SEARCH · ONE CONTENT GRAPH

From one term to its
lessons, evidence, and runs.

This is more than a title filter: it indexes bilingual content, source locators, stable IDs, experiment variables, and trace event types.

144 public records7 entity types2 content languages
144 resultsShowing all public records
  1. Lessons

    The Minimal Agent Loop

    Turn one model call into a loop with an explicit stop condition.

    s01-agent-loopS01agent-loop
  2. Lessons

    Streaming and Events

    Convert streaming deltas into observable, replayable events.

    s02-events-streamingS02agent-loopevent-stream
  3. Lessons

    Tool Registry and Dispatch

    Validate, execute, and record model-requested tool calls.

    s03-tool-dispatchS03agent-looptool-dispatch
  4. Lessons

    File, Shell, and Edit Tools

    Put workspace boundaries and explicit allowlists around side-effecting tools.

    s04-workspace-toolsS04workspace-toolspath-boundary
  5. Lessons

    Instructions and Project Discovery

    Compose hierarchical instructions from the workspace root to the current directory.

    s05-instructionsS05instruction-discovery
  6. Lessons

    Context Budgets and Truncation

    Make input pruning and tool-output truncation explicit policies.

    s06-context-budgetS06context-budgettool-output-truncation
  7. Lessons

    Session Event Logs, Replay, and Branching

    Rebuild derived state from append-only events and create lineage-preserving branches.

    s07-session-replayS07event-streamsession-replay
  8. Lessons

    Context Compaction and Checkpoints

    Replace long visible history with a provenance-bearing semantic checkpoint while retaining the factual log.

    s08-context-compactionS08context-budgetcontext-compaction
  9. Lessons

    Memory and On-Demand Skills

    Move durable facts and full capability instructions out of resident context, retrieving them only when relevant.

    s09-memory-skillsS09memory-retrievalextension-runtime
  10. Lessons

    Approval Policy and Exact Authorization

    Insert fail-closed allow, ask, and deny decisions between model requests and real effects.

    s10-approval-policyS10execution-policy
  11. Lessons

    OS Sandboxes and Network Boundaries

    Compile semantic authorization into a minimal capability envelope enforceable by an isolation backend.

    s11-sandbox-networkS11execution-policyos-sandbox
  12. Lessons

    Project Trust and Prompt Injection

    Separate workspace identity, explicit trust, and instruction-versus-data authority.

    s12-project-trustS12project-trustinstruction-discovery
  13. Lessons

    Git Checkpoints, Diffs, and Rollback

    Create a clean baseline before effects, review the exact diff, and roll back only explicit paths.

    s13-checkpoint-rollbackS13checkpoint-rollbackevent-stream
  14. Mechanisms

    Agent Loop

    Own the repeated transition from user intent to model response, tool effects, updated context, and an explicit stop reason.

    agent-looploopcore
  15. Mechanisms

    Checkpoint, Diff, and Rollback

    Create a version-bound baseline before edits, review a scoped state delta, and restore only the explicitly selected paths when recovery is needed.

    checkpoint-rollbackreliabilitycore
  16. Mechanisms

    Client–Server Protocol

    Expose the harness loop, sessions, events, approvals, and configuration to terminal, IDE, desktop, or automation clients through a stable transport boundary.

    client-server-protocolinterfacescore
  17. Mechanisms

    Context Budget

    Allocate finite input space across instructions, conversation, tools, memory, and output reserve.

    context-budgetcontextcore
  18. Mechanisms

    Context Compaction

    Replace an over-budget model-visible history with a smaller projection while preserving enough state, provenance, and recent work to continue safely.

    context-compactioncontextcore
  19. Mechanisms

    Event Stream

    Translate provider deltas and harness transitions into ordered, replayable events.

    event-streamobservabilitycore
  20. Mechanisms

    Execution Policy

    Decide which tool actions may run, which require a human or policy decision, and which must be constrained by an operating-system boundary.

    execution-policysafetycore
  21. Mechanisms

    Extension Runtime

    Load and execute hooks, plugins, skills, custom tools, or provider adapters without collapsing extension code into the core loop.

    extension-runtimeextensibilitycore
  22. Mechanisms

    Instruction Discovery

    Discover, order, bound, and attribute project instructions before model execution.

