Struct CapabilitiesV3 

pub struct CapabilitiesV3 {

    pub schema_version: String,
    pub summary: String,
    pub to_describe_to_user: String,
    pub tools: Vec<ToolEntry>,
    pub agent_permitted_families: Option<Vec<String>>,
    pub your_harness_supports_deferred_registration: Option<bool>,
    pub tier: String,
    pub version: String,
    pub features: CapabilityFeatures,
    pub models: CapabilityModels,
    pub permissions: CapabilityPermissions,
    pub hooks: CapabilityHooks,
    pub compaction: CapabilityCompaction,
    pub approval: CapabilityApproval,
    pub transcripts: CapabilityTranscripts,
    pub hnsw: CapabilityHnsw,
    pub kg_backend: Option<String>,
    pub memory_kinds: Vec<String>,
    pub reflection: CapabilityReflection,
    pub skills: CapabilitySkills,
    pub forensic: CapabilityForensic,
    pub governance: CapabilityGovernance,
    pub atomisation: CapabilityAtomisation,
    pub memory_kind_vocab: CapabilityMemoryKindVocab,
    pub confidence_calibration: CapabilityConfidenceCalibration,
    pub provenance_substrate_layer: CapabilityProvenanceSubstrateLayer,
}

v0.7.0 capabilities schema (A1 increment). Additive over Capabilities (v2): the top-level summary field carries a pre-computed, plain-language description of the LLM’s operational tool surface so reasoning-class LLMs converge on accurate first-answer descriptions without having to traverse families[] and count manually.

Wire selection: clients opt in via accept="v3" on the MCP memory_capabilities call, or Accept-Capabilities: v3 over HTTP (HTTP wiring lands with A5). Default response remains v2 until A5 flips the default. v2 stays supported indefinitely.

Increment plan: A2 adds to_describe_to_user, A3 adds per-tool callable_now, A4 adds agent_permitted_families. A5 bumps the default wire shape and seals v3 as the recommended client target.

Fields

schema_version: String

Schema-version discriminator. Always "3" in v0.7.0.

summary: String

Pre-computed plain-language summary of operational access. Carries the loaded vs total tool counts under the active profile plus the three named recovery paths (--profile, memory_load_family, memory_smart_load). Computed at response time from the live profile state — never cached at build time because the count of advertised tools depends on the running server’s --profile flag.

to_describe_to_user: String

v0.7.0 A2 — pre-computed end-user-facing sentence the LLM should repeat verbatim when an end-user asks “what tools do you have?”. Distinct in tone from Self::summary: no MCP jargon, no --profile references, no JSON-RPC vocabulary. Reads as a normal sentence a human would say. See docs/v0.7/canonical-phrasings.md for the canonical substitution template + worked examples per profile.

tools: Vec<ToolEntry>

v0.7.0 A3 — per-tool array carrying name, family, loaded, and callable_now. callable_now combines profile-side loaded-state with the [mcp.allowlist] agent-can-call decision so an LLM that caches this manifest can skip a doomed JSON-RPC call rather than discovering -32601 the hard way. Order matches tool_definitions()’s registration walk so a sequential reader gets a stable presentation.

agent_permitted_families: Option<Vec<String>>

v0.7.0 A4 — list of family names this agent is permitted to access via the [mcp.allowlist] gate. Present (with possibly an empty array) only when the allowlist is configured AND an agent_id was provided. Absent when the allowlist is disabled or no agent_id was provided — that absence is meaningful, not a drift, hence Option<Vec<String>> + skip_serializing_if.

LLMs that keep a per-agent manifest cache can use this to short-circuit family-level decisions without iterating tools[] and counting unique families.

your_harness_supports_deferred_registration: Option<bool>

v0.7.0 B4 — whether the active MCP harness exposes tools registered after the initial tools/list to the LLM. Computed at response time from the harness detected at the initialize.clientInfo.name handshake (see crate::harness).

Some(true) only for Claude Code today (deferred registration via ToolSearch). Some(false) for every other named harness. None (omitted from the wire via skip_serializing_if) when no clientInfo was captured — typically HTTP callers, or an MCP client that issued memory_capabilities before initialize (malformed but defensively handled by absence).

