Skip to content

Landscape register

Primary-source comparison of execution and orchestration mechanisms relevant to UH. Each row records a public source URL, version or commit where available, retrieval date, quotation and adoption rationale. The machine-readable coordinator_check field records the source-quotation check, not product acceptance or a guarantee of current upstream behavior. Re-verify monthly or when the row’s recheck_trigger applies.

Rows: 118. Source check: 102 verified, 16 unverified.

System Mechanism Source Version Stance Check Recheck
Claude Code Custom subagents declare a model, tools, permission mode, maxTurns, effort, and other frontmatter; each runs in an independent context window.
“The --agents flag accepts JSON with a prompt field plus these [frontmatter] fields: … maxTurns … effort …”
https://code.claude.com/docs/en/sub-agents (, 2026-09-15) v2.1.261+ (version references visible; page otherwise current) adopt: Declarative per-worker model, tools, effort and turn cap map well; budgets beyond turns are absent. unverified monthly
OpenAI Codex CLI Codex supports subagents for focused delegated work and configuration profiles can select model and effort; the fetched CLI page exposes delegation but not per-worker budgets.
“Split up a larger investigation — subagents: Ask Codex to delegate focused work to specialized agents, then bring their findings back into the main terminal session.”
https://learn.chatgpt.com/docs/codex/cli (, 2026-09-15) current docs (meta version not displayed) adapt: Delegation exists, but a complete per-worker contract is not documented on this page. verified monthly
oh-my-pi The README describes first-class task fan-out into isolated worktrees, typed results, and per-subagent tool surfaces.
“Split a job across workers and get typed results back.”
https://github.com/can1357/oh-my-pi (, 2026-09-15) main README as fetched; moving branch adopt: Typed worker outputs and isolation directly map to UH contracts; the README does not enumerate all budget fields. verified monthly
SWE-agent SWE-agent is governed by a single YAML configuration and accepts a user-selected language model; per-worker budget/seed contract is not exposed in README.
“Configurable & fully documented: Governed by a single yaml file”
https://github.com/SWE-agent/SWE-agent (, 2026-09-15) main README as fetched; project warns mini-SWE-agent supersedes it adapt: Central configuration is a useful seed for worker contracts, but required fields are not unified. verified monthly
Paperclip Paperclip models agent roles, reporting lines, responsibilities, goals, and task ancestry. Issues carry company, project, goal, and parent links so assigned agents receive the organizational why.
“Goal Alignment — Every task traces back to the organization mission. Agents know what to do and why.”
https://raw.githubusercontent.com/paperclipai/paperclip/master/README.md (repo_source, 2026-09-15) master @ 6cfe4acff7284fc330e3e24ef0f4b0ac878628c4 adapt: Adopt goal ancestry and roles, but add explicit per-worker budgets and seed fields. verified monthly
Inngest AgentKit AgentKit networks combine agents, shared state, and a router; each agent may use its own model and provider. Agent definitions include role-oriented name, description, system prompt, model, and tools.
“Networks are Systems of Agents. Use Networks to create powerful AI workflows by combining multiple Agents.”
https://agentkit.inngest.com/concepts/networks (official_docs, 2026-09-15) docs current; no release pin shown adapt: Reuse composable team roles and model selection, adding explicit effort/time/memory budgets. verified AgentKit network documentation
oh-my-codex OMX provides a workflow layer around Codex with explicit commands for clarification, planning, durable goals, and team execution. It supports model/reasoning flags and worktree launch isolation.
“It keeps Codex as the execution engine and makes it easier to start a stronger Codex session by default and run one consistent workflow from clarification to completion.”
https://raw.githubusercontent.com/Yeachan-Heo/oh-my-codex/main/README.md (repo_source, 2026-09-15) main; README has no release version adapt: Reuse durable objective/checkpoint workflow, adding a formal per-worker contract. unverified oh-my-codex release
System Mechanism Source Version Stance Check Recheck
Claude Code Stop and StopFailure hooks run at turn boundaries, and TeammateIdle can send feedback that keeps a teammate working; the page documents no denial-budget or repeated-failure supervisor.
“TeammateIdle: runs when a teammate is about to go idle. Exit with code 2 to send feedback and keep the teammate working.”
https://code.claude.com/docs/en/hooks (, 2026-09-15) current docs (no release version displayed) adapt: Lifecycle hooks provide a seam, but stop reasons and resume policy remain application logic. unverified monthly
Cursor CLI Cursor has interactive, plan, and ask modes plus non-interactive print mode; sessions can continue/resume, but no denial/failure supervisor loop is documented.
“Resume previous conversations: agent ls, agent resume, agent --continue, agent --resume="chat-id-here"”
https://cursor.com/docs/cli/overview.md (, 2026-09-15) current docs (no release version displayed) adapt: Resume exists, while stop-reason policy and bounded retry are absent. unverified monthly
Hermes Agent Hermes approval timeout denies by default, cron/unattended modes can deny instantly, and its hardline blocklist cannot be overridden.
“When no response is given within the timeout, the command is denied by default (fail-closed).”
https://hermes-agent.nousresearch.com/docs/user-guide/security (, 2026-09-15) current (Docusaurus version=current) adopt: Fail-closed approval and hard stops map directly; repeated-failure/stall supervision remains external. verified monthly
OpenHands Agent Canvas starts conversations and automations, including schedules/webhooks, and Agent Server runs multiple agents; stop/retry supervisor semantics are not stated.
“Create automations and workflows … Run on a schedule or in response to webhook events”
https://github.com/All-Hands-AI/OpenHands (, 2026-09-15) 1.18.0 image example; current repo adapt: Scheduling/dispatch exists but bounded denial/stall supervisor logic is not documented. unverified monthly
Paperclip Paperclip supports agent pause, resume, and terminate controls, plus watchdogs and automatic recovery for orphaned runs. Its governance section describes recovery actions and review gates rather than a denial-budget policy.
“Governance with rollback. Approval gates are enforced, config changes are revisioned, and bad changes can be rolled back safely.”
https://raw.githubusercontent.com/paperclipai/paperclip/master/README.md (repo_source, 2026-09-15) master @ 6cfe4acff7284fc330e3e24ef0f4b0ac878628c4 adapt: Reuse lifecycle controls but retain UH-specific repeated-failure, stall, and hard-stop semantics. verified monthly
Temporal Temporal durably records workflow event history so execution can resume after crashes or outages. Activities have retry policies and heartbeat details can checkpoint progress for a later attempt.
“If a failure occurs, the Workflow Execution can resume from the last recorded event, ensuring that progress isn’t lost.”
https://docs.temporal.io/temporal (official_docs, 2026-09-15) main @ 097ce3a12f118d209bb7b8e1ebb9e5d96e961eb6 adapt: Adopt durable replay and activity recovery but implement UH-specific denial and stall policies above it. verified Temporal durable execution documentation
LangGraph LangGraph interrupts suspend graph execution and persist exact state in a checkpointer until external input arrives. Re-invoking with Command resume continues the same thread.