    instruction-discoverycontextcore
  23. Mechanisms

    Memory Retrieval

    Store selected, source-attributed facts outside the immediate context and retrieve only relevant records for the current turn.

    memory-retrievalcontextcore
  24. Mechanisms

    Network Boundary

    Constrain outbound destinations, protocols, resolution, redirects, and credential flow independently from whether a tool is semantically approved.

    network-boundarysafetycore
  25. Mechanisms

    Operating-System Sandbox

    Enforce filesystem, process, and related capability limits below semantic tool policy so compromised or incorrect code cannot exceed its execution envelope.

    os-sandboxsafetycore
  26. Mechanisms

    Workspace Path Boundary

    Reject absolute paths and traversal before a file tool touches the filesystem.

    path-boundarysafetycore
  27. Mechanisms

    Project Trust and Instruction Authority

    Bind explicit trust to a stable workspace identity and decide which project inputs may become instructions instead of ordinary untrusted data.

    project-trustsafetycore
  28. Mechanisms

    Retry and Recovery

    Classify failures, bound retries, avoid duplicate side effects, and preserve enough state for a safe continuation or explicit pause.

    retry-recoveryreliabilitycore
  29. Mechanisms

    Session Replay and Branching

    Rebuild derived session state from an append-only event log and create lineage-preserving branches without rewriting the parent run.

    session-replayreliabilitycore
  30. Mechanisms

    Subagent Orchestration

    Delegate bounded work into isolated contexts, control capabilities and lifecycle, and merge results without confusing ownership or workspace state.

    subagent-orchestrationorchestrationemerging
  31. Mechanisms

    Tool Dispatch

    Validate, authorize, execute, normalize, and record model-requested tool calls.

    tool-dispatchtoolscore
  32. Mechanisms

    Tool Output Truncation

    Bound oversized tool results while preserving provenance, size, and a useful preview.

    tool-output-truncationcontextcore
  33. Mechanisms

    Workspace Tools

    Expose file, command, and edit effects through explicit, inspectable tool boundaries.

    workspace-toolstoolscore
  34. Agents

    Claude Code

    Anthropic's terminal coding agent; this Tier B profile is limited to official documentation and the public distribution repository because the product core is not published there.

    claude-codeTier Bpartialcli
  35. Agents

    Codex

    OpenAI's coding agent across terminal and other product surfaces; published snapshots keep official product behavior separate from the open-source CLI implementation.

    codexTier Bopencli
  36. Agents

    Grok Build

    SpaceXAI's open-source Rust coding agent with TUI, headless, and ACP surfaces; this Tier B profile maps a pinned repository sync while keeping runtime and sandbox claims explicitly bounded.

    grok-buildTier Bopencli
  37. Agents

    OpenCode

    An open-source, multi-provider coding agent with terminal, desktop, server, and ACP surfaces; this Tier B profile maps a pinned source revision without claiming Native runtime verification.

    opencodeTier Bopencli
  38. Agents

    Pi

    A deliberately minimal, multi-provider terminal coding harness whose agent core, session tree, compaction, tools, RPC mode, and extension runtime are published in one monorepo.

    piTier Bopencli
  39. Agents

    Reasonix

    An open Go coding-agent engine shared by terminal, desktop, browser, and ACP surfaces, with cache-aware context maintenance, policy gates, checkpoints, plugins, and subagent orchestration.

    reasonixTier Bopencli
  40. Agents

    Reference Harness

    The project's dependency-free, deterministic teaching harness and controlled-experiment baseline; it is not a production coding-agent product.

    reference-harnessTier Copensdk
  41. Snapshots

    Claude Code · 2.1.226-docs-boundary

    Official-documentation snapshot aligned with public distribution tag v2.1.226. No claim in this snapshot asserts the closed core's file layout, hidden prompts, or internal state machine.

    claude-code-2026-08-10-docs2026-08-10clicurrent
  42. Snapshots

    Codex · docs-observation

    This is an official-documentation snapshot, not a pinned binary or source commit. Version-specific implementation claims remain intentionally unknown.

    codex-2026-08-10-cli2026-08-10clicurrent
  43. Snapshots

    Codex · source@89a335e

    Clean source inspection pinned to one public commit. The snapshot does not claim that every hosted Codex surface runs this exact revision, and it has not yet earned Tier A because no native experiment is attached.

