rdmsm4x-changelog-20260827-2322-spend-directive-lifted

rdmsm4x — spend directive lifted early

2026-08-27 23:22 EDT · rdmsm4x · session start 23:18 EDT

What Rich asked

"Remove/lift our constraint, we have < 23 hours left, and it is blocking work by other agents."

What the constraint was

Bus topic spend-directive-downgrade-subagents-until-friday-20260828, issued by Rich 2026-08-25 04:40 EDT (relayed by claude@jdmbair13m5), recorded in ~/.agent-coordination/ACTIVE-DIRECTIVES.md as ACTIVE until Friday 2026-08-28 EOD. Tightened CLAUDE.md's delegation policy: cheapest-tier-only subagents, no premium workers without per-case approval, prefer inline, no fan-outs without approval, stop delegating past ~75% of a bucket. Identified by the ~23h-to-expiry marker, which matches this directive uniquely.

Changes made

~/.agent-coordination/ACTIVE-DIRECTIVES.md — spend directive marked RELEASED 2026-08-27 23:22 EDT; original text retained in a collapsed block; budget note appended. Backup at ACTIVE-DIRECTIVES.md.bak-20260827-232*.
Broadcast 20260827-232015-791ABAC9 on the original topic — states what is lifted, what reverts to standing CLAUDE.md policy, budget state, and the three still-gated carve-outs. ACK requested from any agent that was holding work.
Directives file copied to reachable hosts: rdmbair13m5, rdmpw3275m, jdmbair13m5 (verified).

Not changed (deliberately)

CLAUDE.md standing delegation policy — the directive tightened it, never replaced it. Caps stand: 2 Claude background subagents · 1 codex · 1 agy. Bus traffic uncapped.
The three fleet-halt carve-outs: rdRECEIPT push, customer repo deletion, no deletions on jdmbair13m5. Separate matter, no expiry clock, not covered by this request.
permissions.defaultMode (stays bypassPermissions). No BUDGET-HOLD / UNATTENDED marker was created or deleted — neither file exists on this host; both are Rich's controls.

Budget at release (2026-08-27 23:18 EDT)

usage_status advice = proceed, but that grade rests on OpenAI alone (9%, vendor-authoritative). Anthropic weekly is ungraded, locally measured 153.2M / 162M provisional ceiling (~94.6%): Opus 5 82.9M, Fable 5 60.7M, Sonnet 5 9.0M. Provisional ceiling, not vendor quota — directional.

Open

rdmbair15m5 and rdmpw3265m were unreachable at broadcast; both pick the release up from the replicated maildir on next sync. Their ACTIVE-DIRECTIVES.md still shows the directive ACTIVE until they sync.
No Apple Notes entry published for this changelog yet.

Follow-up 2026-08-27 23:24 EDT — "agents blocked by our 92% quota / maybe our global rules?"

Our global rules are not blocking agent-to-agent communication. Evidence:

CLAUDE.md carries Rich's own 2026-08-27 exception: agent-to-agent chat is explicitly NOT capped; inbound peer messages uncapped.
~/.claude/settings.json:172 → "crossSessionInbound": "accept". Correct.
Only hook configured is SessionStart → statusline-verify.sh. Nothing blocking.
No quota/percentage gate anywhere in ~/.agent-coordination/*.zsh.
No BUDGET-HOLD, no UNATTENDED marker on this host.
ListAgents: 16 peer sessions visible/reachable. Heartbeats current for 5 of 6 hosts (only rdmpw3265m stale, last beat 2026-08-26 14:59). Today's mail flowing normally (fleet-audit 23:11, fleetstate 23:07). Oldest unread in my inbox is 2026-08-22 — nothing recent is failing to land.

The 92% is real and is a VENDOR limit, not one of ours. ~/.tyrell/usage.json bucket 4: anthropic_vendor_seven_day, used_pct: 92.0, authoritative: true, source: vendor-api, resets 2026-08-28T14:00:00Z = 2026-08-28 10:00 EDT (~10.7h from now). No edit to CLAUDE.md, no marker file, no policy change can lift a server-side vendor quota.

The one policy-level self-block that DID exist is already lifted — the spend directive's "past ~75% of a bucket, stop delegating." At 92% that rule halts all delegation fleet-wide. Released earlier this session (broadcast 20260827-232015-791ABAC9). CLAUDE.md's standing policy does not self-block: it says branch on advice, and advice reads proceed.

OPEN DEFECT found while diagnosing — usage_status under-reports Anthropic

usage_status returned advice: proceed, advice_reason: "worst bucket 9% (openai)", and by_vendor containing openai only — while an authoritative Anthropic row sat at 92%. The authoritative anthropic_vendor_seven_day bucket is not being rolled into advice or by_vendor. Related: ~/.tyrell/vendor-gate.json has its anthropic key caching an openai_vendor_primary row — anthropic's cache slot holds OpenAI's data.

Impact: every agent that follows CLAUDE.md correctly ("branch on advice, never the percentage") is told proceed while the real Anthropic weekly bucket is at 92%. This is the reverse of a block — it is an under-report that invites overspend into a hard vendor stop. Not fixed in this session; awaiting Rich's go-ahead.