Since 2026-10-06 the Codex lanes run gpt-6.1-sol, not gpt-5.5. GPT-5.5 leaves Codex for ChatGPT sign-in on 2026-10-14, so rig-gitops#923 and #933 moved every gpt-5.5 model line in the tables below. That covers the Codex lanes' main model and the providers.codex-cli.model settings. Everything else on this page is the 2026-09-25 snapshot.

Rig agent model and effort inventory

Snapshot of 2026-09-25, before tuning. The Mac Mini produced this read-only inventory (job rig-models, effort high). The orchestrator re-checked its key claims: the OAuth mapping, resumed runs dropping --model, and the image's Claude Code 2.1.281. Citation prefixes: gitops: = dashecorp/rig-gitops 9b65b3f, rar: = rig-agent-runtime, rc: = rig-conductor, c3: = this repo.

Changed since, same day: - Every Review-E Claude leg now runs claude-sonnet-5 (rig-gitops#855, verified in-cluster). - llm.effort now reaches claude-cli, claude-tmux and codex-cli, and resumed claude-cli runs keep --model and --system-prompt (rig-agent-runtime#749). Effort is still unset everywhere. - codex-cli no longer throws when passMcpToCli is off (rig-agent-runtime#750).

Tuning plan: one fleet at a time, each change measured against the 30-day baseline (2026-08-26 to 2026-09-25): Dev-E 88 PRs, 82% merged, 64% approved on the first Review-E verdict; Review-E 698 reviews, of which 394 approved, 145 changes requested, 101 commented and 58 dismissed. Next: Dev-E and the planner to claude-opus-5-5, then effort per role (see Mac Mini delegation for the routing principle).

Read-only inventory, 2026-09-25. Nothing was changed in any repo.

Sources and how to read the citations

Prefix	Path under ~/repos/	State read
gitops:	rig-gitops-fresh/	branch main-sync = 9b65b3fd (.git/refs/heads/main-sync:1) = dashecorp/rig-gitops main 9b65b3f. This is what Flux deploys.
rar:	dashecorp-rig-agent-runtime/	The working tree is on fix/issue-744-review-guard-checks-pr = 15ceb0f (.git/HEAD:1). The deployed image is node-sha-008ed87b… (gitops:apps/dev-e/rig-agent-helmrelease.yaml:22) = local main-sync (.git/refs/heads/main-sync:1). I could not diff them because git is permission-gated. The only #744 code cited here is src/closed-assignment-guard.js; treat that one as possibly undeployed. The chart itself is pulled from main (gitops:apps/dev-e/rig-agent-source.yaml:9-10), which is newer than the image.
ae:	automate-e/	main = 008ed87b (.git/refs/heads/main:1). This is the rig-agent-runtime checkout that iBuild-E actually runs on the Mac Mini (c3:docs/infrastructure/hosts.md:137).
rc:	rig-conductor-rigbuild/	branch design-2394 (.git/HEAD:1). It is not certain this equals the deployed conductor.
c3:	claude-3-944/	a worktree

Deployment scope: Flux reconciles apps/{dev-e,review-e,rig-planner,rig-assistant,dashecorp,stigjohnny} on GCP (gitops:clusters/gcp/kustomization.yaml:5-20). apps/dev-e and apps/review-e are also reconciled on the legacy Dell cluster (gitops:clusters/dell/kustomization.yaml:6-7, gitops:README.md:33), where the pods have sat Pending since May (gitops:docs/unschedulable-pod-watch.md:44). apps/dev-e-staging is inert: staging-rig is suspend: true (gitops:clusters/gcp/staging-rig.yaml:17).

1. Inventory

Row legend: - Effort: no row can set effort, reasoning or thinking. See the "Effort" note below the table. - Chain: the provider chain the runtime actually walks. It is [provider, ...fallbackProviders, ...keys(providers)] with duplicates removed (rar:src/agent/provider-chain.js:35-60), so every key under llm.providers is a fallback, not just the entries in fallbackProviders. Each hop uses that provider's providers.<p>.model override (rar:src/agent/provider-chain.js:62-72). - Auth: every in-cluster claude pod gets CLAUDE_CODE_OAUTH_TOKEN, mapped from secret key anthropic-api-key (rar:charts/rig-agent-runtime/templates/deployment.yaml:80-84). The chart never sets ANTHROPIC_API_KEY (gitops:docs/claude-tmux-pilot.md:31). Codex pods with CODEX_DEVICE_AUTH=true drop OPENAI_API_KEY and use the ChatGPT login in CODEX_HOME/auth.json (rar:src/agent/providers/codex-auth.js:46-57). - Cost: "flat" means subscription quota: the Anthropic Max plan (gitops:docs/whitepaper/cost-framework.md:263) or the ChatGPT plan. The dollar figures the rig reports for these are counterfactual, not billed (gitops:docs/claude-tmux-pilot.md:52, rar:src/agent/providers/claude-tmux.js:120-133).

