pmcp_server_toolkit/

tools.rs

// Net-new code for Phase 83 TKIT-07.
// Lands the `[[tools]]`-config-driven synthesizer that turns config rows into
// `ToolInfo` + `Arc<dyn ToolHandler>` pairs.

//! `[[tools]]` → `ToolInfo` + `Arc<dyn ToolHandler>` synthesizer.
//!
//! Net-new code for Phase 83 TKIT-07. Turns curated `[[tools]]` config entries
//! into complete pmcp [`ToolInfo`] + [`Arc<dyn ToolHandler>`] pairs with zero
//! per-tool Rust handlers.
//!
//! # Invariants enforced
//!
//! - **JSON Schema object envelope, CHECKED HERE.** Every synthesized
//!   [`ToolInfo`] carries an `input_schema` with `"type": "object"`, an explicit
//!   `properties` map, a `required` array, and `"additionalProperties": false`
//!   (`true` only when `[server.validation] additional_properties` opts out).
//!   Unknown argument keys are refused by
//!   `pmcp::server::schema_validation::validate_input`, called from the
//!   `ValidatingToolHandler` decorator that `enforce_input_schema` wraps
//!   around every handler this module pushes — **in the toolkit, before the
//!   backend call**. Threat T-83-05-02.
//!
//!   Both halves matter, and stating only the first is what made this comment a
//!   defect for two phases. Core `pmcp`'s `tools/call` dispatch does **not**
//!   validate request arguments against a tool's declared `inputSchema`; that
//!   wiring is deliberately deferred (Phase 128 D-01), so a comment locating the
//!   enforcement "upstream" described a mitigation that did not exist. It exists
//!   now, and it exists *here*.
//!
//!   Condition: the `input-validation` feature must be on. It is on via this
//!   crate's `default` feature set and by that route ALONE — measured: `http =
//!   ["dep:reqwest", "dep:url", "dep:openapiv3", "dep:serde_yaml", "dep:base64",
//!   "dep:regex", "dep:tokio", "pmcp/streamable-http"]` and `openapi-code-mode =
//!   ["http", "code-mode", "pmcp-code-mode/js-runtime"]`, neither of which names
//!   it. Those two umbrellas do NOT forward it, so a
//!   `--no-default-features --features http` build — exactly the shape a
//!   consumer reaches for — turns it off, and then
//!   `enforce_input_schema` is the identity function and says so in a
//!   `tracing::warn!`. Second condition: `[server.validation]
//!   enforce_input_schema` must not be `false`.
//!
//!   Backed by `tests/input_validation_acceptance.rs`'s
//!   `input_validation_refuses_undeclared_argument_without_contacting_upstream`
//!   (row 10) and `input_validation_refuses_absent_arguments_when_required_declared`
//!   (row 9), each of which fails if the `validate_input` call is removed, plus
//!   `input_validation_accepts_compliant_call_with_one_upstream_request` (row 11)
//!   as the passing control that catches an over-refusing decorator.
//! - **`handler.metadata()` returns `Some(ToolInfo)`.** Phase 82's `tool_arc`
//!   consumes `handler.metadata()` at registration; returning `None` would
//!   silently degrade the schema enforcement to "anything goes" (RESEARCH
//!   §Risks #2 — threat T-83-05-01).
//! - **Constructors, never struct-literals.** Both [`ToolInfo`] and
//!   [`ToolAnnotations`] are `#[non_exhaustive]` (PATTERNS §Pattern C). The
//!   synthesizer uses [`ToolInfo::with_annotations`] / [`ToolInfo::new`] and
//!   the [`ToolAnnotations::new()`]-then-`.with_*` fluent builder.
//! - **Cognitive complexity ≤25 per function.** Decomposed into
//!   [`build_input_schema`], [`build_param_property`], and [`build_annotations`]
//!   per Phase 75 D-03 + PATTERNS §Pattern G. No `#[allow]` annotations.

use std::sync::Arc;

use async_trait::async_trait;
use pmcp::server::ToolHandler;
use pmcp::types::{ToolAnnotations, ToolInfo};
use pmcp::RequestHandlerExtra;
use serde_json::{json, Map, Value};

use crate::config::{AnnotationsDecl, ParamDecl, ServerConfig, ToolDecl, ValidationSection};
use crate::error::Result;
use crate::policy::ToolkitHooks;
// Every use of the trait object itself lives inside a `#[cfg(feature =
// "input-validation")]` item (`ValidatingToolHandler`'s field and `wrap`'s
// parameter), so an ungated import is an `unused_imports` warning — and a hard
// error — in the `--no-default-features --features http` build this module's own
// header names as the shape a consumer reaches for.
#[cfg(feature = "input-validation")]
use crate::policy::ArgumentValidator;
use crate::sql::SqlConnector;

#[cfg(feature = "http")]
use crate::error::ToolkitError;
#[cfg(feature = "http")]
use crate::http::{HttpConnector, Operation, Parameter, ParameterLocation};

#[cfg(feature = "openapi-code-mode")]
use crate::code_mode::HttpCodeExecutor;
#[cfg(feature = "openapi-code-mode")]
use pmcp_code_mode::ExecutionConfig;

/// Type alias for one synthesized tool tuple: `(name, ToolInfo, Arc<dyn ToolHandler>)`.
///
/// Exists so [`synthesize_from_config`]'s return type does not trip
/// `clippy::type_complexity` while preserving the exact `(name, ToolInfo, Arc)`
/// shape consumers register with `pmcp::ServerBuilder::tool_arc` (PATTERNS §9).
pub type SynthesizedTool = (String, ToolInfo, Arc<dyn ToolHandler>);

/// Synthesize one `ToolInfo` + handler per `[[tools]]` config entry.
///
/// Each returned tuple is `(name, ToolInfo, Arc<dyn ToolHandler>)` and is
/// ready to feed into `pmcp::ServerBuilder::tool_arc(name, handler)`. The
/// `ToolInfo` carries the full input schema (synthesized from
/// `[[tools.parameters]]`) and `ToolAnnotations` (from `[tools.annotations]`)
/// so the builder's metadata cache will never fall back to the empty schema.
///
/// # Errors
///
/// Returns [`crate::ToolkitError::Synth`] if a tool declaration is internally
/// inconsistent. The Plan 05 GREEN body never produces this error path —
/// synthesis is total over the parsed [`ServerConfig`] surface — but the
/// `Result` return is kept for forward compatibility with Plan 06 (code-mode
/// wiring) and Phase 84 (SQL backend resolution).
///
/// # Example
///
/// ```
/// use pmcp_server_toolkit::config::ServerConfig;
/// use pmcp_server_toolkit::tools::synthesize_from_config;
///
/// let cfg = ServerConfig::default();
/// let synthesized = synthesize_from_config(&cfg).unwrap();
/// assert_eq!(synthesized.len(), 0);
/// ```
pub fn synthesize_from_config(config: &ServerConfig) -> Result<Vec<SynthesizedTool>> {
    synthesize_inner(config, None, &ToolkitHooks::default())
}

/// [`synthesize_from_config`] with registered E2 hooks (Phase 128).
///
/// PUBLIC, not `pub(crate)`: `crates/pmcp-openapi-server` is a DIFFERENT crate and
/// calls the free synthesizers directly rather than going through
/// [`crate::ServerBuilderExt`], so a crate-private variant would leave the shipped
/// OpenAPI binary unable to pass hooks at all — a registration surface that exists
/// and cannot be reached from the deployment that most needs it.
///
/// # Errors
///
/// As [`synthesize_from_config`].
pub fn synthesize_from_config_and_hooks(
    config: &ServerConfig,
    hooks: &ToolkitHooks,
) -> Result<Vec<SynthesizedTool>> {
    synthesize_inner(config, None, hooks)
}

/// Synthesize tools that execute against a wired [`SqlConnector`] (Phase 84
/// CONN-01 / D-06). ADDITIVE variant alongside [`synthesize_from_config`] — the
/// existing API is unchanged and all P83 callers compile without modification.
///
/// Each synthesized [`SynthesizedToolHandler`] holds the shared `connector`, so
/// its `handle()` body calls [`SqlConnector::execute`] with the tool's declared
/// `sql` + the named parameters extracted from the validated args. When a tool
/// declares `ui_resource_uri`, the synthesized [`ToolInfo`] also carries widget
/// metadata so pmcp core's `with_widget_enrichment` populates `structuredContent`
/// (D-06) — that flip lives in the shared [`synthesize_inner`] helper and so
/// fires for both entry points.
///
/// # Errors
///
/// Returns [`crate::ToolkitError::Synth`] if a tool declaration is internally
/// inconsistent. Synthesis is total over the parsed [`ServerConfig`] surface —
/// the connector is threaded into each handler for runtime use, not consulted at
/// synthesis time.
///
/// # Example
///
/// ```no_run
/// use std::sync::Arc;
/// use pmcp_server_toolkit::config::ServerConfig;
/// use pmcp_server_toolkit::sql::SqlConnector;
/// use pmcp_server_toolkit::tools::synthesize_from_config_with_connector;
///
/// fn build(connector: Arc<dyn SqlConnector>) {
///     let cfg = ServerConfig::default();
///     let tools = synthesize_from_config_with_connector(&cfg, connector).unwrap();
///     assert_eq!(tools.len(), 0);
/// }
/// ```
pub fn synthesize_from_config_with_connector(
    config: &ServerConfig,
    connector: Arc<dyn SqlConnector>,
) -> Result<Vec<SynthesizedTool>> {
    synthesize_inner(config, Some(connector), &ToolkitHooks::default())
}

/// [`synthesize_from_config_with_connector`] with registered E2 hooks (Phase 128).
///
/// PUBLIC for the reason given on [`synthesize_from_config_and_hooks`].
///
/// # Errors
///
/// As [`synthesize_from_config_with_connector`].
pub fn synthesize_from_config_with_connector_and_hooks(
    config: &ServerConfig,
    connector: Arc<dyn SqlConnector>,
    hooks: &ToolkitHooks,
) -> Result<Vec<SynthesizedTool>> {
    synthesize_inner(config, Some(connector), hooks)
}

/// Shared synthesizer body for both [`synthesize_from_config`] (no connector)
/// and [`synthesize_from_config_with_connector`] (connector wired).
///
/// Keeps the two public entry points one-liners so the widget_meta flip (D-06)
/// and the handler construction logic are not duplicated. Decomposed per
/// PATTERNS §Pattern G — the per-tool body delegates to [`build_input_schema`],
/// [`build_annotations`], and [`apply_widget_meta`] to stay under cog 25.
fn synthesize_inner(
    config: &ServerConfig,
    connector: Option<Arc<dyn SqlConnector>>,
    hooks: &ToolkitHooks,
) -> Result<Vec<SynthesizedTool>> {
    let validation = &config.server.validation;
    let mut out = Vec::with_capacity(config.tools.len());
    for decl in &config.tools {
        let info = build_tool_info(decl, validation);
        let handler: Arc<dyn ToolHandler> = Arc::new(SynthesizedToolHandler {
            info: info.clone(),
            decl: decl.clone(),
            connector: connector.clone(),
        });
        // Push site 1 of 3 (SQL handler) — D1 enforcement + E2 validator.
        let handler = enforce_input_schema(handler, &info, decl, validation, hooks);
        out.push((decl.name.clone(), info, handler));
    }
    Ok(out)
}

/// Enforce a synthesized tool's declared `inputSchema` before its inner handler
/// runs (Phase 128, D1 / T-128-01).
///
/// # Enforcing function
///
/// `pmcp::server::schema_validation::validate_input`, called from BOTH
/// [`ValidatingToolHandler::handle`] and [`ValidatingToolHandler::handle_output`].
/// Backed by `tests/input_validation_acceptance.rs` acceptance rows 8–11, which
/// fail if either call is removed.
///
/// Wrapping happens at every handler push site in this module, so all four public
/// entry points inherit the enforcement with none forgotten.
///
/// With the `input-validation` feature OFF this is the identity function, and the
/// opt-out is logged once per synthesized tool at synthesis time: a validation rule
/// that is off must never read as on.
///
/// # `[server.validation] enforce_input_schema = false`
///
/// That flag is an operator opt-out from the SCHEMA CHECK, not from the decorator.
/// `ValidatingToolHandler` carries an `enforce_schema: bool` and skips only the
/// `validate_input` call, so a future explicitly-registered argument validator
/// living in the same decorator keeps running. Turning off one enforcement must
/// never silently turn off another — and losing an explicit custom validator
/// because schema enforcement was disabled is the worst form of that class, since
/// the operator turned off A and lost B without being told.
///
/// The decorator is constructed when schema enforcement is ON **or** a validator is
/// registered for this tool. The registry is [`ToolkitHooks`], consulted below via
/// [`ToolkitHooks::argument_validator_for`], so the decorator is skipped entirely
/// only when `enforce_input_schema = false` AND no validator is registered for this
/// tool name — which preserves the "no added allocation for a server that uses
/// neither" property.
///
/// Until Phase 128 plan 09 this paragraph read "No validator registry exists yet
/// (it lands with E2), so `has_registered_validator` is `false` today". That
/// sentence outlived the registry by one plan and is kept named here rather than
/// silently deleted: it is the documented-but-absent defect this phase exists to
/// close, INVERTED — a comment in the very function that closes the gap still
/// describing the machinery as absent. The binding it named no longer exists.
///
/// Backed by `tools::argument_validator_seam::a_registered_validator_still_runs_with_enforce_input_schema_false`,
/// which fails if the `registered_validator.is_none()` conjunct is dropped from the
/// early return, and by
/// `tools::argument_validator_seam::a_registered_validator_refuses_a_combination_the_schema_permits`,
/// which fails if the lookup is removed.
#[cfg(feature = "input-validation")]
fn enforce_input_schema(
    handler: Arc<dyn ToolHandler>,
    info: &ToolInfo,
    decl: &ToolDecl,
    validation: &ValidationSection,
    hooks: &ToolkitHooks,
) -> Arc<dyn ToolHandler> {
    // Phase 128 E2 — the registry lookup that replaced plan 03's
    // `let has_registered_validator = false;` placeholder.
    let registered_validator = hooks.argument_validator_for(&decl.name);
    if !validation.enforce_input_schema && registered_validator.is_none() {
        tracing::warn!(
            tool = %decl.name,
            "[server.validation] enforce_input_schema = false: this tool's arguments are \
             NOT checked against its declared inputSchema before the backend call"
        );
        return handler;
    }
    ValidatingToolHandler::wrap(
        handler,
        info,
        decl,
        validation.enforce_input_schema,
        registered_validator,
    )
}

