Handoff Cost

Handoff Cost

This page gives the cost of one /dotclaude:handoff and the cost of one /compact. Both runs start from the same session. Read it to compare the two costs, and the sources of each number. The source of the numbers is the measurement of 2026-10-12, with Claude Code 2.1.296. The record is .dotclaude/spec/changes/1.0.0/evidence/handoff-cost.md. The index of the evidence is Evidence.

What the Numbers Compare

• The handoff figures are for the whole handoff, which has two model calls.
• The compaction figures are for the /compact step and its follow-up turn.
• The context figures are not the same kind of value. For the handoff, the context is at the last call of each step. For the compaction, the context after is postTokens of the compact boundary. The follow-up value is at the first call of the follow-up. Read the context rows as two different points. Do not read them as one measure.

Setup

• Date: 2026-10-12, from 17:32 UTC to 17:34 UTC.
• Claude Code: 2.1.296, from claude --version.
• Model: claude-sonnet-5-5, with a context window of 1,000,000 tokens, as the record names it.
• Checkout: branch dotclaude-1-0-0, tip 627d1d22.
• Each run used just sandbox, with a new sandbox from just sandbox-clean.
• The project of the sandbox held a copy of the asks, decide, guard, and handoff slices, with the tests removed, in one commit.
• The sandbox had no config/settings.json. So mainHandover had its default, manual. The default is DEFAULT_MAIN_HANDOVER in src/plugins/dotclaude/slices/handoff/limits.ts.
• Four claude runs ran in total:
  • One base run.
  • One handoff fork.
  • One compaction fork.
  • One follow-up.

Commands

All runs used just sandbox --model claude-sonnet-5-5 --output-format stream-json --verbose. The base run read every file under src/asks and src/decide with the Read tool. The handoff fork ran /dotclaude:handoff. The compaction fork ran /compact. The compaction fork received the follow-up Continue the work.

• The base run read 31 files. Its last call had a context of 81,316 tokens (20,245 cache read plus 61,071 cache write).
• --fork-session worked. The two forks started from the same context.
• The handoff wrote the handoff file (1,585 bytes). It ran /clear. The resume turn in the new session read the file and answered. The handoff needed no follow-up turn of its own.
• The compaction fork needed the follow-up turn.
• The report command was CLAUDE_CONFIG_DIR=<sandbox>/config plugins/dotclaude-sessions/bin/dotclaude-usage --days 1 --window 1000000.

Sources

• Source S: the usage and total_cost_usd fields of the result events in the stream-json output.
• Source T: the cost-state and compact_boundary entries, and the usage of the assistant entries, in the transcripts of the sandbox. A forked session starts with the cost-state of its base. So a step is the cost-state after it minus the cost-state before it.
• Source U: the output of dotclaude-usage, used as a check only.

Handoff

Measure	Handoff step	Resume step	Total	Source
Input tokens (not cached)	6	4	10	S: usage.input_tokens
Cache read tokens	164,242	25,946	190,188	S: usage.cache_read_input_tokens
Cache write tokens	2,576	5,874	8,450	S: usage.cache_creation_input_tokens
Output tokens	869	147	1,016	S: usage.output_tokens
Model calls	2	2	4	T: four assistant ids
Cost (API list price)	$0.0354	$0.0276	$0.0630	S: total_cost_usd
Context, last call of the step	82,927	18,487	18,487	T: input plus cache read plus cache write
Time of the step	6.7 s	2.4 s	9.1 s	S: duration_ms

Compaction

Measure	/compact step	Follow-up step	Total	Source
Input tokens (not cached)	59,526	2	59,528	T: cost-state deltas
Cache read tokens	20,245	12,613	32,858	T: cost-state deltas
Cache write tokens	3,542	15,803	19,345	T: cost-state deltas
Output tokens	3,049	391	3,440	T: cost-state deltas
Model calls	not shown	1	not shown	T: the transcript has no assistant entry for the summary call
Cost (API list price)	$0.1604	$0.0684	$0.2288	S: total_cost_usd deltas
Context, first call of the follow-up	7,835 (postTokens)	28,418	28,418	T: compact_boundary and the follow-up usage
Time of the step	18.8 s	4.4 s	23.2 s	T: durationMs, and S: duration_ms

Comparison

• On this session, the handoff cost $0.0630. The compaction with its follow-up cost $0.2288.
• The summary call of the compaction read 59,526 tokens that the cache did not serve. It wrote 3,049 output tokens.
• The handoff made no such call. Its turn read the cached context of 164,242 tokens over two calls.
• The context after the handoff was smaller than the context after the compaction. This holds for the values in the tables only, because they are at different calls.
• The handoff took 9.1 s. The compaction with its follow-up took 23.2 s.
• The sums of dotclaude-usage do not match these tables. A fork copies the history of its base session. The tool counts each model call once. It does not show the call of /compact. Source U was a check. It was not the source of the tables.

What Is Not Measured

• The page has one session of one size (about 81,000 tokens), one model, and one run of each step. The numbers have no spread.
• The call of /compact is not in the transcript as a message. Its tokens come from the difference of cost-state. Its call count is unknown.
• The cost is the API list price that Claude Code prints. It is not a subscription limit. Source S and source T come from the same client, so they are not two independent meters.
• The quality of the continued work is not measured. This page does not rate the answers of the resume turn or the follow-up.
• Cache expiry between steps is not measured. The steps ran within two minutes, with a cache of one hour. A handoff or a compaction after a long pause pays more cache writes.

Related Pages

• Usage-Evidence: the plans, prices, and usage of dotclaude 1.0.0 over one measured week.