Base fleet (GCP, live)

Fleet / variant (AGENT_ID)	Active provider	Model	Chain actually walked	Context	Max turns / timeout	Replicas / KEDA	Auth → cost	Where set
Dev-E node, claude (dev-e-node)	claude-tmux (pilot)	claude-opus-4-7[1m]	claude-tmux → claude-cli (opus-4-7[1m]) → codex-cli (gpt-5.5)	[1m]	300 is ignored because tmux passes no --max-turns (rar:src/agent/providers/claude-tmux.js:359). Limits: 1,200,000 ms, idle-stall 360,000 ms.	no replicaCount, so chart default 1 (rar:charts/rig-agent-runtime/values.yaml:4). KEDA 0..1 on signal:dev-e list (gitops:apps/dev-e/scaledobject.yaml:12-19).	OAuth → flat. The codex hop has no credential: no CODEX_DEVICE_AUTH, and dev-e-secrets has no openai-api-key key (gitops:apps/dev-e/dev-e-secrets.sops.yaml:7-12).	gitops:apps/dev-e/rig-agent-helmrelease.yaml:196-212
Dev-E node, codex (dev-e-node-codex)	codex-cli	gpt-5.5	codex-cli → claude-cli (claude-sonnet-4-6)	n/a	150 is ignored: codex passes no max-turns (rar:src/agent/providers/codex-cli.js:598-638). Limit 1,200,000 ms.	no replicaCount (1). KEDA 0..1. The comment says "pinned to 0 by paused-replicas", but the annotations: block is empty, so it is live (gitops:apps/dev-e/scaledobject-codex.yaml:6,14).	device-auth, so ChatGPT plan → flat (…-codex.yaml:277-278). The claude hop is OAuth → flat.	gitops:apps/dev-e/rig-agent-helmrelease-codex.yaml:197-210
Dev-E dotnet, claude (dev-e-dotnet)	claude-cli	claude-opus-4-7[1m]	claude-cli only (no providers)	[1m]	300 / 1,200,000 ms	default 1. KEDA 0..1 (gitops:apps/dev-e/dev-e-dotnet-scaledobject.yaml:12-13).	OAuth → flat	gitops:apps/dev-e/dotnet-helmrelease.yaml:137-144
Dev-E dotnet, codex (dev-e-dotnet-codex)	codex-cli	gpt-5.5	codex-cli only	n/a	150 (ignored) / 1,200,000 ms	replicaCount: 1, KEDA min 1, so always on (gitops:apps/dev-e/dev-e-dotnet-scaledobject-codex.yaml:24-25)	ChatGPT → flat	gitops:apps/dev-e/dotnet-helmrelease-codex.yaml:143-150
Dev-E python, claude (dev-e-python)	claude-cli	claude-opus-4-7[1m]	claude-cli only	[1m]	300 / 1,200,000 ms	default 1, KEDA 0..1	OAuth → flat	gitops:apps/dev-e/python-helmrelease.yaml:137-144
Dev-E python, codex (dev-e-python-codex)	codex-cli	gpt-5.5	codex-cli only	n/a	150 (ignored) / 1,200,000 ms	replicaCount: 0 plus KEDA paused-replicas: "0", so off (…python-helmrelease-codex.yaml:21, …dev-e-python-scaledobject-codex.yaml:7)	ChatGPT → flat	gitops:apps/dev-e/python-helmrelease-codex.yaml:143-150