    codex-2026-08-10-source2026-08-10clicurrent
  44. Snapshots

    Codex · source@89a335e

    Separate server-surface snapshot for the public App Server transport and JSON-RPC dispatch boundary; it is not evidence about a particular desktop or IDE client release.

    codex-2026-08-10-app-server2026-08-10servercurrent
  45. Snapshots

    Grok Build · 1.0.0@75e73f3

    Pinned public repository commit 75e73f3; SOURCE_REV records upstream monorepo revision a61c32b12a2b400f212221cd8762e05f9b36828d. Sandbox claims preserve the source-visible unsupported/apply-failure fallback and do not imply universal fail-closed enforcement.

    grok-build-2026-08-10-source2026-08-10clicurrent
  46. Snapshots

    OpenCode · 1.18.16@941e71d

    Clean source inspection pinned to OpenCode package version 1.18.16 and one immutable commit. The profile separates semantic permission rules from OS isolation and remains Tier B until a reviewed Native experiment exists.

    opencode-2026-08-10-source2026-08-10clicurrent
  47. Snapshots

    Pi · 0.84.1@936aff0

    Clean source inspection of package version 0.84.1 pinned to one commit. Negative safety claims describe the documented core boundary, not every third-party Pi extension. A native experiment is still required for Tier A.

    pi-2026-08-10-source2026-08-10clicurrent
  48. Snapshots

    Pi · 0.84.1@936aff0

    RPC-specific snapshot. Its JSONL transport is source-mapped independently from Pi's interactive CLI behavior.

    pi-2026-08-10-rpc2026-08-10servercurrent
  49. Snapshots

    Reasonix · 1.22.0@6376bd4

    Clean source inspection pinned to the commit whose release ledger identifies 1.22.0. This snapshot maps the CLI engine only; desktop, browser, and ACP behavior is not assumed identical unless a claim names the shared boundary. A native experiment is still required for Tier A.

    reasonix-2026-08-10-source2026-08-10clicurrent
  50. Snapshots

    Reasonix · 1.22.0@6376bd4

    ACP-specific snapshot for NDJSON JSON-RPC framing, event projection, and permission requests. Runtime interoperability remains untested in this project.

    reasonix-2026-08-10-acp2026-08-10acpcurrent
  51. Snapshots

    Reference Harness · 0.1.0

    A controlled, scripted baseline used to validate the Lab pipeline. It must not be compared with a native vendor agent or presented as production-grade isolation.

    reference-harness-2026-08-10-controlled2026-08-10sdkcurrent
  52. Claims

    Claude Code's official Agent Teams documentation describes experimental lead-and-teammate coordination with independent context windows, a shared task list, direct messaging, and no automatic worktree isolation.

    claude-code · subagent-orchestration · reviewed

    claude-code-doc-agent-teamssubagent-orchestrationcliofficial-doc
  53. Claims

    Claude Code's official documentation separates permission rules, which decide tool and resource access, from OS-level sandbox enforcement for Bash filesystem and network access, and recommends using both as defense in depth.

    claude-code · execution-policy · reviewed

    claude-code-doc-execution-policyexecution-policycliofficial-doc
  54. Claims

    Claude Code's official hook contract exposes lifecycle events including pre- and post-tool use, permission requests, session start/end, and stop; PreToolUse hooks can deny a call before normal permission-mode checks, while allow decisions cannot override stricter deny rules.

    claude-code · extension-runtime · reviewed

    claude-code-doc-hooksextension-runtimecliofficial-doc
  55. Claims

    Claude Code's official documentation describes subagents with separate context windows, configurable prompts, tools, models, permissions, hooks, skills, depth and turn limits, and optional temporary Git worktree isolation.

    claude-code · subagent-orchestration · reviewed

    claude-code-doc-subagentssubagent-orchestrationcliofficial-doc
  56. Claims

    Official Codex CLI documentation presents one terminal loop for exploring code, planning changes, editing files, running local tools, steering an active turn, and continuing follow-up work in the same session.

    codex · agent-loop · reviewed

    codex-cli-terminal-loopagent-loopcliofficial-doc
  57. Claims

    At the pinned Codex source revision, run_turn repeatedly samples the model, executes requested function calls, feeds results into later samples, accepts steering input, and stops when follow-up work is no longer required or a stop/error path wins.

