docs(spike): record the xhigh executor truncation spike #73

Merged

crenshawdev merged 1 commit from cadence/spike-xhigh-truncation into main

2026-07-28 18:38:28 +00:00

crenshawdev commented

2026-07-28 18:37:44 +00:00

(Migrated from github.com)

Tracks the spike write-up that settles the max_tokens question gating the cad-executor row in #63.

.planning/spikes/ was untracked, so #63 could only reference the evidence as a plain path. This puts it in the repo.

Verdict: invalidated. The constraint #63 states is real, the consequence it claims is not.

max_tokens is genuinely unsettable and unobservable for a subagent, and it binds thinking and response text together with no way to disable thinking at xhigh or max. That part of #63 was accurate.
But it bounds one API response, not a phase. Commits are Bash tool calls, so a truncated response is one whose tool call never issued: a stalled turn you re-dispatch, not the partial commit and corrupted phase #63 described.
Measured, not argued: a realistic executor turn cost 26,521 output tokens and issued its tool call intact; a deliberately maximal response reached 36,015 and stopped on end_turn; and stop_reason: "max_tokens" occurs zero times across 33,165 assistant responses in 668 transcripts.

Docs only. No code, no behaviour change.

Closes nothing. #63 stays open with its body and measurements updated, and #72 was filed for the maxTurns gap the spike turned up in passing.

Tracks the spike write-up that settles the `max_tokens` question gating the `cad-executor` row in #63. `.planning/spikes/` was untracked, so #63 could only reference the evidence as a plain path. This puts it in the repo. **Verdict: invalidated.** The constraint #63 states is real, the consequence it claims is not. - `max_tokens` is genuinely unsettable and unobservable for a subagent, and it binds thinking and response text together with no way to disable thinking at `xhigh` or `max`. That part of #63 was accurate. - But it bounds one API response, not a phase. Commits are Bash tool calls, so a truncated response is one whose tool call never issued: a stalled turn you re-dispatch, not the partial commit and corrupted phase #63 described. - Measured, not argued: a realistic executor turn cost 26,521 output tokens and issued its tool call intact; a deliberately maximal response reached 36,015 and stopped on `end_turn`; and `stop_reason: "max_tokens"` occurs zero times across 33,165 assistant responses in 668 transcripts. Docs only. No code, no behaviour change. Closes nothing. #63 stays open with its body and measurements updated, and #72 was filed for the `maxTurns` gap the spike turned up in passing.