supercode-harness 0.4.13

//! §4.6 "Preset validation" (`docs/composable-harness/COMPOSABLE-HARNESS-
//! DESIGN.md`) — golden tests for the six reserved presets, P2 of the
//! composable-harness migration (design §5.2 phase **P2**, item 3: "Golden
//! tests per preset").
//!
//! Each test resolves `extends = "<preset>"` through
//! [`supercode_harness::configfile::resolve`] and asserts (a) load-bearing resolved
//! [`Config`] fields, (b) the module activation set, and (c) §4.6's own
//! per-preset verdict — which warnings fire and which don't, matched
//! against the design doc's mechanical re-validation table.

use std::collections::BTreeMap;

use supercode_harness::configfile::{resolve, ResolveOptions};
use supercode_harness::{ApprovalPolicy, SandboxPolicy};

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}"))
}

fn assert_modules(modules: &BTreeMap<String, bool>, on: &[&str], off: &[&str]) {
    for name in on {
        assert_eq!(
            modules.get(*name),
            Some(&true),
            "expected module `{name}` ON"
        );
    }
    for name in off {
        assert_ne!(
            modules.get(*name),
            Some(&true),
            "expected module `{name}` OFF"
        );
    }
}

// ---------------------------------------------------------------------------
// pi-core — §4.1. §4.6 verdict: PASS, one named residual (C3 mandatory,
// fires by design — "pi ships no popups/sandbox").
// ---------------------------------------------------------------------------
#[test]
fn pi_core_resolves_with_only_the_c3_mandatory_warning() {
    let r = resolve_preset("pi-core");

    // (a) load-bearing Config fields.
    assert_eq!(r.config.effort.as_deref(), Some("medium"));
    assert_eq!(r.config.max_tool_output_bytes, Some(51200));
    assert!(r.config.load_project_context);
    assert_eq!(
        r.config.tool_schema_tier,
        supercode_harness::SchemaTier::Full
    );
    assert_eq!(r.config.compact_after_messages, Some(0));

    // (b) module activation set — pi's kept-list ON, first-party omissions OFF.
    assert_modules(
        &r.modules,
        &[
            "trust",
            "session_tree",
            "session_share",
            "server",
            "plugins",
            "tui",
        ],
        &[
            "tools_search",
            "mcp",
            "subagents",
            "permissions",
            "plan_mode",
            "todos",
            "tools_background",
            "tools_web",
            "checkpoint",
            "memory",
            "hooks",
            "deferred_tools",
            "cache",
            "reduction",
            "model_catalog",
            "model_oauth",
            "tools_apply_patch",
            "tools_persistent_shell",
            "tools_question",
            "lsp",
            "formatters",
            "structured_output",
            "telemetry",
            "integrations",
        ],
    );

    // (c) §4.6 verdict: exactly the C3 mandatory warning, nothing else.
    assert_eq!(
        r.warnings.len(),
        1,
        "pi-core should carry exactly the C3 mandatory warning: {:?}",
        r.warnings
    );
    assert!(r.warnings[0].contains("C3"));
    assert_eq!(r.config.sandbox, SandboxPolicy::DangerFullAccess);
    assert_eq!(r.config.approval, ApprovalPolicy::Never);
}

// ---------------------------------------------------------------------------
// cc-parity — §4.2. §4.6 verdict: PASS, C6 resolved via
// background_prompts="parent" (schema key, not prose); no C3 (approval !=
// never); no other conflicts fire.
// ---------------------------------------------------------------------------
#[test]
fn cc_parity_resolves_with_zero_warnings_c6_satisfied_via_parent() {
    let r = resolve_preset("cc-parity");

    assert_eq!(r.config.model, "anthropic/claude-opus-4-8");
    assert_eq!(r.config.effort.as_deref(), Some("medium"));
    assert_eq!(r.config.approval, ApprovalPolicy::Untrusted);
    assert!(r.config.auto_approved_tools.contains("read_file"));