Review-E, claude (review-e)	claude-tmux	claude-sonnet-4-6 (#646). No [1m] since 2026-06-18.	claude-tmux → claude-cli (sonnet-4-6) → codex-cli (gpt-5.5)	200K	150 (ignored under tmux) / 1,200,000 ms, idle 360,000 ms	replicaCount: 1 (…rig-agent-helmrelease.yaml:21). KEDA 0..1 on signal:review-e (gitops:apps/review-e/scaledobject.yaml:18-26).	OAuth → flat. The codex hop has no credential (same reason as Dev-E node).	gitops:apps/review-e/rig-agent-helmrelease.yaml:202-218
Review-E, codex (AGENT_ID also review-e)	codex-cli	gpt-5.5	codex-cli → claude-cli (sonnet-4-6, #648)	n/a	75 (ignored) / 1,200,000 ms	replicaCount: 1 (…-codex.yaml:23). KEDA 0..1 on the same signal:review-e list (gitops:apps/review-e/scaledobject-codex.yaml:17-26).	No CODEX_DEVICE_AUTH (…-codex.yaml:243-265), so the env passes through unchanged and auth depends on whether auth.json exists on the PVC. Comments call the plan "ChatGPT Pro", "Team" and "Business" (…-codex.yaml:21,206, scaledobject.yaml:7). → flat	gitops:apps/review-e/rig-agent-helmrelease-codex.yaml:205-218
Planner, claude (rig-planner)	claude-cli	claude-opus-4-7[1m]	claude-cli → codex-cli (gpt-5.5)	[1m]	50 / 1,200,000 ms. Temperature 0.1 is inert: CLIs get no temperature (see §2).	replicaCount: 1, KEDA paused-replicas: "1" (manual pin) (gitops:apps/rig-planner/keda-scaledobject.yaml:9)	OAuth → flat	gitops:apps/rig-planner/rig-agent-helmrelease.yaml:118-132
Planner, codex (AGENT_ID rig-planner-codex, but the heartbeat agentId is rig-planner)	codex-cli	gpt-5.5	codex-cli → claude-cli (sonnet-4-6)	n/a	30 / 1,200,000 ms	replicaCount: 0, paused 0: off (…-codex.yaml:21, keda-scaledobject-codex.yaml:7)	ChatGPT → flat	gitops:apps/rig-planner/rig-agent-helmrelease-codex.yaml:115-129
Rig-E, tenant chat agent (rig-e)	openai-api → RunPod vLLM	Qwen/Qwen2.5-32B-Instruct-AWQ	openai-api only. No fallback, deliberately (…rig-e-helmrelease.yaml:258-265).	16,384 tokens, about 13,400 usable (c3:docs/infrastructure/gpu-brain.md:33)	15 / 300,000 ms	replicaCount: 1, no KEDA (gitops:apps/dashecorp/kustomization.yaml:21-25)	OPENAI_API_KEY = the pod's VLLM key (gitops:apps/dashecorp/rig-e-secrets.sops.yaml.example:12-13) → per GPU-hour: at most $2.00/h, last seen $1.59/h, about $0.24–0.53 per conversation (c3:docs/infrastructure/gpu-brain.md:31-37)	gitops:apps/dashecorp/rig-e-helmrelease.yaml:237-267
rig-assistant, the "manager". It is not Rig-E (gitops:apps/dashecorp/rig-e-helmrelease.yaml:3).	claude_code backend (claude -p)	claude-sonnet-4-6 via ASSISTANT_LLM__MODEL. The app default is claude-opus-4-8.	none visible	not visible	not visible. The rig-assistant code is not in the sources.	Deployment replicas: 1 (gitops:apps/rig-assistant/deployment.yaml:7)	CLAUDE_CODE_OAUTH_TOKEN, the operator's token (…rig-assistant-secrets.sops.yaml:9) → flat	gitops:apps/rig-assistant/deployment.yaml:45-54

