Improve session navigation, model discovery, and background commands

[snimu]

Summary

• Simplify the agents view with compact model IDs, total costs, expandable child sessions, and immediate session opening.
• Open local file links and highlight file references and command options.
• Refresh Prime Inference models from the live catalog, with offline fallback and team-scoped private models.
• Let root agents create independent sibling sessions with provider-scoped credentials.
• Send Shell message received notices through steering at the next safe boundary, without interrupting running tools.
• Stop owned recovery launches when another authenticated supervisor wins.

Validation

• Focused Node22 shell steering, supervisor ownership, and runtime/UI tests passed.
• Python 3.13: 57 shell tests and 104 REPL tests passed.
• npm run check passed.

Limitations

• A supervisor recovery launch with no confirmed authenticated winner is left untouched to avoid stopping an active supervisor.
• Native Windows execution remains unverified locally.
• A timed-out sibling-session create may leave an idle, unprompted session. It may remain if idle eviction is disabled.
• Reading an already-completed asyncio wrapper’s cached result may still trigger a completion notice. Awaiting the original handle in its creating cell suppresses it.

Supersedes #2079, #2073, #1145, #1928, #1996, #2053, and #2087.

RES-1308, RES-1314, RES-1275, ENG-5282, RES-1239, RES-1256

---

Note

Medium Risk
Touches model catalog merging, auth cache migration, daemon session creation, and injected steering messages—important runtime paths with fallbacks, but broad surface area and behavior changes for operators.

Overview
This PR extends Prime Inference model discovery end-to-end: a shared catalog parser in pi-ai reads live /models pricing and specs (with validation and safe code generation via render-models.ts), while the agent merges bundled models with a disk cache and background refresh (PI_OFFLINE=1 skips network). Authorized private routes now return full model metadata, not just ID sets, and the private-auth cache format stores catalog entries accordingly.

The agents view is redesigned around a compact table (session, model on every row, activity, cost/age), inactive sessions collapse by default (alt+i), subagents expand from agent rows (alt+right), and Enter no longer blocks on a loading anchor—opening the selected row always works.

RLM/runtime: depth-0 daemon sessions gain rlm.create_session for sibling top-level sessions with inherited provider credentials; detached bash() process groups emit steering “Shell message received” notices at the next safe boundary while keeping the kernel/session resident until completion. Editor and message UI highlight @path references and --flags (including queued previews).

Daemon: supervisor replacement launches verify authenticated ownership before treating a child as the winner, and kill losing contenders.

^{Reviewed by Cursor Bugbot for commit 5eee936. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add background Bash completion tracking, live Prime Inference model refresh, compact agents view, and RLM root session creation

• Background Bash handles now track their creating REPL cell and send bash.completed notifications only when the handle was not awaited by that cell. Sessions stay resident while background Bash work is active (bash.py, repl.py, agent-session.ts).
• Prime Inference models now refresh from the public models endpoint with bundled and disk-cache fallbacks. Private authorization caches complete model definitions (names, pricing, limits, modalities) instead of just IDs (prime-inference-model-catalog.ts, model-registry.ts, prime-inference-models.ts).
• The agents view is reworked into compact sectioned tables with shared Session/Model/Activity/Cost/Age columns, inactive-session collapsing (Alt+I), subagent expansion (Alt+Right), and a keyboard actions panel (agents-view-mode.ts).
• Daemon-backed RLM root sessions can be created via rlm.create_session, which validates inputs, transfers credentials, creates a resident depth-0 session, and returns a typed RLMCreateSessionHandle (daemon-mode.ts, rlm/init.py).
• Prompt token highlighting now covers @-references and -- options in editors, queued previews, and sent user messages via grapheme-aware masking (prompt-highlight.ts, user-message.ts).
• Behavioral Change: isPrivatePrimeInferenceModel now uses shared classification (case-insensitive internal/, dev/, or colon-containing IDs) instead of only the internal/ prefix; replacement supervisor launch validates PID-bound authenticated claims and kills competing replacements.

^{Macroscope summarized 5eee936.}

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 96f3620. Configure here.}