“When an interrupt is triggered, LangGraph saves the graph state using its persistence layer and waits indefinitely until you resume execution.”
https://docs.langchain.com/oss/python/langgraph/interrupts (official_docs, 2026-09-15) docs current; main metadata request rate-limited 2026-09-15 adopt: The durable pause/resume primitive maps directly to supervisor intervention and human gates. verified LangGraph interrupts documentation
Mastra Mastra workflows compose typed steps and provide built-in suspension, resumption, streaming results, and time travel for replaying or retrying steps. Workflow state is shared and can persist across suspend/resume.
“Running workflows to execute the entire sequence, with built-in support for suspension, resumption, and streaming results.”
https://mastra.ai/docs/workflows/overview (official_docs, 2026-09-15) main @ b61b64dfa4341d18d2e22ccb2435020024750dcd adapt: Reuse typed workflow suspension but supply UH supervisor stop reasons and denial budgets. verified Mastra workflows documentation
Inngest AgentKit AgentKit networks run a router-agent loop and expose maxIter to cap iterations; the router may return undefined to stop. Shared state records outputs for later routing.
“A Network can specify an optional maxIter setting to limit the number of iterations.”
https://agentkit.inngest.com/concepts/networks (official_docs, 2026-09-15) docs current; no release pin shown adapt: Iteration caps help bound loops but do not encode UH’s stop reason taxonomy or resume notes. verified AgentKit network documentation
GSD-2 / GSD Pi The original gsd-build/gsd-2 repository is an archive redirect; current GSD Pi describes autonomous workflows that plan, implement, verify, and advance milestones and tasks. It stores summaries and validation evidence locally.
“This repository is no longer the active home for GSD 2 development.”
https://raw.githubusercontent.com/gsd-build/gsd-2/main/README.md (repo_source, 2026-09-15) archived redirect; moved to open-gsd/gsd-pi defer: The requested repository is archived, so behavior must be reverified against the moved project before adoption. verified open-gsd/gsd-pi release
oh-my-claudecode OMC’s Team surface uses a staged team-plan, team-prd, team-exec, team-verify, and team-fix loop. It also documents persistent execution and a lead-session responsibility for final approval.
“Team runs as a staged pipeline: team-plan → team-prd → team-exec → team-verify → team-fix (loop).”
https://raw.githubusercontent.com/Yeachan-Heo/oh-my-claudecode/main/README.md (repo_source, 2026-09-15) v4.1.7 (version stated in README) adapt: Reuse staged execution and verification, adding explicit denial and stall budgets. verified oh-my-claudecode release

B3: Per-tool-call guard and tamper detection

Section titled “B3: Per-tool-call guard and tamper detection”
System Mechanism Source Version Stance Check Recheck
Claude Code PreToolUse hooks match tool names and can return permissionDecision deny plus a human-readable reason before execution.
“The PreToolUse event fires. The hook can deny the call, but staying silent doesn’t approve it.”
https://code.claude.com/docs/en/hooks (, 2026-09-15) current docs (no release version displayed) adopt: A per-call deny hook exists; UH still needs path-only normalization and tamper evidence. verified monthly
OpenAI Codex CLI Codex config exposes approval_policy and sandbox_mode; sandbox policy distinguishes read-only, workspace-write, and danger-full-access.
“sandbox_mode
read-only workspace-write danger-full-access Sandbox policy for filesystem and network access during command execution.” https://learn.chatgpt.com/docs/config-file/config-reference (, 2026-09-15)
Cursor CLI The CLI exposes a sandbox flag with enabled/disabled modes and a menu to control network access, but no documented per-tool path matcher.
“Configure command execution settings with /sandbox or the --sandbox <mode> flag (enabled or disabled).”
https://cursor.com/docs/cli/overview.md (, 2026-09-15) current docs (no release version displayed) adapt: Coarse sandbox toggle only; UH needs path-only matching and deny explanations. verified monthly
oh-my-pi Stream rules use regex matches to abort a stream, inject a rule, and retry from the same point.
“A regex match aborts the stream mid-token, injects the rule as a system reminder, and retries from the same point.”
https://github.com/can1357/oh-my-pi (, 2026-09-15) main README as fetched; moving branch adopt: Tamper/course-correction seam exists; per-tool path guard must be separately verified. verified monthly
Hermes Agent Hermes applies case-insensitive fnmatch deny globs before YOLO/approval and returns BLOCKED without execution; it warns this is not a complete OS sandbox.
“A denied command returns a BLOCKED error to the agent telling it not to retry or rephrase. Nothing runs.”
https://hermes-agent.nousresearch.com/docs/user-guide/security (, 2026-09-15) current (Docusaurus version=current) adopt: Substitute-action deny reason and layered guard are strong; path-only filesystem matching is not the mechanism. verified monthly
goose The public goose README establishes a native agent that can install, execute, edit, and test; no public fetched mechanism in scope documented path-level tool guards.
“install, execute, edit, and test with any LLM”
https://github.com/block/goose (, 2026-09-15) main README as fetched; moving branch defer: Execution exists but the requested deny-reason/path matcher contract was not evidenced. verified monthly
Model Context Protocol MCP defines tools/list and tools/call, schemas, annotations, and an explicit warning that tool annotations are untrusted unless from trusted servers. It recommends a human able to deny invocations.
“For trust & safety and security, there SHOULD always be a human in the loop with the ability to deny tool invocations.”
https://modelcontextprotocol.io/specification/2025-06-18/server/tools (official_docs, 2026-09-15) 2025-06-18 specification adapt: Use MCP tool discovery and denial semantics, adding UH path-only matching and tamper detection. verified MCP specification revision
Landscape gap (searched scope) Across Paperclip, Orca, Temporal, LangGraph, Mastra, AgentKit, OTel, Langfuse, AgentOps, Braintrust, Promptfoo, SWE-bench, AgentEvals, MCP, GSD Pi, OMC, and OMX, the fetched mechanisms expose tools, approval, or roots but no documented combination of path-only matching, substitute-action deny reasons, and tamper detection.
“Clients MUST consider tool annotations to be untrusted unless they come from trusted servers.”
https://modelcontextprotocol.io/specification/2025-06-18/server/tools (official_docs, 2026-09-15) 2025-06-18 specification landscape_gap: Searched the full named scope for path-only guards and tamper detection; found no system documenting all three together. verified monthly

B4: Credential scoping, deny-read, preflight

Section titled “B4: Credential scoping, deny-read, preflight”
System Mechanism Source Version Stance Check Recheck
Claude Code Sandbox credential rules protect declared files and environment variables, while denyRead/allowRead path rules control filesystem visibility.
“The sandbox.credentials setting declares credential files and environment variables to protect from sandboxed commands.”
https://code.claude.com/docs/en/sandboxing (, 2026-09-15) v2.1.187+ for credentials; current page adapt: Credential and path policy exist, but per-worker identity/scoping and controller preflight are not one contract. verified monthly
OpenAI Codex CLI workspace-write can add writable roots, and project config is loaded only for trusted projects; this scopes filesystem access but is not a credential-per-worker scheme.