Tenant fleets (GCP)

Fleet	Provider / model	Chain	Turns	State
dashecorp Dev-E node (dev-e-node, tenant)	claude-cli / claude-opus-4-7[1m]	claude-cli only	300	replicaCount: 0 (gitops:apps/dashecorp/dev-e-helmrelease.yaml:30,220-231). KEDA scales 0→1 on signal:…@dashecorp, with no pause annotation.
dashecorp Review-E	claude-cli / claude-sonnet-4-6	claude-cli only	150	replicaCount: 0 (gitops:apps/dashecorp/review-e-helmrelease.yaml:34,198-208), KEDA 0..1
dashecorp Planner	claude-cli / claude-opus-4-7[1m]	claude-cli → codex-cli	50	replicaCount: 0 (gitops:apps/dashecorp/planner-helmrelease.yaml:31,128-141), KEDA 0..1
stigjohnny Dev-E / Planner / Review-E / Review-E-codex	opus-4-7[1m] / opus-4-7[1m] / opus-4-7[1m] / gpt-5.5 (claude hop opus-4-7)	Each includes codex-cli, or claude-cli for the codex variant.	300 / 50 / 150 / 75	All gated: replicaCount: 0 plus paused-replicas: "0" (gitops:apps/stigjohnny/{dev-e,planner,review-e,review-e-codex}-{helmrelease,scaledobject}*.yaml). Review-E here never got the #646 trim: it is on claude-opus-4-7[1m] (gitops:apps/stigjohnny/review-e-helmrelease.yaml:216,224), and the codex leg's claude hop is claude-opus-4-7 (…review-e-helmrelease-codex.yaml:224).

Off-cluster

Agent	Provider / model	Chain	Turns / timeout	Auth → cost	Where set
iBuild-E (Mac Mini, what actually runs)	rig-agent-runtime claude-cli / claude-opus-4-7[1m]	claude-cli → codex-cli (gpt-5.4, device-auth)	200 / 1,200,000 ms	Host CLAUDE_CODE_OAUTH_TOKEN (c3:docs/infrastructure/hosts.md:115) → flat	ae:character.json:79-98. The process is node src/index.js in automate-e (c3:docs/infrastructure/hosts.md:137). This is a local host file, not gitops.
iBuild-E, legacy agent-runner config named in the brief	no model key, so the CLI default	none	--max-turns 50 hard-coded. --max-budget-usd 2.0, or 10.0 for /opus.	OAuth → flat	c3:agent-runner/agents/ibuild-e.json:1-18, c3:agent-runner/src/session-manager.js:78-92. The deployed daemon is automate-e, not this. The Mini also carries an uncommitted edit to this file (c3:docs/infrastructure/hosts.md:118).
Volt-E / Pi-E / iClaw-E (agent-runner, outside the rig)	"model": "opus" / none / none	none	50	OAuth → flat	c3:agent-runner/agents/{volt-e,pi-e,iclaw-e}.json (per subagent read: volt-e.json:12)

Effort / reasoning column (all rows): not settable

• claude-cli: argv carries --model, --system-prompt, --max-turns and permission/MCP flags only (rar:src/agent/providers/claude-cli.js:189-221). There is no --effort, no MAX_THINKING_TOKENS and no --fallback-model anywhere in src/.
• claude-tmux: [claude, '--model', m, '--dangerously-skip-permissions'] (rar:src/agent/providers/claude-tmux.js:359).
• codex-cli: exec --json … --model m -C cwd --search. Its only -c flags are for mcp_servers.* (rar:src/agent/providers/codex-cli.js:598-659). There is no model_reasoning_effort and no config.toml is written.
• anthropic SDK: no thinking field (rar:src/agent/providers/anthropic-sdk.js:47-54).
• openai-api: only temperature (rar:src/agent/providers/openai-api.js:338).
• iBuild-E agent-runner: none (c3:agent-runner/src/session-manager.js:82-92).
• The baked Claude settings file rar:config/claude-settings.json (copied at rar:Dockerfile.base:107) has no model, effort or thinking key.

[1m]

It is passed verbatim as --model (rar:src/agent/providers/claude-cli.js:202, claude-tmux.js:359). The runtime does not parse or strip it and sends no beta header. It appears only as a pricing key (rar:src/usage.js:12). Whether claude-opus-5-5[1m] or claude-sonnet-5[1m] is a valid CLI alias is unverified.

2. Levers

A. Gitops value only (HelmRelease character.llm.*, extraEnv, replicaCount, KEDA). Takes effect on the next Flux reconcile, about 10 min.