    assert_modules(
        &r.modules,
        &[
            "tools_search",
            "tools_web",
            "tools_question",
            "todos",
            "plan_mode",
            "subagents",
            "tools_background",
            "permissions",
            "permissions.rules",
            "permissions.protected_paths",
            "trust",
            "mcp",
            "deferred_tools",
            "hooks",
            "memory",
            "checkpoint",
            "session_tree",
            "model_catalog",
            "tui",
        ],
        &[
            "tools_apply_patch",
            "tools_persistent_shell",
            "lsp",
            "formatters",
            "session_share",
            "server",
            "reduction",
            "cache",
            "structured_output",
            "model_oauth",
            "permissions.sandbox", // module 12 (OS enforcement) explicitly off
        ],
    );

    // No C3 (approval=untrusted != never), no C1 (apply_patch off), no D-7
    // (read_file active), no D-9 (small_model set), C6 satisfied literally
    // via background_prompts="parent" — zero warnings.
    assert!(
        r.warnings.is_empty(),
        "cc-parity should resolve with zero warnings: {:?}",
        r.warnings
    );
}

// ---------------------------------------------------------------------------
// cx-parity — §4.3. §4.6 verdict: PASS. D-7 degrades to a warning (only
// `bash` active, no `read_file`); no C3; C6 satisfied only via the S8
// judgment-call (approval=model_requested), which this resolver records as
// an explicit warning rather than silent pass.
// ---------------------------------------------------------------------------
#[test]
fn cx_parity_resolves_with_d7_and_c6_s8_warnings_only() {
    let r = resolve_preset("cx-parity");

    assert_eq!(
        r.harness.core.tools.enabled.as_deref(),
        Some(&["bash".to_string(), "view_image".to_string()][..])
    );
    assert_eq!(r.harness.core.shell_env_snapshot, Some(true));
    // P5-1: approval="model_requested" is now a real ApprovalPolicy variant
    // (design §3.2 S8, built this unit) — cx-parity resolves to its
    // INTENDED posture (ModelRequested) instead of the pre-P5-1 fail-safe
    // to Untrusted.
    assert_eq!(r.config.approval, ApprovalPolicy::ModelRequested);
    assert_eq!(r.config.sandbox, SandboxPolicy::WorkspaceWrite);

    assert_modules(
        &r.modules,
        &[
            "tools_persistent_shell",
            "tools_apply_patch",
            "todos",
            "tools_web",
            "tools_background",
            "subagents",
            "deferred_tools",
            "structured_output",
            "permissions",
            "permissions.rules",
            "permissions.protected_paths",
            "trust",
            "mcp",
            "hooks",
            "model_catalog",
            "tui",
        ],
        &[
            "tools_search",
            "tools_question",
            "plan_mode",
            "memory",
            "checkpoint",
            "session_tree",
            "session_share",
            "lsp",
            "formatters",
            "server",
            "reduction",
            "cache",
            "model_oauth",
        ],
    );

    // C1 does NOT fire: edit_file/write_file simply aren't in
    // core.tools.enabled, so nothing co-advertises with apply_patch.
    assert!(!r.warnings.iter().any(|w| w.contains("C1")));
    // C3 does NOT fire (sandbox=workspace_write, approval=model_requested).
    assert!(!r.warnings.iter().any(|w| w.contains("C3")));
    // D-7 fires (bash-only read pathway).
    let d7: Vec<_> = r.warnings.iter().filter(|w| w.contains("D-7")).collect();
    assert_eq!(
        d7.len(),
        1,
        "expected exactly one D-7 warning: {:?}",
        r.warnings
    );
    // C6 fires as the S8 judgment-call warning (not a silent pass, not an error).
    let c6: Vec<_> = r.warnings.iter().filter(|w| w.contains("C6")).collect();
    assert_eq!(
        c6.len(),
        1,
        "expected exactly one C6 warning: {:?}",
        r.warnings
    );
    assert!(c6[0].contains("S8"));
    assert_eq!(
        r.warnings.len(),
        2,
        "expected exactly D-7 + C6, nothing else: {:?}",
        r.warnings
    );
}

