Skip to content

graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs

Python

Models: retired gateway aliases leave the pickers; defaults and rungs move up a generation

justrachc41effb86ca8
diff --git a/spec/kernels/score.json b/spec/kernels/score.json
index 63934bcf..566e27a3 100644
--- a/spec/kernels/score.json
+++ b/spec/kernels/score.json
@@ -13421,13 +13421,13 @@
"source": "needle"
},
{
- "id": "gpt-5.5",
+ "id": "gpt-6-astra",
"tier": "frontier",
"source": "needle"
},
{
- "id": "gpt-5.6-sol",
- "tier": "frontier",
+ "id": "gpt-6-sol",
+ "tier": "mid",
"source": "needle"
},
{
@@ -13436,7 +13436,7 @@
"source": "needle"
},
{
- "id": "gpt-5.6-luna",
+ "id": "gpt-6-luna",
"tier": "small",
"source": "needle"
},
diff --git a/spec/ref/score.py b/spec/ref/score.py
index 056578a5..583fc5c4 100644
--- a/spec/ref/score.py
+++ b/spec/ref/score.py
@@ -90,7 +90,9 @@ NEEDLES: tuple[tuple[str, Tier], ...] = (
("nano", "small"),
("terra", "mid"),
("luna", "small"),
+ ("-sol", "mid"),
("opus", "frontier"),
+ ("gpt-6", "frontier"),
("gpt-5", "frontier"),
("deepseek-v4", "frontier"),
("grok-4", "frontier"),
@@ -106,10 +108,10 @@ NEEDLES: tuple[tuple[str, Tier], ...] = (
# id, expected needle tier (unknown = price fallback, out of the cube)
MODELS: tuple[tuple[str, Tier], ...] = (
("claude-opus-4-8", "frontier"),
- ("gpt-5.5", "frontier"),
- ("gpt-5.6-sol", "frontier"),
+ ("gpt-6-astra", "frontier"),
+ ("gpt-6-sol", "mid"),
("gpt-5.6-terra", "mid"),
- ("gpt-5.6-luna", "small"),
+ ("gpt-6-luna", "small"),
("deepseek-v4-pro", "frontier"),
("grok-4.3", "frontier"),
("claude-haiku-4-5", "small"),
@@ -218,6 +220,8 @@ def check_properties() -> int:
raise ValueError("flash-beats-family")
if class_of("gpt-5.6-terra") != "mid":
raise ValueError("terra-beats-family")
+ if class_of("gpt-6-sol") != "mid" or class_of("gpt-6-astra") != "frontier":
+ raise ValueError("sol-beats-family")
if class_of("deepseek-v4-flash") != "small":
raise ValueError("deepseek-flash-small")
check_diagram()
diff --git a/src/catalog_price.zig b/src/catalog_price.zig
index 38e70db4..c9c3f0ed 100644
--- a/src/catalog_price.zig
+++ b/src/catalog_price.zig
@@ -129,13 +129,13 @@ test "a gateway catalog price wins over the baked snapshot; other providers are
defer pricing.active_model_table = saved;
pricing.active_model_table = &.{
.{ .provider = "codegraff", .name = "gemini-3.8-flash", .context = 1_048_576, .price = .{ .in = 1, .out = 2, .cache = 0.1 } },
- .{ .provider = "codegraff", .name = "gpt-5.5", .context = 1_050_000 },
+ .{ .provider = "codegraff", .name = "gpt-6-sol", .context = 1_050_000 },
};
const live = pricing.priceForProvider("codegraff", "gemini-3.8-flash").?;
try std.testing.expect(approx(live.in, 1) and approx(live.out, 2));
// No catalog price: the baked gateway row still answers.
- const baked = pricing.priceForProvider("codegraff", "gpt-5.5").?;
- try std.testing.expect(approx(baked.in, 4));
+ const baked = pricing.priceForProvider("codegraff", "gpt-6-sol").?;
+ try std.testing.expect(approx(baked.in, 2.2));
// A direct provider never reads the gateway's catalog.
try std.testing.expect(pricing.priceForProvider("google", "gemini-3.8-flash") == null or !approx(pricing.priceForProvider("google", "gemini-3.8-flash").?.in, 1));
}
diff --git a/src/commands_misc.zig b/src/commands_misc.zig
index a6d90770..09e0ff6d 100644
--- a/src/commands_misc.zig
+++ b/src/commands_misc.zig
@@ -403,7 +403,7 @@ pub fn tryHandle(root: *Agent, keys: *Keys, arena: Allocator, line: []const u8,
}
if (printed_header) try out.print("{s} copy an id above: /model lmstudio <id>{s}\n", .{ style.dim, style.reset });
}
- try out.print("{s}tip: /model <name> (fuzzy) · /model <provider> (its default) · /model <provider> <model> (pin, e.g. 'codex gpt-5.5'){s}\n", .{ style.dim, style.reset });
+ try out.print("{s}tip: /model <name> (fuzzy) · /model <provider> (its default) · /model <provider> <model> (pin, e.g. 'codex gpt-6-sol'){s}\n", .{ style.dim, style.reset });
try out.flush();
return true;
}
diff --git a/src/commands_model.zig b/src/commands_model.zig
index ea94e820..b01d2120 100644
--- a/src/commands_model.zig
+++ b/src/commands_model.zig
@@ -446,7 +446,7 @@ pub fn tryHandle(root: *Agent, keys: *Keys, arena: Allocator, line: []const u8,
var mbuf: [320]u8 = undefined; // the message, or (first 16 bytes) just the size
const staged = stageImagePath(root, path);
switch (staged) {
- .no_vision => try out.print("⚠ {s} can't see images — switch to a vision model first, e.g. /model claude-opus-4-8 or /model gpt-5.5\n", .{root.provider.model}),
+ .no_vision => try out.print("⚠ {s} can't see images — switch to a vision model first, e.g. /model claude-opus-4-8 or /model gpt-6-sol\n", .{root.provider.model}),
.ok => |o| try out.print("📎 attached {s} ({s}) — sent with your next message\n", .{ path, vision.fmtBytes(mbuf[0..16], o.bytes) }),
// Sized, distinct lines instead of "missing, or larger than 5MB" (#349).
.too_large, .not_found, .read_error => try out.print("{s}\n", .{vision.stageMessage(&mbuf, staged, path)}),
diff --git a/src/fleet.zig b/src/fleet.zig
index 3221871c..56c83cdd 100644
--- a/src/fleet.zig
+++ b/src/fleet.zig
@@ -66,7 +66,7 @@ pub const AgentType = struct {
// validated here — resolution happens per spawn against the live catalog
// (subagent_pin.zig) so an unavailable pin degrades to the session
// default instead of failing a persona load or a spawn mid-fleet.
- model: ?[]const u8 = null, // frontmatter `model: gpt-5.6-terra`
+ model: ?[]const u8 = null, // frontmatter `model: gpt-6-luna`
tier: ?Tier = null, // frontmatter `tier: mid`
effort: ?root.ReasoningEffort = null, // frontmatter `effort: max` — reasoning depth, an axis independent of model/tier (vocabulary: subagent_pin.parseEffort)
};
diff --git a/src/pipeline_score.zig b/src/pipeline_score.zig
index d9952f6c..04ed8986 100644
--- a/src/pipeline_score.zig
+++ b/src/pipeline_score.zig
@@ -119,7 +119,7 @@ test "#296: a pipeline stage accrues exactly ONE fitness row, celled on its slot
var stat: StageStat = .{};
for (0..5) |_| stat.noteAttempt();
for (0..4) |_| stat.noteOk();
- try std.testing.expect(captureStage(codex("gpt-5.6-terra"), "transform each file", "transform each file", "Apply the codemod; keep behavior identical.", &stat));
+ try std.testing.expect(captureStage(codex("gpt-6-sol"), "transform each file", "transform each file", "Apply the codemod; keep behavior identical.", &stat));
const rows = out.writer.buffered();
try std.testing.expectEqual(@as(usize, 1), std.mem.count(u8, rows, "\"kind\":\"score\""));
@@ -130,7 +130,7 @@ test "#296: a pipeline stage accrues exactly ONE fitness row, celled on its slot
try std.testing.expectEqualStrings("migration", cells[0].shape);
try std.testing.expectEqualStrings("transform", cells[0].role);
try std.testing.expectEqualStrings("mid", cells[0].tier);
- try std.testing.expectEqualStrings("gpt-5.6-terra", cells[0].model);
+ try std.testing.expectEqualStrings("gpt-6-sol", cells[0].model);
try std.testing.expectEqual(@as(u32, 1), cells[0].n);
try std.testing.expectApproxEqAbs(@as(f64, 0.8), cells[0].mean(), 1e-9);
}
@@ -150,20 +150,20 @@ test "#296: rounds with no signal file nothing — unreached, all-fail, off-voca
// A stage no item reached says nothing about its genome.
var unreached: StageStat = .{};
- try std.testing.expect(!captureStage(codex("gpt-5.6-terra"), "transform each file", "", "g", &unreached));
+ try std.testing.expect(!captureStage(codex("gpt-6-sol"), "transform each file", "", "g", &unreached));
// Total failure is skipped, not filed as a cell-poisoning 0.
var all_fail: StageStat = .{};
all_fail.noteAttempt();
- try std.testing.expect(!captureStage(codex("gpt-5.6-terra"), "transform each file", "", "g", &all_fail));
+ try std.testing.expect(!captureStage(codex("gpt-6-sol"), "transform each file", "", "g", &all_fail));
// An off-vocabulary stage runs fine but stays uncelled, like a phase.
var clean: StageStat = .{};
clean.noteAttempt();
clean.noteOk();
- try std.testing.expect(!captureStage(codex("gpt-5.6-terra"), "polish the files", "some-persona", "g", &clean));
+ try std.testing.expect(!captureStage(codex("gpt-6-sol"), "polish the files", "some-persona", "g", &clean));
// The fleet master switch gates local capture the same way it gates
// scoreVariants.
