@dreb/coding-agent 2.47.0 → 2.49.0

package/docs/dashboard.md

@@ -120,7 +120,7 @@ networking window above.
 | Screen | What it does |
 |---|---|
 | **Fleet** | Home. Live-first: one grid of every live session at the top — status chip (● running / ◆ needs-attention / ○ idle / ✕ error), project path, activity line, live subagent lines, tasks progress, ctx%, model, terminal provider-error reason, last activity. Live cards keep a deterministic order by project path, then session start time; needs-attention cards badge the browser tab without jumping around. Below the grid: past sessions grouped by project, three compact rows per group with an "all N on disk" expander, resume and delete. |
-| **Session view** | Full chat drill-in. Markdown streaming transcript (text, thinking blocks with expand preference, inline provider/API failures with partial output preserved, agent-result cards, tool cards with bespoke read/write/edit/bash bodies plus full expandable inputs, markdown-rendered results for markdown-contract tools like subagent/skill/web_fetch/suggest_next, and inline tool-result images, compaction/branch summaries, custom messages), per-message copy, tasks panel, subagent strip, status line with elapsed time plus ■ stop and compaction/retry aborts, a persistent session-header live indicator, and an info bar with cwd, branch, session name, token breakdown, cost/(sub)/daily rollup, ctx%, median tok/s, and a stats popover. Composer supports auto-grow, history, `/` autocomplete from `get_commands`, image attach/paste with sent images retained as user-message previews, queued-message chips with restore-all, steer/follow-up modes, and suggest-next. The ⋯ menu covers export HTML, compact, rename, fork-from-message, loaded context, and tool expand/collapse. Session names update live from manual rename or auto-naming. Extension UI requests for select/confirm/input/editor render as modals; rich `ask`/`ask_user` questions render inline with Markdown-formatted question text, choices, optional free text, an in-card Stop agent action, Escape-to-stop, and the authoritative auto-stop countdown. Pending questions set needs-attention state and use the existing hidden-page notification path. Extension notifications render as toasts. |
+| **Session view** | Full chat drill-in. Markdown streaming transcript (text, thinking blocks with expand preference, inline provider/API failures with partial output preserved, agent-result cards, tool cards with bespoke read/write/edit/bash bodies plus full expandable inputs, markdown-rendered results for markdown-contract tools like subagent/skill/web_fetch/suggest_next, and inline tool-result images, compaction/branch summaries, custom messages), per-message copy, tasks panel, a bounded scrollable subagent panel that lists every retained agent newest-first with full running/done counts, status line with elapsed time plus ■ stop and compaction/retry aborts, a persistent session-header live indicator, and an info bar with cwd, branch, session name, token breakdown, cost/(sub)/daily rollup, ctx%, median tok/s, and a stats popover. Composer supports auto-grow, history, `/` autocomplete from `get_commands`, image attach/paste with sent images retained as user-message previews, queued-message chips with restore-all, steer/follow-up modes, and suggest-next. The ⋯ menu covers export HTML, compact, rename, fork-from-message, loaded context, and tool expand/collapse. Session names update live from manual rename or auto-naming. Extension UI requests for select/confirm/input/editor render as modals; rich `ask`/`ask_user` questions render inline with Markdown-formatted question text, choices, optional free text, an in-card Stop agent action, Escape-to-stop, and the authoritative auto-stop countdown. Pending questions set needs-attention state and use the existing hidden-page notification path. Extension notifications render as toasts. |
 | **Subagent view** | Read-only transcript of a background agent: live events via the RPC relay, hydrated from the agent's on-disk session log (`/subagents/:agentId/messages`) so the transcript survives browser reloads. Shows the task, streaming output, tool activity, and any safe Dispatch Arbiter changed/unchanged/failure records with the final agent/model/thinking. No raw arbiter output is displayed or transported. No composer — subagents can't be steered yet; the parent session controls them. |
 | **Files** | Host-wide browser with places shortcuts (home, /tmp, project roots), breadcrumbs to `/`, new-folder, download, drop-zone/picker upload with explicit collision prompts, and "new session here" on any directory. It also shows the **effective global nested-context trust** for the displayed canonical directory: untrusted, trusted by that root, inherited from a granting root, or global expert trust-all. You can trust the displayed folder and descendants, or untrust the actual granting root; untrusting an inherited folder removes that root's trust for all descendants. |
 | **Settings** | Persistent defaults (default model, thinking level, steering/follow-up queue modes, auto-compaction, auto-retry) via `get_settings`/`set_settings` — validation errors are shown verbatim. The global-only Dispatch Arbiter card exposes enable/disable, exact authenticated model selection, thinking, guide path, and readiness guidance; model-less enablement is blocked and RPC/runtime validation remains fail-closed. Entering Settings flushes pending writes and reloads durable global + project settings, so external edits appear; read, parse, or write failures fail loudly instead of showing stale settings. The global-only nested-context policy lists every explicit trusted root for audit and revoke, offers a simple add-by-path control, and includes a prominently warned expert trust-all toggle; the Files view remains the primary place to grant trust while browsing. Most defaults seed new sessions; context-trust changes are observed by active main/subagent processes for future lazy loads, but cannot remove already injected content. Dashboard-local preferences (always expand thinking, transcript image display mode, needs-attention notification permission) live in the browser, alongside an appearance section: a theme gallery of eight curated themes (entropist.ca, Dim, Solarized, Gruvbox, Caves of Qud, Van Gogh, and the colorblind-safe Okabe-Ito and Paul Tol) with live preview cards and a system/light/dark mode selector, saved per browser. Shows the current rotating pairing code on the host/local dashboard, plus the paired-devices list with unpair. |
@@ -128,7 +128,7 @@ networking window above.
 