“Codex loads project-scoped config files only when you trust the project.”
https://learn.chatgpt.com/docs/config-file/config-reference (, 2026-09-15) current docs (meta version not displayed) adapt: Trust and writable-root controls are useful; credential deny-read preflight remains missing. verified monthly
Hermes Agent Hermes lists MCP credential filtering and hardened backend path allowlists among its security layers.
“MCP credential filtering — environment variable isolation for MCP subprocesses”
https://hermes-agent.nousresearch.com/docs/user-guide/security (, 2026-09-15) current (Docusaurus version=current) adapt: Credential filtering is present but per-worker scoping/preflight evidence is not shown. verified monthly
alibaba/OpenSandbox OpenSandbox has a credential vault for secure outbound injection without exposing real secrets to workloads.
“Credential Vault: Secure credential injection for sandbox outbound requests without exposing real secrets to workloads.”
https://github.com/opensandbox-group/OpenSandbox (, 2026-09-15) main README as fetched; moving branch adopt: Secret injection boundary is explicit; per-worker identity/preflight remains to integrate. verified monthly
Paperclip Paperclip has instance and company secrets, encrypted local storage, and scoped secret injection for explicitly needing runs. It describes roles, permissions, and company boundaries for agents.
“Sensitive values stay out of prompts unless a scoped run explicitly needs them.”
https://raw.githubusercontent.com/paperclipai/paperclip/master/README.md (repo_source, 2026-09-15) master @ 6cfe4acff7284fc330e3e24ef0f4b0ac878628c4 adapt: Use scoped secret injection as a baseline, adding UH fail-closed path preflight and deny-read checks. verified monthly
Model Context Protocol MCP roots let clients expose filesystem boundaries to servers and require clients to validate root URIs and permissions. The security guidance says servers should respect root boundaries and validate paths.
“Roots define the boundaries of where servers can operate within the filesystem.”
https://modelcontextprotocol.io/specification/2025-06-18/client/roots (official_docs, 2026-09-15) 2025-06-18 specification adapt: Roots inform worker scope, but UH still needs per-worker credentials and sensitive-path deny-read. verified MCP specification revision

B5: Deliverable scans and completion marker

Section titled “B5: Deliverable scans and completion marker”
System Mechanism Source Version Stance Check Recheck
Pi Pi documents a session/state-management core and its repository highlights tests, tooling, and reproducible release source archives, but no deliverable scan contract.
“AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI”
https://github.com/badlogic/pi-mono (, 2026-09-15) main README as fetched; moving branch defer: The toolkit is extensible but the named deliverable invariants are not a built-in mechanism. verified monthly
SWE-agent SWE-agent targets automated fixes in real repositories and benchmark batch mode; the fetched README does not specify artifact invariant scans.
“SWE-agent takes a GitHub issue and tries to automatically fix it”
https://github.com/SWE-agent/SWE-agent (, 2026-09-15) main README as fetched; project warns mini-SWE-agent supersedes it defer: Deliverable production exists, but UH scan invariants are not evidenced. verified monthly
Aider Aider automatically lints/tests after edits and can fix detected problems; this is validation but not the requested path/secret/last-line deliverable scan.
“Automatically lint and test your code every time aider makes changes.”
https://github.com/Aider-AI/aider (, 2026-09-15) main README as fetched; moving branch adapt: Post-edit validation maps partially to deliverable scans. verified monthly
Turso AgentFS AgentFS records every file operation, tool call, and state change in SQLite, enabling queryable history and audit-oriented scans.
“Every file operation, tool call, and state change is recorded in a SQLite database file.”
https://github.com/tursodatabase/agentfs (, 2026-09-15) main README as fetched; moving branch adopt: Audit trail supports deliverable verification, though named scan rules are UH-owned. verified monthly
Promptfoo Promptfoo evaluates prompts, models, and applications through declarative test cases, configurable metrics, matrix views, and structured results. It also supports automated red-team probes and risk reports.
“promptfoo is an open-source CLI and library for evaluating and red-teaming LLM apps.”
https://www.promptfoo.dev/docs/ (official_docs, 2026-09-15) docs current (meta version current) adapt: Adopt declarative deliverable/evaluation checks but add UH-specific path, secret, and terminal-line scans. unverified Promptfoo docs current release
SWE-bench SWE-bench evaluates generated patches against real GitHub issues in reproducible Docker environments. Its v5 CLI records run IDs and evaluation logs/results for later reporting.
“SWE-bench is a benchmark for evaluating large language models on real world software issues collected from GitHub.”
https://raw.githubusercontent.com/SWE-bench/SWE-bench/main/README.md (repo_source, 2026-09-15) main; v5 CLI visible in README adapt: Use containerized reproducibility and run IDs as a model for deliverable validation evidence. verified SWE-bench CLI release
GSD-2 / GSD Pi GSD Pi documents artifacts such as plans, summaries, validation notes, and reports, plus structured JSON quick-task results and meaningful exit codes. It maintains local .gsd state and reviewable markdown projections.
“Produces artifacts such as plans, summaries, validation notes, and reports.”
https://raw.githubusercontent.com/open-gsd/gsd-pi/main/README.md (repo_source, 2026-09-15) v1.20.0 adapt: Use artifact and validation-note concepts while adding UH terminal-line and secret-like scans. verified GSD Pi release
Landscape gap (searched scope) The named systems provide evaluation, tracing, or artifact reports, but the fetched sources do not document one deliverable scan combining dangling paths, fingerprint phrases, secret-like strings, stray files, and a required DONE/BLOCKED final line.
“Score outputs automatically by defining metrics.”
https://www.promptfoo.dev/docs/ (official_docs, 2026-09-15) docs current (meta version current) landscape_gap: Searched named evaluation and orchestration systems; found no source documenting the complete filesystem-and-terminal deliverable scan. unverified monthly
System Mechanism Source Version Stance Check Recheck
Landscape search No scoped system source fetched documents cost attribution by stop reason as a required first-class field; nearest systems expose generic usage/billing or sandbox lifecycle only.
“Agent Hub … roster shows current activity and usage for every subagent.”
https://github.com/can1357/oh-my-pi (, 2026-09-15) main README as fetched; moving branch landscape_gap: Searched: cost attribution stop reason, usage by denial/stall/OOM, per-worker billing; no scoped source showed the complete join. unverified monthly
Paperclip Paperclip tracks token and cost data by company, agent, project, goal, issue, provider, and model. Budget policies have warning thresholds and hard stops, and overspend pauses agents and cancels queued work.
“Token and cost tracking by company, agent, project, goal, issue, provider, and model.”
https://raw.githubusercontent.com/paperclipai/paperclip/master/README.md (repo_source, 2026-09-15) master @ 6cfe4acff7284fc330e3e24ef0f4b0ac878628c4 adapt: Reuse attribution dimensions but add mandatory stop-reason attribution and unknown-cost states. verified monthly
Langfuse Langfuse records usage and USD cost per LLM generation, either ingested from provider responses or inferred from model definitions. Ingested values take priority, and model definitions can include pricing tiers.