Lever	Key	Notes
Model per fleet	llm.model and llm.providers.<p>.model	The model is required and has no default (rar:src/character.js:92-95). Change both, or a fallback hop silently runs the old model (the reason for #648).
Provider	llm.provider	Values: claude-cli / claude-tmux / codex-cli / openai-api
Fallback chain	llm.fallbackProviders and the set of keys under llm.providers	Removing a codex-cli: key removes codex from the chain. Removing a key is the only way to stop a claude pod from falling to a codex hop that has no credentials.
Turns	llm.maxTurns	Effective only on claude-cli (and openai-api/anthropic). It is not used by claude-tmux or codex (rar:src/agent/providers/claude-tmux.js:359, codex-cli.js:598-638).
Timeouts	llm.timeoutMs, llm.idleStallMs	rar:src/agent/providers/claude-cli.js:277, claude-tmux.js:368,477, codex-cli.js:435
Capacity	replicaCount (0 is now honoured, rar:charts/rig-agent-runtime/templates/deployment.yaml:12-16) and autoscaling.keda.sh/paused-replicas	Max is 1 per fleet because of the shared PVC (gitops:apps/review-e/scaledobject.yaml:19-21)
CLI env passthrough (hypothesis, verify first)	extraEnv	The claude child inherits the whole pod env (rar:src/agent/providers/claude-cli.js:151-158; claude-tmux uses the same builder, claude-tmux.js:324). Any env-based Claude Code setting could therefore be set with no code change, for example MAX_THINKING_TOKENS, or ANTHROPIC_MODEL to pin the model on the --resume path (§5). Whether the pods' CLI version honours a given variable must be checked against that version first.
Temperature	llm.temperature	Inert for every CLI agent. Only the openai-api and anthropic providers read it (rar:src/agent/providers/openai-api.js:338, anthropic-sdk.js:50).
Routing (conductor config, not gitops)	issue labels provider:codex / provider:claude, and routeTierToCodexFirst in the target repo's .rig-agent.yaml	rc:src/ConductorE.Api/Program.cs:1264-1320, rc:src/ConductorE.Core/UseCases/TierCodexRouter.cs:63-91

B. Needs a runtime code change (rig-agent-runtime, then an image-pin bump in gitops)

Change	Where
Pass --effort (claude)	Both argv branches at rar:src/agent/providers/claude-cli.js:189-206, and rar:src/agent/providers/claude-tmux.js:359. Config plumbing is free: llm.effort or llm.providers.<p>.effort already flows through resolveCharacterForProvider (rar:src/agent/provider-chain.js:62-72) and the raw-JSON ConfigMap (rar:charts/rig-agent-runtime/templates/configmap.yaml:8-9).
Pass -c model_reasoning_effort=… (codex)	Next to the -c loop at rar:src/agent/providers/codex-cli.js:633-635, fed from the caller at :416-424
Keep --model on --resume	rar:src/agent/providers/claude-cli.js:189-197
Honour maxTurns under claude-tmux and codex	rar:src/agent/providers/claude-tmux.js:359, codex-cli.js:598-638
Pricing entries for 5.x models	rar:src/usage.js:4-15 (unknown models fall to $1/$5), rar:src/agent/providers/claude-tmux.js:128-130 (non-sonnet/haiku are priced as Opus 15/75), and the conductor's rc:src/ConductorE.Core/UseCases/AnthropicPricing.cs:24-65 (unknown models priced as sonnet-4-5)
Report cost once, with the model that ran	rar:src/stream-consumer.js:1076-1085 duplicates the provider's own reportTokenUsage (claude-cli.js:431-439, claude-tmux.js:556-562) and labels it with the top-level character.llm.model, not the fallback hop's model
Pin the Claude Code / codex CLI versions	rar:Dockerfile.base:69 installs unpinned @anthropic-ai/claude-code @openai/codex. Node, dotnet and python images build FROM …:base (rar:Dockerfile.node:3). claude-opus-5-5 needs CLI 2.1.280 or later (c3:docs/infrastructure/hosts.md:114), and the in-pod version is unknown from source.
Distinct identity for the Review-E codex pod	A gitops AGENT_ID change, but coupled to the conductor: the router expects review-e-codex (rc:src/ConductorE.Core/UseCases/QuotaAwareReviewRouter.cs:45-48), and today both pods report review-e (gitops:apps/review-e/rig-agent-helmrelease-codex.yaml:63,246).

