Timing distribution · SyFI TraceLab

Session

one continuous trace of work, often spanning multiple requests or problems.

Request

one user input through the agent's final response.

Agent step

one model call inside a request.

User-initiated step

an agent step started by user input.

Tool-triggered step

an agent step started by a tool result.

Problem

Of the wall-clock time a coding agent consumes, how much is the human thinking, the LLM generating, and the tools executing — per session, per request, per step, and per individual latency?

The time-domain sibling of session_cost_distribution. Computes the data behind tab:timing_distribution (src/04_SessionContext.tex): for each granularity and category, avg / p50 / p90 / p99 per unit plus the category’s share of total time where a block has a meaningful total (same Avg/P50/P90/P99 + % layout as the cost table). The category set differs by granularity because human thinking is a between-request quantity:

Per session — Total elapsed (wall-clock first→last timing event), with session-total Human thinking, LLM generation, and Tool execution shares.
Per session, human capped (1h) — the same session units, but each human idle gap is clamped to one hour before summing. This is the prompt-cache-TTL view used for cache-relevant time.
Per request — Total (response time) (turn e2e) = LLM generation + Tool execution + possible overlap. No human term: human wait sits between requests, never inside one.
Per step — LLM generation vs Tool execution only (one round has no human term, no e2e).
Per individual latency — the strictly-positive human-input waits, positive observable per-round generation spans, and positive per-tool effective latencies. These rows line up with the human-wait, generation-time, and tool-latency CDF/summary views.

Tables

Per session

Metric	Avg	P50	P90	P99	% time
Total elapsed	8.2h	5.1m	5.9h	206.5h
Human thinking	7.6h	0.0s	4.1h	190.0h	92.3%
LLM generation	16.1m	2.0m	26.7m	3.9h	3.3%
Tool execution	23.5m	14.9s	26.1m	6.9h	4.8%

Per session, human capped (1h)

Metric	Avg	P50	P90	P99	% time
Total	1.8h	5.5m	3.4h	27.2h
Human thinking	1.2h	0.0s	2.3h	18.3h	64.3%
LLM generation	16.1m	2.0m	26.7m	3.9h	14.5%
Tool execution	23.5m	14.9s	26.1m	6.9h	21.2%

Per request

Metric	Avg	P50	P90	P99	% time
Total (response time)	4.3m	38.3s	6.4m	43.9m
LLM generation	1.7m	28.8s	3.7m	13.6m	41.0%
Tool execution	2.5m	0.3s	2.3m	34.4m	59.8%

Per step

Metric	Avg	P50	P90	P99	% time
LLM generation	11.5s	4.9s	20.2s	1.3m	40.7%
Tool execution	16.8s	0.1s	10.0s	3.2m	59.3%

Per individual latency

Metric	Avg	P50	P90	P99
Human thinking	46.7m	1.4m	20.6m	13.9h
LLM generation	13.2s	5.7s	22.2s	1.4m
Tool execution	18.4s	0.3s	13.6s	3.6m

Table 1Per-session, per-request, per-step, and individual latency time by category; % time is each category’s share.

A coding session is mostly idle, waiting on the human (the paper’s tab:timing_distribution). Human thinking is 92.3% of session wall-clock, dwarfing LLM generation (3.3%) and tool execution (4.8%); most sessions are short — the median is a single request with no inter-request gap — but a heavy tail of sessions left open for hours or days (session p99 elapsed ≈ 206h) accumulates most of that idle. Capping each gap at one hour (the cache-relevant budget) drops the human share to 64.3%, with generation and tool rising to 14.5% and 21.2%. The individual latency block uses the same positive values as the CDF/summary views: human waits have p50 1.4 min and p90 20.6 min, LLM generation spans have p50 5.7s and p90 22.2s, and positive tool latencies have p50 0.3s and p90 13.6s. Inside an individual request the human term vanishes and tool execution leads generation, 59.8% vs 41.0% of the 2,783h of total response time; an average request runs 4.3 min end to end (median 38s, p90 6.4 min), and per active step the model spends ~11.5s generating and ~16.8s in tools.

Reference

Definitions

LLM generation (per step) — observable generation span, latest qualifying input event → last model-output event; identical to llm_generation/generation_time_cdf and the per-round generation in human_in_the_loop/user_turn_decomposition.
Tool execution (per step) — sum of strictly-positive effective tool latency (tool_internal_latency_ms else tool_wall_latency_ms); identical to tool_calls/tool_latency_distribution.
Human thinking — previous event of any type → next user_message, strictly positive. The per-session row sums these gaps by session, so sessions with no second user message contribute 0s; the individual latency block’s human row reports the positive gap distribution directly and matches human_in_the_loop/human_input_wait.
Request e2e matches user_turn_decomposition turn-for-turn. Generation and tool execution can overlap (concurrent tools, generation streaming during a tool call), so their shares may slightly exceed the measured e2e total.
Request — one user turn (same turn state machine as user_turn_decomposition, user_turn_response_time, session_internal_counts, session_cost_distribution). Step — one LLM round. Session — one session_id (4,258 sessions have a positive wall-clock span; the rest are single-timestamp and dropped).

Running it

uv run python artifacts/session/session_timing_distribution/analyze.py -i trace/syfi_coding_trace.jsonl
uv run python artifacts/session/session_timing_distribution/analyze.py            # default merged trace

Outputs

session_timing_distribution.tex — the merged single-column timing table (Avg / P50 / P90 / P99
- % time) for the paper.
session_timing_distribution.md — GFM Markdown mirror of the table, rendered on the web detail page.
headline.json — the few headline numbers for the Overview gallery card.
stdout — merged + per-provider (Claude / Codex) per-category percentiles and time shares.

Headline numbers (public trace)

Sessions are mostly idle: human thinking is 92.3% of session wall-clock (avg 7.6h of an 8.2h session; medians are tiny — a single-request session has no inter-request gap). The long idle tail (session p99 ≈ 206h) is what pushes prompt prefixes past the cache TTL.
Positive human input waits match the CDF: p50 1.4 min, p90 20.6 min, p99 13.9h. This is the event-level view; it differs from the per-session median because 58.5% of sessions have no positive inter-request human wait.
Individual LLM/tool distributions match their summaries: observed generation spans have p50 5.7s and p90 22.2s; positive tool effective latencies have p50 0.3s and p90 13.6s.
Within a request, tool execution dominates, not generation: tool 59.8% vs generation 41.0% of the 2,782.7h of total response time (the two slightly overlap, hence shares can exceed 100%).
Avg response time: 4.3 min / request (p50 38s, p90 6.4 min); avg active work 11.5s generation + 16.8s tool per step.

The session human share is consistent across providers under the provider-agnostic definition (Claude 89.9%, Codex 94.3%) — earlier the trigger-based definition undercounted Codex (81.6%) and spilled ~13% into an “Other” residual, which is why that residual row was removed.

No figures.