/// The `input-validation`-off half of [`enforce_input_schema`]: the identity
/// function, with the absence logged once per synthesized tool.
///
/// A `#[cfg]` SIBLING rather than one function with two inner blocks, because the
/// off arm reads none of the schema inputs and the single-function form needed an
/// `#[allow(unused_variables)]` — which this module's own header rules out ("No
/// `#[allow]` annotations"). This is the same sibling-pair idiom
/// `check_tool_input_schema_compiles`, `warn_if_pattern_checking_unavailable`,
/// `spec_placeholder_rules` and `lint_declared_cap_above_placeholder_floor`
/// already use in this crate for exactly this situation.
#[cfg(not(feature = "input-validation"))]
fn enforce_input_schema(
    handler: Arc<dyn ToolHandler>,
    _info: &ToolInfo,
    decl: &ToolDecl,
    _validation: &ValidationSection,
    hooks: &ToolkitHooks,
) -> Arc<dyn ToolHandler> {
    tracing::warn!(
        tool = %decl.name,
        "the `input-validation` feature is OFF: this tool's arguments are NOT checked \
         against its declared inputSchema before the backend call"
    );
    // `ToolkitHooks::argument_validator_for` is NOT feature-gated, so
    // `with_argument_validator` compiles and appears to succeed in this build —
    // while `ValidatingToolHandler`, the only thing that can RUN a registered
    // validator, is gated away. Naming the drop is the difference between an
    // enforcement that is off and one that silently reads as on; the schema warning
    // above does not cover it, because a registered validator is a SEPARATE switch
    // (see `ValidatingToolHandler::enforce_schema`).
    if hooks.argument_validator_for(&decl.name).is_some() {
        tracing::warn!(
            tool = %decl.name,
            "an E2 ArgumentValidator IS registered for this tool and is DISCARDED: \
             running it requires the `input-validation` feature"
        );
    }
    handler
}

/// Decorator that refuses a `tools/call` whose arguments violate the tool's
/// declared `inputSchema`, WITHOUT invoking the inner handler.
///
/// Crate-private by design: the validator and the value-free refusal renderer both
/// live in core `pmcp` (D-01), so this type holds no schema logic of its own.
#[cfg(feature = "input-validation")]
struct ValidatingToolHandler {
    inner: Arc<dyn ToolHandler>,
    /// The synthesized `ToolInfo`'s `input_schema`, verbatim.
    input_schema: Value,
    /// `input_schema.to_string()`, computed ONCE at synthesis so the hot
    /// `tools/call` path never re-serializes the whole schema to hit the core
    /// validator cache.
    schema_key: String,
    /// Declared parameter names in declaration order — the ONLY names a refusal
    /// message may echo (SC-7).
    declared: Vec<String>,
    /// Whether the SCHEMA CHECK runs, from `[server.validation]
    /// enforce_input_schema`.
    ///
    /// A SEPARATE switch from the decorator's existence on purpose: when the E2
    /// argument-validator registry lands in the same decorator, an operator who
    /// turns schema enforcement off must not silently lose an
    /// explicitly-registered validator too.
    enforce_schema: bool,
    /// The E2 [`ArgumentValidator`] registered for this tool, if any (Phase 128).
    ///
    /// Looked up by tool NAME once at synthesis, so the `tools/call` path does no
    /// map lookup. `None` is the overwhelmingly common case and costs one
    /// `Option` discriminant test per call.
    validator: Option<Arc<dyn ArgumentValidator>>,
}

#[cfg(feature = "input-validation")]
impl ValidatingToolHandler {
    /// Wrap `handler`, taking the schema from the already-built `info` and the
    /// declared parameter names from `decl`.
    fn wrap(
        inner: Arc<dyn ToolHandler>,
        info: &ToolInfo,
        decl: &ToolDecl,
        enforce_schema: bool,
        validator: Option<Arc<dyn ArgumentValidator>>,
    ) -> Arc<dyn ToolHandler> {
        let input_schema = info.input_schema.clone();
        let schema_key = input_schema.to_string();
        Arc::new(Self {
            inner,
            input_schema,
            schema_key,
            declared: decl.parameters.iter().map(|p| p.name.clone()).collect(),
            enforce_schema,
            validator,
        })
    }

    /// Validate `args`, mapping any violation to a value-free
    /// `pmcp::Error::ToolRejected`. Kept a separate helper so both trait entry
    /// points stay one-liners and well under the cog-25 gate.
    ///
    /// # Why a tool-level rejection and not `Error::Validation`
    ///
    /// A refused argument is the CALLER's to fix, and `Error::ToolRejected` is the
    /// SDK's variant for exactly that: its own docs name "schema-mismatched input"
    /// as an intended use, and `execute_code` already returns it for a bad approval
    /// token. `Error::Validation` surfaces as a JSON-RPC error, which reads to a
    /// model as a server fault and gives it nothing to retry with. Measured on the
    /// 2.21.0 stack: 14 of 15 refusals in a 24-probe matrix reached the client as
    /// protocol errors, and the Triarii UMLS server wrote an adapter to remap them.
    /// The message is unchanged and stays value-free; only the channel moves.
    ///
    /// Returns `Ok(())` without consulting the schema when `enforce_schema` is
    /// `false` — skipping the CHECK, not the decorator, so anything else this
    /// decorator does still happens.
    fn check(&self, args: &Value) -> pmcp::Result<()> {
        self.check_schema(args)?;
        self.check_registered_validator(args)
    }

    /// The D1 declared-schema check.
    ///
    /// Returns `Ok(())` without consulting the schema when `enforce_schema` is
    /// `false` — skipping the CHECK, not the decorator, so the E2 validator below
    /// still runs.
    fn check_schema(&self, args: &Value) -> pmcp::Result<()> {
        use pmcp::server::schema_validation::{render_refusal, validate_input};

        if !self.enforce_schema {
            return Ok(());
        }
        validate_input(&self.input_schema, Some(args), Some(&self.schema_key)).map_err(
            |violations| {
                let declared: Vec<&str> = self.declared.iter().map(String::as_str).collect();
                pmcp::Error::tool_rejected(render_refusal(&violations, &declared), None)
            },
        )
    }

    /// The E2 registered-validator check (Phase 128).
    ///
    /// Called by [`Self::check`] STRICTLY AFTER [`Self::check_schema`] returns
    /// `Ok`, which is the E2 ordering contract: a custom rule never sees arguments
    /// the declared schema already refused (T-128-41). Inverting the two calls is
    /// what `a_registered_validator_is_not_invoked_when_the_schema_refuses` fails
    /// on.
    ///
    /// The refusal message is the VALIDATOR's own, authored outside this crate, so
    /// unlike `render_refusal` above the toolkit cannot guarantee it is value-free.
    /// That residual is documented on [`crate::policy::ArgumentRefusal`].
    fn check_registered_validator(&self, args: &Value) -> pmcp::Result<()> {
        let Some(validator) = self.validator.as_ref() else {
            return Ok(());
        };
        validator
            .validate(args)
            .map_err(|refusal| pmcp::Error::tool_rejected(refusal.message().to_string(), None))
    }
}

#[cfg(feature = "input-validation")]
#[async_trait]
impl ToolHandler for ValidatingToolHandler {
    async fn handle(&self, args: Value, extra: RequestHandlerExtra) -> pmcp::Result<Value> {
        self.check(&args)?;
        self.inner.handle(args, extra).await
    }

    fn metadata(&self) -> Option<ToolInfo> {
        self.inner.metadata()
    }

    /// Validates, then DELEGATES to the inner handler's own `handle_output`.
    ///
    /// Implementing only `handle` would silently replace an inner handler's
    /// `handle_output` override with the trait default
    /// (`handle(..).map(ToolOutput::Payload)`), stripping that handler's ability to
    /// own its `CallToolResult` envelope. A decorator must not narrow the contract
    /// of what it wraps — and both entry points must validate, or an inner handler
    /// that overrides `handle_output` becomes an unvalidated path.
    async fn handle_output(
        &self,
        args: Value,
        extra: RequestHandlerExtra,
    ) -> pmcp::Result<pmcp::server::ToolOutput> {
        self.check(&args)?;
        self.inner.handle_output(args, extra).await
    }
}

/// Flip widget metadata onto `info` when the declaration carries a
/// `ui_resource_uri` (D-06 / REVIEWS M1).
///
/// Uses the feature-independent [`ToolInfo::with_meta_entry`] surface to insert
/// `_meta.ui.resourceUri`. This is the verified-correct API: `with_widget_meta`
/// is gated on pmcp's `mcp-apps` feature (which the toolkit does not enable),
/// whereas `with_meta_entry` is always available and produces the `ui.resourceUri`
/// shape that `ToolInfo::widget_meta()` recognises — so pmcp core's
/// `with_widget_enrichment` populates `structuredContent`. Annotations on `info`
/// are preserved (chained, not reconstructed).
fn apply_widget_meta(info: ToolInfo, decl: &ToolDecl) -> ToolInfo {
    match decl.ui_resource_uri.as_deref() {
        Some(uri) => info.with_meta_entry("ui", json!({ "resourceUri": uri })),
        None => info,
    }
}

/// Build the [`ToolInfo`] for a synthesized tool from its declaration.
///
/// The schema + annotations + widget-meta sequence is identical for every tool
/// kind (single-call HTTP, SQL, and script), so it lives here once — keeping the
/// `#[non_exhaustive]` [`ToolInfo`] constructor discipline (the `with_annotations`
/// vs `new` arms) in a single place rather than copy-pasted per synthesizer.
fn build_tool_info(decl: &ToolDecl, validation: &ValidationSection) -> ToolInfo {
    let schema = build_input_schema(decl, validation);
    let annotations = build_annotations(decl.annotations.as_ref());
    let base = match annotations {
        Some(ann) => {
            ToolInfo::with_annotations(decl.name.clone(), decl.description.clone(), schema, ann)
        },
        None => ToolInfo::new(decl.name.clone(), decl.description.clone(), schema),
    };
    apply_widget_meta(base, decl)
}

/// Build the JSON Schema `properties` + `required` envelope from a
/// `[[tools.parameters]]` list.
///
/// Decomposed from [`synthesize_from_config`] to keep cognitive complexity ≤25
/// (Phase 75 D-03 + PATTERNS §Pattern G).
///
/// `pub(crate)` since Phase 128 SC-2: [`crate::config::ServerConfig::validate`]
/// builds each tool's schema through THIS function and compiles it at config time,
/// so the schema the compile gate checks is byte-identical to the one the runtime
/// validator enforces. A second construction path there would let the gate pass a
/// schema the server never serves.
///
/// Takes the whole [`ToolDecl`] rather than just its parameters because the D3
/// default cap is POSITION-scoped, and position is a property of the tool's `path`
/// and `method` — see [`crate::config::ToolDecl::param_position`].
///
/// `properties` and `required` are emitted in `[[tools.parameters]]` declaration
/// order, so two synthesis runs over one config produce byte-identical output.
pub(crate) fn build_input_schema(decl: &ToolDecl, validation: &ValidationSection) -> Value {
    let mut props = Map::new();
    let mut required = Vec::new();
    for p in &decl.parameters {
        let mut prop = build_param_property(p);
        apply_position_cap(&mut prop, p, decl.param_position(&p.name), validation);
        props.insert(p.name.clone(), prop);
        if p.required {
            required.push(Value::String(p.name.clone()));
        }
    }
    json!({
        "type": "object",
        "properties": props,
        "required": required,
        "additionalProperties": validation.additional_properties,
    })
}

/// Apply the D3 position-scoped default length cap to ONE already-built property
/// object (Phase 128).
///
/// Emits `maxLength = validation.default_max_length` only when ALL of:
/// the parameter's effective type is `string`; it declares no `max_length` of its
/// own; `default_max_length` is non-zero; and `position` is `Path` or `Query`.
/// BODY position emits nothing — capping a `POST` payload's free-text field is
/// exactly the breakage D-05 exists to avoid, and `ServerConfig::lint` surfaces
/// those instead.
///
/// A DECLARED `max_length` is never overridden and never merged with the default in
/// either direction: the author's number wins outright, and the always-on
/// `PLACEHOLDER_MAX_LENGTH` floor is what bounds a URL segment regardless.
///
/// The four conditions live in `crate::config::default_cap_applies` rather than
/// here, so `lint()` and this function cannot disagree about which parameters are
/// covered — a disagreement would make `lint()` report a parameter as uncapped
/// while the schema caps it, or the reverse.
///
/// Its own free function (not inlined into [`build_input_schema`]) to keep that
/// function's cognitive complexity well under the 25 gate.
fn apply_position_cap(
    prop: &mut Value,
    p: &ParamDecl,
    position: crate::config::ParamPosition,
    validation: &ValidationSection,
) {
    if crate::config::default_cap_applies(p, position, validation) {
        prop["maxLength"] = json!(validation.default_max_length);
    }
}

/// Build a single JSON Schema property object from a [`ParamDecl`].
///
/// Per-parameter constraints (`minimum`, `maximum`, `maxLength`, `minLength`,
/// `pattern`, `format`, `maxItems`, `items`, `default`, `enum`) are folded in only
/// when present — an undeclared keyword is ABSENT from the emitted schema, never
/// present-and-empty, because an empty `pattern` matches everything and an empty
/// `items` object constrains nothing while both read like rules.
///
/// The param's `param_type` defaults to `"string"` when omitted in TOML to match
/// JSON Schema's permissive default.
fn build_param_property(p: &ParamDecl) -> Value {
    let ty = p.param_type.as_deref().unwrap_or("string");
    let mut prop = json!({ "type": ty });
    if let Some(desc) = &p.description {
        prop["description"] = Value::String(desc.clone());
    }
    if let Some(min) = p.minimum {
        prop["minimum"] = json!(min);
    }
    if let Some(max) = p.maximum {
        prop["maximum"] = json!(max);
    }
    if let Some(max_len) = p.max_length {
        prop["maxLength"] = json!(max_len);
    }
    if let Some(min_len) = p.min_length {
        prop["minLength"] = json!(min_len);
    }
    if let Some(pattern) = &p.pattern {
        prop["pattern"] = Value::String(pattern.clone());
    }
    if let Some(format) = &p.format {
        prop["format"] = Value::String(format.clone());
    }
    if let Some(max_items) = p.max_items {
        prop["maxItems"] = json!(max_items);
    }
    if let Some(items) = &p.items {
        prop["items"] = build_items_property(items);
    }
    if let Some(default) = &p.default {
        // toml::Value serializes losslessly into serde_json::Value via serde.
        if let Ok(v) = serde_json::to_value(default) {
            prop["default"] = v;
        }
    }
    if let Some(enum_vals) = &p.enum_values {
        if let Ok(v) = serde_json::to_value(enum_vals) {
            prop["enum"] = v;
        }
    }
    prop
}

/// Build the OBJECT-form JSON Schema `items` value from an [`ItemsDecl`]
/// (Phase 128, D2).
///
/// A separate free function for two reasons. First, cognitive complexity:
/// [`build_param_property`] gained five arms in this phase and RESEARCH assumption
/// A4 (that they stay under cog 25) was explicitly unmeasured, so the construction
/// lives outside it. Second, and more importantly, there is exactly ONE place that
/// can decide the SHAPE of `items`, and it returns a [`Value::Object`]
/// unconditionally. The array form of `items` is a draft-07 tuple construct that
/// does NOT compile under the Draft 2020-12 pin, and a schema that does not compile
/// takes the whole tool's validator down — so "never an array here" is a safety
/// property, not a style preference, and it is enforced by this function having no
/// code path that builds a JSON array.
fn build_items_property(items: &crate::config::ItemsDecl) -> Value {
    let ty = items.item_type.as_deref().unwrap_or("string");
    let mut prop = json!({ "type": ty });
    if let Some(max_len) = items.max_length {
        prop["maxLength"] = json!(max_len);
    }
    if let Some(pattern) = &items.pattern {
        prop["pattern"] = Value::String(pattern.clone());
    }
    prop
}