    codex · agent-loop · reviewed

    codex-source-agent-loopagent-loopclisource
  58. Claims

    The pinned Codex App Server accepts JSON-RPC requests and notifications over selectable stdio, Unix-socket, or WebSocket transports and dispatches typed requests through dedicated processors.

    codex · client-server-protocol · reviewed

    codex-source-app-serverclient-server-protocolserversource
  59. Claims

    The pinned Codex turn loop checks context limits before and after sampling and can replace history through local or remote compaction before resuming the same turn.

    codex · context-compaction · reviewed

    codex-source-auto-compactioncontext-compactionclisource
  60. Claims

    The pinned Codex protocol defines typed events for turn lifecycle, streamed content, tools, approvals, compaction, hooks, and collaborative-agent activity, and the turn implementation emits those events through the session boundary.

    codex · event-stream · reviewed

    codex-source-event-streamevent-streamclisource
  61. Claims

    The pinned Codex core models approval decisions separately from sandbox transformation, with policy-dependent outcomes that can forbid, request approval for, skip approval for, or sandbox a tool attempt.

    codex · execution-policy · reviewed

    codex-source-execution-policyexecution-policyclisource
  62. Claims

    At the pinned revision, Codex assembles selected skills and plugins into turn context and runs typed lifecycle hooks around session start, tool use, permissions, stop, compaction, and session end.

    codex · extension-runtime · reviewed

    codex-source-extension-runtimeextension-runtimeclisource
  63. Claims

    The pinned Codex core registers multi-agent tools for spawning and communicating with child agents and implements spawn through a dedicated agent-control boundary rather than inside the main sampling loop.

    codex · subagent-orchestration · reviewed

    codex-source-subagent-routingsubagent-orchestrationclisource
  64. Claims

    The pinned Codex core separates model-visible tool specifications in ToolRouter from runtime registration and collision handling in ToolRegistry, then resolves each parsed call to its registered runtime.

    codex · tool-dispatch · reviewed

    codex-source-tool-routertool-dispatchclisource
  65. Claims

    Grok Build's pinned ACP library implements line-oriented JSON-RPC transport and normalization, and the shell projects Agent activity into ACP session updates for editor or process clients.

    grok-build · client-server-protocol · reviewed

    grok-build-source-acp-interfaceclient-server-protocolclisource
  66. Claims

    Grok Build's pinned session actor multiplexes commands and asynchronous events, explicitly starts turns, streams sampler output, dispatches tool calls, and emits lifecycle updates through completion, failure, or cancellation.

    grok-build · agent-loop · reviewed

    grok-build-source-agent-loopagent-loopclisource
  67. Claims

    Grok Build's pinned Agent uses a configurable compaction policy with context thresholds, optional model and memory-flush settings, wall-clock limits, and an optional two-pass mode.

    grok-build · context-compaction · reviewed

    grok-build-source-context-compactioncontext-compactionclisource
  68. Claims

    Grok Build's pinned file-event utilities write versioned session events as append-oriented JSONL, and the tracker emits typed turn, tool, permission, and terminal lifecycle records.

    grok-build · event-stream · reviewed

    grok-build-source-event-logevent-streamclisource
  69. Claims

    Grok Build's pinned source separates semantic allow/ask/deny policy from OS sandbox activation; the manager exposes requested and active state, uses supported platform backends, and may explicitly continue unsandboxed when unsupported or application fails.

    grok-build · execution-policy · reviewed

    grok-build-source-execution-policyexecution-policyclisource
  70. Claims

    Grok Build's pinned extension surface includes plugins, prompt skills, MCP tools, and lifecycle hooks, with pre-tool gates able to modify or block operations.

    grok-build · extension-runtime · reviewed

    grok-build-source-extension-runtimeextension-runtimeclisource
  71. Claims

    Grok Build's pinned sampler retry module produces explicit decisions for rate limits, transient failures, context limits, image fallback, and client rebuilds, using retry-after or exponential jitter while leaving fatal outcomes non-retryable.

    grok-build · retry-recovery · reviewed

    grok-build-source-retry-policyretry-recoveryclisource
  72. Claims

    Grok Build's pinned source resolves named subagents and resume identity, while task execution uses separate child session identity and explicit isolation or worktree metadata rather than an inline prompt macro.

