Reference · generated from sources.json
Sources and evidence
Open-source implementation claims point to pinned commits. Product comparisons point to official documentation. Interpretations are labeled, and unknown internals stay unknown.
§0
How to read the labels
SOURCE is pinned repository code or documentation. DOCUMENTED is a current first-party product page. OBSERVED requires a reproducible public interaction. INTERPRETATION is the primer’s design reading. UNKNOWN marks a boundary public evidence does not resolve.
The user-supplied article “Every agent harness runs into the same limit…” served as a question and claim map for context-management topics. It is not used as authority for open-source implementation details or reverse-engineered closed-source internals.
§1
DeepSeek Harness · pinned source
DeepSeek Harness README
DeepSeek Harness Architecture
Agent Turn And Step Lifecycle
Tool Execution Pipeline
DeepSeek concrete agent loop
DeepSeek turn, step, and request construction
DeepSeek ToolRuntime
DeepSeek capability seams
DeepSeek session subsystem
DeepSeek compaction subsystem
DeepSeek spill storage subsystem
DeepSeek session persistence subsystem
DeepSeek skills subsystem
DeepSeek subagent subsystem
DeepSeek dynamic extensions subsystem
§2
Pi · pinned source
Pi Agent Harness README
Pi Agent Core README
Pi agent loop
Pi Agent prompt and continue entry points
Pi Agent Core stream-function seam
Pi coding-agent provider composition
Pi coding-agent read tool
Pi coding-agent truncation utilities
Pi coding-agent compaction
Pi coding-agent AgentSession
Pi coding-agent session manager
Pi coding-agent extensions
Pi coding-agent skills
Pi subagent extension example
Pi containerization patterns
§3
Codex · official documentation
Codex CLI
Agent approvals and security
Custom instructions with AGENTS.md
Codex subagents
Build skills for Codex
§4
Claude Code · official documentation
Claude Code tools reference
Claude Code permissions
Claude Code custom subagents
Claude Code skills
Claude Code context window
Claude Code hooks guide
§5
Evidence method
Symphony Service Specification
Effective harnesses for long-running agents
Harness design for long-running application development
Evaluating the impact of LSP-based code intelligence on coding agents
Public comparison evidence boundary
§6
Claim registry
| Claim ID | Statement | Class | Sources |
|---|---|---|---|
general-harness-boundary |
A harness supplies control flow, tool mediation, context and state management, and policy around model inference. | INTERPRETATION | dsh-architecture · pi-agent-readme |
general-turn-loop |
A tool-using agent turn alternates harness-controlled model calls with validated tool execution and recorded observations until a stop condition is reached. | INTERPRETATION | dsh-lifecycle · pi-agent-loop |
general-control-spectrum |
Agent behavior combines deterministic mechanisms, configurable policy, uncertain environmental effects, and probabilistic model proposals. | INTERPRETATION | dsh-architecture · pi-agent-readme |
general-common-anatomy |
The model adapter, context builder, loop, tool runtime, policy layer, session state, and extension surface are recurring harness subsystem roles even when package boundaries differ. | INTERPRETATION | dsh-architecture · pi-agent-readme |
general-kernel-control-plane |
A small request-cycle kernel can sit inside a larger control plane that owns context, policy, durable state, coordination, isolation, observability, and cross-run recovery. | INTERPRETATION | dsh-architecture · pi-agent-loop · openai-symphony-spec |
general-completion-verification |
Completion verification should compare the finished artifact with explicit acceptance criteria and produce independent evidence rather than rely on the implementing agent's declaration. | INTERPRETATION | anthropic-effective-long-running · anthropic-long-running-design |
general-tool-evaluation |
Tool opportunity and presentation, model adoption, downstream task outcome, and operating cost are separate evaluation questions. | INTERPRETATION | nuanced-lsp-evaluation |
general-fresh-process-recovery |
Long-running work benefits from durable task, progress, evidence, and continuation artifacts that let a fresh process recover without relying on one lossy summary. | INTERPRETATION | anthropic-effective-long-running · openai-symphony-spec |
general-harness-assumption-half-life |
Harness components can encode assumptions about model weaknesses, so their value should be retested as models improve rather than treated as permanent architecture. | DOCUMENTED | anthropic-long-running-design |
native-loop-mappings |
The abstract request cycle maps to DeepSeek steps, Cordis services and split event planes, and to Pi's reusable run loop, message-conversion boundary, injected stream function, and coding-session continuation logic. | SOURCE | dsh-architecture · dsh-agent-turn-source · dsh-tool-pipeline · dsh-subagent-subsystem · pi-agent-readme · pi-agent-loop · pi-agent-wrapper-source · pi-stream-fn-source · pi-provider-composer-source · pi-agent-session · pi-subagent-example |
general-context-working-set |
A model context is a bounded working set assembled from a larger body of transcript, instructions, tool observations, and durable state. | INTERPRETATION | dsh-architecture · pi-compaction |
general-capability-authority |
Tool registration, model visibility, schema validity, authorization, confinement, execution, and effect verification are distinct control questions. | INTERPRETATION | dsh-tool-pipeline · pi-readme |
general-delegation-semantics |
Plans, skills, and subagents are harness conventions whose context inheritance, tools, authority, and result limits vary by implementation. | INTERPRETATION | dsh-architecture · codex-agents · claude-subagents |
dsh-plugin-architecture |
DeepSeek Harness composes its major runtime elements as Cordis plugins. | SOURCE | dsh-readme · dsh-architecture |
dsh-developer-preview |
The pinned DeepSeek Harness version is a developer preview with expected compatibility-breaking changes. | SOURCE | dsh-readme |
dsh-agent-runtime |
DeepSeek AgentLoop creates or resumes agents, while each ReactLoopAgent holds the per-session live runtime state. | SOURCE | dsh-agent-loop-source · dsh-agent-turn-source |
dsh-turn-step |
A DeepSeek step is one model request plus tools; a turn contains zero or more steps. | SOURCE | dsh-architecture · dsh-lifecycle |
dsh-inbox-semantics |
DeepSeek follow-up, steering, and injected context use distinct inbox paths and wake or admission behavior. | SOURCE | dsh-agent-turn-source |
dsh-request-error-recovery |
DeepSeek closes a failed step before request-error handling, and retries only when recovery changes the model-visible surface. | SOURCE | dsh-lifecycle · dsh-compaction-subsystem |
dsh-model-visible-logged |
DeepSeek derives model history from the session log and requires model-visible input to be reconstructable from it. | SOURCE | dsh-architecture · dsh-agent-loop-source |
dsh-tool-execution-pipeline |
DeepSeek tool execution separates policy, guards, approval, dispatch, post-processing, finalization, and durable results. | SOURCE | dsh-tool-pipeline · dsh-tools-source |
dsh-capability-seam-design |
DeepSeek models a swappable capability as a service definition, provider, and consumer, commonly connected through Cordis composition. | SOURCE | dsh-architecture · dsh-capability-seams |
dsh-event-sourced-session |
DeepSeek uses an append-only session event log as the durable basis for derived model history and replay-oriented views. | SOURCE | dsh-architecture · dsh-session-subsystem |
dsh-context-transforms |
DeepSeek compaction can replace a selected surface range with a summary, optional pruning can replace large tool results, and spill storage can retain oversized exact text behind a locator. | SOURCE | dsh-compaction-subsystem · dsh-spill-subsystem |
dsh-skill-loading |
DeepSeek presents a bounded skill catalog to the model and loads a full permitted skill body on demand through the skill tool. | SOURCE | dsh-skills-subsystem |
dsh-subagent-context |
DeepSeek subagent providers distinguish spawn from fork conversation seeding, while child scopes, tools, services, and authority are separate contracts. | SOURCE | dsh-subagent-subsystem |
dsh-dynamic-extensions |
DeepSeek dynamic extensions define versioned Cordis packages with explicit host/client activation, approval, run identity, and retraction lifecycle. | SOURCE | dsh-extensions-subsystem |
pi-three-layers |
Current Pi separates provider API, reusable agent core, and coding-agent integration. | SOURCE | pi-readme · pi-agent-readme |
pi-agent-message-boundary |
Pi converts flexible AgentMessage values to LLM messages at the model-call boundary. | SOURCE | pi-agent-readme · pi-agent-loop |
pi-tool-execution-modes |
Pi supports sequential and parallel tool execution with per-tool sequential forcing. | SOURCE | pi-agent-readme · pi-agent-loop |
pi-read-limits |
The pinned Pi coding-agent read tool supports offset and limit and caps text at 2,000 lines or 50 KiB. | SOURCE | pi-read-tool · pi-truncate |
pi-compaction-defaults |
The pinned Pi coding-agent defaults reserve 16,384 tokens and target 20,000 recent tokens during compaction. | SOURCE | pi-compaction |
pi-compaction-preparation |
Pi compaction preparation separates summarized history from retained context, carries earlier summaries forward, preserves split-turn prefixes, and records file operations for the next summary. | SOURCE | pi-compaction |
pi-compaction-trigger |
Pi automatic compaction triggers when estimated context exceeds the context window minus the configured reserve. | SOURCE | pi-compaction |
pi-no-built-in-permissions |
Pi states that it has no built-in permission system and runs with the launching process's permissions by default. | SOURCE | pi-readme |
pi-agent-session-integration |
Pi coding-agent AgentSession wraps the reusable agent loop with model/auth selection, resources, queues, retries, compaction, extension events, and session UX. | SOURCE | pi-agent-loop · pi-agent-session |
pi-session-branching |
Pi session entries form an append-only branchable tree whose active context path can include compaction and branch-summary entries. | SOURCE | pi-session-manager |
pi-extension-surface |
Pi coding-agent extensions can register tools, commands, event hooks, providers, renderers, and resource contributions across a documented lifecycle. | SOURCE | pi-extensions-doc |
pi-skill-loading |
Pi discovers skill names and descriptions for the system prompt and loads full SKILL.md content on demand through read or an explicit skill command. | SOURCE | pi-skills-doc |
pi-subagent-extension-pattern |
Pi demonstrates subagents as a coding-agent extension that starts separate Pi processes with isolated contexts and explicit tool/model configuration. | SOURCE | pi-subagent-example · pi-extensions-doc |
codex-public-surfaces |
Codex publicly documents approval and sandbox controls, project instructions, skills, and subagent workflows. | DOCUMENTED | codex-cli · codex-security · codex-project-instructions · codex-agents · codex-skills |
claude-public-surfaces |
Claude Code publicly documents tools, permissions, skills, isolated subagents, and context behavior. | DOCUMENTED | claude-tools · claude-permissions · claude-subagents · claude-skills · claude-hooks · claude-context |
closed-internals-unknown |
Public evidence does not establish the internal implementations of current Codex or Claude Code. | UNKNOWN | public-comparisons |