/// Build [`ToolAnnotations`] from an optional `[tools.annotations]` block.
///
/// Per PATTERNS §Pattern C, the constructor + fluent builder is used (never a
/// struct literal — [`ToolAnnotations`] is `#[non_exhaustive]`). The `cost_hint`
/// field has no `ToolAnnotations` accessor and is therefore not propagated at
/// this layer; it lives on the toolkit's [`AnnotationsDecl`] and is consumed
/// by future plans that surface cost into rate-limiting policy.
fn build_annotations(decl: Option<&AnnotationsDecl>) -> Option<ToolAnnotations> {
    let d = decl?;
    let a = ToolAnnotations::new()
        .with_read_only(d.read_only_hint)
        .with_destructive(d.destructive_hint)
        .with_idempotent(d.idempotent_hint)
        .with_open_world(d.open_world_hint);
    Some(a)
}

// -----------------------------------------------------------------------------
// SynthesizedToolHandler — crate-private
// -----------------------------------------------------------------------------

/// Crate-private handler wrapping a synthesized [`ToolInfo`].
///
/// [`ToolHandler::metadata`] MUST return `Some(self.info.clone())` — Phase 82's
/// `tool_arc` consumes `handler.metadata()` at registration time; returning
/// `None` would cause the builder to fall back to an empty schema (RESEARCH
/// §Risks #2 — threat T-83-05-01). The unit + property tests in
/// [`crate::tools`] and `tests/tool_synthesis_props.rs` lock this in.
///
/// `handle()` reads the declared `sql`, extracts named parameters from the
/// validated args, and calls [`SqlConnector::execute`] when a `connector` is
/// wired (handlers built via [`synthesize_from_config_with_connector`]).
/// Handlers built via the no-connector [`synthesize_from_config`] carry
/// `connector = None` and return an explicit `Err` on invocation — preserving
/// P83 behaviour where the no-connector path was test-only (T-84-03-05). The
/// `decl` is held so the handler can read `sql` / `ui_resource_uri` /
/// `parameters` without re-walking the config.
struct SynthesizedToolHandler {
    info: ToolInfo,
    decl: ToolDecl,
    /// `Some` only for handlers built via [`synthesize_from_config_with_connector`].
    connector: Option<Arc<dyn SqlConnector>>,
}

/// Extract the named `(name, value)` parameter pairs the connector binds from,
/// filtering the caller's `args` against the declared parameter list
/// (T-84-03-01: only declared parameter names reach `execute()`; an extra key is
/// dropped HERE, and separately refused one layer out).
///
/// The refusal is not "upstream". It is
/// `pmcp::server::schema_validation::validate_input` inside the
/// `ValidatingToolHandler` decorator that [`enforce_input_schema`] wraps around
/// this handler at its push site, so an undeclared key normally never reaches this
/// function at all. Conditional on the `input-validation` feature AND on
/// `[server.validation]` leaving both `enforce_input_schema` (`true`) and
/// `additional_properties` (`false`) at their defaults — with `additional_properties
/// = true` an undeclared key is ADMITTED by the schema and this filter is the only
/// thing that keeps it out of `execute()`. That is why the drop stays here rather
/// than being replaced by an assertion: the two layers close the case under
/// different configurations. Backed by
/// `tests/input_validation_acceptance.rs::input_validation_refuses_undeclared_argument_without_contacting_upstream`
/// for the schema layer; the filter layer is covered by this module's
/// `extract_params` unit tests.
///
/// Found by this phase's SC-6 candidate-phrase sweep, not by the plan's named
/// list: the phrasing "rejects them upstream" is the same defect class as
/// the retired `HttpToolHandler::handle` claim one function-group away, which
/// located the same envelope check "upstream" in the same way.
///
/// The retired phrasings are deliberately NOT quoted verbatim anywhere in `src/`.
/// This phase's SC-6 gates grep for each retired phrase and require a count of
/// zero, and a gate that a historical citation can satisfy is a gate that has
/// stopped distinguishing a surviving claim from a note about one — the first
/// draft of THIS comment quoted one of them and turned the gate red on itself,
/// which is the cheapest possible demonstration that the gate is sensitive. The
/// full before/after text lives in `128-10-SUMMARY.md`.
///
/// When the caller omits an optional parameter that declares a `default`, the
/// default is applied so the bound SQL sees a concrete value. Without this an
/// omitted `:limit` / `:offset` would bind as unbound `NULL` and SQLite rejects
/// `LIMIT NULL` with a "datatype mismatch" — so the declared default is the
/// difference between a working and a broken tool call (the reference
/// `search_tracks` / `list_artists` calls rely on it).
///
/// An EXPLICIT JSON `null` for a declared-default parameter is treated the SAME
/// as an omitted parameter — the declared default is applied (85-10 WR-02
/// secondary fix). Without the `is_null` filter a caller sending
/// `{"limit": null}` would bind `LIMIT NULL` and SQLite would reject the query
/// with "datatype mismatch", even though the tool declares `default = 20`.
fn extract_named_params(decl: &ToolDecl, args: &Value) -> Vec<(String, Value)> {
    decl.parameters
        .iter()
        .filter_map(|p| {
            args.get(&p.name)
                // Explicit JSON `null` falls through to the declared default,
                // exactly like an omitted key (no `LIMIT NULL` bind).
                .filter(|v| !v.is_null())
                .cloned()
                .or_else(|| {
                    p.default
                        .as_ref()
                        .and_then(|d| serde_json::to_value(d).ok())
                })
                .map(|v| (p.name.clone(), v))
        })
        .collect()
}

#[async_trait]
impl ToolHandler for SynthesizedToolHandler {
    async fn handle(&self, args: Value, _extra: RequestHandlerExtra) -> pmcp::Result<Value> {
        let sql = self.decl.sql.as_deref().ok_or_else(|| {
            pmcp::Error::Internal(format!("tool '{}' has no `sql` declared", self.info.name))
        })?;
        let connector = self.connector.as_ref().ok_or_else(|| {
            pmcp::Error::Internal(format!(
                "tool '{}' requires connector wiring — build via synthesize_from_config_with_connector",
                self.info.name
            ))
        })?;
        let named_params = extract_named_params(&self.decl, &args);
        // T-84-03-02: format!("{e}") uses ConnectorError::Display, which Plan 01
        // Task 2 guarantees does not echo credentials.
        let rows = connector
            .execute(sql, &named_params)
            .await
            .map_err(|e| pmcp::Error::Internal(format!("connector error: {e}")))?;
        Ok(Value::Array(rows))
    }

    fn metadata(&self) -> Option<ToolInfo> {
        Some(self.info.clone())
    }
}

// -----------------------------------------------------------------------------
// Single-call HTTP synthesizer (Phase 90 OAPI-02a) — feature `http`
// -----------------------------------------------------------------------------

/// Synthesize one `ToolInfo` + handler per single-call `[[tools]]` entry,
/// executing against a wired [`HttpConnector`] (Phase 90 OAPI-02a / D-01).
///
/// Mirrors [`synthesize_from_config_with_connector`] (the SQL analog) in shape.
/// For each `[[tools]]` where [`ToolDecl::is_script_tool`] is `false` and a
/// `path` + `method` pair is present, an [`Operation`] is built from the path
/// template (the `{...}` segments become path parameters), the declared
/// `[[tools.parameters]]` (non-path params become query params; `POST`/`PUT`/
/// `PATCH` carry a request body), and the per-tool `base_url` override — which is
/// reflected onto the [`Operation`] so the connector targets the per-tool host
/// (never silently dropped, Codex MEDIUM). The synthesized [`ToolInfo`] uses the
/// EXISTING [`build_input_schema`] (object envelope + `additionalProperties:false`)
/// and [`build_annotations`] helpers; the handler calls
/// [`HttpConnector::execute`] and returns the JSON.
///
/// # Script-tool seam (Plan 05)
///
/// A `script` tool encountered here returns a typed [`ToolkitError::Synth`] — it
/// is an EXPLICIT, clearly-marked seam, NOT a silent skip and NOT a `todo!()`.
/// Plan 05 widens this function's signature (adding the shared `http_exec` +
/// `exec_config`) and fills the `is_script_tool()` arm with a `ScriptToolHandler`
/// branch; that change is a localized, anticipated edit because the seam is
/// surfaced here.
///
/// # Errors
///
/// Returns [`ToolkitError::Synth`] when a `[[tools]]` entry is a `script` tool
/// (Plan 05 seam) or is neither a valid single-call (missing `path` OR `method`)
/// nor a script tool (T-90-03-04 negative validation — an ill-formed tool is
/// rejected, never silently registered).
#[cfg(feature = "http")]
pub fn synthesize_from_config_with_http_connector(
    config: &ServerConfig,
    connector: Arc<dyn HttpConnector>,
) -> Result<Vec<SynthesizedTool>> {
    // No script-tool builder is supplied on this (single-call-only) entry point,
    // so the `is_script_tool()` arm of [`synthesize_http_inner`] returns the
    // typed Plan 05 seam error. The OpenAPI Code Mode build calls
    // [`synthesize_from_config_with_http_connector_and_scripts`], which supplies
    // a [`ScriptToolHandler`] builder so a `script` tool synthesizes a real
    // handler over the shared engine (OAPI-02b / D-01 / D-02).
    synthesize_from_config_with_http_connector_and_hooks(
        config,
        connector,
        &ToolkitHooks::default(),
    )
}

/// [`synthesize_from_config_with_http_connector`] with registered E2 hooks
/// (Phase 128).
///
/// PUBLIC for the reason given on [`synthesize_from_config_and_hooks`].
///
/// # Errors
///
/// As [`synthesize_from_config_with_http_connector`].
#[cfg(feature = "http")]
pub fn synthesize_from_config_with_http_connector_and_hooks(
    config: &ServerConfig,
    connector: Arc<dyn HttpConnector>,
    hooks: &ToolkitHooks,
) -> Result<Vec<SynthesizedTool>> {
    // No script-tool builder is supplied on this (single-call-only) entry point,
    // so the `is_script_tool()` arm of [`synthesize_http_inner`] returns the
    // typed Plan 05 seam error. The OpenAPI Code Mode build calls
    // [`synthesize_from_config_with_http_connector_and_scripts`], which supplies
    // a [`ScriptToolHandler`] builder so a `script` tool synthesizes a real
    // handler over the shared engine (OAPI-02b / D-01 / D-02).
    synthesize_http_inner(
        config,
        connector,
        |decl| {
            Err(ToolkitError::Synth(format!(
                "tool '{}' is a script tool — script tools require the `openapi-code-mode` \
                 feature (use synthesize_from_config_with_http_connector_and_scripts)",
                decl.name
            )))
        },
        hooks,
    )
}

/// Synthesize single-call AND script `[[tools]]` against a wired
/// [`HttpConnector`] plus a shared [`HttpCodeExecutor`] + [`ExecutionConfig`]
/// (Phase 90 OAPI-02b / D-01 / D-02).
///
/// This is the OpenAPI Code Mode entry point (gated `openapi-code-mode`): it
/// adds the `http_exec` + `exec_config` the script-tool path needs, threading
/// the SAME `HttpCodeExecutor` instance that feeds Code Mode (D-02 — one engine,
/// two surfaces). Single-call tools synthesize exactly as in
/// [`synthesize_from_config_with_http_connector`]; a `script` tool synthesizes a
/// [`ScriptToolHandler`] that compiles + runs the embedded JS through the SAME
/// `PlanCompiler` + `PlanExecutor` + `HttpCodeExecutor` seam Code Mode uses,
/// with NO validate/token cycle (admin-authored, `ExecutionConfig`-bounded —
/// Pitfall 7).
///
/// The binary (Plan 06) supplies `http_exec` (built once over the resolved
/// backend `base_url` + auth provider) and `exec_config` (from the
/// `[code_mode.limits]` / defaults: `max_api_calls=50`, `max_loop_iterations=100`,
/// `timeout_seconds=30`).
///
/// # Errors
///
/// Returns [`ToolkitError::Synth`] when a `[[tools]]` entry is neither a valid
/// single-call (missing `path` OR `method`) nor a script tool (T-90-03-04
/// negative validation), or when a script tool fails to build its `ToolInfo`.
#[cfg(feature = "openapi-code-mode")]
pub fn synthesize_from_config_with_http_connector_and_scripts(
    config: &ServerConfig,
    connector: Arc<dyn HttpConnector>,
    http_exec: HttpCodeExecutor,
    exec_config: ExecutionConfig,
) -> Result<Vec<SynthesizedTool>> {
    synthesize_from_config_with_http_connector_and_scripts_and_hooks(
        config,
        connector,
        http_exec,
        exec_config,
        &ToolkitHooks::default(),
    )
}

/// [`synthesize_from_config_with_http_connector_and_scripts`] with registered E2
/// hooks (Phase 128).
///
/// This is THE variant `crates/pmcp-openapi-server`'s `build_server` calls, which
/// is why it is PUBLIC rather than `pub(crate)`: that crate reaches the
/// synthesizer directly and never through [`crate::ServerBuilderExt`], so a
/// crate-private variant would ship E2 as present-but-inert on this phase's
/// principal HTTP surface (T-128-39b).
///
/// # Errors
///
/// As [`synthesize_from_config_with_http_connector_and_scripts`].
#[cfg(feature = "openapi-code-mode")]
pub fn synthesize_from_config_with_http_connector_and_scripts_and_hooks(
    config: &ServerConfig,
    connector: Arc<dyn HttpConnector>,
    http_exec: HttpCodeExecutor,
    exec_config: ExecutionConfig,
    hooks: &ToolkitHooks,
) -> Result<Vec<SynthesizedTool>> {
    let validation = &config.server.validation;
    // Script tools re-check the class policy on every resolved request, when
    // the config sets an operation-class key.
    let http_exec = match config.code_mode.as_ref() {
        Some(section) if !section.class_keys_set().is_empty() => {
            http_exec.with_class_policy(section)
        },
        _ => http_exec,
    };
    synthesize_http_inner(
        config,
        connector,
        |decl| {
            let handler =
                ScriptToolHandler::new(decl, http_exec.clone(), exec_config.clone(), validation)?;
            let info = handler.tool_info.clone();
            let arc: Arc<dyn ToolHandler> = Arc::new(handler);
            Ok((info, arc))
        },
        hooks,
    )
}

/// Refuse a curated tool the `[code_mode]` class policy refuses: a script
/// tool by its calls, a single-call tool by its `(method, path)`. A tool that
/// is neither is left to the synthesizer's own validation. No gate (no class
/// key set) refuses nothing.
#[cfg(feature = "openapi-code-mode")]
fn check_curated_tool(
    gate: Option<&crate::code_mode::ClassGate>,
    decl: &ToolDecl,
) -> std::result::Result<(), crate::error::ConfigValidationError> {
    let Some(gate) = gate else {
        return Ok(());
    };
    if let Some(script) = decl.script.as_deref() {
        return gate.check_script(&decl.name, script);
    }
    match (decl.method.as_deref(), decl.path.as_deref()) {
        (Some(method), Some(path)) => gate.check_single_call(&decl.name, method, path),
        _ => Ok(()),
    }
}