    grok-build · subagent-orchestration · reviewed

    grok-build-source-subagent-runtimesubagent-orchestrationclisource
  73. Claims

    Grok Build's pinned ToolBridge owns a finalized registry, exposes model-facing definitions, supports built-in and MCP-backed registrations, and routes calls through typed registry dispatch.

    grok-build · tool-dispatch · reviewed

    grok-build-source-tool-bridgetool-dispatchclisource
  74. Claims

    OpenCode's pinned session prompt service repeatedly handles pending subtasks and compaction, resolves the active Agent and tools, invokes the model processor, and stops only at an explicit terminal condition.

    opencode · agent-loop · reviewed

    opencode-source-agent-loopagent-loopclisource
  75. Claims

    OpenCode's pinned source exposes a headless HTTP server, while the terminal UI can use an embedded worker and RPC-forward global server events rather than owning a separate Agent core.

    opencode · client-server-protocol · reviewed

    opencode-source-client-serverclient-server-protocolclisource
  76. Claims

    OpenCode's pinned compaction service measures context pressure, constructs a summary turn, preserves recent context, and separately truncates oversized historical tool output.

    opencode · context-compaction · reviewed

    opencode-source-context-compactioncontext-compactionclisource
  77. Claims

    OpenCode's pinned event bridge enriches events with location and aggregate identity, assigns version and sequence metadata for durable records, and forwards public events through the global bus.

    opencode · event-stream · reviewed

    opencode-source-event-bridgeevent-streamclisource
  78. Claims

    OpenCode's pinned permission service uses last-match wildcard rules, defaults unresolved requests to ask, and represents deny, one-time approval, persistent approval, and rejection explicitly.

    opencode · execution-policy · reviewed

    opencode-source-execution-policyexecution-policyclisource
  79. Claims

    OpenCode's pinned source discovers local skills, loads configured plugins, and imports MCP tools into the common tool-resolution boundary, with trust and permission still configuration-dependent.

    opencode · extension-runtime · reviewed

    opencode-source-extension-runtimeextension-runtimeclisource
  80. Claims

    OpenCode's pinned retry policy classifies provider and HTTP failures, honors retry-after metadata, applies bounded exponential delay, and excludes non-retryable failures such as context overflow.

    opencode · retry-recovery · reviewed

    opencode-source-retry-policyretry-recoveryclisource
  81. Claims

    OpenCode's pinned task tool limits nesting, derives child permissions, creates parent-linked child sessions, and supports foreground or background tasks with explicit metadata.

    opencode · subagent-orchestration · reviewed

    opencode-source-subagent-sessionssubagent-orchestrationclisource
  82. Claims

    OpenCode's pinned tool registry combines built-in, custom, plugin, and MCP definitions, permission-filters visibility, validates arguments, and dispatches registered handlers with normalized output metadata.

    opencode · tool-dispatch · reviewed

    opencode-source-tool-registrytool-dispatchclisource
  83. Claims

    Pi's pinned agent core uses an inner loop for model/tool/steering turns and an outer loop for queued follow-up messages, emitting explicit agent, turn, message, and tool lifecycle events throughout.

    pi · agent-loop · reviewed

    pi-source-agent-loopagent-loopclisource
  84. Claims

    Pi triggers compaction against a configurable reserve threshold, summarizes the older branch, keeps recent entries, tracks file operations, and preserves the full append-only session history outside the reduced model context.

    pi · context-compaction · reviewed

    pi-source-context-compactioncontext-compactionclisource
  85. Claims

    Pi explicitly does not provide an in-process permission system or built-in sandbox for model-invoked filesystem, process, network, or credential access; isolation must come from a container, VM, micro-VM, or other operating-system boundary.

    pi · execution-policy · reviewed

    pi-source-execution-boundaryexecution-policyclisource
  86. Claims

    Pi's coding-agent layer uses an extension runner to dispatch lifecycle events and lets extensions add or replace tools, commands, UI, providers, context transforms, and compaction behavior without modifying the low-level agent loop.

    pi · extension-runtime · reviewed

    pi-source-extension-runtimeextension-runtimeclisource
  87. Claims

    Pi intentionally omits a built-in subagent orchestration policy; its official project guidance points users to separate Pi processes or extensions so delegation semantics remain outside the minimal core.