main_mod.g_fleet = false;
- try std.testing.expect(!captureStage(codex("gpt-5.6-terra"), "transform each file", "", "g", &clean));
+ try std.testing.expect(!captureStage(codex("gpt-6-sol"), "transform each file", "", "g", &clean));
try std.testing.expectEqual(@as(usize, 0), std.mem.count(u8, out.writer.buffered(), "\"kind\":\"score\""));
}
diff --git a/src/pricing_gateway.zig b/src/pricing_gateway.zig
index d52761fb..1d4187d2 100644
--- a/src/pricing_gateway.zig
+++ b/src/pricing_gateway.zig
@@ -17,12 +17,10 @@ pub const rows = [_]ModelPrice{
.{ .name = "glm-5.2", .in = 1.68, .out = 5.28, .cache = 0.312, .cache_write_multiplier = 1 },
.{ .name = "glm-5.3", .in = 1.68, .out = 5.28, .cache = 0.312, .cache_write_multiplier = 1 },
.{ .name = "glm-5.3-flash", .in = 0.075, .out = 0.25, .cache = 0.015, .cache_write_multiplier = 1 },
- .{ .name = "gpt-5.5", .in = 4.0, .out = 20.0, .cache = 0.4, .cache_write_multiplier = 1, .high_at = 272_000, .high_in = 8.0, .high_out = 30.0, .high_cache = 0.8 },
- .{ .name = "gpt-5.6", .in = 4.0, .out = 20.0, .cache = 0.4, .cache_write_multiplier = 1, .high_at = 272_000, .high_in = 8.0, .high_out = 30.0, .high_cache = 0.8 },
- .{ .name = "gpt-5.6-luna", .in = 0.2, .out = 1.2, .cache = 0.02, .cache_write_multiplier = 1, .high_at = 272_000, .high_in = 0.4, .high_out = 1.7999999999999998, .high_cache = 0.04 },
- .{ .name = "gpt-5.6-sol", .in = 4.0, .out = 20.0, .cache = 0.4, .cache_write_multiplier = 1, .high_at = 272_000, .high_in = 8.0, .high_out = 30.0, .high_cache = 0.8 },
- .{ .name = "gpt-5.6-terra", .in = 2.0, .out = 12.0, .cache = 0.2, .cache_write_multiplier = 1, .high_at = 272_000, .high_in = 4.0, .high_out = 18.0, .high_cache = 0.4 },
- .{ .name = "gpt-6-astra", .in = 10.0, .out = 50.0, .cache = 1.0, .cache_write_multiplier = 1 },
+ .{ .name = "gpt-6-astra", .in = 11.0, .out = 55.0, .cache = 1.1, .cache_write_multiplier = 1, .high_at = 272_000, .high_in = 22.0, .high_out = 82.5, .high_cache = 2.2 },
+ .{ .name = "gpt-6-luna", .in = 0.11, .out = 0.55, .cache = 0.011, .cache_write_multiplier = 1, .high_at = 272_000, .high_in = 0.22, .high_out = 0.825, .high_cache = 0.022 },
+ .{ .name = "gpt-6-sol", .in = 2.2, .out = 11.0, .cache = 0.22, .cache_write_multiplier = 1, .high_at = 272_000, .high_in = 4.4, .high_out = 16.5, .high_cache = 0.44 },
+ .{ .name = "gpt-6.1-sol", .in = 2.0, .out = 10.0, .cache = 0.1, .cache_write_multiplier = 1, .high_at = 272_000, .high_in = 4.0, .high_out = 15.0, .high_cache = 0.2 },
.{ .name = "grok-4.6", .in = 2.0, .out = 6.0, .cache = 0.5, .cache_write_multiplier = 1, .high_at = 200_000, .high_in = 4.0, .high_out = 12.0, .high_cache = 1.0 },
.{ .name = "grok-4.7", .in = 2.0, .out = 6.0, .cache = 0.5, .cache_write_multiplier = 1, .high_at = 200_000, .high_in = 4.0, .high_out = 12.0, .high_cache = 1.0 },
.{ .name = "grok-build", .in = 1.0, .out = 2.0, .cache = 0.2, .cache_write_multiplier = 1 },
diff --git a/src/pricing_models.zig b/src/pricing_models.zig
index b0954e06..a7aa875a 100644
--- a/src/pricing_models.zig
+++ b/src/pricing_models.zig
@@ -60,6 +60,7 @@ pub const rows = [_]ModelInfo{
.{ .provider = "codex", .name = "gpt-6-astra", .context = 272_000, .default_effort = "medium", .default_verbosity = "low" },
.{ .provider = "codex", .name = "gpt-6-sol", .context = 272_000, .default_effort = "medium", .default_verbosity = "low" },
.{ .provider = "codex", .name = "gpt-6.1-sol", .context = 272_000, .default_effort = "low", .default_verbosity = "low" },
+ .{ .provider = "codex", .name = "gpt-6-luna", .context = 272_000 },
.{ .provider = "codex", .name = "gpt-5.6-sol", .context = 272_000 },
.{ .provider = "codex", .name = "gpt-5.6-terra", .context = 272_000 },
.{ .provider = "codegraff", .name = "muse-spark-1.2", .context = 262_144 },
@@ -119,11 +120,9 @@ pub const rows = [_]ModelInfo{
.{ .provider = "codegraff", .name = "deepseek-v4-flash", .context = 1_000_000 },
.{ .provider = "codegraff", .name = "minimax-m3", .context = 1_000_000 },
.{ .provider = "codegraff", .name = "gpt-6-astra", .context = 1_050_000 },
- .{ .provider = "codegraff", .name = "gpt-5.6-sol", .context = 1_050_000 },
- .{ .provider = "codegraff", .name = "gpt-5.6-terra", .context = 1_050_000 },
- .{ .provider = "codegraff", .name = "gpt-5.6-luna", .context = 1_050_000 },
- .{ .provider = "codegraff", .name = "gpt-5.6", .context = 1_050_000 },
- .{ .provider = "codegraff", .name = "gpt-5.5", .context = 1_050_000 },
+ .{ .provider = "codegraff", .name = "gpt-6-sol", .context = 1_050_000 },
+ .{ .provider = "codegraff", .name = "gpt-6.1-sol", .context = 1_050_000 },
+ .{ .provider = "codegraff", .name = "gpt-6-luna", .context = 1_050_000 },
.{ .provider = "codegraff", .name = "kimi-k2.6", .context = 262_144 },
.{ .provider = "codegraff", .name = "grok-4.7", .context = 500_000 },
.{ .provider = "codegraff", .name = "grok-4.6", .context = 500_000 },
diff --git a/src/pricing_tests.zig b/src/pricing_tests.zig
index 3b5ef87d..a9828b2b 100644
--- a/src/pricing_tests.zig
+++ b/src/pricing_tests.zig
@@ -324,12 +324,12 @@ test "pricing Astra has metered catalog entries and cache write premium" {
}
test "pricing gateway aliases do not inherit vendor rates or cache writes" {
- const gateway = pricing.priceForProvider("codegraff", "gpt-5.5").?;
- const direct = pricing.priceForProvider("openai", "gpt-5.5").?;
- try std.testing.expectEqual(@as(f64, 4), gateway.in);
- try std.testing.expectEqual(@as(f64, 5), direct.in);
+ const gateway = pricing.priceForProvider("codegraff", "gpt-6-sol").?;
+ const direct = pricing.priceForProvider("openai", "gpt-6-sol").?;
+ try std.testing.expectApproxEqAbs(@as(f64, 2.2), gateway.in, 1e-9);
+ try std.testing.expectEqual(@as(f64, 2), direct.in);
const astra = pricing.priceForProvider("codegraff", "gpt-6-astra").?;
- try std.testing.expectApproxEqAbs(@as(f64, 1), pricing.usdForUsage(astra, astra.name, 0, 0, 100_000, 0), 1e-9);
+ try std.testing.expectApproxEqAbs(@as(f64, 1.1), pricing.usdForUsage(astra, astra.name, 0, 0, 100_000, 0), 1e-9);
}
test "pricing gateway tally respects priced and free rows without vendor fallback" {
@@ -338,7 +338,9 @@ test "pricing gateway tally respects priced and free rows without vendor fallbac
try std.testing.expectApproxEqAbs(@as(f64, 0.13086), tally.usd, 1e-9);
const zero = pricing.priceForProvider("codegraff", "ling-3.0-flash-fin").?;
try std.testing.expectEqual(@as(f64, 0), pricing.usdFor(zero, 100, 100, 100));
- try std.testing.expect(pricing.priceForProvider("codegraff", "gpt-6-sol") == null);
+ // A retired gateway alias has no gateway row and never borrows the vendor's.
+ try std.testing.expect(pricing.priceFor("gpt-5.5") != null);
+ try std.testing.expect(pricing.priceForProvider("codegraff", "gpt-5.5") == null);
}
test "pricing gateway overrides name-only runtime prices and long bands" {
diff --git a/src/provider.zig b/src/provider.zig
index 676fd24c..f005762e 100644
--- a/src/provider.zig
+++ b/src/provider.zig
@@ -79,7 +79,7 @@ pub const provider_specs = [_]ProviderSpec{
.{ .id = "anthropic", .display_name = "Anthropic", .kind = .anthropic, .auth = .x_api_key, .url = "https://api.anthropic.com/v1/messages", .env_key = "ANTHROPIC_API_KEY", .default_model = "claude-opus-5-5", .catalog = .anthropic, .models_url = "https://api.anthropic.com/v1/models?limit=1000" },
.{ .id = "codegraff", .display_name = "Codegraff", .kind = .openai, .auth = .bearer, .url = "https://gateway.codegraff.com/v1/chat/completions", .env_key = "CODEGRAFF_API_KEY", .default_model = "mimo-v2.6-pro", .login = .codegraff_device, .catalog = .openai, .models_url = "https://gateway.codegraff.com/v1/models", .takes_effort = true },
.{ .id = "deepseek", .display_name = "DeepSeek", .kind = .openai, .auth = .bearer, .url = "https://api.deepseek.com/chat/completions", .env_key = "DEEPSEEK_API_KEY", .default_model = "deepseek-v4-pro", .takes_effort = true },
- .{ .id = "openai", .display_name = "OpenAI", .kind = .responses, .auth = .bearer, .url = "https://api.openai.com/v1/responses", .env_key = "OPENAI_API_KEY", .default_model = "gpt-5.6" },
+ .{ .id = "openai", .display_name = "OpenAI", .kind = .responses, .auth = .bearer, .url = "https://api.openai.com/v1/responses", .env_key = "OPENAI_API_KEY", .default_model = "gpt-6-sol" },
// Google AI Studio on its first-party Interactions API — the wire Google
// says every new model and agent capability launches on, not the
// OpenAI-compatibility shim. A turn POSTs one Interaction and reads back
@@ -135,7 +135,7 @@ pub const provider_specs = [_]ProviderSpec{
// — it's the OAuth access token read from CODEX_HOME/auth.json at startup
// (see loadCodexAuth), the same on-disk-credential trick used for the
// codegraff gateway key in ~/forge/.credentials.json.
- .{ .id = "codex", .display_name = "Codex (ChatGPT)", .kind = .responses, .auth = .bearer, .url = "https://chatgpt.com/backend-api/codex/responses", .env_key = "CODEX_DISABLED", .default_model = "gpt-5.6-sol", .login = .codex_device, .sub_login = true, .catalog = .codex },
+ .{ .id = "codex", .display_name = "Codex (ChatGPT)", .kind = .responses, .auth = .bearer, .url = "https://chatgpt.com/backend-api/codex/responses", .env_key = "CODEX_DISABLED", .default_model = "gpt-6-sol", .login = .codex_device, .sub_login = true, .catalog = .codex },
};
/// Optional workspace-local router loaded from `.graff/.config.router`.
diff --git a/src/provider_codegraff_tests.zig b/src/provider_codegraff_tests.zig
index 72fcffcd..c084e52f 100644
--- a/src/provider_codegraff_tests.zig
+++ b/src/provider_codegraff_tests.zig
@@ -25,7 +25,7 @@ test "Codegraff selects the gateway wire and endpoint per model" {
test "Codegraff-only credentials route native aliases without live discovery" {
var keys: provider.Keys = .{ .values = @splat(null) };
try std.testing.expect(keys.set("codegraff", "key", .environment));
- for ([_][]const u8{ "gpt-5.6", "grok-4.6" }) |model| {
+ for ([_][]const u8{ "gpt-6-sol", "grok-4.6" }) |model| {
const got = try keys.providerFor(model);
try std.testing.expectEqualStrings("codegraff", got.id);
try std.testing.expectEqual(provider.Provider.Kind.responses, got.kind);
@@ -256,5 +256,5 @@ test "GPT6 Sol fallback routes preserve explicit provider and existing defaults"
try std.testing.expectEqualStrings("gpt-6-sol", selected.model);
try std.testing.expectEqual(provider.Provider.Kind.responses, selected.kind);
}
- try std.testing.expect(!pricing.providerModelInTable("codegraff", "gpt-6-sol"));
+ try std.testing.expect(pricing.providerModelInTable("codegraff", "gpt-6-sol"));
}
diff --git a/src/provider_routing_tests.zig b/src/provider_routing_tests.zig
index 77af6f2f..7016b44b 100644
--- a/src/provider_routing_tests.zig
+++ b/src/provider_routing_tests.zig
@@ -33,7 +33,7 @@ test "providerFor (#294): a catalogued model with no keyed provider fails instea
// A model served by several providers still falls through to whichever is
// keyed — losing one credential must not break a model another can serve.
- try std.testing.expectEqualStrings("openai", (try no_codex.providerFor("gpt-5.6-terra")).id);
+ try std.testing.expectEqualStrings("openai", (try no_codex.providerFor("gpt-6-luna")).id);
// The gateway fallback survives for genuinely UNCATALOGUED models, which is
// all it was ever meant to cover.
diff --git a/src/route_phase_tests.zig b/src/route_phase_tests.zig
index 06341d0c..23fb547b 100644
--- a/src/route_phase_tests.zig
+++ b/src/route_phase_tests.zig
@@ -19,24 +19,24 @@ fn codex(model: []const u8) Provider {
}
/// The same synthetic sheet #372's demonstration uses: on the SHEET ALONE
-/// terra outscores luna and is not dominated by it, so anything that moves
+/// sol outscores luna and is not dominated by it, so anything that moves
/// below is provably the lived evidence talking, not the priors.
const demo_sheet = [_]bench.Entry{
- .{ .model = "gpt-5.6-terra", .effort = "medium", .score = 0.60, .cost = 0.9 }, // effective $0.90
- .{ .model = "gpt-5.6-luna", .effort = "max", .score = 0.40, .cost = 0.61 }, // effective $0.61
+ .{ .model = "gpt-6-sol", .effort = "medium", .score = 0.60, .cost = 0.9 }, // effective $0.90
+ .{ .model = "gpt-6-luna", .effort = "max", .score = 0.40, .cost = 0.61 }, // effective $0.61
};
/// Lived scores from a `research` shape's `sweep` phase: the ladder's mid rung
-/// (terra) did badly there, the cheap rung (luna) did well. Exactly the rows
+/// (sol) did badly there, the cheap rung (luna) did well. Exactly the rows
/// route_trace.captureScore writes — no recipe rows, so bench_priors' own
/// #374 sheet feedback is deliberately NOT what is being exercised here.
const demo_archive =
- \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
- \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
- \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
+ \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
+ \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
;
/// Install `demo_archive` + `demo_sheet` as the session's live policy, with
@@ -82,26 +82,26 @@ test "role prior: search roles ride the small rung; judged roles keep the sessio
const saved = Env.install(a, "", true);
defer saved.restore();
- const find = phase.forPhase(codex("gpt-5.6-terra"), .review, "find the defects", &.{}, true);
- try std.testing.expectEqualStrings("gpt-5.6-luna", find.provider.model);
+ const find = phase.forPhase(codex("gpt-6-sol"), .review, "find the defects", &.{}, true);
+ try std.testing.expectEqualStrings("gpt-6-luna", find.provider.model);
try std.testing.expect(find.pin != null);
try std.testing.expect(find.role_prior);
// The source stays the ladder's — a hand rule must not claim evidence.
try std.testing.expectEqual(policy.Source.ladder, find.sourceFor(false));
- const sweep = phase.forPhase(codex("gpt-5.6-terra"), .research, "sweep the sources", &.{}, true);
- try std.testing.expectEqualStrings("gpt-5.6-luna", sweep.provider.model);
+ const sweep = phase.forPhase(codex("gpt-6-sol"), .research, "sweep the sources", &.{}, true);
+ try std.testing.expectEqualStrings("gpt-6-luna", sweep.provider.model);
- const verify = phase.forPhase(codex("gpt-5.6-terra"), .review, "verify each finding", &.{}, true);
+ const verify = phase.forPhase(codex("gpt-6-sol"), .review, "verify each finding", &.{}, true);
try std.testing.expect(verify.pin == null);
- try std.testing.expectEqualStrings("gpt-5.6-terra", verify.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", verify.provider.model);
// A user-chosen seat is never moved by the prior, same as #372's rule.
const pinned = Env.install(a, "", false);
defer pinned.restore();
- const kept = phase.forPhase(codex("gpt-5.6-terra"), .review, "find the defects", &.{}, true);
+ const kept = phase.forPhase(codex("gpt-6-sol"), .review, "find the defects", &.{}, true);
try std.testing.expect(kept.pin == null);
- try std.testing.expectEqualStrings("gpt-5.6-terra", kept.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", kept.provider.model);
}
test "#376 uniformRole: a phase routes only when all of its workers agree on one role" {
@@ -128,19 +128,19 @@ test "#376 DEMONSTRATION: an observed phase re-seats EVERY worker; an unobserved
const env = Env.install(arena_state.allocator(), demo_archive, true);
defer env.restore();
- // The session descended the #291 ladder to terra; the (research, sweep)
+ // The session descended the #291 ladder to sol; the (research, sweep)
// cell has measured luna as strictly better value there.
