feat(ux): surface memory and skill saves in the activity summary

[rodboev]

Thinking Path

• Memory writes and skill mutations happen inside collapsed tool cards, with no visual signal to the user that persistent state changed. The issue (Notify user when agent saves memory or skills #3340) asks for a notification so users know what was saved.
• The maintainer endorsed "the existing toast/activity surface" as the natural home. The Activity group summary ("Activity: N tools") is the closest existing pattern, always visible above each response.
• Rather than a badge, pill, or separate toggle, the simplest approach is to append memory/skill counts directly to the Activity label text with a comma: "Activity: 4 tools, 1 memory saved". No styling changes, no emoji, consistent with the existing label typography.
• Detection checks tc.name in {memory, skill_manage}, tc.args.action in {save, create, update, upsert}, and requires tc.done !== false and !tc.is_error, so running or failed calls are excluded from the count.

What Changed

• static/ui.js: Added _MEMORY_SKILL_TOOLS and _MEMORY_SKILL_WRITE_ACTIONS constant sets, _isMemorySkillWrite(tc) detection helper (gates on completed non-error calls). Modified _syncToolCallGroupSummary to count memory/skill write cards and append ", N memories saved" / ", N skills updated" to the Activity label text. Memory and skill write counts are subtracted from the total so the tool count reflects only non-memory tools (e.g., 3 visible cards = "2 tools, 1 memory saved"). Modified buildToolCard to attach _tcData = tc on the row element so the summary function can inspect tool name and args.
• tests/: Static source assertions verifying key code patterns (_MEMORY_SKILL_TOOLS, _isMemorySkillWrite, label suffix text, _tcData).

Why It Matters

Users get immediate feedback in the collapsed Activity summary when the agent saves a memory or updates a skill. The count is visible at a glance without expanding the group, and clicking to expand reveals the full tool card with entry details.

Verification

pytest tests/test_memory_skill_badge.py -v --timeout=60
pytest tests/ -v --timeout=60
# Manual: start a session, prompt the agent to save a memory
# Verify: Activity label reads "Activity: N tools, 1 memory saved"

Risks / Follow-ups

• _tcData attaches a reference to the tc object on the DOM element. This is a common pattern in the codebase (e.g., row.dataset.liveTid), but a future maintainer might prefer a WeakMap. Low risk since tool cards are short-lived DOM nodes.
• The suffix counts completed writes only; in-progress memory saves show as regular tool cards until done.
• A richer display (expandable memory content preview, like ChatGPT's "Saved to memory" block) is a natural follow-up. This PR establishes the detection infrastructure; the display can evolve.
• Review/undo affordance (mentioned in the issue) is out of scope. It needs a backend handle on the saved entry.

Screenshots

• Before (Activity group with no memory indication):

---

• After (label reads "Activity: 1 tool, 1 memory saved"):

Model Used

Claude Opus 4.8 via Claude Code CLI

Closes #3340

[greptile-apps]

Greptile Summary

This PR surfaces memory and skill write operations in the Activity group label, so users can see at a glance that persistent state changed without expanding the tool card group. Detection is gated on completed, non-error tool calls whose action/command field is in the write-action set.

• Adds _isMemorySkillWrite helper and two constant sets to static/ui.js; augments _syncToolCallGroupSummary to append , N memories saved / , N skills updated to the Activity label; attaches tc as _tcData on each tool card row for the summary function to inspect.
• Adds tests/test_memory_skill_badge.py with four source-text grep tests (already flagged in prior review threads as too shallow to catch logic regressions).

Confidence Score: 5/5

Safe to merge; the change is a purely additive UI label enhancement with no writes to state, no new network calls, and no HTML injection.

The detection helper and label-building logic are simple, narrowly scoped, and only affect the collapsed Activity group label text. Tool card rendering, streaming, and session state are untouched. The identified concerns are theoretical and do not affect the happy path in the current DOM structure.

No files require special attention beyond the already-discussed test coverage gap in tests/test_memory_skill_badge.py.

Important Files Changed

Filename	Overview
static/ui.js	Adds _isMemorySkillWrite helper, two constant sets, and augments _syncToolCallGroupSummary to append memory/skill write counts to the Activity label; attaches _tcData reference on each tool card row element.
tests/test_memory_skill_badge.py	New test file with four tests that grep ui.js for static string presence; no behavioral or DOM-level assertions, so logic regressions would not be caught.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[tool_start event] -->|buildToolCard tc.done=false| B[row._tcData = tc added to DOM]
    B --> C[_syncToolCallGroupSummary]
    C --> D{_isMemorySkillWrite}
    D -->|done===false| E[otherCount - Label: Activity N tools]
    F[tool_complete event] -->|buildToolCard tc.done=true replaces row| G[row._tcData updated]
    G --> H[_syncToolCallGroupSummary]
    H --> I{_isMemorySkillWrite}
    I -->|name=memory| J[memCount++]
    I -->|name=skill_manage| K[skillCount++]
    I -->|other| L[otherCount++]
    J --> M[Compose label]
    K --> M
    L --> M
    M --> N[Activity N tools X memories saved]

Loading

_{Reviews (3): Last reviewed commit: "fix(ux): strip leading comma from Activi..." | Re-trigger Greptile}

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.

[rodboev]

Edited to note that we decrement the tool count by the number of memories saved, since we now break out the memories.

[rodboev]

The leading-comma bug in the memory-only path was a good catch. Fixed in b6fe6a3: when otherCount is 0, the label now reads "Activity: 1 memory saved" instead of "Activity, 1 memory saved".

[nesquena-hermes]

Shipped in v0.51.282 (Release IX) 🎉

Thank you @rodboev — your Activity-summary memory/skill-save feature is now live, and the underlying idea + infrastructure (per-row classification, label suffix, tool-count subtraction) were exactly right. We made two correctness fixes while absorbing it:

1. Detection vocabulary — the action sets {save, create, update, upsert} didn't match the real agent tool enums (memory.action = add/replace/remove; skill_manage.action = create/patch/edit/delete/write_file/remove_file), so the counter never actually fired on a real save. Rewrote into per-tool predicates with the correct enums (saves = memory add/replace + skill create/patch/edit/write_file; deletions excluded so the "saved"/"updated" verbs stay accurate).

2. Snapshot/restore persistence — classification on row._tcData (a JS property) doesn't survive the live tool-call group's outerHTML/innerHTML snapshot+restore on session switch, so the suffix silently vanished after switching away and back. Now also stamped as durable data-memory-save / data-skill-update attributes with a fallback read in the summary.

Also swapped the static source-string assertions for a node-driven behavioral test (11 cases) and verified live: counter fires, survives a real HTML round-trip identically, no double-count, and sessions with no memory/skill writes render the unchanged "Activity: N tools".

Your authorship is preserved via a Co-authored-by trailer on the release commit. Closing as merged; closes #3340.