    pi · subagent-orchestration · reviewed

    pi-source-orchestration-boundarysubagent-orchestrationclisource
  88. Claims

    Pi RPC mode accepts typed commands as LF-delimited JSON on stdin and emits responses plus agent events as LF-delimited JSON on stdout, including controls for prompting, steering, sessions, compaction, retry, models, and tools.

    pi · client-server-protocol · reviewed

    pi-source-rpc-modeclient-server-protocolserversource
  89. Claims

    Pi combines a live typed AgentEvent stream with an append-only JSONL session tree whose entries carry id and parentId, allowing the active context to move between branches without deleting prior history.

    pi · event-stream · reviewed

    pi-source-session-eventsevent-streamclisource
  90. Claims

    Pi validates tool arguments, offers before- and after-tool interception points, and executes a batch sequentially or in parallel according to the configured strategy and individual tool constraints.

    pi · tool-dispatch · reviewed

    pi-source-tool-dispatchtool-dispatchclisource
  91. Claims

    Reasonix ACP uses an NDJSON JSON-RPC 2.0 connection over stdin/stdout, runs requests and notifications concurrently, maps engine events to session updates, and forwards gated tool decisions as client permission requests.

    reasonix · client-server-protocol · reviewed

    reasonix-source-acpclient-server-protocolacpsource
  92. Claims

    The pinned Reasonix engine runs a bounded tool-round loop that captures a cache prefix shape, streams a provider turn, commits only a clean terminal attempt, then either finalizes or dispatches the committed tool calls before the next round.

    reasonix · agent-loop · reviewed

    reasonix-source-agent-loopagent-loopclisource
  93. Claims

    Reasonix treats compaction as a low-frequency cache-reset point, uses multiple context thresholds and a recent-tail budget, and records prefix-shape diagnostics that distinguish system, tool-schema, and content-rewrite changes.

    reasonix · context-compaction · reviewed

    reasonix-source-context-maintenancecontext-compactionclisource
  94. Claims

    Reasonix designates an append-only events JSONL file as transcript authority and keeps derived context, event-index, display-index, conflict, checkpoint, and job artifacts in separate sidecars.

    reasonix · event-stream · reviewed

    reasonix-source-event-logevent-streamclisource
  95. Claims

    Reasonix separates permission and approval posture from OS-level shell confinement: approval modes preserve explicit deny and fresh-decision rules, while an enforced but unavailable shell sandbox fails closed instead of running unwrapped.

    reasonix · execution-policy · reviewed

    reasonix-source-execution-policyexecution-policyclisource
  96. Claims

    Reasonix plugin manifests can contribute hooks and MCP servers, while lazy MCP placeholders keep cached tool metadata stable and defer process startup or live schema reconciliation until needed.

    reasonix · extension-runtime · reviewed

    reasonix-source-extension-runtimeextension-runtimeclisource
  97. Claims

    Reasonix replays one frozen sampling payload across bounded stream-recovery attempts, commits only the selected clean attempt, and separately breaks repeated identical write-like failures after a small threshold unless a safe state recheck clears the signature.

    reasonix · retry-recovery · reviewed

    reasonix-source-retry-recoveryretry-recoveryclisource
  98. Claims

    Reasonix constructs a filtered child tool registry, supports a read-only child profile, caps recursive delegation by depth, hides parent-only orchestration and job tools, and gives each child a private temporary-session boundary.

    reasonix · subagent-orchestration · reviewed

    reasonix-source-subagentssubagent-orchestrationclisource
  99. Claims

    Reasonix separates batch scheduling from single-call execution, so concurrency decisions, per-call context, capability gates, tool invocation, results, and failure bookkeeping cross explicit boundaries.

    reasonix · tool-dispatch · reviewed

    reasonix-source-tool-dispatchtool-dispatchclisource
  100. Claims

    In the controlled reference experiment, the project harness validates and executes one declared read_file call, records the successful result between two model turns, and produces the same normalized trace fingerprint across two repetitions.

    reference-harness · tool-dispatch · reviewed

    reference-harness-controlled-tool-roundtriptool-dispatchsdkreproduced
  101. Experiments

    Reference Harness approval binding probe

    Does one approval grant authorize only the exact declared write and leave an auditable result?

