@animalabs/connectome-host 0.3.10 → 0.4.0

package/CHANGELOG.md

@@ -1,7 +1,88 @@
 # Changelog
 
+## 0.4.0
+
+### Changed
+
+- **context-manager ^0.6.0** — the fatal coverage invariant: a compile
+  refuses (`OverBudgetError` / `UncoveredDropError`) rather than shipping a
+  context with silently-dropped messages, and recall-pair pricing includes
+  reasoning carriers (fixes the permanent compile wedge / silent middle loss
+  on carrier-bearing stores). Default `overBudgetGraceRatio` is now 0.02.
+- **agent-framework ^0.7.0** — host-side recovery for context refusals: the
+  OverBudget drain breaker also kicks for `UncoveredDropError`, and a
+  `context-refusal` ops alert fires immediately (fleet-watch) with the
+  recovery knobs named. Plus the context-settings preview surface and the
+  workspace read cap.
+
+### Added
+
+- `compressionMaxTokens` recipe passthrough — cap compression output for
+  models with low output ceilings (2c78936).
+
 ## Unreleased
 
+### Fixed
+
+- **TUI bug sweep** (#64): operator-safety and observability fixes.
+  - `/quit` confirm no longer treats arbitrary input as consent — only an
+    explicit `y`/`yes` (or re-typed `/quit`) kills fleet children, `d`
+    detaches, anything else cancels; a typed-through message is restored to
+    the input (paste referents intact) instead of discarded. Ctrl+C now goes
+    through the same confirmation; a second Ctrl+C force-quits.
+  - `/checkpoint` records the message position and `/restore` branches back
+    to it (previously restored to the branch head — rolling back nothing);
+    repeat restores at the same position are a no-op, and an unreachable
+    position degrades to the branch head with an explicit note.
+  - Session switch fully resets TUI observability state (tree aggregator,
+    stream subscriptions, per-agent caches) — fleet subtrees no longer
+    freeze after `/session switch`.
+  - Memory: peek logs / transcripts / scrollback capped, and detached
+    renderables are `destroy()`ed so their native text buffers are actually
+    freed (the fleet view leaked one buffer per line per 500ms repaint).
+  - Agent-name resolution is exact (`shortAgentName`, fork `-d{depth}`
+    scheme included) instead of substring matching that cross-wired agents
+    with prefix-overlapping names; peek tails no longer clip the newest
+    lines; fleet-view kill/restart failures are surfaced; per-round context
+    size (`ctx:`) and session totals (`Σ`) are separate status segments;
+    synesthete summaries moved off the render path and back off 30s after
+    a failed call instead of retrying at 2 Hz.
+  - Smaller UX: peek works on finished subagents (final runtime shown),
+    fork `done` summaries always print a chat line, Esc/Ctrl+B work from
+    the fleet view, paste placeholders survive `]` in the pasted text,
+    `/help` documents `/find` and `/branchto`, `/clear` with arguments
+    clears.
+
+### Docs
+
+- Synced stale documentation with the current build: repos marked public
+  (AGENT-ONBOARDING), `forking-knowledge-miner` → `connectome-host`
+  naming, webui default port corrected to 7340, DEV-ENVIRONMENT
+  branch/version table refreshed (all feature branches merged),
+  LOCUS-ROUTING and both root plan docs marked implemented.
+
+### Changed
+
+- **Tool-bloat reduction**: subscription-gc's `set_channel_idle_limit` /
+  `list_channel_idle_limits` tools folded into `agent_settings` as the
+  `channel_idle_limits` field (per-entry merge; number / `"off"` /
+  `"default"`-or-null to clear), following the reasoning-controls
+  precedent. The old tool names remain routable (undeclared), so agent
+  muscle memory keeps working; agents just no longer carry the two extra
+  tool schemas. `get` also reports read-only `channel_idle_default`,
+  `channel_idle_counters`, and `channel_idle_pinned`, preserving what
+  `list_channel_idle_limits` exposed. Updates are all-or-nothing: a patch
+  with any invalid entry applies none of its entries.
+- **GC pins split from agent overrides**: ChannelModeModule now holds
+  debounced channels open via an internal `pin_channel_idle_limit` verb
+  and a separate pins layer, instead of writing an `"off"` override.
+  Consequences: a blanket `agent_settings reset` clears only agent-set
+  limits — it can no longer silently re-enable auto-close on a channel in
+  debounced mode — and a pre-existing agent override now survives a
+  debounced→mentions round-trip rather than being reset to default.
+  (Pins persisted by earlier builds as `"off"` overrides stay agent-level
+  until the next mode change re-asserts them as pins.)
+
 ## 0.3.10 — 2026-07-21
 