“Langfuse tracks the usage and cost of every LLM call in your application.”
https://langfuse.com/docs/observability/features/token-and-cost-tracking (official_docs, 2026-09-15) main @ 8505c9e320bd4071de7a39b23245ea94fed932ab adapt: Adopt measured-versus-inferred precedence but add stop-reason dimensions and unknown never zero. verified Langfuse cost tracking documentation
AgentOps AgentOps automatically instruments supported agent frameworks and records each execution as a session. Its dashboard presents execution duration, event breakdowns, LLM calls, and traces for debugging.
“After setting up AgentOps, each execution of your program is recorded as a session.”
https://docs.agentops.ai/v2/introduction (official_docs, 2026-09-15) v2 docs current; repo main metadata checked 2026-09-15 adapt: Use session-level observability as evidence, but require explicit provider cost truth and stop reasons. verified AgentOps v2 documentation
System Mechanism Source Version Stance Check Recheck
Landscape search No scoped system source fetched derives a memory cap from declared parallelism; resource limits are generally explicit per sandbox or platform.
“resourceLimits … memory …”
https://raw.githubusercontent.com/opensandbox-group/OpenSandbox/main/docs/guides/secure-container.md (, 2026-09-15) main (raw source, moving) landscape_gap: Searched: memory cap derived parallelism, worker concurrency memory formula, agent budget allocation; found explicit limits, not derivation. unverified monthly
Landscape gap (searched scope) The fetched orchestration systems expose iteration caps, state, queues, or concurrency settings, but none documents deriving a worker memory cap from declared parallelism as a first-class policy.
“A Network can specify an optional maxIter setting to limit the number of iterations.”
https://agentkit.inngest.com/concepts/networks (official_docs, 2026-09-15) docs current; no release pin shown landscape_gap: Searched all named orchestration systems for parallelism-derived memory accounting; found iteration/concurrency controls but no such derivation. verified monthly
System Mechanism Source Version Stance Check Recheck
Claude Code Agent teams use independent context windows, direct teammate messaging, a shared task list, and hooks for TeammateIdle/TaskCompleted quality gates.
“Agent teams let you coordinate multiple Claude Code instances working together.”
https://code.claude.com/docs/en/agent-teams (, 2026-09-15) v2.1.178 (explicitly stated) adopt: Multiple workers and wave-like task coordination are native; model-family diversity is configurable but not enforced. verified monthly
OpenAI Codex CLI The config reference supports review_model and the CLI supports dedicated review against changes without modifying the tree.
“Run a dedicated review against uncommitted changes, a commit, or a base branch. Codex reports prioritized findings without modifying your working tree.”
https://learn.chatgpt.com/docs/codex/cli (, 2026-09-15) current docs (meta version not displayed) adopt: Review is first-class; multiple model-family reviewer waves are not guaranteed. verified monthly
Cursor CLI Print mode can review changes with an output format, and Cloud Agent handoff continues work asynchronously.
“Use print mode for non-interactive scenarios like scripts, CI pipelines, or automation.”
https://cursor.com/docs/cli/overview.md (, 2026-09-15) current docs (no release version displayed) defer: Review/remote continuation are present, but reviewer breadth and independent model families are not described. verified monthly
oh-my-pi A reviewer model runs in an advisor role, reads every main-agent turn, and injects notes inline.
“Pair a reviewer model to the ‘advisor’ role and it reads every turn the main agent takes, injecting notes inline.”
https://github.com/can1357/oh-my-pi (, 2026-09-15) main README as fetched; moving branch adopt: Independent reviewer is native; wave packets and multiple model families need adaptation. verified monthly
OpenHands Agent Canvas can connect to OpenHands, Claude Code, Codex, Gemini, or any ACP agent backend, enabling heterogeneous review workers.
“Use with OpenHands, Claude Code, Codex, Gemini, or any agent with Agent-Client Protocol (ACP).”
https://github.com/All-Hands-AI/OpenHands (, 2026-09-15) 1.18.0 image example; current repo adapt: Backend diversity is native; review-wave breadth and packet contracts are not. verified monthly
Mastra Mastra scorers evaluate agent outputs and workflow-step outputs with model-graded, rule-based, and statistical methods. Sampling, trace evaluation, stored scores, and verdict gates support continuous quality checks.
“Scorers are automated tests that evaluate Agents outputs using model-graded, rule-based, and statistical methods.”
https://mastra.ai/docs/evals/overview (official_docs, 2026-09-15) main @ b61b64dfa4341d18d2e22ccb2435020024750dcd adapt: Use scorer and gate ideas, extending them to multiple independent reviewer families and wave packets. verified Mastra evals documentation
Braintrust Braintrust evaluations define data, a task, and scoring functions, then compare approaches and catch regressions. The platform supports traces, human feedback, datasets, and experiments in a workflow.
“Each evaluation consists of three components: Data, Task, Scores.”
https://www.braintrust.dev/docs/evaluation-quickstart (official_docs, 2026-09-15) docs current; no release pin shown adapt: Use evaluator structure and regression comparisons, extending to independent reviewer model families. unverified Braintrust evaluation documentation
Promptfoo Promptfoo’s GitHub Action compares before and after evaluations on pull requests and posts result links. Its red-team documentation describes systematic adversarial generation, automated scoring, and CI/CD integration.
“On every pull request that modifies a prompt, the action will automatically run a full comparison.”
https://www.promptfoo.dev/docs/integrations/github-action/ (official_docs, 2026-09-15) docs current (meta version current) adapt: Reuse before/after quality evidence, adding multiple reviewer breadth and wave-level packets. verified Promptfoo GitHub Action release
SWE-bench SWE-bench provides gold/reference and prediction evaluation modes, result caching by run and instance IDs, and parallel workers. It measures patch resolution rather than reviewer diversity.
“The run summary is written to logs/evaluation/<run_id>/results.json.”
https://raw.githubusercontent.com/SWE-bench/SWE-bench/main/README.md (repo_source, 2026-09-15) main; v5 CLI visible in README adapt: Adopt reproducible result packets while adding independent reviewer model-family breadth. verified SWE-bench CLI release
AgentEvals AgentEvals evaluates agent trajectories, including intermediate tool-call steps, with strict/unordered/subset/superset matching and LLM-as-judge evaluators. Results include a score and reasoning/comment.
“This package contains a collection of evaluators and utilities for evaluating the performance of your agents, with a focus on agent trajectory.”
https://raw.githubusercontent.com/langchain-ai/agentevals/main/README.md (repo_source, 2026-09-15) main; commit metadata rate-limited 2026-09-15 adopt: Trajectory evaluation directly supports review breadth and tool-structure evidence. verified AgentEvals release
oh-my-claudecode OMC supports teams of Codex, Gemini, Antigravity, Grok, Cursor, and Claude workers and assigns role-oriented tasks such as critic, code-reviewer, security-reviewer, and test-engineer. Final approval remains a lead-session responsibility.
