Competitive analysis
August 20, 2026 ยท View on GitHub
Research date: 2026-08-20
This review compares product boundaries and public APIs. No third-party source implementation is copied.
DeepSeek Harness ecosystem
| Project | Primary job | Overlap | Gap relative to dsh-tool-squeeze |
|---|---|---|---|
| DSH spill policy (official) | Store oversized plain text and show a head/tail preview | Prevents huge inline results | Content-agnostic; a critical error in the middle can be absent from the preview |
| DSH tool-result pruner (official) | Prune old tool results during compaction | Reduces later context pressure | Acts after results have already entered context; not format-aware |
| billion-context-dsh | Model-driven active context pruning as a compaction backend | Context reduction | Conversation/context lifecycle scope, not deterministic pre-context tool-result processing |
| dsh-context | Inspect context composition, history, events, and token cost | Observability | Observe/manage rather than transform one tool result before model consumption |
| dsh-session-audit | Analyze completed session actions, failures, and usage | Reports tool behavior | Post-session analytics, not inline result compression |
| dsh-usage-stats | Usage trends, heatmaps, model breakdown, export | Token accounting | Measures usage after the fact |
| dsh-agent-budget | Enforce durable token/deadline budgets | Controls total spend | Admission control rather than output transformation |
| dsh-trace | Export turns, model steps, and tool calls to telemetry | Observability | Records events rather than shrinking content |
Public references:
- https://github.com/deepseek-ai/deepseek-harness
- https://github.com/bowenliang123/dsh-context
- https://github.com/bwndlct/dsh-session-audit
- https://github.com/lanlandeli/dsh-usage-stats
- https://github.com/vibeinging/dsh-agent-budget
- https://github.com/vibeinging/dsh-trace
Other agent ecosystems
Headroom
Headroom is a broad compression platform: library, proxy, MCP server, multiple content routers, retrieval, optional learned/ML methods, cache alignment, memory, and cross-agent context. It demonstrates demand for evidence-aware compression and reversible retrieval, but it is much broader than one native Harness hook. The dsh plugin differentiates through a tiny TypeScript-only, DSH-native, deterministic path with no proxy or Python service.
Reference: https://github.com/headroomlabs-ai/headroom (Apache-2.0).
token-saver
token-saver wraps command execution and supplies many command-specific parsers. It has strong CLI specialization and reproducible compression baselines. The MVP deliberately does not compete on parser count: it handles any Harness tool, including MCP and non-shell results, at the official result lifecycle seam.
Reference: https://github.com/ppgranger/token-saver (Apache-2.0).
MCP result compression and context pruning
MCP compression servers generally require an explicit tool call or a proxy layer and may add another service boundary. Session compaction systems operate on accumulated history. Both can be useful, but neither guarantees that the first giant result is transformed before its first model-context insertion.
Product boundary
dsh-tool-squeeze is not another conversation compactor. Its narrow promise is:
Deterministic, evidence-preserving compression of a completed tool result before that large presentation repeatedly consumes model context.
Its defensible value is the combination of:
- native
tools/post-executeintegration; - preservation of canonical tool values;
- official spill-store retention of the original presentation;
- conservative log/JSON/HTML/text processors;
- visible omission and token-estimate notices;
- fail-open behavior and per-tool bypass;
- no LLM, proxy, external service, or core patch.
The project should not claim universal losslessness. Content compression is necessarily lossy; safety comes from conservative activation, explicit notices, evidence rules, and a retrievable original.