Context window balloons at fresh-session start (~40-70K baseline tokens) and full-file re-injection on every auto-compact

**Environment:** Claude Code 2.1.259 (macOS Darwin 24.5.0, arm64), VS Code extension 2.1.259, Sonnet 5 (Pro plan)

**Summary:** Two related but distinct issues, both confirmed with direct evidence from session transcript files (`~/.claude/projects/*/*.jsonl`), not speculation.

**Issue A — Fresh-session-start baseline (~20-35% of window before any real work):**
Measured the very first assistant turn's `usage` field across 10 independent fresh sessions (5 in one project, 5 in another): totals ranged 40,613-43,344 tokens in one project and 44,252-69,534 in another. Notably, `cache_read_input_tokens` was **identically 7,298** across 5 separate sessions in the same project — a fixed, cached harness-level cost. No `type:"file"` attachments were present at session start in any sample, ruling out "recently read files" as the cause. Tried 4 fixes across earlier sessions (disabling zero-usage plugins, narrowing global MCP server scope, isolating a specific plugin's SessionStart hook, trimming personal `CLAUDE.md`) — none moved this baseline more than a few percent. Best available explanation (by elimination, not direct measurement): cumulative cost of per-session skill-catalog/agent-type-listing/deferred-tool-name-list/MCP-instruction injection that Claude Code sends on every session start regardless of plugin count.

**Issue B — Full-file re-injection on every auto-compact:**
Scanned all 92 session files in one project for `isCompactSummary` markers: found 42 real auto-compact events. In each, following the compact summary, Claude Code re-attaches: (1) the compact summary itself, (2) a subset of files recently touched via `Read` as **full content** (`attachment.type:"file"`, averaging ~7.8KB per event) rather than the lightweight `compact_file_reference` given to other touched files, (3) `deferred_tools_delta`, `agent_listing_delta`, `mcp_instructions_delta` (present in 41/42 events), (4) SessionStart:compact hook output including a plugin's full skill text. This appears to be **not proportional to actual new content** — it's a re-send of largely-unchanged state on every single compact, which compounds over long sessions. Cross-project check: auto-compact never fired in shorter-lived project sessions (longest observed: 3,692 lines) — only in the one project with marathon 30K+ line sessions — consistent with the ~83.5% documented auto-compact threshold, but the *reinjection cost itself* at each compact appears avoidable/reducible.

**Ask:**
1. Is there a way to make compact's per-file full/reference decision configurable, or default more conservatively to references?
2. Can agent-listing/deferred-tools/MCP-instruction blocks be sent as true deltas (only what changed) instead of full resends at every compact?
3. Is the ~40-70K fixed session-start cost expected to scale with plugin/skill count, and is there a lighter-weight catalog-reference mode instead of full injection?

Start with the session transcript files at ~/.claude/projects/*/*.jsonl and reproduce fresh-session and auto-compact usage, checking usage, isCompactSummary, attachment.type, compact_file_reference, and the delta blocks. Done means determining whether full-file reinjection and session-start catalog, tool, and MCP costs can be made configurable or reduced to true deltas, with evidence from comparable transcripts.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.