mars
MARS-773
Move root CLAUDE.md content into per-lane seat files (Elazar 3b)
Done low
cmcoder-mars-cc
Elazar ruling 3b, CLI 2026-09-28: a project-root CLAUDE.md auto-loads for every seat beneath it, so Mars's 2829-byte root was paid 11x per turn while seat files were 179-279 byte stubs - nearly all per-agent content was inherited and untailored. Root cut to a 502-byte identity statement; content moved down into each seat's own CLAUDE.md, scoped to that lane. Measured per-turn drop: db -56pct, audit -64, pm -63, researcher -68, coder -34.
Activity
-
Pushed 5457699, v2.82.14, live-verified: /api/health reports 2.82.14 and remote main package.json reads 2.82.14 (read via gh api at check time, not the local checkout). Audit PASS:5457699 shallow, no findings, from audit-general-ca - audit-general-cc is down and the hub forwarded. Audit confirmed all nine .claude/rules dirs resolve, seat identities match .agent-name, and lane content landed in the right seats; it also caught that the old root line 'Only coder/audit agents push' was stale against current commons (audit never pushes) and is correctly gone. TWO COVERAGE LIMITS, both stated by the reviewer and neither closed: it was a documentation-only source review - nobody has loaded a FRESH Claude Code session in any of these seats to confirm the new files actually load as intended, and the reviewer did not independently verify the live/remote version (I did). The byte figures are file sizes, not measured context usage. Rollout to pluto/venus/terra/enamel dispatched to monitor-context-cc with the two traps found here: a relay/proxy seat must carry NO commons symlinks (Mars's proxy was silently paying 23KB/turn), and per-seat CLAUDE.md files are often untracked, so --only must name each file explicitly or it sweeps the seat's settings and rules in with it.
refactor
2w ago by coder-mars-cc
2w ago
2026-09-28 08:56