@gtrabanco/pi-nan-agentic-workflow

The optional NaN.builders preset for @gtrabanco/pi-agentic-workflow — the provider profile, its four cost-posture ladders, and the quota-tier routing policy, in one small independently versioned package (issue #388).

The main pi package ships provider-neutral: with only it installed, recommendedModels: true is a documented no-op and every command routes on your session's model. Installing this companion next to it — with no config change — restores NaN-aware routing: the built-in nan profiles appear in the resolution chain (listed in /agentic-workflow-settings, protected by the built-in-not-editable rule), and the ladders below apply.

Flexible, but opinionated by default. Every choice on this page is a config key you can override. With nothing configured, the preset routes everything through the economy posture: the ladder tuned to deliver a reasonable quality on most tasks while living almost entirely inside the free monthly allowances. If you want something else, say so once in config — the nearest instruction wins, and nothing in the preset fights you. One distinction worth reading before you pick: the models are not all processed in the same place — see Data residency for what runs on the EU cluster and the manual overrides that keep each posture EU-compliant.

Install

# From a Pi session (the main package is a prerequisite):
pi install @gtrabanco/pi-nan-agentic-workflow

The package is pure data plus one registration entry: on load it registers the preset into the main package's provider-preset seam. Nothing else changes.

Choosing a posture

The preset ships four named profiles — four answers to "how much should model quality cost me?". Select one (or a fallback list) with the generic presetProfile config key, which the main package documents:

{ "presetProfile": "economy" }

An ordered array is the interesting one: first entry = most preferred. When a profile's route is unusable, the next entry is probed — so a list rides down the catalog as capacity runs out:

{ "presetProfile": ["quality", "balanced", "economy", "no-credits-continue-working"] }

• Profile names are quality, balanced, economy, no-credits-continue-working; the bare chain entry nan always resolves the preset's default (economy), and each named entry travels as nan.<profile> in the chain.
• Unknown names are skipped; if a selection names nothing the catalog carries, the default posture is used — a typo never silently drops NaN out of the chain.
• Your own profiles win. Declare a profile under the same chain name in profiles (e.g. "nan.economy") and your definition replaces the preset's entirely; add it to profileOrder and it takes precedence like any user profile (feature 63). Want a hand-tuned ladder? It composes.
• tierRouting still applies (see below): it is the opt-out for the preset's tier policy, independent of which posture is selected.
• tierRouting: false restores the selected profile's exact published ladder — no silent quality regression for users who disable routing.

When you are out of credits (balance zero, quotas drained), the no-credits-continue-working posture keeps the workflow alive on the models that never consume credits: the two unlimited cluster models, plus the flash tier while its monthly allowances hold — because an unused free allowance is yours whether or not you have credits. The honest options when that posture's verdicts also thin out:

1. Buy the smallest pack — 1,000 credits for €24.70 is roughly 1.6–1.9 billion tokens on the flash tier; a subscription month (7,000 credits) is more than the flash allowances can burn through in typical agentic work.
2. Wait for the billing reset — every quota resets with each billing period; the unlimited cluster models are not affected by the wait.
3. Keep working in no-credits-continue-working meanwhile — judgment-heavy verdicts defer to a human, mechanical work continues uninterrupted.

Plans, credits, and what actually costs money

All chatting models draw from one pool of credits (NaN Cloud, sign-up via this referral link; full catalog: nan.builders/docs/models). The billing model, in one paragraph:

• Subscriptions renew each billing period: €70 → 7,000 credits/month, €200 → 20,000 credits/month. Credit packs are one-off purchases — 1,000 cr (€24.70), 5,000 cr (€116.85), 25,000 cr (€550.05), the 5% member discount already applied; custom top-ups €5–1,000. Pack credits do not renew; all credits expire 6 months after purchase. Auto-recharge is off by default.
• Free allowances reset each billing period and are consumed before a single credit is touched. Combined they give the flash tier roughly 6.5 billion tokens/month for free on the flash ladders — which is why the default posture rarely burns credits at all.
• Credits pay for: any flash-tier overflow beyond an allowance, the premium glm5.3, and the four paid "larger" models from token one.
• Rate limits per key: 60 requests/min; concurrency 7 simultaneous (base plan) / 10 (premium) across all models, with per-model caps of 7/10 (glm5.3, glm5.3-flash, deepseek-v4-flash, qwen3.8-flash) and 5 (mimo-v2.6-flash, qwen3.6, gemma4); 1.5M tokens/min for the latter tier of models where a tpm cap applies. glm5.3 additionally caps at 400M tokens per rolling 4-hour window — a heavy coding-agent run reaches that first, so the ladders below keep premium rungs for judgment slots and never for volume.