/// Shared synthesizer body for the single-call HTTP entry points.
///
/// `build_script_tool` is invoked for each `script` tool: the single-call-only
/// entry point passes a closure that returns the typed Plan 05 / `openapi-code-mode`
/// seam error, while the OpenAPI Code Mode entry point passes a closure that
/// constructs a [`ScriptToolHandler`]. Decomposed per PATTERNS §Pattern G to keep
/// the per-tool loop body under cog ≤25.
#[cfg(feature = "http")]
fn synthesize_http_inner(
    config: &ServerConfig,
    connector: Arc<dyn HttpConnector>,
    mut build_script_tool: impl FnMut(&ToolDecl) -> Result<(ToolInfo, Arc<dyn ToolHandler>)>,
    hooks: &ToolkitHooks,
) -> Result<Vec<SynthesizedTool>> {
    let validation = &config.server.validation;
    // A curated tool the `[code_mode]` class policy refuses could never
    // succeed, so it fails the boot. Applies only when the config sets an
    // operation-class key (see `ClassGate::for_curated_tools`).
    #[cfg(feature = "openapi-code-mode")]
    let class_gate = crate::code_mode::ClassGate::for_curated_tools(config);
    let mut out = Vec::with_capacity(config.tools.len());
    for decl in &config.tools {
        #[cfg(feature = "openapi-code-mode")]
        check_curated_tool(class_gate.as_ref(), decl)?;
        if decl.is_script_tool() {
            let (info, handler) = build_script_tool(decl)?;
            // Push site 2 of 3 (SCRIPT tool) — D1 enforcement. This site is
            // SEPARATE from the HTTP push below because of the `continue`; wrapping
            // only the HTTP push would leave every script tool unvalidated.
            let handler = enforce_input_schema(handler, &info, decl, validation, hooks);
            out.push((decl.name.clone(), info, handler));
            continue;
        }

        // Single-call requires BOTH path and method. A `[[tools]]` that is
        // neither a valid single-call nor a script tool is rejected (T-90-03-04).
        let (path, method) = match (decl.path.as_deref(), decl.method.as_deref()) {
            (Some(p), Some(m)) => (p, m),
            _ => {
                return Err(ToolkitError::Synth(format!(
                    "tool '{}' is not a valid single-call tool: both `path` and `method` are required",
                    decl.name
                )));
            },
        };

        let operation = build_operation(path, method, decl);
        let info = build_tool_info(decl, validation);
        let handler: Arc<dyn ToolHandler> = Arc::new(HttpToolHandler {
            info: info.clone(),
            operation,
            connector: connector.clone(),
        });
        // Push site 3 of 3 (single-call HTTP handler) — D1 enforcement + E2 validator.
        let handler = enforce_input_schema(handler, &info, decl, validation, hooks);
        out.push((decl.name.clone(), info, handler));
    }
    Ok(out)
}

/// Build the [`Operation`] for a single-call tool from its `path` template,
/// `method`, declared parameters, and per-tool `base_url`.
///
/// Path parameters are the `{...}` segments of the path template. Every other
/// declared `[[tools.parameters]]` becomes a QUERY parameter, or — when the method
/// carries a request body (`POST`/`PUT`/`PATCH`) — a
/// [`ParameterLocation::Body`] parameter that
/// [`crate::http::HttpClient`]'s `build_body` folds into the JSON payload. The
/// per-tool `base_url` is reflected onto the [`Operation`] (Codex MEDIUM — never
/// dropped).
///
/// # Agreement with `ToolDecl::param_position` — structural, not documentary
///
/// Both splits this function performs read a shared helper, so neither can drift
/// from [`crate::config::ToolDecl::param_position`]:
///
/// - the PATH split and `param_position`'s `Path` arm both read
///   [`crate::config::path_placeholder_names`];
/// - the QUERY/BODY split and `param_position`'s `Query`/`Body` arms both read
///   `crate::config::method_carries_request_body`, which is also the sole source of
///   [`Operation::has_request_body`].
///
/// That second agreement is Phase 128 CR-02/CR-03. Before it, this function marked
/// every non-path declared parameter `ParameterLocation::Query` regardless of
/// method while `param_position` classified a mutating tool's as `Body`, and the
/// divergence was documented here as deliberate. It was not survivable: a `POST`
/// tool's declared parameters travelled in the URL — so the D3 cap was withheld
/// from values that genuinely land in a request line — and `build_body`, which
/// collects only args ABSENT from `operation.parameters`, found nothing, so no
/// payload reached the backend at all once `additionalProperties: false` began to
/// be enforced.
///
/// D-05 is still honoured, and now by the routing rather than despite it: a
/// `Body`-located parameter really is a payload field, and
/// `crate::config::default_cap_applies` emits no default `maxLength` for it.
#[cfg(feature = "http")]
fn build_operation(path: &str, method: &str, decl: &ToolDecl) -> Operation {
    let method_upper = method.to_uppercase();
    let path_param_names: Vec<&str> = crate::config::path_placeholder_names(path).collect();

    let mut parameters = Vec::with_capacity(decl.parameters.len());
    // Path params (template `{...}` segments) — always required.
    //
    // Phase 128 D4(b): this loop iterates the TEMPLATE, not `decl.parameters`, so
    // the declared narrowing is looked up BY NAME and falls back to "no declared
    // rules" when the template names a segment the config never declared. Such a
    // parameter still gets the unconditional character floor and the always-on cap
    // at substitution time — it simply gets no narrowing on top (D-10).
    for name in &path_param_names {
        let declared = decl.parameters.iter().find(|p| p.name == **name);
        parameters.push(
            Parameter::new((*name).to_string(), ParameterLocation::Path, true).with_rules(
                declared.and_then(|p| p.pattern.clone()),
                declared.and_then(|p| p.max_length),
                // D-11: the config is the ONE legitimate source of this permission.
                declared.is_some_and(|p| p.allow_slash),
            ),
        );
    }
    // Remaining declared params → the QUERY STRING, or the JSON BODY when the
    // method carries one (Phase 128 CR-02).
    //
    // `has_request_body` and this location are derived from the SAME
    // `method_carries_request_body` predicate `ToolDecl::param_position` reads, so
    // a `Body`-located parameter on a body-less request is not constructible and
    // the D3 cap's scope cannot drift from the request's actual routing.
    let has_request_body = crate::config::method_carries_request_body(&method_upper);
    let non_path_location = if has_request_body {
        ParameterLocation::Body
    } else {
        ParameterLocation::Query
    };
    for p in &decl.parameters {
        if path_param_names.iter().any(|n| *n == p.name) {
            continue;
        }
        parameters.push(
            Parameter::new(p.name.clone(), non_path_location, p.required).with_rules(
                p.pattern.clone(),
                p.max_length,
                p.allow_slash,
            ),
        );
    }

    Operation {
        method: method_upper,
        path: path.to_string(),
        parameters,
        has_request_body,
        base_url: decl.base_url.clone(),
    }
}

/// Crate-private handler for a single-call HTTP tool (Phase 90 OAPI-02a).
///
/// Holds the synthesized [`ToolInfo`], the built [`Operation`], and the shared
/// [`HttpConnector`]. [`ToolHandler::metadata`] returns `Some(self.info.clone())`
/// (the same RESEARCH §Risks #2 invariant the SQL handler upholds); `handle()`
/// calls [`HttpConnector::execute`] and returns the JSON response.
#[cfg(feature = "http")]
struct HttpToolHandler {
    info: ToolInfo,
    operation: Operation,
    connector: Arc<dyn HttpConnector>,
}

#[cfg(feature = "http")]
#[async_trait]
impl ToolHandler for HttpToolHandler {
    async fn handle(&self, args: Value, _extra: RequestHandlerExtra) -> pmcp::Result<Value> {
        // T-90-03-01: arg injection is bounded by TWO checks, both of which run
        // before this handler's connector call and both of which live in this
        // repository rather than "upstream":
        //
        // (1) the object-envelope schema (additionalProperties:false) is checked by
        //     `pmcp::server::schema_validation::validate_input`, called from the
        //     `ValidatingToolHandler` decorator that `enforce_input_schema` wraps
        //     around this handler at its push site — so a call carrying an
        //     undeclared key never reaches `handle` at all. Conditional on the
        //     `input-validation` feature and on `[server.validation]
        //     enforce_input_schema` not being `false`. Backed by
        //     `tests/input_validation_acceptance.rs::input_validation_refuses_undeclared_argument_without_contacting_upstream`,
        //     which fails if the decorator or its `validate_input` call is removed.
        //
        // (2) path substitution touches only declared `{params}`, and EVERY
        //     substituted value additionally faces
        //     `pmcp::server::schema_validation::validate_path_placeholder` inside
        //     `HttpClient::substitute_path`, with the composed result facing
        //     `validate_resolved_path` through `check_composed_path` before
        //     dispatch. Backed by
        //     `tests/curated_path_injection.rs::curated_path_injection_refuses_traversal_via_a_placeholder_value`
        //     and `..._refuses_a_query_via_a_placeholder_value`, with
        //     `..._accepts_a_compliant_call_with_one_upstream_request` as the
        //     passing control.
        //
        // Until Phase 128 this comment located the envelope check somewhere
        // "upstream" of here (the retired phrasing is quoted in full in
        // `128-10-SUMMARY.md` and deliberately nowhere in `src/`, so the SC-6 grep
        // gate keeps distinguishing a surviving claim from a note about one).
        // Core `pmcp`'s `tools/call` dispatch does not validate request arguments
        // against a declared `inputSchema` — that wiring is deliberately deferred
        // (D-01) — so the claim named a mitigation that did not exist. Restated
        // rather than deleted, because as of this phase it is true and local.
        // The connector's Display is redaction-safe (T-90-01-01); no URL/credential
        // reaches the client error.
        //
        // Phase 128 E1: `execute_for_tool`, not `execute`, so a registered
        // `RequestPolicy` is told WHICH tool the call came from. An `Operation`
        // describes an endpoint and cannot carry a tool name; the default trait
        // body delegates to `execute`, so an out-of-repo connector is unaffected.
        self.connector
            .execute_for_tool(&self.info.name, &self.operation, &args)
            .await
            .map_err(|e| pmcp::Error::Internal(format!("connector error: {e}")))
    }

    fn metadata(&self) -> Option<ToolInfo> {
        Some(self.info.clone())
    }
}

// -----------------------------------------------------------------------------
// Script-tool handler (Phase 90 OAPI-02b / D-01 / D-02) — feature
// `openapi-code-mode`
// -----------------------------------------------------------------------------

/// Crate-private handler for a **script** `[[tools]]` entry (OAPI-02b / D-01).
///
/// A script tool runs admin-authored embedded JS through the EXACT SAME
/// `pmcp_code_mode` engine that Code Mode uses (D-02 — one engine, two
/// surfaces): [`pmcp_code_mode::PlanCompiler`] compiles the `script` to an
/// execution plan, then [`pmcp_code_mode::PlanExecutor`] over the shared
/// [`HttpCodeExecutor`] walks it. The client's validated `args` are bound to the
/// `args` variable BEFORE the script runs — identical to the `JsCodeExecutor`
/// path's `set_variable("args", …)`, which is what makes the engine-parity proof
/// (Plan 05 Task 2) hold byte-for-byte.
///
/// # No token cycle (Pitfall 7 / T-90-05-01)
///
/// A script tool is admin-authored + trusted (like a `sql=` curated query), so it
/// skips the Code Mode validation + HMAC-token gate entirely. It is bounded ONLY by
/// the [`ExecutionConfig`] caps (`max_api_calls`, `max_loop_iterations`,
/// `timeout_seconds`) the [`PlanExecutor`](pmcp_code_mode::PlanExecutor) enforces,
/// and by the `PlanCompiler`-accepted JS subset (no `eval` / FFI).
///
/// # Feature gate (RESEARCH Pitfall 4)
///
/// Gated `openapi-code-mode` (the umbrella that forwards
/// `pmcp-code-mode/js-runtime`) — `PlanCompiler` / `PlanExecutor` are NOT in
/// scope under bare `code-mode`, so the light / curated-only build (`http
/// code-mode`) compiles without this type (single-call only).
#[cfg(feature = "openapi-code-mode")]
struct ScriptToolHandler {
    /// The admin-authored script, compiled ONCE at synthesis (the body is fixed
    /// content). Executed per `handle` over a fresh
    /// [`PlanExecutor`](pmcp_code_mode::PlanExecutor).
    plan: pmcp_code_mode::ExecutionPlan,
    /// The SAME executor instance that feeds Code Mode (D-02). Cloned per request
    /// to construct a fresh [`PlanExecutor`](pmcp_code_mode::PlanExecutor).
    http_exec: HttpCodeExecutor,
    /// The execution bounds (Pitfall 7 — the only limit on an admin script).
    exec_config: ExecutionConfig,
    /// The synthesized `ToolInfo` (object-envelope schema from
    /// `[[tools.parameters]]`, `additionalProperties:false`).
    ///
    /// `args` are checked against THIS schema before the script runs (T-90-05-03),
    /// by `pmcp::server::schema_validation::validate_input` inside the
    /// `ValidatingToolHandler` decorator that [`enforce_input_schema`] wraps
    /// around this handler at its push site — so a refusal happens before
    /// [`ToolHandler::handle`] is entered and therefore before
    /// `PlanExecutor::execute` makes any backend call.
    ///
    /// Conditional on the `input-validation` feature and on `[server.validation]
    /// enforce_input_schema` not being `false`; with either off,
    /// [`enforce_input_schema`] logs the opt-out and returns the handler undecorated.
    ///
    /// Backed by
    /// `tests/script_tool.rs::script_tool_refuses_a_schema_violating_arg_before_the_script_runs`,
    /// which asserts the refusal AND that the mock backend observed zero requests —
    /// it fails if the decorator or its `validate_input` call is removed. The
    /// sibling `script_tool_args_max_lines_binding_is_honored` asserts argument
    /// BINDING, which is a different (also true) property; before this phase it was
    /// the only row carrying this threat ID, and a binding assertion is not a
    /// validation assertion.
    tool_info: ToolInfo,
}

