dsh-call-shrink

In a DSH agent environment a single model turn often spawns multiple parallel sub-tasks—tool calls, title generation, repeated streams—each of which triggers a real upstream API call. When the requests are near-duplicates or identical, tokens and billable usage balloon for no additional value. dsh-call-shrink intercepts these calls at the Stream layer and compresses them to the minimum number of upstream requests through deduplication, caching, and batch merging, without changing the conversation outcome.

The core is a three-tier compression pipeline. First, a stable-fingerprint cache replays historical responses (24-hour TTL, 4096 entries, disk-persisted). Second, in-flight requests with the same fingerprint share a single upstream result. Third, a batch-merge window collects up to N independent requests within 10 seconds and bundles them into one upstream call—no item limit, no character-based chunking. The merge protocol wraps each sub-conversation in [[TASK n]] delimiters; the model emits structured JSON per task, and any parse failure gracefully falls back to sequential retries so no answer is lost. Title-generation requests are short-circuited to zero calls, and a cost-awareness constraint is injected into the system prompt to encourage the model to batch parallel tool calls proactively.

Built for developers running DSH agent workloads with multi-session or parallel sub-agent patterns who care about API spend. After installation, a real-time panel in the session header displays today's usage against budget, dedup hits, cache hits, and the number of calls saved—pairing with a daily budget dashboard so every upstream call earns its keep.