Post-tool latency span (#10662)

The audited defect: in Codex rollout journals the interval AFTER a tool result is recorded and BEFORE the next model-emitted record grows with session age and context size (median ~11s at tool-result ordinal 1-20, ~21s at 101-200, with the hotspot at 50k-100k input tokens). internal/codexlifecycle/posttool.go makes that interval a first-class span instead of an unnamed hole in the timeline.

Definition

One PostToolSpan per function_call_output / custom_tool_call_output record:

call ──ToolMS──▶ output ──GapMS──▶ next model-emitted record

A model-emitted record is function_call, custom_tool_call, token_count, task_complete, or turn_aborted (the same set decomposeTimeline anchors on; ARecord.PayloadKind recovers custom calls, so no JSON is re-parsed). A trailing result with no next model-emitted record is skipped and counted as tail_skipped in the report: an unclosed interval measures nothing.

Disjointness (the no-double-counting witness)

ToolMS measures call → its own output; GapMS measures that output → the next model-emitted record. The intervals share exactly one endpoint and cannot overlap, so ToolMS + GapMS always equals the call → next-record interval. A slow tool can never be booked as post-tool model latency. Interior compacted records become subspans (pre_compaction, then one compaction segment per edge) that tile GapMS exactly.

Closed vocabularies

Context band, from the first token_count with input_tokens > 0 at or after the gap end (that token_count reports the usage of the request that consumed this tool result): unobserved, lt10k, 10k_25k, 25k_50k, 50k_100k, gte100k.
Ordinal bucket, per-session 1-based tool-result ordinal: 1_20, 21_50, 51_100, 101_200, gte201.
Attribution (correlation): tool_slow when ToolMS >= GapMS (the wait was the tool), stall_capped when GapMS > 300s (only the first 300s is model-active, the remainder idle, same rule as decomposeTimeline), compaction_in_gap when a compaction fired inside a non-stall gap, else model_reasoning.

Correlation, not causation

Journal timestamps cannot separate provider TTFT from gateway queueing or harness scheduling inside GapMS. Attribution tokens are CORRELATION aids over observable structure; the report never claims a causal latency split. Live-path emit belongs to #10636 and the timing inventory to #10621.

Running it

fak session-audit posttool --here --json

Flags mirror fak session-audit codex: --root DIR (default ~/.codex/sessions, honoring CODEX_HOME), --cwd DIR|--here, --max N. The text renderer prints overall gap/tool percentiles plus one line per non-empty band and ordinal bucket in canonical order, each with the tool_p50 control beside the gap percentiles. Regression corpus and expected values: internal/codexlifecycle/testdata/posttool/issue-10662/.

Post-tool latency span: tool_result_recorded to next_model_item (#10662)

The first-class post-tool span in internal/codexlifecycle: definition, disjointness rule, closed band and attribution vocabularies, and the fak session-audit posttool readout.