#[cfg(feature = "openapi-code-mode")]
impl ScriptToolHandler {
    /// Build a [`ScriptToolHandler`] from a script `[[tools]]` declaration,
    /// the shared [`HttpCodeExecutor`], and the [`ExecutionConfig`] bounds.
    ///
    /// The `tool_info` is built from `[[tools.parameters]]` via the SAME
    /// [`build_input_schema`] / [`build_annotations`] / [`apply_widget_meta`]
    /// helpers the single-call path uses, so a script tool's `args` are checked
    /// against an identically-shaped schema (object envelope,
    /// `additionalProperties:false` unless `[server.validation]
    /// additional_properties` opts out) by the same enforcer — one
    /// [`enforce_input_schema`] call per push site, one
    /// `pmcp::server::schema_validation::validate_input` inside
    /// `ValidatingToolHandler`. There is no second validator for script tools;
    /// see [`ScriptToolHandler::tool_info`] for the condition and the acceptance
    /// row.
    ///
    /// `validation` is threaded in so a script tool's schema is built under the SAME
    /// `[server.validation]` policy as every other tool kind — a script tool whose
    /// parameters escaped the D3 cap would be a hole in exactly the surface this
    /// phase closes. A script tool's parameters are BODY position (it declares no
    /// `path`/`method`), so in practice the cap does not apply to them; the point is
    /// that the decision is made by one rule rather than by which synthesizer ran.
    ///
    /// # Errors
    ///
    /// Returns [`ToolkitError::Synth`] if the declaration carries no `script`
    /// (a defensive guard — callers route only `is_script_tool()` entries here),
    /// or if the script fails to compile (surfaced here at server build time,
    /// failing fast rather than on the first tool call).
    fn new(
        decl: &ToolDecl,
        http_exec: HttpCodeExecutor,
        exec_config: ExecutionConfig,
        validation: &ValidationSection,
    ) -> Result<Self> {
        let script = decl.script.clone().ok_or_else(|| {
            ToolkitError::Synth(format!(
                "tool '{}' has no `script` body — not a script tool",
                decl.name
            ))
        })?;
        // Compile the admin-authored JS ONCE at synthesis time — the script is
        // fixed content, so compiling per request would re-run a full SWC parse
        // on the hot path (the PlanCompiler-accepted subset, no eval / FFI, is
        // the static bound). A compile error surfaces here at server build.
        let plan = pmcp_code_mode::PlanCompiler::with_config(&exec_config)
            .compile_code(&script)
            .map_err(|e| {
                ToolkitError::Synth(format!(
                    "tool '{}' script failed to compile: {e}",
                    decl.name
                ))
            })?;
        let tool_info = build_tool_info(decl, validation);
        Ok(Self {
            plan,
            // Phase 128 E1: this handler owns a per-tool CLONE of the shared
            // executor, which is the one place a script tool's name can be
            // attached — `HttpExecutor::execute_request` is a `pmcp-code-mode`
            // trait method and carries no tool name.
            http_exec: http_exec.with_tool_label(&decl.name),
            exec_config,
            tool_info,
        })
    }
}

#[cfg(feature = "openapi-code-mode")]
#[pmcp_code_mode::async_trait]
impl ToolHandler for ScriptToolHandler {
    /// Run the pre-compiled admin-authored script over the shared engine, binding
    /// the validated `args` to the `args` variable (D-02 — identical to the
    /// `JsCodeExecutor` path's `set_variable("args", …)`).
    async fn handle(&self, args: Value, extra: RequestHandlerExtra) -> pmcp::Result<Value> {
        // (1) Execute the plan (compiled once in `new`) over a PER-REQUEST clone
        //     of the shared HttpCodeExecutor (D-02), threading the captured
        //     inbound MCP token (Plan 90-10 / OAPI-03 / OAPI-05) so an
        //     `oauth_passthrough` backend forwards it. Bounded by ExecutionConfig
        //     (Pitfall 7 — no token cycle, only these caps).
        let mut executor = pmcp_code_mode::PlanExecutor::new(
            crate::code_mode::request_executor_from_extra(&self.http_exec, &extra),
            self.exec_config.clone(),
        );
        // (2) Bind the client args to `args` (T-90-05-03) — byte-identical to
        //     compile_and_execute's set_variable("args", …). These args have
        //     ALREADY been checked against `self.tool_info.input_schema` by
        //     `pmcp::server::schema_validation::validate_input`, in the
        //     `ValidatingToolHandler` decorator wrapped around this handler — a
        //     violating call never reaches this line. That is a fact about the
        //     DECORATOR, not about this function: `handle` performs no validation
        //     of its own and must not be read as if it did. Condition and
        //     acceptance row: see `ScriptToolHandler::tool_info`.
        executor.set_variable("args", args);

        let result = executor
            .execute(&self.plan)
            .await
            .map_err(|e| pmcp::Error::Internal(format!("script execution failed: {e}")))?;
        Ok(result.value)
    }

    fn metadata(&self) -> Option<ToolInfo> {
        Some(self.tool_info.clone())
    }
}

// -----------------------------------------------------------------------------
// Tests — Plan 05 Task 1 (RED) → GREEN in Task 2
// -----------------------------------------------------------------------------

#[cfg(test)]
mod tests {
    use super::*;
    use crate::config::{
        AnnotationsDecl, ItemsDecl, ParamDecl, ServerConfig, ServerSection, ToolDecl,
        ValidationSection,
    };
    use serde_json::Value;

    /// Construct a minimal `ServerConfig` that satisfies `validate()` (non-empty
    /// `name` + `version`) so the synthesizer path is the system under test —
    /// not the parser/validator from Plan 04.
    fn cfg_with_tools(tools: Vec<ToolDecl>) -> ServerConfig {
        ServerConfig {
            server: ServerSection {
                name: "demo".to_string(),
                version: "0.1.0".to_string(),
                ..Default::default()
            },
            tools,
            ..Default::default()
        }
    }

    #[test]
    fn empty_tools_returns_empty_vec() {
        let cfg = cfg_with_tools(vec![]);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        assert_eq!(out.len(), 0);
    }

    #[test]
    fn one_tool_no_params_yields_object_schema() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "ping".to_string(),
            description: Some("Ping the server".to_string()),
            parameters: vec![],
            annotations: None,
            ..Default::default()
        }]);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        assert_eq!(out.len(), 1);
        let (name, info, _handler) = &out[0];
        assert_eq!(name, "ping");
        assert_eq!(info.name, "ping");
        assert_eq!(info.description.as_deref(), Some("Ping the server"));
        let schema = &info.input_schema;
        assert_eq!(schema["type"], Value::String("object".to_string()));
        assert_eq!(schema["properties"], serde_json::json!({}));
        assert_eq!(schema["required"], serde_json::json!([]));
        assert_eq!(schema["additionalProperties"], Value::Bool(false));
    }

    #[test]
    fn required_and_optional_params_partitioned() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "search".to_string(),
            description: Some("Search".to_string()),
            parameters: vec![
                ParamDecl {
                    name: "query".to_string(),
                    param_type: Some("string".to_string()),
                    description: Some("the search query".to_string()),
                    required: true,
                    ..Default::default()
                },
                ParamDecl {
                    name: "max_results".to_string(),
                    param_type: Some("integer".to_string()),
                    description: Some("maximum result count".to_string()),
                    required: false,
                    default: Some(toml::Value::Integer(100)),
                    minimum: Some(1.0),
                    maximum: Some(1000.0),
                    ..Default::default()
                },
            ],
            ..Default::default()
        }]);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        let (_, info, _) = &out[0];
        let schema = &info.input_schema;
        assert_eq!(schema["required"], serde_json::json!(["query"]));
        let props = schema["properties"].as_object().expect("object");
        assert_eq!(props["query"]["type"], "string");
        assert_eq!(props["max_results"]["type"], "integer");
        assert_eq!(props["max_results"]["minimum"], serde_json::json!(1.0));
        assert_eq!(props["max_results"]["maximum"], serde_json::json!(1000.0));
        assert_eq!(props["max_results"]["default"], serde_json::json!(100));
    }

    #[test]
    fn param_max_length_propagates() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "echo".to_string(),
            description: Some("Echo".to_string()),
            parameters: vec![ParamDecl {
                name: "text".to_string(),
                param_type: Some("string".to_string()),
                description: Some("input text".to_string()),
                required: true,
                max_length: Some(256),
                ..Default::default()
            }],
            ..Default::default()
        }]);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        let (_, info, _) = &out[0];
        assert_eq!(
            info.input_schema["properties"]["text"]["maxLength"],
            serde_json::json!(256)
        );
    }

    #[test]
    fn annotations_round_trip_via_fluent_builder() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "destroy_all".to_string(),
            description: Some("Destroy all data (test)".to_string()),
            parameters: vec![],
            annotations: Some(AnnotationsDecl {
                read_only_hint: false,
                destructive_hint: true,
                idempotent_hint: false,
                open_world_hint: false,
                cost_hint: Some("high".to_string()),
            }),
            ..Default::default()
        }]);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        let (_, info, _) = &out[0];
        let ann = info.annotations.as_ref().expect("annotations");
        assert_eq!(ann.read_only_hint, Some(false));
        assert_eq!(ann.destructive_hint, Some(true));
        assert_eq!(ann.idempotent_hint, Some(false));
        assert_eq!(ann.open_world_hint, Some(false));
    }

    /// REVIEWS H1 (in-plan widget_meta flip — no SqliteConnector dependency).
    ///
    /// When a `[[tools]]` entry declares `ui_resource_uri`, the synthesized
    /// `ToolInfo` must carry widget metadata so pmcp core's
    /// `with_widget_enrichment` (gated on `info.widget_meta().is_some()`)
    /// populates `structuredContent` (D-06). The flip lives in the shared
    /// `synthesize_inner` helper, so it fires for BOTH entry points; this test
    /// exercises it via the no-connector `synthesize_from_config` path.
    #[test]
    fn widget_meta_flips_when_ui_resource_uri_present() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "widget_tool".to_string(),
            description: Some("renders a widget".to_string()),
            ui_resource_uri: Some("ui://test".to_string()),
            ..Default::default()
        }]);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        let (_, info, _) = &out[0];
        assert!(
            info.widget_meta().is_some(),
            "ui_resource_uri set ⇒ widget_meta() must be Some so D-06 structuredContent fires"
        );
    }

    /// REVIEWS H1 negative case — a tool WITHOUT `ui_resource_uri` must NOT
    /// carry widget metadata (T-84-03-03: no accidental flip on non-widget
    /// tools).
    #[test]
    fn widget_meta_absent_when_ui_resource_uri_none() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "plain_tool".to_string(),
            description: Some("no widget".to_string()),
            ui_resource_uri: None,
            ..Default::default()
        }]);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        let (_, info, _) = &out[0];
        assert!(
            info.widget_meta().is_none(),
            "ui_resource_uri absent ⇒ widget_meta() must be None (no accidental flip)"
        );
    }

    #[tokio::test]
    async fn synthesized_handler_metadata_returns_some() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "ping".to_string(),
            description: Some("ping".to_string()),
            parameters: vec![],
            annotations: None,
            ..Default::default()
        }]);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        let (_, expected_info, handler) = &out[0];
        let actual = handler.metadata();
        assert!(
            actual.is_some(),
            "RESEARCH §Risks #2 invariant: SynthesizedToolHandler::metadata() MUST return Some(ToolInfo)"
        );
        assert_eq!(actual.unwrap().name, expected_info.name);
    }

    /// A `[[tools]]` declaration with one defaulted `limit` param (default=20),
    /// used to exercise [`extract_named_params`]'s default / explicit-null logic.
    fn decl_with_limit_default() -> ToolDecl {
        ToolDecl {
            name: "search".to_string(),
            description: Some("Search".to_string()),
            sql: Some("SELECT * FROM t LIMIT :limit".to_string()),
            parameters: vec![ParamDecl {
                name: "limit".to_string(),
                param_type: Some("integer".to_string()),
                description: Some("row limit".to_string()),
                required: false,
                default: Some(toml::Value::Integer(20)),
                ..Default::default()
            }],
            ..Default::default()
        }
    }

    #[test]
    fn extract_named_params_applies_default_when_absent() {
        // `{}` → declared default (20) is bound (the reference search/list calls
        // rely on this so an omitted :limit never binds NULL).
        let decl = decl_with_limit_default();
        let params = extract_named_params(&decl, &serde_json::json!({}));
        assert_eq!(params, vec![("limit".to_string(), serde_json::json!(20))]);
    }

    #[test]
    fn extract_named_params_explicit_null_applies_default() {
        // 85-10 WR-02 secondary fix: an EXPLICIT JSON null must NOT bind
        // `LIMIT NULL` — it falls through to the declared default exactly like
        // an omitted key.
        let decl = decl_with_limit_default();
        let params = extract_named_params(&decl, &serde_json::json!({ "limit": null }));
        assert_eq!(
            params,
            vec![("limit".to_string(), serde_json::json!(20))],
            "explicit null must apply the declared default, not bind LIMIT NULL"
        );
    }

    #[test]
    fn extract_named_params_explicit_value_overrides_default() {
        // A concrete value wins over the default.
        let decl = decl_with_limit_default();
        let params = extract_named_params(&decl, &serde_json::json!({ "limit": 5 }));
        assert_eq!(params, vec![("limit".to_string(), serde_json::json!(5))]);
    }

    // -------------------------------------------------------------------------
    // Phase 128 D2 — the six new `ParamDecl` keywords reach `inputSchema`
    // -------------------------------------------------------------------------

    /// Fetch the synthesized property object for `param` of the single tool in
    /// `tools`.
    fn prop_of(tools: Vec<ToolDecl>, param: &str) -> Value {
        let cfg = cfg_with_tools(tools);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        let (_name, info, _handler) = &out[0];
        info.input_schema["properties"][param].clone()
    }

    /// D2 / SC-2: `pattern`, `minLength`, `format` and `maxItems` all reach the
    /// emitted `inputSchema`.
    #[test]
    fn input_schema_emits_d2_scalar_keywords() {
        let prop = prop_of(
            vec![ToolDecl {
                name: "lookup".to_string(),
                parameters: vec![ParamDecl {
                    name: "region".to_string(),
                    param_type: Some("string".to_string()),
                    required: true,
                    pattern: Some("^[A-Z]{3}$".to_string()),
                    min_length: Some(3),
                    format: Some("uuid".to_string()),
                    max_items: Some(5),
                    ..Default::default()
                }],
                ..Default::default()
            }],
            "region",
        );
        assert_eq!(prop["pattern"], serde_json::json!("^[A-Z]{3}$"));
        assert_eq!(prop["minLength"], serde_json::json!(3));
        assert_eq!(prop["format"], serde_json::json!("uuid"));
        assert_eq!(prop["maxItems"], serde_json::json!(5));
    }

    /// D2 / RESEARCH Finding 1h: `items` is emitted in OBJECT form. The array
    /// (draft-07 tuple) form does not compile under the Draft 2020-12 pin and
    /// would take the whole tool's validator down.
    #[test]
    fn input_schema_emits_items_as_object_never_array() {
        let prop = prop_of(
            vec![ToolDecl {
                name: "batch".to_string(),
                parameters: vec![ParamDecl {
                    name: "codes".to_string(),
                    param_type: Some("array".to_string()),
                    required: true,
                    items: Some(ItemsDecl {
                        item_type: Some("string".to_string()),
                        max_length: Some(8),
                        pattern: Some("^[a-z]+$".to_string()),
                    }),
                    max_items: Some(10),
                    ..Default::default()
                }],
                ..Default::default()
            }],
            "codes",
        );
        assert!(
            prop["items"].is_object(),
            "items must be an OBJECT, got: {}",
            prop["items"]
        );
        assert!(
            !prop["items"].is_array(),
            "array-form items does not compile under the 2020-12 pin"
        );
        assert_eq!(
            prop["items"],
            serde_json::json!({
                "type": "string",
                "maxLength": 8,
                "pattern": "^[a-z]+$",
            })
        );
    }

