Frame loss between master agents and subagents
Measured 2026-08-14 from local Claude Code transcripts. This finding is recomputable with:
go run ./cmd/framevisibility --homes C:\Users\USER --project C--work-fak --out <scratch>\fold.json
The committed verb emits aggregate numbers plus per-session counts; raw transcript text, prompts, paths inside tool arguments, and account credentials are never copied into the output. The raw fold remains local and untracked.
Result
| Question | Measured result |
|---|---|
| Subagent events reconstructable in the master frame | 13,485 of 636,730 (2.12%) |
| Generous upper bound, crediting events whose existence is only countable | 233,970 of 636,730 (36.75%) |
| Decision-relevant events in the master’s verbose stream | 117,902 of 319,028 (36.96%) |
Operating envelope: 25 Claude homes, 777 root sessions, 9,572 subagent transcripts, 3,356 explicit spawn events, and 317,191 classified tool calls (96,706 master; 220,485 subagent). This exceeds the issue’s requested envelope of 20 transcripts, 50 spawns, and 500 classified calls.
Reproducible rubric
The exact executable rubric is internal/framevisibility/fold.go:
- A subagent event is visible only when its normalized text is represented in the master’s transcript. A tool call or tool result that exists only in a descendant transcript is not visible.
- The generous bound additionally credits events whose existence can be inferred from a returned count without reconstructing their content.
- A master event is decision-relevant when it is a tool error, repository mutation, mutating shell command, delegation control event, or terminal answer. The broad shell matcher deliberately makes 36.96% an upper bound.
- Thinking, progress narration, read-only calls, and successful read results are not decision-relevant under this operator-action rubric.
A synthetic transcript test in internal/framevisibility/fold_test.go proves that
hidden descendant calls stay outside master visibility while the returned result
is credited.
Interpretation
This CONFIRMS #6572’s claim that watching only the master’s verbose tool stream is a strongly biased sample: the strict view reconstructs only 2.12% of descendant events. It also confirms that verbose master output is not a proxy for the whole fleet: even with an intentionally generous rubric, 63.04% of master events are not decision-relevant.
The result weakens any stronger claim that the master knows nothing about its subagents: under the generous existence-only interpretation, up to 36.75% of subagent events leave some trace. That is still not enough to reconstruct the actual calls or dead ends.
Limits and invalidating assumptions
- The corpus is local fak work, not a random sample of every agent workload.
- Text normalization may credit coincidental repeated text; therefore 2.12% is an upper estimate of semantic reconstructability, not proof of causal forwarding.
- Subagent files are deduplicated by basename to avoid nested copies. If basenames are not unique within a session, the denominator is understated.
- A provider that starts forwarding structured descendant events into the master transcript would invalidate the central result; rerun the verb after such a format change.
- If the operator’s true decision rubric treats ordinary read calls as relevant, master relevance rises and this classification should be revised before reuse.