@circuitwall/jarela 2.2.1 → 2.3.0

package/.next/standalone/docs/adoption/changesets.json

@@ -0,0 +1,309 @@
+{
+  "schema_version": 1,
+  "changesets": [
+    {
+      "version": "2.0.1",
+      "previous_version": "2.0.0",
+      "date": "2026-10-11",
+      "hints": [],
+      "commits": [
+        "chore(release): bump to 2.0.1",
+        "fix(npm): remove oversized runtime assets"
+      ],
+      "prompts": [],
+      "tools": []
+    },
+    {
+      "version": "2.0.2",
+      "previous_version": "2.0.1",
+      "date": "2026-10-11",
+      "hints": [],
+      "commits": [
+        "chore(release): bump to 2.0.2",
+        "fix(npm): exclude nested server source maps"
+      ],
+      "prompts": [],
+      "tools": []
+    },
+    {
+      "version": "2.0.3",
+      "previous_version": "2.0.2",
+      "date": "2026-10-12",
+      "hints": [],
+      "commits": [
+        "chore(release): bump to 2.0.3",
+        "fix(npm): remove maps before npm packing"
+      ],
+      "prompts": [],
+      "tools": []
+    },
+    {
+      "version": "2.0.4",
+      "previous_version": "2.0.3",
+      "date": "2026-10-12",
+      "hints": [],
+      "commits": [
+        "chore(release): bump to 2.0.4",
+        "fix(npm): use installed onnx runtime dependencies"
+      ],
+      "prompts": [],
+      "tools": []
+    },
+    {
+      "version": "2.0.5",
+      "previous_version": "2.0.4",
+      "date": "2026-10-13",
+      "hints": [],
+      "commits": [
+        "chore(release): bump to 2.0.5",
+        "fix(providers): detect streams that end without events",
+        "fix(http): fix gzip response decoding",
+        "docs(site): update guides, theme, and architecture diagrams",
+        "refactor(bridges): add capability-driven bridge modules",
+        "fix(security): update vulnerable media parser",
+        "refactor(modules): split built-in provider and package registries"
+      ],
+      "prompts": [],
+      "tools": []
+    },
+    {
+      "version": "2.1.0",
+      "previous_version": "2.0.5",
+      "date": "2026-10-14",
+      "hints": [],
+      "commits": [
+        "chore(release): bump to 2.1.0",
+        "fix(release): add model download retries and cache",
+        "fix(chat): render grouped summary topic previews",
+        "fix(workflows): validate legacy checklist items",
+        "docs(pricing): document aggregator data sources",
+        "feat(core): add verified workflows, pricing signals, and summaries"
+      ],
+      "prompts": [
+        {
+          "id": "static.agent.tool-usage-sop",
+          "source": "lib/agents/prepare/system-prompt.ts",
+          "change": "changed",
+          "added_lines": [
+            "0. Classify the request. Informational answers may be direct; read/retrieve requests need a read tool; write/change requests need a state-changing tool; destructive requests also need the required confirmation. Use only these operation labels when describing work: read, write, execute, send, schedu…"
+          ],
+          "removed_lines": [
+            "0. Classify the request. Informational answers may be direct; read/retrieve requests need a read tool; write/change requests need a state-changing tool; destructive requests also need the required confirmation. Use only these operation labels when describing work: read, write, execute, send, schedu…"
+          ],
+          "omitted_lines": 0
+        },
+        {
+          "id": "system.Tool usage SOP",
+          "source": "lib/agents/prepare/system-prompt.ts",
+          "change": "changed",
+          "added_lines": [
+            "0. Classify the request. Informational answers may be direct; read/retrieve requests need a read tool; write/change requests need a state-changing tool; destructive requests also need the required confirmation. Use only these operation labels when describing work: read, write, execute, send, schedu…"
+          ],
+          "removed_lines": [
+            "0. Classify the request. Informational answers may be direct; read/retrieve requests need a read tool; write/change requests need a state-changing tool; destructive requests also need the required confirmation. Use only these operation labels when describing work: read, write, execute, send, schedu…"
+          ],
+          "omitted_lines": 0
+        }
+      ],
+      "tools": [
+        {
+          "name": "refresh_model_pricing",
+          "change": "changed",
+          "description_changed": true,