    /// D2 edge (empty): a `ParamDecl` carrying NONE of the six new keywords emits
    /// exactly the five it emits today — no empty `pattern` string and no empty
    /// `items` object appears.
    #[test]
    fn input_schema_omits_d2_keywords_when_undeclared() {
        let prop = prop_of(
            vec![ToolDecl {
                name: "legacy".to_string(),
                parameters: vec![ParamDecl {
                    name: "count".to_string(),
                    param_type: Some("integer".to_string()),
                    description: Some("how many".to_string()),
                    required: false,
                    minimum: Some(1.0),
                    maximum: Some(10.0),
                    max_length: Some(4),
                    ..Default::default()
                }],
                ..Default::default()
            }],
            "count",
        );
        let obj = prop.as_object().expect("property object");
        let mut keys: Vec<&str> = obj.keys().map(String::as_str).collect();
        keys.sort_unstable();
        assert_eq!(
            keys,
            vec!["description", "maxLength", "maximum", "minimum", "type"],
            "exactly the five pre-D2 keywords, and no more"
        );
        for absent in ["pattern", "minLength", "format", "items", "maxItems"] {
            assert!(
                obj.get(absent).is_none(),
                "undeclared keyword {absent} must not be emitted"
            );
        }
    }

    /// D2 edge (ordering): two synthesis runs over ONE config produce
    /// byte-identical `input_schema` values, so `properties` and `required` follow
    /// `[[tools.parameters]]` declaration order deterministically.
    #[test]
    fn input_schema_is_byte_identical_across_two_synthesis_runs() {
        let tools = vec![ToolDecl {
            name: "search".to_string(),
            parameters: vec![
                ParamDecl {
                    name: "zebra".to_string(),
                    param_type: Some("string".to_string()),
                    required: true,
                    pattern: Some("^z".to_string()),
                    ..Default::default()
                },
                ParamDecl {
                    name: "alpha".to_string(),
                    param_type: Some("string".to_string()),
                    required: true,
                    min_length: Some(1),
                    ..Default::default()
                },
                ParamDecl {
                    name: "middle".to_string(),
                    param_type: Some("integer".to_string()),
                    required: false,
                    ..Default::default()
                },
            ],
            ..Default::default()
        }];
        let cfg = cfg_with_tools(tools);
        let first = synthesize_from_config(&cfg).expect("synthesize")[0]
            .1
            .input_schema
            .to_string();
        let second = synthesize_from_config(&cfg).expect("synthesize")[0]
            .1
            .input_schema
            .to_string();
        assert_eq!(first, second, "synthesis must be byte-deterministic");
        // Declaration order, not sorted order.
        assert!(
            first.find("\"zebra\"").unwrap() < first.find("\"alpha\"").unwrap(),
            "properties must follow declaration order: {first}"
        );
    }

    /// D2 edge (adjacency): `min_length == max_length` accepts exactly that length
    /// and refuses one code point either side, through the SAME core validator the
    /// D1 decorator uses.
    #[cfg(feature = "input-validation")]
    #[test]
    fn min_length_equal_to_max_length_accepts_exactly_that_length() {
        use pmcp::server::schema_validation::validate_input;

        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "exact".to_string(),
            parameters: vec![ParamDecl {
                name: "code".to_string(),
                param_type: Some("string".to_string()),
                required: true,
                min_length: Some(3),
                max_length: Some(3),
                ..Default::default()
            }],
            ..Default::default()
        }]);
        let out = synthesize_from_config(&cfg).expect("synthesize");
        let schema = &out[0].1.input_schema;

        validate_input(schema, Some(&serde_json::json!({ "code": "abc" })), None)
            .expect("exactly three code points must be accepted");
        validate_input(schema, Some(&serde_json::json!({ "code": "ab" })), None)
            .expect_err("two code points must be refused");
        validate_input(schema, Some(&serde_json::json!({ "code": "abcd" })), None)
            .expect_err("four code points must be refused");
    }

    // -------------------------------------------------------------------------
    // Phase 128 D3 — position-scoped default cap and `[server.validation]`
    // -------------------------------------------------------------------------

    /// Synthesize `tools` under `validation` and return the property object for
    /// `param` of the FIRST tool.
    fn prop_under(tools: Vec<ToolDecl>, validation: ValidationSection, param: &str) -> Value {
        let mut cfg = cfg_with_tools(tools);
        cfg.server.validation = validation;
        let out = synthesize_from_config(&cfg).expect("synthesize");
        out[0].1.input_schema["properties"][param].clone()
    }

    /// A single-call `GET` tool declaring one path parameter and one query
    /// parameter, both uncapped strings.
    fn get_tool_with_path_and_query() -> Vec<ToolDecl> {
        vec![ToolDecl {
            name: "line_status".to_string(),
            path: Some("/lines/{line_id}/status".to_string()),
            method: Some("GET".to_string()),
            parameters: vec![
                ParamDecl {
                    name: "line_id".to_string(),
                    param_type: Some("string".to_string()),
                    required: true,
                    ..Default::default()
                },
                ParamDecl {
                    name: "detail".to_string(),
                    param_type: Some("string".to_string()),
                    required: false,
                    ..Default::default()
                },
            ],
            ..Default::default()
        }]
    }

    /// D3: a PATH-position string with no `max_length` is capped at the configured
    /// default.
    #[test]
    fn default_cap_applies_to_path_position_string() {
        let prop = prop_under(
            get_tool_with_path_and_query(),
            ValidationSection::default(),
            "line_id",
        );
        assert_eq!(prop["maxLength"], serde_json::json!(256));
    }

    /// D3: a QUERY-position string (non-path parameter of a `GET` tool) with no
    /// `max_length` is capped at the configured default.
    #[test]
    fn default_cap_applies_to_query_position_string() {
        let prop = prop_under(
            get_tool_with_path_and_query(),
            ValidationSection::default(),
            "detail",
        );
        assert_eq!(prop["maxLength"], serde_json::json!(256));
    }

    /// D3 / D-05 / T-128-13d: a BODY-position string on a mutating single-call tool
    /// receives NO cap. Capping a `POST` payload's free-text field is exactly the
    /// breakage review note C objected to.
    #[test]
    fn default_cap_never_applies_to_body_position_string() {
        let tools = vec![ToolDecl {
            name: "add_comment".to_string(),
            path: Some("/issues/{id}/comments".to_string()),
            method: Some("POST".to_string()),
            parameters: vec![
                ParamDecl {
                    name: "id".to_string(),
                    param_type: Some("string".to_string()),
                    required: true,
                    ..Default::default()
                },
                ParamDecl {
                    name: "body_text".to_string(),
                    param_type: Some("string".to_string()),
                    required: true,
                    ..Default::default()
                },
            ],
            ..Default::default()
        }];
        let path_prop = prop_under(tools.clone(), ValidationSection::default(), "id");
        assert_eq!(
            path_prop["maxLength"],
            serde_json::json!(256),
            "the path parameter of a POST tool IS still capped"
        );
        let body_prop = prop_under(tools, ValidationSection::default(), "body_text");
        assert!(
            body_prop.get("maxLength").is_none(),
            "a POST payload field must receive no default cap, got: {body_prop}"
        );
    }

    /// D3 edge (adjacency): a DECLARED `max_length` is never overridden and never
    /// merged with the default, in either direction.
    #[test]
    fn declared_max_length_wins_over_the_default_cap() {
        let mut tools = get_tool_with_path_and_query();
        tools[0].parameters[0].max_length = Some(12);
        let prop = prop_under(tools, ValidationSection::default(), "line_id");
        assert_eq!(prop["maxLength"], serde_json::json!(12));
    }

    /// D3 edge (empty): `default_max_length = 0` emits no `maxLength` in ANY
    /// position and disables the cap entirely.
    #[test]
    fn default_max_length_zero_disables_the_cap_in_every_position() {
        let validation = ValidationSection {
            default_max_length: 0,
            ..Default::default()
        };
        for param in ["line_id", "detail"] {
            let prop = prop_under(get_tool_with_path_and_query(), validation.clone(), param);
            assert!(
                prop.get("maxLength").is_none(),
                "{param} must carry no maxLength when the default is 0, got: {prop}"
            );
        }
    }

    /// D3 edge (boundary + encoding): the cap is counted in Unicode CODE POINTS,
    /// not bytes — so exactly `default_max_length` multi-byte characters are
    /// accepted and one more is refused, through the SAME core validator.
    #[cfg(feature = "input-validation")]
    #[test]
    fn default_cap_boundary_is_counted_in_code_points_not_bytes() {
        use pmcp::server::schema_validation::validate_input;

        let validation = ValidationSection {
            default_max_length: 8,
            ..Default::default()
        };
        let mut cfg = cfg_with_tools(get_tool_with_path_and_query());
        cfg.server.validation = validation;
        let out = synthesize_from_config(&cfg).expect("synthesize");
        let schema = &out[0].1.input_schema;

        // Each `é` is TWO bytes and ONE code point. Eight of them are 16 bytes.
        let at_limit: String = "é".repeat(8);
        let over_limit: String = "é".repeat(9);
        assert_eq!(
            at_limit.len(),
            16,
            "the fixture must actually be multi-byte"
        );
        validate_input(
            schema,
            Some(&serde_json::json!({ "line_id": at_limit })),
            None,
        )
        .expect("exactly 8 code points must be accepted even though they are 16 bytes");
        validate_input(
            schema,
            Some(&serde_json::json!({ "line_id": over_limit })),
            None,
        )
        .expect_err("9 code points must be refused");
    }

    /// `additional_properties = true` flips the emitted envelope, re-opening the
    /// unknown-argument class for this server.
    #[test]
    fn additional_properties_opt_out_flips_the_envelope() {
        let mut cfg = cfg_with_tools(get_tool_with_path_and_query());
        cfg.server.validation = ValidationSection {
            additional_properties: true,
            ..Default::default()
        };
        let out = synthesize_from_config(&cfg).expect("synthesize");
        assert_eq!(
            out[0].1.input_schema["additionalProperties"],
            Value::Bool(true)
        );

        let mut cfg = cfg_with_tools(get_tool_with_path_and_query());
        cfg.server.validation = ValidationSection::default();
        let out = synthesize_from_config(&cfg).expect("synthesize");
        assert_eq!(
            out[0].1.input_schema["additionalProperties"],
            Value::Bool(false),
            "the default must remain a closed envelope"
        );
    }

    /// T-128-13c: `enforce_input_schema = false` skips the SCHEMA CHECK, not the
    /// decorator. This test pins the separation the E2 argument-validator registry
    /// will depend on: with the flag off an undeclared argument is ACCEPTED (schema
    /// off), while the inner handler's own rule still REFUSES (the non-schema
    /// enforcement is untouched). Without this test the two collapse back together
    /// on the next refactor.
    #[cfg(feature = "input-validation")]
    #[tokio::test]
    async fn enforce_input_schema_false_skips_the_check_not_the_decorator() {
        /// Stands in for an explicitly-registered argument validator: a rule that
        /// lives INSIDE the decorated stack and is not the JSON Schema check.
        struct RefusingInner;

        #[async_trait]
        impl ToolHandler for RefusingInner {
            async fn handle(
                &self,
                _args: Value,
                _extra: RequestHandlerExtra,
            ) -> pmcp::Result<Value> {
                Err(pmcp::Error::Validation(
                    "the registered validator refused".to_string(),
                ))
            }
        }

        let decl = ToolDecl {
            name: "guarded".to_string(),
            parameters: vec![ParamDecl {
                name: "declared".to_string(),
                param_type: Some("string".to_string()),
                required: false,
                ..Default::default()
            }],
            ..Default::default()
        };
        let info = build_tool_info(&decl, &ValidationSection::default());
        let undeclared = serde_json::json!({ "not_declared_at_all": "x" });

        // (a) enforcement ON: the SCHEMA refuses before the inner handler runs, so
        //     the message is the schema refusal, not the inner one.
        let on = ValidatingToolHandler::wrap(Arc::new(RefusingInner), &info, &decl, true, None);
        let err = on
            .handle(undeclared.clone(), RequestHandlerExtra::default())
            .await
            .expect_err("an undeclared argument must be refused when enforcement is on");
        assert!(
            !err.to_string().contains("registered validator"),
            "the schema check must run FIRST when enforcement is on: {err}"
        );

        // (b) enforcement OFF: the schema check is skipped (the undeclared argument
        //     is accepted by it), and the inner rule STILL refuses. That is the
        //     separation — A off must not silently turn B off.
        let off = ValidatingToolHandler::wrap(Arc::new(RefusingInner), &info, &decl, false, None);
        let err = off
            .handle(undeclared, RequestHandlerExtra::default())
            .await
            .expect_err("the inner (non-schema) rule must still refuse");
        assert!(
            err.to_string().contains("registered validator"),
            "with the schema check off, the refusal must come from the inner rule: {err}"
        );
    }
}

// -----------------------------------------------------------------------------
// Tests — Phase 128 D4(b): `build_operation` carries the declared placeholder
// rules onto every `Parameter`.
// -----------------------------------------------------------------------------

/// Named `build_operation` deliberately: libtest matches the FULL test path, so a
/// test placed in the crate's existing `mod tests` would be
/// `tools::tests::build_operation_…` and this plan's `--lib tools::build_operation`
/// verify filter would select ZERO tests while exiting 0. As a SIBLING of `tests`
/// at the `tools` module level the filter resolves as written. Do not fold these
/// into `mod tests`.
#[cfg(all(test, feature = "http"))]
mod build_operation {
    use super::*;

    /// A single-call `GET` tool on `path` carrying `parameters`.
    fn decl(path: &str, parameters: Vec<ParamDecl>) -> ToolDecl {
        ToolDecl {
            name: "t".to_string(),
            description: Some("t".to_string()),
            path: Some(path.to_string()),
            method: Some("GET".to_string()),
            parameters,
            ..Default::default()
        }
    }

    fn param_named<'a>(op: &'a Operation, name: &str) -> &'a Parameter {
        op.parameters
            .iter()
            .find(|p| p.name == name)
            .unwrap_or_else(|| panic!("parameter {name} present"))
    }

    /// A path parameter declaring a `pattern` produces a `Parameter` carrying it.
    #[test]
    fn build_operation_carries_a_declared_path_pattern() {
        let d = decl(
            "/content/{version}",
            vec![ParamDecl {
                name: "version".to_string(),
                param_type: Some("string".to_string()),
                required: true,
                pattern: Some("^C[0-9]+$".to_string()),
                ..Default::default()
            }],
        );
        let op = super::build_operation("/content/{version}", "GET", &d);
        let p = param_named(&op, "version");
        assert_eq!(p.location, ParameterLocation::Path);
        assert_eq!(p.pattern.as_deref(), Some("^C[0-9]+$"));
    }

    /// A query parameter declaring `max_length = 64` produces a `Parameter`
    /// carrying `max_length: Some(64)`.
    #[test]
    fn build_operation_carries_a_declared_query_max_length() {
        let d = decl(
            "/search",
            vec![ParamDecl {
                name: "q".to_string(),
                param_type: Some("string".to_string()),
                max_length: Some(64),
                ..Default::default()
            }],
        );
        let op = super::build_operation("/search", "GET", &d);
        let p = param_named(&op, "q");
        assert_eq!(p.location, ParameterLocation::Query);
        assert_eq!(p.max_length, Some(64));
    }