Where the numbers come from. Every tariff, allowance and limit quoted on this page restates the recorded snapshot at catalog/nan-catalog-2026-11.jsonc — source URLs, capture date and the docs-vs-quote winner rule in its provenance header. When the provider changes anything (prices, allowances, the catalog), update that snapshot first; test/catalog-parity.test.mjs then fails on every tier map, ladder and README row that disagrees, walking you through the rest.

The fleet

Tariffs are credits per 1M tokens after the free allowance (November 2026 tariff table). typical is the provider's published blended figure — treat it as an order of magnitude, not a quote.

Tool calling is OpenAI-tools-native only on glm5.3, glm5.3-flash, deepseek-v4-flash and mimo-v2.6-flash. qwen3.6, qwen3.8-flash and gemma4 document tool calling in XML format — fine for single-shot text work, not for the executor path's OpenAI-tools harness unless the harness is XML-aware. The executor rungs therefore stay on the four function-calling models. Run the tool-calling smoke test in docs/workflow/GOLDEN_FIXTURE.md before promoting any model into that path.

Reasoning control per model (per the API reference — verified against the November docs, superseding the older "cannot be turned off" notes):

Reasoning traces arrive in message.reasoning_content. Cap any agentic loop's max_tokens with headroom, stream long outputs, and note that a reasoning-only stream is closed by the platform (a finish_reason: "length" on an empty delta, not billed) if a model plans without converging.

Data residency — what runs where

Not every model in the tables above processes under the same legal roof. The chat models split in two, and the split is the one marked in the catalog:

• Runs on the EU cluster. glm5.3, glm5.3-flash, deepseek-v4-flash, qwen3.8-flash, qwen3.6 and gemma4 are served from the operator's own infrastructure — Helmcode S.L., Spain (platform policy supervised by the Spanish DPA, the AEPD), all of them ZDR: prompts and completions are not stored.
• External provider — do not assume EU processing. mimo-v2.6-flash, kimi-k3.1-pro, qwen3.8-max, deepseek-v4-pro and hy4 are served by third-party providers. The catalog documents zero data retention for them, but it does not name the third parties or say where their GPUs sit. If EU processing is crucial for you (a regulatory duty, a data-processing agreement, a sector rule), treat these five as not EU-compliant until you verify the serving route yourself — and therefore as rungs to strip out of the ladders.

The preset ships no separate EU profiles — the flexible way to get them is the mechanism that already exists: a user-defined profile shadowing a preset chain name wins over the preset entirely. Paste the blocks below into your pi-agentic-workflow.json and every posture stays EU-only, whatever presetProfile selects. Keep tierRouting exactly as you had it; it keeps routing the rest of the chain (the default profile and anything not shadowed).

{
  "profiles": {
    "nan.quality": {
      "default": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
      "commands": {
        "audit-pr": { "model": ["nan/glm5.3", "nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "high" },
        "product-audit": { "model": ["nan/glm5.3", "nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "high" },
        "unit-lane": { "model": ["nan/glm5.3", "nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "high" },
        "review-change": { "model": ["nan/glm5.3-flash", "nan/glm5.3", "nan/deepseek-v4-flash"], "thinking": "high" },
        "triage-issue": { "model": ["nan/glm5.3-flash", "nan/glm5.3", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "init-workspace": { "model": ["nan/glm5.3-flash", "nan/glm5.3", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "advance": { "model": ["nan/glm5.3-flash", "nan/glm5.3", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "execute-phase": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "fold-findings": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "discover-repository-state": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "resolve-repository-state": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "audit-docs": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" },
        "log-session": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" },
        "workflow-status": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" }
      }
    },
    "nan.balanced": {
      "default": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash"], "thinking": "inherit" },
      "commands": {
        "audit-pr": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash", "nan/glm5.3"], "thinking": "high" },
        "product-audit": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash", "nan/glm5.3"], "thinking": "high" },
        "unit-lane": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash", "nan/glm5.3"], "thinking": "high" },
        "review-change": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "high" },
        "triage-issue": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "init-workspace": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "advance": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "execute-phase": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash"], "thinking": "inherit" },
        "fold-findings": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash"], "thinking": "inherit" },
        "discover-repository-state": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash"], "thinking": "inherit" },
        "resolve-repository-state": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash"], "thinking": "inherit" },
        "audit-docs": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" },
        "log-session": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" },
        "workflow-status": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" }
      }
    },
    "nan.economy": {
      "default": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
      "commands": {
        "audit-pr": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "high" },
        "product-audit": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "high" },
        "unit-lane": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "high" },
        "review-change": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "high" },
        "triage-issue": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "init-workspace": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "advance": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "inherit" },
        "audit-docs": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" },
        "log-session": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" },
        "workflow-status": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" }
      }
    },
    "nan.no-credits-continue-working": {
      "default": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash", "nan/qwen3.8-flash", "nan/qwen3.6"], "thinking": "inherit" },
      "commands": {
        "audit-pr": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash", "nan/qwen3.6"], "thinking": "high" },
        "product-audit": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash", "nan/qwen3.6"], "thinking": "high" },
        "unit-lane": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash"], "thinking": "high" },
        "triage-issue": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash", "nan/qwen3.8-flash"], "thinking": "inherit" },
        "init-workspace": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash", "nan/qwen3.8-flash"], "thinking": "inherit" },
        "advance": { "model": ["nan/glm5.3-flash", "nan/deepseek-v4-flash", "nan/qwen3.8-flash"], "thinking": "inherit" },
        "execute-phase": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash", "nan/qwen3.8-flash", "nan/qwen3.6"], "thinking": "inherit" },
        "fold-findings": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash", "nan/qwen3.8-flash", "nan/qwen3.6"], "thinking": "inherit" },
        "discover-repository-state": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash", "nan/qwen3.8-flash", "nan/qwen3.6"], "thinking": "inherit" },
        "resolve-repository-state": { "model": ["nan/deepseek-v4-flash", "nan/glm5.3-flash", "nan/qwen3.8-flash", "nan/qwen3.6"], "thinking": "inherit" },
        "audit-docs": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" },
        "log-session": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" },
        "workflow-status": { "model": ["nan/qwen3.6", "nan/gemma4"], "thinking": "inherit" }
      }
    }
  }
}