Track B’s runtime loaders (B1 memory_load_family, B2 memory_smart_load) key off this bit to shape their to_invoke text — on false harnesses they advise the LLM to ask the operator for a --profile <family> restart rather than expect the new tools to appear mid-session.

tier: String

version: String

features: CapabilityFeatures

models: CapabilityModels

permissions: CapabilityPermissions

hooks: CapabilityHooks

compaction: CapabilityCompaction

approval: CapabilityApproval

transcripts: CapabilityTranscripts

hnsw: CapabilityHnsw

kg_backend: Option<String>

v0.7 J1 — knowledge-graph backend tag forwarded from the v2 projection. Some("age" | "cte") once the SAL handle is threaded through AppState; None while no SAL adapter is wired. Skipped from the JSON wire when None so older clients that don’t know the field round-trip cleanly.

memory_kinds: Vec<String>

L1-1 (v0.7.0) — typed memory-kind set. Forwarded from the v2 projection’s memory_kinds field. Always ["observation", "reflection"] for v0.7.0.

L3-5 honesty note. The grand-slam spec called for a third "goal" kind here, but the crate::models::memory::MemoryKind enum in this binary only carries Observation and Reflection. Per the operator’s “every reported field maps to real implementation” directive, the v3 surface reports exactly what the substrate enforces — the goal kind is deferred to the tracker (a4f8d465) for a v0.8.0 wave that lands the enum variant + migration + write-path coverage. Reporting it here today would be theatrical.

reflection: CapabilityReflection

v0.7.0 L3-5 — recursive-learning capability surface. Every sub-field anchors a real implementation in this binary; see CapabilityReflection for the per-field audit anchors.

skills: CapabilitySkills

v0.7.0 L3-5 — Agent-Skills capability surface. Lists the seven registered memory_skill_* MCP tools; the round-trip guarantee is pinned by tests/skill_test.rs. See CapabilitySkills.

forensic: CapabilityForensic

v0.7.0 L3-5 — forensic-evidence CLI surface. Names the three driver verbs that this binary actually ships (verify-reflection-chain, export-forensic-bundle, verify-forensic-bundle). See CapabilityForensic.

governance: CapabilityGovernance

v0.7.0 L3-5 — substrate-rules governance surface. Honestly labelled "operator_signed" because the L1-6 loader refuses to honour unsigned rules. See CapabilityGovernance.

atomisation: CapabilityAtomisation

v0.7.0 WT-1-G — atomisation capability surface. Names the six operator-facing atomisation surfaces (tool / cli / auto / recall_preference / forensic / curator) plus the derives_from link relation that anchors atom → parent lineage. See CapabilityAtomisation for the per-field implementation anchor map.

Additive over the L3-5 surface — pre-WT-1-G v3 payloads still deserialise cleanly (the default_capability_atomisation helper resolves to the current-implementation snapshot for any payload missing the field).

memory_kind_vocab: CapabilityMemoryKindVocab

v0.7.x Form 6 (issue #759) — Batman-taxonomy memory-kind vocabulary capability surface. Names the recall-filter + auto-classify surfaces shipped under Form 6 and enumerates the substrate’s full set of recognised memory_kind values. See CapabilityMemoryKindVocab.

Additive over the WT-1-G surface — pre-Form-6 v3 payloads deserialise cleanly via the default_capability_memory_kind_vocab helper.

confidence_calibration: CapabilityConfidenceCalibration

v0.7.0 Form 5 (issue #758) — confidence-calibration capability surface. Names the five operator-facing Form-5 substrates (auto_derive / shadow_mode / freshness_decay / calibration_cli / calibration_tool) plus the signals_schema wire-shape discriminator. See CapabilityConfidenceCalibration for the per-field anchor map.

Additive over the WT-1-G surface — pre-Form-5 v3 payloads still deserialise cleanly because of the default_capability_confidence_calibration helper.

provenance_substrate_layer: CapabilityProvenanceSubstrateLayer

v0.7.0 #973 Item C — narrative summary of the substrate’s do-calculus posture.

Trait Implementations

impl Clone for CapabilitiesV3

fn clone(&self) -> CapabilitiesV3

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for CapabilitiesV3

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for CapabilitiesV3

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for CapabilitiesV3

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.