Model, delegation & spend policy v3.0 — consolidation + fleet deploy

Host: rdmsm4x (lead) · Session: dev-db [982a7e] · Started 13:32 EDT, closed 13:58 EDT, 2026-08-25 America/New_York

Five overlapping CLAUDE.md sections that answered the same question differently were consolidated into one, the three-way lead-tier contradiction was resolved, the unenforceable budget backstop was replaced with a real MCP tool, and the result was deployed to 5 of 6 fleet hosts across all three agent harnesses.

Why

~/.claude/CLAUDE.md carried five sections governing the same decision, written on five different dates: Spend & scope discipline (v2.4, 08-13), Model & delegation policy (v2.6, 08-15), Second opinions via agy (v2.9, 08-22), Unattended escalation to Fable 5 (v1.0, 08-25), External-budget workers (v1.0, 08-25). They conflicted:

1. Lead tier stated three ways. v2.3 "the main thread never downgrades to save cost" → v2.6 "right-size the lead down to the lowest capable tier" → Fable v1.0 "default tier is Opus 5 high." A stale v2.3 artifact was still sitting at ~/.claude/_fleet-model-policy-block.md.
2. Concurrency caps in three places with different numbers (2 background agents / 2 codex + 1 agy / amended to 1 codex).
3. "Unattended" was undefined yet gated the entire Fable 5 pre-authorization — an agent at 03:00 had no way to determine which mode it was in.
4. The budget backstop was unenforceable. "When a usage bucket passes ~75%, stop delegating" — nothing on any host could read a usage bucket. Searched ~/.claude for usage/quota/budget telemetry: none exists; /usage is interactive-only. Six Macs were following a rule none could evaluate.
5. Fleet drift. rdmsm4x was 743 lines (Aug 25); the other four reachable hosts were 623 lines (Aug 23), missing three whole sections. Codex and agy had no model/delegation policy at all — the policy governing cross-vendor delegation had never been given to the vendors doing the work.

Decisions (Rich, this session)

• Lead tier: Opus 5 high is the default, explicitly right-sized down to Sonnet-class for routine work; Max on demonstrated difficulty; Fable 5 trigger-gated only.
• Structure: consolidate all five into one section rather than patching in place.
• Escalation gate: the ask is the attended/unattended detector. On a trigger the agent always asks; an answer is binding; a 60s timeout or an unavailable AskUserQuestion (headless/cron) means unattended → escalate and log. Once unattended is established, don't re-ask. Rich's added exception: once unattended, the stuck-after-~3-attempts trigger escalates with no ask at all. Any new message from Rich resets the session to attended.
  • An idle-clock threshold (the 8h variant considered) was rejected: during an active overnight run that clock only grows while the agent works, so a long threshold locks escalation during exactly the hours it exists to cover, and unlocks once Rich is awake.
• Budget: build real usage monitoring into Tyrell today; use a stop-flag file in the interim.

What changed

Tyrell usage_status MCP tool — built by the tyrell.app build kickoff session

Coordinated via SendMessage; proposed the contract, that session implemented and shipped it (99f47e5, 8e49d94). Contract: stateless re-read per call; advice: proceed|conserve|hold|unknown

• advice_reason; observed_at always present; unknown first-class (used_pct null, stale degrades to unknown rather than carrying forward); ~/.agent-coordination/BUDGET-HOLD forces hold with the file's first line as the reason; thresholds (60/75) live in the tool so agents branch on the enum, not a percentage. Schema is bucket-name agnostic so agy/ChatGPT buckets ride the same shape.

Known limit (v1): no live vendor telemetry exists. A collector writes ~/.tyrell/usage.json; until one exists the honest answer is unknown = "proceed but announce spend before fan-outs".

Independently verified here, not taken on report — driving the built binary directly over stdio:

• tools/list returns usage_status alongside the four pre-existing tools.
• No sources → {"advice":"unknown","source":"none","observed_at":"2026-08-25T17:51:03Z"}.
• With BUDGET-HOLD present → {"advice":"hold","advice_reason":"<file's first line>","source":"manual-flag"}.
• Test flag removed afterwards; confirmed absent.