“Cursor also supports reviewer-style roles (critic, code-reviewer, security-reviewer, test-engineer).”
https://raw.githubusercontent.com/Yeachan-Heo/oh-my-claudecode/main/README.md (repo_source, 2026-09-15) v4.1.7 (version stated in README) adapt: Use distinct reviewer roles and model families as a breadth precedent, adding wave packet schemas. verified oh-my-claudecode release
oh-my-codex OMX’s canonical workflow includes plan, ultragoal, team, code-review, and ultraqa surfaces, with team used when a story needs coordinated parallel work. It recommends worktree isolation for concurrent sessions.
“Use $team inside that execution path only when a specific Ultragoal story needs coordinated parallel work.”
https://raw.githubusercontent.com/Yeachan-Heo/oh-my-codex/main/README.md (repo_source, 2026-09-15) main; README has no release version adapt: Adopt selective team shapes and worktree isolation, adding independent reviewer-family requirements. verified oh-my-codex release
System Mechanism Source Version Stance Check Recheck
Claude Code In plan mode a teammate sends a plan approval request to the lead, but Claude Code approves it as soon as it arrives without lead review.
“When a teammate finishes planning, it sends a plan approval request to the lead. Claude Code approves the plan in the lead’s session as soon as the request arrives, without the lead reviewing it.”
https://code.claude.com/docs/en/agent-teams (, 2026-09-15) current docs (version references visible) reject: This explicitly fails a human promotion gate. verified monthly
OpenAI Codex CLI approval_policy controls pauses before command execution and approvals_reviewer can be user or auto_review.
“approvals_reviewer
user auto_review Who reviews eligible approval prompts under on-request or granular approval policies.” https://learn.chatgpt.com/docs/config-file/config-reference (, 2026-09-15) current docs (meta version not displayed)
Paperclip Paperclip provides board approval workflows, review/approval stages, decision tracking, and audit logging. It presents approval gates as enforced governance rather than merely advisory review.
“Governance & Approvals — Board approval workflows, execution policies with review/approval stages, decision tracking, budget hard-stops.”
https://raw.githubusercontent.com/paperclipai/paperclip/master/README.md (repo_source, 2026-09-15) master @ 6cfe4acff7284fc330e3e24ef0f4b0ac878628c4 adopt: The enforced human-gate model directly matches promotion authority requirements. verified monthly
LangGraph LangGraph documents approval workflows by interrupting before critical actions and routing on a true or false resume payload. The framework exposes the pending payload to a caller for human review.
“Approval workflows: Pause before executing critical actions (API calls, database changes, financial transactions).”
https://docs.langchain.com/oss/python/langgraph/interrupts (official_docs, 2026-09-15) docs current; main metadata request rate-limited 2026-09-15 adapt: Use the interrupt pattern but add approver identity and immutable approval records. verified LangGraph interrupts documentation
Model Context Protocol MCP sampling requires clients to control model access, selection, and permissions, with human review before request forwarding and response delivery. Clients should return a user-rejected error.
“There SHOULD always be a human in the loop with the ability to deny sampling requests.”
https://modelcontextprotocol.io/specification/2025-06-18/client/sampling (official_docs, 2026-09-15) 2025-06-18 specification adopt: MCP’s explicit approval boundary supports rejecting auto-promotion under human gates. verified MCP specification revision
System Mechanism Source Version Stance Check Recheck
Landscape search No scoped system source fetched defines liveness from pending tool-call structure rather than event freshness; most lifecycle docs use idle/timeout/events.
“A Sandbox is considered active if any of the following are true: … an active command … stdin … open TCP connection”
https://modal.com/docs/guide/sandboxes.md (, 2026-09-15) current docs (no release version displayed) landscape_gap: Searched: pending tool call liveness, event freshness heartbeat, stalled agent detector; found activity/timeout signals, not pending-structure liveness. unverified monthly
Temporal Temporal Activity workers report heartbeats, and the service retains the last heartbeat details for the next retry attempt. Heartbeats are progress/checkpoint signals, not proof that an arbitrary worker is making useful progress.
“The last recorded Heartbeat details are made available to the Activity function on the next attempt.”
https://docs.temporal.io/activities (official_docs, 2026-09-15) main @ 097ce3a12f118d209bb7b8e1ebb9e5d96e961eb6 adapt: Use heartbeats as one input, but derive UH liveness from pending tool-call structure. verified Temporal activity heartbeat documentation
System Mechanism Source Version Stance Check Recheck
Claude Code In-process and split-pane teams let the lead select a teammate and type a direct message; Ctrl+T exposes the task list.
“Each teammate is a full, independent Claude Code session. You can message any teammate directly to give additional instructions, ask follow-up questions, or redirect their approach.”
https://code.claude.com/docs/en/agent-teams (, 2026-09-15) current docs (version references visible) adopt: Direct steering is documented and interactive. verified monthly
Cursor CLI Interactive mode lets a user approve commands and Cloud Agent handoff moves the conversation to a running cloud agent.
“Start a conversational session with the agent to describe your goals, review proposed changes, and approve commands.”
https://cursor.com/docs/cli/overview.md (, 2026-09-15) current docs (no release version displayed) adapt: User steering/approval is interactive, but a documented running-worker control channel is thin. verified monthly
oh-my-pi Agent Hub shows activity/usage, live transcripts, and controls to steer, revive, or kill workers.
“Open one to read its live transcript, type a steering message, revive a parked worker, or kill a stuck one.”
https://github.com/can1357/oh-my-pi (, 2026-09-15) main README as fetched; moving branch adopt: Direct steering and lifecycle control are explicit. verified monthly
Hermes Agent Hermes supports Ctrl+C/new-message interrupt-and-redirect in CLI and stop/new messages on gateway platforms.
“Interrupt current work
Ctrl+C or send a new message” https://github.com/NousResearch/hermes-agent (, 2026-09-15) main README as fetched; moving branch adopt: A steering path exists across CLI and messaging. verified
System Mechanism Source Version Stance Check Recheck
oh-my-pi The project publishes native install paths for macOS, Linux, and Windows and says completions derive from live CLI metadata.
“macOS · Linux · Windows · bun ≥ 1.3.14”
https://github.com/can1357/oh-my-pi (, 2026-09-15) main README as fetched; moving branch adopt: Cross-platform installation/path handling is explicit; artifact path semantics still need UH normalization. verified monthly
OpenHands The Docker backend guide gives distinct macOS/Linux and Windows PowerShell mount commands, mapping host project roots to /projects.
“The backend … can access any project under the mounted /projects path.”
https://docs.openhands.dev/openhands/usage/agent-canvas/backend-setup/docker (, 2026-09-15) current docs (no release version displayed) adopt: Cross-platform artifact/workspace mapping is documented. unverified monthly
Aider Aider provides Windows CI and documents terminal usage; the fetched README focuses on repo editing and Git rather than portable artifact-path normalization.
“aider is AI pair programming in your terminal”
https://github.com/Aider-AI/aider (, 2026-09-15) main README as fetched; moving branch defer: Cross-platform use is implied by project workflows, but the UH artifact contract is not documented. verified monthly
Landscape search No single scoped system source defines a portable artifact-path contract spanning Windows, POSIX, worktrees, and sandbox mounts; products document platform-specific mounts instead.