- const base = codex("gpt-5.6-terra");
+ const base = codex("gpt-6-sol");
const seat = phase.forPhase(base, .research, "sweep the repo", &sweep_niches, true);
try std.testing.expect(seat.pin != null);
- try std.testing.expectEqualStrings("gpt-5.6-luna", seat.provider.model);
- try std.testing.expectEqualStrings("gpt-5.6-luna", seat.pin.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", seat.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", seat.pin.?.model);
// PHASE-UNIFORM: one seat, so every worker of the phase resolves to the
// same model and files under the same cell — with or without a
// root-authored genome. That is the property #290 needs preserved.
for (sweep_niches) |niche| {
- try std.testing.expectEqualStrings("gpt-5.6-luna", seat.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", seat.provider.model);
const cell = seat.cellOf(niche);
try std.testing.expectEqual(policy.Shape.research, cell.shape);
try std.testing.expectEqualStrings("sweep", cell.role);
@@ -153,7 +153,7 @@ test "#376 DEMONSTRATION: an observed phase re-seats EVERY worker; an unobserved
// including today's `source` attribution.
const quiet = phase.forPhase(base, .feature, "implement the fix", &.{"implementer"}, true);
try std.testing.expect(quiet.pin == null);
- try std.testing.expectEqualStrings("gpt-5.6-terra", quiet.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", quiet.provider.model);
try std.testing.expectEqual(policy.Source.ladder, quiet.sourceFor(false));
try std.testing.expectEqual(policy.Source.workflow_override, quiet.sourceFor(true));
}
@@ -170,7 +170,7 @@ test "#376: the re-seated phase's trace says learned-policy on every worker" {
defer route_trace.g_test_sink = null;
// Drive the exact emission workflow.zig makes, once per worker.
- const seat = phase.forPhase(codex("gpt-5.6-terra"), .research, "sweep the repo", &sweep_niches, true);
+ const seat = phase.forPhase(codex("gpt-6-sol"), .research, "sweep the repo", &sweep_niches, true);
for (sweep_niches) |niche| {
route_trace.emitSpawnProvider(std.testing.io, null, niche, seat.provider, seat.cellOf(niche), seat.sourceFor(false), null, niche);
}
@@ -180,7 +180,7 @@ test "#376: the re-seated phase's trace says learned-policy on every worker" {
while (lines.next()) |ln| : (n += 1) {
const ev = try std.json.parseFromSliceLeaky(std.json.Value, a, ln, .{});
try std.testing.expectEqualStrings("agent_route", ev.object.get("type").?.string);
- try std.testing.expectEqualStrings("gpt-5.6-luna", ev.object.get("resolved_model").?.string);
+ try std.testing.expectEqualStrings("gpt-6-luna", ev.object.get("resolved_model").?.string);
try std.testing.expectEqualStrings("learned-policy", ev.object.get("source").?.string);
try std.testing.expectEqualStrings("research", ev.object.get("shape").?.string);
try std.testing.expectEqualStrings("sweep", ev.object.get("role").?.string);
@@ -196,7 +196,7 @@ test "#376: the re-seated phase's trace says learned-policy on every worker" {
test "#376: an explicitly chosen worker model outranks the phase policy" {
var arena_state = std.heap.ArenaAllocator.init(std.testing.allocator);
defer arena_state.deinit();
- const base = codex("gpt-5.6-terra");
+ const base = codex("gpt-6-sol");
// --subagent-model: the session named the worker model, so the route is
// `session-default`, not the ladder's, and evidence never overrides a
@@ -206,7 +206,7 @@ test "#376: an explicitly chosen worker model outranks the phase policy" {
defer env.restore();
const seat = phase.forPhase(base, .research, "sweep the repo", &sweep_niches, true);
try std.testing.expect(seat.pin == null);
- try std.testing.expectEqualStrings("gpt-5.6-terra", seat.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", seat.provider.model);
try std.testing.expectEqual(policy.Source.session_default, seat.sourceFor(false));
}
// A plain inherit-the-root (no worker provider at all) is the same story:
@@ -225,20 +225,20 @@ test "#376: the phase seat can only trade DOWN on cost, and never leaves the pro
defer arena_state.deinit();
const a = arena_state.allocator();
- // The mirror image of the demonstration: terra is the loved model but it
+ // The mirror image of the demonstration: sol is the loved model but it
// is also the DEARER one, so a phase already seated on luna can never be
- // moved up to it however well terra scored in the cell.
+ // moved up to it however well sol scored in the cell.
const upward =
- \\{"kind":"score","prompt_sha":"aa","score":1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"aa","score":1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"bb","score":0,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
- \\{"kind":"score","prompt_sha":"bb","score":0,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
+ \\{"kind":"score","prompt_sha":"aa","score":1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"aa","score":1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"bb","score":0,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
+ \\{"kind":"score","prompt_sha":"bb","score":0,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
;
const env = Env.install(a, upward, true);
defer env.restore();
- const seat = phase.forPhase(codex("gpt-5.6-luna"), .research, "sweep the repo", &sweep_niches, true);
+ const seat = phase.forPhase(codex("gpt-6-luna"), .research, "sweep the repo", &sweep_niches, true);
try std.testing.expect(seat.pin == null);
- try std.testing.expectEqualStrings("gpt-5.6-luna", seat.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", seat.provider.model);
// And whatever it decides, the seat is provider-local by construction: the
// id, the credential and the account all ride across unchanged, so no
@@ -248,7 +248,7 @@ test "#376: the phase seat can only trade DOWN on cost, and never leaves the pro
// explicit `tier` ask, which a phase task never makes.
const env2 = Env.install(a, demo_archive, true);
defer env2.restore();
- const base = codex("gpt-5.6-terra");
+ const base = codex("gpt-6-sol");
const moved = phase.forPhase(base, .research, "sweep the repo", &sweep_niches, true);
try std.testing.expect(moved.pin != null);
try std.testing.expectEqualStrings(base.id, moved.provider.id);
@@ -261,7 +261,7 @@ test "#376: an uncelled phase, and a bootstrap install, keep today's behavior ex
var arena_state = std.heap.ArenaAllocator.init(std.testing.allocator);
defer arena_state.deinit();
const a = arena_state.allocator();
- const base = codex("gpt-5.6-terra");
+ const base = codex("gpt-6-sol");
{
const env = Env.install(a, demo_archive, true);
defer env.restore();
@@ -269,7 +269,7 @@ test "#376: an uncelled phase, and a bootstrap install, keep today's behavior ex
// evidence that WOULD have moved a `sweep` phase moves nothing.
const seat = phase.forPhase(base, .research, "ponder", &.{ "researcher", "review the diff" }, true);
try std.testing.expect(seat.pin == null);
- try std.testing.expectEqualStrings("gpt-5.6-terra", seat.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", seat.provider.model);
}
{
// An install with no archive at all: zero cells — the learned layer
@@ -280,7 +280,7 @@ test "#376: an uncelled phase, and a bootstrap install, keep today's behavior ex
defer env.restore();
const seat = phase.forPhase(base, .research, "sweep the repo", &sweep_niches, true);
try std.testing.expect(seat.pin != null);
- try std.testing.expectEqualStrings("gpt-5.6-luna", seat.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", seat.provider.model);
try std.testing.expect(seat.role_prior);
try std.testing.expectEqual(policy.Source.ladder, seat.sourceFor(false));
}
diff --git a/src/route_policy_tests.zig b/src/route_policy_tests.zig
index f38848f1..ca38a4bc 100644
--- a/src/route_policy_tests.zig
+++ b/src/route_policy_tests.zig
@@ -26,28 +26,28 @@ fn codex(model: []const u8) Provider {
/// (0.60 vs 0.40) and is not dominated by it, so any re-seating that happens
/// is provably the lived evidence talking, not the priors.
const demo_sheet = [_]bench.Entry{
- .{ .model = "gpt-5.6-terra", .effort = "medium", .score = 0.60, .cost = 0.9 }, // effective $0.90
- .{ .model = "gpt-5.6-luna", .effort = "max", .score = 0.40, .cost = 0.61 }, // effective $0.61
+ .{ .model = "gpt-6-sol", .effort = "medium", .score = 0.60, .cost = 0.9 }, // effective $0.90
+ .{ .model = "gpt-6-luna", .effort = "max", .score = 0.40, .cost = 0.61 }, // effective $0.61
};
/// Lived scores from a `research` shape's `sweep` phase: terra (the ladder's
/// mid rung) did badly, luna did well. Written in the exact self-describing
/// `kind:"score"` shape route_trace.captureScore emits.
const demo_archive =
- \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
- \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
- \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
+ \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
+ \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
;
test "#372 trace line: every source kind renders its own attribution, absent fields render -" {
var buf: [320]u8 = undefined;
// The line the issue specifies, field for field.
try std.testing.expectEqualStrings(
- "shape=research role=sweep tier=mid resolved_model=gpt-5.6-terra source=ladder policy_or_genome_id=ladder:codex",
- policy.formatDecision(&buf, .{ .shape = .research, .role = "sweep", .tier = .mid, .resolved_model = "gpt-5.6-terra", .source = .ladder, .policy_id = "ladder:codex" }),
+ "shape=research role=sweep tier=mid resolved_model=gpt-6-sol source=ladder policy_or_genome_id=ladder:codex",
+ policy.formatDecision(&buf, .{ .shape = .research, .role = "sweep", .tier = .mid, .resolved_model = "gpt-6-sol", .source = .ladder, .policy_id = "ladder:codex" }),
);
// Each source kind is spelled with hyphens on the wire, and each is
// distinguishable — that is the whole point of the field.
@@ -82,14 +82,14 @@ test "#372 trace: the --json agent_route event carries the same six fields" {
route_trace.g_test_sink = &sink;
defer route_trace.g_test_sink = null;
- route_trace.emitSpawn(std.testing.io, null, "sweep the callers", "codex", "gpt-5.6-terra", .{ .shape = .research, .role = "sweep" }, .ladder, "");
+ route_trace.emitSpawn(std.testing.io, null, "sweep the callers", "codex", "gpt-6-sol", .{ .shape = .research, .role = "sweep" }, .ladder, "");
const line = std.mem.trimEnd(u8, sink.writer.buffered(), "\n");
try std.testing.expect(std.mem.indexOfScalar(u8, line, '\n') == null); // one complete object per line
const ev = try std.json.parseFromSliceLeaky(std.json.Value, arena_state.allocator(), line, .{});
try std.testing.expectEqualStrings("agent_route", ev.object.get("type").?.string);
try std.testing.expectEqualStrings("research", ev.object.get("shape").?.string);
try std.testing.expectEqualStrings("sweep", ev.object.get("role").?.string);
- try std.testing.expectEqualStrings("gpt-5.6-terra", ev.object.get("resolved_model").?.string);
+ try std.testing.expectEqualStrings("gpt-6-sol", ev.object.get("resolved_model").?.string);
try std.testing.expectEqualStrings("ladder", ev.object.get("source").?.string);
try std.testing.expectEqualStrings("sweep the callers", ev.object.get("task").?.string);
// The tier is READ BACK off the resolved model, so an inherited route —
@@ -127,9 +127,9 @@ test "#372: the shape is read off the phase titles the root model authored" {
try std.testing.expectEqualStrings("review", policy.roleOf("do the thing", "review the diff"));
try std.testing.expectEqualStrings("", policy.roleOf("do the thing", "some-persona"));
// tierOf is the ladder inverse the trace needs for an unpinned worker.
- try std.testing.expectEqual(tier_ladder.Tier.mid, policy.tierOf("codex", "gpt-5.6-terra").?);
+ try std.testing.expectEqual(tier_ladder.Tier.mid, policy.tierOf("codex", "gpt-6-sol").?);
try std.testing.expectEqual(tier_ladder.Tier.frontier, policy.tierOf("codex", "gpt-6-astra").?);
- try std.testing.expect(policy.tierOf("codex", "gpt-5.6-sol") == null);
+ try std.testing.expect(policy.tierOf("codex", "gpt-6.1-sol") == null);
try std.testing.expect(policy.tierOf("codex", "claude-opus-4-8") == null); // off this ladder
try std.testing.expectEqualStrings("", policy.tierLabelFor("xai", "grok-4.3")); // no ladder at all
}
@@ -139,12 +139,12 @@ test "#372 foldCells: only self-describing score rows become observations" {
defer arena_state.deinit();
const a = arena_state.allocator();
const archive =
- \\{"kind":"score","prompt_sha":"aa","score":0.5,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"aa","score":0.7,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.5,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.7,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
\\{"kind":"score","prompt_sha":"zz","score":0.9}
\\{"kind":"score","prompt_sha":"zz","score":0.9,"shape":"research","role":"sweep","tier":"mid"}
- \\{"kind":"score","prompt_sha":"zz","score":40,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-luna"}
- \\{"kind":"recipe","prompt_sha":"aa","model":"gpt-5.6-luna","effort":"max"}
+ \\{"kind":"score","prompt_sha":"zz","score":40,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-luna"}
+ \\{"kind":"recipe","prompt_sha":"aa","model":"gpt-6-luna","effort":"max"}
\\not json at all
;
const cells = policy.foldCells(a, archive);
@@ -154,7 +154,7 @@ test "#372 foldCells: only self-describing score rows become observations" {
try std.testing.expectEqualStrings("research", cells[0].shape);
try std.testing.expectEqualStrings("sweep", cells[0].role);
try std.testing.expectEqualStrings("mid", cells[0].tier);
- try std.testing.expectEqualStrings("gpt-5.6-terra", cells[0].model);
+ try std.testing.expectEqualStrings("gpt-6-sol", cells[0].model);
try std.testing.expectEqual(@as(u32, 2), cells[0].n);
try std.testing.expectApproxEqAbs(@as(f64, 0.6), cells[0].mean(), 1e-9);
// An archive with nothing in it is the bootstrap case: zero cells, so
@@ -166,8 +166,8 @@ test "#372 DEMONSTRATION: a dominated mid rung is re-seated, an unobserved cell
var arena_state = std.heap.ArenaAllocator.init(std.testing.allocator);
defer arena_state.deinit();
const cells = policy.foldCells(arena_state.allocator(), demo_archive);
- const ladder_mid = tier_ladder.forProvider("codex").?.mid.?; // today's answer: gpt-5.6-terra
- try std.testing.expectEqualStrings("gpt-5.6-terra", ladder_mid);
+ const ladder_mid = tier_ladder.forProvider("codex").?.mid.?; // today's answer: gpt-6-sol
+ try std.testing.expectEqualStrings("gpt-6-sol", ladder_mid);
// BASELINE — the sheet alone, no lived evidence: terra outscores luna on
// the sheet, so nothing dominates it and today's ladder answer stands.
