feat(omo-opencode): render background task notifications as markdown

[Zaldaryon]

Summary

• Background task notifications are written in markdown, and a client that renders only assistant output prints them verbatim. The person watching the session reads the raw <system-reminder> block, **ID:**, backticks and all, sitting next to assistant text that is rendered.
• The notification part now carries metadata.render set to "markdown", which lets a client render it. Clients that ignore the hint behave exactly as they do today, and the model receives the same text either way.

Changes

• packages/utils/src/internal-initiator-marker.ts: adds createRenderedInternalAgentTextPart, which is createInternalAgentTextPart plus metadata: { render: "markdown" }. The existing helper is untouched, so task prompts sent to child sessions keep their current shape.
• packages/omo-opencode/src/features/background-agent/parent-wake-prompt-dispatch.ts: sendParentWakePrompt builds the notification part with the new helper on both the reply and the no-reply path.
• Tests for the helper, for the hint surviving withInternalNoReplyMarker, and for the part that ParentWakeNotifier actually dispatches.

Why the hint and not synthetic: true: synthetic hides the part, which is right for text meant only for the model. A completion notice is also for the person watching, so it needs to stay visible and be readable.

QA & Evidence

• What was tested: bun run typecheck
  Observed result: exit 0, no diagnostics
  Artifact: console output
  Why sufficient: covers the new export and both call sites across every workspace project

• What was tested: bun run build
  Observed result: build: all steps completed, exit 0
  Artifact: console output
  Why sufficient: the generated bundles rebuild cleanly with the change; regenerated dist files are deliberately left out of this PR

• What was tested: bun test packages/omo-opencode/src/shared/internal-initiator-marker.test.ts packages/omo-opencode/src/features/background-agent/parent-wake-notification-render-hint.test.ts
  Observed result: 27 pass, 0 fail, 54 expect() calls
  Artifact: console output
  Why sufficient: the five new cases pin the hint on the helper, through the no-reply marker, and on the part ParentWakeNotifier dispatches, next to the existing marker cases that guard the untouched helper

• What was tested: the root bun test suite, on this branch and on the parent commit
  Observed result: 15735 pass, 43 skip, 9 fail here against 15729 pass, 43 skip, 10 fail on the base. No test fails only on this branch
  Artifact: the failing test names and the diff are in a comment on this PR
  Why sufficient: it separates pre-existing environment failures from anything this change could have caused

Risks & Residuals

• A client that does not read metadata renders the notification exactly as it does now. Mitigated: the hint is additive and nothing branches on it inside this repo.
• createInternalAgentContinuationTextPart sets its own metadata for compaction. Not applicable: it builds on the untouched base helper and is synthetic, so it is hidden regardless.
• The reader side does not exist upstream yet. Accepted: feat(tui): let injected text parts opt into markdown rendering anomalyco/opencode#43319 adds it, and this change is useful on its own the moment any client honors the hint. Neither PR depends on the other landing first.

Automated Checks

bun run typecheck
bun run build
bun test

Related Issues

Reader side of the same problem: anomalyco/opencode#43319 and anomalyco/opencode#43318.

[github-actions]

All contributors have signed the CLA. Thank you! ✅
_{Posted by the CLA Assistant Lite bot.}

[Zaldaryon]

I have read the CLA Document and I hereby sign the CLA

[Zaldaryon]

I have read the CLA Document and I hereby sign the CLA

[Zaldaryon]

Root suite result, as promised in the PR body.

On this branch: 15735 pass, 43 skip, 9 fail across 15787 tests in 2064 files.

Nine failures sounded bad, so I ran the same suite on the parent commit and diffed the failing test names. The base has 10 failures. Every failure on this branch also fails on the base, and no test fails only here:

$ comm -23 branch.fails base.fails    # unique to this branch
(nothing)

The eight shared failures, deduplicated:

#given PowerShell shell detection in injectGitMasterConfig > #then emits $env: prefix syntax in pwsh code block
injectGitMasterConfig > uses PowerShell routing when PowerShell is detected
shell-env > detectShellType > #given PSModulePath is set without SHELL #then returns powershell
shell-env > detectShellType > #given Windows platform without PSModulePath #then returns cmd
#given the generated Codex installer #when release versions are synchronized #then its embedded package version matches the root release version
omo-ai payload verifier > #then it passes with no missing-artifact error
/doctor > #given repeated model-not-found reflection failures #then reflection health and remediation are reported
claude-code-agent-loader > loadOpencodeGlobalAgents > returns empty object when pointed at dir without agents/

Four are shell detection asserting Windows behavior, two check packaging artifacts that a fork checkout does not carry, and two are assertions on local state. My environment is WSL2 on Linux with Bun 1.3.14, so I read all eight as environment rather than as anything this branch touches. Say the word if any of them is meant to pass on Linux and I will look closer.

The tenth failure, preflightMemoryModels > #given a launcher whose grandchild holds the output pipes, failed on the base and passed here. It looks timing dependent rather than related to this change.