feat: add native session goals

[jorgitin02]

Issue for this PR

Closes #27167

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Adds native per-session goals to OpenCode. Goals are persisted server-side, exposed through session HTTP APIs and the generated JS SDK, and shared by the TUI and app. The runtime injects defensive hidden goal context, accounts token and wall-clock usage, auto-continues active goals only when idle, pauses on abort, and budget-limits when token budget is exhausted. Model tools are limited to reading, explicit create, and marking complete.

How did you verify your code works?

• cd packages/opencode && bun typecheck
• cd packages/opencode && bun test --timeout 30000 test/session/goal.test.ts
• cd packages/opencode && bun test --timeout 30000 test/session/goal.test.ts test/server/httpapi-session.test.ts test/session/prompt.test.ts --test-name-pattern 'SessionGoal|session goal routes|goal continuation|goal tools'
• cd packages/opencode && bun test --timeout 30000 test/session/snapshot-tool-race.test.ts
• cd packages/app && bun typecheck
• cd packages/app && bun test --preload ./happydom.ts src/components/prompt-input/submit.test.ts src/context/global-sync/event-reducer.test.ts src/context/global-sync/session-cache.test.ts
• cd packages/sdk/js && bun typecheck
• cd packages/desktop && bun typecheck
• git diff --check
• Pre-push hook ran bun turbo typecheck
• Ran codex review --base refs/remotes/upstream/dev; fixed two P2 budget-accounting findings and added regression tests.

Screenshots / recordings

Not attached. The UI changes are compact slash/status surfaces; app and TUI behavior is covered by targeted tests.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[TaQuangKhoi]

so much of code changes

[jorgitin02]

@codex review

[Cheny5863]

⚠️ 请注意：此 PR 修改了系统 Prompt，这会导致 DeepSeek 的 Prompt 缓存（Context Caching）失效/命中失败。

English:
⚠️ Please note: This PR modifies the system prompt, which will cause DeepSeek's prompt caching to miss.

[Cheny5863]

There is another problem:

False “no progress” auto-pause stopped goal continuation prematurely

Summary

When a session had an active goal, autoContinueGoal could incorrectly decide that a synthetic continuation turn made no progress, immediately set the goal to paused, and stop injecting further continuation prompts. From the user’s perspective, the agent “stopped on its own” or /goal resume appeared to do nothing (goal went back to paused within ~200ms).

This was not an LLM/provider failure. The goal loop exited before a new LLM turn was enqueued.

Root cause

After a synthetic goal-continuation user message (Continue working toward the active session goal.), auto-pause only inspected the latest assistant message. Pure text responses (e.g. a status summary asking “need to continue?”) do not count as progress — only tools/patches/subtasks do:

    const assistantMadeGoalProgress = (message: MessageV2.WithParts) =>
      message.info.role === "assistant" &&
      message.parts.some((part) => {
        if (part.type === "patch" || part.type === "subtask") return true
        if (part.type !== "tool") return false
        if (part.tool === "get_goal") return false
        return part.state.status === "completed" || part.state.status === "running" || part.state.status === "pending"
      })

Original buggy condition (simplified):

// BEFORE fix — only looked at the latest assistant
if (
  isGoalContinuationMessage(latestUser) &&
  latestAssistant.finish &&
  !assistantMadeGoalProgress(latestAssistant)  // ← too narrow
) {
  await goals.update({ sessionID, status: "paused" })
  return
}

Failure scenario

Typical message sequence:

1. Synthetic user: Continue working toward the active session goal.
2. Assistant A: runs bash / other tools (real progress)
3. Assistant B (latest): text-only summary, e.g. ## Status summary… need to continue?

On idle, autoContinueGoal (or /goal resume → continueGoal) saw:

• Latest user = old synthetic continuation message
• Latest assistant = B (text only) → assistantMadeGoalProgress === false

So it logged auto_pause_no_progress and paused the goal — even though assistant A in the same stretch had already used tools.

Observed in logs:

goal.continue reason=start
goal.continue reason=auto_pause_no_progress latestAssistantProgress=false
goal.status from=active to=paused

No inject_synthetic_prompt, no new LLM call.

Why normal chat still worked

A real user message creates a new turn and does not hit this auto-pause branch the same way. The bug specifically affected the goal idle-continuation loop, not arbitrary user prompts.

Fix

Treat a continuation stretch as “no progress” only if none of the last N messages (N=5) contain tool/patch/subtask progress:

        if (
          Option.isSome(latestUser) &&
          Option.isSome(latestAssistant) &&
          latestAssistant.value.info.role === "assistant" &&
          isGoalContinuationMessage(latestUser.value) &&
          latestUser.value.info.id < latestAssistant.value.info.id &&
          latestAssistant.value.info.finish &&
          !recentProgress
        ) {
          yield* goalContinueLog(sessionID, "auto_pause_no_progress", messageContext)
          yield* goals.update({ sessionID, status: "paused" }).pipe(
            Effect.catchIf(NotFoundError.isInstance, () => Effect.void),
            Effect.ignore,
          )
          return
        }

Where recentProgress is derived from the last 5 messages:

    const recentMessagesMadeGoalProgress = (messages: MessageV2.WithParts[]) =>
      messages.some((message) => assistantMadeGoalProgress(message))

After the fix, resume logs show recentProgress=true and inject_synthetic_prompt instead of immediate re-pause.

[Cheny5863]

Used it for a bit and everything looks good so far—no other issues found. The program runs great, thanks for the PR!

[jorgitin02]

⚠️ 请注意：此 PR 修改了系统 Prompt，这会导致 DeepSeek 的 Prompt 缓存（Context Caching）失效/命中失败。

English: ⚠️ Please note: This PR modifies the system prompt, which will cause DeepSeek's prompt caching to miss.

Both issues should be fixed now.