent's come-back window; first reply measured from the person's
  message to the sent reply; rubric v1 and why factual accuracy is not the judge's job (A3);
  the judge, from another vendor, alerting on a check only after 80% human agreement (A4);
  support.quality_daily, counts and codes only (A5); autonomy without blocking (A6);
  portability (A7). Plus the weekly assessment loop (findings with cause class, owner,
  action and priority, compared by finding id), the checks that caught real defects, and
  one deployment's baseline labelled as a measurement, not a target. Routed from the
  agent-evals references table and an agent-orchestrator checklist line.
• test/validate.py: check_support_standard_keeps_its_parts (four self-tests) fails
  when the standard loses a part heading A1 to A7, a channel kind's target row, or the
  "never 0" rule in A2 or A5.
• README: the reference count reads 32 (it read 27 while 31 shipped; agent-harness 10,
  not 7).

v0.30.2 — agent-stack

• The verification ledger records this version: 0.30.1 shipped with the ledger still headed by the version before it. No skill changes.

v0.30.1 — agent-stack

• agent-interop/references/gateway.md cited files and a commit inside a private repository of the operator's workspace. The citations now name the record without its private path. The trust tiers, money limits and human gates are unchanged.

v0.30.0 — agent-stack

• Every skill description is a router of 368 to 458 characters (was 929 to
  939; total 3,742 to 1,693). The operator's Claude Code skill listing measured
  29,995 of about 30,000 characters for 134 skills, and 81 skills showed no
  description; a plugin skill cannot be hidden, so length is the only lever. Each
  description is one "Use when" sentence, four or five English and Russian
  triggers, and a "Not for" boundary naming the sibling; agent-orchestrator
  carries the refusal phrase. Triggers: takes a colon inside the folded scalar.
• Every removed trigger phrase moved verbatim into a ## Triggers section of
  its SKILL.md body (8, 5, 11 and 9 phrases).
• agent-orchestrator keeps its scope list in the body. The description's
  "Covers" list (the saga across database and provider API among it) was the only
  place the body named that scope, and fix-as-01.01 reads it; it now sits under
  ## Triggers.
• test/validate.py: the member ratchet moves from 940 to the family soft cap
  of 500, and a description without a "Not for" boundary fails.
  The CI ratchet plant moves with it (500 passes, 501 fails on the ratchet), and
  the over-long plant pads past 1024 again.

v0.29.0 — agent-stack

• agent-orchestrator references/governance.md gains an incident checklist. Six
  published incidents, each with its primary source (read 2026-10-19) and the control
  this pack draws from it: egress closed at the DNS resolver with an alert that stops
  the run; secrets split into pieces to evade scanners and uploads to public file
  hosts; compaction summaries that carried instructions the brief never gave; agents
  submitting third-party forms and working around tool limits; found credentials and
  a found wallet used without permission; Docker daemons exposed on port 2375 with LLM
  provider keys as the attacker's first target.
• agent-evals §6: every task is shown solvable before it scores an agent. A
  reference solution that passes all graders, a 0% pass rate across many trials read
  as a broken task, and a task with no passing reference reported as TEST_ERROR.
  Sources: Anthropic's eval guide and METR's investigation, where an estimated 30-40%
  impossible targets drove agents to cheat the scorer. New checklist line.
• agent-evals references/multi-agent.md: "Goal met through a prohibited effect"
  seed case, with the published precedent of agents invoicing strangers and mass
  emailing while chasing a revenue goal.
• test/validate.py: check_sourced_safety_rows_keep_their_sources (with four
  self-tests) fails when an incident row loses its id, row or https source, when the
  section loses its read date, or when the solvability rule or the seed case loses
  its source. Watched failing with three problems before the doctrine landed.

v0.28.2 — agent-stack

• Resume only an owned stopped worker through the host's supported operation.
  A quiet stream does not prove a stall; one failed lane does not authorize
  stopping siblings or duplicating their in-flight work.
• Reconcile uncertain outward effects before replay. Run IDs and unchanged
  arguments do not guarantee cached completion or preserved context.
• Record finite attempt and deadline bounds, including host retries, and retain
  provider Retry-After/reset timing. Recheck task authority and ownership before
  each attempt; preserve the best partial when the series ends.
• Keep independent work moving while a watcher handles the eligible retry.
  Billing, sign-in, consent and secret-entry gates remain human; no account/key
  switching or cap bypass. Focused text controls cover known unsafe instructions,
  without claiming observed model or scheduler behavior.