Fork-replay one recorded span from a failed agent run and get an attribution verdict.
Scored across 2 tools
The tools occupy distinct phases: fork_safety_check is a non-executing forensic pre-check, while replay_span actually fork-replays a span for attribution. They still both target 'one recorded span,' so an agent could occasionally hesitate between checking safety and requesting replay, but the descriptions make the boundary reasonably clear.
Both names use snake_case and are readable, but the conventions are slightly mixed: fork_safety_check is a noun-compound name while replay_span follows a verb_noun pattern. This is a minor deviation rather than a confusing inconsistency.
Two tools is borderline thin for a span-level replay/fork service. Each tool is substantive, but the surface feels minimal for the workflow implied by the server name.
There is no tool to discover or list recorded spans/runs, inspect run context, or batch-replay multiple spans. An agent must already know the exact span to check or replay, creating significant workflow gaps.
Examine ONE recorded span (or a run's span set) forensically — canonicalization, env binding, completeness, stub coverage — and return the fork-safety verdict (FORK-SAFE-LIVE / FORK-SAFE-STUBBED / FORK-DEGRADED / FORK-UNSAFE) WITHOUT executing the span or forking anything. $0.01 per span checked on the hosted surface.
| Name | Required | Description | Default |
|---|---|---|---|
| span | No | One normalized span record. | |
| spans | No | A recorded run's span records; without span_id every span is checked. | |
| span_id | No |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations present, the description carries the full burden and does well: it discloses that nothing is executed or forked, enumerates the four possible verdict strings, and warns of a $0.01 per-span charge on the hosted surface. Missing only auth/rate-limit or failure-mode detail.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
A single dense sentence front-loads the action and enumeration, followed by one clause on non-execution and one on cost. No filler, and the verdict list is placed where a scanning agent will see it.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
With no output schema, the description correctly compensates by naming the exact return values (the four FORK-* verdicts). For a nested-object forensic tool this is largely sufficient; only edge behavior on invalid or mixed spans is left unexplained.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 67% and the description maps the key choice (single span vs a run's span set) and reiterates the span_id scoping rule already in the schema. It adds framing but no format or validation detail for the undocumented portion of the nested span object.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
Uses a specific verb ("Examine ... forensically") against a specific resource (ONE recorded span or a run's span set) and enumerates the dimensions checked. It also explicitly separates itself from the sibling by stating it works WITHOUT executing or forking anything, unlike replay_span.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The phrase "WITHOUT executing the span or forking anything" establishes when this tool applies versus an execution/replay path, and "ONE recorded span (or a run's span set)" clarifies scope. It stops short of naming replay_span as the alternative, leaving the routing inference implicit.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Fork-replay ONE recorded span from a failed agent run and return an attribution verdict (determinism vs tool drift vs env drift vs world-state vs model decision), without re-running the whole run.
| Name | Required | Description | Default |
|---|---|---|---|
| span | No | One normalized span record. | |
| spans | No | A recorded run's span records; span_id selects the one to fork. | |
| span_id | No |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden. It usefully discloses the re-execution scope ('without re-running the whole run') and enumerates the verdict categories, which is real behavioral context. However, it omits mutation/reversibility, permissions, and any precondition that a recorded run's spans must be supplied.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
A single dense sentence that front-loads the action, the scope, and the output, with zero filler. Nothing could be trimmed without losing information.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a tool with nested object inputs, a confusing 0 required parameters, no annotations, and no output schema, the description covers purpose and the verdict output well. It still leaves preconditions (which inputs are actually required, whether any state is mutated) unaddressed.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 67%: 'span' and 'spans' are documented in-schema, while 'span_id' is not. The description reinforces single-span selection ('ONE recorded span', 'span_id selects the one to fork') but adds no format, default, or required/optional clarification, so it only marginally exceeds the schema.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
States a specific verb+resource (fork-replay ONE recorded span) and the resulting artifact (an attribution verdict), with a precise scope qualifier ('from a failed agent run', 'without re-running the whole run'). It is clear what the tool does, though there are no siblings to differentiate against.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies the debugging context ('from a failed agent run', 'ONE recorded span') but never states explicit when-to-use vs when-not or prerequisites, e.g. that the run must already be recorded. Usage is inferable rather than prescribed.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
fork_safety_checkreplay_spanEvidence-bound second-opinion audit of an agent conclusion against caller-supplied evidence.
Compaction witness for agents: pin invariants, name dropped spans, re-inject verbatim or HALT.
Replay supplied records against old and new bounded JSON Schema draft-07 contracts. Receive exact...
Parallel-burst abuse testing for agent APIs with replayable evidence packs.
Add one secure layer between your agents and this server.