26bd8dachangelog v2.16.1 is 6 commits behind this build — last written 8d ago, so 6 changes to what the site shows are unrecorded
Posted unattended (--auto); not yet reviewed by @jwildfire.
Selection under the goal contract. The init ran lean (fresh hand-off from the 20:59 wrapup, one delta agent, ideas-inbox pass), then the --auto selection fetched all four goal issues live. The hand-off's top item — open.gismo#6, verified and scoped — turned out to be ineligible: goal #79's body bars every app-goal implementation increment until the #34 plan rewrite lands, and flags the missing obotclaw App install on open.gismo besides. Selection fell through to charts #78, whose boundary — "small, well-scoped backlog items beat big anchors at night" — pointed at the sv#54 tail; obot-policy explain confirmed safety.viz is an auto-profile repo with branch + draft-PR + standard-lane-merge granted. That a goal's prose overrode the carried todo list, correctly and without a human in the loop, is exactly what the goals-as-issues design was for.
sv#54 closed out (sv#122). The four items, each traced to its upstream hep-explorer issue: full measure names with on-line abbreviations in the participant-profile spaghetti (PPRF-SPAG-004, upstream #290) · R/nR primary-source references linked from the Clinical guide (HEP-DOC-001, #335) · a Log Base control offering log10 decades or log2 doublings (HEP-CTRL-017, #112 — implemented as a tick generator, not a transform: log positions are base-independent, the base only picks which multiples the gridlines land on) · per-axis manual min/max inputs reusing the sv#85 axis-limits contract (HEP-AXIS-001..004, #238). A side-fix landed with it: the site's mdInline was truncating any URL containing a parenthesised journal PII. 1110 unit + 225 e2e tests green; merged on the standard lane as obotclaw after CI passed, and CI + Pages ran green again on the merge commit. Requirement #88 now has six of its eight sub-issues closed — sv#49 (exposure track) and sv#55 remain.
An evidence gotcha caught before it shipped. The canonical baseline refresh initially kept four stale participant-profile captures: Playwright's --update-snapshots only rewrites a baseline when the comparison fails, and safety.viz's 0.02 maxDiffPixelRatio let the changed legend pass inside tolerance — CI green, evidence page quietly showing the old chart. The sibling caught it visually, deleted the captures so the canonical run regenerated them fresh, and the trap is now recorded in the renderer-done-gate memory: a green evidence run is not proof the pictures moved.
Guardrail precision, measured on a live run. Three friction events, none of them the work itself failing, all now tracked: the auto-mode classifier denied the obotclaw discussion-reply lane, so the ideas-inbox pass could not post its triage reply on discussion #141 — skipped per the no-workaround rule, watermark held, filed as obot.agent#74 · the merge-gate-guard hook false-positived on a spawn briefing that merely mentioned the raw merge command in prose while forbidding it — obot.agent#58 confirmed live, one lost round-trip · the obotclaw App token's missing actions: write recurred (the evidence workflow dispatch fell back to personal auth mid-unattended-run, second session in a row) — promoted from memory-only to obot.agent#73. Both new trackers carry proposed mitigations per the risk convention.
bb8cad3, milestone v1.2.0).actions: write — unattended evidence refresh falls back to human auth → obot.agent#73.--check mergeability).ideas-file --reply + allowlisted lane), then post the owed #141 reply.actions: write on the obotclaw App — closes the recurring personal-auth fallback.This entry was drafted by Claude Code using Fable 5 in an unattended session — unreviewed.