The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
fix(skills): correct the headroom guidance in context-budget
robinryukb620c0d605df
diff --git a/skills/context-budget/SKILL.md b/skills/context-budget/SKILL.mdindex d54140263..4900391d1 100644--- a/skills/context-budget/SKILL.md+++ b/skills/context-budget/SKILL.md@@ -43,39 +43,48 @@ Scan all component directories and estimate token consumption:- Estimate schema overhead at ~500 tokens per tool- Flag: servers with >20 tools, servers that wrap simple CLI commands (`gh`, `git`, `npm`, `supabase`, `vercel`)-**Session scaffolding** (the transcript itself)+**Session scaffolding** (measured from the transcript)-Everything above is loaded once at startup. A *reconnect* re-injects a second copy of much of it into-the session transcript as `attachment` lines, and that cost is charged to the same context window.-Measure it directly rather than estimating:+Everything above is loaded once at startup. A *reconnect* re-injects much of it again as `attachment`+lines in the session transcript — and that re-injection is charged to the same context window.```sh-# on the newest transcript for this project-f=$(ls -t ~/.claude/projects/"${PWD//\//-}"/*.jsonl | head -1)+f=$(ls -t ~/.claude/projects/"${PWD//\//-}"/*.jsonl 2>/dev/null | head -1)+[ -n "$f" ] || { echo "no transcript for this project yet"; exit 0; }python3 - "$f" <<'EOF'import json,systot=att=0-for line in open(sys.argv[1]):- tot+=len(line)+for line in open(sys.argv[1], encoding="utf-8", errors="replace"):+ if not line.strip(): continue+ tot+=len(line.encode())try:- if json.loads(line).get('type')=='attachment': att+=len(line)- except: pass-print(f"transcript {tot}B | scaffolding {att}B ({att*100//max(tot,1)}%) | conversation {tot-att}B")+ if json.loads(line).get("type")=="attachment": att+=len(line.encode())+ except Exception: pass # a malformed line counts as conversation, never crashes+if tot==0: print("empty transcript"); raise SystemExit+pct = att*100.0/tot+print(f"transcript {tot}B | scaffolding {att}B ({pct:.1f}%) | rest {tot-att}B")+if att*10 > tot*4: print(f"WARNING: scaffolding is {pct:.1f}% of the transcript (>40%)")EOF```-Measured on a session that had reconnected once: **33,884 of 49,830 bytes — 68% scaffolding**, broken-down as `skill_listing` 14.6KB, `mcp_instructions_delta` 8.1KB, `deferred_tools_delta` 7.1KB,-`agent_listing_delta` 2.6KB. The conversation itself was 15.9KB.+Measured on a session that had reconnected once: **33,884 of 49,830 bytes — 68% scaffolding**+(`skill_listing` 14.6KB, `mcp_instructions_delta` 8.1KB, `deferred_tools_delta` 7.1KB,+`agent_listing_delta` 2.6KB) against 15.9KB of conversation.-- Flag: scaffolding >40% of the transcript — the session is paying more for its own inventory than for- the work-- This is the strongest argument for the reductions elsewhere in this skill: trimming a skill or an- MCP server saves tokens *once per reconnect*, not once per session-- It also means transcript file size is a poor proxy for how full a context window is. Count only- non-`attachment` lines when reporting "how much room is left"+**What this number is, and what it is not.**-**CLAUDE.md** (project + user-level)+- It *is* a cost signal. Scaffolding is real context spend, and it is rebuilt on every reconnect —+ so trimming a skill or an MCP server saves tokens **once per reconnect**, not once per session.+ That strengthens every recommendation elsewhere in this skill.+- It is **not** a headroom estimate, in either direction. Attachments *are* charged, so excluding+ them overstates free space. And after a compaction the pre-compaction turns remain in the JSONL+ while the model's active context holds only the much smaller summary — so counting persisted bytes+ overstates what is actually loaded. **A transcript is a durable log, not a view of the context+ window; do not report remaining room from its size.**+- Report it as two separate figures — scaffolding share of the transcript, and conversation bytes —+ and treat both as diagnostics rather than as a fullness gauge.++**CLAUDE.md** (project + user-level)**CLAUDE.md** (project + user-level)- Count tokens per file in the CLAUDE.md chain- Flag: combined total >300 lines