deeplossless

A durable execution-state runtime for long AI coding sessions. It preserves exact evidence, avoids repeated execution, and gives agents a rebuildable working state. It sits as an OpenAI-compatible proxy between your client and the DeepSeek API.

cargo install deeplossless
deeplossless --api-key sk-...
# Point any OpenAI-compatible client at https://localhost:8080/v1

Long context windows are useful. Durable execution state is different.

---

Quick Start

# Try without API key — runs a local demo
deeplossless demo

# Proxy mode: set once
export DEEPSEEK_API_KEY=sk-...
deeplossless

# Or let the proxy extract your key from the first request

OpenAI-compatible clients: point base_url to https://localhost:8080/v1.

What It Does

Long coding session (3 tasks, 86 turns)

Vanilla Agent                          DeepLossless Runtime
────────────────────────────────────── ──────────────────────────────────────
21,070 tokens                          13,500 tokens
14 repeated replans                    5 replans
8 repeated failures                    3 failures
11 repo rereads                        9 rereads avoided

                                       ↓36% total tokens
                                       ↓64% replanning
                                       ↓62% repeated failures

Try it yourself — no API key needed:

git clone https://github.com/gordonlu/deeplossless.git && cd deeplossless
cargo test --test long_session_benchmark -- --nocapture
cargo test --test simulated_session -- --nocapture

What Gets Reused

• Repeated deterministic tool calls — exact cached results, invalidated when dependencies change
• File observations — versioned evidence instead of stale conversational recollection
• Failed attempts — preserves what failed and why as evidence, without forcing an old fix
• Plans — persists typed execution state and revalidates dependencies
• Working context — recalls and compacts projections while Ground Truth stays exact

Configuration

Argument	Default	Description
--host	127.0.0.1	Listen address
--port	8080	Listen port
--api-key	DEEPSEEK_API_KEY env	DeepSeek API key
--admin-key	ADMIN_KEY env	Admin key for LCM endpoints
--upstream	https://api.deepseek.com	Upstream API base URL
--db-path	~/.deeplossless/lcm.db	SQLite database path
--rate-limit	100	Max requests/second
--summarizer-model	deepseek-flash	Model for optional background summarization
--dry-run	disabled	Save request bodies, skip upstream
--log-dir	disabled	Per-request JSON logging
--record	disabled	Record raw request/response for protocol debugging
--passthrough	disabled	Pure byte-level passthrough (no pipeline, no context injection)
--tls-cert	auto-generated	Custom TLS certificate (PEM)
--tls-key	auto-generated	Custom TLS private key (PEM)
--lcm-context	off	Enable DAG context injection into system messages
--lcm-context-tokens	1024	Token budget for LCM context injection. Set via body lcm_max_tokens per-request
--no-lcm-context	off	Disable LCM context injection entirely
--no-cache-normalize	off	Disable system prompt date stripping (on by default)
--no-pipeline	off	Skip context injection but still capture reasoning
--no-header-mod	off	Use upstream headers as-is
--dag-threshold	0.80	Compaction trigger (fraction of context window)
--summarizer-budget	1000	Max LLM summarizer calls per session (0=unlimited)
--http-port	8081	Plain HTTP port for sandboxed agents
--workspace	auto-detected	Project path for stable conversation identity across restarts
--audit-mode	full	Audit logging: full, onerror (buffer, flush on failure), off
--snapshot-mode	manual	DAG snapshots: auto, manual (via API), off
--onerror-ring-size	50	OnError audit ring buffer size

TLS is always on. A self-signed certificate is auto-generated at ~/.deeplossless/. Run deeplossless trust once to configure SSL_CERT_FILE so OpenSSL-based tools (Codex, curl, etc.) trust the certificate.

Set via RUNTIME_PROFILE=minimal|efficient|exploratory|autonomous|custom.

Codex + DeepSeek

# 1. Trust the certificate (one-time setup)
deeplossless trust
setx SSL_CERT_FILE "%USERPROFILE%\.deeplossless\cert.pem"
# Restart your terminal after this

# 2. Start the proxy
deeplossless

# 3. Codex config (~/.codex/config.toml)
[model_providers.localproxy]
name = "deeplossless"
base_url = "https://localhost:8080/v1"
wire_api = "responses"    # for Codex Responses API

# 4. Run
codex

DeepSeek Responses requests stay native end-to-end. DeepLossless records exact provider items, derives execution state and working context internally, then forwards native Responses JSON/SSE without translating it through Chat Completions.

Claude Code (Anthropic Messages API)

deeplossless
# Claude Code detects the proxy automatically via SSL_CERT_FILE

# ~/.claude/claude_desktop_config.json or .codex/config.toml
[model_providers.localproxy]
name = "deeplossless"
base_url = "https://localhost:8080/v1"
wire_api = "anthropic"    # Anthropic Messages API format

The proxy translates Anthropic Messages API → DeepSeek Chat Completions, including tool use, content blocks (text/thinking/tool_use/tool_result), and streaming SSE with proper content block lifecycle events. Reasoning content from DeepSeek is exposed as thinking blocks for Claude clients.

LCM API Endpoints

Endpoint	Method	Description
/v1/lcm/current	GET	Current conversation ID
/v1/lcm/grep/{id}	GET	Search past context by query
/v1/lcm/inject	POST	Inject DAG context into last user message
/v1/lcm/chat/completions	POST	Chat completions with automatic context injection
/v1/lcm/sessions/{id}/system-prompt	GET	Deduplicated system prompt history
/v1/lcm/sessions/{id}/events	GET	Execution events for a session
/v1/lcm/sessions/{id}/patches	GET	File observation patches
/v1/lcm/latency	GET	Recent upstream request latency records
/v1/lcm/latency/summary	GET	Aggregated latency statistics (p50/p95/p99)
/v1/lcm/cache/stability	GET	Cache stability diagnostics
/v1/lcm/status/{conv_id}	GET	DAG health and execution status
/v1/lcm/runtime/report	GET	Session performance report

Context injection merges into the last user message (not pushed as a new message), preserving message sequence and avoiding agent state machine disruption.

Session Report

curl -sk https://localhost:8080/v1/lcm/runtime/report?label=fix+build

# deeplossless session report: fix build
50 turns · 180s duration · 42% cache reuse
21 exact cache reuses · 3 recorded failure loops

Runtime Profiles

Profile	Cache	Retries	Context	Budget	Best for
Minimal	100%	1	20%	30%	Budget-conscious
Efficient	80%	2	50%	60%	Daily coding
Exploratory	50%	3	80%	80%	Debugging
Autonomous	30%	5	100%	95%	Complex tasks

WebUI

Execution forensics viewer — see what the AI actually did.

git clone https://github.com/gordonlu/deeplossless-ui.git
cd deeplossless-ui && npm install && npm run dev

Opens at http://localhost:3000 — connects to a running deeplossless instance.

Tech Docs

• Technical Reference — architecture, API reference, protocol tests, runtime strategies
• Architecture Documents — invariants, lifecycle, replay model, dependency authority
• Runtime Event Schema — frozen event contract
• Dependency Model — dependency taxonomy and authority boundaries

Requirements

• Rust 1.85+
• DeepSeek API key (for proxy mode; benchmarks run without)

License

MIT