    reference-approval-binding-v1controlledcomplete2 runs
  102. Experiments

    Reference Harness checkpoint and rollback probe

    Can the Reference Harness restore the declared file to a clean checkpoint after a reviewed temporary change?

    reference-checkpoint-rollback-v1controlledcomplete2 runs
  103. Experiments

    Reference Harness context budget and truncation probe

    Does the Reference Harness enforce a declared context budget and record deterministic tool-output truncation?

    reference-context-budget-v1controlledcomplete2 runs
  104. Experiments

    Reference Harness context compaction probe

    Can the Reference Harness replace older turns with a deterministic checkpoint while preserving the verified conclusion?

    reference-context-compaction-v1controlledcomplete2 runs
  105. Experiments

    Reference Harness memory retrieval probe

    Can the Reference Harness retrieve a task-relevant memory source and select the declared skill without loading unrelated memory?

    reference-memory-retrieval-v1controlledcomplete2 runs
  106. Experiments

    Reference Harness project trust probe

    Can the Reference Harness distinguish trusted policy from untrusted project and tool text before instruction assembly?

    reference-project-trust-v1controlledcomplete2 runs
  107. Experiments

    Reference Harness sandbox and network policy probe

    Can the Reference Harness record a destination-scoped network decision while guaranteeing that the teaching backend performs no real network effect?

    reference-sandbox-network-v1controlledcomplete2 runs
  108. Experiments

    Reference Harness replay and branch probe

    Can append-only events deterministically rebuild session state and create a branch from an explicit sequence boundary?

    reference-session-replay-v1controlledcomplete2 runs
  109. Experiments

    Reference Harness stream normalization probe

    Can the Reference Harness normalize a fragmented model stream into one deterministic event sequence without depending on hidden model state?

    reference-stream-normalization-v1controlledcomplete2 runs
  110. Experiments

    Reference Harness controlled tool roundtrip

    Can the registry-backed reference runner reproduce one read-only tool roundtrip, including the same normalized event trace, from a pinned fixture and scenario?

    reference-tool-roundtrip-v1controlledcomplete2 runs
  111. Traces

    Deterministic tool roundtrip (format example)

    synthetic trace · 4 events

    example-tool-roundtripsyntheticformat fixture
  112. Traces

    Reference Harness approval binding probe · reference-approval-binding-v1-run-001

    controlled trace · 11 events

    reference-approval-binding-v1-run-001controlledreference-approval-binding-v1
  113. Traces

    Reference Harness approval binding probe · reference-approval-binding-v1-run-002

    controlled trace · 11 events

    reference-approval-binding-v1-run-002controlledreference-approval-binding-v1
  114. Traces

    Reference Harness checkpoint and rollback probe · reference-checkpoint-rollback-v1-run-001

    controlled trace · 15 events

    reference-checkpoint-rollback-v1-run-001controlledreference-checkpoint-rollback-v1
  115. Traces

    Reference Harness checkpoint and rollback probe · reference-checkpoint-rollback-v1-run-002

    controlled trace · 15 events

    reference-checkpoint-rollback-v1-run-002controlledreference-checkpoint-rollback-v1
  116. Traces

    Reference Harness context budget and truncation probe · reference-context-budget-v1-run-001

    controlled trace · 10 events

    reference-context-budget-v1-run-001controlledreference-context-budget-v1
  117. Traces

    Reference Harness context budget and truncation probe · reference-context-budget-v1-run-002

    controlled trace · 10 events

    reference-context-budget-v1-run-002controlledreference-context-budget-v1
  118. Traces

    Reference Harness context compaction probe · reference-context-compaction-v1-run-001

    controlled trace · 10 events

    reference-context-compaction-v1-run-001controlledreference-context-compaction-v1
  119. Traces

    Reference Harness context compaction probe · reference-context-compaction-v1-run-002

    controlled trace · 10 events

    reference-context-compaction-v1-run-002controlledreference-context-compaction-v1
  120. Traces

    Reference Harness memory retrieval probe · reference-memory-retrieval-v1-run-001

    controlled trace · 13 events

    reference-memory-retrieval-v1-run-001controlledreference-memory-retrieval-v1
  121. Traces

    Reference Harness memory retrieval probe · reference-memory-retrieval-v1-run-002

    controlled trace · 13 events

    reference-memory-retrieval-v1-run-002controlledreference-memory-retrieval-v1
  122. Traces