@@ -178,7 +178,7 @@ test "#372 DEMONSTRATION: a dominated mid rung is re-seated, an unobserved cell
// luna now matches-and-beats terra's quality for $0.61 against $0.90, so
// it DOMINATES it and the (research, sweep) mid rung is re-seated.
const reseated = policy.learnedRungIn(cells, &demo_sheet, "codex", .{ .shape = .research, .role = "sweep" }, ladder_mid);
- try std.testing.expectEqualStrings("gpt-5.6-luna", reseated.?);
+ try std.testing.expectEqualStrings("gpt-6-luna", reseated.?);
// SPARSE — a cell nobody has ever scored has no basis to move anything,
// so it falls through the whole hierarchy to today's ladder answer.
@@ -197,30 +197,30 @@ test "#372 hierarchy: exact cell decides, a silent cell borrows its shape, a sil
// A DIFFERENT role in the same shape has no rows of its own, so it
// borrows the shape-level pool (`sweep`'s evidence) — rung two.
const cells = policy.foldCells(a, demo_archive);
- try std.testing.expectEqualStrings("gpt-5.6-luna", policy.learnedRungIn(cells, &demo_sheet, "codex", .{ .shape = .research, .role = "synthesize" }, ladder_mid).?);
+ try std.testing.expectEqualStrings("gpt-6-luna", policy.learnedRungIn(cells, &demo_sheet, "codex", .{ .shape = .research, .role = "synthesize" }, ladder_mid).?);
// But an exact cell that HAS evidence decides for itself, including
// deciding to keep the ladder's answer — it must not fall through to a
// shape-level pool that would have overridden it. Here `synthesize` says
// terra is excellent and luna is poor, so terra survives.
const keeps =
- \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
- \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
- \\{"kind":"score","prompt_sha":"cc","score":1,"shape":"research","role":"synthesize","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"cc","score":1,"shape":"research","role":"synthesize","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"dd","score":0,"shape":"research","role":"synthesize","tier":"small","model":"gpt-5.6-luna"}
- \\{"kind":"score","prompt_sha":"dd","score":0,"shape":"research","role":"synthesize","tier":"small","model":"gpt-5.6-luna"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.2,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
+ \\{"kind":"score","prompt_sha":"bb","score":0.95,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
+ \\{"kind":"score","prompt_sha":"cc","score":1,"shape":"research","role":"synthesize","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"cc","score":1,"shape":"research","role":"synthesize","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"dd","score":0,"shape":"research","role":"synthesize","tier":"small","model":"gpt-6-luna"}
+ \\{"kind":"score","prompt_sha":"dd","score":0,"shape":"research","role":"synthesize","tier":"small","model":"gpt-6-luna"}
;
const split = policy.foldCells(a, keeps);
try std.testing.expect(policy.learnedRungIn(split, &demo_sheet, "codex", .{ .shape = .research, .role = "synthesize" }, ladder_mid) == null);
- try std.testing.expectEqualStrings("gpt-5.6-luna", policy.learnedRungIn(split, &demo_sheet, "codex", .{ .shape = .research, .role = "sweep" }, ladder_mid).?);
+ try std.testing.expectEqualStrings("gpt-6-luna", policy.learnedRungIn(split, &demo_sheet, "codex", .{ .shape = .research, .role = "sweep" }, ladder_mid).?);
// A single lucky run is SPARSE (min_model_obs) and moves nothing.
const one_run =
- \\{"kind":"score","prompt_sha":"aa","score":0.1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"bb","score":1,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
+ \\{"kind":"score","prompt_sha":"aa","score":0.1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"bb","score":1,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
;
try std.testing.expect(policy.learnedRungIn(policy.foldCells(a, one_run), &demo_sheet, "codex", .{ .shape = .research, .role = "sweep" }, ladder_mid) == null);
}
@@ -232,14 +232,14 @@ test "#372: the learned policy can only ever trade DOWN on cost, never up" {
// is also the CHEAP one, so asking about the `small` rung can never be
// answered with the pricier terra however well terra scored.
const archive =
- \\{"kind":"score","prompt_sha":"aa","score":1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"aa","score":1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-5.6-terra"}
- \\{"kind":"score","prompt_sha":"bb","score":0,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
- \\{"kind":"score","prompt_sha":"bb","score":0,"shape":"research","role":"sweep","tier":"small","model":"gpt-5.6-luna"}
+ \\{"kind":"score","prompt_sha":"aa","score":1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"aa","score":1,"shape":"research","role":"sweep","tier":"mid","model":"gpt-6-sol"}
+ \\{"kind":"score","prompt_sha":"bb","score":0,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
+ \\{"kind":"score","prompt_sha":"bb","score":0,"shape":"research","role":"sweep","tier":"small","model":"gpt-6-luna"}
;
const cells = policy.foldCells(arena_state.allocator(), archive);
const small = tier_ladder.forProvider("codex").?.small.?;
- try std.testing.expectEqualStrings("gpt-5.6-luna", small);
+ try std.testing.expectEqualStrings("gpt-6-luna", small);
try std.testing.expect(policy.learnedRungIn(cells, &demo_sheet, "codex", .{ .shape = .research, .role = "sweep" }, small) == null);
}
@@ -260,34 +260,34 @@ test "#372: a learned cell reroutes a TIER pin, and an explicit model pin still
bench.g_entries = &demo_sheet;
policy.g_cells = policy.foldCells(arena_state.allocator(), demo_archive);
- const root = codex("gpt-5.6-sol");
+ const root = codex("gpt-6-astra");
const cell: Cell = .{ .shape = .research, .role = "sweep" };
// A `tier: mid` ask in this cell now resolves to the model the cell
// measured as strictly better value, and says so.
const learned = pin_mod.resolveIn(root, .{ .tier = .mid }, cell);
- try std.testing.expectEqualStrings("gpt-5.6-luna", learned.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", learned.provider.?.model);
try std.testing.expectEqual(pin_mod.Source.learned_policy, learned.source);
try std.testing.expectEqualStrings("codex", learned.provider.?.id); // provider-local, always
// The SAME ask in an unmeasured cell keeps today's ladder answer.
const unmeasured = pin_mod.resolveIn(root, .{ .tier = .mid }, .{ .shape = .feature, .role = "implement" });
- try std.testing.expectEqualStrings("gpt-5.6-terra", unmeasured.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", unmeasured.provider.?.model);
try std.testing.expectEqual(pin_mod.Source.ladder, unmeasured.source);
// AN EXPLICIT USER PIN OUTRANKS EVERYTHING LEARNED. Same cell, same
// session, but the call named a model: the learned layer is never even
// consulted, and the trace attributes the choice to the human.
- const explicit = pin_mod.resolveIn(root, .{ .model = "gpt-5.6-terra" }, cell);
- try std.testing.expectEqualStrings("gpt-5.6-terra", explicit.provider.?.model);
+ const explicit = pin_mod.resolveIn(root, .{ .model = "gpt-6-sol" }, cell);
+ try std.testing.expectEqualStrings("gpt-6-sol", explicit.provider.?.model);
try std.testing.expectEqual(pin_mod.Source.explicit_pin, explicit.source);
// A persona's exact `model:` frontmatter pin likewise wins, and is
// attributed to the persona rather than to the user's call.
- const persona = pin_mod.resolveIn(root, .{ .model = "gpt-5.6-terra", .from_persona = true }, cell);
- try std.testing.expectEqualStrings("gpt-5.6-terra", persona.provider.?.model);
+ const persona = pin_mod.resolveIn(root, .{ .model = "gpt-6-sol", .from_persona = true }, cell);
+ try std.testing.expectEqualStrings("gpt-6-sol", persona.provider.?.model);
try std.testing.expectEqual(pin_mod.Source.persona, persona.source);
// And the cell-less entry point (every pre-#372 caller) is unchanged.
- try std.testing.expectEqualStrings("gpt-5.6-terra", pin_mod.resolve(root, .{ .tier = .mid }).provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", pin_mod.resolve(root, .{ .tier = .mid }).provider.?.model);
}
test "#372 round trip: what captureVariant writes is exactly what foldCells learns from" {
@@ -305,8 +305,8 @@ test "#372 round trip: what captureVariant writes is exactly what foldCells lear
// Three judged `sweep` variants on the mid rung, three on the small one —
// the capture the scoring path makes, with no hand-written JSON anywhere.
- for (0..3) |_| route_trace.captureVariant(codex("gpt-5.6-terra"), .research, "sweep the repo", "sweep", "aa", 0.2);
- for (0..3) |_| route_trace.captureVariant(codex("gpt-5.6-luna"), .research, "sweep the repo", "sweep", "bb", 0.95);
+ for (0..3) |_| route_trace.captureVariant(codex("gpt-6-sol"), .research, "sweep the repo", "sweep", "aa", 0.2);
+ for (0..3) |_| route_trace.captureVariant(codex("gpt-6-luna"), .research, "sweep the repo", "sweep", "bb", 0.95);
const cells = policy.foldCells(a, out.writer.buffered());
try std.testing.expectEqual(@as(usize, 2), cells.len);
@@ -315,12 +315,12 @@ test "#372 round trip: what captureVariant writes is exactly what foldCells lear
try std.testing.expectEqualStrings("research", c.shape);
try std.testing.expectEqualStrings("sweep", c.role);
try std.testing.expectEqual(@as(u32, 3), c.n);
- try std.testing.expectEqualStrings(if (std.mem.eql(u8, c.model, "gpt-5.6-terra")) "mid" else "small", c.tier);
+ try std.testing.expectEqualStrings(if (std.mem.eql(u8, c.model, "gpt-6-sol")) "mid" else "small", c.tier);
}
// ...and the policy reaches the same verdict as the hand-written archive.
try std.testing.expectEqualStrings(
- "gpt-5.6-luna",
- policy.learnedRungIn(cells, &demo_sheet, "codex", .{ .shape = .research, .role = "sweep" }, "gpt-5.6-terra").?,
+ "gpt-6-luna",
+ policy.learnedRungIn(cells, &demo_sheet, "codex", .{ .shape = .research, .role = "sweep" }, "gpt-6-sol").?,
);
}
diff --git a/src/router_catalog_tests.zig b/src/router_catalog_tests.zig
index 85423252..c192531c 100644
--- a/src/router_catalog_tests.zig
+++ b/src/router_catalog_tests.zig
@@ -424,7 +424,7 @@ test "router discovery replaces only its provider slice" {
try std.testing.expect(pricing.providerModelInTable("codegraff", "kimi-k3"));
try std.testing.expect(!pricing.providerModelInTable("codegraff", "claude-opus-4.8"));
try std.testing.expect(pricing.providerModelInTable("anthropic", "claude-opus-4-8"));
- try std.testing.expect(pricing.providerModelInTable("codex", "gpt-5.6-sol"));
+ try std.testing.expect(pricing.providerModelInTable("codex", "gpt-6-sol"));
}
test "startup catalog loads fan out concurrently with Kimi" {
@@ -450,7 +450,7 @@ test "providerFor (#377): family-prefixed spelling routes to the direct provider
try std.testing.expectEqualStrings("kimi", p.id);
try std.testing.expectEqualStrings("k3", p.model);
// Exact catalog names keep absolute priority over the family alias.
- try std.testing.expectEqualStrings("gpt-5.6-sol", (try all.providerFor("gpt-5.6-sol")).model);
+ try std.testing.expectEqualStrings("gpt-6-sol", (try all.providerFor("gpt-6-sol")).model);
// Without the kimi credential, Moonshot's platform id `kimi-k3` is an
// exact catalog row — not the gateway.
var values: [provider.provider_specs.len]?[]const u8 = @splat("k");
@@ -499,11 +499,11 @@ test "xAI defaults to the Responses wire; GRAFF_XAI_WIRE=chat opts out (#502)" {
try std.testing.expectEqual(Provider.Kind.responses, moved.kind);
try std.testing.expectEqualStrings(provider.xai_responses_url, moved.url);
// codex keeps in-stream compaction — never the explicit endpoint.
- const codex = try all.providerById("codex", "gpt-5.6-sol");
+ const codex = try all.providerById("codex", "gpt-6-sol");
try std.testing.expect(codex.serverCompactUrl() == null);
// every other provider is untouched by the knob: openai keeps its native
// Responses wire (default since v0.0.250) and its own endpoint.