The blocks above are deliberately complete: a user-defined profile replaces the preset's derivation wholesale — including tier routing, which no longer applies to a shadowed chain name — so every command is spelled out, with the mechanical rows pinned to the zero-credit EU chain by hand. Three boundaries to know:

• Spawned review passes are separate. The per-pass model declarations (passes in the config, issue #201) route review-change's spawned passes independently of the command ladder — if you pin EU routes for commands, pin the passes you actually use the same way.
• Spawned task classes are not user-overridable today. A low-stakes class (summarization, triage, mechanical-logging, catalog-step, conduction) already resolves to the zero-credit EU chain under tier routing — EU-safe on every posture. A strength-sensitive class (review-axis, audit-verdict, planning) resolves from the selected posture's published default ladder, which still contains external-provider rungs in quality and balanced (the mimo-v2.6-flash design rung) — that route is not reachable by either knob above. If EU-only spawned planning matters for you, it is a real gap: say so and it becomes a unit.
• ZDR is not residency. All ten models document zero data retention, which bounds what is kept — it says nothing about where the GPUs processing your prompts sit. The two claims live in different sections of the catalog on purpose.

The preferred ladders

The four postures side by side, per task class. Rungs are ordered: the router probes down the chain and the first usable rung serves the command; profiles differ in how eagerly they spend credits and in how far a judgment rung is allowed to stretch.

Why these picks (independent evidence first, vendor claims flagged):

• glm5.3-flash is the fleet's best credit-per-quality workhorse with independent verification: it matches glm5.3 on Terminal-Bench 2.1 (84.3 vs 84-ish) and on real-world agentic work (GDPval-AA Elo within the margin of error), at roughly a tenth of the cost per task. Cache-heavy agent loops (an executor is exactly that) reach the cluster minimum, ~4.4 credits/1M at a 60% cache hit rate; MIT-licensed, NaN-hosted GPUs.
• kimi-k3.1-pro posts the strongest independently measured agentic coding numbers in the fleet (67% Terminal-Bench 4.0 per Artificial Analysis; GDPval-AA 1,846 Elo, third overall) — which is why it leads the quality posture's executor row despite being the most expensive model here (500 credits/1M output, plus the fleet's only cache-write fee). Cost-no-object is the posture's contract.
• glm5.3 remains the long-horizon specialist (open-source SOTA on Terminal-Bench 3.0, the fleet's best CyberGym score) and needs the premium plan; its always-on reasoning and 400M/4h rolling window make it a judgment rung, never a volume rung.
• deepseek-v4-pro's vendor claims are the least independently replicated of the large tier: NIST measured 74.0 on SWE-bench Verified against the claimed 80.6, an independent Terminal-Bench run scored 54.68 against the claimed 87.9, and its Artificial Analysis Intelligence Index sits one point above deepseek-v4-flash. It stays in the ladders only as an escalation rung (balanced) — never as a recommendation to start from, and never in the cost-minimal posture.
• mimo-v2.6-flash is the cheapest model in real blended use (0.6 typical, ~1,666M tokens per 1,000 credits, its reasoning is now fully switchable) and a different model family from the executors, which is what makes it valuable as the independent reviewer rung. Its caveat: it is served as an external provider (not NaN's own GPUs), caps at 5 concurrent requests, and independent agent-suite results are mixed where vendor ones are strong.
• qwen3.8-max: an elite exam-style reasoner (GPQA 92+) that is not a top-tier coding agent (DeepSWE 56.6, mid-pack agentic rank). At 20.2 credits/1M it wins no task class in this workflow, so no ladder routes to it; documented here so its absence is a decision, not an oversight.
• hy4: the cheapest of the large tier (181M tokens per 1,000 credits) and the long-document specialist, but with only vendor-run evidence at this time and a 64K output cap. Treat it as a candidate for long-document passes pending a golden-fixture smoke run, not a default rung.
• qwen3.6 and gemma4 cost no credits under any plan and never burn out — they are the zero-credit floor every ladder's mechanical row stands on (qwen3.6's OpenAI-tools smoke pass is recorded in GOLDEN_FIXTURE.md; set reasoning off with none so a log line pays zero thinking tokens). They are previous generation (the provider itself recommends deepseek-v4-flash to start from), so they stay off the executor layer except as the last rung in the no-credits posture, where their output is only ever re-checked reasoning.