“PowerShell uses backticks (`) for line continuation instead of backslashes.”
https://docs.openhands.dev/openhands/usage/agent-canvas/backend-setup/docker (, 2026-09-15) current docs (no release version displayed) landscape_gap: Searched: cross-platform artifact path contract, Windows POSIX path normalization, portable worktree artifact; found platform-specific instructions, not one contract. verified monthly
Model Context Protocol MCP roots require file:// URIs and can list multiple repositories, making filesystem boundaries transportable as URI values. The spec requires clients to validate root URIs against path traversal.
“This MUST be a file:// URI in the current specification.”
https://modelcontextprotocol.io/specification/2025-06-18/client/roots (official_docs, 2026-09-15) 2025-06-18 specification adapt: Use URI roots for portable artifacts, adding platform-specific normalization and canonicalization. verified MCP specification revision

X1: Owned process tree and kill on controller loss

Section titled “X1: Owned process tree and kill on controller loss”
System Mechanism Source Version Stance Check Recheck
E2B E2B sandboxes expose timeout, pause/resume, kill, and automatic pause on expiry; the lifecycle ends on timeout or OOM.
“When their timeout expires, they can automatically pause to save resources — preserving their full state so you can resume at any time.”
https://docs.e2b.dev/sandbox.md (, 2026-09-15) current docs (no release version displayed) adopt: Lifecycle/resource termination is native; controller-loss process ownership is not guaranteed by this source. verified monthly
alibaba/OpenSandbox OpenSandbox command API accepts per-sandbox resource limits and lifecycle timeout, with distributed scheduling through Docker/Kubernetes runtimes.
“Built-in lifecycle management … local runs and large-scale distributed scheduling.”
https://github.com/opensandbox-group/OpenSandbox (, 2026-09-15) main README as fetched; moving branch adapt: Resource/lifecycle controls exist; owned process tree and controller-loss kill need platform configuration. unverified monthly
Daytona Daytona sandboxes allocate vCPU, RAM, and disk and expose lifecycle APIs; the public repo states development moved private and releases stopped after v0.190.0.
“This repository is no longer maintained.”
https://github.com/daytonaio/daytona (, 2026-09-15) v0.190.0 (license link; repository archived/no further releases) defer: Resource controls exist, but re-verification must move to the private successor or current docs. verified monthly
Modal Sandboxes Modal sandboxes have created/scheduled/started/finished lifecycle states, max lifetime, idle timeout, OOM termination, and explicit terminate.
“Finished — The Sandbox has stopped running … timeout or idle timeout was reached, or an out-of-memory condition occurred.”
https://modal.com/docs/guide/sandboxes.md (, 2026-09-15) current docs (no release version displayed) adopt: Lifecycle and resource stop reasons are explicit; controller-loss ownership is not. unverified monthly
kubernetes-sigs/agent-sandbox Agent Sandbox manages stateful singleton workloads with lifecycle operations including scheduled deletion, pausing, and resuming.
“Lifecycle Management: The Sandbox controller manages the lifecycle of the pod, including creation, scheduled deletion, pausing and resuming.”
https://github.com/kubernetes-sigs/agent-sandbox (, 2026-09-15) v1.0.2 installation example; current repository adopt: Controller lifecycle is native; per-worker memory/kill-on-controller-loss policy is Kubernetes-dependent. verified monthly
Landscape search No scoped system source fetched combines owned controller process-tree tracking, a hard memory cap, and guaranteed kill on controller loss across local and remote backends.
“agent-sandbox follows the Kubernetes controller pattern”
https://github.com/kubernetes-sigs/agent-sandbox (, 2026-09-15) v1.0.2 landscape_gap: Searched: controller loss kill process tree, supervisor-owned process group memory cap, cross-backend kill guarantee; found platform-specific lifecycle controllers only. verified monthly
System Mechanism Source Version Stance Check Recheck
Landscape search No scoped system source fetched attests measured route/model/credential usage as an immutable per-worker record; several expose usage or telemetry separately.
“the roster shows current activity and usage for every subagent”
https://github.com/can1357/oh-my-pi (, 2026-09-15) main README as fetched; moving branch landscape_gap: Searched: route attestation, measured usage attestation, immutable model/provider per worker; no complete evidence chain found. verified monthly
System Mechanism Source Version Stance Check Recheck
Claude Code Bash sandboxing enforces filesystem/network restrictions at the OS level; failIfUnavailable can convert unsupported/missing sandbox into a hard failure.
“By default, if the sandbox cannot start because dependencies are missing or the platform is unsupported, Claude Code shows a warning and runs commands without sandboxing.”
https://code.claude.com/docs/en/sandboxing (, 2026-09-15) current docs (no release version displayed) adapt: Isolation tiers and fail-closed mode exist, but native Windows is unsupported and the default is fail-open. verified monthly
OpenAI Codex CLI Codex provides read-only/workspace-write/danger-full-access sandbox modes plus native Windows sandbox selection.
“windows.sandbox
unelevated elevated Windows-only native sandbox mode when running Codex natively on Windows.” https://learn.chatgpt.com/docs/config-file/config-reference (, 2026-09-15) current docs (meta version not displayed)
Pi Pi explicitly says filesystem, process, network, and credential permissions are inherited from the launching process and recommends external sandboxing/containerization.
“Pi does not include a built-in permission system for restricting filesystem, process, network, or credential access.”
https://github.com/badlogic/pi-mono (, 2026-09-15) main README as fetched; moving branch reject: The documented default is host permissions, not an isolation tier. verified monthly
Hermes Agent Hermes supports local, Docker, SSH, Singularity, Modal, Daytona, and Vercel Sandbox backends; container isolation is a documented layer.
“Runs anywhere, not just your laptop — Seven terminal backends — local, Docker, SSH, Singularity, Modal, Daytona, and Vercel Sandbox.”
https://github.com/NousResearch/hermes-agent (, 2026-09-15) main README as fetched; moving branch adopt: Multiple isolation backends exist, with backend-specific semantics. verified monthly
goose goose runs as a native desktop/CLI/API agent on macOS, Linux, and Windows; the fetched README does not claim a sandbox boundary.
“A native desktop app for macOS, Linux, and Windows. A full CLI for terminal workflows. An API to embed it anywhere.”
https://github.com/block/goose (, 2026-09-15) main README as fetched; moving branch defer: Platform portability is clear, but isolation must be supplied/verified elsewhere. verified monthly
OpenHands Agent Canvas can run an agent backend in Docker with only mounted projects and settings exposed; the guide warns privileged mode weakens isolation.
“The agent server can execute arbitrary shell commands inside the container.”
https://docs.openhands.dev/openhands/usage/agent-canvas/backend-setup/docker (, 2026-09-15) current docs (no release version displayed) adapt: Container boundary exists, with an explicit privileged-mode risk. verified monthly
E2B E2B runs AI-generated code in secure isolated cloud sandboxes controlled through SDKs.
