Prompt Thermometer — Token Injection Report

01 · The Gap

59× undercount

13×

Scanner undercounts the cold floor by 13×

Static file scan produces ~1,747 tokens. The real cold injection floor — measured from cache_read_input_tokens immediately after a compaction reset — is 22,890 tokens. The full session warm ceiling reaches 103K–107K, giving the 59× ratio against accumulated conversation cache.

Scanner estimate

1,747

tokens · static disk scan

Cold injection floor

22,890

tokens · post-compaction reset

Session peak observed

107,237

tokens · accumulated warm ceiling

The scanner reads files. The harness injects systems.

CLAUDE.md + MEMORY.md are readable from disk. MCP tool schemas (Premiere has 200+ tools), loaded skill definitions, the agent type registry, and the harness envelope are injected programmatically at session start — never written to disk. The 21,143-token cold gap is entirely those invisible sources.

02 · Compaction Windows

Six resets.
One pattern.

Window 1 · 23:37

20,795

Peak 107,237 before reset

Window 2 · 00:17

22,890

First stable floor read

Window 3 · 00:58

22,890

Peak 81,471 before reset

Window 4 · 01:06

22,890

Peak 104,437 before reset

Window 5 · 02:17

22,890

Peak 104,437 before reset

Window 6 · 02:30

22,890

Current — still climbing

Timestamp (UTC)	cache_read	cache_write	input	output	Event
23:37:11	20,795	24,149	20,983	233	Session open
23:38:29	69,118	5,036	2	592	Cache warming
23:55:37	93,086	354	1	528	—
00:06:22	99,055	710	1	558	Pre-compaction
00:17:46	22,890	77,290	3	262	COMPACTION
00:19:38	100,180	1,346	1	499	Warming again
00:56:07	107,237	141	3	100	Session peak
00:58:06	22,890	38,577	1	89	COMPACTION
01:53:03	104,095	120	1	200	—
02:17:53	22,890	57,676	1	110	COMPACTION
02:29:51	103,642	144	3	216	Latest turn

03 · Scanner Breakdown

What the tool
actually found

Source	Est. tokens	Method	Status
user-global CLAUDE.md	~1,200	tiktoken cl100k disk read	FOUND
project CLAUDE.md	~150	tiktoken cl100k disk read	FOUND
MEMORY.md	—	slug lookup failed	BUG · NOT FOUND
Skills (disk)	~397	agents/ fallback post-fix	PARTIAL
MCP tool schemas	—	runtime injection only	UNREADABLE
Harness envelope	—	internal, not exposed	UNOBSERVABLE
Scanner total	~1,747	readable sources only	real floor: 22,890

04 · What's Missing

21,143 tokens
unaccounted

01

Skills & Agents

Skill text is injected at runtime by the harness — not loaded from .claude/skills/ as static text. The disk scanner reads the file but the actual injection content may differ (pre-processed, filtered, versioned). Estimate: several thousand tokens per active skill.

02

MCP Tool Schemas

Each connected MCP server sends its full tool schema list at session start. Premiere Pro alone has 200+ tools. Those JSON schemas are injected into the system prompt but never written to disk. With all profiles loaded: likely 10,000–18,000 tokens.

03

Harness Envelope

The Claude Code harness wraps every session with internal instructions, agent type listings, hook configs, tool descriptions, and system-reminder content. Injected programmatically. Structurally unobservable from outside the process.

05 · Fixes Applied

Three bugs.
One session.

01

MEMORY.md slug resolution · Resolved

The scanner was building the memory path from the project slug using a path-normalization rule that produced the wrong key. Fixed in sources.py — now resolves the correct UUID-based project directory and finds MEMORY.md reliably.
02

Skills fallback to agents dir · Applied

When .claude/skills/ is empty or missing, the scanner now falls back to .claude/agents/ and reads agent definition files as the best available proxy for injected skill content. Recovered ~397 tokens of skill coverage previously reported as zero.
03

--mcp-observed-tokens flag · Added

The scanner now accepts --mcp-observed-tokens N to manually pass in the cache_read_input_tokens value from the JSONL. When provided, the report shows the observed floor alongside the estimate and computes the residual gap (observed − estimated = unaccounted injection).

06 · Next Steps

Where to go
from here

Measurement strategy

The most reliable thermometer reading is cache_read_input_tokens at the first turn after compaction. That 22,890 number is the true injection floor. Measure against that — not the warm ceiling — and compute the residual to isolate what the scanner still can't see.

MCP delta experiment

Run one session with mcp.ps1 profile dev (no Adobe/Premiere/Resolve). Check cold floor post-compaction. If it drops from 22,890 to ~10K–14K, MCP schemas are confirmed as the dominant cost. That delta = schema injection per profile.

Profile comparisons

Current reading (22,890) includes all active MCPs for the AI-Workflow profile. Switch to dev or post-production and measure again. Difference isolates schema cost per profile.

Cost per session

At Sonnet 4.6 pricing, 22,890 cache-read tokens/turn ≈ $0.0023/turn at $0.10/MTok. A 100-turn session costs ~$0.23 in injection alone. A warm 103K-token floor costs ~$0.010/turn → ~$1.03/100 turns.

TOKENINJECTIONFLOOR

59× undercount

Six resets.One pattern.

What the toolactually found

21,143 tokensunaccounted

Three bugs.One session.

Where to gofrom here

TOKEN
INJECTION
FLOOR

Six resets.
One pattern.

What the tool
actually found

21,143 tokens
unaccounted

Three bugs.
One session.

Where to go
from here