跪拜 Guibai
← All articles
AI Programming

SoduWork Carries Pi, Qwen Code, and Google Harness² Through One Conversation: Rules Survive, Billing Does Not

By 硅徒 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Teams juggling multiple CLI agents can stop re-explaining business rules at every model switch, but the handoff context is hard-truncated, ACP-style harnesses produce no billable usage data, and a Linux-only CPU-sampling assumption silently stalls the scheduler on macOS, so relay cost and reliability stay open questions for production.

Summary

A relay experiment ran three AI harnesses inside one SoduWork conversation (session `01a1d12b-7ee4-71b4-9ea5-3e5667a1c7c4`) against 120 cloud gateway fault tickets. Pi 1.4.0 cleaned the data and set four rules (R1–R4) in 8.59s; Qwen Code 0.21.8 produced a regional scorecard in 16.85s that matched 4/4 rules without re-pasting thresholds; Google Harness² 0.2.0 finished a 190-character decision brief in 27s after its first attempt was tripped by a path-safety guardrail.

The handoff works through SoduWork's segmented-conversation bridge: it snapshots the old harness, forces the same working directory, and auto-injects a truncated context summary when the new harness starts. The summary keeps only the trailing 28,000 characters of message text and the last 12 tool records, so early details can fall away.

Cost traces show the relay is serviceable but incomplete. Pi's leg consumed 71,043 tokens (76.4% cache-hit) at about ¥0.048; Google Harness² needed 119,312 tokens with just 6.1% cache-hit across 11 model requests (¥0.232). Qwen Code's ACP events are marked `contextOnly` and sit outside the billing path, so that leg's real cost is unknown. A macOS-only scheduler bug also stalled the first switch: loadavg-derived CPU estimates capped at 100% and blocked `claim_next` even at zero concurrency.

Takeaways
— SoduWork's segmented-conversation bridge keeps Pi 1.4.0, Qwen Code 0.21.8, and Google Harness² 0.2.0 working in one session with a shared working directory and an auto-generated handoff summary.
— Pi filtered 120 tickets to 96 in 8.59s (about ¥0.048), removing 24 change-freeze-window tickets with zero score mismatches.
— The first Pi→Qwen Code switch stalled because macOS has no /proc/stat; loadavg 20.59 on a 10-core machine capped the estimated CPU at 100% and blocked claim_next even with no tasks running. A local zero-concurrency patch unblocked the run.
— Qwen Code produced regional_scorecard.json in 16.85s meeting 4/4 rules: total impact 3195.12, 2 RED_FREEZE / 1 AMBER_WATCH / 1 GREEN_NORMAL, with dual leaderboard output per R3.
— Google Harness²'s first attempt was blocked by its path-whitelist guardrail during the planning stage; a retry that dropped the file-write and output the brief in the body took 27s and 8 model requests.
— Harness² sent 11 model requests across both rounds: 119,312 tokens at only 6.1% cache-hit rate, roughly 2.4x Pi's rough-filter cost on the successful round and 4.8x counting the failed round.
— Qwen Code's usage events are contextOnly: 53,016 context tokens, output 0, and meta.json usage 0 — none of it billable, so that leg's real cost is unaccounted.
— The handoff summary brief() truncates at 6,000 characters per message with the trailing 28,000 kept, and at 12 tool records with 1,000 characters per argument and the trailing 6,000 kept.
— The official comparison case ran a 5-round relay where all three round-5 harnesses kept 4/4 accumulated rules without re-stating them; a 336-character manual summary in a fresh conversation also kept 4/4.
Conclusions

The 4/4 pass proves state carried over, not that a fresh harness independently re-derived the rules: R1's filtering and R2's scoring already lived in Pi's JSON, so Qwen Code validated continuity of artifacts rather than reasoning from scratch.

The macOS scheduler stall is a platform assumption that will keep biting local-dev AI tooling: Linux-only /proc/stat logic silently inflates CPU on Darwin and blocks the scheduler even when nothing is running.

ACP-style harnesses like Qwen Code currently emit contextOnly usage events that the billing path explicitly excludes, so a relay's real cost can vanish from the ledger even as context tokens pile up.

Google Harness² sent 8 model requests to produce a 190-character brief at an 11.8% cache-hit rate, and its own path guardrail killed the first attempt; heavy reasoning pipelines inflate both the cost and the failure modes of a relay.

Concepts & terms
Segmented Conversation
SoduWork's mechanism for switching AI harnesses inside one conversation: it snapshots the previous harness's messages, tools, and usage receipts, namespaces tool-call IDs per session, forces the same working directory, and injects a truncated progress summary into the new harness's context.
Harness
An adapter or agent base in SoduWork that wraps a model behind a common interface, such as the desktop-native Pi, an external coding CLI like Qwen Code connected via ACP, or a reasoning pipeline like Google Harness² that runs seven-stage text reasoning before an independent executor.
ACP (Agent Client Protocol)
The protocol SoduWork's Runner uses to connect external coding agents like Qwen Code. Its context updates arrive as contextOnly events that the billing path explicitly excludes, so they cannot be treated as billable token consumption.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