enhance(fleet-runtime): upgrade fleet Claude Code v2.1.81 -> >=v2.1.89 — eliminates #1 error class (Read-before-Edit, 105+/24h) #580

Open
opened 2026-06-21 09:21:49 +00:00 by jason.woltje · 0 comments
Owner

Problem

The agent fleet runs Claude Code v2.1.81. The single dominant error class in the nightly harvest is "File has not been read yet. Read it first before writing to it." (Read-before-Edit) — agents Edit a file they viewed via cat/sed/Bash without a separate Read, the Edit is rejected, and a model recovery turn is burned re-reading. Pure wall-clock waste, fleet-wide.

Evidence (nightly enhance harvest, 24h ending 2026-06-21)

Raw per-session counts (NOT the digest's deduped cluster count, which undercounts — see note):

session Read-before-Edit errors
feacca06 39
ff505fb3 19
bad51cfc 18
9bb32ffd 16
0f1d4711 13
total (5 sessions) 105+

This is the #1 error signature across the fleet, every nightly window.

Root cause + fix (well-defined)

Claude Code v2.1.89 changelog (Apr 1): "Edit works on files viewed via sed/cat without separate Read." That single change eliminates this entire error class. v2.1.89 also rolls up cumulative fixes v2.1.82–v2.1.88: prompt-cache-miss on long sessions, memory leaks, autocompact thrash loop (stops after 3 refills), rate-limit crash loop. Latest at last release-check was v2.1.92.

The fleet has been on v2.1.81 and "strongly recommended to update" since early April (multiple OpenBrain release-checks) — but the upgrade was never scheduled. This issue exists to convert that standing recommendation into a tracked, evidence-backed action.

Proposed action

Coordinated fleet upgrade of Claude Code from v2.1.81 to >= v2.1.89 (prefer latest stable, v2.1.92). This is a rollout across agent hosts/launchers — Mos to own sequencing and blast radius (large change; one lever at a time).

Why it matters (north star: wall-clock)

105+ failed Edits/recovery turns per 24h across the fleet is a continuous, compounding wall-clock tax that a one-time version bump erases — plus the cache/memory/crash-loop fixes improve long-session stability fleet-wide.

Routing note

Filed in mosaicstack/stack as the framework/runtime owner. If fleet CC version is better owned by host provisioning in usc/infrastructure, please re-route. Treat as a SEPARATE structural lever from the wrapper batch (#550/#559) and the host-prereq item (#561).

Acceptance

  • Fleet agent hosts run Claude Code >= v2.1.89.
  • "File has not been read yet" drops toward ~0 in the next nightly harvest window.

Note on the evidence count

The nightly digest's "Error clusters" section showed only "8x" for this signature — it dedups by normalized signature and undercounts high-frequency repeats. True magnitude (105+) was confirmed by raw grep -oc across the top-error sessions. Captured as a harvester gotcha (OpenBrain dcc6238a); a future low-priority fix could make the digest report true frequency.


Filed by the nightly fleet-enhancement pass (enhance/ae4). Deduped against OpenBrain (version-currency already recorded; no prior issue tracked the upgrade). Merge/dispatch authority remains with Mos/coordinator (AGENTS.md rule 38).

## Problem The agent fleet runs **Claude Code v2.1.81**. The single dominant error class in the nightly harvest is **"File has not been read yet. Read it first before writing to it."** (Read-before-Edit) — agents Edit a file they viewed via `cat`/`sed`/Bash without a separate `Read`, the Edit is rejected, and a model recovery turn is burned re-reading. Pure wall-clock waste, fleet-wide. ## Evidence (nightly enhance harvest, 24h ending 2026-06-21) Raw per-session counts (NOT the digest's deduped cluster count, which undercounts — see note): | session | Read-before-Edit errors | |---|---| | feacca06 | 39 | | ff505fb3 | 19 | | bad51cfc | 18 | | 9bb32ffd | 16 | | 0f1d4711 | 13 | | **total (5 sessions)** | **105+** | This is the #1 error signature across the fleet, every nightly window. ## Root cause + fix (well-defined) Claude Code **v2.1.89** changelog (Apr 1): *"Edit works on files viewed via sed/cat without separate Read."* That single change eliminates this entire error class. v2.1.89 also rolls up cumulative fixes v2.1.82–v2.1.88: prompt-cache-miss on long sessions, memory leaks, autocompact thrash loop (stops after 3 refills), rate-limit crash loop. Latest at last release-check was v2.1.92. The fleet has been on v2.1.81 and "strongly recommended to update" since early April (multiple OpenBrain release-checks) — but the upgrade was never scheduled. This issue exists to convert that standing recommendation into a tracked, evidence-backed action. ## Proposed action Coordinated fleet upgrade of Claude Code from v2.1.81 to >= v2.1.89 (prefer latest stable, v2.1.92). This is a rollout across agent hosts/launchers — Mos to own sequencing and blast radius (large change; one lever at a time). ## Why it matters (north star: wall-clock) 105+ failed Edits/recovery turns per 24h across the fleet is a continuous, compounding wall-clock tax that a one-time version bump erases — plus the cache/memory/crash-loop fixes improve long-session stability fleet-wide. ## Routing note Filed in `mosaicstack/stack` as the framework/runtime owner. If fleet CC version is better owned by host provisioning in `usc/infrastructure`, please re-route. Treat as a SEPARATE structural lever from the wrapper batch (#550/#559) and the host-prereq item (#561). ## Acceptance - Fleet agent hosts run Claude Code >= v2.1.89. - "File has not been read yet" drops toward ~0 in the next nightly harvest window. ## Note on the evidence count The nightly digest's "Error clusters" section showed only "8x" for this signature — it dedups by normalized signature and undercounts high-frequency repeats. True magnitude (105+) was confirmed by raw `grep -oc` across the top-error sessions. Captured as a harvester gotcha (OpenBrain dcc6238a); a future low-priority fix could make the digest report true frequency. --- _Filed by the nightly fleet-enhancement pass (enhance/ae4). Deduped against OpenBrain (version-currency already recorded; no prior issue tracked the upgrade). Merge/dispatch authority remains with Mos/coordinator (AGENTS.md rule 38)._
jason.woltje added the fleet-enhancement label 2026-06-21 09:21:49 +00:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: mosaicstack/stack#580