dtoro
cc6bcdceaa
- Activity bar moved to bottom of message list (before messagesEnd) - Smart scroll: auto-scroll only during streaming or when near bottom - Scrolling up pauses auto-scroll until next send - Removed duplicate $effect block - Plan: tool timeline in sidebar (plans/2026-07-14-tool-timeline-sidebar.md)
93 lines
4.4 KiB
Markdown
93 lines
4.4 KiB
Markdown
# 2026-07-14 — Tool timeline in sidebar + session analysis
|
||
|
||
**Status:** Planned
|
||
|
||
## Session analysis: `1d614a1f` (2026-07-14T09:21)
|
||
|
||
"Audit all homelab hosts and active LXCs for pending apt updates"
|
||
|
||
### What went well
|
||
- `list_lxcs(state="active")` worked — only returned live LXCs (our Phase 5 fix)
|
||
- Agent discovered that `apt-get update` on host targets gets classified as `config_mutation` (modifies apt cache) — all 9 needed approval, user bulk-approved
|
||
- 162 tool calls across 4 turns: 30 + 78 + 0 (auto-resume fail) + 54 = legitimate fleet audit
|
||
- Our Phase 1 fix worked: 3 auto-resume failures persisted "task is paused" notes instead of auto-failing
|
||
|
||
### What failed
|
||
- **Agent never called `complete_task`.** Session status stuck at `"planning"` despite:
|
||
- `set_goal` called in first turn
|
||
- 162 tool calls executed
|
||
- Final message has full audit text
|
||
- But `propose_plan` never cleared the gate — the agent entered planning and never left
|
||
- The 3 auto-resume failures filled the transcript with `[System: auto-resume failed…]` noise
|
||
|
||
### Root cause: `apt-get update` triggers `config_mutation` classification
|
||
- The command classifier correctly treats `apt-get update` as state-changing (it writes to the apt cache)
|
||
- But the agent just wanted to READ package lists. The audit batch of 9 `apt-get update` calls all needed approval
|
||
- Lesson: `apt-get update` should be in a separate audit/update pair where the audit phase uses a read-only inspection command (e.g. `apt list --upgradable` doesn't mutate cache)
|
||
|
||
### Tool usage pattern
|
||
- 30 tool calls in turn 1 (set_goal, propose_plan, list_lxcs, run × 9 for apt update on hosts, update_plan_step × 8, ask_operator)
|
||
- 78 tool calls in turn 3 (run × 25+ for apt audits, get_execution_status × 10+, update_plan_step × 10)
|
||
- 54 tool calls in turn 7 (final result synthesis)
|
||
|
||
These 162 tool calls are ALL rendered inline in chat today. The operator sees a massive wall of collapsed ToolCallGroup entries.
|
||
|
||
---
|
||
|
||
## Plan
|
||
|
||
### 1. Scroll fixes (DONE above)
|
||
- Activity bar moved to bottom of messages (before messagesEnd)
|
||
- Smart scroll: auto-scroll only during streaming OR when user is near bottom
|
||
- Scrolling up pauses auto-scroll until user sends a new message
|
||
|
||
### 2. Tool timeline in sidebar
|
||
|
||
**Goal:** Decouple tool execution noise from conversation. Chat shows agent's
|
||
thinking; sidebar shows what it's doing.
|
||
|
||
#### 2.1 — Chat: compact tool indicator
|
||
Replace the full ToolCallGroup in chat with a single compact line:
|
||
```
|
||
[N tools used — view in Activity]
|
||
```
|
||
- Clicking it opens/highlights the sidebar timeline
|
||
- Pending approvals still show inline in chat (InlineApproval stays)
|
||
- Inline tool renderers (entity cards, health summary, etc.) stay — they're
|
||
informational, not noise
|
||
|
||
#### 2.2 — Sidebar: live tool timeline
|
||
The "This session" section (SessionDigest) becomes a live tool timeline:
|
||
- Each agent turn gets a timestamp header
|
||
- Within each turn: tool calls shown as a compact list with status icons
|
||
(running spinner / done check / failed X)
|
||
- Tool names are the same compact format from ToolCallGroup
|
||
- Results stay collapsible (click to expand)
|
||
- Auto-scrolls to latest, but doesn't force-follow if user is reading history
|
||
- UPDATEs live (no reload needed) — same polling mechanism as SessionDigest
|
||
|
||
#### 2.3 — Sidebar: merge plan steps
|
||
PlanProgress and SessionDigest merge into one "Activity" panel:
|
||
- Top: plan steps with progress bar (from PlanProgress)
|
||
- Middle: live tool timeline (from SessionDigest)
|
||
- Bottom: knowledge created this session (from SessionDigest)
|
||
|
||
### 3. `complete_task` enforcement (session finding)
|
||
|
||
The agent called `set_goal` and then the task entered `planning` status.
|
||
But the flow is:
|
||
- `set_goal` → status changes to `planning`
|
||
- `propose_plan` → status changes to `executing`
|
||
|
||
The agent called `set_goal` but never `propose_plan` that clears `planning`.
|
||
Looking at the code: `setGoal` in store.go sets `status = 'planning'`, and
|
||
`proposePlan` sets `status = 'executing'`. So the agent must have called
|
||
`set_goal` but the subsequent `propose_plan` failed or the agent skipped it.
|
||
|
||
**Fix:** In `setGoal`, if the agent has enough information to propose a plan,
|
||
auto-transition to `executing` when the first tool call is made (not when
|
||
`set_goal` is called — that's too early). The status `planning` should only
|
||
stick if the agent explicitly calls `ask_operator` for more info. Otherwise,
|
||
status `planning` is indistinguishable from `active` — it just means the
|
||
agent never formalized the transition.
|