- const oai = try all.providerById("openai", "gpt-5.6");
+ const oai = try all.providerById("openai", "gpt-6-sol");
try std.testing.expectEqual(Provider.Kind.responses, oai.kind);
try std.testing.expect(std.mem.indexOf(u8, oai.url, "api.openai.com") != null);
}
diff --git a/src/schema_agents.zig b/src/schema_agents.zig
index 8ba417a2..7d48bebf 100644
--- a/src/schema_agents.zig
+++ b/src/schema_agents.zig
@@ -5,7 +5,7 @@ pub const subagent_spec = ToolSpec{
.name = "subagent",
.desc = "Delegate a sidecar task; children have file, shell, codedb and skill tools, no parent context, and cannot spawn children. Brief once with context, prompt, established_facts, scope and deliverable. Keep critical-path work local; use disjoint write sets for parallel tasks. Do not set model without a user request or task-specific reason. agent selects a persona; system_prompt overrides it. retained=true preserves history and workspace for follow-up; otherwise the worker is ephemeral. Interactive calls run in background by default; run_in_background=false waits for the child's report on every surface. Headless calls wait by default. agent_message queues feedback without starting a turn; subagent_resume continues a completed worker. agent_output fetches background results; wait_ms>0 waits in headless mode. Do not poll.",
.schema =
- \\{"type": "object", "properties": {"description": {"type": "string", "description": "Short label for logs, 3-5 words"}, "prompt": {"type": "string", "description": "Numbered steps and search hints; the complete task. Optional context/established_facts/scope/deliverable render as headed sections around this"}, "context": {"type": "string", "description": "Optional: repo and hard constraints the child must not re-derive"}, "established_facts": {"type": "string", "description": "Optional: observed facts, paths, and line numbers the child must treat as given"}, "scope": {"type": "string", "description": "Optional: what not to touch or broaden into"}, "deliverable": {"type": "string", "description": "Optional: required report shape; default is files changed, verified, skipped, open questions"}, "agent": {"type": "string", "description": "Optional: a named agent type (reviewer, researcher, implementer, skeptic, or a .harness/agents/ name) whose persona the child runs with"}, "system_prompt": {"type": "string", "description": "Optional: replace the child's system prompt with a custom one (overrides agent)"}, "model": {"type": "string", "description": "Optional: run THIS child on an exact model name available on the current provider (e.g. \"gpt-5.6-luna\"). Highest precedence — beats the persona's own pin and the session default. Ignored (spawn still runs on the session default, with a trace note) if the name is unknown, ambiguous, or not served by this provider — except that a logged-in flat-rate subscription serving the exact name honors it (sub-first routing; the login is the standing consent). For a portable 'cheapest seat' ask, tier is still the better lever."}, "tier": {"type": "string", "enum": ["frontier", "mid", "small"], "description": "Optional: run THIS child on a rung of the current provider's model ladder instead of an exact name — frontier (planner/judge class), mid (reviewers, implementers, researchers), small (mechanical transform/extract/label). Prefer this over model: it stays correct when the root model changes. If the provider has no ladder or no model at that rung, sub-first routing seats the child on a logged-in flat-rate subscription's rung when one exists (no login → the pin is ignored); model wins if both are given."}, "effort": {"type": "string", "enum": ["low", "medium", "high", "xhigh", "max", "none"], "description": "Optional: reasoning effort for THIS child — beats the persona's effort: frontmatter; unpinned children run at medium. An axis independent of model/tier: an effort-only override keeps the persona's model pin. none disables thinking only on supported binary-thinking routes and is rejected before child inference elsewhere. An off-vocabulary value is ignored with a trace note."}, "isolation": {"type": "string", "enum": ["shared_cwd", "worktree"], "description": "Optional: independent / fan-out children default to \"worktree\" (own branch + checkout; clean unchanged trees are removed unless retained=true; dirty or uniquely committed trees are kept + reported). Use \"shared_cwd\" only for same-branch collaboration (review + fix, a pipeline). isolation_fallback:true is the only way a failed git worktree add may share the caller tree."}, "isolation_fallback": {"type": "boolean", "description": "Optional: if isolation:\"worktree\" fails to set up (not a git repo, git error), run in the shared working tree instead of failing the spawn. Default false (fails the spawn on setup failure)."}, "retained": {"type": "boolean", "description": "Optional: persist conversation and workspace for subagent_resume (default false)."}, "run_in_background": {"type": "boolean", "description": "Optional: true starts asynchronously and immediately returns an agent id; false explicitly waits for the child report on every surface. Defaults to true in interactive sessions and false in headless sessions. Do not poll; agent_output(wait_ms>0) blocks until it finishes. A spawn at the live concurrency cap is refused; do not spawn a child to clean up a finished worktree."}}, "required": ["description", "prompt"]}
+ \\{"type": "object", "properties": {"description": {"type": "string", "description": "Short label for logs, 3-5 words"}, "prompt": {"type": "string", "description": "Numbered steps and search hints; the complete task. Optional context/established_facts/scope/deliverable render as headed sections around this"}, "context": {"type": "string", "description": "Optional: repo and hard constraints the child must not re-derive"}, "established_facts": {"type": "string", "description": "Optional: observed facts, paths, and line numbers the child must treat as given"}, "scope": {"type": "string", "description": "Optional: what not to touch or broaden into"}, "deliverable": {"type": "string", "description": "Optional: required report shape; default is files changed, verified, skipped, open questions"}, "agent": {"type": "string", "description": "Optional: a named agent type (reviewer, researcher, implementer, skeptic, or a .harness/agents/ name) whose persona the child runs with"}, "system_prompt": {"type": "string", "description": "Optional: replace the child's system prompt with a custom one (overrides agent)"}, "model": {"type": "string", "description": "Optional: run THIS child on an exact model name available on the current provider (e.g. \"gpt-6-luna\"). Highest precedence — beats the persona's own pin and the session default. Ignored (spawn still runs on the session default, with a trace note) if the name is unknown, ambiguous, or not served by this provider — except that a logged-in flat-rate subscription serving the exact name honors it (sub-first routing; the login is the standing consent). For a portable 'cheapest seat' ask, tier is still the better lever."}, "tier": {"type": "string", "enum": ["frontier", "mid", "small"], "description": "Optional: run THIS child on a rung of the current provider's model ladder instead of an exact name — frontier (planner/judge class), mid (reviewers, implementers, researchers), small (mechanical transform/extract/label). Prefer this over model: it stays correct when the root model changes. If the provider has no ladder or no model at that rung, sub-first routing seats the child on a logged-in flat-rate subscription's rung when one exists (no login → the pin is ignored); model wins if both are given."}, "effort": {"type": "string", "enum": ["low", "medium", "high", "xhigh", "max", "none"], "description": "Optional: reasoning effort for THIS child — beats the persona's effort: frontmatter; unpinned children run at medium. An axis independent of model/tier: an effort-only override keeps the persona's model pin. none disables thinking only on supported binary-thinking routes and is rejected before child inference elsewhere. An off-vocabulary value is ignored with a trace note."}, "isolation": {"type": "string", "enum": ["shared_cwd", "worktree"], "description": "Optional: independent / fan-out children default to \"worktree\" (own branch + checkout; clean unchanged trees are removed unless retained=true; dirty or uniquely committed trees are kept + reported). Use \"shared_cwd\" only for same-branch collaboration (review + fix, a pipeline). isolation_fallback:true is the only way a failed git worktree add may share the caller tree."}, "isolation_fallback": {"type": "boolean", "description": "Optional: if isolation:\"worktree\" fails to set up (not a git repo, git error), run in the shared working tree instead of failing the spawn. Default false (fails the spawn on setup failure)."}, "retained": {"type": "boolean", "description": "Optional: persist conversation and workspace for subagent_resume (default false)."}, "run_in_background": {"type": "boolean", "description": "Optional: true starts asynchronously and immediately returns an agent id; false explicitly waits for the child report on every surface. Defaults to true in interactive sessions and false in headless sessions. Do not poll; agent_output(wait_ms>0) blocks until it finishes. A spawn at the live concurrency cap is refused; do not spawn a child to clean up a finished worktree."}}, "required": ["description", "prompt"]}
,
};
diff --git a/src/schema_protocol.zig b/src/schema_protocol.zig
index eb822f7a..f5d37177 100644
--- a/src/schema_protocol.zig
+++ b/src/schema_protocol.zig
@@ -37,7 +37,7 @@ pub const flags =
pub const json =
\\{
\\ "transport": "newline-delimited JSON over stdin/stdout (--json)",
- \\ "request": "one JSON object per line: {\"type\":\"user\",\"text\":\"...\",\"images\":[{\"type\":\"image_url\",\"url\":\"https://...\"},{\"type\":\"image_base64\",\"media_type\":\"image/png\",\"data\":\"...\"}]} sends a native multimodal user turn (images optional, max 16). Only user input may set this field; {\"type\":\"review\",\"text\":\"target or instructions\"} runs one isolated read-only review turn with fresh model-visible history and no edits, delegation, workflows, or web access; {\"type\":\"answer\",\"text\":\"...\",\"cancelled\":false,\"call_id\":\"optional\"} answers an active ask_user event; {\"type\":\"cancel\"} interrupts an active turn; {\"type\":\"set_system_prompt\",\"text\":\"...\",\"append\":false} replaces (or with append=true extends) the system prompt between turns and acks with a system_prompt event; {\"type\":\"set_model\",\"model\":\"gpt-5.5\",\"provider\":\"codex\"}, {\"type\":\"compact\"}, {\"type\":\"set_mode\",\"mode\":\"plan|normal\"}, and {\"type\":\"set_agent\",\"id\":\"reviewer\"}, {\"type\":\"set_effort\",\"level\":\"low|medium|high|xhigh|max|ultra\"}, {\"type\":\"set_fast\",\"on\":true}, and {\"type\":\"set_ultracode\",\"on\":true} are live control requests acked by model/compact/mode/agent/effort/fast/ultracode events. For compatibility, set_model also accepts legacy {\"name\":\"provider model|model\"}. NOTE: the system prompt heads the KV-cached prefix, so any mutation invalidates the cache for the whole conversation (per Manus context-engineering lessons) — set it at spawn when possible and mutate only at task boundaries",
+ \\ "request": "one JSON object per line: {\"type\":\"user\",\"text\":\"...\",\"images\":[{\"type\":\"image_url\",\"url\":\"https://...\"},{\"type\":\"image_base64\",\"media_type\":\"image/png\",\"data\":\"...\"}]} sends a native multimodal user turn (images optional, max 16). Only user input may set this field; {\"type\":\"review\",\"text\":\"target or instructions\"} runs one isolated read-only review turn with fresh model-visible history and no edits, delegation, workflows, or web access; {\"type\":\"answer\",\"text\":\"...\",\"cancelled\":false,\"call_id\":\"optional\"} answers an active ask_user event; {\"type\":\"cancel\"} interrupts an active turn; {\"type\":\"set_system_prompt\",\"text\":\"...\",\"append\":false} replaces (or with append=true extends) the system prompt between turns and acks with a system_prompt event; {\"type\":\"set_model\",\"model\":\"gpt-6-sol\",\"provider\":\"codex\"}, {\"type\":\"compact\"}, {\"type\":\"set_mode\",\"mode\":\"plan|normal\"}, and {\"type\":\"set_agent\",\"id\":\"reviewer\"}, {\"type\":\"set_effort\",\"level\":\"low|medium|high|xhigh|max|ultra\"}, {\"type\":\"set_fast\",\"on\":true}, and {\"type\":\"set_ultracode\",\"on\":true} are live control requests acked by model/compact/mode/agent/effort/fast/ultracode events. For compatibility, set_model also accepts legacy {\"name\":\"provider model|model\"}. NOTE: the system prompt heads the KV-cached prefix, so any mutation invalidates the cache for the whole conversation (per Manus context-engineering lessons) — set it at spawn when possible and mutate only at task boundaries",
\\ "score_request": "{\"type\":\"score\",\"prompt_sha\":\"<16 hex>\",\"score\":0.7,\"notes\":\"...\",\"parent_sha\":\"<16 hex, optional>\"} appends an evaluation record for an agent/prompt variant to this run's file under .graff/trajectories (the aggregate append-only DGM-style archive; prompt_sha = first 8 bytes of SHA-256 of the system prompt, hex; parent_sha records which prompt this variant was mutated from — the lineage edge DGM parent selection counts children with) and acks with a score event",
\\ "sequencing": "every event object below carries a monotonic per-session \"seq\" as its FIRST field: 1-based, gap-free, one id per emitted line. It survives a resume - the counter is persisted with the conversation as event_seq and restored by --resume, so a REPLACEMENT process continues the numbering instead of reissuing ids a consumer already saw. `graff serve` forwards the child's seq unchanged and uses it for ?from=N replay",
\\ "events": [
diff --git a/src/scoring.zig b/src/scoring.zig
index 3ca4c9f1..68a3babd 100644
--- a/src/scoring.zig
+++ b/src/scoring.zig
@@ -39,10 +39,12 @@ pub fn providerClass(model: []const u8) []const u8 {
.{ .needle = "-mini", .tier = "small" }, // -mini suffix, not "minimax"/"gemini"
.{ .needle = "lite", .tier = "small" },
.{ .needle = "nano", .tier = "small" },
- // gpt-5.x rungs (#291): family-agnostic like "haiku"/"flash" so a
- // future gpt-5.7-terra keeps its tier; must precede "gpt-5".
+ // GPT rungs (#291): family-agnostic like "haiku"/"flash" so a later
+ // generation keeps its tier; must precede "gpt-6"/"gpt-5". Sol is the
+ // mid rung under Astra on the GPT-6 ladder.
