Skip to content

Track Oh-My-Pi 16.3.10 upstream release #1197

Description

@a5c-ai

Upstream version check

  • Agent: Oh-My-Pi / OMP CLI
  • Previous graph version: 16.3.4
  • Latest upstream version: 16.3.10
  • Package checked: @oh-my-pi/pi-coding-agent
  • Release notes: https://github.com/can1357/oh-my-pi/releases/tag/v16.3.10
  • Package/homepage metadata: npm view @oh-my-pi/pi-coding-agent@latest returned 16.3.10, homepage https://omp.sh, repository github.com/can1357/oh-my-pi
  • Published: 2026-07-06T07:52:59Z from GitHub release metadata

Release research

I checked npm latest for @oh-my-pi/pi-coding-agent and read the official GitHub v16.3.10 release notes. This supersedes the existing 16.3.6 tracking issue because upstream advanced through 16.3.10 before this daily run completed.

Real changes identified

  • Empty completion handling: Ollama/Ollama Cloud EOS-only completions now retry empty stops with a single output token before the agent loop can halt silently.
  • Strict tool fallback: Claude Sonnet 5 requests through feature-gated gateways such as Azure Foundry or OpenAI-compatible relays now classify strict-tool rejection and retry without strict tools, remembering the downgrade for the session.
  • Model discovery: LiteLLM rich discovery now ignores unusable sentinel placeholders and continues to /v2/model/info for real models.
  • TUI/subagent performance: running subagent HUD display is limited to 8 items with a summary; registry/observer updates are coalesced; nested subagent trees are capped per level; obsolete task partial-result spinner repaint loop is disabled.
  • Browser cleanup: browser run-end/tab-close aborts are observed and routine teardown aborts are downgraded to log lines rather than causing unhandled rejections that exit the shared session.
  • RPC/session stats: get_session_stats responses now include context-window usage so RPC clients can render context meters.
  • Vision model startup: cached llama.cpp vision models refresh /props metadata before being exposed as text-only.
  • IRC wake/wait: IRC-woken yielded subagents reset stale yield termination state and irc wait drains pending IRC asides before queued-interrupt aborts.
  • Terminal/process cleanup: terminal restore is registered on the normal CLI fatal path; expected cleanup errors can be marked and fatal cleanup has a 10s deadline.

Likely affected system surfaces

  • AgentVersion graph: add agentVersion:omp:16-3-10 with concrete retry, strict-tool fallback, TUI, browser cleanup, RPC, vision, IRC, and cleanup notes.
  • Stream/tool retry handling: strict-tool rejection fallback and empty-stop retry mean adapter retry classification should account for nonfatal strict-tool downgrades and empty completion retries.
  • Provider/model discovery: LiteLLM placeholder filtering and llama.cpp vision metadata refresh affect model catalog/discovery assumptions.
  • TUI/session observers: high-subagent rendering and observer update coalescing may affect snapshot tests or UI event expectations.
  • Browser tool lifecycle: browser teardown aborts should be treated as expected cleanup instead of fatal session errors.
  • RPC clients: get_session_stats now includes context-window usage, so RPC schemas/context meter renderers should be updated if they currently drop that field.
  • IRC integration: wake/wait semantics changed for yielded subagents and queued interrupts.
  • Process cleanup: postmortem expected-cleanup markers and 10s cleanup deadline affect crash/exit handling assumptions.

Install/package changes

No package-name or install method migration was published. Continue to use npm install -g @oh-my-pi/pi-coding-agent for the upstream package.

Migration steps

No user migration steps were published in the release notes.

Assimilation checklist

  • Add AgentVersion record agentVersion:omp:16-3-10 with evidence from the official v16.3.10 release notes.
  • Review OMP retry classification: model strict-tool rejections should downgrade/retry without strict tools rather than fail the session.
  • Review empty completion handling: Ollama/Ollama Cloud EOS-only empty stops should be represented as retryable before loop halt.
  • Update RPC/session capability metadata for get_session_stats context-window usage.
  • Review browser lifecycle handling so routine browser teardown aborts are not treated as fatal adapter errors.
  • Review TUI/subagent observer assumptions for bounded HUD display and coalesced progress repaints.
  • Review model discovery metadata for LiteLLM placeholder filtering and llama.cpp vision /props refresh.
  • Review IRC integration notes for yielded subagent wake and irc wait pending-aside behavior.

Metadata

Metadata

Assignees

Labels

agent-catalogAgent catalog and graph query packageagent-version-updateAgent CLI version assimilation trackingatlasAtlas graph and catalog packageeffort:mediumMedium implementation effortenhancementNew feature or requestgraph-updateAtlas graph update workpriority:mediumMedium priority issue with meaningful workflow impactready-for-devTriaged and ready for developmentrisk:mediumMedium implementation risk; requires focused validation

Type

No type

Fields

No fields configured for issues without a type.

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions