rdmsm4x-changelog-20260904-1847-sidebar-cleanup-pr-sweep-rtty-regression

Swept every open PR in the fleet (there was one, and it merged under its owner while under review), archived 115 stale Claude Code sidebar threads reversibly, retitled the untitled live threads, and filed ISSUE-20260904-13 — a confirmed regression that survives the merged RTTy #10.

Window: 2026-09-04 18:34:21 EDT → 18:47:36 EDT · host rdmsm4x · session c8b04c2f (dev-57) · claude@rdmsm4x Request: "auto-fix/auto-merge all open/pending PRs, delegate sub-agents to review all, rename all current threads, archive disconnected threads — I'd rather the sidebar be clean." Rich sent the same prompt to at least two sessions (dev-de got a copy with an extra framing sentence: RTTy release, ReplicantDB, Tyrell, ECS libraries / ecs0lib).

What changed

PR sweep — result

Exactly ONE open PR across all 91 remote-backed repos under ~/dev — richhdoty/RTTy#10 (Build 49 flows-banner split). Verified three ways: gh search prs --owner, --author, --involves richhdoty. It was a DRAFT whose body said "Not merging until both are green" with its archive-matrix gate still running (pid 87523, started 18:21:29). I did not merge under its owner. The owner's cloud session took it out of draft and merged it at 22:37:24Z (18:37 EDT) as 2629abc; main is now fbede7c.

The regression that survives the merge (ISSUE-20260904-13)

Independent read-only review (one Opus subagent, 155K tokens, 59 tool calls, 8 min) of the full 13-file diff against merge base e0011cf:

• CONFIRMED HIGH. ApplicationFlowSampleRunner.swift:68 still feeds sourceWasPartial from ecs0lib's rolled-up NetworkTrafficCoverage.isPartial, which ORs in ecs0lib's own eviction caps (evictedProcessTotals/evictedEndpointTotals, limits 8,192 / 32,768). Those counters are cumulative and reset only at teardown, so one eviction makes isPartial true for the life of the subscription → rateBaseline=nil and previous=nil every sample → all rates render "—" permanently, orange banner permanent, message blames the kernel. PR #10 removed the ~90 s trigger and left a sticky hours-to-days one. The branch's own docs/coverage-banner-measurement-20260904.md (lines 149-153, 334-341) identified this before the code was written. No test can see it: all tests inject Coverage directly, bypassing the runner mapping.
• Confirmed clean: every isPartial call site moved (lines 140 and 160 both — a half-move would have computed rates against a stale baseline); generated trust chain verifies 26/26 + 8/8 + 3/3 with zero digest mismatches (no archive-time signing risk); version literals fully propagated; 18 affected tests pass; 800 XCTest methods repo-wide.
• Two lesser: 125 pt band comment computes 8×13+7×3=125 with zero slack while the sibling commit adds a caption into the same VStack (scrolls, nothing lost, reason recorded is false); boundingCaption says "Showing top 256" above ≤8 rows.
• Owner notified (rtty-e9 / 0fbacd80, active) at 18:45 EDT with the fix shape: three coverage buckets, not two. Not fixed by me — RTTy is live-owned.

Sidebar cleanup — selection rule (INCLUSION)

Moved only if ALL held: last real activity (timestamp inside the transcript) < 2026-08-31; not in ~/.claude/sessions/; not in any resume pin; a real project thread (subagents/, wf_* untouched). Moved set spans 2026-08-07 → 2026-08-30, 0.24 GB. mtime was rejected as the signal: 95 files share one 2026-09-03 21:16 mtime from a bulk copy; only 9 had genuine Sep-3 activity. Result: real project threads in the sidebar 154 → 39. Undo: ~/.claude/projects-archive/20260904-sidebar-cleanup/RESTORE.sh (exact manifest, nothing else).

Thread titles

Mechanism: append {"type":"custom-title","customTitle":…,"sessionId":…} to the transcript; verified on my own session first. Trap: grep -c '"type":"custom-title"' over-counts 25:3 because the string recurs inside recorded tool output — use jq 'select(.type=="custom-title")'. dev-de retitled six other live threads at 18:45 EDT in a consistent "rdmsm4x: next … (timestamp)" format in parallel with me — redundant, not conflicting; appends are additive and the last record wins. All 9 transcript-backed live threads now carry accurate titles. dev-df has no transcript to title.

Also measured, not acted on

• net/grafana has 2 unpushed commits owned by live session dev-9e, one of which (d327491, SEC-20260904-01) removes a fleet SSH password from push_to_host.exp. The commit itself records that the password remains in history on origin/backup/fleet; repo is PRIVATE. Rotation is Rich's action. Not pushed — live-owned, and pushing changes nothing about exposure.
• 39 repos show ?? docs/ (untracked generated docs) — noise, not work. sites/dev.ecs0.net has 96 dirty entries (in-flight wiki regeneration, includes this session's ticket HTML export).
• Budget: usage_status advice=hold driven solely by OpenAI at 100%; Anthropic 57% 5h / 12% 7d, proceed. One Anthropic subagent spawned; announced here.

Own measurement errors this session (caught before narrating)

find -newermt '-30 minutes' silently matched nothing on BSD find; ps -p | tail -1 returned tail's exit code; substring grep inflated the title count 8×; sorting stat %H:%M without the date. Same class every time — a check returning a plausible value for a question it did not ask.

How to undo

• Sidebar: RESTORE.sh above.
• Titles: truncate each transcript to the byte offset in undo-offsets.txt (scratchpad), or append a new custom-title.
• Ticket: ticket resolve ISSUE-20260904-13 if the owner disputes it.

Outstanding — Rich

1. Rotate the fleet SSH password (SEC-20260904-01). It is in git history on three remotes of a private repo.
2. ISSUE-20260904-13 needs a decision on where the split lives (ecs0lib exposing the terms vs RTTy reading the counters).
3. The offline Remote Control entries in ListAgents (≈23) are server-side registrations; nothing local removes them.