.{ .needle = "terra", .tier = "mid" },
.{ .needle = "luna", .tier = "small" },
+ .{ .needle = "-sol", .tier = "mid" },
.{ .needle = "opus", .tier = "frontier" },
.{ .needle = "gpt-6", .tier = "frontier" }, // astra and later GPT-6 flagships
.{ .needle = "gpt-5", .tier = "frontier" }, // sol and bare gpt-5.x stay frontier
@@ -261,10 +263,10 @@ test "score signing message is canonical and cross-language stable" {
test "providerClass buckets models by capability tier" {
try std.testing.expectEqualStrings("frontier", providerClass("claude-opus-4-8"));
try std.testing.expectEqualStrings("frontier", providerClass("gpt-5.5"));
- try std.testing.expectEqualStrings("frontier", providerClass("gpt-5.6-sol"));
try std.testing.expectEqualStrings("frontier", providerClass("gpt-6-astra"));
- try std.testing.expectEqualStrings("mid", providerClass("gpt-5.6-terra")); // rung needle beats "gpt-5"
- try std.testing.expectEqualStrings("small", providerClass("gpt-5.6-luna"));
+ try std.testing.expectEqualStrings("mid", providerClass("gpt-6-sol")); // rung needle beats "gpt-6"
+ try std.testing.expectEqualStrings("mid", providerClass("gpt-6.1-sol"));
+ try std.testing.expectEqualStrings("small", providerClass("gpt-6-luna"));
try std.testing.expectEqualStrings("frontier", providerClass("deepseek-v4-pro"));
try std.testing.expectEqualStrings("frontier", providerClass("grok-4.3"));
try std.testing.expectEqualStrings("small", providerClass("claude-haiku-4-5"));
diff --git a/src/subagent_mimo_tests.zig b/src/subagent_mimo_tests.zig
index b3b5bfdd..eb23b326 100644
--- a/src/subagent_mimo_tests.zig
+++ b/src/subagent_mimo_tests.zig
@@ -35,8 +35,8 @@ test "MiMo small workers stay on Flash despite a Luna subscription" {
const inherited = selection.resolveSubagentProvider(keys, base, null, null, false, false).?;
try std.testing.expectEqualStrings("mimo-v2.6-flash", inherited.model);
// An exact human pin still wins; this is a default, not a model ban.
- const exact = pin.forSpawn(base, request(arena.allocator(), "{\"model\":\"gpt-5.6-luna\"}"), true);
- try std.testing.expectEqualStrings("gpt-5.6-luna", exact.provider.?.model);
+ const exact = pin.forSpawn(base, request(arena.allocator(), "{\"model\":\"gpt-6-luna\"}"), true);
+ try std.testing.expectEqualStrings("gpt-6-luna", exact.provider.?.model);
try std.testing.expect(selection.resolveSubagentProvider(keys, base.withModel("mimo-v2.6-flash"), null, null, false, false) == null);
}
@@ -120,7 +120,7 @@ test "MiMo routing does not cross to a metered provider without an explicit requ
defer bench.g_keys = saved_keys;
var keys: provider.Keys = .{ .values = @splat("test-key") };
bench.g_keys = &keys;
- const base = try keys.providerById("codex", "gpt-5.6-sol");
+ const base = try keys.providerById("codex", "gpt-6-sol");
try std.testing.expectError(error.CrossProviderConsentRequired, selection.subagentProvider(keys, base, "xiaomi", "mimo-v2.6-flash", false));
const selected = try selection.subagentProvider(keys, base, "xiaomi", "mimo-v2.6-flash", true);
try std.testing.expectEqualStrings("xiaomi", selected.id);
diff --git a/src/subagent_pin_tests.zig b/src/subagent_pin_tests.zig
index b222e02f..5c1801f3 100644
--- a/src/subagent_pin_tests.zig
+++ b/src/subagent_pin_tests.zig
@@ -21,7 +21,7 @@ fn obj(a: Allocator, s: []const u8) std.json.ObjectMap {
/// credentials; withModel only ever reads `id`/`api_key`/`account`/`source`
/// off it.
fn codexBase() Provider {
- return .{ .id = "codex", .kind = .responses, .auth = .bearer, .url = "https://example.invalid", .api_key = "k", .model = "gpt-5.6-sol", .context = 272_000 };
+ return .{ .id = "codex", .kind = .responses, .auth = .bearer, .url = "https://example.invalid", .api_key = "k", .model = "gpt-6-astra", .context = 272_000 };
}
test "persona frontmatter parse: model/tier land on AgentType, a bad tier is dropped" {
@@ -37,14 +37,14 @@ test "persona frontmatter parse: model/tier land on AgentType, a bad tier is dro
defer fleet.g_agent_types = saved;
fleet.g_agent_types = &.{
.{ .name = "implementer", .desc = "", .prompt = "x", .tier = Tier.parse("mid") },
- .{ .name = "reviewer", .desc = "", .prompt = "x", .model = "gpt-5.6-terra" },
+ .{ .name = "reviewer", .desc = "", .prompt = "x", .model = "gpt-6-sol" },
.{ .name = "typo", .desc = "", .prompt = "x", .tier = Tier.parse("cheap") }, // off-vocabulary
.{ .name = "plain", .desc = "", .prompt = "x" },
};
try std.testing.expectEqual(Tier.mid, pin_mod.personaPin("implementer").tier.?);
try std.testing.expect(pin_mod.personaPin("implementer").model == null);
- try std.testing.expectEqualStrings("gpt-5.6-terra", pin_mod.personaPin("reviewer").model.?);
+ try std.testing.expectEqualStrings("gpt-6-sol", pin_mod.personaPin("reviewer").model.?);
// A persona whose `tier:` did not parse contributes nothing at all — the
// spawn keeps the session default rather than inheriting a guess.
try std.testing.expect(pin_mod.personaPin("typo").isNone());
@@ -54,8 +54,8 @@ test "persona frontmatter parse: model/tier land on AgentType, a bad tier is dro
try std.testing.expect(pin_mod.personaPin("").isNone());
// …and the pin reaches a spawn that only names the persona.
- try std.testing.expectEqualStrings("gpt-5.6-terra", pin_mod.forSpawn(codexBase(), obj(a, "{\"agent\":\"implementer\"}"), true).provider.?.model);
- try std.testing.expectEqualStrings("gpt-5.6-terra", pin_mod.forSpawn(codexBase(), obj(a, "{\"agent\":\"reviewer\"}"), true).provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", pin_mod.forSpawn(codexBase(), obj(a, "{\"agent\":\"implementer\"}"), true).provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", pin_mod.forSpawn(codexBase(), obj(a, "{\"agent\":\"reviewer\"}"), true).provider.?.model);
try std.testing.expect(pin_mod.forSpawn(codexBase(), obj(a, "{\"agent\":\"typo\"}"), true).provider == null);
try std.testing.expect(pin_mod.forSpawn(codexBase(), obj(a, "{\"agent\":\"plain\"}"), true).provider == null);
}
@@ -72,12 +72,12 @@ test "researcher is the small/luna search seat on Codex, including Astra roots"
try std.testing.expectEqual(Tier.small, pin_mod.personaPin("researcher").tier.?);
const sol = pin_mod.forSpawn(codexBase(), obj(a, "{\"agent\":\"researcher\"}"), true);
- try std.testing.expectEqualStrings("gpt-5.6-luna", sol.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", sol.provider.?.model);
var astra = codexBase();
astra.model = "gpt-6-astra";
const from_astra = pin_mod.forSpawn(astra, obj(a, "{\"agent\":\"researcher\"}"), true);
- try std.testing.expectEqualStrings("gpt-5.6-luna", from_astra.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", from_astra.provider.?.model);
}
test "spawn override resolution: exact name, alias, ladder rung, and no-op" {
@@ -87,29 +87,29 @@ test "spawn override resolution: exact name, alias, ladder rung, and no-op" {
const base = codexBase();
// Exact catalog name.
- const exact = pin_mod.forSpawn(base, obj(a, "{\"model\":\"gpt-5.6-luna\"}"), true);
+ const exact = pin_mod.forSpawn(base, obj(a, "{\"model\":\"gpt-6-luna\"}"), true);
try std.testing.expectEqual(pin_mod.Outcome.pinned, exact.outcome);
- try std.testing.expectEqualStrings("gpt-5.6-luna", exact.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", exact.provider.?.model);
try std.testing.expectEqualStrings("codex", exact.provider.?.id); // provider-local, always
try std.testing.expectEqualStrings("k", exact.provider.?.api_key); // same credential
// The context window is re-derived for the new model, not inherited.
try std.testing.expect(exact.provider.?.context != 0);
// Abbreviation: the same resolver --subagent-model uses (unique substring).
- const alias = pin_mod.forSpawn(base, obj(a, "{\"model\":\" 5.6-terra \"}"), true); // whitespace trimmed
- try std.testing.expectEqualStrings("gpt-5.6-terra", alias.provider.?.model);
+ const alias = pin_mod.forSpawn(base, obj(a, "{\"model\":\" 6.1-sol \"}"), true); // whitespace trimmed
+ try std.testing.expectEqualStrings("gpt-6.1-sol", alias.provider.?.model);
// Ladder rung instead of a name.
const rung = pin_mod.forSpawn(base, obj(a, "{\"tier\":\"small\"}"), true);
try std.testing.expectEqual(pin_mod.Outcome.pinned, rung.outcome);
- try std.testing.expectEqualStrings("gpt-5.6-luna", rung.provider.?.model);
- try std.testing.expectEqualStrings("gpt-5.6-terra", pin_mod.forSpawn(base, obj(a, "{\"tier\":\"mid\"}"), true).provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", rung.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-sol", pin_mod.forSpawn(base, obj(a, "{\"tier\":\"mid\"}"), true).provider.?.model);
- // Codex frontier is Astra; a Sol child asking for frontier is a real pin.
+ // Codex frontier is Astra: an Astra root asking for frontier, or naming
+ // its own model, is a no-op.
const flagship = pin_mod.forSpawn(base, obj(a, "{\"tier\":\"frontier\"}"), true);
- try std.testing.expectEqual(pin_mod.Outcome.pinned, flagship.outcome);
- try std.testing.expectEqualStrings("gpt-6-astra", flagship.provider.?.model);
- const same = pin_mod.forSpawn(base, obj(a, "{\"model\":\"gpt-5.6-sol\"}"), true);
+ try std.testing.expectEqual(pin_mod.Outcome.same, flagship.outcome);
+ const same = pin_mod.forSpawn(base, obj(a, "{\"model\":\"gpt-6-astra\"}"), true);
try std.testing.expectEqual(pin_mod.Outcome.same, same.outcome);
try std.testing.expect(same.provider == null);
@@ -127,36 +127,36 @@ test "precedence: spawn param > persona pin > session default" {
const saved = fleet.g_agent_types;
defer fleet.g_agent_types = saved;
fleet.g_agent_types = &.{
- .{ .name = "pinned-frontier", .desc = "", .prompt = "x", .model = "gpt-5.6-sol" },
+ .{ .name = "pinned-frontier", .desc = "", .prompt = "x", .model = "gpt-6-astra" },
.{ .name = "pinned-mid", .desc = "", .prompt = "x", .tier = .mid },
- .{ .name = "both", .desc = "", .prompt = "x", .model = "gpt-5.6-terra", .tier = .small },
+ .{ .name = "both", .desc = "", .prompt = "x", .model = "gpt-6-sol", .tier = .small },
};
// The session default is whatever the caller already resolved — here the
// #291 ladder's answer for a sol root, i.e. terra.
- const session_default: Provider = .{ .id = "codex", .kind = .responses, .auth = .bearer, .url = "", .api_key = "k", .model = "gpt-5.6-terra", .context = 272_000 };
+ const session_default: Provider = .{ .id = "codex", .kind = .responses, .auth = .bearer, .url = "", .api_key = "k", .model = "gpt-6-sol", .context = 272_000 };
// 1. An explicit spawn param beats the persona, in BOTH directions —
// including lowering a persona that pinned itself to the flagship,
// which is the case a merge-instead-of-replace rule would break.
try std.testing.expectEqualStrings(
- "gpt-5.6-luna",
+ "gpt-6-luna",
pin_mod.forSpawn(session_default, obj(a, "{\"agent\":\"pinned-frontier\",\"tier\":\"small\"}"), true).provider.?.model,
);
try std.testing.expectEqualStrings(
- "gpt-5.6-sol",
- pin_mod.forSpawn(session_default, obj(a, "{\"agent\":\"pinned-mid\",\"model\":\"gpt-5.6-sol\"}"), true).provider.?.model,
+ "gpt-6-astra",
+ pin_mod.forSpawn(session_default, obj(a, "{\"agent\":\"pinned-mid\",\"model\":\"gpt-6-astra\"}"), true).provider.?.model,
);
// 2. No spawn param → the persona's pin applies over the session default.
try std.testing.expectEqualStrings(
- "gpt-5.6-sol",
+ "gpt-6-astra",
pin_mod.forSpawn(session_default, obj(a, "{\"agent\":\"pinned-frontier\"}"), true).provider.?.model,
);
// 3. Neither → the session default is kept untouched.
try std.testing.expect(pin_mod.forSpawn(session_default, obj(a, "{\"prompt\":\"y\"}"), true).provider == null);
// Within one level, an exact model beats a tier (both on the call…
try std.testing.expectEqualStrings(
- "gpt-5.6-sol",
- pin_mod.forSpawn(session_default, obj(a, "{\"model\":\"gpt-5.6-sol\",\"tier\":\"small\"}"), true).provider.?.model,
+ "gpt-6-astra",
+ pin_mod.forSpawn(session_default, obj(a, "{\"model\":\"gpt-6-astra\",\"tier\":\"small\"}"), true).provider.?.model,
);
// …and both on the persona).
try std.testing.expectEqual(pin_mod.Outcome.same, pin_mod.forSpawn(session_default, obj(a, "{\"agent\":\"both\"}"), true).outcome);
@@ -254,14 +254,14 @@ test "sub-first routing: a logged-in flat-rate sub outranks metered for explicit
// does not serve falls through to a logged-in sub that serves it exactly
// — the same standing consent as the tier path, rescuing a pin that used
// to silently no-op.
