supercode-harness 0.4.11

//! §5.2 "P3 — Module registry wired into the loop and `ToolRegistry`"
//! (`docs/composable-harness/COMPOSABLE-HARNESS-DESIGN.md`) — the P3 test
//! suite the design's own AC calls for:
//!
//! - a per-preset EXACT activation-set test (all 35 [`ModuleId`]s, not a
//!   sampled on/off subset — `composable_presets.rs` already covers the
//!   sampled form for §4.6's warning verdicts; this file is the exhaustive
//!   companion the P3 mandate specifically asks for);
//! - "disabled ⇒ contributes nothing", parameterized over every module with
//!   a tool contribution (§5.3 risk 5's shape);
//! - `todos`-off/`update_plan` and `skills`/`read_file`/`bash` cases (D-7);
//! - the mandatory risk-2 mitigation: `module_registry = false` (default)
//!   makes [`ToolRegistry::from_config`] byte-identical to
//!   [`ToolRegistry::with_builtins`];
//! - C1 warning cases (`tools_apply_patch` co-advertised with
//!   `edit_file`/`write_file` without per-model bits);
//! - SECURITY carry-forward case-sensitivity rejection tests for
//!   `[capabilities.permissions.*]`.

use std::collections::BTreeSet;

use supercode_harness::configfile::{resolve, HarnessConfig, ResolveOptions};
use supercode_harness::{Config, ModuleActivation, ModuleId, ToolRegistry};

fn resolve_preset(name: &str) -> supercode_harness::configfile::Resolved {
    let top = format!("extends = \"{name}\"\n");
    resolve(&top, None, &ResolveOptions::default())
        .unwrap_or_else(|e| panic!("preset `{name}` failed to resolve: {e}"))
}

/// Assert the FULL 35-module activation set matches `expected` exactly — not
/// a sampled on/off subset. A module present in neither list is asserted
/// OFF (§3.1: every module defaults to `enabled = false`).
fn assert_exact_activation(hc: &HarnessConfig, expected: &[ModuleId]) {
    let act = ModuleActivation::from_harness(hc);
    let expected_set: BTreeSet<ModuleId> = expected.iter().copied().collect();
    for &m in ModuleId::ALL {
        let want = expected_set.contains(&m);
        let got = act.is_active(m);
        assert_eq!(
            got, want,
            "module `{m}`: expected active={want}, got active={got}"
        );
    }
    assert_eq!(
        act.len(),
        expected_set.len(),
        "activation set size mismatch (dedup check)"
    );
}

// ---------------------------------------------------------------------------
// Per-preset EXACT activation sets — transcribed directly from the compiled
// preset TOML in `crates/harness/src/presets.rs` (the actual resolver input,
// not the prose "on/off at a glance" tables, which only list load-bearing
// subsets per design §4's own convention).
// ---------------------------------------------------------------------------

#[test]
fn pi_core_exact_activation_set() {
    let r = resolve_preset("pi-core");
    assert_exact_activation(
        &r.harness,
        &[
            ModuleId::Trust,
            ModuleId::SessionTree,
            ModuleId::SessionShare,
            ModuleId::Server,
            ModuleId::Plugins,
            ModuleId::Tui,
        ],
    );
}

#[test]
fn cc_parity_exact_activation_set() {
    let r = resolve_preset("cc-parity");
    assert_exact_activation(
        &r.harness,
        &[
            ModuleId::ToolsSearch,
            ModuleId::ToolsWeb,
            ModuleId::ToolsQuestion,
            ModuleId::Todos,
            ModuleId::PlanMode,
            ModuleId::Subagents,
            ModuleId::ToolsBackground,
            ModuleId::PermissionsApprovals,
            ModuleId::PermissionsRules,
            ModuleId::PermissionsProtectedPaths,
            ModuleId::Trust,
            ModuleId::McpClient,
            ModuleId::DeferredTools,
            ModuleId::Hooks,
            ModuleId::Memory,
            ModuleId::Checkpoint,
            ModuleId::SessionTree,
            ModuleId::ModelCatalog,
            ModuleId::Tui,
        ],
    );
}

