You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
perf(hooks): desktop-lane measurements still open: typos-format per edit (B6), hook cancels after #6628 (B10), rate-limit readings across a rotation (B12) #6687
Three measurements from the audit need the desktop lanes (native Windows and the desktop WSL lane) and were left open. Each decides whether a plugin change is worth making; none approves a code change by itself.
Next step: recount cancels per hook script on both desktop lanes from transcripts and --debug logs once the lanes run a version containing perf(harness-ops): run the per-session event log hook in the background #6628. Settle from the Claude Code docs or a repro whether the harness deletes an orphaned version directory while a live session still references it (a plugin cannot change that behavior).
B12: rate-limit-guard and context-guard readings around an account rotation
Evidence: desktop WSL lane showed 24 holds for a usage reset and 46 handoff suggestions at 18-22% context (unverified).
Decision recorded. The owner approved applying independently verified recommendations on 2026-10-12.
B10: does the harness delete an orphaned version directory while a live session references it? → no; it removes it 14 days after marking it orphaned (added by verification)
Basis: https://code.claude.com/docs/en/plugins/loading 'Cleanup of previous versions' fetched 2026-10-12
Three measurements from the 2026-10-08/09 audit (typos-format per-edit cost on native Windows, hook_cancelled counts after #6628, rate-limit readings across an account rotation) still need the desktop lanes. #6628 merged 2026-10-08, so B10 is now measurable.
Verification notes:
route_reason 'only the owner can drive' is not true for B6 and B10: FLEET.md lists reach paths to both desktop lanes, and the user's instructions say to run what the agent can run.
owner_action omits B10's docs question, which the plugins loading doc already settles: an updated version's old directory is removed 14 days after the .orphaned_at marker, so a session that loaded it keeps running.
This was generated by AI during an unattended host run over /fleet:reach (melo-desk-001, both desktop lanes). Both lanes run harness-ops 4.4.0, which is later than #6628.
B6: typos-format per-edit cost on native Windows
Verdict: no change. The per-edit cost no longer exists. typos-format 0.11.4 registers only SessionStart and Stop hooks, so it runs once per turn, not once per edit.
The measurement was one headless claude.exe -p --debug-file session (Claude Code 2.1.309, no concurrent agents). It made 18 separate Edit calls, 6 each to .md, .py and .txt, in a scratch git repository. Timings come from the debug log's Hooks: ... finished with status N (Nms) lines:
Event
Plugin
n
p50 ms
p90 ms
max ms
PreToolUse:Edit
guardrails
18
239
261
318
PreToolUse:Edit
context-budget
18
53
60
68
PostToolBatch
eol-normalizer
17
340
367
500
PostToolBatch
markdown-format
17
78
381
388
PostToolBatch
ruff-format
17
87
345
350
PostToolBatch
biome-format
17
96
99
101
PostToolBatch
go-format
17
73
81
85
PostToolBatch
bash-format
17
61
65
69
Stop (once)
typos-format
1
143
The p90 spikes for markdown-format and ruff-format come from the edits to their own extensions. With concurrent agents was not measured: the old 917 ms figure belonged to a per-edit hook that no longer exists. guardrails' 239 ms PreToolUse:Edit is #4390's subject.
This was generated by AI during an unattended host run over /fleet:reach (melo-desk-001, both desktop lanes). Both lanes run harness-ops 4.4.0, which is later than #6628.
Verdict: no change. Cancels are rare on both desktop lanes. This is from one pass over the desktop OTEL hot store (the collector binds 127.0.0.1 only, so the store holds this machine's lanes alone) (hook_execution_complete events, num_cancelled), 2026-10-22 to 2026-10-24, Claude Code 2.1.305 to 2.1.310:
Lane
hook events
hooks run
cancelled
events with a cancel
non-blocking errors
desktop WSL
174,006
997,174
85 (0.009%)
29
3,192
desktop Windows
179
2,272
0
0
0
The WSL cancels by event: PreToolUse:Bash 41, PostToolUse:Bash 25, PostToolBatch 14, PostToolUseFailure:Bash 3, PreToolUse:Edit 2.
Activity
github-actions commentedon Oct 9, 2026
Thanks for the report. Automated triage handed this issue to a maintainer.
Automated triage. A maintainer reviews it.
kyle-sexton commentedon Oct 13, 2026
Decision recorded. The owner approved applying independently verified recommendations on 2026-10-12.
Basis: https://code.claude.com/docs/en/plugins/loading 'Cleanup of previous versions' fetched 2026-10-12
Three measurements from the 2026-10-08/09 audit (typos-format per-edit cost on native Windows, hook_cancelled counts after #6628, rate-limit readings across an account rotation) still need the desktop lanes. #6628 merged 2026-10-08, so B10 is now measurable.
Verification notes:
Verified by an independent fresh-context check (doubtful, risk low): gh issue view 6687 (body; only triage comment, no measurements posted); gh pr view 6628: MERGED 2026-10-08T02:36:49Z; issue #6156 open; ~/.config/fleet/FLEET.md desktop endpoints; plugins/loading doc cleanup section
Admission to an agent lane stays with the work-class rules and owner ratification.
kyle-sexton commentedon Oct 25, 2026
This was generated by AI during an unattended host run over
/fleet:reach(melo-desk-001, both desktop lanes). Both lanes run harness-ops 4.4.0, which is later than #6628.B6: typos-format per-edit cost on native Windows
Verdict: no change. The per-edit cost no longer exists. typos-format 0.11.4 registers only
SessionStartandStophooks, so it runs once per turn, not once per edit.The measurement was one headless
claude.exe -p --debug-filesession (Claude Code 2.1.309, no concurrent agents). It made 18 separate Edit calls, 6 each to.md,.pyand.txt, in a scratch git repository. Timings come from the debug log'sHooks: ... finished with status N (Nms)lines:The p90 spikes for markdown-format and ruff-format come from the edits to their own extensions. With concurrent agents was not measured: the old 917 ms figure belonged to a per-edit hook that no longer exists. guardrails' 239 ms PreToolUse:Edit is #4390's subject.
kyle-sexton commentedon Oct 25, 2026
This was generated by AI during an unattended host run over
/fleet:reach(melo-desk-001, both desktop lanes). Both lanes run harness-ops 4.4.0, which is later than #6628.B10: hook cancels after #6628
Verdict: no change. Cancels are rare on both desktop lanes. This is from one pass over the desktop OTEL hot store (the collector binds
127.0.0.1only, so the store holds this machine's lanes alone) (hook_execution_completeevents,num_cancelled), 2026-10-22 to 2026-10-24, Claude Code 2.1.305 to 2.1.310:The WSL cancels by event: PreToolUse:Bash 41, PostToolUse:Bash 25, PostToolBatch 14, PostToolUseFailure:Bash 3, PreToolUse:Edit 2.
B12
Not measured: no account rotation happened during this run, and I didn't trigger one. This issue stays open for B12 alone.
Outputs are kept locally under
.work/host-runs/6687/.