    Reference Harness project trust probe · reference-project-trust-v1-run-001

    controlled trace · 8 events

    reference-project-trust-v1-run-001controlledreference-project-trust-v1
  123. Traces

    Reference Harness project trust probe · reference-project-trust-v1-run-002

    controlled trace · 8 events

    reference-project-trust-v1-run-002controlledreference-project-trust-v1
  124. Traces

    Reference Harness sandbox and network policy probe · reference-sandbox-network-v1-run-001

    controlled trace · 13 events

    reference-sandbox-network-v1-run-001controlledreference-sandbox-network-v1
  125. Traces

    Reference Harness sandbox and network policy probe · reference-sandbox-network-v1-run-002

    controlled trace · 13 events

    reference-sandbox-network-v1-run-002controlledreference-sandbox-network-v1
  126. Traces

    Reference Harness replay and branch probe · reference-session-replay-v1-run-001

    controlled trace · 7 events

    reference-session-replay-v1-run-001controlledreference-session-replay-v1
  127. Traces

    Reference Harness replay and branch probe · reference-session-replay-v1-run-002

    controlled trace · 7 events

    reference-session-replay-v1-run-002controlledreference-session-replay-v1
  128. Traces

    Reference Harness stream normalization probe · reference-stream-normalization-v1-run-001

    controlled trace · 7 events

    reference-stream-normalization-v1-run-001controlledreference-stream-normalization-v1
  129. Traces

    Reference Harness stream normalization probe · reference-stream-normalization-v1-run-002

    controlled trace · 7 events

    reference-stream-normalization-v1-run-002controlledreference-stream-normalization-v1
  130. Traces

    Reference Harness controlled tool roundtrip · reference-tool-roundtrip-v1-run-001

    controlled trace · 9 events

    reference-tool-roundtrip-v1-run-001controlledreference-tool-roundtrip-v1
  131. Traces

    Reference Harness controlled tool roundtrip · reference-tool-roundtrip-v1-run-002

    controlled trace · 9 events

    reference-tool-roundtrip-v1-run-002controlledreference-tool-roundtrip-v1
  132. Traces

    S01 The Minimal Agent Loop · Golden Trace

    golden trace · 5 events

    run-s01-agent-loop-goldengoldenformat fixture
  133. Traces

    S02 Streaming and Events · Golden Trace

    golden trace · 7 events

    run-s02-events-streaming-goldengoldenformat fixture
  134. Traces

    S03 Tool Registry and Dispatch · Golden Trace

    golden trace · 9 events

    run-s03-tool-dispatch-goldengoldenformat fixture
  135. Traces

    S04 File, Shell, and Edit Tools · Golden Trace

    golden trace · 9 events

    run-s04-workspace-tools-goldengoldenformat fixture
  136. Traces

    S05 Instructions and Project Discovery · Golden Trace

    golden trace · 5 events

    run-s05-instructions-goldengoldenformat fixture
  137. Traces

    S06 Context Budgets and Truncation · Golden Trace

    golden trace · 10 events

    run-s06-context-budget-goldengoldenformat fixture
  138. Traces

    S07 Session Event Logs, Replay, and Branching · Golden Trace

    golden trace · 7 events

    run-s07-session-replay-goldengoldenformat fixture
  139. Traces

    S08 Context Compaction and Checkpoints · Golden Trace

    golden trace · 10 events

    run-s08-context-compaction-goldengoldenformat fixture
  140. Traces

    S09 Memory and On-Demand Skills · Golden Trace

    golden trace · 13 events

    run-s09-memory-skills-goldengoldenformat fixture
  141. Traces

    S10 Approval Policy and Exact Authorization · Golden Trace

    golden trace · 11 events

    run-s10-approval-policy-goldengoldenformat fixture
  142. Traces

    S11 OS Sandboxes and Network Boundaries · Golden Trace

    golden trace · 13 events

    run-s11-sandbox-network-goldengoldenformat fixture
  143. Traces

    S12 Project Trust and Prompt Injection · Golden Trace

    golden trace · 8 events

    run-s12-project-trust-goldengoldenformat fixture
  144. Traces

    S13 Git Checkpoints, Diffs, and Rollback · Golden Trace

    golden trace · 15 events

    run-s13-checkpoint-rollback-goldengoldenformat fixture