+          "description": "Refresh provider pricing pages and separate OpenRouter/LiteLLM reference feeds, then recalculate dashboard estimates using provider-page rates only. Returns rate changes, source health, confidence, and current-versus-recorded dashboard totals. These are estimates, not invoice amounts.",
+          "previous_description": "Refresh the configured official LLM pricing sources and recalculate dashboard estimates using the latest rates. Returns rate changes, source health, confidence, and current-versus-recorded dashboard totals. These are estimates, not invoice amounts.",
+          "params_added": [],
+          "params_removed": [],
+          "params_changed": [],
+          "required_added": [],
+          "required_removed": []
+        },
+        {
+          "name": "workflow_progress",
+          "change": "changed",
+          "description_changed": true,
+          "description": "Report structured progress for a substantial multi-step task. Use only when the task has several dependent phases, likely spans turns, or needs explicit verification; skip simple tasks and work already covered by native tool progress. For agent-created workflows, send the full current items snapshot on every call and reuse the same workflow_id. Keep secrets and private data out of workflow fields. Mark items done only with observed validation evidence; use needs_attention for blockers; phase=complete requires every item done or skipped. Every item needs input, expected output, and concrete va…",
+          "previous_description": "Report structured progress for a substantial multi-step task. Use only when the task has several dependent phases, likely spans turns, or needs explicit verification; skip simple tasks and work already covered by native tool progress. For agent-created workflows, send the full current items snapshot on every call and reuse the same workflow_id. Keep secrets and private data out of workflow fields. Mark items done only after verifying the action; use needs_attention for blockers; phase=complete requires every item done or skipped. The system-owned version_adoption workflow has its own durable …",
+          "params_added": [
+            "validation_evidence"
+          ],
+          "params_removed": [],
+          "params_changed": [
+            "items"
+          ],
+          "required_added": [],
+          "required_removed": []
+        }
+      ]
+    },
+    {
+      "version": "2.2.0",
+      "previous_version": "2.1.0",
+      "date": "2026-10-15",
+      "hints": [],
+      "commits": [
+        "chore(release): bump to 2.2.0",
+        "feat(embeddings): add on-demand local model download",
+        "fix(build): exclude foreign runtimes from standalone bundles"
+      ],
+      "prompts": [],
+      "tools": []
+    },
+    {
+      "version": "2.2.1",
+      "previous_version": "2.2.0",
+      "date": "2026-10-15",
+      "hints": [],
+      "commits": [
+        "chore(release): bump to 2.2.1",
+        "fix(api): skip browser http cache for client requests",
+        "feat(models): add model enable toggle and atomic default",
+        "fix(github-copilot): remove retired pat route"
+      ],
+      "prompts": [],
+      "tools": []
+    },
+    {
+      "version": "2.3.0",
+      "previous_version": "2.2.1",
+      "date": "2026-10-16",
+      "hints": [
+        "If an agent's instructions or skills tell it to leave fact capture to compaction or to avoid `memory_upsert`, remove that guidance, since agents now save facts during the turn.",
+        "If a scheduled task or watcher prompt asks the agent to write memory facts, keep that in the agent's own turn; automation channel summaries no longer extract facts."
+      ],
+      "commits": [
+        "chore(repo): remove dead exports and the unused deepseek package",
+        "fix(chat): restore thread activity and bound compaction per thread",
+        "fix(chat): render automation recaps in the warm summary card",
+        "fix(chat): restore the compacting state after switching agents",
+        "fix(chat): allow compacting bridge-only threads, report via toasts",
+        "fix(scripts): align make targets and install the current working tree",
+        "feat(agents): add topic pipeline to automation channel summaries",
+        "feat(agents): add validated topic pipeline for warm compaction",
+        "feat(agents): add proactive, validated memory saving for agents",