 ### Added

package/HEADLESS-FLEET-PLAN.md

@@ -1,6 +1,6 @@
 # Headless Daemon Mode & Fleet Orchestration — Plan
 
-Status: draft • Authors: conversation between @tengro and Claude, 2026-04-21
+Status: implemented (all phases shipped — see [UNIFIED-TREE-PLAN.md](./UNIFIED-TREE-PLAN.md) "Implementation outcome") • Authors: conversation between @tengro and Claude, 2026-04-21
 
 ## Goal
 

package/UNIFIED-TREE-PLAN.md

@@ -1,5 +1,7 @@
 # Unified Tree Plan: Fleet Children as Subagents in the TUI
 
+**Status: shipped** — all phases landed; see "Implementation outcome" below. The per-phase headers preserve the plan as written.
+
 Goal: collapse the two-paradigm UI (`SubagentModule`'s in-process tree vs. `FleetModule`'s flat child list) into a single tree-rendering paradigm where fleet children appear and behave like subagents — same readouts (context tokens, phase, tool-call count), same unfold-children affordance, same row component.
 
 Companion to [HEADLESS-FLEET-PLAN.md](./HEADLESS-FLEET-PLAN.md), which defines the underlying IPC.

package/docs/AGENT-ONBOARDING.md

@@ -22,7 +22,7 @@ A connectome agent is:
 - an **`.env`** with secrets and per-deploy config;
 - a **chronicle** (`data/`) — the event-sourced memory store, optionally seeded by importing
   a prior conversation;
-- run by the **connectome-host** (`forking-knowledge-miner`, run via `bun src/index.ts <recipe> --headless`);
+- run by the **connectome-host** (`bun src/index.ts <recipe> --headless`);
 - under a **systemd --user** service, with a sibling **terminal-sessions** daemon for the shell tool.
 
 It reaches the model through **membrane** (Anthropic-format adapter; can target a gateway via
@@ -164,10 +164,10 @@ shell. First, the **hard prerequisites** (install / confirm before anything else
 or their box owner if unsure):
 - **git**, **Bun** (`curl -fsSL https://bun.sh/install | bash`), **node ≥ 20** (via `nvm`);
 - a **Rust toolchain** (`rustup`) — `chronicle` has a native (napi) component built from Rust;
-- **access to the private `anima-research` / `antra-tess` GitHub orgs.** The connectome
-  packages are **not public**. If `git clone` / `bun install` can't reach them, *stop* — the
-  user must be granted org access (or handed the code) first. There's no way around this; it's
-  the gate before all the mechanics.
+- the `anima-research` / `antra-tess` GitHub repos. These are **public** (as of 2026-07;
+  they previously required org access), and the `@animalabs/*` npm packages install without
+  auth. If a `git clone` / `bun install` fails on access, check whether the repo moved
+  rather than assuming a permissions gate.
 