- const mrouted = pin_mod.forSpawn(dsv, obj(a, "{\"model\":\"gpt-5.6-luna\"}"), true);
+ const mrouted = pin_mod.forSpawn(dsv, obj(a, "{\"model\":\"gpt-6-luna\"}"), true);
try std.testing.expectEqual(pin_mod.Outcome.model_sub_routed, mrouted.outcome);
try std.testing.expectEqualStrings("codex", mrouted.provider.?.id);
- try std.testing.expectEqualStrings("gpt-5.6-luna", mrouted.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", mrouted.provider.?.model);
try std.testing.expect(mrouted.outcome.describe().len > 0);
// …under the identical gates: an explicit --subagent-provider
// (sub_ok=false) forbids the hop…
- try std.testing.expectEqual(pin_mod.Outcome.unknown_model, pin_mod.forSpawn(dsv, obj(a, "{\"model\":\"gpt-5.6-luna\"}"), false).outcome);
+ try std.testing.expectEqual(pin_mod.Outcome.unknown_model, pin_mod.forSpawn(dsv, obj(a, "{\"model\":\"gpt-6-luna\"}"), false).outcome);
// …and the match is strictly exact/alias — a substring must NOT cross a
// provider boundary to find a model the pin did not name.
try std.testing.expectEqual(pin_mod.Outcome.unknown_model, pin_mod.forSpawn(dsv, obj(a, "{\"model\":\"luna\"}"), true).outcome);
@@ -271,7 +271,7 @@ test "sub-first routing: a logged-in flat-rate sub outranks metered for explicit
const nologin = pin_mod.forSpawn(dsv, obj(a, "{\"tier\":\"small\"}"), true);
try std.testing.expectEqual(pin_mod.Outcome.pinned, nologin.outcome);
try std.testing.expectEqualStrings("deepseek-v4-flash", nologin.provider.?.model);
- try std.testing.expectEqual(pin_mod.Outcome.unknown_model, pin_mod.forSpawn(dsv, obj(a, "{\"model\":\"gpt-5.6-luna\"}"), true).outcome);
+ try std.testing.expectEqual(pin_mod.Outcome.unknown_model, pin_mod.forSpawn(dsv, obj(a, "{\"model\":\"gpt-6-luna\"}"), true).outcome);
// Same family through a codegraff login: stay on gateway flash, not luna.
bench.g_keys = &keys;
@@ -294,14 +294,14 @@ test "exact pin: the child's own provider wins over a logged-in sub serving the
if (std.mem.eql(u8, spec.id, "codex")) keys.values[i] = "tok";
}
bench.g_keys = &keys;
- // openai's own catalog serves gpt-5.6-luna — the pin resolves locally
+ // openai's own catalog serves gpt-6-luna — the pin resolves locally
// (.pinned), never hopping to the codex sub that serves the same name:
// a locally-served pin crosses no boundary and moves no billing.
- const oai: Provider = .{ .id = "openai", .kind = .openai, .auth = .bearer, .url = "", .api_key = "k", .model = "gpt-5.6", .context = 1_050_000 };
- const got = pin_mod.forSpawn(oai, obj(a, "{\"model\":\"gpt-5.6-luna\"}"), true);
+ const oai: Provider = .{ .id = "openai", .kind = .openai, .auth = .bearer, .url = "", .api_key = "k", .model = "gpt-6-astra", .context = 1_050_000 };
+ const got = pin_mod.forSpawn(oai, obj(a, "{\"model\":\"gpt-6-luna\"}"), true);
try std.testing.expectEqual(pin_mod.Outcome.pinned, got.outcome);
try std.testing.expectEqualStrings("openai", got.provider.?.id);
- try std.testing.expectEqualStrings("gpt-5.6-luna", got.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", got.provider.?.model);
}
test "two logins split the tiers: k3 is mid, luna is the mechanical rung (#471)" {
@@ -328,7 +328,7 @@ test "two logins split the tiers: k3 is mid, luna is the mechanical rung (#471)"
const want = [_]struct { tier: []const u8, pid: []const u8, model: []const u8 }{
.{ .tier = "frontier", .pid = "codex", .model = "gpt-6-astra" },
.{ .tier = "mid", .pid = "kimi", .model = "k3" },
- .{ .tier = "small", .pid = "codex", .model = "gpt-5.6-luna" },
+ .{ .tier = "small", .pid = "codex", .model = "gpt-6-luna" },
};
for (want) |w| {
const body = std.fmt.allocPrint(a, "{{\"tier\":\"{s}\"}}", .{w.tier}) catch unreachable;
@@ -427,7 +427,7 @@ test "without a login, a pin never crosses a provider boundary" {
.{ .id = "deepseek", .kind = .openai, .auth = .bearer, .url = "", .api_key = "k", .model = "deepseek-v4-pro", .context = 128_000 },
};
for (bases) |base| {
- for ([_][]const u8{ "{\"tier\":\"mid\"}", "{\"tier\":\"small\"}", "{\"model\":\"gpt-5.6-luna\"}", "{\"model\":\"claude-haiku-4-5\"}" }) |args| {
+ for ([_][]const u8{ "{\"tier\":\"mid\"}", "{\"tier\":\"small\"}", "{\"model\":\"gpt-6-luna\"}", "{\"model\":\"claude-haiku-4-5\"}" }) |args| {
const got = pin_mod.forSpawn(base, obj(a, args), true);
const p = got.provider orelse continue;
try std.testing.expectEqualStrings(base.id, p.id);
@@ -445,8 +445,8 @@ test "promotion carries a persona's model/tier policy, and it round-trips" {
defer fleet.g_agent_types = saved;
fleet.g_agent_types = &.{
.{ .name = "worker", .desc = "", .prompt = "old genome", .isolation = .worktree, .tier = .mid },
- .{ .name = "exact", .desc = "", .prompt = "old genome", .model = "gpt-5.6-luna" },
- .{ .name = "luna-max", .desc = "", .prompt = "old genome", .model = "gpt-5.6-luna", .effort = .max },
+ .{ .name = "exact", .desc = "", .prompt = "old genome", .model = "gpt-6-luna" },
+ .{ .name = "luna-max", .desc = "", .prompt = "old genome", .model = "gpt-6-luna", .effort = .max },
.{ .name = "bare", .desc = "", .prompt = "old genome" },
};
@@ -457,10 +457,10 @@ test "promotion carries a persona's model/tier policy, and it round-trips" {
const policy = pin_mod.personaPolicyFrontmatter(a, "worker");
try std.testing.expectEqualStrings("isolation: worktree\ntier: mid\n", policy);
- try std.testing.expectEqualStrings("model: gpt-5.6-luna\n", pin_mod.personaPolicyFrontmatter(a, "exact"));
+ try std.testing.expectEqualStrings("model: gpt-6-luna\n", pin_mod.personaPolicyFrontmatter(a, "exact"));
// #292 follow-up: the effort pin is policy too — promote must not un-pin
// it, and the emitted line round-trips through the loader's parse.
- try std.testing.expectEqualStrings("model: gpt-5.6-luna\neffort: max\n", pin_mod.personaPolicyFrontmatter(a, "luna-max"));
+ try std.testing.expectEqualStrings("model: gpt-6-luna\neffort: max\n", pin_mod.personaPolicyFrontmatter(a, "luna-max"));
var round_eff: ?pin_mod.Effort = null;
var eff_lines = std.mem.tokenizeScalar(u8, pin_mod.personaPolicyFrontmatter(a, "luna-max"), '\n');
while (eff_lines.next()) |ln| {
@@ -489,8 +489,8 @@ test "promotion carries a persona's model/tier policy, and it round-trips" {
test "withModel: rebuilds the model-derived fields, keeps the credential (#292)" {
const base = codexBase();
- const luna = base.withModel("gpt-5.6-luna");
- try std.testing.expectEqualStrings("gpt-5.6-luna", luna.model);
+ const luna = base.withModel("gpt-6-luna");
+ try std.testing.expectEqualStrings("gpt-6-luna", luna.model);
try std.testing.expectEqualStrings("codex", luna.id);
try std.testing.expectEqualStrings("k", luna.api_key);
try std.testing.expectEqual(base.kind, luna.kind);
@@ -521,14 +521,14 @@ test "effort pin: spawn param, persona frontmatter, and axis independence (#292
defer fleet.g_agent_types = saved;
fleet.g_agent_types = &.{
// The tweet-shaped persona: an exact model AND a depth — "Luna Max".
- .{ .name = "luna-worker", .desc = "", .prompt = "x", .model = "gpt-5.6-luna", .effort = .max },
+ .{ .name = "luna-worker", .desc = "", .prompt = "x", .model = "gpt-6-luna", .effort = .max },
.{ .name = "thinker", .desc = "", .prompt = "x", .effort = .xhigh },
};
// Persona effort reaches a spawn that only names the persona, riding
// alongside the persona's model pin.
const luna = pin_mod.forSpawn(base, obj(a, "{\"agent\":\"luna-worker\"}"), true);
- try std.testing.expectEqualStrings("gpt-5.6-luna", luna.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", luna.provider.?.model);
try std.testing.expectEqual(pin_mod.Effort.max, luna.effort.?);
try std.testing.expectEqual(pin_mod.EffortOutcome.pinned, luna.effort_outcome);
@@ -538,11 +538,11 @@ test "effort pin: spawn param, persona frontmatter, and axis independence (#292
// AXIS INDEPENDENCE — an effort-only call keeps the persona's model pin…
const eff_only = pin_mod.forSpawn(base, obj(a, "{\"agent\":\"luna-worker\",\"effort\":\"high\"}"), true);
- try std.testing.expectEqualStrings("gpt-5.6-luna", eff_only.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", eff_only.provider.?.model);
try std.testing.expectEqual(pin_mod.Effort.high, eff_only.effort.?);
// …and a model-only call keeps the persona's effort.
- const mdl_only = pin_mod.forSpawn(base, obj(a, "{\"agent\":\"luna-worker\",\"model\":\"gpt-5.6-terra\"}"), true);
- try std.testing.expectEqualStrings("gpt-5.6-terra", mdl_only.provider.?.model);
+ const mdl_only = pin_mod.forSpawn(base, obj(a, "{\"agent\":\"luna-worker\",\"model\":\"gpt-6-sol\"}"), true);
+ try std.testing.expectEqualStrings("gpt-6-sol", mdl_only.provider.?.model);
try std.testing.expectEqual(pin_mod.Effort.max, mdl_only.effort.?);
// An effort pin alone never touches the model: provider stays null.
diff --git a/src/subagent_selection.zig b/src/subagent_selection.zig
index 501922c7..ade30e91 100644
--- a/src/subagent_selection.zig
+++ b/src/subagent_selection.zig
@@ -63,10 +63,9 @@ pub fn modelForProvider(provider_id: []const u8, query: []const u8) ?[]const u8
/// for this provider's ladder — an off-ladder or unrecognized root model
/// returns null rather than guessing at the nearest rung.
///
-/// GPT-6 Astra stays on Astra. Demoting it onto gpt-5.6-sol/terra is a
-/// SWE-bench regression (Astra-low beats Sol-high). Unpinned workers inherit;
-/// `tier: small` still seats luna for cheap search. A Sol root still drops
-/// one rung onto terra.
+/// GPT-6 roots keep their model: demoting Astra onto Sol is a SWE-bench
+/// regression (Astra-low beats Sol-high). Unpinned workers inherit;
+/// `tier: mid` / `tier: small` still seat Sol / Luna when asked.
fn defaultLadderModel(root: provider_mod.Provider) ?[]const u8 {
const ladder = tier_ladder.forModel(root.id, root.model) orelse return null;
if (std.mem.startsWith(u8, root.model, "gpt-6")) return null;
@@ -76,8 +75,7 @@ fn defaultLadderModel(root: provider_mod.Provider) ?[]const u8 {
return null;
}
}
- if (std.mem.eql(u8, root.model, ladder.frontier) or std.mem.eql(u8, root.model, "gpt-5.6-sol"))
- return ladder.mid orelse ladder.small;
+ if (std.mem.eql(u8, root.model, ladder.frontier)) return ladder.mid orelse ladder.small;
if (ladder.mid) |mid| {
if (std.mem.eql(u8, root.model, mid)) return ladder.small;
}
@@ -163,10 +161,10 @@ test "worker selection stays provider-local unless explicitly allowed" {
if (std.mem.eql(u8, spec.id, "codex") or
std.mem.eql(u8, spec.id, "kimi")) value.* = "test-token";
}
- const root = try keys.providerById("codex", "gpt-5.6-sol");
- const terra = try subagentProvider(keys, root, null, "5.6-terra", false);
- try std.testing.expectEqualStrings("codex", terra.id);
- try std.testing.expectEqualStrings("gpt-5.6-terra", terra.model);
+ const root = try keys.providerById("codex", "gpt-6-astra");
+ const luna = try subagentProvider(keys, root, null, "gpt-6-luna", false);
+ try std.testing.expectEqualStrings("codex", luna.id);
+ try std.testing.expectEqualStrings("gpt-6-luna", luna.model);
try std.testing.expectError(
error.CrossProviderConsentRequired,
subagentProvider(keys, root, "kimi", "k3", false),
@@ -176,7 +174,7 @@ test "worker selection stays provider-local unless explicitly allowed" {
try std.testing.expectEqualStrings("k3", kimi.model);
try std.testing.expectError(
error.ModelNotOnProvider,
- subagentProvider(keys, root, null, "gpt-5.6", false),
+ subagentProvider(keys, root, null, "deepseek-v4-pro", false),
);
try std.testing.expectError(
error.UnknownModel,
@@ -194,14 +192,12 @@ test "defaultLadderModel: descends one rung per provider, bottom rung inherits"
return .{ .id = id, .kind = .openai, .auth = .bearer, .url = "", .api_key = "", .model = model, .context = 0 };
}
}.p;
- // codex: sol -> terra -> luna -> (bottom, inherit)
- try std.testing.expectEqualStrings("gpt-5.6-terra", defaultLadderModel(mk("codex", "gpt-5.6-sol")).?);
- try std.testing.expect(defaultLadderModel(mk("codex", "gpt-6-astra")) == null);
- try std.testing.expectEqualStrings("gpt-5.6-luna", defaultLadderModel(mk("codex", "gpt-5.6-terra")).?);
- try std.testing.expect(defaultLadderModel(mk("codex", "gpt-5.6-luna")) == null);
- // openai: same shape, sibling names
- try std.testing.expectEqualStrings("gpt-5.6-terra", defaultLadderModel(mk("openai", "gpt-5.6")).?);
- try std.testing.expect(defaultLadderModel(mk("openai", "gpt-6-astra")) == null);
+ // codex / openai: GPT-6 roots keep their model (Astra-low beats Sol-high);
+ // tier pins still reach Sol and Luna.
