dsh-agent-frugality

Long-running multi-agent sessions quietly accumulate three mechanism-level failure modes: subagents re-read the same content without shared memory, wasting tokens; soft prompt rules vanish after context compaction; and expensive frontier models end up reviewing their own work at full price. dsh-agent-frugality addresses each of these with mechanism-level defenses backed by independent evidence.

The plugin is built on four core capabilities. The read-ledger content-hashes every read-type tool call into a per-agent ledger, making duplicate reads visible and quantifiable. The immutable-core section re-renders frugality rules on every prompt assembly so context compression cannot erase them. The completion gate requires a host-configured verify command to exit 0 before a done claim is accepted, with a maximum of two anti-loop objections. The review lane delegates independent review to a cheap-model family, keeping the expensive main model from auditing its own output at full cost.

The design philosophy is measure first, intervene second: the ledger is on by default, dedup is off, so you see your real duplicate rate before opting in. The plugin has zero external dependencies, never blocks or rewrites agent output, and stores no file content in logs—only hashes and JSONL event records. It is suited for teams and solo builders running long multi-agent DSH sessions, agent plugin authors, and anyone who wants token spend to be an auditable number rather than a black box.