+        "fix(chat): group older summaries into collapsible topic rows",
+        "feat(agents): add auto-retry for failed workflow step contracts",
+        "feat(chat): add per-step retry to workflow checklists",
+        "feat(lifecycle): add release changesets for version adoption",
+        "fix(chat): keep boundary locate from snapping to bottom"
+      ],
+      "prompts": [
+        {
+          "id": "static.agent.compaction-group",
+          "source": "lib/agents/compaction-pipeline.ts",
+          "change": "added",
+          "added_lines": [],
+          "removed_lines": [],
+          "omitted_lines": 0
+        },
+        {
+          "id": "static.agent.compaction-group-automation",
+          "source": "lib/agents/compaction-pipeline.ts",
+          "change": "added",
+          "added_lines": [],
+          "removed_lines": [],
+          "omitted_lines": 0
+        },
+        {
+          "id": "static.agent.compaction-topic-recap",
+          "source": "lib/agents/compaction-pipeline.ts",
+          "change": "added",
+          "added_lines": [],
+          "removed_lines": [],
+          "omitted_lines": 0
+        },
+        {
+          "id": "static.agent.compaction-topic-recap-automation",
+          "source": "lib/agents/compaction-pipeline.ts",
+          "change": "added",
+          "added_lines": [],
+          "removed_lines": [],
+          "omitted_lines": 0
+        },
+        {
+          "id": "static.agent.conversation-summary",
+          "source": "lib/agents/conversation-summary.ts",
+          "change": "changed",
+          "added_lines": [
+            "A topic groups every consecutive message about one subject and must",
+            "span at least two messages (a request and its reply); fold a stray",
+            "single message into the adjacent topic instead of giving it its own."
+          ],
+          "removed_lines": [],
+          "omitted_lines": 0
+        },
+        {
+          "id": "static.agent.tool-usage-sop",
+          "source": "lib/agents/prepare/system-prompt.ts",
+          "change": "changed",
+          "added_lines": [
+            "0. Classify the request. Informational answers may be direct; read/retrieve requests need a read tool; write/change requests need a state-changing tool; destructive requests also need the required confirmation. Use only these operation labels when describing work: read, write, execute, send, schedu…"
+          ],
+          "removed_lines": [
+            "0. Classify the request. Informational answers may be direct; read/retrieve requests need a read tool; write/change requests need a state-changing tool; destructive requests also need the required confirmation. Use only these operation labels when describing work: read, write, execute, send, schedu…"
+          ],
+          "omitted_lines": 0
+        },
+        {
+          "id": "system.Memory & recall",
+          "source": "lib/agents/prepare/system-prompt.ts",
+          "change": "changed",
+          "added_lines": [
+            "You have long-term memory across sessions, shared with every other Jarela agent. A record you save is available to other agents immediately. Background compaction also extracts facts later, but it runs late and can miss details, so it is a backstop, not your save step.",
+            "Proactive save procedure — run it whenever information fits the active memory policy, without waiting to be asked:",
+            "1. Trigger: the user states a preference, a fact about themselves or their work, a decision, a constraint, or a correction to something saved; or a tool result verifies a durable fact. Routine status, one-off chatter, secrets, credentials, and raw transcripts never qualify.",
+            "2. Check: memory_search (or memory_read for a known key) for an existing record on the same subject. Update that stable key rather than creating a duplicate or contradictory copy.",
+            "3. Write: call memory_upsert in the same turn, before your final reply, with namespace='facts', a stable descriptive key, standalone content, concise tags, confidence='explicit' for user-stated information ('verified' when tool-backed, 'inferred' only when useful and reversible), observed_at when t…",
+            "4. Validate: the memory_upsert result must contain ok:true and the key you sent. If it failed, correct the arguments and retry once; if it still fails, tell the user it was not saved.",