    /// A template segment the config never declared carries NO declared rules and
    /// `allow_slash: false` — the path loop iterates the TEMPLATE, so it must look
    /// the `ParamDecl` up by name and fall back cleanly when there is none.
    #[test]
    fn build_operation_leaves_rules_absent_for_an_undeclared_template_segment() {
        let d = decl("/content/{version}", vec![]);
        let op = super::build_operation("/content/{version}", "GET", &d);
        let p = param_named(&op, "version");
        assert_eq!(p.pattern, None);
        assert_eq!(p.max_length, None);
        assert!(!p.allow_slash);
        assert!(p.required, "a template path parameter stays required");
    }

    /// D-11: `allow_slash` reaches the `Parameter` from the server's OWN config —
    /// the one legitimate source.
    #[test]
    fn build_operation_carries_allow_slash_from_the_config() {
        let d = decl(
            "/files/{subpath}",
            vec![ParamDecl {
                name: "subpath".to_string(),
                param_type: Some("string".to_string()),
                required: true,
                allow_slash: true,
                ..Default::default()
            }],
        );
        let op = super::build_operation("/files/{subpath}", "GET", &d);
        assert!(param_named(&op, "subpath").allow_slash);
    }

    /// The three rule fields round-trip into the core rules type the substitution
    /// point reads.
    #[cfg(feature = "input-validation")]
    #[test]
    fn build_operation_rules_reach_placeholder_rules() {
        let d = decl(
            "/files/{subpath}",
            vec![ParamDecl {
                name: "subpath".to_string(),
                param_type: Some("string".to_string()),
                required: true,
                pattern: Some("^[a-z/]+$".to_string()),
                max_length: Some(128),
                allow_slash: true,
                ..Default::default()
            }],
        );
        let op = super::build_operation("/files/{subpath}", "GET", &d);
        let rules = param_named(&op, "subpath").placeholder_rules();
        assert_eq!(rules.declared_pattern, Some("^[a-z/]+$"));
        assert_eq!(rules.declared_max_length, Some(128));
        assert!(rules.allow_slash);
    }

    /// Phase 128 CR-02 — a body-bearing method routes its non-path declared
    /// parameters to [`ParameterLocation::Body`], not to the query string.
    ///
    /// Before the fix every one of these was `Query`, so the payload
    /// `build_body` assembles was empty and the values travelled in the URL.
    #[test]
    fn build_operation_routes_a_post_non_path_param_to_the_body() {
        for method in ["POST", "PUT", "PATCH", "post"] {
            let mut d = decl(
                "/issues/{id}/comments",
                vec![
                    ParamDecl {
                        name: "id".to_string(),
                        param_type: Some("string".to_string()),
                        required: true,
                        ..Default::default()
                    },
                    ParamDecl {
                        name: "body_text".to_string(),
                        param_type: Some("string".to_string()),
                        required: true,
                        ..Default::default()
                    },
                ],
            );
            d.method = Some(method.to_string());
            let op = super::build_operation("/issues/{id}/comments", method, &d);
            assert!(
                op.has_request_body,
                "{method} carries a request body by definition"
            );
            assert_eq!(param_named(&op, "id").location, ParameterLocation::Path);
            assert_eq!(
                param_named(&op, "body_text").location,
                ParameterLocation::Body,
                "{method}: a non-path declared parameter must be routed to the payload"
            );
            assert_eq!(
                op.body_parameters().len(),
                1,
                "{method}: exactly the one non-path parameter is a body parameter"
            );
            assert!(
                op.query_parameters().is_empty(),
                "{method}: nothing may travel in the query string as well"
            );
        }
    }

    /// The other half of the same rule: a method that carries NO request body has
    /// nowhere but the URL to put a non-path parameter, so it stays `Query`. Without
    /// this row the fix above is satisfiable by routing everything to the body,
    /// which would drop a `GET` tool's parameters entirely.
    #[test]
    fn build_operation_keeps_a_body_less_method_non_path_param_in_the_query() {
        for method in ["GET", "HEAD", "DELETE", "OPTIONS"] {
            let mut d = decl(
                "/search",
                vec![ParamDecl {
                    name: "q".to_string(),
                    param_type: Some("string".to_string()),
                    required: true,
                    ..Default::default()
                }],
            );
            d.method = Some(method.to_string());
            let op = super::build_operation("/search", method, &d);
            assert!(!op.has_request_body, "{method} carries no request body");
            assert_eq!(
                param_named(&op, "q").location,
                ParameterLocation::Query,
                "{method}: a non-path parameter must stay in the query string"
            );
            assert!(
                op.body_parameters().is_empty(),
                "{method}: a body-less request can have no body parameter"
            );
        }
    }

    /// Phase 128 CR-03 — the D3 cap's POSITION and the request's ROUTING agree for
    /// EVERY method this connector can send, so the cap can no longer be withheld
    /// from a value that travels in the query string.
    ///
    /// This is the row that would have caught the original defect: it does not
    /// assert a particular location, it asserts the PAIRING. Before the fix
    /// `param_position` said `Body` for `POST` while `build_operation` said `Query`,
    /// and `apply_position_cap` emitted nothing for `Body` — so a `POST` parameter
    /// reached the request line with no `maxLength`.
    ///
    /// Fails on removal: point either side at its own method list again and the
    /// `POST`/`PUT`/`PATCH` rows (or the `OPTIONS` row) go red.
    #[test]
    fn param_position_agrees_with_the_built_parameter_location_for_every_method() {
        use crate::config::ParamPosition;

        // Every method `HttpClient::convert_method` accepts.
        for method in [
            "GET", "POST", "PUT", "PATCH", "DELETE", "HEAD", "OPTIONS", "patch",
        ] {
            let mut d = decl(
                "/things/{id}",
                vec![
                    ParamDecl {
                        name: "id".to_string(),
                        param_type: Some("string".to_string()),
                        required: true,
                        ..Default::default()
                    },
                    ParamDecl {
                        name: "note".to_string(),
                        param_type: Some("string".to_string()),
                        required: false,
                        ..Default::default()
                    },
                ],
            );
            d.method = Some(method.to_string());
            let op = super::build_operation("/things/{id}", method, &d);

            for p in &op.parameters {
                let expected_location = match d.param_position(&p.name) {
                    ParamPosition::Path => ParameterLocation::Path,
                    ParamPosition::Query => ParameterLocation::Query,
                    ParamPosition::Body => ParameterLocation::Body,
                };
                assert_eq!(
                    p.location,
                    expected_location,
                    "{method}: `{}` is {:?} for the D3 cap but {:?} on the wire — the cap's \
                     scope and the request's routing must not disagree",
                    p.name,
                    d.param_position(&p.name),
                    p.location
                );
            }

            // And the routing is self-consistent: a Body location exists exactly
            // when the request carries a body.
            assert_eq!(
                op.has_request_body,
                !op.body_parameters().is_empty(),
                "{method}: a Body-located parameter on a body-less request would be \
                 silently dropped"
            );
        }
    }

    /// The consequence CR-03 is really about: the D3 default `maxLength` is emitted
    /// for a parameter that travels in the query string and withheld from one that
    /// travels in the JSON payload — checked against the BUILT location rather than
    /// against a method list, so the two cannot drift.
    #[test]
    fn the_default_cap_is_emitted_exactly_where_the_value_travels_in_the_url() {
        use crate::config::{default_cap_applies, ValidationSection};

        let validation = ValidationSection::default();
        assert_ne!(
            validation.default_max_length, 0,
            "the default cap must be on for this row to mean anything"
        );

        for method in ["GET", "HEAD", "DELETE", "OPTIONS", "POST", "PUT", "PATCH"] {
            let param = ParamDecl {
                name: "note".to_string(),
                param_type: Some("string".to_string()),
                required: false,
                ..Default::default()
            };
            let mut d = decl("/things", vec![param.clone()]);
            d.method = Some(method.to_string());
            let op = super::build_operation("/things", method, &d);
            let location = param_named(&op, "note").location;
            let capped = default_cap_applies(&param, d.param_position("note"), &validation);

            assert_eq!(
                capped,
                location == ParameterLocation::Query,
                "{method}: an uncapped string in a {location:?} position is the CR-03 defect \
                 when that position is the query string, and the D-05 requirement when it is \
                 the payload"
            );
        }
    }
}

// -----------------------------------------------------------------------------
// Tests — Phase 90 OAPI-02a single-call HTTP synthesizer (feature `http`)
// -----------------------------------------------------------------------------

#[cfg(all(test, feature = "http"))]
mod synth_http_tests {
    use super::*;
    use crate::config::{ParamDecl, ServerConfig, ServerSection, ToolDecl};
    use crate::http::{HttpConnector, HttpConnectorError, Operation};
    use pmcp::RequestHandlerExtra;
    use serde_json::{json, Value};
    use std::sync::{Arc, Mutex};

    /// A mock [`HttpConnector`] that records the [`Operation`] it last received
    /// and returns a fixed JSON payload — so a synthesized handler can be
    /// invoked without any network.
    struct MockHttpConnector {
        last: Mutex<Option<Operation>>,
        payload: Value,
    }

    impl MockHttpConnector {
        fn new(payload: Value) -> Arc<Self> {
            Arc::new(Self {
                last: Mutex::new(None),
                payload,
            })
        }
    }

    #[async_trait]
    impl HttpConnector for MockHttpConnector {
        async fn execute(
            &self,
            operation: &Operation,
            _args: &Value,
        ) -> std::result::Result<Value, HttpConnectorError> {
            *self.last.lock().unwrap() = Some(operation.clone());
            Ok(self.payload.clone())
        }
        fn base_url(&self) -> &str {
            "https://mock.example.com"
        }
    }

    fn cfg_with_tools(tools: Vec<ToolDecl>) -> ServerConfig {
        ServerConfig {
            server: ServerSection {
                name: "demo".to_string(),
                version: "0.1.0".to_string(),
                ..Default::default()
            },
            tools,
            ..Default::default()
        }
    }

    /// (1) A single-call `[[tools]]` with a `{id}` path param synthesizes a
    /// `ToolInfo` whose input schema marks `id` required (object envelope), and
    /// the handler (wired to a mock connector) returns the mocked JSON.
    #[tokio::test]
    async fn synth_http_single_call_path_param_required_and_handler_returns_json() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "line_status".to_string(),
            description: Some("Line status".to_string()),
            path: Some("/Line/{id}/Status".to_string()),
            method: Some("GET".to_string()),
            parameters: vec![ParamDecl {
                name: "id".to_string(),
                param_type: Some("string".to_string()),
                required: true,
                ..Default::default()
            }],
            ..Default::default()
        }]);
        let connector = MockHttpConnector::new(json!({ "status": "Good Service" }));
        let out = synthesize_from_config_with_http_connector(&cfg, connector.clone())
            .expect("synthesize");
        assert_eq!(out.len(), 1);
        let (name, info, handler) = &out[0];
        assert_eq!(name, "line_status");
        let schema = &info.input_schema;
        assert_eq!(schema["type"], "object");
        assert_eq!(schema["required"], json!(["id"]));
        assert_eq!(schema["additionalProperties"], Value::Bool(false));

        let extra = RequestHandlerExtra::default();
        let result = handler
            .handle(json!({ "id": "victoria" }), extra)
            .await
            .expect("handle");
        assert_eq!(result, json!({ "status": "Good Service" }));

        // The operation carried the `{id}` path param as a Path parameter.
        let op = connector
            .last
            .lock()
            .unwrap()
            .clone()
            .expect("operation recorded");
        let path_params: Vec<&str> = op
            .path_parameters()
            .iter()
            .map(|p| p.name.as_str())
            .collect();
        assert_eq!(path_params, vec!["id"]);
    }

    /// (2) A `POST` tool routes non-path args to the request body
    /// (`has_request_body` true) and the non-path param is NOT a path param.
    #[tokio::test]
    async fn synth_http_post_sets_request_body() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "create_item".to_string(),
            description: Some("Create".to_string()),
            path: Some("/items".to_string()),
            method: Some("post".to_string()),
            parameters: vec![ParamDecl {
                name: "title".to_string(),
                param_type: Some("string".to_string()),
                required: true,
                ..Default::default()
            }],
            ..Default::default()
        }]);
        let connector = MockHttpConnector::new(json!({ "ok": true }));
        let out = synthesize_from_config_with_http_connector(&cfg, connector.clone())
            .expect("synthesize");
        let (_, _, handler) = &out[0];
        let extra = RequestHandlerExtra::default();
        handler
            .handle(json!({ "title": "widget" }), extra)
            .await
            .expect("handle");
        let op = connector
            .last
            .lock()
            .unwrap()
            .clone()
            .expect("operation recorded");
        assert_eq!(op.method, "POST");
        assert!(op.has_request_body, "POST must carry a request body");
        assert!(op.path_parameters().is_empty());
    }

    /// (3) A tool with path + query params lands them in the right schema slots /
    /// `Operation` parameter locations.
    #[tokio::test]
    async fn synth_http_path_and_query_param_slots() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "search".to_string(),
            description: Some("Search".to_string()),
            path: Some("/repos/{owner}/issues".to_string()),
            method: Some("GET".to_string()),
            parameters: vec![
                ParamDecl {
                    name: "owner".to_string(),
                    param_type: Some("string".to_string()),
                    required: true,
                    ..Default::default()
                },
                ParamDecl {
                    name: "state".to_string(),
                    param_type: Some("string".to_string()),
                    required: false,
                    ..Default::default()
                },
            ],
            ..Default::default()
        }]);
        let connector = MockHttpConnector::new(json!([]));
        let out = synthesize_from_config_with_http_connector(&cfg, connector.clone())
            .expect("synthesize");
        let (_, _, handler) = &out[0];
        let extra = RequestHandlerExtra::default();
        handler
            .handle(json!({ "owner": "rust-lang", "state": "open" }), extra)
            .await
            .expect("handle");
        let op = connector
            .last
            .lock()
            .unwrap()
            .clone()
            .expect("operation recorded");
        let path_params: Vec<&str> = op
            .path_parameters()
            .iter()
            .map(|p| p.name.as_str())
            .collect();
        assert_eq!(path_params, vec!["owner"]);
        let query_params: Vec<&str> = op
            .query_parameters()
            .iter()
            .map(|p| p.name.as_str())
            .collect();
        assert_eq!(query_params, vec!["state"]);
    }

    /// (4) A per-tool `base_url` is reflected in the synthesized `Operation`
    /// (Codex MEDIUM — not dropped).
    #[tokio::test]
    async fn synth_http_per_tool_base_url_reflected() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "other_host".to_string(),
            description: Some("Other host".to_string()),
            path: Some("/ping".to_string()),
            method: Some("GET".to_string()),
            base_url: Some("https://other.example.com/v2".to_string()),
            ..Default::default()
        }]);
        let connector = MockHttpConnector::new(json!({ "pong": true }));
        let out = synthesize_from_config_with_http_connector(&cfg, connector.clone())
            .expect("synthesize");
        let (_, _, handler) = &out[0];
        let extra = RequestHandlerExtra::default();
        handler.handle(json!({}), extra).await.expect("handle");
        let op = connector
            .last
            .lock()
            .unwrap()
            .clone()
            .expect("operation recorded");
        assert_eq!(
            op.base_url.as_deref(),
            Some("https://other.example.com/v2"),
            "per-tool base_url must be reflected on the Operation, not dropped"
        );
    }