#[test]
fn cx_parity_exact_activation_set() {
    let r = resolve_preset("cx-parity");
    assert_exact_activation(
        &r.harness,
        &[
            ModuleId::ToolsPersistentShell,
            ModuleId::ToolsApplyPatch,
            ModuleId::Todos,
            ModuleId::ToolsWeb,
            ModuleId::ToolsBackground,
            ModuleId::Subagents,
            ModuleId::DeferredTools,
            ModuleId::StructuredOutput,
            ModuleId::PermissionsApprovals,
            ModuleId::PermissionsRules,
            ModuleId::PermissionsProtectedPaths,
            ModuleId::Trust,
            ModuleId::McpClient,
            ModuleId::McpServer,
            ModuleId::Hooks,
            ModuleId::ModelCatalog,
            ModuleId::Tui,
        ],
    );
}

#[test]
fn oc_parity_exact_activation_set() {
    let r = resolve_preset("oc-parity");
    assert_exact_activation(
        &r.harness,
        &[
            ModuleId::ToolsSearch,
            ModuleId::Todos,
            ModuleId::ToolsWeb,
            ModuleId::Subagents,
            ModuleId::ToolsApplyPatch,
            ModuleId::PermissionsApprovals,
            ModuleId::PermissionsRules,
            ModuleId::Trust,
            ModuleId::McpClient,
            ModuleId::Plugins,
            ModuleId::Lsp,
            ModuleId::Formatters,
            ModuleId::Checkpoint,
            ModuleId::SessionShare,
            ModuleId::Server,
            ModuleId::ModelCatalog,
            ModuleId::Tui,
        ],
    );
}

#[test]
fn token_saver_exact_activation_set() {
    let r = resolve_preset("token-saver");
    // Inherits pi-core's set (extends = "pi-core") plus its own four
    // overrides — nothing pi-core turns off is touched.
    assert_exact_activation(
        &r.harness,
        &[
            ModuleId::Trust,
            ModuleId::SessionTree,
            ModuleId::SessionShare,
            ModuleId::Server,
            ModuleId::Plugins,
            ModuleId::Tui,
            ModuleId::Reduction,
            ModuleId::DeferredTools,
            ModuleId::Cache,
            ModuleId::ModelCatalog,
        ],
    );
}

#[test]
fn supercode_default_exact_activation_set() {
    let r = resolve_preset("supercode-default");
    // pi-core's set MINUS {trust, session_tree, session_share, server,
    // plugins} PLUS the six with_builtins() extras (tools_search,
    // tools_apply_patch, tools_persistent_shell, todos) + notify — the
    // literal "no config file" default (design §4 intro, S10 fix).
    assert_exact_activation(
        &r.harness,
        &[
            ModuleId::Tui,
            ModuleId::ToolsSearch,
            ModuleId::ToolsApplyPatch,
            ModuleId::ToolsPersistentShell,
            ModuleId::Todos,
            ModuleId::Notify,
        ],
    );
}

// ---------------------------------------------------------------------------
// Mandatory risk-2 mitigation (§5.3 risk 2): `module_registry = false`
// (default) ⇒ `ToolRegistry::from_config` is byte-identical to
// `ToolRegistry::with_builtins()` — same names, same order.
// ---------------------------------------------------------------------------

fn tool_names(r: &ToolRegistry) -> Vec<String> {
    r.iter().map(|t| t.name().to_string()).collect()
}

#[test]
fn module_registry_off_by_default_matches_with_builtins_exactly() {
    let config = Config::default();
    assert!(!config.module_registry, "module_registry must default off");
    let from_config = ToolRegistry::from_config(&config);
    let builtins = ToolRegistry::with_builtins();
    assert_eq!(tool_names(&from_config), tool_names(&builtins));
}

#[test]
fn module_registry_off_matches_with_builtins_even_with_a_nonempty_activation_set() {
    // Even if a caller populated `module_activation`/`core_tools_enabled`
    // without flipping the flag, the flag alone gates the new path — the
    // risk-2 mitigation is "flag off ⇒ identical", full stop.
    let r = resolve_preset("cc-parity");
    let mut config = r.config;
    assert!(!config.module_registry);
    config.module_activation = ModuleActivation::from_harness(&r.harness);
    let from_config = ToolRegistry::from_config(&config);
    let builtins = ToolRegistry::with_builtins();
    assert_eq!(tool_names(&from_config), tool_names(&builtins));
}

/// The resolver's own [`HarnessConfig::from_toml_str`] + `[experimental]
/// module_registry = true` end-to-end: `supercode-default` resolved with the
/// flag ON must still produce the SAME tool set as `with_builtins()` — it's
/// defined to reproduce today's unfiltered default stack (design §4 intro),
/// so flipping the flag on for exactly this preset must not observably
/// change the registry.
#[test]
fn supercode_default_with_module_registry_on_still_matches_with_builtins() {
    let top = "extends = \"supercode-default\"\n[experimental]\nmodule_registry = true\n";
    let r = resolve(top, None, &ResolveOptions::default()).expect("resolves");
    assert!(r.config.module_registry);
    let from_config = ToolRegistry::from_config(&r.config);
    let builtins = ToolRegistry::with_builtins();
    assert_eq!(tool_names(&from_config), tool_names(&builtins));
}