// ---------------------------------------------------------------------------
// oc-parity — §4.4. §4.6 verdict: PASS. C1 resolved via per_model +
// model_catalog bits; trust ON as a deliberate safety deviation (oc itself
// lacks a trust gate, D-10 forces it here); no C3 (approval != never). The
// four narrative residuals §4.6 names (trust-on, 3 C5 rule-translation
// deviations, plan_mode-off rationale, checkpoint partial-coverage) are
// design-document-level notes, not resolver-emitted warnings in P2's
// implemented scope (C5 rule translation and checkpoint sandbox-coverage
// are not resolver-checked — see validate_modules's doc comment) — so this
// resolver's own warnings list is empty, which is the correct outcome for
// what P2 actually checks.
// ---------------------------------------------------------------------------
#[test]
fn oc_parity_resolves_with_zero_warnings_trust_on_as_deviation() {
    let r = resolve_preset("oc-parity");

    assert_eq!(r.config.approval, ApprovalPolicy::OnRequest);
    assert_eq!(r.config.sandbox, SandboxPolicy::DangerFullAccess);

    assert_modules(
        &r.modules,
        &[
            "tools_search",
            "todos",
            "tools_web",
            "subagents",
            "tools_apply_patch",
            "permissions",
            "permissions.rules",
            "trust", // deliberate deviation (N7): oc lacks this, D-10 forces it on
            "mcp",
            "plugins",
            "lsp",
            "formatters",
            "checkpoint",
            "session_share",
            "server",
            "model_catalog",
            "tui",
        ],
        &[
            "plan_mode", // S18: opencode's plan_enter/plan_exit are deny-by-default
            "tools_question",
            "tools_background",
            "session_tree",
            "memory",
            "hooks",
            "deferred_tools",
            "cache",
            "reduction",
            "structured_output",
            "model_oauth",
            "permissions.protected_paths", // oc does .env protection via rules, not this module
        ],
    );

    assert!(
        r.warnings.is_empty(),
        "oc-parity should resolve with zero resolver-checked warnings: {:?}",
        r.warnings
    );
}

// ---------------------------------------------------------------------------
// token-saver — §4.5. §4.6 verdict: PASS, inherits pi-core's chain (D-10
// plugins→trust still met) and its one C3 residual; D-9 met directly
// (small_model set); no new conflict from the reduction/deferred/cache
// stack (that's what the `cache` module is the referee for).
// ---------------------------------------------------------------------------
#[test]
fn token_saver_extends_pi_core_chain_and_inherits_only_c3() {
    let r = resolve_preset("token-saver");

    // extends chain resolves root-first.
    assert_eq!(
        r.preset_chain,
        vec!["pi-core".to_string(), "token-saver".to_string()]
    );

    // C9: global minimal tier, edit_file pinned back to full.
    assert_eq!(
        r.config.tool_schema_tier,
        supercode_harness::SchemaTier::Minimal
    );
    assert_eq!(
        r.config.schema_tier_for("edit_file"),
        supercode_harness::SchemaTier::Full
    );
    let reduction = &r.config.reduction_policy;
    assert_eq!(reduction.stale_reads, Some(true));
    assert_eq!(reduction.diff_reads, Some(true));
    assert_eq!(reduction.tool_input_elision, Some(true));
    assert_eq!(reduction.supersede, Some(true));
    assert_eq!(reduction.normalize_output, Some(true));
    assert_eq!(reduction.image_redaction, Some(true));
    assert_eq!(reduction.span_summaries, Some(true));

    // D-9 met via small_model — no fallback warning.
    assert_modules(
        &r.modules,
        &["reduction", "deferred_tools", "cache", "model_catalog"],
        &[],
    );
    assert!(!r.warnings.iter().any(|w| w.contains("D-9")));

    // C3 inherited unchanged from pi-core (permissions untouched by
    // token-saver's own overrides).
    assert_eq!(
        r.warnings.len(),
        1,
        "token-saver should inherit exactly pi-core's C3 warning: {:?}",
        r.warnings
    );
    assert!(r.warnings[0].contains("C3"));