    /// (5) NEGATIVE: a `[[tools]]` missing `method` (and without `script`) is
    /// rejected with a typed `ToolkitError` (T-90-03-04 — never silently
    /// registered).
    #[test]
    fn synth_http_missing_method_rejected() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "broken".to_string(),
            description: Some("missing method".to_string()),
            path: Some("/items".to_string()),
            method: None,
            ..Default::default()
        }]);
        let connector = MockHttpConnector::new(json!(null));
        let err = synthesize_from_config_with_http_connector(&cfg, connector)
            .err()
            .expect("ill-formed single-call tool must be rejected");
        assert!(matches!(err, ToolkitError::Synth(_)));
    }

    /// (5b) NEGATIVE: on the single-call-only entry point, a `script` tool is
    /// rejected with a typed `ToolkitError` pointing at the `openapi-code-mode`
    /// script path — NOT a silent skip, NOT a panic. (The OpenAPI Code Mode
    /// entry point `synthesize_from_config_with_http_connector_and_scripts`
    /// synthesizes a real `ScriptToolHandler` — proven in `script_tool` tests.)
    #[test]
    fn synth_http_script_tool_without_engine_is_rejected() {
        let cfg = cfg_with_tools(vec![ToolDecl {
            name: "scripted".to_string(),
            description: Some("script tool".to_string()),
            script: Some("await api.get('/x')".to_string()),
            ..Default::default()
        }]);
        let connector = MockHttpConnector::new(json!(null));
        let err = synthesize_from_config_with_http_connector(&cfg, connector)
            .err()
            .expect("script tool on the single-call-only entry point must be rejected");
        match err {
            ToolkitError::Synth(msg) => {
                assert!(
                    msg.contains("openapi-code-mode"),
                    "seam message must point at the openapi-code-mode script path: {msg}"
                );
            },
            other => panic!("expected Synth error, got {other:?}"),
        }
    }
}

// -----------------------------------------------------------------------------
// Phase 128 E2 — the per-tool ArgumentValidator seam.
//
// A SIBLING of `mod tests`, selected by the `--lib tools::` filter. It drives the
// PUBLIC `*_and_hooks` synthesizer so the assertions hold through the surface an
// out-of-crate caller (`pmcp-openapi-server`) actually uses.
// -----------------------------------------------------------------------------

/// E2 ordering, the schema opt-out, and the empty case.
#[cfg(all(test, feature = "input-validation"))]
mod argument_validator_seam {
    use super::synthesize_from_config_and_hooks;
    use crate::config::ServerConfig;
    use crate::policy::{ArgumentRefusal, ArgumentValidator, ToolkitHooks};
    use pmcp::RequestHandlerExtra;
    use serde_json::{json, Value};
    use std::sync::atomic::{AtomicUsize, Ordering};
    use std::sync::Arc;

    /// Counts its invocations, then refuses when `end < start`. The refusal is a
    /// FIXED string: it names the rule, never a value.
    struct EndAfterStart {
        calls: Arc<AtomicUsize>,
    }

    impl ArgumentValidator for EndAfterStart {
        fn validate(&self, args: &Value) -> Result<(), ArgumentRefusal> {
            self.calls.fetch_add(1, Ordering::SeqCst);
            let start = args.get("start").and_then(Value::as_i64);
            let end = args.get("end").and_then(Value::as_i64);
            match (start, end) {
                (Some(s), Some(e)) if e < s => {
                    Err(ArgumentRefusal::new("`end` must not precede `start`"))
                },
                _ => Ok(()),
            }
        }
    }

    /// A `[[tools]]` whose declared schema permits any two integers, so the
    /// cross-field rule is one only E2 can express.
    fn cfg(enforce: bool) -> ServerConfig {
        let toml = format!(
            r#"
[server]
name = "range"
version = "0.1.0"

[server.validation]
enforce_input_schema = {enforce}

[[tools]]
name = "range_query"
description = "Query a range"
sql = "SELECT 1"

[[tools.parameters]]
name = "start"
type = "integer"
required = true

[[tools.parameters]]
name = "end"
type = "integer"
required = true
"#
        );
        ServerConfig::from_toml_strict_validated(&toml).expect("parse")
    }

    fn extra() -> RequestHandlerExtra {
        RequestHandlerExtra::default()
    }

    fn hooks(calls: &Arc<AtomicUsize>) -> ToolkitHooks {
        ToolkitHooks::default().with_argument_validator(
            "range_query",
            Arc::new(EndAfterStart {
                calls: Arc::clone(calls),
            }),
        )
    }

    #[tokio::test]
    async fn a_registered_validator_refuses_a_combination_the_schema_permits() {
        let calls = Arc::new(AtomicUsize::new(0));
        let tools = synthesize_from_config_and_hooks(&cfg(true), &hooks(&calls)).expect("synth");
        let (_name, _info, handler) = &tools[0];
        let err = handler
            .handle(json!({ "start": 10, "end": 2 }), extra())
            .await
            .expect_err("the validator refuses");
        assert!(
            err.to_string().contains("`end` must not precede `start`"),
            "the refusal must carry the validator's own message, got: {err}"
        );
        assert_eq!(calls.load(Ordering::SeqCst), 1);
    }

    /// The E2 refusal takes the same TOOL-LEVEL channel as the D1 schema refusal.
    /// A cross-field rule is as much the caller's to fix as a bad type is, and a
    /// model can only act on a tool result, not on a JSON-RPC error.
    #[tokio::test]
    async fn a_validator_refusal_is_a_tool_level_rejection_not_a_protocol_error() {
        let calls = Arc::new(AtomicUsize::new(0));
        let tools = synthesize_from_config_and_hooks(&cfg(true), &hooks(&calls)).expect("synth");
        let (_name, _info, handler) = &tools[0];
        let err = handler
            .handle(json!({ "start": 10, "end": 2 }), extra())
            .await
            .expect_err("the validator refuses");
        assert!(
            matches!(err, pmcp::Error::ToolRejected { .. }),
            "a validator refusal must be a tool-level rejection, got: {err:?}"
        );
        assert!(
            err.to_string().contains("`end` must not precede `start`"),
            "the validator's own message must survive the channel change, got: {err}"
        );
    }

    #[tokio::test]
    async fn a_registered_validator_allows_a_valid_combination() {
        let calls = Arc::new(AtomicUsize::new(0));
        let tools = synthesize_from_config_and_hooks(&cfg(true), &hooks(&calls)).expect("synth");
        let (_name, _info, handler) = &tools[0];
        // No SQL connector is wired, so the INNER handler errors — which is itself
        // the proof that the validator allowed the call through to it.
        let err = handler
            .handle(json!({ "start": 2, "end": 10 }), extra())
            .await
            .expect_err("no connector is wired");
        assert!(
            !err.to_string().contains("must not precede"),
            "the validator must NOT have refused, got: {err}"
        );
        assert_eq!(calls.load(Ordering::SeqCst), 1);
    }

    /// T-128-41: the ordering is the contract. A validator must never see
    /// arguments that failed the declared schema.
    #[tokio::test]
    async fn a_registered_validator_is_not_invoked_when_the_schema_refuses() {
        let calls = Arc::new(AtomicUsize::new(0));
        let tools = synthesize_from_config_and_hooks(&cfg(true), &hooks(&calls)).expect("synth");
        let (_name, _info, handler) = &tools[0];
        // `start` is a declared integer; a string violates the schema.
        let err = handler
            .handle(json!({ "start": "ten", "end": 2 }), extra())
            .await
            .expect_err("D1 refuses");
        assert!(
            !err.to_string().contains("must not precede"),
            "D1 must be the refuser, not E2, got: {err}"
        );
        assert_eq!(
            calls.load(Ordering::SeqCst),
            0,
            "the validator was invoked on arguments the schema already refused"
        );
    }

    /// The joint test also present in plan 03: turning off one enforcement must
    /// never silently turn off another.
    #[tokio::test]
    async fn a_registered_validator_still_runs_with_enforce_input_schema_false() {
        let calls = Arc::new(AtomicUsize::new(0));
        let tools = synthesize_from_config_and_hooks(&cfg(false), &hooks(&calls)).expect("synth");
        let (_name, _info, handler) = &tools[0];

        // Half one: the SCHEMA check is off, so an undeclared key is accepted.
        let err = handler
            .handle(json!({ "start": 2, "end": 10, "undeclared": 1 }), extra())
            .await
            .expect_err("no connector is wired");
        assert!(
            !err.to_string().contains("undeclared"),
            "with enforce_input_schema=false an undeclared key must be accepted, got: {err}"
        );

        // Half two: the VALIDATOR still refuses.
        let err = handler
            .handle(json!({ "start": 10, "end": 2 }), extra())
            .await
            .expect_err("the validator refuses");
        assert!(
            err.to_string().contains("`end` must not precede `start`"),
            "a registered validator must survive the schema opt-out, got: {err}"
        );
        assert_eq!(calls.load(Ordering::SeqCst), 2);
    }

    #[tokio::test]
    async fn a_tool_with_no_registered_validator_runs_d1_only() {
        let tools =
            synthesize_from_config_and_hooks(&cfg(true), &ToolkitHooks::default()).expect("synth");
        let (_name, _info, handler) = &tools[0];
        // D1 still refuses a schema violation.
        handler
            .handle(json!({ "start": "ten", "end": 2 }), extra())
            .await
            .expect_err("D1 refuses");
        // And a schema-valid call reaches the inner handler (which has no connector).
        let err = handler
            .handle(json!({ "start": 10, "end": 2 }), extra())
            .await
            .expect_err("no connector is wired");
        assert!(!err.to_string().contains("must not precede"));
    }

    /// `handle_output` must validate too, or an inner handler that overrides it
    /// becomes a path with schema enforcement and no custom rule.
    #[tokio::test]
    async fn handle_output_runs_the_validator_as_well() {
        let calls = Arc::new(AtomicUsize::new(0));
        let tools = synthesize_from_config_and_hooks(&cfg(true), &hooks(&calls)).expect("synth");
        let (_name, _info, handler) = &tools[0];
        let err = handler
            .handle_output(json!({ "start": 10, "end": 2 }), extra())
            .await
            .expect_err("the validator refuses");
        assert!(
            err.to_string().contains("`end` must not precede `start`"),
            "handle_output must run the validator too, got: {err}"
        );
        assert_eq!(calls.load(Ordering::SeqCst), 1);
    }
}

/// The once-at-startup enforcement report's line formats (Phase 128 D-07).
#[cfg(all(test, feature = "input-validation"))]
mod enforcement_report {
    use crate::config::ServerConfig;
    use crate::policy::{
        render_validation_report, ArgumentRefusal, ArgumentValidator, ReportLevel, ToolkitHooks,
    };
    use serde_json::Value;
    use std::sync::Arc;

    struct Never;
    impl ArgumentValidator for Never {
        fn validate(&self, _args: &Value) -> Result<(), ArgumentRefusal> {
            Err(ArgumentRefusal::new("disabled"))
        }
    }

    fn cfg(extra_validation: &str) -> ServerConfig {
        let toml = format!(
            r#"
[server]
name = "report"
version = "0.1.0"

[server.validation]
{extra_validation}

[[tools]]
name = "get_thing"
description = "Get a thing"
path = "/things/{{id}}"
method = "GET"

[[tools.parameters]]
name = "id"
type = "string"
required = true
pattern = "^[0-9]+$"
"#
        );
        ServerConfig::from_toml_strict_validated(&toml).expect("parse")
    }

    fn texts(lines: &[crate::policy::ReportLine]) -> String {
        lines
            .iter()
            .map(|l| l.text.clone())
            .collect::<Vec<_>>()
            .join("\n")
    }

    #[test]
    fn a_fully_enforcing_config_states_that_no_opt_out_is_active() {
        let lines = render_validation_report(&cfg(""), &ToolkitHooks::default());
        let joined = texts(&lines);
        assert!(
            joined.contains("input validation: schema_check=ON"),
            "{joined}"
        );
        assert!(
            joined.contains("tool 'get_thing' enforces"),
            "one line per tool: {joined}"
        );
        assert!(
            joined.contains("no [server.validation] opt-out is active"),
            "the report must state the ABSENCE of an opt-out, not stay silent: {joined}"
        );
        assert!(
            joined.contains("E1 RequestPolicy registered=false"),
            "{joined}"
        );
        assert!(
            joined.contains("no E2 ArgumentValidator is registered"),
            "{joined}"
        );
        assert!(
            lines.iter().all(|l| l.level == ReportLevel::Info),
            "a fully enforcing config emits no warning"
        );
    }

    #[test]
    fn every_active_opt_out_is_reported_at_warn_level() {
        let lines = render_validation_report(
            &cfg("enforce_input_schema = false\ndefault_max_length = 0\nadditional_properties = true"),
            &ToolkitHooks::default(),
        );
        let joined = texts(&lines);
        assert!(joined.contains("schema_check=OFF"), "{joined}");
        let warns: Vec<&str> = lines
            .iter()
            .filter(|l| l.level == ReportLevel::Warn)
            .map(|l| l.text.as_str())
            .collect();
        assert!(
            warns.len() >= 4,
            "an enforcement that is OFF must never read as on; got {warns:#?}"
        );
        assert!(
            warns.iter().any(|w| w.contains("opt-out ACTIVE")),
            "{warns:#?}"
        );
    }

    #[test]
    fn a_validator_for_an_undeclared_tool_name_is_warned_not_refused() {
        let hooks = ToolkitHooks::default()
            .with_argument_validator("get_thing", Arc::new(Never))
            .with_argument_validator("typoed_name", Arc::new(Never));
        let lines = render_validation_report(&cfg(""), &hooks);
        let joined = texts(&lines);
        assert!(
            joined.contains("E2 ArgumentValidator registered for get_thing, typoed_name"),
            "{joined}"
        );
        let warns: Vec<&str> = lines
            .iter()
            .filter(|l| l.level == ReportLevel::Warn)
            .map(|l| l.text.as_str())
            .collect();
        assert_eq!(warns.len(), 1, "exactly the typo warns: {warns:#?}");
        assert!(warns[0].contains("'typoed_name'"), "{:?}", warns[0]);
        assert!(warns[0].contains("will never run"), "{:?}", warns[0]);
    }

    /// SC-7: the report is built from declarations, so it cannot carry request
    /// data. Asserted rather than trusted, because the log is a channel.
    #[test]
    fn the_report_never_echoes_an_argument_value() {
        let lines = render_validation_report(&cfg(""), &ToolkitHooks::default());
        let joined = texts(&lines);
        for forbidden in ["Bearer", "app_key", "super-secret"] {
            assert!(!joined.contains(forbidden), "{joined}");
        }
    }
}