+            "5. Cite: end your reply with one short line naming each saved key, for example 'Saved to memory: facts/answer-format-preference'. Omit the line when nothing was saved."
+          ],
+          "removed_lines": [
+            "You have long-term memory across sessions.",
+            "- Use memory_upsert proactively for structured records. Use namespace='facts', concise tags, confidence='explicit' for user-stated information, observed_at when the information was true, and expires_at for temporary facts. Keep memory_write only for legacy/free-form notes."
+          ],
+          "omitted_lines": 0
+        },
+        {
+          "id": "system.Tool usage SOP",
+          "source": "lib/agents/prepare/system-prompt.ts",
+          "change": "changed",
+          "added_lines": [
+            "0. Classify the request. Informational answers may be direct; read/retrieve requests need a read tool; write/change requests need a state-changing tool; destructive requests also need the required confirmation. Use only these operation labels when describing work: read, write, execute, send, schedu…"
+          ],
+          "removed_lines": [
+            "0. Classify the request. Informational answers may be direct; read/retrieve requests need a read tool; write/change requests need a state-changing tool; destructive requests also need the required confirmation. Use only these operation labels when describing work: read, write, execute, send, schedu…"
+          ],
+          "omitted_lines": 0
+        }
+      ],
+      "tools": [
+        {
+          "name": "memory_upsert",
+          "change": "changed",
+          "description_changed": true,
+          "description": "Store a durable, structured memory record for proactive semantic recall. Use namespace='facts' for information that should surface in future conversations. Store explicit user preferences, verified facts, decisions, constraints, and reusable project context; do not store transient chat details or secrets. Use an expiry for temporary facts, and status='archived' to soft-retire a record instead of deleting it. Save proactively in the same turn the information appears so other agents can use it immediately; do not defer to background compaction.",
+          "previous_description": "Store a durable, structured memory record for proactive semantic recall. Use namespace='facts' for information that should surface in future conversations. Store explicit user preferences, verified facts, decisions, constraints, and reusable project context; do not store transient chat details or secrets. Use an expiry for temporary facts, and status='archived' to soft-retire a record instead of deleting it.",
+          "params_added": [],
+          "params_removed": [],
+          "params_changed": [],
+          "required_added": [],
+          "required_removed": []
+        },
+        {
+          "name": "workflow_progress",
+          "change": "changed",
+          "description_changed": true,
+          "description": "Report structured progress for a substantial multi-step task. Use only when the task has several dependent phases, likely spans turns, or needs explicit verification; skip simple tasks and work already covered by native tool progress. For agent-created workflows, send the full current items snapshot on every call and reuse the same workflow_id. Keep secrets and private data out of workflow fields. Mark items done only with observed validation evidence; use needs_attention for blockers; phase=complete requires every item done or skipped. A rejected update, or a step left checking or needs_atte…",
+          "previous_description": "Report structured progress for a substantial multi-step task. Use only when the task has several dependent phases, likely spans turns, or needs explicit verification; skip simple tasks and work already covered by native tool progress. For agent-created workflows, send the full current items snapshot on every call and reuse the same workflow_id. Keep secrets and private data out of workflow fields. Mark items done only with observed validation evidence; use needs_attention for blockers; phase=complete requires every item done or skipped. Every item needs input, expected output, and concrete va…",
+          "params_added": [
+            "needs_user"
+          ],
+          "params_removed": [],
+          "params_changed": [],
+          "required_added": [],
+          "required_removed": []
+        }
+      ]
+    }
+  ]
+}

