Prune-guard validation — 4/5, guard WORKS (full trajectories, no deaths); the fail is a capture-group near-miss; found + fixed a 2× token-estimate skew

What was run

regex-log × 5 with the rebuilt context-hook pruning guard at COMPACT_AT_TOKENS=15000 — the same experiment that went 0/5 the previous run (…132733) when the guard still used the fatal ctx.compact().

Results

trial	result	agent	turns	peak ctx	prunes
P6fbmcP	PASS	14m55s	59	25.7k	48
dsnFToC	FAIL	4m21s	15	24.6k	7
i7F2eTe	PASS	2m09s	12	28.5k	0
mPwCvjE	PASS	8m09s	38	23.3k	26
z7DGFc6	PASS	5m24s	8	31.1k	6

The guard is validated. Pruning fired in 4/5 trials (87 prune events total), every trial ran a complete trajectory to a clean final stop (vs dying at turn 6–7 under the old guard), and the 59-turn/48-prune trial shows an agent surviving a long grind because old context kept being trimmed. 4/5 matches regex-log's historical ~90% (27/30 across K≥5 runs).

The one failure is NOT pruning damage

dsnFToC's regex matched exactly the 9 expected dates — but with two capturing groups, so re.findall returned tuples (('2025-01-09','2025-01-09')) instead of strings. Classic §4 near-miss (one assertion, formatting of the match output); the agent's final summary shows full knowledge of the task, nothing it needed had been pruned.

NEW: chars/4 underestimates real tokens ~2× on this content — calibration added

Trial i7F2eTe peaked at 28.5k real tokens with zero prunes: its chars/4 estimate never crossed 15000. Measuring per-request usage vs estimate across all 5 trials: real/est = 1.7–2.1 (log files — dates, IPs, punctuation — tokenize at ~2 chars/token). Raw, that skew would have made the 180k-est default fire at ~340–380k real, far past the 262144 cap → the guard would have missed the very 400s it exists to prevent on dense-content tasks.

Fix (landed right after this run): the guard now calibrates a real-per-estimated ratio from the requests' actual usage while the history is still unpruned (frozen after the first prune, clamped to [1,3], max over the run), and the threshold is therefore ~real tokens. Default restored to COMPACT_AT_TOKENS=200000 (>60k real headroom). Node tests cover ratio-2 triggering, ratio-1 non-triggering, and the clamp. Prune log lines now include the ratio; make failures shows prune counts as pr:N.

Next

Re-run this same K=5 at 15000 to confirm the calibrated trigger (expect prunes in ~5/5 trials incl. short ones, pass ≥4/5), then the two untested 400-victims at defaults: circuit-fibsqrt, make-mips-interpreter.