 ### Dispatch Arbiter observability
 
-When the global Dispatch Arbiter is enabled, the dashboard consumes the typed `subagent_arbitration` RPC event. The matching background-agent card is updated to the final selected agent before child events arrive; the parent chip shows final model/thinking or a failure marker, and the subagent drill-in lists ordered records (including chain steps). Runtime snapshots carry the same safe records so refresh/resync during a live process does not revert to the requested identity.
+When the global Dispatch Arbiter is enabled, the dashboard consumes the typed `subagent_arbitration` RPC event. The matching background-agent card is updated to the final selected agent before child events arrive; the parent panel row shows final model/thinking or a failure marker, and the subagent drill-in lists ordered records (including chain steps). Runtime snapshots carry the same safe records so refresh/resync during a live process does not revert to the requested identity.
 
 Only host-validated proposed/final tuples, changed fields, status, step, and bounded host errors cross RPC/SSE. Arbiter prompts, raw output, reasoning, tasks, guides, and parent excerpts never reach dashboard protocol state. The safe record is separately persisted in the parent session as a non-context custom entry; child transcript hydration remains sourced from the child log.
 
@@ -374,8 +374,10 @@ Then open `https://hostname.tailXXXX.ts.net:<port>` on the phone.
 Background subagents are first-class:
 
 - Fleet cards show running/done counts and live agent lines.
-- The session view shows a chip strip — one chip per background agent; click
-  to drill into its live transcript.
+- The session view shows every retained background agent in a bounded,
+  scrollable panel, ordered newest-first with the full running/done count in its
+  summary. It uses the same native collapse pattern as the task tracker, starts
+  collapsed on mobile, and keeps every row available for transcript drill-in.
 - The drill-in view streams the child's events in real time via the
   `background_agent_event` relay (see [RPC events](rpc.md#event-types)) and
   hydrates from the agent's on-disk session log on mount, so transcripts
@@ -387,7 +389,8 @@ Single breakpoint at 700px. At <=700px, fleet cards stack; long session names,
 status chips, project paths, activity and subagent text, and past-session
 labels wrap within their cards or rows rather than spilling off-screen. The
 session view prioritizes read-and-steer (model/thinking switchers collapse into
-⋯, tasks default collapsed), and the file table shows name + download only.
+⋯, and task/subagent panels default collapsed), and the file table shows name +
+download only.
 Composer modes, abort, and needs-attention affordances are never reduced away
 — steering a running agent from a phone is the primary remote use case.
 

package/docs/providers.md

@@ -52,6 +52,7 @@ Use `/logout` to clear credentials. Tokens are stored in `~/.dreb/agent/auth.jso
 - `/login` uses the Kimi Code OAuth subscription endpoint at `https://api.kimi.com/coding/v1`.
 - `KIMI_API_KEY` uses Kimi For Coding's Anthropic-compatible API at `https://api.kimi.com/coding`.
 - Built-in OAuth models are `kimi-for-coding` (default, 262k context), `k3` (1M context), and `kimi-for-coding-highspeed` (262k context). Model availability is plan-dependent: the `k3` 1M-context ID and the `kimi-for-coding-highspeed` ID are only exposed when the subscription includes them, and `kimi-for-coding-highspeed` runs at roughly 6× speed for 3× quota usage. On login/refresh, dreb asks the Kimi API which models the subscription is entitled to and updates context, reasoning, image, tool-use, protocol, and thinking-effort metadata. Compatible newly discovered IDs are templated conservatively; if discovery fails, the static list remains available.
+- The OAuth `k3` model is context-tiered: the Kimi endpoint serves it as `k3-256k` (256k context, cheaper) and `k3` (1M context, stated to consume 2× the quota), and upgrading does not invalidate the prompt cache. The cheaper variant is exclusive to the Kimi for Coding OAuth endpoint — the pay-per-token Moonshot AI Platform does not expose it. dreb starts every session on the `k3-256k` wire ID and automatically upgrades to `k3` once the session context passes the 256k cutoff (256k minus the default compaction reserve) instead of compacting. A user-lowered compaction threshold compacts before the cutoff and thus effectively disables the upgrade. Customizing `k3`'s context window via `models.json` disables automatic tiering entirely, and compacted or fresh sessions return to the cheaper tier.
 - OAuth requests use the current Kimi Code device identity contract and share its stable `~/.kimi-code/device_id`. Extra low-precedence headers can be supplied with newline-separated `KIMI_CODE_CUSTOM_HEADERS` values.
 - The OAuth coding endpoint accepts OpenAI-style multimodal content arrays with base64 `image_url` data URLs. dreb keeps a conservative 32k output-token cap because the managed model catalog does not currently advertise a per-model output limit.
 - Moonshot Open Platform uses a different base URL (`https://api.moonshot.ai/v1`); don't assume both routes expose identical behavior.

package/docs/rpc.md

@@ -1617,6 +1617,7 @@ Response:
 | `length_retry` | Response hit the token limit; retrying with a larger budget |
 | `auto_compaction_start` | Auto-compaction begins |
 | `auto_compaction_end` | Auto-compaction completes |
+| `context_window_upgrade` | Wire model tier auto-upgraded (e.g. Kimi K3 256k → 1M); includes `provider`, `modelId`, `fromContextWindow`, `toContextWindow` |
 | `auto_retry_start` | Auto-retry begins (after transient error) |
 | `auto_retry_end` | Auto-retry completes (success or final failure) |
 | `background_agent_start` | Background subagent launched (includes `sessionDir`) |

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@dreb/coding-agent",
-	"version": "2.47.0",
+	"version": "2.49.0",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"type": "module",
 	"drebConfig": {