 The repos (clone all as siblings under one dir, canonically `~/connectome-local/`):
 
@@ -203,7 +203,7 @@ errors and logs `[webui] listening`. Then continue with the per-agent install di
 
 **Pick non-colliding ports.** Every agent on the box needs a unique:
 - **shell-daemon port** (`SESSION_SERVER_PORT`, default 3100)
-- **webui port** (default 7342)
+- **webui port** (module default 7340; binds `0.0.0.0` and requires `basicAuth` unless set to loopback)
 
 Check what's taken (`ss -tlnp`) and pick the next pair (e.g. 3101 / 7343).
 
@@ -381,7 +381,7 @@ Two `systemd --user` units in `~/.config/systemd/user/`:
 
 1. **`terminal-sessions.service`** — the shell daemon, `--host localhost --headless --token
    ${SESSION_SERVER_TOKEN}`, env `SESSION_SERVER_PORT`. `enable` + `start` it first.
-2. **`<agent>-agent.service`** — `bun .../forking-knowledge-miner/src/index.ts
+2. **`<agent>-agent.service`** — `bun .../connectome-host/src/index.ts
    .../recipes/<agent>.json --headless`, `EnvironmentFile=...env`, `Environment=PATH=<node bin>:<bun bin>:...`,
    `After=/Wants=terminal-sessions.service`, `Restart=always`.
 
@@ -472,7 +472,8 @@ robust fallback strategy.
 
 ## Appendix — quick reference
 
-- **Run host:** `bun forking-knowledge-miner/src/index.ts <recipe.json> --headless`
+- **Run host:** `bun connectome-host/src/index.ts <recipe.json> --headless` (legacy
+  checkouts may still name the directory `forking-knowledge-miner`)
 - **Reach loopback webui:** `ssh -L <port>:localhost:<port> <login-user>@<box>` → `http://localhost:<port>`
 - **Live makeup:** `GET /debug/context/makeup` (basic auth) — segments + exact total tokens
 - **Compiled context:** `GET /debug/context`

package/docs/DEPLOYMENTS.md

@@ -9,9 +9,9 @@ end-to-end, see [`AGENT-ONBOARDING.md`](./AGENT-ONBOARDING.md).
 
 ## TL;DR
 
-An agent deployment is the Connectome stack (`forking-knowledge-miner` host +
+An agent deployment is the Connectome stack (`connectome-host` +
 `agent-framework` + MCPL servers) plus a per-agent install dir. Agents run
-**headless** (`bun forking-knowledge-miner/src/index.ts <recipe> --headless`),
+**headless** (`bun connectome-host/src/index.ts <recipe> --headless`),
 each as its own OS user, supervised by launchd (macOS) or systemd-user (Linux).
 
 > **Model integrity matters.** Each agent has a *correct* model and must stay on

package/docs/DEV-ENVIRONMENT.md

@@ -1,6 +1,6 @@
 # Connectome Dev Environment — Setup Guide
 
-**Status:** Working notes, uncommitted. Snapshot of the dev layout as of 2026-05-30.
+**Status:** Working notes. Snapshot of the dev layout as of 2026-07-22 (originally 2026-05-30; all feature branches in the original table have since merged to `main`).
 **Goal:** Reproduce the current development state — every component as a local git
 checkout, wired together so the host runs against editable source.
 
@@ -24,17 +24,22 @@ that holds a recipe + `.env` + chronicle data and references the code by absolut
 
 All cloned as siblings under `~/connectome-local/`:
 