// ---------------------------------------------------------------------------
// "Disabled ⇒ contributes nothing" — parameterized over every module with a
// tool contribution (§5.3 risk 5's shape: "disabled ⇒ contributes nothing"
// is one parameterized test).
// ---------------------------------------------------------------------------

struct ToolModuleCase {
    module_toml: &'static str,
    tool_name: &'static str,
}

const TOOL_MODULE_CASES: &[ToolModuleCase] = &[
    ToolModuleCase {
        module_toml: "[capabilities.todos]\nenabled = true\n",
        tool_name: "update_plan",
    },
    ToolModuleCase {
        module_toml: "[capabilities.tools_apply_patch]\nenabled = true\n",
        tool_name: "apply_patch",
    },
    ToolModuleCase {
        module_toml: "[capabilities.tools_persistent_shell]\nenabled = true\n",
        tool_name: "shell",
    },
];

fn resolve_module_registry_on(extra: &str) -> supercode_harness::Config {
    let top = format!("[experimental]\nmodule_registry = true\n{extra}");
    resolve(&top, None, &ResolveOptions::default())
        .expect("resolves")
        .config
}

#[test]
fn disabled_module_contributes_no_tool_enabled_module_does() {
    for case in TOOL_MODULE_CASES {
        // Module OFF (nothing set beyond the flag): tool absent.
        let off_config = resolve_module_registry_on("");
        let off_registry = ToolRegistry::from_config(&off_config);
        assert!(
            off_registry.get(case.tool_name).is_none(),
            "`{}` must be ABSENT from the registry when its module is off",
            case.tool_name
        );

        // Module ON: tool present.
        let on_config = resolve_module_registry_on(case.module_toml);
        let on_registry = ToolRegistry::from_config(&on_config);
        assert!(
            on_registry.get(case.tool_name).is_some(),
            "`{}` must be PRESENT in the registry when its module is on",
            case.tool_name
        );
    }
}

#[test]
fn tools_search_off_means_none_of_list_dir_glob_search_are_registered() {
    let config = resolve_module_registry_on("");
    let registry = ToolRegistry::from_config(&config);
    for name in ["list_dir", "glob", "search"] {
        assert!(
            registry.get(name).is_none(),
            "`{name}` must be absent when tools_search is off"
        );
    }
}

#[test]
fn tools_search_on_registers_all_three_by_default() {
    let config = resolve_module_registry_on("[capabilities.tools_search]\nenabled = true\n");
    let registry = ToolRegistry::from_config(&config);
    for name in ["list_dir", "glob", "search"] {
        assert!(
            registry.get(name).is_some(),
            "`{name}` must be present when tools_search is on with default sub-flags"
        );
    }
}

#[test]
fn tools_search_sub_flags_narrow_individually() {
    let config = resolve_module_registry_on(
        "[capabilities.tools_search]\nenabled = true\nlist_dir = false\nglob = false\n",
    );
    let registry = ToolRegistry::from_config(&config);
    assert!(registry.get("list_dir").is_none());
    assert!(registry.get("glob").is_none());
    assert!(
        registry.get("search").is_some(),
        "content_search sub-flag stayed at its true default"
    );
}

#[test]
fn core_tools_enabled_list_gates_the_default_four() {
    // cx-parity's `core.tools.enabled = ["bash", "view_image"]` — no
    // read_file/write_file/edit_file at all (the C1 resolution: disabling
    // edit/write advertising).
    let config = resolve_module_registry_on("[core.tools]\nenabled = [\"bash\"]\n");
    let registry = ToolRegistry::from_config(&config);
    assert!(registry.get("bash").is_some());
    assert!(registry.get("read_file").is_none());
    assert!(registry.get("write_file").is_none());
    assert!(registry.get("edit_file").is_none());
}

// ---------------------------------------------------------------------------
// D-7 / skills prompt section: only appears when `core.skills.enabled` AND a
// read pathway (`read_file` or `bash`) is active — the design's own
// illustration of "a disabled module contributes no prompt sections".
// ---------------------------------------------------------------------------

fn system_prompt_for(top_toml: &str) -> String {
    let config = resolve(top_toml, None, &ResolveOptions::default())
        .expect("resolves")
        .config;
    let provider: Box<dyn supercode_harness::Provider> =
        Box::new(supercode_harness::OpenAiProvider::new(
            "http://127.0.0.1:0",
            "test-key",
            Default::default(),
        ));
    let agent = supercode_harness::Agent::with_provider(config, provider);
    agent.history()[0].content.clone().unwrap_or_default()
}

