SkillWeave v1.5.0

SkillWeave v1.5.0 — execution that survives a crash

Long agent runs fail in the middle, and until now recovering one meant parsing transcripts.
This release reconstructs a run's execution graph from persistent context instead, bounds
token budgets per profile, and escalates irreversible decisions to an operator before any
state is mutated rather than after.

Highlights

• A crashed orphan, worker or coordinator is recovered from persistent context, not from
  transcript parsing.
• Token budgets are declared per profile and enforced, so a crash loop cannot double-charge.
• Irreversible choices escalate before the mutation, not after it.

Added

• Crash recovery — orphan, worker and coordinator recovery reconstruct execution DAGs from
  persistent context rather than brittle transcript parsing.
• Context limits and checkpointing — configurable token profiles (for example 120k
  no_new_task, 150k checkpoint, 170k stop) bound execution budgets.
• Operator agent escalation — operator_agent carries strict deterministic decision
  scopes; an irreversible choice escalates before state is mutated.
• Persistent policy — idempotent compensation, backoff and retry budgets, so a crash loop
  does not double-charge tokens.
• Remote workspace sandbox — distributed execution honours the skillweave-workspace
  contract without leaking orchestration state.
• Soak and endurance testing — multi-hour degradation limits, and a DoubleStartGuard
  against a second process starting over a live run.

Upgrade

pip install --upgrade skillweave. Token profiles ship with defaults; set them explicitly if
your runs are long enough to approach a context ceiling.

Full changelog: CHANGELOG.md · Diff: v1.4.0...v1.5.0