-| Dir | GitHub repo | Branch | Ver | Role |
+| Dir | GitHub repo | Branch | Ver (2026-07-22) | Role |
 |---|---|---|---|---|
-| `forking-knowledge-miner` | `anima-research/connectome-host` | `feat/tui-mem-stats-and-strategy-passthrough` | 0.3.0 | the host app (run via **bun**) |
-| `agent-framework` | `anima-research/agent-framework` | `fix/lazy-register-locus-and-send-failure-marker` | 0.5.0 | host runtime: gate, MCPL orchestration, locus routing, `think` |
-| `discord-mcpl` | `anima-research/discord-mcpl` | `main` | 0.1.0 | Discord surface (MCPL server) |
-| `heartbeat-mcpl` | `anima-research/heartbeat-mcpl` | `main` | 0.1.0 | periodic self-wake (MCPL server) |
-| `terminal-sessions-mcp` | `antra-tess/terminal-sessions-mcp` ⚠️ | `main` | 1.0.1 | shell: session daemon (ws://localhost:3100) + per-agent MCP stdio frontend |
-| `membrane` | `antra-tess/membrane` ⚠️ | `fix/before-request-hook-on-stream-paths` | 0.5.47 | LLM client lib — **single shared instance required** |
-| `context-manager` | `anima-research/context-manager` | `feat/adaptive-resolution` | 0.4.0 | context compilation / autobiographical memory |
-| `chronicle` | `anima-research/chronicle` | `main` | 0.1.1 | record / chronicle store |
-| `mcpl-core-ts` | `anima-research/mcpl-core-ts` | `main` | 0.1.0 | MCPL protocol types (discord-mcpl path dep) |
+| `forking-knowledge-miner` | `anima-research/connectome-host` | `main` | 0.3.10 | the host app (run via **bun**) |
+| `agent-framework` | `anima-research/agent-framework` | `main` | 0.6.10 | host runtime: gate, MCPL orchestration, locus routing, `think` |
+| `discord-mcpl` | `anima-research/discord-mcpl` | `main` | 0.1.4 | Discord surface (MCPL server) |
+| `heartbeat-mcpl` | `anima-research/heartbeat-mcpl` | `main` | 0.1.3 | periodic self-wake (MCPL server) |
+| `terminal-sessions-mcp` | `antra-tess/terminal-sessions-mcp` ⚠️ | `main` | 1.6.0 | shell: session daemon (ws://localhost:3100) + per-agent MCP stdio frontend |
+| `membrane` | `antra-tess/membrane` ⚠️ | `main` | 0.5.74 | LLM client lib — **single shared instance required** |
+| `context-manager` | `anima-research/context-manager` | `main` | 0.5.14 | context compilation / autobiographical memory |
+| `chronicle` | `anima-research/chronicle` | `main` | 0.2.7 | record / chronicle store |
+| `mcpl-core-ts` | `anima-research/mcpl-core-ts` | `main` | 0.2.1 | MCPL protocol types |
+
+> Versions drift; treat the column as a dated snapshot. The published npm
+> releases now track `main` closely (typically within a patch), so the
+> stock `bun install` path is sufficient for host-level work — use the
+> checkout layout below when editing the libraries themselves.
 
 > ⚠️ **Org note:** `terminal-sessions-mcp` and `membrane` still live under the
 > personal `antra-tess` org, not `anima-research`. Consider migrating them for
@@ -51,19 +56,15 @@ All cloned as siblings under `~/connectome-local/`:
 ```bash
 mkdir -p ~/connectome-local && cd ~/connectome-local
 
-git clone -b feat/tui-mem-stats-and-strategy-passthrough \
-  git@github.com:anima-research/connectome-host.git forking-knowledge-miner
-git clone -b fix/lazy-register-locus-and-send-failure-marker \
-  git@github.com:anima-research/agent-framework.git
-git clone -b main  git@github.com:anima-research/discord-mcpl.git
-git clone -b main  git@github.com:anima-research/heartbeat-mcpl.git
-git clone -b main  git@github.com:antra-tess/terminal-sessions-mcp.git
-git clone -b fix/before-request-hook-on-stream-paths \
-  git@github.com:antra-tess/membrane.git
-git clone -b feat/adaptive-resolution \
-  git@github.com:anima-research/context-manager.git
-git clone -b main  git@github.com:anima-research/chronicle.git
-git clone -b main  git@github.com:anima-research/mcpl-core-ts.git
+git clone git@github.com:anima-research/connectome-host.git forking-knowledge-miner
+git clone git@github.com:anima-research/agent-framework.git
+git clone git@github.com:anima-research/discord-mcpl.git
+git clone git@github.com:anima-research/heartbeat-mcpl.git
+git clone git@github.com:antra-tess/terminal-sessions-mcp.git
+git clone git@github.com:antra-tess/membrane.git
+git clone git@github.com:anima-research/context-manager.git
+git clone git@github.com:anima-research/chronicle.git
+git clone git@github.com:anima-research/mcpl-core-ts.git
 ```
 
 ---
@@ -82,8 +83,9 @@ terminal-sessions-mcp
 
 In each: `npm install && npm run build` (build script is `tsc`).
 
