telos-sdk 0.1.2

TELOS — cache-friendly prompt protocol: stable prefix, tiered bands, ephemeral tail, layered adapters, anchored marks.

A portable, cache-friendly protocol for LLM agent context.

_{One canonical IR — your tools, system, turns, and memory — runs unchanged across Anthropic, OpenAI, DeepSeek, vLLM, and SGLang, with KV-cache hits preserved across turns and cost reported in absolute dollars.}

Quickstart · Engines · Why · Three things · Protocol · User Guide

_{📖  English · Simplified Chinese}

---

⬢  30-second start

from telos import Bridge, load_engine, load_harness

harness = load_harness("openclaw")          # or "hermes"
engine  = load_engine("anthropic")          # or "openai" / "deepseek"

ir = harness.parse(raw_request, session_id="task-001",
                   engine="anthropic", model="claude-opus-4-7",
                   expected_turns=20)

bridge = Bridge(ir, engine)
plan   = bridge.mark()        # let the engine decide BP / routing-key
wire   = bridge.emit()        # get the wire request, ready to send

response = call_llm(wire)     # your own HTTP client
report   = bridge.absorb_usage(response)
print(report.cache_read, report.raw_input)

Full end-to-end run: telos/demo.py — python -m telos.demo.

---

⬢  What it looks like

_{Every call's normalized usage lands in a jsonl log, aggregated into a single-file HTML dashboard.
It counts absolute dollars saved — not ratios you can game by shrinking the denominator.}

---

⬢  Engines — one IR, five backends

TELOS is normative: it defines how context should be represented, and engines align by capability. One TelosIR landing on different engines is degraded deterministically by the adapter — never silently, never lossily in meaning.

Capability	Anthropic 4.6+	OpenAI 4+/5.x	DeepSeek V3+	vLLM	SGLang
explicit BP / anchors	✓ (≤4)	✗	✗	✓	✓
explicit prewarm	✓	✗	✗	✓	✓
routing key	✗	prompt_cache_key	✗	cache_salt	affinity_key
cache probe / segment evict	✗	✗	✗	✓	✓
fork-and-replace	✗	✗	✗	partial	✓

Bidirectional capability (BidirectionalEngineAdapter, only on open-source inference engines): cooperative_fold() lets the server keep the prefix KV untouched and recompute only the summary tail — a closed API's fold is a client-side rewrite that forces the server to re-prefill the whole span every time. Full matrix in protocol §6.

---

⬢  Why TELOS exists — stop being a tenant in someone else's agent

Every team putting agents into production hits the same four walls.
TELOS is the single answer to all four — one canonical representation of agent context.

🔒  Context no longer locked to one vendor "Switch models to run the same task, and you start from scratch." TelosIR is an engine-agnostic, serializable, portable representation of context. An Anthropic session moves verbatim to DeepSeek, to your own vLLM — adapters degrade deterministically, without losing meaning. 📁 telos/ir.py · telos/engine/	💸  Stop paying twice for the same prefix "Twenty turns in, every turn re-prefills the identical prefix." Three-color bands — PIN · FOLD · DROP — plus one ordering invariant keep the "base" resident in the KV cache. Memory is pulled on demand, not stuffed whole into every prompt. 📁 telos/bridge.py · telos/refpool.py
🧾  Cost you can see, counted in absolute $ "All you get is a ratio, diluted by whatever denominator." Every call's normalized usage lands in a jsonl log, aggregated into a single-file HTML dashboard. It counts absolutes: cache_read, cost saved — ratios can be gamed by shrinking the denominator; absolute $ cannot. 📁 scripts/build_savings_dashboard.py	🎛  You hold the controller "Handing a task to the agent that's better at it — can't be done." Because context is portable, you genuinely hold the controller: one task dispatched across harnesses, taking the best of each. TELOS provides mechanism, never policy — it never decides for you, never burns an LLM call on routing. 📁 telos/harness/

In one sentence: TELOS is the canonical representation of the only durable asset
in the agent stack — context. Your context; harnesses are just hired help.

---

⬢  One representation, three things

TELOS — Greek τέλος, "purpose, end"; also read as "stone tablet." The inscription on a tablet's base is carved once and lasts a lifetime; the lines added turn by turn above can be wiped anytime, but never touch the base. And the tablet means a second thing — the tablet is yours: it can be carried to any scribe (harness), any printing house (engine).