package/.next/standalone/lib/skills/builtins/memory-organization/SKILL.md

@@ -37,6 +37,22 @@ Every update to an existing key keeps a dated snapshot of the prior record in
 its `history`, and older free-form or v1 entries are upgraded to the current
 schema automatically the first time they're read — no manual migration step.
 
+## Proactive Saving
+
+Memory is shared by every Jarela agent, so save durable information in the
+turn it appears instead of relying on background compaction, which runs later
+and can miss details.
+
+1. **Trigger:** the user states a preference, fact, decision, constraint, or
+   correction; or a tool result verifies a durable fact.
+2. **Check:** search for an existing record on the same subject and reuse its
+   key.
+3. **Write:** call `memory_upsert` before the final reply.
+4. **Validate:** the result must contain `ok: true` and the key you sent;
+   retry once on failure, then tell the user it was not saved.
+5. **Cite:** end the reply with `Saved to memory: facts/<key>` for each saved
+   key.
+
 ## Retrieval And Consolidation
 
 1. Let proactive recall supply matching facts on normal user turns. Use

package/.next/standalone/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@circuitwall/jarela",
-  "version": "2.2.1",
-  "description": "Jarela — local chat interface for LangGraph agents (multi-provider, single-process, SQLite-backed).",
+  "version": "2.3.0",
+  "description": "Jarela — local chat interface for LangGraph agents (multi-provider, single-process, SQLite-backed).",
   "license": "Apache-2.0",
   "author": "Andrew Ge Wu",
   "homepage": "https://github.com/CircuitWall/jarela#readme",
@@ -116,6 +116,7 @@
     "check:impact": "node scripts/check-change-impact.mjs",
     "pricing:snapshot": "node scripts/snapshot-pricing.mjs",
     "prompts:dump": "node scripts/dump-prompts.mjs",
+    "release:changeset": "node scripts/release-changeset.mjs",
     "docs:sync-assets": "node scripts/sync-docs-assets.mjs",
     "docs:dev": "npm run docs:sync-assets && mkdocs serve",
     "docs:build": "npm run docs:sync-assets && mkdocs build",
@@ -143,7 +144,6 @@
     "@huggingface/transformers": "^4.3.1",
     "@langchain/cohere": "^1.1.0",
     "@langchain/core": "^1.2.12",
-    "@langchain/deepseek": "^1.1.13",
     "@langchain/google-genai": "^2.3.2",
     "@langchain/langgraph": "^1.4.17",
     "@langchain/langgraph-checkpoint": "^1.1.5",

package/CHANGELOG.md

@@ -7,6 +7,31 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [2.3.0] - 2026-10-16
+
+### Added
+
+- **Compact earlier messages as validated topics: messages are grouped into topics, each topic gets a recap, and the recaps are merged in code. Every step's output is checked in code and retried once before falling back. Scheduled tasks, watchers and bridges are compacted with the same topic pipeline.**
+- **Agents save new facts to memory in the same turn with `memory_upsert`, check that the save succeeded, and cite the saved key, instead of waiting for background compaction.**
+- **Retry a single workflow step from the chat or the version-adoption checklist. Steps that fail validation retry automatically up to 3 times each.**
+- **Record a changeset of prompt and tool changes with every release, so version adoption can show what changed and adapt settings, task prompts, instructions and skills after an upgrade.**
+- **Show what a thread is doing (thinking, responding, preparing context, compacting) after switching agents or reloading, including turns started by a scheduled task, watcher, bridge or another device.**
+
+### Fixed
+
+- **Stop compaction after a time limit and allow only one compaction per thread at a time. The limit defaults to 10 minutes and is set with `JARELA_COMPACTION_TIMEOUT_MS`. `/compact` now says when a thread is already compacting or the attempt timed out.**
+- **Allow compacting agents whose only messages come from bridges, watchers or scheduled tasks, and show those recaps in the summary card instead of a summary that stays pending.**
+- **Report compaction results as toast notifications, and keep showing the compacting state after switching agents.**
+- **Show older summaries as collapsible topic rows instead of one long block.**
+- **Keep "locate boundary" and message deep links from snapping back to the bottom.**
+- **Install the current working tree, including uncommitted changes, from both `make.ps1` and `Makefile`, and reinstall dependencies when `package-lock.json` changes.**
+- **Remove the unused `@langchain/deepseek` dependency; DeepSeek still runs through the OpenAI-compatible adapter.**
+
+### Adoption
+
+- If an agent's instructions or skills tell it to leave fact capture to compaction or to avoid `memory_upsert`, remove that guidance, since agents now save facts during the turn.
+- If a scheduled task or watcher prompt asks the agent to write memory facts, keep that in the agent's own turn; automation channel summaries no longer extract facts.
+
 ## [2.2.1] - 2026-10-15
 