-> `discord-mcpl` depends on `@connectome/mcpl-core` via `file:../mcpl-core-ts`
-> (a path dep), so build `mcpl-core-ts` first.
+> `discord-mcpl` now depends on `@animalabs/mcpl-core` as a regular npm
+> dependency (the old `file:../mcpl-core-ts` path dep is gone). A sibling
+> `mcpl-core-ts` checkout is only needed when changing mcpl-core itself.
 
 ---
 

package/docs/LOCUS-ROUTING-DESIGN.md

@@ -1,7 +1,10 @@
 # Conversational Locus & Output Routing — Design Note
 
-**Status:** Agreed design, implementation deferred.
-**Date:** 2026-05-29
+**Status:** Implemented — the locus now lives in `agent-framework`
+(`src/mcpl/channel-registry.ts`; `routeSpeech` fires at turn completion in
+`framework.ts`) and has grown beyond this note (global-locus fallback for
+heartbeats, fork/home-channel semantics). Kept as design rationale.
+**Date:** 2026-05-29 (status updated 2026-07-21)
 **Context:** Surfaced while wiring periodic heartbeats for an agent instance
 running headless on a Linux VPS.
 
@@ -95,6 +98,9 @@ discord-mcpl patch.
 
 ## Bugs found in passing (separate fixes, worth upstreaming)
 
+*(Update 2026-07-21: #3 has been fixed upstream — `agent-framework`'s
+feature-set-manager now normalizes array→record on ingest.)*
+
 1. **discord-mcpl `connect()` race** — resolved on `login()` (before gateway
    READY), so `registerDiscordChannels` enumerated an empty `guilds.cache`.
    *Fixed*: `connect()` now awaits the `ready` event. (Not yet

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@animalabs/connectome-host",
-  "version": "0.3.10",
+  "version": "0.4.0",
   "description": "General-purpose agent TUI host with recipe-based configuration",
   "type": "module",
   "scripts": {
@@ -14,10 +14,10 @@
     "postinstall": "test -d web && npm --prefix web install && npm --prefix web run build || true"
   },
   "dependencies": {
-    "@animalabs/agent-framework": "^0.6.8",
+    "@animalabs/agent-framework": "^0.7.0",
     "@animalabs/chronicle": "^0.2.0",
-    "@animalabs/context-manager": "^0.5.13",
-    "@animalabs/membrane": "^0.5.73",
+    "@animalabs/context-manager": "^0.6.0",
+    "@animalabs/membrane": "^0.5.75",
     "@opentui/core": "^0.1.82"
   },
   "devDependencies": {

package/src/commands.ts

@@ -151,7 +151,9 @@ export function handleCommand(command: string, app: AppContext): CommandResult {
           { text: '  /restore <name>        Restore to checkpoint', style: 'system' },
           { text: '  /branches              List Chronicle branches', style: 'system' },
           { text: '  /checkout <name>       Switch to branch', style: 'system' },
-          { text: '  /history               Show state transitions', style: 'system' },
+          { text: '  /history [n]           Show state transitions (last n)', style: 'system' },
+          { text: '  /find <text>           Search messages for text', style: 'system' },
+          { text: '  /branchto <msgId>      Branch from a specific message', style: 'system' },
           { text: '  /mcp list              List MCPL servers', style: 'system' },
           { text: '  /mcp add <id> <cmd>    Add/overwrite server', style: 'system' },
           { text: '  /mcp remove <id>       Remove a server', style: 'system' },
@@ -747,7 +749,13 @@ function handleCheckpoint(app: AppContext, name?: string): CommandResult {
   if (!cm) return { lines: [{ text: 'No agent context manager.', style: 'system' }] };
 