~/.claude/CLAUDE.md → v3.0

New single section "Model, delegation & spend (v3.0)" at line 40, 221 lines, seven subsections: tier ladder (table) · usage_status budget check · 10 delegation rules · Fable triggers + gate · worker tiers across vendors · agy specifics · the delegation test · revision log v2.3→v3.0.

Removed (content preserved in the consolidation, nothing dropped):

• lines 40–91 Spend & scope + Model & delegation policy
• lines 437–504 the agy-second-opinions v2.9 delimited block (markers included; verified no script injects that block — it was a one-off hand recovery on 08-23, so removing the markers is safe)
• lines 653–743 Unattended escalation + External-budget workers

Header bumped v2.8 → v3.0; line-3 preamble updated to name the consolidated policy and to state it binds Codex and agy too.

Fleet deployment

New script: ~/dev/fleet/maintenance/scripts/deploy_model_policy_v3.zsh (v1.0).

• ~/.claude/CLAUDE.md: whole-file copy from rdmsm4x, divergence-gated — a host is overwritten only if it has zero lines the pre-deploy canonical lacked; otherwise skipped and reported, never clobbered. All four reachable peers measured zero unique lines.
• ~/.codex/AGENTS.md and ~/.gemini/GEMINI.md: the policy injected as a delimited fleet-model-delegation-spend v3.0 block via the existing _inject_block.py, so peer-written content outside the markers survives byte for byte. The block carries an orientation preamble telling non-Claude harnesses they are a worker tier — §3/§5/§7 bind them, §1/§2/§4 are context.
• Per-host backups written before every write: CLAUDE.md.bak-pre-v3-<ts>, AGENTS.md.bak-pre-v3-<ts>, GEMINI.md.bak-pre-v3-<ts>.

Result — 5/6 hosts, all three context files each:

Verified independently of the deploy script's own report, by SSH to rdmbair15m5 and rdmpw3275m: v3.0 header present, exactly one ## Model, delegation & spend (v3.0 heading, zero superseded section headings remaining, and usage_status referenced 5× in each of the three context files.

Reproduce

# policy present and superseded sections gone, any host
grep -c '^## Model, delegation & spend (v3.0' ~/.claude/CLAUDE.md          # -> 1
grep -c '^## Spend & scope discipline\|^## Unattended escalation' ~/.claude/CLAUDE.md   # -> 0

# usage_status, both paths (rdmsm4x)
B=~/dev/apps/Tyrell/.build/release/tyrell-mcp
printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"v","version":"1"}}}' \
  '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"usage_status","arguments":{}}}' | $B | tail -1

# re-deploy / catch up a host
zsh ~/dev/fleet/maintenance/scripts/deploy_model_policy_v3.zsh --dry-run
zsh ~/dev/fleet/maintenance/scripts/deploy_model_policy_v3.zsh --host jdmbair13m5

Undo

cp ~/.claude/CLAUDE.md.bak-pre-v3-modelpolicy-20260825-135234 ~/.claude/CLAUDE.md   # rdmsm4x
# peers: ~/.claude/CLAUDE.md.bak-pre-v3-20260825-135552
# codex/agy: ~/.codex/AGENTS.md.bak-pre-v3-20260825-135552, ~/.gemini/GEMINI.md.bak-pre-v3-20260825-135552

Outstanding — owner actions

1. jdmbair13m5 has not received v3.0 (asleep at deploy time; the Tyrell session reported the same host unreachable for its own deploy). Re-run the deploy with --host jdmbair13m5 when it wakes. Until then that host's agents follow the superseded 08-23 policy.
2. No usage collector exists, so usage_status returns unknown everywhere. The enforceable control today is ~/.agent-coordination/BUDGET-HOLD, which is Rich's to set — agents are explicitly forbidden from creating or deleting it.
3. Second ChatGPT Pro account not authed ([email protected]); codex remains authed to one.
4. Grok SuperHeavy joins the worker tiers when live; codex caps revisit at that point.
5. No second opinion obtained on this revision — not routed to agy. The changes are policy text plus a reversible, backed-up file deployment, not a data-destroying operation.

No secrets were written to any file, message, or note in this session.