#[test]
fn skills_off_means_no_skills_section() {
    let top = "[experimental]\nmodule_registry = true\n[core.prompts]\nfoo = \"bar {args}\"\n";
    let prompt = system_prompt_for(top);
    assert!(
        !prompt.contains("# Skills"),
        "skills off must not add a section"
    );
}

#[test]
fn skills_on_with_read_file_active_adds_the_section() {
    let top = "[experimental]\nmodule_registry = true\n\
[core.skills]\nenabled = true\n\
[core.prompts]\nfoo = \"bar {args}\"\n";
    let prompt = system_prompt_for(top);
    assert!(prompt.contains("# Skills"), "expected a skills section");
    assert!(prompt.contains("foo"));
}

#[test]
fn skills_on_with_only_bash_active_still_adds_the_section_d7_warn_not_error() {
    let top = "[experimental]\nmodule_registry = true\n\
[core.tools]\nenabled = [\"bash\"]\n\
[core.skills]\nenabled = true\n\
[core.prompts]\nfoo = \"bar {args}\"\n";
    // D-7 (S3-amended): bash-only degrades to a resolver WARNING, not a hard
    // failure — the skills section still appears (cx-parity's own shape).
    let resolved =
        resolve(top, None, &ResolveOptions::default()).expect("resolves (warn, not error)");
    assert!(resolved.warnings.iter().any(|w| w.contains("D-7")));
    let provider: Box<dyn supercode_harness::Provider> =
        Box::new(supercode_harness::OpenAiProvider::new(
            "http://127.0.0.1:0",
            "test-key",
            Default::default(),
        ));
    let agent = supercode_harness::Agent::with_provider(resolved.config, provider);
    assert!(agent.history()[0]
        .content
        .as_deref()
        .unwrap_or("")
        .contains("# Skills"));
}

#[test]
fn skills_on_with_no_read_pathway_at_all_is_a_hard_resolve_error() {
    let top = "[experimental]\nmodule_registry = true\n\
[core.tools]\nenabled = []\n\
[core.skills]\nenabled = true\n";
    let err = resolve(top, None, &ResolveOptions::default())
        .expect_err("no read_file/bash at all must hard-error (D-7)");
    assert!(format!("{err}").contains("core.skills"));
}

#[test]
fn skills_gated_by_module_registry_flag_even_when_enabled_on_config() {
    // Flag OFF: even with `skills_enabled = true` set directly on a
    // hand-built `Config` (bypassing the resolver entirely), the assembly
    // site must not add the section — `module_registry` is the master gate
    // for ALL P3 behavior (§5.3 risk 2), checked before `skills_enabled` is
    // ever consulted.
    let config = supercode_harness::Config::builder()
        .skills_enabled(true)
        .core_tools_enabled(vec!["read_file".to_string()])
        .prompt("foo", "bar {args}")
        .build();
    assert!(!config.module_registry);
    let provider: Box<dyn supercode_harness::Provider> =
        Box::new(supercode_harness::OpenAiProvider::new(
            "http://127.0.0.1:0",
            "test-key",
            Default::default(),
        ));
    let agent = supercode_harness::Agent::with_provider(config, provider);
    assert!(
        !agent.history()[0]
            .content
            .as_deref()
            .unwrap_or("")
            .contains("# Skills"),
        "module_registry off must suppress the skills section regardless of skills_enabled"
    );
}

// ---------------------------------------------------------------------------
// C1 warning cases (§2.2 conflict 1): `tools_apply_patch` co-advertised with
// `edit_file`/`write_file` without per-model bits.
// ---------------------------------------------------------------------------

#[test]
fn c1_fires_when_apply_patch_and_edit_write_coexist_without_bits() {
    let top = "[core.tools]\nenabled = [\"read_file\", \"bash\", \"edit_file\", \"write_file\"]\n\
[capabilities.tools_apply_patch]\nenabled = true\n";
    let r = resolve(top, None, &ResolveOptions::default()).expect("resolves");
    assert!(
        r.warnings.iter().any(|w| w.contains("C1")),
        "expected a C1 warning: {:?}",
        r.warnings
    );
}