   const branch = cm.currentBranch();
-  app.branchState.checkpoints.set(name, { branchName: branch.name });
+  // A checkpoint is a *position*, not just a branch: record the id of the
+  // last message so /restore can branch back to this exact point even after
+  // the branch head has moved on. branchName alone would restore to the
+  // branch HEAD — i.e. not roll anything back.
+  const { messages } = cm.queryMessages({});
+  const messageId = messages.length > 0 ? messages[messages.length - 1]!.id : undefined;
+  app.branchState.checkpoints.set(name, { branchName: branch.name, messageId });
 
   return { lines: [{ text: `Checkpoint "${name}" saved at branch ${branch.name} (head: ${branch.head}).`, style: 'system' }] };
 }
@@ -774,12 +782,50 @@ function handleRestore(app: AppContext, name?: string): CommandResult {
   const cm = getAgentCM(app.framework);
   if (!cm) return { lines: [{ text: 'No agent context manager.', style: 'system' }] };
 
-  const target = point.branchName;
+  // Restore to the recorded *position*. If the checkpoint captured a message
+  // id, branch at that message (same machinery as /undo) — switching to the
+  // stored branch name alone would land on its head, which by now may
+  // include everything the user wanted to roll back. Checkpoints from before
+  // the messageId field (or taken on an empty branch) fall back to the head.
+  let target = point.branchName;
+  let exact = false;
+  if (point.messageId) {
+    // A checkpoint is a POSITION, so position equality is the whole
+    // "already there" test. Comparing branch names too would kill the guard
+    // after the first restore (you'd be on restore-{name}-{ts}, never the
+    // original branch again) and every repeat /restore would mint another
+    // branch pointing at the same message.
+    const { messages } = cm.queryMessages({});
+    const atCheckpoint = messages.length > 0
+      && messages[messages.length - 1]!.id === point.messageId;
+    if (atCheckpoint) {
+      return { lines: [{ text: `Already at checkpoint "${name}".`, style: 'system' }] };
+    }
+    try {
+      // branchAt returns the new branch's NAME (the only thing switchBranch accepts).
+      target = cm.branchAt(point.messageId, `restore-${name}-${Date.now()}`);
+      exact = true;
+      // The record follows the position: keep the newest branch that holds
+      // the checkpoint message as the fallback for future restores.
+      app.branchState.checkpoints.set(name, { branchName: target, messageId: point.messageId });
+    } catch {
+      // branchAt resolves ids against the CURRENT branch's view of the log;
+      // after e.g. /undo the checkpoint message may not be reachable from
+      // here. Degrade to the checkpoint's branch head (the pre-messageId
+      // behavior) rather than failing the restore outright.
+      target = point.branchName;
+    }
+  }
 
   // switchBranch is async — strategy reinit needs to complete before next op.
   const asyncWork = Promise.resolve(cm.switchBranch(target))
     .then(() => ({
-      lines: [{ text: `Restored to checkpoint "${name}" (branch: ${target}).`, style: 'system' as const }],
+      lines: [
+        { text: `Restored to checkpoint "${name}" (branch: ${target}).`, style: 'system' as const },
+        ...(point.messageId && !exact
+          ? [{ text: '  (note: exact checkpoint position unreachable from the current branch — restored to the branch head instead)', style: 'system' as const }]
+          : []),
+      ],
       branchChanged: true,
     }))
     .catch(err => ({

