[testikun]

Problem

OpenPI's subagent timing spans child Session history, individual executions, result delivery and the parent host's ability to continue. We need a current comparison of the mainstream designs before choosing lifecycle changes. Related: #196 and #512.

Value

Give the design discussion a source-pinned comparison of OpenCode V2, Kimi Code, Claude Code, Codex and DeepSeek Harness. Distinguish mechanisms already present on main from proposed changes and historical issue evidence.

Approach

Add a draft research record and index entry. Compare context, execution identity, foreground/background calls, continuation, notifications, authority and recovery. Recommend retaining Pi-native child Sessions, clarify Session/run/delivery/host boundaries, and record design ablations. This proposal is not an accepted Decision.

OpenPI is pinned to main 31b2d045; four peer repositories are pinned to exact commits, and Claude Code is scoped to official rolling documentation. Forty downloaded source files and sixteen peer permalinks were verified against bounded receipts outside Git. No peer harness or model benchmark was executed.

Validation

• bun run check: passed, including a run with Node 24.22.1 and Bun 1.3.14.
• bun run test: passed in an isolated PI_CODING_AGENT_DIR; Node: 2,421 passed / 8 skipped; UI: 107 files / 1,640 passed.
• The first run inherited non-default personal UI settings and failed six renderer/footer assertions. All eleven cases in the three affected files pass with isolated settings. No implementation, test assertion or personal configuration was changed.
• Source receipts, reference definitions, documentation contract and git diff --check: passed.

Impact

Documentation only. User-visible runtime behavior, model-facing tools/context, runtime lifecycle, persisted configuration/data and compatibility: unchanged. The recommendations remain draft; the reported timing symptom has not been reproduced on the pinned main.