Cost model

Wyren’s distiller runs continuously in the background. Naively, that’s wasteful. Wyren uses a three-tier pipeline to keep cost minimal — or zero under the preferred execution path.

Naive baseline (what we DON’T do)

Without tiering, every Stop hook triggers a Sonnet 4.6 call:

• Per call: ~10k input tokens (transcript slice + memory + prompt) + ~2k output.
• Sonnet 4.6: $3/M input + $15/M output → ~$0.06 per call.
• 3 teammates × 4 active hours × 12 triggers/hr = 144 calls = ~$8.60 total.

Cheap in absolute terms, but most turns contribute nothing to memory (tool-use loops, file reads, small tweaks). The fix is tiering.

Tier 0 — Local scoring filter (FREE)

Before any API call, lib/filter.mjs scores the transcript slice. Scoring runs entirely in Node — no API call, no cost.

Text signals (weighted, capped per category):

Category	Weight	Examples
Decision language	3	decided, we're going with, chose, agreed
Rejection / failure	3	rejected, doesn't work, tried X but, abandoned
Deliberate hacks	3	workaround, hack, hardcoded, stub, mock
Scope signals	2	out of scope, deferred, descoped, dropping
Open questions	2	open question, still deciding, TBD, revisit
Maintenance flags	2	TODO, FIXME, before launch
Weak signals	1	actually, instead, broken, for now

Edit tool calls (Edit / Write / MultiEdit) are scored separately at weight 3 per call, capped at weight × 4.

Structural signals (scored on raw JSONL lines, not rendered text):

• Session length ≥ 10 turns: +2; ≥ 20 turns: +4 total
• Average user message length > 200 chars: +2 (explains context or decisions)
• File edits ≥ 3: +2; ≥ 8: +4 total

Rule: total score must reach the threshold (default 3, overridable via WYREN_TIER0_THRESHOLD) to proceed. A single high-value signal word is enough; weak words alone need reinforcement. ~60–70% of triggers die here for free.

Tier 1 — Claude Haiku 4.5 (routine distillation)

The remaining 30–40% of triggers go to Haiku 4.5 (claude-haiku-4-5-20251001):

• Pricing: $1/M input + $5/M output → **$0.02 per call**.
• Latency: ~2 s P50 — background completes before the next trigger.
• Quality: sufficient for incremental updates. The task is merging a small delta into a well-structured existing file, not generating from scratch.

This is the only model the Stop hook ever spawns automatically.

Tier 2 — Claude Sonnet 4.6 (deep cleanup)

Sonnet is designed for a slower cadence to correct Haiku drift over long sessions:

• Once per hour after any Tier 1 runs
• When memory.md exceeds 60 lines → forced re-compression
• On explicit /wyren-handoff

Not yet automated. The standalone distiller.mjs accepts --model <id>, so you can invoke Sonnet manually with wyren distill --model claude-sonnet-4-6 --push — but the Stop hook always uses Haiku. No timer or line-count trigger exists yet. /wyren-handoff runs the normal distiller (Haiku) with --push.

Haiku drift on long sessions is a real concern — Tier 2 automation is planned follow-up work, not a current capability. See Future.

Prompt caching (SDK path only)

When using the Anthropic SDK (not claude -p):

• System prompt → 5 min cache, ~$0 on repeat reads.
• Current memory.md → 5 min cache, ~$0 on repeat reads.
• Only the transcript slice varies per call — small incremental cost.

Effective cached per-call cost: ~$0.003 on Haiku, ~$0.008 on Sonnet.

Revised total — what actually ships

With Tier 0 + Tier 1 (no Tier 2 automation):

• 144 raw triggers (3 teammates × 4 active hours × 12/hr).
• Tier 0 kills ~70% → ~43 triggers reach the API.
• All 43 go to Haiku. Under claude -p these draw from each teammate’s Claude Code quota — no separate billing.
• If the SDK path is used instead with prompt caching: 43 × $0.003 = ~$0.13 total for the session.

Compared to the naive baseline of $8.60 — roughly 65× cheaper, and $0 under the preferred claude -p path.

The zero-billing path

When using claude -p headless (the default), distillation draws from each teammate’s existing Claude Code quota — no separate API billing surface at all.

The tiered SDK path only matters for cost-conscious deployments where teammates do not have Claude Code subscriptions.

Free-only fallback

If a teammate has no Claude Code auth AND no API key (edge case), Tier 0 scoring alone produces a degraded but non-zero memory:

• Bullet-list every signal-word match with surrounding context.
• No semantic cleanup — verbose and noisy.
• Still better than nothing on a team with one paid member distilling on everyone else’s behalf.

Which tiers ship?

Tier	Status
Tier 0 (weighted scoring filter)	✅ Shipped in lib/filter.mjs. Mandatory — runs before every API call. Kills ~70% of triggers.
Tier 1 (Haiku 4.5)	✅ Shipped as default in distiller.mjs. Invoked via claude -p --model claude-haiku-4-5-20251001.
Tier 2 (Sonnet on timer / size threshold)	❌ Not automated. distiller.mjs --model <id> accepts Sonnet manually; no automatic trigger.
SDK prompt caching	❌ Not shipped. Current path uses claude -p exclusively.