#[test]
fn c1_does_not_fire_with_per_model_bits_and_model_catalog_on() {
    let top = "[core.tools]\nenabled = [\"read_file\", \"bash\", \"edit_file\", \"write_file\"]\n\
[capabilities.tools_apply_patch]\nenabled = true\nper_model = true\n\
[capabilities.model_catalog]\nenabled = true\n";
    let r = resolve(top, None, &ResolveOptions::default()).expect("resolves");
    assert!(
        !r.warnings.iter().any(|w| w.contains("C1")),
        "expected no C1 warning with per_model bits + model_catalog on: {:?}",
        r.warnings
    );
}

#[test]
fn c1_does_not_fire_when_apply_patch_is_the_only_write_path() {
    // cx-parity's own shape: edit_file/write_file simply not enabled at all.
    let top = "[core.tools]\nenabled = [\"bash\"]\n\
[capabilities.tools_apply_patch]\nenabled = true\nper_model = true\n";
    let r = resolve(top, None, &ResolveOptions::default()).expect("resolves");
    assert!(!r.warnings.iter().any(|w| w.contains("C1")));
}

#[test]
fn supercode_default_carries_the_c1_warning_faithfully() {
    // §4.6's own verdict: supercode-default legitimately fires C1 (today's
    // actual with_builtins() posture has no per-model filtering at all).
    let r = resolve_preset("supercode-default");
    assert!(r.warnings.iter().any(|w| w.contains("C1")));
}

// ---------------------------------------------------------------------------
// SECURITY carry-forward: case-sensitive, deny-unknown-fields rejection for
// `[capabilities.permissions.*]` — a wrong-case key must never be silently
// honored.
// ---------------------------------------------------------------------------

#[test]
fn case_mismatched_sandbox_key_is_rejected_with_a_warning_not_honored() {
    let top = "[capabilities.permissions]\nenabled = true\nSandbox = \"danger_full_access\"\n";
    let r = resolve(top, None, &ResolveOptions::default()).expect("resolves (warns, not errors)");
    assert!(
        r.warnings.iter().any(|w| w.contains("SECURITY")),
        "expected a SECURITY case-sensitivity warning: {:?}",
        r.warnings
    );
    // Never honored: the lowercase-only reader never sees `Sandbox`, so the
    // resolved posture keeps the safe default (danger_full_access is
    // ALREADY the fail-safe default in this codebase's config model, so the
    // meaningful assertion is the resolver's `permissions.sandbox` MODULE
    // activation, which must NOT read the case-mismatched key as if it
    // named a real table).
    assert!(!supercode_harness::modules::ModuleId::PermissionsSandbox.is_active(&r.harness));
}

#[test]
fn case_mismatched_tier_key_inside_sandbox_table_is_rejected() {
    let top = "[capabilities.permissions]\nenabled = true\n\
[capabilities.permissions.sandbox]\nenabled = true\nTier = \"read_only\"\n";
    let r = resolve(top, None, &ResolveOptions::default()).expect("resolves (warns, not errors)");
    assert!(
        r.warnings.iter().any(|w| w.contains("SECURITY")),
        "expected a SECURITY case-sensitivity warning: {:?}",
        r.warnings
    );
}

#[test]
fn correctly_cased_permissions_keys_never_trip_the_security_warning() {
    let top = "[capabilities.permissions]\nenabled = true\napproval = \"untrusted\"\nsandbox = \"read_only\"\nauto_approved_tools = [\"read_file\"]\n\
[capabilities.permissions.rules]\nenabled = true\ndeny = []\nask = []\nallow = [\"*\"]\n\
[capabilities.permissions.protected_paths]\nenabled = true\npaths = [\".git/**\"]\n";
    let r = resolve(top, None, &ResolveOptions::default()).expect("resolves");
    assert!(
        !r.warnings.iter().any(|w| w.contains("SECURITY")),
        "correctly-cased keys must never trip the SECURITY warning: {:?}",
        r.warnings
    );
}

#[test]
fn every_reserved_preset_never_trips_the_security_warning() {
    // The six compiled-in presets are themselves the ground truth for
    // "correctly cased" — a regression here would mean a design-doc TOML
    // transcription typo slipped past P2's own golden tests.
    for name in supercode_harness::presets::RESERVED_PRESET_NAMES {
        let r = resolve_preset(name);
        assert!(
            !r.warnings.iter().any(|w| w.contains("SECURITY")),
            "preset `{name}` unexpectedly tripped the SECURITY warning: {:?}",
            r.warnings
        );
    }
}