Session names and active subagents

Select a session in Projects, the mosaic, Sessions, or detailed Accounts, then press n or click Rename / suggest title. Enter a visible name and save. Suggest title (Ctrl+G) prepares a short name from the saved native title or a brief redacted task summary. Review it and save. Use native name clears the override.

Esc cancels the draft and returns to the selected session without saving, hiding or stopping it. While a provider terminal is attached, its keyboard input cannot trigger console actions such as renaming, hiding a mosaic tile or changing the background selection. Detach before using console shortcuts.

A visible name is separate from the internal session name. Renaming preserves the session ID, provider conversation ID, tmux socket, session HOME, account binding, workspace and checkout. Native history titles are preserved. The displayed name uses this priority: the devctrl override, the native title, then the original internal name. The browser and terminal console use the same helper, including shared session cards, saved-session history, mosaic entries and opening progress. Only the owner may rename; a shared terminal control grant does not grant ownership.

Suggestions do not send a prompt to the running agent, start another conversation, interrupt a turn, restart the CLI or invoke a paid model. The host reads bounded native metadata locally, using the official Codex title index format. It prefers a useful Claude custom-title / ai-title or Codex title-index entry, with the latest substantive saved user task as a fallback. Claude text blocks and Codex user messages, response items and paginated user-message items are supported. Later continuations such as “continue” or its localized equivalents do not replace the task. Injected environment and agent instructions, operator notifications, CLI commands, tool results and known meta messages are excluded.

The local extractor retains a short task clause of at most ten words and 80 characters. Credentials, URLs, paths and email addresses are removed before a suggestion leaves the account. It does not infer a task from a repository name or a tool such as bash, and rejects generic native names such as “api bash”. The controller also filters previously saved suggestions. If useful context is unavailable or outside the bounded read window, Suggest title explains that there is not enough saved task context and preserves the draft name for manual editing. Suggestions are excerpts of saved task context, not model-generated summaries of the entire conversation.

Native capabilities were checked against the current primary documentation. Codex `/rename` changes a saved chat's name, and Codex App Server provides thread/name/set. Claude's CLI reference describes --name and /rename. devctrl runs standalone provider terminals and does not own an App Server connection to those active CLIs. Its rename action therefore changes the devctrl display override. It never types /rename into the live composer or edits native history files.

Each managed session in Projects, the mosaic, Sessions and detailed Accounts has a separate SUBAGENTS row. Its short status stays visible when a long task title is truncated. Press j, or click the row in the browser, to inspect native child IDs, active/inactive/unknown state, observation time and parent relationships supplied by the provider. The panel scrolls with arrows, PgUp/PgDn or the mouse wheel. It reads the existing snapshot without attaching, sending input or restarting the provider.

The separate MODEL row shows the main session's observed native model. The same row and j details are used in the terminal and browser console. Claude's official status-line payload supplies the selected model.id, including a switch before the next response; its bounded model file contains only the account, native conversation, process generation, model and timestamps. Without a fresh status-line observation, the exact parent's latest native assistant response can supply its last observed model. Codex uses the exact rollout's latest turn_context.model after the current process started. In j, the source distinguishes the native selected model from the last native turn/response model. A Codex selection or resume without a new persisted turn cannot prove a new model: resuming alone does not update its rollout. Updates appear when the provider publishes new native evidence; devctrl never types /model or opens a second provider process to query it.

Account, conversation and process-generation mismatches hide a model, and observations expire after 60 seconds without a refresh. Restart/exit clears the parent observation. A resumed process cannot reuse model evidence recorded before it started. Launch configuration and model-routing advice are not presented as the current model. Additional CLIs whose native model telemetry is unsupported show MODEL · unknown. Shared session cards include only the display model/source/timestamps, excluding the account and native conversation identity. No paid request is made to name or discover models.

Long Codex turns can push their turn_context out of a short transcript tail. The model reader searches backward through at most 8 MiB within a 500 ms budget, then inspects bounded appends to the same owned rollout. Its cache contains only the sanitized model and file position, validates the native conversation on each read, and fences records by account, UID, generation and the current process-start boundary. Ordinary output appends retain the last native turn model; a later model record replaces it. A changed identity, truncated/replaced file, unobserved append gap or exhausted scan budget remains unknown rather than borrowing another process's model.