 ### Fixed

package/api/client.ts

@@ -523,6 +523,12 @@ export const api = {
           body: JSON.stringify({ action }),
           cache: "no-store",
         }),
+      retryItem: (itemId: string) =>
+        request<VersionAdoptionState>("/lifecycle/adoption", {
+          method: "POST",
+          body: JSON.stringify({ action: "retry", item_id: itemId }),
+          cache: "no-store",
+        }),
     },
   },
 
@@ -712,6 +718,9 @@ export const api = {
     },
     delete: (thread_id: string) =>
       request<{ deleted: boolean }>(`/threads/${thread_id}`, { method: "DELETE" }),
+    // What the thread is doing now (run phase, queued turns, compaction).
+    activity: (thread_id: string) =>
+      request<import("./types").ThreadActivity>(`/threads/${thread_id}/activity`, { cache: "no-store" }),
     // Abort the active agent run on this thread. The server's stream loop
     // catches the resulting AbortError and emits an `error` + `done` so
     // any UI queue-drain hook (e.g. ChatView) still fires.

package/api/types.ts

@@ -168,6 +168,7 @@ export interface VersionAdoptionChecklistItem {
   validation: string;
   status: VersionAdoptionChecklistStatus;
   validation_evidence?: string;
+  attempts?: number;
   reason: string;
   affected_files: string[];
   action?: "read_skill" | "review_permissions" | "review_schedule" | "review_watcher" | "read_instructions" | "review_tools" | "review_api" | "review_persistence" | "review_release";
@@ -408,6 +409,28 @@ export interface ThreadDetail extends ThreadSummary {
   // that a compaction is in flight.
   pending_hot_since?: string | null;
   pending_hot_since_seq?: number | null;
+  // A user-triggered /compact is running for this thread. Lets a client that
+  // remounts mid-compaction (agent switch, reload) restore the busy state.
+  manual_compaction_pending?: boolean;
+  activity?: ThreadActivity;
+  warm_channel_summaries?: WarmChannelSummary[];
+}
+
+/** What a thread is doing right now — GET /threads/:id/activity. */
+export interface ThreadActivity {
+  run: { phase: string; label: string; started_at: number; agent_id: string | null } | null;
+  queue_depth: number;
+  compaction: { kind: "manual" | "auto" | "refresh"; boundary_seq: number | null; started_at: number; deadline_at: number } | null;
+  busy: boolean;
+}
+
+/** Recap of one automation channel (bridge, watcher, scheduled task) committed
+ *  at the current boundary — ADR-0044. The chat recap is `warm_summary`. */
+export interface WarmChannelSummary {
+  channel: string;
+  summary: string;
+  computed_at: string;
+  topics: SummaryTopicSegment[] | null;
 }
 
 export interface SummaryTopicFact {
@@ -415,6 +438,9 @@ export interface SummaryTopicFact {
   content: string;
   tags: string[];
   confidence: "explicit" | "inferred" | "verified";
+  /** `namespace/key` of the memory record, once memorized. */
+  key?: string;
+  origin?: "agent" | "compaction";
 }
 
 export interface SummaryTopicSegment {
@@ -423,6 +449,8 @@ export interface SummaryTopicSegment {
   end_at: string;
   recap: string;
   facts: SummaryTopicFact[];
+  /** Messages folded into this topic across compaction passes. */
+  message_count?: number;
 }
 
 export interface ThreadContextPin {
@@ -435,6 +463,7 @@ export interface ThreadContextPin {
   warm_summary_source_messages: number | null;
   warm_summary_source_chars: number | null;
   warm_summary_topics: SummaryTopicSegment[] | null;
+  warm_channel_summaries?: WarmChannelSummary[];
   pending_hot_since?: string | null;
   pending_hot_since_seq?: number | null;
 }

package/app/api/v1/lifecycle/adoption/route.ts

@@ -4,12 +4,17 @@ import { validateBody } from "@/lib/api/responses";
 import {
   getVersionAdoptionState,
   updateVersionAdoptionState,
+  VersionAdoptionRequestError,
 } from "@/lib/stores/version-adoption";
 
