fix: collapse tool calls during streaming to prevent truncation hiding agent text

[wangyuyan-agent]

Summary

Fixes #242 — streaming truncation hides agent interaction prompts, causing thread deadlock.

┌──────────────────────────────────────────────────────────────────┐
│ BEFORE — streaming phase (bug)                                   │
│                                                                  │
│  compose_display() output:                                       │
│  ┌────────────────────────────────────────────────────────────┐  │
│  │ ✅ `Spawning agent crew`                                   │  │
│  │ ✅ `Searching for '(?i)openclaw'`                          │  │
│  │ ✅ `Running: find / -maxdepth 4 -iname "*openclaw*"`      │  │
│  │ ✅ `Running: curl -s "https://api.github.com/search/...`  │  │
│  │ ✅ `Running: curl -s "https://api.github.com/repos/...`   │  │
│  │ ✅ `Running: curl -s "https://api.github.com/repos/...`   │  │
│  │ ✅ `Running: echo "=== TOP CONTRIBUTORS ===" ; curl...`   │  │
│  │ ✅ `Running: echo "=== Cosmos/IBC ===" ; curl -s ...`     │  │
│  │ ✅ `Reading listing claw-info, listing openclaw`           │  │
│  │ ✅ `Reading README.md:1, LICENSE:1-30`                     │  │
│  │ ✅ `Running: node -e "const p = require(...)"`             │  │
│  │ ✅ `Reading tool_d7d3ac0c...:1-50`                        │  │
│  │ ✅ `Reading 2026-04-02.md:1-60, gateway-lifecycle...`     │  │
│  │ ✅ `Summarizing`                                           │  │
│  │ ✅ `Summarizing`                                           │  │
│  │                                                            │  │
│  │  "I found issues in config-7. Should I fix it?"            │  │
│  └────────────────────────────────────────────────────────────┘  │
│       ▲                                                          │
│       │  truncate_chars(&content, 1900) keeps FIRST 1900 chars   │
│       │  15 tool lines ≈ 1200+ chars → agent text GONE           │
│       │                                                          │
│  User sees: 15 lines of ✅ + "…"                                 │
│  Agent waits for reply it will never get → DEADLOCK              │
└──────────────────────────────────────────────────────────────────┘

┌──────────────────────────────────────────────────────────────────┐
│ AFTER — streaming phase (fix)                                    │
│                                                                  │
│  compose_display(..., streaming=true):                            │
│  ┌────────────────────────────────────────────────────────────┐  │
│  │ ✅ 15 tool(s) completed                    (~30 chars)     │  │
│  │                                                            │  │
│  │ I found issues in config-7. Should I fix it?               │  │
│  │                                            ← VISIBLE ✅    │  │
│  │                                                            │  │
│  │           (1800+ chars remaining for text)                 │  │
│  └────────────────────────────────────────────────────────────┘  │
│                                                                  │
│  Fallback: if STILL over 1900 → tail-priority truncation         │
│            keeps LAST 1900 chars → latest text always visible    │
└──────────────────────────────────────────────────────────────────┘

┌──────────────────────────────────────────────────────────────────┐
│ AFTER — final message (unchanged)                                │
│                                                                  │
│  compose_display(..., streaming=false):                           │
│  ┌────────────────────────────────────────────────────────────┐  │
│  │ ✅ `Spawning agent crew`                                   │  │
│  │ ✅ `Searching for '(?i)openclaw'`                          │  │
│  │ ✅ `Running: find / ...`                                   │  │
│  │ ... (all tools fully expanded)                             │  │
│  │ ✅ `Summarizing`                                           │  │
│  │                                                            │  │
│  │ I found issues in config-7. Should I fix it?               │  │
│  └────────────────────────────────────────────────────────────┘  │
│  split_message() handles >2000 chars → multiple Discord msgs     │
└──────────────────────────────────────────────────────────────────┘

Collapsing rules (streaming only):
  finished ≤ 3  →  individual lines (no change)
  finished > 3  →  "✅ N tool(s) completed"
  with failures →  "✅ 3 · ❌ 2 tool(s) completed"
  running       →  always individual (guarded at same threshold)

Changes

Single file: src/discord.rs + removal of unused truncate_chars in src/format.rs

1. Threshold-based tool collapsing in compose_display()

Added a streaming mode that collapses finished tool entries when count exceeds TOOL_COLLAPSE_THRESHOLD (3):

• ≤ 3 finished tools → shown individually (no UX regression)
• > 3 finished tools → ✅ N tool(s) completed (one line)
• Failed tools → counted separately: ✅ 3 · ❌ 2 tool(s) completed
• Running tools → always shown individually, with same threshold guard for parallel edge case (🔧 N more running)

Why 3? A tool line averages 40–80 chars. 3 lines ≈ 120–240 chars = 6–13% of the 1900-char budget, leaving 1660+ for agent text. Also the practical "glanceable" limit.

2. Tail-priority truncation (defense-in-depth)

Streaming edit now keeps the last 1900 chars instead of the first, with …(truncated) indicator. Even if collapsing isn't enough, the user always sees the most recent agent output.

3. Final message unchanged

Final edit uses full expansion (streaming: false) — split_message() handles multi-message delivery, so truncation isn't a risk. Collapsing is only for the streaming phase.

Test results

7 new unit tests added (33 total, all passing):

• Threshold boundary (3 vs 4 tools)
• Mixed completed + failed count accuracy
• Running tools alongside collapsed finished tools
• Parallel running tools guard
• Non-streaming full expansion
• Multi-byte character safety (CJK, emoji)

Screenshots

2 tools — individual display (below threshold):

15 tools — collapsed + agent text visible:

18 tools — collapsed: