Citation Map

Maps every factual claim in final documents back to a registered source in source_registry.md. No final architecture recommendation may be made unless the supporting research is traceable here.

Format

[CLAIM-XXX] "claim text" → [SRC-XXX] (file/section reference) — URL
  └─ Used in: <document path>

Claims

Codebase Study: Open Responses (SRC-007)

[CLAIM-001] "Open Responses is a self-hosted drop-in replacement for OpenAI's Responses API" → SRC-007 (README.md, line 5) └─ Used in: None

[CLAIM-002] "Open Responses core is a single main.go file (~80KB)" → SRC-007 (main.go, file size) └─ Used in: None

[CLAIM-003] "Open Responses supports Docker Compose V1 and V2 compatibility" → SRC-007 (CLAUDE.md, Docker Compose section) └─ Used in: None

Codebase Study: Codex (SRC-008)

[CLAIM-004] "Codex has 128 Rust crates in its workspace" → SRC-008 (codex-rs/ directory listing, verified: 128 Cargo.toml files on 2026-06-23) └─ Used in: None

[CLAIM-005] "Codex enforces no history rewrite, bounded items, and 10K token cap per context item" → SRC-008 (AGENTS.md, lines 96-100) └─ Used in: None

[CLAIM-006] "Codex prefers modules under 500 LoC, 800 max" → SRC-008 (AGENTS.md, lines 50-54) └─ Used in: None

[CLAIM-007] "Codex uses cross-platform sandboxing: Seatbelt (macOS), Bubblewrap/Landlock (Linux), restricted tokens (Windows)" → SRC-008 (codex-rs/core/README.md) └─ Used in: 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md

[CLAIM-008] "Codex explicitly warns against adding code to codex-core due to bloat" → SRC-008 (AGENTS.md, lines 73-84) └─ Used in: None

Codebase Study: LiteLLM (SRC-009)

[CLAIM-009] "LiteLLM supports 100+ LLM providers in OpenAI format" → SRC-009 (README.md, line 7) └─ Used in: None

[CLAIM-010] "LiteLLM provider translation uses BaseConfig with transform_request() and transform_response()" → SRC-009 (ARCHITECTURE.md, lines 355-366) └─ Used in: None

[CLAIM-011] "LiteLLM's main.py is 342KB, router.py is 512KB, utils.py is 403KB" → SRC-009 (file sizes in litellm/ directory, verified 2026-06-23) └─ Used in: None

[CLAIM-012] "LiteLLM achieves 8ms P95 latency at 1K RPS" → SRC-009 (README.md, line 62) └─ Used in: None

[CLAIM-013] "LiteLLM is used by Stripe, Netflix, Google ADK, OpenAI Agents SDK" → SRC-009 (README.md, lines 66-76) └─ Used in: None

[CLAIM-014] "LiteLLM uses Prisma with PostgreSQL and Redis for persistence/caching" → SRC-009 (ARCHITECTURE.md, lines 186-206) └─ Used in: None

Codebase Study: OpenRouter SDK (SRC-010)

[CLAIM-015] "OpenRouter SDK provides access to 400+ language models through a unified API" → SRC-010 (OVERVIEW.md, line 3) └─ Used in: None

[CLAIM-016] "OpenRouter SDK is auto-generated from OpenAPI specs via Speakeasy" → SRC-010 (CLAUDE.md, line 7) └─ Used in: None

[CLAIM-017] "OpenRouter SDK has three tool types: regular, generator, and manual" → SRC-010 (CLAUDE.md, lines 130-134) └─ Used in: None

[CLAIM-018] "OpenRouter SDK supports async parameter resolution where any parameter can be a function" → SRC-010 (CLAUDE.md, lines 137-141) └─ Used in: None

Model Landscape Research (OpenRouter model pages, 2026-06-23)

[CLAIM-019] "Grok 4.3 pricing is $1.25 input / $2.50 output per 1M tokens with 1M context window" → OpenRouter (og:description meta tag) └─ Used in: None

[CLAIM-020] "GLM 5.2 pricing is $0.98 input / $3.08 output per 1M tokens with 1M context, 20 providers" → OpenRouter (og:description meta tag) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-021] "Claude Fable 5 pricing is $10 input / $50 output per 1M tokens, Mythos-class, 128K max output" → OpenRouter (og:description meta tag) └─ Used in: None

[CLAIM-022] "Nemotron 3 Ultra has 55B active / 550B total params (MoE), $0.50 input / $2.20 output, 1M context" → OpenRouter (og:description meta tag) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-023] "Kimi K2.7 Code activates 32B parameters out of ~1T total, $0.68 input / $3.41 output, 262K context" → OpenRouter (og:description meta tag + model description) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-024] "20x pricing spread exists between cheapest ($0.50 Nemotron) and most expensive ($10 Claude) frontier input pricing" → Calculated from CLAIM-019 through CLAIM-022 └─ Used in: None

[CLAIM-025] "5 of 8 frontier reasoning models support 1M token context windows" → Aggregated from OpenRouter model pages └─ Used in: None

Context Management & Evolution (SRC-002: Hermes, SRC-008: Codex)

[CLAIM-026] "Hermes context_compressor.py is 2,650 lines implementing a 5-phase compression pipeline" → SRC-002 (agent/context_compressor.py, full file) └─ Used in: None

[CLAIM-027] "Hermes uses a 13-section structured summary template for compaction" → SRC-002 (agent/context_compressor.py, lines 1527–1600) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-028] "Hermes implements in-place compaction (PR #38763) as alternative to session rotation" → SRC-002 (agent/conversation_compression.py, lines 527–556) └─ Used in: None

[CLAIM-029] "Hermes uses a DB-backed compression lock to prevent concurrent session forks" → SRC-002 (agent/conversation_compression.py, lines 349–444) └─ Used in: None

[CLAIM-030] "Hermes iterative summarization stores _previous_summary and updates it instead of regenerating" → SRC-002 (agent/context_compressor.py, lines 867–868, 1602–1616, 1684–1685) └─ Used in: None

[CLAIM-031] "Hermes tail protection uses a token-budgeted backward walk (~20K tokens default)" → SRC-002 (agent/context_compressor.py, lines 2231–2333) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-032] "Hermes anti-thrashing skips compression after 2 consecutive passes saving <10%" → SRC-002 (agent/context_compressor.py, lines 967–976) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-033] "Hermes tool output pruning has 3 passes: dedup (MD5 hash), summarize, and truncate arguments" → SRC-002 (agent/context_compressor.py, lines 982–1142) └─ Used in: None

[CLAIM-034] "Hermes MemoryManager allows exactly one built-in and one external memory provider" → SRC-002 (agent/memory_manager.py, lines 314–398) └─ Used in: None

[CLAIM-035] "Hermes session_search uses FTS5 with 4 calling shapes: discovery, scroll, read, browse" → SRC-002 (tools/session_search_tool.py, lines 1–30, 495–616) └─ Used in: None

[CLAIM-036] "Hermes session rotation rolls back to parent session ID if child creation fails" → SRC-002 (agent/conversation_compression.py, lines 603–640) └─ Used in: None

[CLAIM-037] "Hermes summary failure fallback chain: aux model → main model → deterministic fallback → abort" → SRC-002 (agent/context_compressor.py, lines 1691–1828) └─ Used in: None

[CLAIM-038] "Hermes redacts credentials before summarization and again after to catch LLM echo" → SRC-002 (agent/context_compressor.py, lines 1176–1182, 1683) └─ Used in: None

[CLAIM-039] "Codex uses server-side Responses API compaction with per-tool TruncationPolicy" → SRC-008 (codex-rs/core/src/unified_exec/process.rs, line 304; core/src/tools/spec_plan.rs, line 711) └─ Used in: 14_backend_agent_stack_engineering/prompt_context_instruction_engineering.md

[CLAIM-040] "Hermes memory sync runs on a single-worker background ThreadPoolExecutor to avoid blocking turns" → SRC-002 (agent/memory_manager.py, lines 516–572) └─ Used in: None

Prompt, Context & Instruction Engineering (SRC-002: Hermes, SRC-008: Codex, SRC-003: Pi)

[CLAIM-041] "Hermes system prompt uses three tiers: stable (identity, tools, skills), context (AGENTS.md, .cursorrules), volatile (memory, timestamp)" → SRC-002 (agent/system_prompt.py, lines 10–19) └─ Used in: None

[CLAIM-042] "Hermes context file discovery uses a first-match-wins priority cascade: .hermes.md → AGENTS.md → CLAUDE.md → .cursorrules" → SRC-002 (agent/prompt_builder.py, lines 1841–1876) └─ Used in: None

[CLAIM-043] "Hermes dynamic context file truncation uses 6% of context window with 20K floor and 500K ceiling" → SRC-002 (agent/prompt_builder.py, lines 1104–1116) └─ Used in: None

[CLAIM-044] "Hermes truncation preserves 70% head and 20% tail with a recovery marker in the middle" → SRC-002 (agent/prompt_builder.py, lines 1700–1710) └─ Used in: None

[CLAIM-045] "Hermes uses date-only timestamps (strftime '%A, %B %d, %Y') to preserve prefix cache" → SRC-002 (agent/system_prompt.py, lines 448–453) └─ Used in: 14_backend_agent_stack_engineering/prompt_context_instruction_engineering.md

[CLAIM-046] "Hermes edit-format steering maps GPT/Codex to V4A patch mode and Claude/Gemini/open-weight to replace mode" → SRC-002 (agent/coding_context.py, lines 116–132) └─ Used in: None

[CLAIM-047] "Codex AGENTS.md discovery walks from project root to cwd, concatenating all files along the path" → SRC-008 (codex-rs/core/src/agents_md.rs, lines 1–16, 155–241) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-048] "Codex supports AGENTS.override.md as a local override that takes precedence over AGENTS.md" → SRC-008 (codex-rs/core/src/agents_md.rs, lines 37–38, 246–247) └─ Used in: None

[CLAIM-049] "Hermes SubdirectoryHintTracker injects context discovered mid-session into tool results, not system prompt" → SRC-002 (agent/subdirectory_hints.py, lines 57–98) └─ Used in: None

[CLAIM-050] "Hermes subdirectory hints are limited to 8,000 chars and contained to the working directory tree" → SRC-002 (agent/subdirectory_hints.py, lines 36, 169–196) └─ Used in: None

[CLAIM-051] "Hermes coding posture supports 4 modes: auto (default), focus (lean toolset), on (forced), off (disabled)" → SRC-002 (agent/coding_context.py, lines 39–49, 281–298) └─ Used in: None

[CLAIM-052] "Hermes tool-use enforcement is model-family gated to GPT, Codex, Gemini, Gemma, Grok, GLM, Qwen, DeepSeek" → SRC-002 (agent/prompt_builder.py, line 292) └─ Used in: None

[CLAIM-053] "Hermes workspace snapshot includes git branch, upstream, ahead/behind, dirty state, recent 3 commits, manifests, and verify commands" → SRC-002 (agent/coding_context.py, lines 738–789) └─ Used in: None

[CLAIM-054] "Hermes project root detection uses 18+ marker files and 3+ lockfile patterns for package manager inference" → SRC-002 (agent/coding_context.py, lines 74–91) └─ Used in: None

[CLAIM-055] "Hermes DEVELOPER_ROLE_MODELS swaps system to developer role for GPT-5+/Codex at the API boundary" → SRC-002 (agent/prompt_builder.py, line 612) └─ Used in: None

[CLAIM-056] "Pi uses XML tags (, ) for structural boundaries in system prompt" → SRC-003 (packages/coding-agent/src/core/system-prompt.ts, lines 62–67, 155–160) └─ Used in: None

[CLAIM-057] "Hermes skills index uses a two-layer cache: in-process LRU (max 8) + disk snapshot validated by mtime/size manifest" → SRC-002 (agent/prompt_builder.py, lines 1334–1391) └─ Used in: None

[CLAIM-058] "Hermes compact/focus mode demotes non-coding skill categories to names-only but never hides them" → SRC-002 (agent/coding_context.py, lines 449–469) └─ Used in: None

[CLAIM-059] "Hermes defines 14+ platform-specific communication instructions covering WhatsApp, Telegram, Discord, Slack, CLI, WebUI, and others" → SRC-002 (agent/prompt_builder.py, lines 614–818) └─ Used in: None

[CLAIM-060] "Codex uses provenance tracking per instruction entry: source_path, environment_id, cwd" → SRC-008 (codex-rs/core/src/agents_md.rs, lines 472–484) └─ Used in: None

Tool Systems (SRC-002: Hermes, SRC-003: Pi, SRC-008: Codex)

[CLAIM-061] "Hermes ToolRegistry uses module-level auto-discovery via ast.parse to identify files containing registry.register calls before importing them" → SRC-002 (tools/registry.py, lines 29–74) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-062] "Hermes TTL caches check_fn results for 30 seconds to prevent redundant environment probing (like Docker/Modal status checks)" → SRC-002 (tools/registry.py, lines 121–141) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-063] "Hermes ToolRegistry supports override=True plugin registration to replace built-in tools (such as swapping Playwright with a headed Chrome CDP backend)" → SRC-002 (tools/registry.py, lines 247–280) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-064] "Codex traits CoreToolRuntime and ToolExecutor define the interface for native tools, supporting hooks, telemetry tags, and dynamic argument diff streaming" → SRC-008 (codex-rs/core/src/tools/registry.rs, lines 41–145) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-065] "Pi separates its default tool collections into coding-focused tools (read, bash, edit, write) and read-only tools (read, grep, find, ls)" → SRC-003 (packages/coding-agent/src/core/tools/index.ts, lines 138–154) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-066] "Pi uses a pluggable BashOperations interface allowing the bash tool executor to delegate commands to remote shells like SSH" → SRC-003 (packages/coding-agent/src/core/tools/bash.ts, lines 40–58) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-067] "Hermes centralizes tool output truncation limits behind get_tool_output_limits in config.yaml, defaulting to 50KB for bytes, 2000 lines for pagination, and 2000 characters for line length" → SRC-002 (tools/tool_output_limits.py, lines 39–41, 59–89) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-068] "Hermes file path resolution uses absolute, sentinel-free workspace roots (rejecting sentinels like '.', 'auto', 'cwd') to prevent worktree directory misrouting" → SRC-002 (tools/file_tools.py, lines 111–137, 246–257) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-069] "Hermes rejects writing to sensitive paths (/etc/, /boot/, /usr/lib/systemd/, ~/.ssh/, netrc, pgpass, npmrc, pypirc, and ~/.hermes/config.yaml) to prevent security escalations" → SRC-002 (tools/file_tools.py, lines 368–375, 397–424) └─ Used in: 07_tools/built_in_local_tools.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-070] "Hermes checks if the write payload contains line numbers (consecutive digits followed by |) or internal status messages and blocks them to prevent file corruption" → SRC-002 (tools/file_tools.py, lines 597–601, 655–726) └─ Used in: 07_tools/built_in_local_tools.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-071] "Hermes defines HARDLINE_PATTERNS (recursive deletes of root/home, mkfs, dd to raw devices, fork bombs, telinit, shutdown) that cannot be bypassed by YOLO modes" → SRC-002 (tools/approval.py, lines 262–284, 333–343) └─ Used in: 07_tools/built_in_local_tools.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-072] "Hermes normalizes terminal command strings by stripping ANSI escapes, null bytes, shell backslash-escapes, and empty-string literals to resist obfuscation bypasses" → SRC-002 (tools/approval.py, lines 559–591) └─ Used in: 07_tools/built_in_local_tools.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-073] "Hermes terminal tool transforms sudo invocations to sudo -S -p '' and pipes cached or user-provided passwords to stdin to execute elevated commands on TTY-less terminals" → SRC-002 (tools/terminal_tool.py, lines 445–558, 752–800) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-074] "Hermes terminal tool brace-groups backgrounded compound statements (A && { B & }) to prevent hanging stdout pipes and waiting on long-running processes" → SRC-002 (tools/terminal_tool.py, lines 587–749) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-075] "Hermes Session Search tool provides query-based discovery with first/last 3 message bookends and message drill-down centered on specific anchor message IDs" → SRC-002 (tools/session_search_tool.py, lines 5–30) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-076] "Hermes delegate_task blocks child agents from executing clarify, memory, execute_code, send_message, or delegate_task to prevent recursive loops and side effects" → SRC-002 (tools/delegate_tool.py, lines 44–53) └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md, 07_tools/built_in_local_tools.md

[CLAIM-077] "Hermes subagent execution handles TUI deadlock by installing non-interactive auto-deny or auto-approve callbacks into worker threads since they do not inherit the parent's TTY" → SRC-002 (tools/delegate_tool.py, lines 57–112) └─ Used in: None

[CLAIM-078] "Hermes cronjob creator scans user prompts strictly for exfiltration patterns and invisible Unicode characters, while sanitizing and stripping invisible Unicode from pre-vetted skill content" → SRC-002 (tools/cronjob_tools.py, lines 225–278) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-079] "Hermes todo tool maintains session todo items on the agent instance, injecting only active (pending/in_progress) tasks up to a capped count (256) and description length (4000) after compaction" → SRC-002 (tools/todo_tool.py, lines 24–33, 106–138) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-080] "Hermes progressive disclosure skills system lists names and descriptions first, deferredly loading full instructions and linked files on-demand" → SRC-002 (tools/skills_tool.py, lines 9–13, 52–67) └─ Used in: 07_tools/built_in_local_tools.md, 19_final_reports/harness_architecture_specification_report.md

Agent Execution Loops (SRC-001, SRC-002, SRC-003, SRC-004, SRC-005, SRC-008)

[CLAIM-081] "Hermes conversation loop implements dual-limit budgeting (max_iterations and IterationBudget remaining check) combined with a single-turn _budget_grace_call fallback turn" → SRC-002 (agent/conversation_loop.py, lines 589–614) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-082] "Hermes conversation loop drains pending user steering input (_drain_pending_steer) before the model call on each iteration, injecting the steer content into the last tool message to keep message sequence roles valid" → SRC-002 (agent/conversation_loop.py, lines 662–700) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-083] "Hermes executes defensive role-alternation repairs (repair_message_sequence_with_cursor) and sanitizes malformed tool call JSON payloads (_repair_tool_call_arguments) before every LLM request" → SRC-002 (agent/conversation_loop.py, lines 707–737) └─ Used in: None

[CLAIM-084] "Hermes conversation loop features eager provider failovers, switching to fallback providers on network timeouts, invalid responses, rate limits, or safety refusals, with jittered exponential backoffs up to 120s between attempts" → SRC-002 (agent/conversation_loop.py, lines 1280–1417) └─ Used in: 14_backend_agent_stack_engineering/context_management_and_evolution.md

[CLAIM-085] "Pi and OpenClaw agent loops utilize a nested dual while-loop structure: an outer loop that handles follow-up/queued messages, and an inner loop that processes tool execution and steering messages" → SRC-003 (packages/agent/src/agent-loop.ts, lines 170–266) and SRC-001 (packages/agent-core/src/agent-loop.ts, lines 297–429) └─ Used in: None

[CLAIM-086] "Pi and OpenClaw agent loops resolve sequential vs. parallel tool execution modes by scanning the batch of requested tool calls and executing them sequentially if any tool requires sequential execution mode" → SRC-003 (packages/agent/src/agent-loop.ts, lines 380–388) and SRC-001 (packages/agent-core/src/agent-loop.ts, lines 547–589) └─ Used in: None

[CLAIM-087] "Pi and OpenClaw agent loops expose beforeToolCall and afterToolCall lifecycle hooks that can block tool execution or rewrite tool results dynamically" → SRC-003 (packages/agent/src/agent-loop.ts, lines 581–605, 682–702) and SRC-001 (packages/agent-core/src/agent-loop.ts, lines 581–605, 682–707) └─ Used in: None

[CLAIM-088] "Codex drives a turn loop (run_turn) that intercepts pending user inputs (get_pending_input) mid-run and triggers pre-turn and mid-turn auto-compaction when context tokens exceed limits" → SRC-008 (codex-rs/core/src/session/turn.rs, lines 215–223, 287–380) └─ Used in: None

[CLAIM-089] "Codex resolves explicitly enabled plugin and skill references in user input to inject turn-scoped instructions (build_skills_and_plugins) before generating prompts" → SRC-008 (codex-rs/core/src/session/turn.rs, lines 170–175, 521–684) └─ Used in: None

[CLAIM-090] "LangGraph drives loop ticks (tick and after_tick in PregelLoop) using a Pregel state channel model that evaluates triggers, spawns concurrent task executions, saves checkpoints, and handles node-level error handlers" → SRC-004 (libs/langgraph/langgraph/pregel/_loop.py, lines 592–674, 676–715) └─ Used in: None

[CLAIM-091] "LangGraph supports human-in-the-loop capabilities by raising a GraphInterrupt before or after node execution when matching configured interrupt criteria" → SRC-004 (libs/langgraph/langgraph/pregel/_loop.py, lines 659–664, 707–712) └─ Used in: 04_agent_loops/human_in_the_loop_steering.md

[CLAIM-092] "LangChain's classic AgentExecutor implements a while-loop (_call) bounded by iteration caps and wall-clock execution time limits" → SRC-005 (libs/langchain/langchain_classic/agents/agent.py, lines 1235–1239, 1585–1600) └─ Used in: None

[CLAIM-093] "LangChain AgentExecutor routes output parser errors back to the model as observations by executing a virtual _Exception tool, allowing the agent to self-correct" → SRC-005 (libs/langchain/langchain_classic/agents/agent.py, lines 1322–1361) └─ Used in: None

[CLAIM-094] "LangChain early stopping methods include force (returns static limit message) and generate (sends a final completion request to the model prompting it to synthesize an answer from the intermediate steps)" → SRC-005 (libs/langchain/langchain_classic/agents/agent.py, lines 940–968) └─ Used in: None

[CLAIM-095] "LiteLLM utilizes tenacity declarative retry loops (tenacity.Retrying and tenacity.AsyncRetrying) in completion_with_retries and acompletion_with_retries to recover from transient API timeouts, 429s, or model outages." → SRC-009 (litellm/main.py, lines 5840-5899) └─ Used in: None

[CLAIM-096] "OpenRouter SDK implements an asynchronous while(true) retry loop (retryBackoff in retries.ts) that manages connection error recovery, exponential backoffs, random jitter, and HTTP Retry-After header parsing." → SRC-010 (src/lib/retries.ts, lines 153-195) └─ Used in: None

[CLAIM-097] "assistant-ui implements a reconnecting SSE stream parser (openPiEventStream in eventSource.ts) with a nested outer reconnection while loop, an inner stream reader while loop, and a chunk framing loop." → SRC-006 (packages/react-pi/src/client/eventSource.ts, lines 144-192) └─ Used in: None

[CLAIM-098] "Open Responses' CLI tool (main.go) uses Go iteration loops inside Cobra command runners to check service health, tail docker logs, and parse configurations." → SRC-007 (main.go, lines 1805-1851) └─ Used in: None

[CLAIM-099] "Hermes's Docker environment (docker.py) runs tini or catatonit as PID 1 to prevent zombie processes, and runs an asynchronous orphan container reaper to clean up exited containers." → SRC-002 (tools/environments/docker.py, lines 138-237, 858-870) └─ Used in: 07_tools/built_in_local_tools.md

[CLAIM-100] "Hermes's Daytona environment (daytona.py) supports persistent sandboxes and implements file sync batching via sandbox.fs.upload_files() multipart uploads to bypass per-file TLS/HTTP handshake overhead." → SRC-002 (tools/environments/daytona.py, lines 89-130, 160-180) └─ Used in: None

[CLAIM-101] "Docker and Podman provide open-source container isolation; Firecracker implements KVM-based microVMs with sub-10ms boot times; gVisor intercepts system calls via a userspace Sentry kernel." → SRC-002 (tools/environments/docker.py) and Firecracker └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-102] "Building container sandboxes from scratch requires system primitives including Linux namespaces (MNT, PID, NET, USER, IPC, UTS), cgroups v2 resource limits, chroot or pivot_root mount isolation, and seccomp filters." → Linux kernel docs └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-103] "Building Firecracker microVMs from scratch requires a compiled minimal vmlinux kernel, an ext4 rootfs disk image, resources configuration, and JSON REST API socket commands." → Firecracker └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-104] "Chrome Extension Bridges connect active browser tabs to local agent runtimes via WebSockets, allowing DOM manipulation, credential inheritance, and bot-detection bypass." → Chrome DevTools Protocol └─ Used in: 19_final_reports/harness_architecture_specification_report.md

Conversation Completions, Anthropic Shapes, and Reranking (SRC-001, SRC-002)

[CLAIM-105] "Hermes's Anthropic adapter extracts system messages and passes them as a top-level system parameter." → SRC-002 (agent/anthropic_adapter.py, lines 2267–2281) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-106] "Hermes's Anthropic adapter enforces role alternation by merging consecutive same-role messages." → SRC-002 (agent/anthropic_adapter.py, lines 2045–2095) └─ Used in: 14_backend_agent_stack_engineering/context_management_and_evolution.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-107] "Hermes's Anthropic adapter converts tool messages to tool_result content blocks inside a user message, merging consecutive tool results." → SRC-002 (agent/anthropic_adapter.py, lines 1909–1968) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-108] "Hermes's Anthropic adapter strips orphaned tool_use and tool_result blocks, demoting thinking blocks to text if their signatures are invalidated." → SRC-002 (agent/anthropic_adapter.py, lines 1988–2044) └─ Used in: 14_backend_agent_stack_engineering/context_management_and_evolution.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-109] "Hermes's Anthropic adapter strips thinking block signatures for third-party endpoints, keeps unsigned thinking for Kimi/DeepSeek, and downgrades unsigned thinking to text for direct Anthropic." → SRC-002 (agent/anthropic_adapter.py, lines 2097–2202) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-110] "Hermes's Anthropic adapter evicts older multimodal screenshots to save context, keeping only the 3 most recent images." → SRC-002 (agent/anthropic_adapter.py, lines 2203–2237) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-111] "Hermes's Anthropic adapter sanitizes responses-only keys like instructions and store to prevent SDK type errors." → SRC-002 (agent/anthropic_adapter.py, lines 2554–2583) └─ Used in: None

[CLAIM-112] "Hermes's Anthropic adapter normalizes tool names under OAuth to double-underscore mcp__ prefix to bypass subscription billing filters." → SRC-002 (agent/anthropic_adapter.py, lines 2406–2422) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-113] "The qmd hybrid retrieval engine uses embeddinggemma-300M-Q8_0 for vector embeddings, qwen3-reranker-0.6b-q8_0 for reranking, and qmd-query-expansion-1.7B for query expansion." → SRC-002 (optional-skills/research/qmd/SKILL.md, lines 65–73) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-114] "The qmd search engine merges sparse and dense queries via Reciprocal Rank Fusion (RRF) with top-rank weight boosts." → SRC-002 (optional-skills/research/qmd/SKILL.md, lines 375–380) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-115] "The qmd search engine blends retrieval and reranking scores dynamically based on candidate rank range, trusting the reranker more for the long tail." → SRC-002 (optional-skills/research/qmd/SKILL.md, lines 381–384) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-116] "Gateways like OpenCode or Hermes support Claude Code integration by loading and refreshing first-party OAuth credentials from macOS Keychain entries ('Claude Code-credentials') and local JSON records." → SRC-002 (agent/anthropic_adapter.py, lines 859–954) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-117] "Gateways spoof Claude Code user-agent details (claude-cli/2.1.74) and inject specific OAuth beta headers to allow third-party tools to route through Anthropic plan billing." → SRC-002 (agent/anthropic_adapter.py, lines 804–813, 2376–2422) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-118] "Open Responses provides a self-hosted drop-in alternative for OpenAI's stateful Responses API, allowing client SDKs to query local models via a unified thread-runner spec." → SRC-007 (open-responses/README.md, lines 17–22) └─ Used in: 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-119] "LiteLLM manages local model execution by automatically spawning background 'ollama serve' subprocesses to host offline models." → SRC-009 (litellm/tests/test_litellm/proxy/proxy_server/test_lifecycle.py, lines 442–457) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-120] "Coding agent systems are bootstrapped CLI-first to validate the core ReAct loops and tools in a bare-metal shell environment before abstracting them into SDK classes or editor frontends." → Developer agent bootstrapping analysis └─ Used in: None

[CLAIM-121] "Transitioning code-executing agents to multi-tenant architectures requires compute isolation (running untrusted code in resource-constrained sandboxes or container microVMs) to prevent tenant file access and exfiltration." → Firecracker, gVisor └─ Used in: None

[CLAIM-122] "Multi-tenant database architectures isolate concurrent session state by provisioning workspace-scoped SQLite files to prevent SQLite database lock contention." → SRC-002 (hermes_state.py) └─ Used in: 14_backend_agent_stack_engineering/context_management_and_evolution.md

[CLAIM-123] "Gateway proxies intercept outbound request streams, verifying tenant JWT auth tokens to track model costs and enforce organizational rate-limiting quotas." → SRC-009 (litellm/proxy/) └─ Used in: None

Channel Connectors & Secure Device-Pairing Protocols (SRC-001)

[CLAIM-124] "The gateway generates a short-lived SetupPayload structure containing the gateway public URL, a single-use bootstrap token, and expiration timestamp." → SRC-001 (extensions/device-pair/index.ts, lines 51-55, 600-610) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-125] "The setup code is serialized and normalized into a URL-safe Base64 string." → SRC-001 (extensions/device-pair/index.ts, lines 462-466) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 14_backend_agent_stack_engineering/prompt_context_instruction_engineering.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-126] "The URL-safe setup code is converted to a QR PNG image using a temporary file system pipeline." → SRC-001 (extensions/device-pair/index.ts, lines 790-800) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md

[CLAIM-127] "The setup QR is dispatched as media over platform-specific channel-specific senders." → SRC-001 (extensions/device-pair/index.ts, lines 613-643, 801-817) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md

[CLAIM-128] "Pairing network policies block cleartext ws:// connections unless loopback or private LAN ranges, otherwise forcing wss://." → SRC-001 (extensions/device-pair/index.ts, lines 260-307) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-129] "Handshake client verification holds requests in a pending queue until approved via /pair approve command." → SRC-001 (extensions/device-pair/index.ts, lines 715-736) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-130] "Twilio SMS dispatches REST POST requests to Twilio Messages API using Basic authentication." → SRC-001 (extensions/sms/src/twilio.ts, lines 521-567) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-131] "SMS webhook security verifies signatures using timing-safe comparisons of computed SHA-1 HMAC hashes." → SRC-001 (extensions/sms/src/twilio.ts, lines 201-218) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-132] "Slack channel connector verifies HMAC-SHA256 signatures, formats rich blocks, and binds thread_ts conversation keys." → SRC-001 (extensions/slack/src/channel.ts, lines 247-269, 524-590) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-133] "Telegram ingress spools incoming updates asynchronously to handle spike traffic without timeouts." → SRC-001 (extensions/telegram/src/telegram-ingress-spool.ts, lines 1-100) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-134] "Telegram forum topic binding creates canonical session threads mapped as chatId:topicId to isolate sub-chats." → SRC-001 (extensions/telegram/src/channel.ts, lines 524-596) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-135] "WhatsApp connector connects using emulated socket loops, captures QR codes, and caches authenticated session keys." → SRC-001 (extensions/whatsapp/src/login-qr.ts, lines 191-244, 313-500) └─ Used in: 12_open_source_voice_integration/channel_connectors_and_pairing.md, 19_final_reports/harness_architecture_specification_report.md

Gotchas, Caching, Tracing & Regex Constraints (SRC-002, SRC-005, SRC-001)

[CLAIM-136] "Prompt caching requires stable prefixes and placing volatile parameters (like time and dynamic chat histories) at the absolute tail of the message array to avoid invalidating cache partitions." → SRC-002 (agent/system_prompt.py, lines 448-453) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-137] "Regex-based parsing of JSON tool payloads is fragile and presents ReDoS backtracking risks, making programmatic parsing (JSON5, AST) or structured LLM schemas preferred." → SRC-002 (agent/conversation_loop.py, lines 707-737; agent/message_sanitization.py, lines 185-279) └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-138] "SQLite database concurrency is throttled by write locks in WAL mode, requiring randomized retry jitter (e.g. 15 retries, 20-150ms backoff) or dedicated single-tenant database files." → SRC-002 (agent/conversation_compression.py, lines 349–444; hermes_state.py, lines 929-977) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-139] "Docker sandbox runtimes require a PID 1 init process like tini or catatonit to reap zombie subprocesses and prevent file descriptor leakage." → SRC-002 (tools/environments/docker.py, lines 138-237, 858-870) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-140] "Tool path validators must resolve paths absolute-only, ignoring sentinel-based prefixes like '.' or 'cwd' to prevent accidental escape of target git worktrees." → SRC-002 (tools/file_tools.py, lines 111–137) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-141] "To prevent shell execution hangs in TTY-less environments, interactive sudo commands must use standard redirect flags (sudo -S -p '') and background processes must be brace-grouped." → SRC-002 (tools/terminal_tool.py, lines 587-749) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-142] "Self-healing loops use virtual tools (like LangChain's _Exception tool) to pass parsing/validation failures back to the LLM for correction, bounded by small loop limits to prevent cost spirals." → SRC-005 (libs/langchain/langchain_classic/agents/agent.py, lines 1322-1361) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-143] "Twilio webhook verification is highly sensitive to reverse proxies terminating SSL or rewriting requested URLs, requiring explicit header parsing to reconstruct absolute origin endpoints." → SRC-001 (extensions/sms/src/twilio.ts, lines 201-218) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-144] "WebSocket-based personal WhatsApp client emulators are prone to session file corruption upon sudden connection drops, requiring transaction-safe double-buffered credentials updates." → SRC-001 (extensions/whatsapp/src/login-qr.ts, lines 191-244) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

Multi-Model Deliberation & Fusion Patterns (SRC-011, SRC-012, SRC-013, SRC-014, SRC-015, SRC-016, SRC-017, SRC-018)

[CLAIM-145] "OpenRouter Fusion dispatches prompts to a panel of 2–8 models in parallel with web search enabled; a judge model produces structured JSON analysis (consensus, contradictions, blind spots, unique insights) without merging; the primary model uses the analysis to write the final answer." → SRC-011 (Plugin Configuration section) └─ Used in: 06_subagents/README.md, 06_subagents/multi_model_deliberation_and_swarms.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-146] "OpenRouter Fusion uses x-openrouter-fusion-depth headers to prevent recursive fusion invocations; panel and judge models cannot invoke fusion a second time, bounding deliberation to a single level." → SRC-011 (Recursion protection section) └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-147] "The Mixture-of-Agents paper (Wang et al., ICLR 2025) demonstrated that an open-source-only MoA using Llama, Qwen, and similar models achieved state-of-the-art results on AlpacaEval 2.0, outperforming GPT-4 as a standalone model, by leveraging the collaborativeness property where LLMs produce better outputs when referencing prior responses." → SRC-014 └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-148] "Together AI provides a minimal ~50-line Python reference implementation of the Mixture-of-Agents architecture using the together Python SDK to orchestrate proposer and aggregator models." → SRC-018 └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md

[CLAIM-149] "Karpathy's llm-council implements a 3-stage council workflow: (1) independent opinions sent to multiple LLMs simultaneously, (2) anonymized peer review where models rank and critique each other, (3) chairman synthesis compiling the final answer from all opinions and critiques." → SRC-016 └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md

[CLAIM-150] "Council/debate implementations require anonymizing model identifiers to prevent lab-bias (GPT judges favoring GPT outputs), rotating challengers to counteract sycophancy, and anti-capitulation prompts to prevent models from changing answers without genuine reasoning." → SRC-016, SRC-017 (Consilium) └─ Used in: 06_subagents/README.md, 06_subagents/multi_model_deliberation_and_swarms.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-151] "Research demonstrates a ~35.9% reduction in hallucination rates via multi-agent consensus compared to single-model setups, as models independently arriving at the same factual claim provide higher confidence than any single assertion." → SRC-014, SRC-015 (DRACO benchmark research) └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-152] "CrewAI implements the Supervisor-Worker swarm pattern via Process.hierarchical configuration, automatically creating a manager agent that delegates sub-tasks to role-specialized worker agents with A2A protocols, built-in memory, and checkpointing." → SRC-012 └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md

[CLAIM-153] "LangGraph StateGraph implements multi-model deliberation as directed graphs with conditional edges for confidence-based routing, reducers to merge parallel branch state, cycles for iterative debate loops, and first-class human-in-the-loop governance gates." → SRC-013 └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md

[CLAIM-154] "The danielrosehill/Awesome-LLM-Council-Projects repository curates an ecosystem of council-derived projects including Consilium (anti-sycophancy CLI), PolyCouncil (local model councils), llm-deliberate (social choice voting), and the judges library (Jury aggregation objects)." → SRC-017 └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md

[CLAIM-155] "Microsoft Agent Framework (MAF) succeeded AutoGen and Semantic Kernel, shifting from unbounded conversational debate ('conversational chaos') to explicit graph-based workflows with strict type safety, durable state, enterprise telemetry, and checkpointing. The community fork AG2 preserves the original AutoGen conversational debate style." → Microsoft, AG2 └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md

[CLAIM-156] "OpenAI Agents SDK (March 2025) succeeded the experimental Swarm framework, adding built-in tracing, guardrails, session management, and observability for production-grade agent-to-agent handoffs." → OpenAI Agents SDK └─ Used in: 06_subagents/multi_model_deliberation_and_swarms.md

[CLAIM-157] "DRACO benchmark testing shows budget fusion panels (e.g., Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro with a frontier judge) outperform standalone frontier models at ~50% cost, and even self-synthesis (pairing a model with itself) produces measurably better outputs than a single run." → SRC-015, SRC-011 (DRACO benchmark, OpenRouter Fusion documentation) └─ Used in: 06_subagents/README.md, 06_subagents/multi_model_deliberation_and_swarms.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

Case Studies: Codex Desktop, VS Code Copilot Chat, Cursor, and Google Antigravity (SRC-019, SRC-020, SRC-021, SRC-022)

[CLAIM-158] "OpenAI Codex Desktop App is designed as a centralized command center supporting parallel multi-threaded project control and dedicated Plan Mode (interviewing/task breakdown) and Goal Mode (long-running autonomous execution)." → [SRC-019] └─ Used in: 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-159] "Codex Desktop App features automated Computer Use with secure local sandboxing (Windows/macOS) and interactive DOM/screenshot annotations allowing agents to perform design-loop visual adjustments." → [SRC-019] └─ Used in: 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-160] "Codex Desktop App integrates git worktrees to support parallel development tracks, routing commands through JSON-RPC over local WebSocket/Unix sockets via codex app-server." → [SRC-019], [SRC-008] └─ Used in: 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-161] "GitHub Copilot Agent Mode in VS Code divides workspace orchestration between the code-first Chat View panel and a dedicated Agents Window designed for multi-project/repository workflows." → [SRC-020] └─ Used in: 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-162] "VS Code Agent Mode executes terminal commands autonomously (such as npm run dev or testing loops), gating executions behind security approval prompts that the user can authorize or configure." → [SRC-020] └─ Used in: 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-163] "VS Code Agent Mode leverages custom tool configurations and Model Context Protocol (MCP) server inputs, requiring Rich Shell Integration to parse terminal outputs accurately." → [SRC-020] └─ Used in: 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-164] "Cursor Composer provides a multi-file edit workspace supporting concurrent inline editor diff previews (red/green highlight annotations) and programmatic change rejection/acceptance." → SRC-021 └─ Used in: 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-165] "Cursor Composer Agent Mode controls execution safety through settings (Ask Every Time, Run in Sandbox, Run Everything) and allow/deny lists defined in project-level .cursor/cli.json configuration files." → SRC-021 └─ Used in: 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-166] "Google Antigravity environment uses an agent-first standalone app and custom IDE fork, offering a dedicated Manager Surface to track parallel asynchronous agent execution." → [SRC-022] └─ Used in: 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-167] "Google Antigravity enforces a strict planning artifact lifecycle, generating and updating implementation_plan.md, task.md, and walkthrough.md files to verify coding steps." → [SRC-022] └─ Used in: 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-168] "Google Antigravity implements interactive browser subagent tools, capturing WebP recordings in the background to demonstrate UI changes." → [SRC-022] └─ Used in: 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

[CLAIM-169] "Google Antigravity utilizes granular tool approvals for run_command execution, write_to_file path access, and supports slash commands like /goal, /schedule, and /grill-me." → [SRC-022] └─ Used in: 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md

Generative UI in Responses, MCP Apps, and MCP UI (SRC-023, SRC-024, SRC-025, SRC-026)

[CLAIM-170] "Generative UI (GenUI) is a front-end design where the model constructs interface layouts on the fly, dynamically updating the frontend based on user intent and conversational context." → [SRC-024], [SRC-025] └─ Used in: 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-171] "OpenAI Responses API and other structured interfaces support Generative UI by enforcing JSON schemas (response_format structures) that outline components and props for client-side rendering." → [SRC-025] └─ Used in: 08_mcps/README.md, 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-172] "Vercel AI SDK provides standard React hook patterns (AI SDK UI) to handle LLM streams, structured outputs, and progressive component streaming." → [SRC-025] └─ Used in: 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-173] "CopilotKit provides an open-source agentic application framework that coordinates client-side React components with agent pipelines using the AG-UI protocol." → [SRC-024] └─ Used in: 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 18_architecture_recommendations/README.md

[CLAIM-174] "Model Context Protocol (MCP) Apps, standardized via SEP-1865, allow remote MCP servers to serve interactive visual widgets and application frames directly to compatible hosts." → [SRC-023] └─ Used in: 08_mcps/README.md, 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-175] "MCP Apps ensure client security by executing generated HTML, CSS, and JavaScript inside sandboxed, isolated iframes rather than directly in the host application's DOM." → [SRC-023] └─ Used in: 08_mcps/README.md, 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-176] "Mastra AI and mcp-use offer TypeScript SDK interfaces to author and expose MCP tool servers containing rich UI widget assets alongside standard data endpoints." → [SRC-026] └─ Used in: 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 18_architecture_recommendations/README.md

[CLAIM-177] "The Model Context Protocol specification introduced a stateless Core in mid-2026, shifting context management to dedicated Extensions like the Tasks framework." → [SRC-023] └─ Used in: 08_mcps/README.md, 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-178] "The Tasks Extension in MCP models long-running asynchronous background executions as durable state machines driven by the client to prevent context bloating." → [SRC-023] └─ Used in: 08_mcps/README.md, 08_mcps/mcp_apps_and_ui.md, 15_frontend_react_vite_agent_stack/README.md, 15_frontend_react_vite_agent_stack/complete_frontend_architecture.md, 15_frontend_react_vite_agent_stack/frontend_architecture_and_ux.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

LibreChat & LibreChat Agents SDK Reference Integration (SRC-027, SRC-028)

[CLAIM-179] "LibreChat manages user-scoped Model Context Protocol (MCP) server connections dynamically via Redis flow state maps and CSRF-binding cookies during callback authentication." → SRC-027 (api/server/routes/mcp.js, lines 104-220, 226-515) └─ Used in: 02_open_source_agent_frameworks/README.md, 08_mcps/README.md, 08_mcps/mcp_apps_and_ui.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-180] "LibreChat implements the Open Responses API standard, converting input lists to internal messages and streaming semantic events like response.output_item.added via Server-Sent Events (SSE)." → SRC-027 (api/server/routes/agents/responses.js, lines 44-84; api/server/controllers/agents/responses.js, lines 284-336) └─ Used in: 15_frontend_react_vite_agent_stack/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-181] "LibreChat controllers utilize Breadth-First Search (BFS) discovery patterns to traverse connected sub-agents dynamically while verifying remote sharing permissions." → SRC-027 (api/server/controllers/agents/responses.js, lines 488-565) └─ Used in: 19_final_reports/harness_architecture_specification_report.md

[CLAIM-182] "LibreChat memory processors filter out skill-primed meta messages from the memory window to protect the agent's personalization engine from instruction injection." → SRC-027 (api/server/controllers/agents/client.js, lines 776-800) └─ Used in: 05_agent_memory/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-183] "LibreChat Agents SDK defines a structured token calibration formula calibrationRatio = cumulativeProviderReported / cumulativeRawSent to align local tokenizer estimates with provider metrics." → SRC-028 (src/run.ts, lines 133-160, 502-510) └─ Used in: 01_open_source_agentic_sdks/README.md, 05_agent_memory/README.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-184] "LibreChat Agents SDK implements dynamic instruction overhead calibration using a 15% variance threshold (CALIBRATION_VARIANCE_THRESHOLD) to update the tool schema token ceiling." → SRC-028 (src/graphs/Graph.ts, lines 97, 1551; docs/summarization-behavior.md, line 47) └─ Used in: 01_open_source_agentic_sdks/README.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-185] "LibreChat Agents SDK processes multi-agent graph topologies via LangGraph Command structures (Command.PARENT), enabling conditional handoffs and sequential transfers." → SRC-028 (src/graphs/MultiAgentGraph.ts, lines 311-381, 401-512) └─ Used in: 02_open_source_agent_frameworks/README.md, 08_mcps/README.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-186] "Handoff reception in the LibreChat Agents SDK is handled by filtering out the transfer tool calls and preceding messages to give the destination agent a clean context window." → SRC-028 (src/graphs/MultiAgentGraph.ts, lines 558-717) └─ Used in: 02_open_source_agent_frameworks/README.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-187] "Observation masking in the Agents SDK replaces consumed ToolMessage payloads (which have subsequent substantive AI reasoning text) with character-limited previews (~300 characters) when context pressure reaches 80%." → SRC-028 (docs/summarization-behavior.md, lines 13-22, 152-162) └─ Used in: 01_open_source_agentic_sdks/README.md, 05_agent_memory/README.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-188] "Mid-run summaries in the Agents SDK are stored on the AgentContext and injected as a HumanMessage on clean state turns, allowing summaries to compete for message budget instead of lowering system instruction ceilings." → SRC-028 (docs/summarization-behavior.md, lines 74-90, 141-150) └─ Used in: 01_open_source_agentic_sdks/README.md, 05_agent_memory/README.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-189] "Conversation Steering allows injecting a HumanMessage dynamically into active execution loops without resetting history context, which requires resolving strict role alternation requirements." → SRC-004 (libs/langgraph/langgraph/types.py, lines 810-934) and SRC-002 (agent/conversation_loop.py, lines 594, 1397, 2664) └─ Used in: 04_agent_loops/README.md, 04_agent_loops/human_in_the_loop_steering.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-190] "Execution loops monitor explicit interrupt flags (like _interrupt_requested) at superstep or turn boundaries to pause operations cleanly before starting subsequent API calls or tool invocations." → SRC-002 (run_agent.py, lines 2400-2446; agent/conversation_loop.py, lines 594, 3557) and SRC-004 (libs/langgraph/langgraph/pregel/_loop.py, lines 659-664) └─ Used in: 04_agent_loops/README.md, 04_agent_loops/human_in_the_loop_steering.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-191] "Christian Vastveit's cascading retry fix in PR #6600 sets a request-local cancellation token during force-close, letting the HTTP generation worker trap transport errors and bubble up a clean InterruptedError instead of retrying." → SRC-002 (tests/agent/test_cascading_interrupt_6600.py, lines 9-23, 51-80) └─ Used in: 04_agent_loops/README.md, 04_agent_loops/human_in_the_loop_steering.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-192] "Interactive governance gates halt loops on security boundaries (e.g. executing terminal commands) and broker approvals over mobile APNS notifications, CLI prompts, or sandbox widgets." → SRC-001 (packages/agent-core/src/exec-approval-ios-push.ts, lines 15-56) and SRC-006 (README.md, lines 358-372) └─ Used in: 04_agent_loops/README.md, 04_agent_loops/human_in_the_loop_steering.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-193] "Session-level tool approvals, such as in Hermes's computer use module, temporary auto-approve subsequent UI actions like click or type for the remainder of the session once an initial permission is granted." → SRC-002 (tools/computer_use/tool.py, lines 143, 293-314) └─ Used in: 04_agent_loops/README.md