③ Sovereignty   You hold the controller — any task, hire any harness, any model, no cage
                      ▲  made possible only by
① Portability   Context / memory is an engine-agnostic, serializable, portable asset
                      ▼  and the same representation also delivers
② Efficiency    Extreme KV-cache hits + on-demand memory; cost in absolute $, visible

③ is the purpose, ① the mechanism, ② the payoff and the wedge. Not parallel — a stack. The iron rule: TELOS provides mechanism, never policy — the moment it decides for you, it has taken the controller back.

---

⬢  Architecture

agent harness ──► TELOS Bridge ──► engine adapter ──► LLM service
   (parse)          (5 primitives)   (capability-aware)

Layer	Files	Responsibility
harness	harness/openclaw.py hermes.py	split envelope, large docs into ref-pool, produce TelosIR
bridge	bridge.py ir.py refpool.py	5 primitives, invariant checks, frozen ref-pool slugs, canonicalize
engine	engine/anthropic.py openai.py deepseek.py	capability-aware Mark, wire serialization, usage parsing

The bridge is pure Python with no LLM SDK dependency. TelosIR is the single data structure that passes between all three layers — frozen, narrow-fielded, engine-agnostic.

---

⬢  One invariant

The whole protocol has exactly one hard constraint. Within each segment (tools / system / a single message), blocks must be in physical order:

PIN*  →  FOLD*  →  DROP*

_{(In a message, tool_result blocks always come first — required by the Anthropic protocol.)}

Band	Meaning	Typical content
PIN	long-lived stable segment	tool definitions, system prompt, the user's current question
FOLD	cacheable but droppable on compact	assistant replies, tool_result, large ref-pool docs
DROP	never enters the cache hash	timestamp, cwd, git status, envelope

Violate it and TelosInvariantError is raised. Everything else is a soft suggestion.

Five primitives  _{(Bridge methods)}

Primitive	Purpose
place(segment, blocks)	put blocks into tools / system / the current message
pin(slug, payload)	write a PIN block into the system segment
mark()	let the engine produce this turn's BP / routing-key plan
fold(slugs= / message_range=, summary=)	fold old turns into ref-pool references
refresh(plan)	once throttling allows, send a max_tokens=0 prewarm (Anthropic only)

ref-pool — a "pointer table" for context

A slug is frozen the moment register() is called: content can change (fold()), the slug cannot. fold() swaps the payload, not the slug → every [ref:slug] reference stays byte-identical → BPs still hit after a fold. This is "portable context" realized in the protocol: stable pointers, flowing content.

---

⬢  Cost you can see · savings dashboard

Every TELOS entry point (gateway / SDK transport) appends each call's normalized usage to a usage_log jsonl, aggregated into a single-file HTML page (zero JS, opens offline):

# one-line install
pip install telos-sdk          # or: brew install telos-sdk (see packaging/)

# auto-detect harnesses, inject config, start the gateway
telos init

# open the live dashboard in your browser
telos dashboard

The dashboard counts absolutes: cumulative cache_read, cost saved = cache_read × (input_price − cache_read_price), token mix, broken down across harness / model / session.

---

⬢  Appendix: R1–R8 protocol-hazard fixes

Review surfaced 8 design hazards in the protocol; the Python implementation fixes all of them:

ID	Problem	Fix location
R1	OpenAI prompt_cache_key only widens slots at ≥15 RPM per key	engine/openai.py :: KEY_RPM_SOFT_CAP = 12 + shard()
R2	Anthropic's 4 BPs cover only head + tail, leaving mid turns uncached	engine/anthropic.py :: _MID_ANCHOR_STRIDE = 19
R3	sub-agent IR and parent IR sharing a session_id	harness/hermes.py — sub-IR parsed independently
R4	after fold(), a Mark slot can land in a folded span	bridge.py :: fold() — re-run mark()
R5	tool field / array order not stably canonicalized	bridge.py :: _canonicalize_ir()
R6	thinking blocks lost across non-tool_result calls	engine/base.py :: thinking_preserved_across_non_tool_result
R7	no explicit priority when Anthropic BP candidates > 4	engine/anthropic.py :: plan_marks priority + truncation
R8	refresh unthrottled, can saturate quota in reverse	bridge.py :: REFRESH_THRESHOLD = 11 adaptive gate

---

⬢  Going deeper

What you want	Where
Get started (install, integration, CLI, troubleshooting)	docs/User-guide.md
Understand the protocol	docs/2026-05-06-telos-protocol.md
See the architecture	docs/ARCHITECTURE.md
See the change history	CHANGELOG.md

---

⬢  License

Apache-2.0 — the protocol core is open source, forever. See LICENSE.