Implement Goals as a thin native /goal task mode

[integry]

Context

This supersedes the unreleased goal control-plane/runtime/UI implementations in #2003-#2015 and PRs #2020, #2059, #2065, #2067, and #2070.

A goal is not a ProPR-authored plan. It is one existing ProPR coding-agent task whose first agent input is the native slash command /goal <objective> (or the provider's documented non-interactive equivalent). The coding agent owns planning, decomposition, parallel subagents, issue/PR structure, status answers, and completion.

Implement this as a thin execution mode over ProPR's existing task machinery. Do not build a second orchestration system.

Reuse these existing primitives

Use and extend the current:

• AgentRegistry and each provider's executeTask container path;
• worktree allocation, branch handling, credentials, image/config mounts, container naming, callbacks, abort fencing, and queue de-duplication;
• worker task state/history metadata for task, session/conversation, container, worktree, model, and lifecycle state;
• Redis/container streaming, taskWatcher, conversation parsers, /api/task/:id/live-details, execution logs, todos, and token usage;
• existing stop/cancel and follow-up mechanisms;
• existing realtime task delivery and polling fallback;
• existing LLM metrics and task duration/history.

Do not add a goal-specific supervisor, lease system, WAL/event/output database, JSON/JSONL authority, socket protocol, container manager, provider-effect ledger, planner, node graph, dependency scheduler, merge scheduler, or synthetic checklist.

Minimal durable model

Add only a small owner-scoped goal record (or an equivalent first-class task discriminator) that references the existing task execution and stores fields not already authoritative elsewhere:

• goal ID, owner, repository, objective, branch/worktree identity;
• immutable selected agent and current/requested model;
• optional max-parallel-tasks and Ultrafix preferences passed as policy to the coding agent;
• desired host state (running, paused, cancelled) and final result state;
• current existing task ID plus provider session/conversation identity;
• passively discovered final PR/artifact references.

Existing task history/output/metrics remain authoritative. Do not copy them into goal-specific tables.

Native execution contract

• Add a goal execution mode to the existing agent task options/runner.
• Initial stdin/input is the exact native goal command with the user's objective. Do not wrap it in ProPR implementation/planning prompts.
• Goal mode preserves provider session state. It must not use current one-shot flags such as Claude --no-session-persistence or Codex --ephemeral.
• On retry, host restart, or container loss, enqueue the same goal task against the same worktree and resume the persisted provider session/conversation. Never start a second goal silently.
• Claude, Codex, and Antigravity use their documented native goal/session-resume capabilities in the existing agent containers. Capability-probe the pinned binary and expose goalCapable: false if a provider/version does not actually support native /goal; do not emulate it with a ProPR loop.
• Persist session/conversation identity through the existing early-session callback before claiming crash recovery.
• Pause stops further native continuation at a safe supported boundary while retaining the session/worktree. Resume re-enqueues the same session. Cancel is terminal and distinct.
• Corrective and canned “what's done?” / “what's left?” messages are ordinary inputs to the same provider session and their replies are ordinary task output.
• A supported model change becomes the requested model for the next safe continuation/resume and records the effective model reported by the provider.
• Parallelism and Ultrafix are text/structured policy supplied once to the coding agent; ProPR does not schedule the resulting work.

API and UI

Add a dedicated Goals navigation/list/create/detail surface, but compose it from existing task APIs and components wherever possible.

• Create selects repository, agent, model, objective, optional parallelism, and Ultrafix preference, then starts exactly one native goal task.
• List/detail joins minimal goal metadata with existing task history and live-details output.
• Terminal output, provider-native todos/checklist, current task, and token usage come directly from existing task parsing.
• Reuse the existing realtime task stream and polling fallback. Do not define a second goal event/replay/socket protocol.
• Controls expose pause/resume/cancel, supported model changes, canned status messages, and free-form corrections.
• Show elapsed/active/paused time from goal/task history and passively associated issues/PRs; do not manufacture an issue hierarchy.
• The final PR is the coding agent's draft epic PR (or its single PR for a small goal), recorded for human review and never auto-merged by goal mode.
• Enforce the same owner/repository authorization used by tasks on every read and mutation.

Recovery

On worker/daemon startup, find nonterminal goal records whose existing task/container is no longer live and re-enqueue a fenced resume using the saved provider session and worktree. Reuse current task/container liveness checks and duplicate-execution protection. Recovery must be idempotent and must not create a new worktree, branch, or provider goal.

Validation

• Unit/contract tests prove goal mode changes provider CLI arguments/input while the normal one-shot path is unchanged.
• Tests prove no ephemeral/no-persistence flags in goal mode and exact session resume for each supported provider.
• A worker/container crash test proves re-enqueue resumes the same goal/task/worktree/session once.
• Existing live-details tests prove terminal output, native todos, current task, and tokens render without a new event store.
• UI tests cover create/list/detail, pause/resume/cancel, model request/effective state, corrective/canned input, provider capability gating, ownership, and reconnect.
• Repository validation remains green.

Delivery

This PR targets 2002-epic-long-running-goals. Keep it reviewable and delete/reuse existing abstractions instead of importing code from the superseded goal branches. The epic branch will later produce the final draft PR to main.

[propr-dev]

🤖 AI processing has started for this issue using codex agent with gpt-5.6-sol model.

I'll analyze the problem and work on a solution. This may take a few minutes.

Processing Details:

• Agent: codex
• Model: gpt-5.6-sol
• Branch: 2071/gpt-5.6-sol-implement-goals-as-a-thin-20260902-1634-psl
• Base Branch: 2002-epic-long-running-goals
• Worktree: issue-2071-20260902-1634-gpt-5.6-sol-psl

🔍 Track Task Execution