LightRSI supports Claude Code, Codex, OpenClaw and DeepSeek Harness. Two other popular open-source coding agents are:
How it would plug in. pi has an in-process extension API, so no proxy is needed. The adapter would reuse the shared code from #95 and #96:
| TokenPilot feature |
pi hook |
| Stable prefix |
before_agent_start system-prompt sections |
| Reduction |
the request-local context hook |
| Eviction (opt-in) |
turn_end context edits, applied before pi's own compaction |
| Recovery |
a native memory_fault_recover tool, since pi has no MCP |
Early numbers. I ran a 4-turn coding session with llama-server, Qwen3-1.7B and a 32k context window, on a prototype. Prompt tokens processed went from 37,465 to 4,994 (−87%). Without the adapter, pi had to compact the conversation; with it, it didn't. I'll re-measure on the final code and put the results in the PR.
Plan: one PR, once #96 has merged. Would you be happy to take it?
LightRSI supports Claude Code, Codex, OpenClaw and DeepSeek Harness. Two other popular open-source coding agents are:
@earendil-works/pi-coding-agent): this issue proposes adding it.How it would plug in. pi has an in-process extension API, so no proxy is needed. The adapter would reuse the shared code from #95 and #96:
before_agent_startsystem-prompt sectionscontexthookturn_endcontext edits, applied before pi's own compactionmemory_fault_recovertool, since pi has no MCPEarly numbers. I ran a 4-turn coding session with llama-server, Qwen3-1.7B and a 32k context window, on a prototype. Prompt tokens processed went from 37,465 to 4,994 (−87%). Without the adapter, pi had to compact the conversation; with it, it didn't. I'll re-measure on the final code and put the results in the PR.
Plan: one PR, once #96 has merged. Would you be happy to take it?