unit-lane spans two rows — pick by the strongest step its triage returns, not by the skill's name. The lane runs its judgment steps (triage, design, plan, evidence, review, close-out) in one session, so the model must cover the highest-class step it will reach; execution phases hand off to execute-phase one turn each at the phase's tier. A lane whose triage includes design reads the Product definition row; a lane of triage → plan with no design reads Planning / triage.

Folding review-change/audit-pr findings routes through the standalone fold-findings skill (falling back to execute-phase's embedded fold cycle where a separate invocation isn't available), and follows the row rule of thumb: a routine/mechanical finding fixes on the Execution row; a subtle finding (logic, security, architecture) fixes on the ladder of the review that found it. The fixing model is never weaker than the one that wrote the original code — otherwise the fix itself needs re-catching on re-review.

Adversarial review (review-change --adversarial N / --synthesize): reviewers never weaker than the model that authored the diff, and never the author's own family. The fleet carries five families across the four postures (GLM, DeepSeek, Moonshot, Xiaomi, Qwen), so decorrelated reviewer pairs are a routing decision, not a purchase: e.g. under Balanced, a deepseek-v4-flash-authored change reviews on mimo-v2.6-flash + glm5.3-flash; under Quality, add glm5.3 or kimi-k3.1-pro as a third. Fan-out caps at ≤3–4 concurrent (mimo caps at 5 by itself; an agentic loop spends one request per tool round-trip, so parallel agents hit 60 rpm quickly).

Tier routing (amendment 1, re-cut for the tariff model)

The November tariffs changed what "saving" means: the flash tier's free allowances are large enough that executor-class work on them is effectively free, while the unlimited cluster models are too small and too XML-tool-bound for the executor path. The tier policy therefore demotes only the mechanical, one-way passes:

• Demoted under routing-on (default): log-session, workflow-status, audit-docs — and spawned task classes classified non-strength-sensitive (summarization, triage, mechanical-logging, catalog-step, conduction). These ride the zero-credit chain: nan/qwen3.6 first, then nan/gemma4 — models that cost no credits under any plan and never run out.
• Everything else runs its profile's ladder verbatim — judgment slots (audit-pr, product-audit, review-change, unit-lane) and executor phases included; their rungs live inside the free allowances by design.
• Opt-out, two levels. Global: "tierRouting": false (or {"enabled": false}). Per skill: {"skills": {"log-session": false}}. With the opt-out applied, the affected routes match the profile's published ladder exactly.
• Tier changes are one-line data edits. The billing map (NAN_MODELS) carries the tariff facts; when the provider changes allowances or prices, edit it and re-cut the ladders beside it.

Example tierRouting in agentic-workflow.json:

{ "tierRouting": { "enabled": true, "skills": { "audit-docs": false } } }

Sources — where these numbers come from

Every claim on this page traces to one of the surfaces below. Consult the source, not this README, when a decision matters; the snapshot file (catalog/nan-catalog-2026-11.jsonc) records what this page was built from and when.

First-party (the provider)

Independent evaluators (the evidence the picks are grounded on)

Vendor-run surfaces (cited only with that label)

The reading rule

Independently verified beats vendor-reported; vendor-reported beats vibes — and an empty third-party section is a finding, not a zero. Before pinning any rung, re-check the current leaderboard: this landscape moves monthly, and the community-voted cluster changes quarterly (qwen3.6 unlimited is "for now", the same as ever).

Verify your catalog first

Run GET /v1/models with your own key and route only to models it actually returns; the exact set depends on your plan (premium for glm5.3) and the current catalog. The community-voted cluster changes quarterly.

Whisper, Kokoro, Rerank, Qwen3 Embedding and Flux 2 Klein are NaN's hosted audio/retrieval/image models — this workflow never routes to them. Model strength above is framed by benchmark evidence with provenance named (independent where it exists, vendor-run marked as such) — re-check against a current leaderboard before pinning; this landscape moves fast.