package/src/framework-agent-config.ts

@@ -55,5 +55,8 @@ export function buildFrameworkAgentConfig(
     ...(recipe.agent.sameRoundThinkTextPolicy !== undefined
       ? { sameRoundThinkTextPolicy: recipe.agent.sameRoundThinkTextPolicy }
       : {}),
+    ...(recipe.agent.proseRouting !== undefined
+      ? { proseRouting: recipe.agent.proseRouting }
+      : {}),
   };
 }

package/src/framework-strategy.ts

@@ -67,6 +67,9 @@ export function buildFrameworkStrategy(
     headWindowTokens: strategyConfig?.headWindowTokens ?? 4000,
     recentWindowTokens: strategyConfig?.recentWindowTokens ?? 30000,
     compressionModel: strategyConfig?.compressionModel ?? model,
+    ...(strategyConfig?.compressionMaxTokens !== undefined
+      ? { compressionMaxTokens: strategyConfig.compressionMaxTokens }
+      : {}),
     autoTickOnNewMessage: true,
     maxMessageTokens: strategyConfig?.maxMessageTokens ?? 10000,
     ...(strategyType === 'frontdesk' ? { timeZone } : {}),

package/src/index.ts

@@ -46,6 +46,7 @@ import { ChannelModeModule } from './modules/channel-mode-module.js';
 import { WebUiModule } from './modules/web-ui-module.js';
 import { ObserversModule } from './modules/observers-module.js';
 import { McplAdminModule } from './modules/mcpl-admin-module.js';
+import { TtsRelayModule } from './modules/tts-relay-module.js';
 import { loadMcplServers, applyAgentOverlay, DEFAULT_CONFIG_PATH, DEFAULT_AGENT_OVERLAY_PATH } from './mcpl-config.js';
 import { SessionManager } from './session-manager.js';
 import { resolveAgentName } from './agent-name.js';
@@ -357,6 +358,14 @@ async function createFramework(
     moduleInstances.push(new ObserversModule({ path: observersPath }));
   }
 
+  // TTS relay tap — opt-in per recipe. Pure trace-bus consumer: mirrors the
+  // agent's streaming generation to a melodeus-tts-relay /bot endpoint and
+  // trims posted messages on voice interruptions. See modules.ttsRelay in
+  // recipe.ts for the config contract (url/token validated at recipe load).
+  if (modules.ttsRelay) {
+    moduleInstances.push(new TtsRelayModule(modules.ttsRelay));
+  }
+
   // Extension-registered modules. Instantiated last so built-in modules keep
   // their historical positions; each factory gets the minimal runtime context
   // plus its declaring extension's config blob.

package/src/modules/channel-mode-module.ts

@@ -200,11 +200,13 @@ export class ChannelModeModule implements Module {
     }
 
     // 3. Pin subscription-gc off so a chatty channel doesn't auto-close
-    //    itself back out of debounced mode.
+    //    itself back out of debounced mode. The pin layer sits above
+    //    agent-level overrides, so agent_settings reset/update can't undo it
+    //    and any pre-existing override survives the mode round-trip.
     steps.push(
-      await this.callToolStep('gc-off', `${this.gcModuleName}--set_channel_idle_limit`, {
+      await this.callToolStep('gc-pin', `${this.gcModuleName}--pin_channel_idle_limit`, {
         channelId,
-        limit: 'off',
+        pinned: true,
       }),
     );
 
@@ -225,11 +227,12 @@ export class ChannelModeModule implements Module {
     // 2. Close ambient traffic through the generic host lifecycle.
     steps.push(await this.callToolStep('close', 'channel_close', { channelId, serverId: this.serverId }));
 
-    // 3. Restore the default auto-close limit.
+    // 3. Release the pin; the channel's agent override (if any) or the
+    //    default limit applies again.
     steps.push(
-      await this.callToolStep('gc-default', `${this.gcModuleName}--set_channel_idle_limit`, {
+      await this.callToolStep('gc-unpin', `${this.gcModuleName}--pin_channel_idle_limit`, {
         channelId,
-        limit: 'default',
+        pinned: false,
       }),
     );