v0.10.0

• Add hr-progress-report for task activity, evidence-based milestone progress
  and remaining wall-clock time, with blockers, waiting conditions and unknowns.

• Use tables for comparisons and repeated records. Show scoped backlog and
  research items in priority order based on the task's goal, expected impact
  and decision value, with readiness and uncertainty stated separately.

• Add an opt-in release-batching modifier to batch approved compatible changes
  with a defined cutoff, preserve repository release requirements, and let
  urgent fixes ship without waiting for unfinished work.

• Add a configuration CLI for browsing and toggling modifiers and skills, choosing question preferences, and previewing validated changes before safe writes.

• Prefer isolation over elevated benchmark priority. Record effective scheduling
  settings equally across comparison arms, preserve deployment settings for
  representative results, and test priority changes in separate sensitivity
  pilots under the host's resource rules.

• Batch approved changes into releases instead of publishing a version for each
  pull request. Record changes under Unreleased until the batch is ready;
  urgent fixes can still ship independently. Review checks accept matching
  Unreleased entries.

v0.9.2

• Make text-only questions explicit: do not call question-card tools unless
  the question-cards modifier is enabled. Setup explains text-only and
  optional card formats before changing the configuration, including
  differences in availability and presentation across hosts, providers and models.
• Preserve complete text offers when cards are enabled, and remove the card
  modifier when a user changes to text only.

v0.9.1

• Observe compressed-swap activity and cgroup ownership when interpreting Linux
  benchmark pressure. Stable disk swap counters alone do not prove no swapping.

v0.9.0

• Rename all shipped skills with the hr- prefix. Update skill invocations,
  personal override directories and frontmatter names, exclusions and host
  links. Rule headings and modifier names are unchanged. No old-name aliases
  are shipped. See the upgrade guide.
• Refuse legacy exclusion names and ambiguous personal override names before
  composition writes output. skills.independent records generic personal
  skills intentionally kept alongside their prefixed counterparts.

v0.8.0

• Add five web UI skills for accessibility, colors, layout, typography and
  interface writing, with their supporting references and discovery metadata.
• Preserve both upstream MIT notices and pinned provenance. Mobile browser
  checks distinguish device emulation from hardware verification, and
  theme-color guidance accounts for platform support.
• UI review verdicts fit within the host's completion rules.

v0.7.3

• Align resource-counter boundaries with measured work. Keep post-run
  observations diagnostic and report admission and monitoring overhead
  separately from workload time. Batch compatible probes when they dominate
  elapsed time.
• When a local phase has unknown resource use or budget, contact the reservation
  owner, host coordinator or operator and record agreed limits before launch.

v0.7.2

• Reserve shared machines only for heavy local phases or deliberately isolated
  local performance measurements. Remote inference, light CLI work and remote
  waits do not need an exclusive slot. Keep timing conditions visible when
  evaluating remote agents and models.

v0.7.1

Swarmail priority guides when a receiver considers pausing its task.
High or urgent mail can justify saving work, handling an authorized
request and resuming. Normal or low priority leaves the pause decision
to the receiver based on content. Priority grants no additional authority.

v0.7.0

• Handoffs now reconcile transcript records and linked task ledgers with an
  offline checker. The receipt records source counts and gaps, item outcomes,
  owned and dependency pull requests, and commits without pull requests.
• The checker reports complete, partial or invalid mechanical coverage.
  It cannot find promises hidden in records classified as context, discover
  unreported sources, or decide whether an outcome is supported semantically.
  An unfinished receipt never delays a low-quota stop.

v0.6.5

• Attribute benchmark CPU, load and resident memory before treating them as
  competing work. Expected own utilization does not reject a run; actual
  resource pressure still does. Use a settled fresh interval between arms and
  investigate uncertain overlap before a decisive comparison.