 export const dynamic = "force-dynamic";
 
 const Body = z.object({
   action: z.enum(["start", "mark_done", "dismiss", "retry"]),
+  item_id: z.string().min(1).max(64).regex(/^[a-zA-Z0-9][a-zA-Z0-9_-]*$/).optional(),
+}).refine((body) => body.item_id === undefined || body.action === "retry", {
+  message: "item_id is only valid with action retry",
+  path: ["item_id"],
 });
 
 export function GET() {
@@ -19,5 +24,12 @@ export function GET() {
 export async function POST(req: NextRequest) {
   const body = await validateBody(req, Body);
   if (body instanceof NextResponse) return body;
-  return NextResponse.json(updateVersionAdoptionState(body.action));
-}
+  try {
+    return NextResponse.json(updateVersionAdoptionState(body.action, undefined, body.item_id));
+  } catch (err) {
+    if (err instanceof VersionAdoptionRequestError) {
+      return NextResponse.json({ error: err.message }, { status: 400 });
+    }
+    throw err;
+  }
+}

package/app/api/v1/models/[name]/compact-threads/route.ts

@@ -23,7 +23,7 @@ import { z } from "zod";
 import { listAgentConfigs } from "@/lib/stores/agent-configs";
 import { getDefaultModelConfig } from "@/lib/stores/model-config";
 import { listThreadsByAgent, getRecentMessagesWindow } from "@/lib/stores/threads";
-import { compactThreadWarmContext } from "@/lib/agents/warm-summary-background";
+import { compactThreadWarmContext, isManualCompactionRunning, trackManualCompaction } from "@/lib/agents/warm-summary-background";
 import { getProvider } from "@/lib/providers";
 import type { ProviderParams } from "@/lib/providers/types";
 import { errorMessage } from "@/lib/utils/error";
@@ -77,17 +77,20 @@ export async function POST(req: NextRequest, { params }: Params) {
       const msgs = getRecentMessagesWindow(t.thread_id, 0, undefined, "foreground");
       if (msgs.length <= keep_last) { skipped += 1; continue; }
       const hotStart = msgs[msgs.length - keep_last]!;
+      if (isManualCompactionRunning(t.thread_id)) { skipped += 1; continue; }
       try {
-        const committed = await compactThreadWarmContext(t.thread_id, hotStart.seq, {
-          expectedHotSinceSeq: t.hot_since_seq ?? null,
-          alignTopicBoundary: false,
-          allowEmptySummary: true,
-          providerOverride: {
-            provider,
-            modelId: using.model_id,
-            params: (using.params ?? {}) as ProviderParams,
-          },
-        });
+        const committed = await trackManualCompaction(t.thread_id, hotStart.seq, (attempt) =>
+          compactThreadWarmContext(t.thread_id, hotStart.seq, {
+            expectedHotSinceSeq: t.hot_since_seq ?? null,
+            alignTopicBoundary: false,
+            allowEmptySummary: true,
+            attempt,
+            providerOverride: {
+              provider,
+              modelId: using.model_id,
+              params: (using.params ?? {}) as ProviderParams,
+            },
+          }));
         if (committed) compacted += 1;
         else skipped += 1;
       } catch (e) {

package/app/api/v1/threads/[thread_id]/activity/route.ts

@@ -0,0 +1,19 @@
+/**
+ * UI-internal — `GET /api/v1/threads/[thread_id]/activity`
+ *
+ * One snapshot of what the thread is doing (agent turn phase, queued turns,
+ * compaction), polled by the chat so it can show and resume activity that
+ * started elsewhere: the scheduler, a watcher, a bridge, or another device.
+ */
+
+import { NextRequest, NextResponse } from "next/server";
+import { getThread } from "@/lib/stores/threads";
+import { getThreadActivity } from "@/lib/agents/thread-activity";
+
+type Params = { params: Promise<{ thread_id: string }> };
+
+export async function GET(_req: NextRequest, { params }: Params) {
+  const { thread_id } = await params;
+  if (!getThread(thread_id)) return NextResponse.json({ error: "Thread not found" }, { status: 404 });
+  return NextResponse.json(getThreadActivity(thread_id), { headers: { "Cache-Control": "no-store" } });
+}