“run AI-generated code in secure isolated sandboxes in the cloud”
https://github.com/e2b-dev/E2B (, 2026-09-15) main README as fetched; moving branch adopt: Cloud isolation is the core product. verified monthly
alibaba/OpenSandbox OpenSandbox supports Docker/Kubernetes runtimes and secure gVisor, Kata, and Firecracker runtimes; server startup validates availability and refuses missing configured runtimes.
“The server validates the configured runtime at startup and will refuse to start if the runtime is unavailable.”
https://raw.githubusercontent.com/opensandbox-group/OpenSandbox/main/docs/guides/secure-container.md (, 2026-09-15) main (raw source, moving) adopt: Fail-closed runtime selection and multiple isolation tiers map well. verified monthly
Daytona Daytona provides isolated composable computers with a dedicated kernel, filesystem, network stack, and allocated resources.
“full composable computers with complete isolation, a dedicated kernel, filesystem, network stack”
https://github.com/daytonaio/daytona (, 2026-09-15) v0.190.0 adopt: Strong sandbox primitive is documented. verified monthly
Modal Sandboxes Modal describes Sandboxes as secure containers for executing untrusted user/agent code and supports runtime creation.
“secure containers for executing untrusted user or agent code on Modal”
https://modal.com/docs/guide/sandboxes.md (, 2026-09-15) current docs (no release version displayed) adopt: Isolation is the core primitive. verified monthly
kubernetes-sigs/agent-sandbox Agent Sandbox is an orchestrator delegating low-level isolation to RuntimeClass-backed gVisor/Kata runtimes.
“It delegates low-level container isolation to secure “Sandbox Runtimes” (like gVisor or Kata Containers)”
https://github.com/kubernetes-sigs/agent-sandbox (, 2026-09-15) v1.0.2 installation example; current repository adopt: Clear separation of orchestration and isolation tier. verified monthly
System Mechanism Source Version Stance Check Recheck
Claude Code SessionStart fires on session begin or resume, and worktree/task state persists for resumed sessions.
“SessionStart
When a session begins or resumes” https://code.claude.com/docs/en/hooks (, 2026-09-15) current docs (no release version displayed) adopt: Resume lifecycle is explicit; UH-specific checkpoint/contract restoration still must be layered. verified
OpenAI Codex CLI The CLI documents codex resume for reopening recent or older local chats and supports repeatable codex exec workflows.
“codex resume: Reopen a recent chat from the current repository, or search across local chats when you need to return to older work.”
https://learn.chatgpt.com/docs/codex/cli (, 2026-09-15) current docs (meta version not displayed) adopt: Session resume is native; supervisor note restoration is outside the documented primitive. verified monthly
Cursor CLI agent resume/continue reopen prior conversations by latest or chat ID.
“Pick up your Cloud Agent tasks on web or mobile at cursor.com/agents.”
https://cursor.com/docs/cli/overview.md (, 2026-09-15) current docs (no release version displayed) adapt: Persistence/resume is available, though durable artifact contracts are not specified. verified monthly
oh-my-pi The README documents on-disk sessions and --resume completion against sessions, plus isolated worktrees for tasks.
“--resume against your on-disk sessions.”
https://github.com/can1357/oh-my-pi (, 2026-09-15) main README as fetched; moving branch adopt: Session persistence is native. verified monthly
Pi Pi is an agent loop with state management and session benchmark/storage packages; the public README also directs users to persistent session tooling.
“AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI”
https://github.com/badlogic/pi-mono (, 2026-09-15) main README as fetched; moving branch adapt: State/session primitives exist, but the exact resume contract requires source-level follow-up. verified monthly
Hermes Agent Hermes provides cross-session isolation and cron persistence paths, while /new, /reset, /retry, and /undo operate session state.
“Cross-session isolation — sessions cannot access each other’s data or state”
https://hermes-agent.nousresearch.com/docs/user-guide/security (, 2026-09-15) current (Docusaurus version=current) adopt: Session boundary is explicit; a supervisor resume note is not standardized. unverified monthly
E2B E2B pauses a sandbox and preserves full state indefinitely for later resume, with explicit timeout control.
“pausing resets the runtime window, and your sandbox’s full state is preserved indefinitely.”
https://docs.e2b.dev/sandbox.md (, 2026-09-15) current docs (no release version displayed) adopt: Persistence/resume is directly documented. verified monthly
Turso AgentFS AgentFS stores files, state, and history in one SQLite file and supports snapshots/restoration for reproducibility.
“Snapshot an agent’s state at any point with cp agent.db snapshot.db. Restore it later to reproduce exact execution states”
https://github.com/tursodatabase/agentfs (, 2026-09-15) main README as fetched; moving branch adopt: Persistence/checkpoint primitive is explicit. verified monthly
Daytona Daytona snapshots preserve state for persistent agent operations across sessions.
“Our stateful environment snapshots enable persistent agent operations across sessions”
https://github.com/daytonaio/daytona (, 2026-09-15) v0.190.0 adopt: Persistence is native, though source continuity is now a recheck risk. verified monthly
Modal Sandboxes Modal recommends Filesystem Snapshots to preserve state beyond 24 hours and restore into a subsequent Sandbox.
“If you need a Sandbox to run for more than 24 hours, we recommend using Filesystem Snapshots to preserve its state”
https://modal.com/docs/guide/sandboxes.md (, 2026-09-15) current docs (no release version displayed) adopt: Checkpoint/resume exists, with explicit lifetime boundary. verified monthly
kubernetes-sigs/agent-sandbox Sandbox CRDs provide stable identity and persistent storage that survive restarts, with pause/resume lifecycle.
“Persistent Storage: Sandboxes can be configured with persistent storage that survives restarts.”
https://github.com/kubernetes-sigs/agent-sandbox (, 2026-09-15) v1.0.2 adopt: Stateful resume primitive is explicit. verified monthly
System Mechanism Source Version Stance Check Recheck
Paperclip Paperclip uses DB-backed wakeup queues and tracked runs that produce structured logs, cost events, session state, and audit trails. Its issues and conversations persist across reboots.
“DB-backed wakeup queue with coalescing, budget checks, workspace resolution, secret injection, skill loading, and adapter invocation.”
https://raw.githubusercontent.com/paperclipai/paperclip/master/README.md (repo_source, 2026-09-15) master @ 6cfe4acff7284fc330e3e24ef0f4b0ac878628c4 adapt: Adopt durable queue and run state, retaining UH’s explicit run/task/attempt identity contract. verified monthly
Temporal Temporal creates durable Activity Executions, dispatches Activity Tasks through queues, and collects results to determine next workflow steps. Retry policies distinguish activity task attempts from workflow execution retries.
“This creates a durable Activity Execution in Temporal that orchestrates the full lifecycle of an Activity.”
https://docs.temporal.io/activities (official_docs, 2026-09-15) main @ 097ce3a12f118d209bb7b8e1ebb9e5d96e961eb6 adopt: Durable execution and attempt-aware retry are strong foundations for UH lifecycle state. verified Temporal activity execution documentation
LangGraph LangGraph checkpointers persist thread-scoped graph state for conversation continuity, human-in-the-loop workflows, time travel, and fault tolerance. A thread_id selects the state to resume.