C. Impossible, or not possible without a design change

• Per-task model or effort by tier. The dispatch payload has no model or effort field (rc:src/ConductorE.Api/Adapters/ValkeyStreamPublisher.cs:164-185). Tiers only gate dispatch and optional codex-first routing (rc:src/ConductorE.Api/Services/TaskSpecConsumerService.cs:71-95). Per-PR "Sonnet default, Opus on T2/T3" (gitops:docs/2026-06-18-review-e-sonnet-cost-trim.md:38) therefore needs conductor plus runtime work, or separate fleets.
• Effort for Rig-E. Qwen2.5-32B-Instruct has no reasoning mode, and its context is capped at 16K by the GPU pod (c3:docs/infrastructure/gpu-brain.md:33).
• Same-provider model fallback (Opus → Sonnet on quota). Only a provider switch changes the model (rar:src/agent.js:119-185).
• iBuild-E via gitops. Its config is a local file on the Mac Mini (ae:character.json) and has to be changed on that host.

3. Outcome data that exists to judge a change

Conductor base: http://rig-conductor-api.rig-conductor.svc.cluster.local:8080. Event wire names are in rc:src/ConductorE.Api/…/SubmitEvent.cs, and records in rc:src/ConductorE.Core/Domain/Events.cs. The conductor citations below come from a delegated read; QuotaAwareReviewRouter.cs, TierCodexRouter.cs and ReviewDispatchRouter.cs were re-read directly.

Role	"Model or effort was wrong" signal	Events	Where to read it
Dev-E	First-pass merge rate, review rounds per PR, CI failures per PR, re-dispatches, stuck/crashed runs, turns and duration per issue	PR_CREATED, MERGED / MERGE_GATE_MERGED, REVIEW_DISPUTED (= CHANGES_REQUESTED), CI_FAILED / BUILD_FAILED, ISSUE_ASSIGNED.Attempt, ISSUE_UNASSIGNED, SLO_BREACHED, AGENT_STUCK, ISSUE_STUCK, CLI_COMPLETED (Reason = completed/crashed/timeout, plus Turns, DurationMs, CostUsd), ISSUE_DONE.DurationMinutes, TESTS_RUN (runtime hook rar:hooks/pre-push-tests.sh:409)	Per issue: GET /api/issues/trace?repo&issueNumber. Rounds: IssueStatus.ReviewRoundsCount, via GET /api/issues/item. GET /api/ci-failures, GET /api/stuck-issues. Turn p50/p90/p95 and cost by agent: GET /api/execution-logs/stats?agentId&days. Per run: GET /api/execution-logs (includes Model). Also GitHub PR reviews and checks.
Review-E	False blocks (CHANGES_REQUESTED followed by a merge with no code change, or an operator override), misreadings (dispute escalations), lost reviews (a dispatch with no verdict), rounds hitting the cap	REVIEW_PASSED, REVIEW_DISPUTED, REVIEW_FAILED, RE_REVIEW_REQUESTED, DISPUTE_ESCALATED. The max-rounds cap is 3 (MAX_REVIEW_ROUNDS). Operator merge without approval is auto-inferred at rc:Program.cs:2037-2084.	Verdicts come from the GitHub pull_request_review webhook when the login contains "review-e" (rc:Program.cs:2347-2464,2573-2581). The runtime no longer emits verdicts (rar:src/review-follow-up.js:1-12). The ground truth is GitHub PR reviews by review-e-bot.
Planner	Re-plans, T3 escalations, sub-issues that fail downstream	TIER_CLASSIFIED, TASK_SPEC_POSTED, EPIC_DECOMPOSED. There is no dedicated re-plan event: count repeated TASK_SPEC_POSTED per issue.	GET /api/issues/trace
All	Quota pressure, fallback use	AGENT_QUOTA_REPORTED, PROVIDER_EXHAUSTED (defined in the conductor but never emitted by the runtime, rar:docs/usage-tracking.md:126), CLI_STARTED.Model, AgentStatus.LastModel	GET /api/agents. Fallback hops are only in CLI_PROGRESS text, throttled to one per 55 s, with "Using provider" messages dropped (rar:src/stream-consumer.js:784-797), plus GET /api/agent-logs/{agentId}?count= (gitops:docs/claude-tmux-pilot.md:68-70) and pod stdout, which rotates out within about 30 min.
Cost	Spend or quota per model	TOKEN_USAGE (on the agent stream, not the issue stream)	GET /api/usage?agentId&repo&days, /api/costs/summary?days, /api/costs/daily?days, /api/costs/issue. None of them group by model.

Gaps that block a clean A/B today: 1. The two Review-E pods share AGENT_ID review-e, so verdicts, heartbeats and costs cannot be split by agent. The only split is ExecutionLog.Model per run. 2. There is no fallback event, so a run that silently fell to another provider's model cannot be told apart in TOKEN_USAGE, which carries the top-level model label. 3. Costs are counterfactual and double-counted (§5). 4. A baseline has to be pulled before changing anything, because the projections are cumulative.