package/app/api/v1/threads/[thread_id]/context-pin/route.test.ts

@@ -27,6 +27,7 @@ vi.mock("@/lib/agents/context-boundary", () => ({
 vi.mock("@/lib/agents/warm-summary-background", () => ({
   pendingCompactionBoundary: () => null,
   pendingCompactionBoundarySeq: () => null,
+  currentChannelSummaries: () => [],
 }));
 vi.mock("@/lib/agents/conversation-summary", () => ({ parseStoredTopics: () => null }));
 

package/app/api/v1/threads/[thread_id]/context-pin/route.ts

@@ -2,7 +2,7 @@ import { NextRequest, NextResponse } from "next/server";
 import { z } from "zod";
 import { getThread, getThreadMessageBySeq } from "@/lib/stores/threads";
 import { moveThreadContextBoundary } from "@/lib/agents/context-boundary";
-import { pendingCompactionBoundary, pendingCompactionBoundarySeq } from "@/lib/agents/warm-summary-background";
+import { currentChannelSummaries, pendingCompactionBoundary, pendingCompactionBoundarySeq } from "@/lib/agents/warm-summary-background";
 import { parseStoredTopics } from "@/lib/agents/conversation-summary";
 
 type Params = { params: Promise<{ thread_id: string }> };
@@ -43,6 +43,7 @@ export async function PATCH(req: NextRequest, { params }: Params) {
     warm_summary_source_messages: updated?.warm_summary_source_messages ?? null,
     warm_summary_source_chars: updated?.warm_summary_source_chars ?? null,
     warm_summary_topics: parseStoredTopics(updated?.warm_summary_topics),
+    warm_channel_summaries: currentChannelSummaries(thread_id, updated?.hot_since_seq),
     pending_hot_since: pendingCompactionBoundary(thread_id),
     pending_hot_since_seq: pendingCompactionBoundarySeq(thread_id),
   });

package/app/api/v1/threads/[thread_id]/route.ts

@@ -18,8 +18,9 @@ import { getMessageUsageByIds } from "@/lib/stores/message-usage";
 import { getAgentConfig } from "@/lib/stores/agent-configs";
 import { getModelConfig, getModelParams } from "@/lib/stores/model-config";
 import { messageToResponse, resolveContextWindowTokens } from "@/lib/api/serializers";
-import { pendingCompactionBoundary, pendingCompactionBoundarySeq } from "@/lib/agents/warm-summary-background";
+import { currentChannelSummaries, isManualCompactionRunning, pendingCompactionBoundary, pendingCompactionBoundarySeq } from "@/lib/agents/warm-summary-background";
 import { parseStoredTopics } from "@/lib/agents/conversation-summary";
+import { getThreadActivity } from "@/lib/agents/thread-activity";
 import { getCheckpointer } from "@/lib/agents/checkpointer";
 import { errorMessage } from "@/lib/utils/error";
 
@@ -84,9 +85,12 @@ export async function GET(req: NextRequest, { params }: Params) {
   return NextResponse.json({
     ...thread,
     warm_summary_topics: parseStoredTopics(thread.warm_summary_topics),
+    warm_channel_summaries: currentChannelSummaries(thread_id, thread.hot_since_seq),
     context_window_tokens: contextWindowTokens,
     pending_hot_since: pendingCompactionBoundary(thread_id),
     pending_hot_since_seq: pendingCompactionBoundarySeq(thread_id),
+    manual_compaction_pending: isManualCompactionRunning(thread_id),
+    activity: getThreadActivity(thread_id),
     // No server-side filtering: clients receive every message with its
     // `category` tag and apply the chat-panel filter toolbar on the
     // render side. Keeping the raw transcript over the wire means audit