“Checkpointers persist a thread’s graph state as checkpoints.”
https://docs.langchain.com/oss/python/langgraph/persistence (official_docs, 2026-09-15) docs current; main metadata request rate-limited 2026-09-15 adapt: Adopt checkpointed state while keeping UH’s run/task/attempt IDs separate from a thread ID. verified LangGraph persistence documentation
Mastra Mastra workflow steps have schemas, IDs, state, suspension/resumption, and Studio status/log views. Time travel can replay individual steps after a run completes.
“After a run completes, replay individual steps to inspect or retry them.”
https://mastra.ai/docs/workflows/overview (official_docs, 2026-09-15) main @ b61b64dfa4341d18d2e22ccb2435020024750dcd adapt: Adopt step-level replay but retain explicit immutable attempt records. verified Mastra workflows documentation
Inngest AgentKit AgentKit state stores chronological interaction history and typed data shared by agents, tools, and routers. The network loop uses state to make subsequent routing decisions.
“The history system maintains a chronological record of all Agent interactions in your Network.”
https://agentkit.inngest.com/concepts/state (official_docs, 2026-09-15) docs current; no release pin shown adapt: Use shared typed state for context, but add durable run/task/attempt lifecycle records. verified AgentKit state documentation
AgentOps AgentOps models executions as sessions and records traces, actions, tool calls, errors, and LLM calls for later drilldown. It is primarily an observability record, not a durable supervisor state machine.
“Most powerful of all is the Session Waterfall. On the left, a time visualization of all your LLM calls, Action events, Tool calls, and Errors.”
https://docs.agentops.ai/v2/introduction (official_docs, 2026-09-15) v2 docs current; repo main metadata checked 2026-09-15 defer: Use trace concepts for observability but do not substitute sessions for authoritative lifecycle state. verified AgentOps v2 documentation
GSD-2 / GSD Pi GSD Pi stores planning and runtime state under .gsd, tracks work in a local database with markdown projections, and supports worktree-aware automation. The README identifies v1.20.0 as the latest release.
“Tracks project state in a local database with markdown projections for review.”
https://raw.githubusercontent.com/open-gsd/gsd-pi/main/README.md (repo_source, 2026-09-15) v1.20.0 adapt: Adopt durable local projections, keeping UH’s attempt identity and cross-platform path contract. verified GSD Pi release
oh-my-codex OMX stores plans, logs, memory, and mode tracking under .omx and describes durable goals with checkpoints. Standard sessions use a single writable session pointer and fail closed on same-root conflicts.
“It keeps project guidance, plans, logs, and state in .omx/.”
https://raw.githubusercontent.com/Yeachan-Heo/oh-my-codex/main/README.md (repo_source, 2026-09-15) main; README has no release version adapt: Use durable local state and fail-closed ownership, separating UH run/task/attempt identities. unverified oh-my-codex release
System Mechanism Source Version Stance Check Recheck
Langfuse Langfuse distinguishes ingested usage/cost from inferred cost, supports custom model definitions and pricing tiers, and explicitly warns that missing reasoning-token usage prevents correct inference for some models.
“When both are available, ingested values take priority over inferred ones.”
https://langfuse.com/docs/observability/features/token-and-cost-tracking (official_docs, 2026-09-15) main @ 8505c9e320bd4071de7a39b23245ea94fed932ab adopt: The explicit distinction between measured and inferred values matches UH cost truth requirements. verified Langfuse cost tracking documentation
System Mechanism Source Version Stance Check Recheck
Model Context Protocol The 2025-06-18 MCP specification requires explicit user consent for data access, operations, tools, and sampling. It also defines response errors such as user rejection for sampling.
“Users must explicitly consent to and understand all data access and operations.”
https://modelcontextprotocol.io/specification/2025-06-18 (official_docs, 2026-09-15) 2025-06-18 specification adopt: Adopt explicit consent and denial as first-class approval records. verified MCP specification revision

X8: Evidence and trace schemas without prompts or secrets

Section titled “X8: Evidence and trace schemas without prompts or secrets”
System Mechanism Source Version Stance Check Recheck
Paperclip Paperclip documents opt-in OpenTelemetry server traces and an anonymous telemetry contract that excludes prompts, issue content, file paths, and secrets. Mutating actions and lifecycle events are recorded as durable activity.
“No personal information, issue content, prompts, file paths, or secrets are ever collected.”
https://raw.githubusercontent.com/paperclipai/paperclip/master/README.md (repo_source, 2026-09-15) master @ 6cfe4acff7284fc330e3e24ef0f4b0ac878628c4 adopt: The explicit exclusion list is a strong evidence-schema privacy precedent. verified monthly
OpenTelemetry GenAI semantic conventions The fetched OpenTelemetry page states that GenAI semantic conventions moved to a dedicated repository and is no longer maintained in the old location. This confirms the convention surface is evolving and version-sensitive.
“GenAI semantic conventions have moved to the OpenTelemetry GenAI semantic conventions repository.”
https://opentelemetry.io/docs/specs/semconv/gen-ai/ (official_docs, 2026-09-15) moved page; destination repository current 2026-09-15 defer: Track the moving standard but keep UH evidence schemas versioned and privacy-filtered independently. verified OpenTelemetry semantic-conventions-genai release
Langfuse Langfuse accepts OTLP traces at a documented endpoint, maps OTel attributes to its data model, and cautions against putting sensitive information in propagated baggage. It supports HTTP JSON and protobuf, not gRPC.
“Do not include sensitive information (passwords, API keys, personal data, etc.) in baggage.”
https://langfuse.com/integrations/native/opentelemetry (official_docs, 2026-09-15) main @ 8505c9e320bd4071de7a39b23245ea94fed932ab adapt: Use OTLP ingestion and propagation guidance, but enforce UH prompt/secret exclusion at emission. verified Langfuse OpenTelemetry integration documentation
Braintrust Braintrust traces expose complete inputs and outputs, token counts, latency, cost, model configuration, and metadata in its logs. Its quickstart directs users to configure SDK tracing and verify traces appear.
“Click into any trace to see the complete input and output, token counts, latency, cost, model configuration, and metadata.”
https://www.braintrust.dev/docs/tracing-quickstart (official_docs, 2026-09-15) docs current; no release pin shown adapt: Its trace fields are useful, but UH must exclude prompts and secrets from evidence schemas by design. verified Braintrust tracing documentation
Model Context Protocol MCP security guidance requires user consent before exposing data and limits server visibility into prompts during sampling. It explicitly treats tool descriptions and annotations as potentially untrusted.
“The protocol intentionally limits server visibility into prompts.”
https://modelcontextprotocol.io/specification/2025-06-18 (official_docs, 2026-09-15) 2025-06-18 specification adopt: Use MCP’s consent and prompt-visibility constraints to shape evidence schemas that exclude sensitive content. verified MCP specification revision