4. Gaps and hypotheses

Each item is a hypothesis to test with the data in §3. None is a recommendation.

#	Hypothesis	Test	Cost / quota side
H1	Dev-E on claude-opus-4-7[1m] → claude-opus-5-5 raises first-pass merge rate and lowers review rounds and CI fails per PR.	Before/after on the node fleet: rate of MERGED without REVIEW_DISPUTED, ReviewRoundsCount, CI_FAILED per PR, and CLI_COMPLETED.Reason. First confirm the in-pod CLI is 2.1.280 or later.	Flat quota (OAuth). Opus-class weekly quota burn is the risk, not dollars. No approval needed for spend unless the credential changes.
H2	Review-E claude-sonnet-4-6 → claude-sonnet-5 (the 2026-09-20 decision) lowers false blocks and misreadings at similar quota use.	False-block rate and DISPUTE_ESCALATED per review, split by ExecutionLog.Model, because the codex pod shares the agent id.	Flat quota. Pricing tables have no sonnet-5 entry, but tmux's startsWith('claude-sonnet') gives it 3/15 (rar:src/agent/providers/claude-tmux.js:128), so its reported cost is only as good as that guess.
H3	Planner opus-4-7[1m] → opus-5-5 (or sonnet-5) cuts re-plans and T3 escalations.	Repeated TASK_SPEC_POSTED per issue; downstream Dev-E success on planned issues	Flat quota
H4	[1m] is not needed for the planner (50 turns, one TaskSpec per run, rig-agent-helmrelease.yaml:122) or Review-E (already dropped). Dropping it on Dev-E may cost quality on large repos.	Compare CLI_COMPLETED crashes and context errors before and after	On a subscription, [1m] may draw quota faster. Untested.
H5	Effort high or xhigh on Dev-E and Review-E (after lever B) reduces rounds and false blocks more cheaply than a model bump.	Same metrics, same model, effort only	Flat quota, more tokens per task
H6	Codex model_reasoning_effort (after lever B) closes the quality gap between the codex and claude variants.	Tier-t1 codex-first issues vs claude issues on merge rate	ChatGPT plan quota; bursts already exhausted the 5h window at about 9 PRs (gitops:apps/review-e/rig-agent-helmrelease.yaml:17-20)
H7	Pinning ANTHROPIC_MODEL via extraEnv stops resumed Dev-E iterations from running a CLI-default model.	CLI_STARTED / ExecutionLog.Model on resumed runs vs fresh runs	None

Which changes raise per-token spend (these need the operator's approval): - Swapping any claude pod's anthropic-api-key secret value to a real API key. - Adding a LiteLLM proxy or paid key (gitops:docs/whitepaper/cost-framework.md:274). - Anything that makes Rig-E's GPU pod stay up longer, or land on a pricier card. - Wider use of OPENAI_API_KEY for memory embeddings (per-token, optional: gitops:apps/dev-e/rig-agent-helmrelease.yaml:288-297).

Every model or effort change on the existing OAuth and device-auth credentials only consumes flat quota.