Child rows show a short task title, the native model when available, token usage and explicit CPU/RAM availability. The active-child preview puts its model before a long title. j identifies the latest native model's source and timestamp, distinguishing a turn/response from a successful spawn; stale or inactive child models are labeled as observations. Later native turn contexts or responses replace the child's earlier model after a switch. Codex prefers the native saved thread title, then a local redacted excerpt of the latest child task, with its native nickname as a fallback. A successful persisted native spawn can supply a short task title before its child file flushes. Claude reads only the exact owning account/session's subagents/agent-ID.jsonl file and matches each row's native agentId, sessionId and sidechain flag. Titles use the same bounded local extractor as session-name suggestions; they do not call a model, inject a prompt or expose full task text.

Codex thread total uses the latest native `token_usage_record.thread_token_usage` whose thread_id matches the child. Cumulative snapshots are never summed, and parent/sibling counters are excluded. Older non-forked histories can use token_count.info.total_token_usage; inherited fork counters remain unavailable without exact child attribution. These are recorded thread counters, not a bill or a subscription quota percentage. Cached input and reasoning output are subsets of their native input/output categories.

Claude latest response shows the latest owned native assistant-message usage observed in the bounded child file, not an estimated cumulative history total. Streamed copies of a message are not added together. Its displayed input includes uncached input, cache reads and cache creation once, following the native cache usage categories. Missing counters/model/title are shown as unavailable. If lifecycle hooks are missing, native child history can still be inspected, but its activity and active total stay unknown.

Neither provider's native thread metadata establishes a unique OS PID for every child. CPU/RAM unavailable is therefore explicit; the parent process's CPU/RAM is not divided among children. No monetary cost is inferred from token counts or subscription usage.

Ordinary tool calls and OS process counts do not count as subagents. Claude uses its official SubagentStart and SubagentStop hooks, with unique agent IDs and separate metadata markers, so concurrent starts do not lose updates. Claude hooks supply child IDs but do not establish each child's native parent; those parents are shown as unavailable. A parent reaching its input prompt does not clear background agents. A new provider process generation cannot inherit an older count.

Codex uses its account’s canonical native rollout store and read-only metadata and events, as defined by the official Codex protocol source. source.subagent.thread_spawn.parent_thread_id (or the native thread_source / parent_thread_id fields) establishes the family. The panel shows direct children with Parent: this session, and descendants with their native parent ID. Native turn starts, completion (including terminal failures), abort events, successful spawns and parent close/resume/wait statuses establish activity. Current paginated histories persist item_completed records containing CollabAgentToolCall / SubAgentActivity; legacy histories can use corresponding native lifecycle events. Transient errors and tool end events are not guaranteed to appear in every provider version. Sibling conversations, unrelated accounts, review and compaction helpers are excluded. A task that finished responding is inactive even if its reusable native thread still exists. An active event from before the current provider process started leaves that child and the total unknown until newer native evidence arrives.

The host reads only the owning account’s canonical state (and Claude’s workspace-scoped account state), filters the exact conversation family, validates file ownership, refuses symlinked rollouts, limits directory traversal and read sizes with a two-second scan budget, and sends the count, source, generation, observation time and at most 256 child records. Each optional task title is at most 80 characters; model IDs and numeric counters are separately bounded. No full subagent prompt, transcript, tool arguments or response content is sent. The Claude collector keeps all active identities before recent inactive history, and marks truncated history explicitly.

Scan limits, unknown formats, incomplete child status or unreadable native state produce SUBAGENTS · unknown. An observation older than 60 seconds becomes unknown (stale), even if the view receives no new snapshot. Stopped and failed sessions show session ended without inventing a zero count. Child states in an unknown/stale/ended panel are labeled as last observations; earlier process generations are hidden. A known zero is shown only from current native evidence. Older count-only reports still show their count while explaining that child identities are unavailable. The host checks the conversation and process generation again when reporting to the controller.

Claude sessions launched with the updated configuration register both subagent hooks. Already-running CLIs with an older hook configuration may show unknown until the user's next normal resume/restart; applying a display name never forces that restart. Native file formats can change between provider versions, so unsupported versions remain visible as unknown telemetry.

The owner API is PUT /v1/sessions/:id/display-name with { "displayName": "Checkout repair" }, or null to reset. Names are at most 80 characters and reject control characters, invisible format characters and line separators. POST /v1/sessions/:id/display-name/suggest accepts an empty object and returns a suggestion without applying it, with source native-title or local-context. If no useful title is available it returns { "displayName": null, "source": "insufficient-context" }. It rejects transcript/prompt payloads. The authenticated owning host reports native counts through PUT /v1/agents/:id/sessions/:sessionId/subagents; observations from another host, conversation or process generation are rejected.