Repository navigation
Releases
Releases · igorxmel/agent-offload
v0.3.0
Added
offload budget: today's executor requests per provider, and where a quota stop was
last seen.- Executor usage is now exact per delegation and per attempt, read from each attempt's own
OpenCode session; overlapping delegations no longer report the same combined numbers. --max-steps N(andmax_stepsin the config): an attempt is stopped after N model
requests and its work is kept as partial, so one attempt cannot spend a whole free
daily allowance.--modelis repeatable: the named models are tried in order.--fallbackgoes on to
other acceptable models after them.
Changed
- With
--model, only the named models are tried. Before, a failure moved on to models
nobody named, which spent the shared free allowance of a parallel delegation.
Fixed
- Read-only roles no longer count the person's own uncommitted edits as the executor's
(#1, reported by @mattabate). - Work done before any executor failure is kept on
<branch>-partial-N, and the verdict
shows the real error line instead of OpenCode's banner. - OpenRouter's per-model "Daily limit reached" is recognised as a quota.
v0.2.0 - parallel, free review, survives free tiers
Everything in this release came out of running the kit on real work: a Rust + Kotlin to-do app with LAN sync, a Windows app and an Android app, built by free models behind hidden checks they never saw.
Highlights
- Parallel delegations - each delegation runs in its own git worktree slot; your tree is never checked out, build output stays warm.
- Free first review -
--role review --on <branch>has a free model read the diff and return findings. - Hidden checks (
--holdout) and resume with a numbered list of discrepancies (--resume). - Stays free - only models with a listed price of zero that can call tools; free-tier keys via
offload allow <provider>/*. - Survives providers - quota stops skip the provider, stalls are killed, and work done before a timeout, stall or quota stop is kept on
<branch>-partial-N. - Executor usage (steps, tokens, cost) in every verdict.
Measured
A/B on one real task, same model, fresh sessions: the orchestrator spent ~20% less with the kit (2.6x fewer output tokens) at about 7x the wall time; both arms passed 101/101 hidden checks. The free budget, not the kit, is now the limit - see the README.
Full list: CHANGELOG.md.