5. Surprising findings

1. The decided move to claude-sonnet-5 never shipped. No sonnet-5, opus-5, fable or 2026-09-20 string exists anywhere in rig-gitops main 9b65b3f (Grep, case-insensitive). Nothing is in the runtime pricing (rar:src/usage.js:4-15), the conductor pricing (rc:…/AnthropicPricing.cs:24-65) or claude-3 docs either. The rig is still entirely on 4.x: opus-4-7[1m], sonnet-4-6, and gpt-5.5 for codex. The only 5.x use is the orchestrator's own Mac Mini delegation (c3:docs/infrastructure/hosts.md:101,116). (Review-E already has no [1m].)
2. The fallback chain is wider than fallbackProviders. Every llm.providers key is appended (rar:src/agent/provider-chain.js:37-41). So Dev-E node and Review-E claude fall claude-tmux → claude-cli → codex-cli (gpt-5.5), and the codex hop has no credential in those pods. Any error at all triggers the next hop (fallbackEligible defaults to true, rar:src/agent/provider-error.js:7,68-71). Rig-E's comment that "claude-cli reads character.llm.model directly; llm.providers only orders the chain" (gitops:apps/dashecorp/rig-e-helmrelease.yaml:258-262) is stale: per-provider models are merged (rar:src/agent/provider-chain.js:62-72).
3. A resumed claude-cli run does not pass --model (rar:src/agent/providers/claude-cli.js:189-197). The stream consumer resumes sessions on retries and iterations (rar:src/stream-consumer.js:839-855). Those runs use whatever model the CLI defaults to, not the configured one. This is a silent model change.
4. maxTurns is dead config on every claude-tmux and codex-cli agent: 7 base-fleet HelmReleases (Dev-E node, Review-E claude, and the five codex variants, two of which are off). Only claude-cli enforces it.
5. temperature is dead config on every CLI agent.
6. The dollar figures behind #646 were counterfactual. Every claude pod runs on subscription OAuth. The "$607/7d, Review-E 86%" figure (gitops:docs/2026-06-18-review-e-sonnet-cost-trim.md:13) is API-equivalent. The Opus→Sonnet trim saved quota, not billed dollars. The cost data is also unreliable:
7. It is double-counted for claude-cli and claude-tmux (§2B).
8. tmux cost counts only the pasted prompt and the final reply at 4 chars/token, not the agentic turns in between (rar:src/agent/providers/claude-tmux.js:125-133,555). Dev-E node and Review-E are therefore under-reported by a wide margin.
9. The conductor re-prices unknown models as sonnet-4-5 when cache tokens are present (rc:…/MartenEventStore.cs:446-459).
10. The "Daily Cost Report" calls /api/usage with no days, so it posts all-time totals, not daily ones (gitops:apps/rig-conductor/daily-cost-report-cronjob.yaml:31).
11. Quota-aware Review-E routing is inert. The conductor chooses between review-e and review-e-codex by quota (rc:src/ConductorE.Core/UseCases/QuotaAwareReviewRouter.cs:45-101, rc:src/ConductorE.Api/Services/ReviewDispatchRouter.cs:38-39). The codex pod calls itself review-e, so review-e-codex never heartbeats, and both pods race on one LPOP list.
12. Contradictory billing comments. The codex Review-E comments say the switch was "instead of metered Anthropic API" (gitops:apps/review-e/rig-agent-helmrelease-codex.yaml:21), while the claude HR and docs say OAuth subscription (…rig-agent-helmrelease.yaml:203, gitops:docs/claude-tmux-pilot.md:31). rar:src/agent.js:221-223 also says tmux uses a metered ANTHROPIC_API_KEY, which contradicts claude-tmux.js:9-14.
13. Stale KEDA comments. dev-e-node-codex says it is "pinned to 0 by paused-replicas", but it has none, so it is live (gitops:apps/dev-e/scaledobject-codex.yaml:6,14). review-e-codex says "Idle: codex variant disabled", yet it is the active reviewer at replicaCount: 1 (gitops:apps/review-e/scaledobject-codex.yaml:18-20). The claude Review-E is described as "replicaCount: 0 keeps it idle" (:9-10) but has been 1 since gitops#476 (…rig-agent-helmrelease.yaml:21).
14. The #646 trim did not reach every copy. stigjohnny Review-E is still claude-opus-4-7[1m], and its codex leg falls back to claude-opus-4-7 (dormant today). Planner-codex and Dev-E-codex claude hops are sonnet-4-6 while their claude siblings are opus. iBuild-E's codex hop is gpt-5.4, not 5.5 (ae:character.json:94).
15. iBuild-E is not ibuild-e.json. The live daemon runs rig-agent-runtime from automate-e with character.json (c3:docs/infrastructure/hosts.md:137, c3:docs/infrastructure/agents.md:31). agent-runner/agents/ibuild-e.json has no model and is not what launchd runs.
16. Chart replicaCount: 0 was ignored until #732 (rar:charts/rig-agent-runtime/templates/deployment.yaml:12-16). Earlier "dormant" fleets actually rendered 1.
17. The Dell legacy cluster still reconciles apps/dev-e and apps/review-e (gitops:clusters/dell/kustomization.yaml:6-7), so any model change also lands there, and those pods are Pending.

Docs

No docs changed: this job was read-only by brief, and this report is the only deliverable. Doc updates these findings would need (§5 items 2, 6, 8 and 9) belong to whichever job makes the corresponding change.