    // token-saver inherits pi-core's kept-list ON (trust/session_tree/
    // session_share/server/plugins, tui) and its first-party-omission OFFs
    // (mcp/permissions/subagents/todos/...) unchanged — only the reduction/
    // deferred_tools/cache/model_catalog stack differs (asserted above).
    assert_modules(
        &r.modules,
        &[
            "trust",
            "session_tree",
            "session_share",
            "server",
            "plugins",
            "tui",
        ],
        &[
            "mcp",
            "permissions",
            "subagents",
            "todos",
            "tools_background",
            "tools_web",
        ],
    );
}

#[test]
fn direct_reduction_gate_overrides_token_saver_preset() {
    let r = resolve(
        "extends = \"token-saver\"\n[capabilities.reduction]\nstale_reads = false\n",
        None,
        &ResolveOptions::default(),
    )
    .expect("direct reduction override resolves");
    assert_eq!(r.config.reduction_policy.stale_reads, Some(false));
    assert_eq!(r.config.reduction_policy.diff_reads, Some(true));
    assert_eq!(r.config.reduction_policy.duplicates, Some(true));
    assert!(r.config.handoff_enabled);
}

#[test]
fn reduction_master_false_disables_the_separate_handoff_consumer() {
    let r = resolve(
        "extends = \"token-saver\"\n[capabilities.reduction]\nenabled = false\n",
        None,
        &ResolveOptions::default(),
    )
    .expect("token-saver with reduction master disabled must resolve");
    assert!(!r.config.handoff_enabled);
}

// ---------------------------------------------------------------------------
// supercode-default — §4 intro (S10 fix). §4.6 verdict: PASS, two named
// residuals: C1 (apply_patch co-advertised with edit_file/write_file, no
// model_catalog bits — "faithful to today's actual unfiltered default
// stack") AND C3 (mandatory, same as pi-core). No plugins→trust edge at all
// (both off, S10's whole point).
// ---------------------------------------------------------------------------
#[test]
fn supercode_default_resolves_with_c1_and_c3_no_plugins_trust_edge() {
    let r = resolve_preset("supercode-default");

    // Core knobs identical to pi-core, unchanged.
    assert_eq!(r.config.effort.as_deref(), Some("medium"));
    assert_eq!(r.config.max_tool_output_bytes, Some(51200));

    assert_modules(
        &r.modules,
        &[
            "tools_search",
            "tools_apply_patch",
            "tools_persistent_shell",
            "todos",
            "notify",
        ],
        &[
            "trust",
            "session_tree",
            "session_share",
            "server",
            "plugins",
            "reduction",
            "permissions",
        ],
    );
    // tui stays on (§1.9 deviation, inherited from pi-core).
    assert_modules(&r.modules, &["tui"], &[]);

    assert_eq!(
        r.warnings.len(),
        2,
        "supercode-default should carry exactly C1 + C3: {:?}",
        r.warnings
    );
    assert!(r.warnings.iter().any(|w| w.contains("C1")));
    assert!(r.warnings.iter().any(|w| w.contains("C3")));

    // No plugins→trust dependency edge is even evaluated (S10: both off) —
    // resolution must not error, and no D-10 message should appear.
    assert!(!r.warnings.iter().any(|w| w.contains("D-10")));
}

/// A no-config-file resolution must land on `supercode-default` semantics
/// (design §4 intro: "supercode with no config file resolves to this
/// preset ... the current default stack's C3 exposure is named and warned
/// about rather than implicit"). This proves the PRESET's own shape matches
/// today's actual `Config::default()`/`with_builtins()` posture — CLI
/// wiring of "no file -> extends=supercode-default" is a separate, later
/// step (see this crate's tests/composable_resolver.rs for the note on
/// what's left to P3).
#[test]
fn supercode_default_matches_todays_actual_default_posture() {
    let r = resolve_preset("supercode-default");
    // Today's literal Config::default() / tools/mod.rs defaults.
    assert_eq!(r.config.sandbox, SandboxPolicy::DangerFullAccess);
    assert_eq!(r.config.approval, ApprovalPolicy::Never);
}