-
v1.101.0 — Pinky RRRR: honest steer attribution
StableAll checks were successfulChecks / Functional checks (push) Successful in 46sreleased this
2026-10-17 03:42:46 +00:00 | 10 commits to main since this releaseOwner finding #60: a steer-delivered transcript message rode the real_user user_prompt shape, so the sender rendered as generic user — B³’s words landed in gen-4 as “user ⧉”.
The cure is an evidence chain, never a guess
- The message’s steer anchor pins the delivery — its body is the verbatim posted bytes.
- The anchor’s send stamp + normalized byte equality prove which steer prompt file carried the words (the
obs-child seam the lane rows know). - The source is the session whose own tool part ran the posting call — it names the target sid AND the verbatim text, inside the send-time window. Latest hit is the poster; a photo-finish between two sessions is unproven → stays
user.
- Machine steers (context guardian, quota throttle, starvation watch, turn-retry) carry their identity by construction — recorded at spawn.
userstays exclusively genuine owner input; legacy rows, cleaned prompt files and dead source sids settle honestly.- Symmetry: agent replies keep their
agentclass — who asked and who answered both render honestly. - The drawer renders the source title in the sender slot (mint kept); the tooltip carries the provenance + sid. Findings persist in a capped album under the steer dir.
QQQQ-F3 fold
The clipid specimen recorded 13 legs while its docs said 11 — the double-recorded Q2/Q7 proofs got their own ids; docstring, labels and Makefile now say 13.
Evidence
- Suite: 1733 ran (+20 new), 5 skipped, all green (version-guard tie resolved by this tag).
- Vault:
rrrr_attrib_cdp.py— 12/12 legs (API + rendered drawer). - Super-Linter: six lanes green.
Downloads
-
Source code (ZIP)
0 downloads
-
Source code (TAR.GZ)
0 downloads