[CLAIM-194] "Nous Hermes implements three distinct edit bypass policies (ask, workspace_session, and session) to decouple manual human checks from autonomous execution options." → SRC-002 (acp_adapter/edit_approval.py, lines 44-47, 148-180) └─ Used in: 04_agent_loops/README.md, 04_agent_loops/human_in_the_loop_steering.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-195] "Regardless of active auto-approve bypass settings, writes targeting sensitive files (such as .env, credentials, or SSH keys) are always hard-gated and must prompt the human for approval." → SRC-002 (acp_adapter/edit_approval.py, lines 44-45, 140-147) └─ Used in: 04_agent_loops/README.md, 04_agent_loops/human_in_the_loop_steering.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-196] "Swarm delegation parameters in Nous Hermes enforce default auto-deny policies for nested subagent tool calls when executing headlessly inside background jobs, unless overridden by explicit subagent auto-approve options." → SRC-002 (tools/delegate_tool.py, lines 67-81, 103-110, 1639) └─ Used in: 04_agent_loops/human_in_the_loop_steering.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-197] "Nous Hermes implements a session-level todo tool enabling agents to create, read, update, and merge checklists to manage multi-step subtasks." → SRC-002 (tools/todo_tool.py, lines 187-227, 240-294) └─ Used in: 05_agent_memory/agent_scratchpads_and_session_memory.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-198] "The Hermes TodoStore encapsulates an in-memory checklist bound to the AIAgent session, capping items at 256 and content lengths at 4000 characters to prevent context window inflation." → SRC-002 (tools/todo_tool.py, lines 24-34, 36-47, 91-96) └─ Used in: 05_agent_memory/agent_scratchpads_and_session_memory.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-199] "Upon context compaction, the Hermes session runtime re-injects the active task checklist list back into the conversation context as a system-like formatting block to preserve task focus." → SRC-002 (tools/todo_tool.py, lines 106-114, 133-138) └─ Used in: 05_agent_memory/README.md, 05_agent_memory/agent_scratchpads_and_session_memory.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-200] "To prevent the agent from repeating resolved actions post-compaction, the TodoStore re-injection logic filters out completed and cancelled checklist items, yielding only pending and in-progress tasks." → SRC-002 (tools/todo_tool.py, lines 124-132) └─ Used in: 05_agent_memory/agent_scratchpads_and_session_memory.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-201] "Autonomous agents create and maintain user-visible workspace check-lists (like TODO.md or NOTES.md) directly in the project directory to track session objectives across ephemeral runs." → SRC-021 (Cursor Composer task listings; Claude Code CLI notes usage) └─ Used in: 05_agent_memory/agent_scratchpads_and_session_memory.md

[CLAIM-202] "Claude Code boots by checking a repository-level CLAUDE.md rule file or progressive sub-folder rules inside .claude/rules/ to load coding guidelines and project-specific testing tools." → SRC-020 (Claude Code memory files setup rules; VS Code instructions search) └─ Used in: 05_agent_memory/agent_scratchpads_and_session_memory.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-203] "Google Antigravity restricts session-specific scratch files, temporary compilations, and experimental scripts to a conversation-locked app data directory (<appDataDir>/brain/<conversation-id>/scratch/) to avoid VCS code tree pollution." → SRC-022 (Antigravity scratch space conventions; App data directory rules) └─ Used in: 05_agent_memory/README.md, 05_agent_memory/agent_scratchpads_and_session_memory.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-204] "Learned user corrections and styling preferences are cached to a local .claude/memory.md file using an auto-memory model, which allows developer auditability via /memory commands." → SRC-020 (Claude Code memory log storage and command schema) └─ Used in: 05_agent_memory/agent_scratchpads_and_session_memory.md, 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-205] "Knowledge Graph memory layers (such as Mem0, Graphiti, or Cognee) extract conversational facts as structured subject-relation-object triplets to build semantic profile graphs." → SRC-002 (plugins/memory/supermemory/README.md, lines 71-72; evermind.ai / mem0.ai docs) └─ Used in: 05_agent_memory/agent_scratchpads_and_session_memory.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-206] "Knowledge Graph stores resolve contradictory facts by traversing timestamped edges and reduce context token consumption up to 90% by substituting raw transcript blocks with relation maps." → evermind.ai / graphiti.ai docs (Zep Graphiti temporal graph edges and compaction data) └─ Used in: 05_agent_memory/agent_scratchpads_and_session_memory.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-207] "Nous Hermes supports local graph-based session recall via its Hindsight memory provider plugin, which executes multi-strategy semantic search and entity graph traversal queries." → SRC-002 (plugins/memory/hindsight/init.py, lines 3-4, 318, 622; README.md, lines 130-131) └─ Used in: 05_agent_memory/agent_scratchpads_and_session_memory.md, 19_final_reports/harness_architecture_specification_report.md

Agent Self-Improvement & Curation Loops (SRC-002, SRC-020, SRC-029, SRC-030)

[CLAIM-208] "Nous Hermes Curator uses an inactivity checker (defaulting to 7 days intervals and 2 hours idle time) rather than a cron daemon to execute background tasks on a separate prompt cache without mutating active sessions." → SRC-002 (agent/curator.py, lines 19-24, 146-160) └─ Used in: 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-209] "Telemetry counts view, use, and patch triggers for agent skills, writing to a sidecar JSON file (.usage.json) rather than file frontmatter to keep user content clear of log chatter." → SRC-002 (tools/skill_usage.py, lines 1-17, 457-473) └─ Used in: 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-210] "Skills created by background forks of the agent are marked with 'created_by': 'agent' or 'agent_created': true inside .usage.json, limiting Curator's scope strictly to agent-created code." → SRC-002 (agent/curator.py, lines 141-160) └─ Used in: 09_skills_md/self_improving_agents_and_learning_loops.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-211] "Unused agent-created skills transition deterministically from active to stale after stale_after_days (default 30) and are archived to ~/.hermes/skills/.archive/ after archive_after_days (default 90)." → SRC-002 (agent/curator.py, lines 32-34, 276-287) └─ Used in: 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-212] "A skill pin protects it from auto-transitions and deletion tool calls (skill_manage(action='delete')), while still permitting updates/patches so the agent can fix coding errors without unpinning." → SRC-002 (tools/skill_manager_tool.py, lines 183-189, 211-235) └─ Used in: 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-213] "Curation LLM consolidation maps candidate skills to cluster groups, merging overlaps into existing or new class-level umbrella skills (e.g. python-helpers), and demoting session bugfixes to subdirectory files like references/ or scripts/." → SRC-002 (agent/curator.py, lines 35-40, 365-433) └─ Used in: 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-214] "Consolidation requires package integrity, ensuring that relative links, asset folders, and support files inside the source skill are either archived together or re-homed and rewritten inside the target umbrella." → SRC-002 (agent/curator.py, lines 35-40, 434-452) └─ Used in: 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-215] "Before execution, agent-created skills are scanned using static AST analysis and vulnerability checkers, blocking executions when threat patterns are resolved." → SRC-002 (tools/skill_manager_tool.py, lines 50-56, 78-102) └─ Used in: 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-216] "The curator takes tarball snapshots (skills.tar.gz) pre-run under .curator_backups/ alongside a manifest.json file, allowing developers to roll back unintended merges and prunings." → SRC-002 (agent/curator.py, lines 112-135) └─ Used in: 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md, 18_architecture_recommendations/README.md, 19_final_reports/harness_architecture_specification_report.md

[CLAIM-217] "Test-Time Self-Improvement (TT-SI) and Recursive Introspection (RISE) utilize execution trace analysis and self-rewards to modify prompt routing and patch code logic on error." → SRC-029, SRC-030 └─ Used in: 09_skills_md/README.md, 09_skills_md/self_improving_agents_and_learning_loops.md