+ for ([_][]const u8{ "codex", "openai" }) |id| {
+ for ([_][]const u8{ "gpt-6-astra", "gpt-6-sol", "gpt-6-luna" }) |model|
+ try std.testing.expect(defaultLadderModel(mk(id, model)) == null);
+ }
// anthropic: opus-5-5 -> sonnet-5-5 -> (bottom) (ADR 0219). #471: a rung
// must be a cheaper seat, so older opus generations are not rungs, and a
// root still on one is off-ladder and inherits.
@@ -227,11 +223,6 @@ test "default ladder: resolves the mid sibling when no flags are given" {
if (std.mem.eql(u8, spec.id, "codex") or std.mem.eql(u8, spec.id, "anthropic"))
value.* = "test-token";
}
- const codex_root = try keys.providerById("codex", "gpt-5.6-sol");
- const worker = resolveSubagentProvider(keys, codex_root, null, null, false, false).?;
- try std.testing.expectEqualStrings("codex", worker.id); // provider-local, unpinned
- try std.testing.expectEqualStrings("gpt-5.6-terra", worker.model);
-
const astra_root = try keys.providerById("codex", "gpt-6-astra");
try std.testing.expect(resolveSubagentProvider(keys, astra_root, null, null, false, false) == null);
@@ -247,7 +238,7 @@ test "default ladder: bottom-rung and off-ladder roots inherit (resolve to null)
if (std.mem.eql(u8, spec.id, "codex") or std.mem.eql(u8, spec.id, "xai"))
value.* = "test-token";
}
- const luna_root = try keys.providerById("codex", "gpt-5.6-luna");
+ const luna_root = try keys.providerById("codex", "gpt-6-luna");
try std.testing.expect(resolveSubagentProvider(keys, luna_root, null, null, false, false) == null);
const xai_root = try keys.providerById("xai", "grok-4.3"); // unlisted family, never guess
@@ -257,9 +248,10 @@ test "default ladder: bottom-rung and off-ladder roots inherit (resolve to null)
test "--no-subagent-tier opts out of the default ladder back to plain inherit" {
var keys: provider_mod.Keys = .{ .values = @splat(null) };
for (provider_mod.provider_specs, &keys.values) |spec, *value| {
- if (std.mem.eql(u8, spec.id, "codex")) value.* = "test-token";
+ if (std.mem.eql(u8, spec.id, "anthropic")) value.* = "test-token";
}
- const root = try keys.providerById("codex", "gpt-5.6-sol");
+ const root = try keys.providerById("anthropic", "claude-opus-5-5");
+ try std.testing.expect(resolveSubagentProvider(keys, root, null, null, false, false) != null);
try std.testing.expect(resolveSubagentProvider(keys, root, null, null, false, true) == null);
}
@@ -268,10 +260,10 @@ test "explicit --subagent-model still wins over the default ladder" {
for (provider_mod.provider_specs, &keys.values) |spec, *value| {
if (std.mem.eql(u8, spec.id, "codex")) value.* = "test-token";
}
- const root = try keys.providerById("codex", "gpt-5.6-sol");
- // The ladder default would be terra; an explicit pin to luna wins.
- const pinned = resolveSubagentProvider(keys, root, null, "gpt-5.6-luna", false, false).?;
- try std.testing.expectEqualStrings("gpt-5.6-luna", pinned.model);
+ const root = try keys.providerById("codex", "gpt-6-astra");
+ // The ladder default would inherit Astra; an explicit pin to luna wins.
+ const pinned = resolveSubagentProvider(keys, root, null, "gpt-6-luna", false, false).?;
+ try std.testing.expectEqualStrings("gpt-6-luna", pinned.model);
}
test "default ladder never crosses a provider boundary implicitly" {
@@ -283,8 +275,8 @@ test "default ladder never crosses a provider boundary implicitly" {
}
const Case = struct { id: []const u8, model: []const u8 };
for ([_]Case{
- .{ .id = "codex", .model = "gpt-5.6-sol" },
- .{ .id = "openai", .model = "gpt-5.6" },
+ .{ .id = "codex", .model = "gpt-6-sol" },
+ .{ .id = "openai", .model = "gpt-6-astra" },
.{ .id = "anthropic", .model = "claude-opus-4-8" },
.{ .id = "deepseek", .model = "deepseek-v4-pro" },
}) |case| {
@@ -298,13 +290,13 @@ test "ladder rungs vs providerClass: documented, intentional disagreement (#291)
// #291 requires ladder tier labels to match scoring.providerClass()
// output for the same models, "or report the discrepancy rather than
// papering over it." The gpt family now agrees end-to-end with the
- // ladder, like claude: "terra"/"luna" are tier-distinct needles in
+ // ladder, like claude: "sol"/"luna" are tier-distinct needles in
// providerClass (as opus/sonnet/haiku are), so pullElites and fleet
- // propose events keyed on the class are no longer rung-blind — a terra
- // root pulls "mid" elites, not elites grown on sol.
- try std.testing.expectEqualStrings("frontier", @import("scoring.zig").providerClass("gpt-5.6-sol"));
- try std.testing.expectEqualStrings("mid", @import("scoring.zig").providerClass("gpt-5.6-terra")); // ladder rung: mid
- try std.testing.expectEqualStrings("small", @import("scoring.zig").providerClass("gpt-5.6-luna")); // ladder rung: small
+ // propose events keyed on the class are no longer rung-blind — a sol
+ // root pulls "mid" elites, not elites grown on astra.
+ try std.testing.expectEqualStrings("frontier", @import("scoring.zig").providerClass("gpt-6-astra"));
+ try std.testing.expectEqualStrings("mid", @import("scoring.zig").providerClass("gpt-6-sol")); // ladder rung: mid
+ try std.testing.expectEqualStrings("small", @import("scoring.zig").providerClass("gpt-6-luna")); // ladder rung: small
try std.testing.expectEqualStrings("frontier", @import("scoring.zig").providerClass("deepseek-v4-pro"));
try std.testing.expectEqualStrings("small", @import("scoring.zig").providerClass("deepseek-v4-flash")); // ladder rung: small
// Claude agrees end-to-end (opus/sonnet/haiku are tier-distinct needles
diff --git a/src/subagent_tier_ladder.zig b/src/subagent_tier_ladder.zig
index 26abba01..9410a5cf 100644
--- a/src/subagent_tier_ladder.zig
+++ b/src/subagent_tier_ladder.zig
@@ -12,7 +12,7 @@
//! `frontier`/`mid`/`small` here are the ladder's OWN rung labels; they are
//! NOT guaranteed to equal scoring.providerClass(name) for that literal
//! model string — see the "ladder rungs vs providerClass" pinned test in
-//! subagent_selection.zig: the gpt-5.6 family and DeepSeek flash now agree
+//! subagent_selection.zig: the GPT-6 family and DeepSeek flash now agree
//! end-to-end with providerClass.
const std = @import("std");
@@ -62,8 +62,8 @@ pub const TierLadder = struct {
};
pub const ladders = [_]TierLadder{
- .{ .provider = "codex", .frontier = "gpt-6-astra", .mid = "gpt-5.6-terra", .small = "gpt-5.6-luna" },
- .{ .provider = "openai", .frontier = "gpt-5.6", .mid = "gpt-5.6-terra", .small = "gpt-5.6-luna" },
+ .{ .provider = "codex", .frontier = "gpt-6-astra", .mid = "gpt-6-sol", .small = "gpt-6-luna" },
+ .{ .provider = "openai", .frontier = "gpt-6-astra", .mid = "gpt-6-sol", .small = "gpt-6-luna" },
// #471: the rungs below frontier must be cheaper SEATS. opus-5-5 ($4/$20)
// and sonnet-5-5 ($2/$10) are the current lineup (ADR 0219); descending to
// an older opus saves nothing, and sonnet-5-5 halves both halves of the bill.
@@ -87,7 +87,7 @@ pub const ladders = [_]TierLadder{
/// Is `rung` a genuinely cheaper SEAT than `frontier` — strictly less $/MTok
/// on the real price sheet? Unknown on either side answers true: an unpriced
-/// model cannot be disproved (codex's `gpt-5.6-sol` has no row at all), and
+/// model cannot be disproved (a model with no price row at all), and
/// #291's descent must keep working there.
///
/// #471: the bench sheet ranks by $/TASK, which folds the seat's rate together
@@ -196,8 +196,8 @@ test "Tier.parse/modelFor: the rung vocabulary #292 pins against (#291 names)" {
const codex = forProvider("codex").?;
try std.testing.expectEqualStrings("gpt-6-astra", codex.modelFor(.frontier).?);
- try std.testing.expectEqualStrings("gpt-5.6-terra", codex.modelFor(.mid).?);
- try std.testing.expectEqualStrings("gpt-5.6-luna", codex.modelFor(.small).?);
+ try std.testing.expectEqualStrings("gpt-6-sol", codex.modelFor(.mid).?);
+ try std.testing.expectEqualStrings("gpt-6-luna", codex.modelFor(.small).?);
// A family with no such rung answers null rather than the nearest rung.
try std.testing.expectEqualStrings("deepseek-v4-flash", forProvider("deepseek").?.modelFor(.small).?);
try std.testing.expectEqualStrings("deepseek-v4-flash", forProvider("codegraff").?.modelFor(.small).?);
@@ -213,10 +213,10 @@ test "#471 a rung must be a cheaper SEAT, not merely a terser one" {
try std.testing.expect(!cheaperSeat("claude-opus-5", "claude-fable-5"));
// sonnet-5 is the real thing: $2/$10 against $5/$25, cheaper on both halves.
try std.testing.expect(cheaperSeat("claude-opus-5", "claude-sonnet-5"));
- // Unknown on either side cannot be disproved, and #291's descent has to
- // keep working on providers whose models carry no price row at all
- // (codex's gpt-5.6-sol is the live example).
- try std.testing.expect(cheaperSeat("gpt-5.6-sol", "gpt-5.6-luna"));
+ // The GPT-6 rungs are real descents; unknown on either side cannot be
+ // disproved, so #291's descent keeps working on unpriced models.
+ try std.testing.expect(cheaperSeat("gpt-6-astra", "gpt-6-sol"));
+ try std.testing.expect(cheaperSeat("gpt-6-sol", "gpt-6-luna"));
try std.testing.expect(cheaperSeat("claude-opus-5", "no-such-model"));
}
diff --git a/src/vision_ask_tests.zig b/src/vision_ask_tests.zig
index 7e3647d9..609b842e 100644
--- a/src/vision_ask_tests.zig
+++ b/src/vision_ask_tests.zig
@@ -25,7 +25,7 @@ fn deepseek() Provider {
}
fn codex() Provider {
- return .{ .id = "codex", .kind = .responses, .auth = .bearer, .url = "", .api_key = "k", .model = "gpt-5.6-sol", .context = 270_000 };
+ return .{ .id = "codex", .kind = .responses, .auth = .bearer, .url = "", .api_key = "k", .model = "gpt-6-astra", .context = 270_000 };
}
/// A Keys table with `logged_in` providers holding a token, installed as the
@@ -89,7 +89,7 @@ test "#380 re-route: an automatic seat that cannot see images moves to one that
try std.testing.expectEqual(policy.Source.vision_ask, ask.source);
// Sub-first: the logged-in flat-rate codex subscription, cheapest rung.
try std.testing.expectEqualStrings("codex", ask.pin.provider.?.id);
- try std.testing.expectEqualStrings("gpt-5.6-luna", ask.pin.provider.?.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", ask.pin.provider.?.model);
try std.testing.expectEqualStrings("vision-ask", ask.source.label());
try std.testing.expect(ask.note() != null);
@@ -202,7 +202,7 @@ test "#380 fast fail: no vision-capable model anywhere refuses the spawn, action
test "#380 candidate search: sub-first, then provider-local under the cost ceiling" {
// Ladder rungs bottom-up, so a re-route descends price rather than seizing
// the frontier model.
- try std.testing.expectEqualStrings("gpt-5.6-luna", va.visionModelFor("codex").?);
+ try std.testing.expectEqualStrings("gpt-6-luna", va.visionModelFor("codex").?);
try std.testing.expectEqualStrings("claude-sonnet-5-5", va.visionModelFor("anthropic").?);
try std.testing.expect(va.visionModelFor("deepseek") == null); // no vision model at any rung
@@ -219,7 +219,7 @@ test "#380 candidate search: sub-first, then provider-local under the cost ceili
withKeys(&keys, &.{"codex"});
const seat = va.visionSeat(deepseek()).?;
try std.testing.expectEqualStrings("codex", seat.id);
- try std.testing.expectEqualStrings("gpt-5.6-luna", seat.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", seat.model);
// An anthropic base with no sub logged in resolves provider-locally, and
// only because sonnet-5-5 is CHEAPER than the opus it replaces.
@@ -328,8 +328,8 @@ test "#380 phase seat: a whole phase moves at once, and only an automatic one" {
// uniformly, and reported as vision-ask rather than learned-policy.
const moved = va.phaseSeat(ladder_seat, &image_tasks, false);
try std.testing.expect(moved.vision);
- try std.testing.expectEqualStrings("gpt-5.6-luna", moved.provider.model);
- try std.testing.expectEqualStrings("gpt-5.6-luna", moved.pin.?.model); // every worker gets it
+ try std.testing.expectEqualStrings("gpt-6-luna", moved.provider.model);
+ try std.testing.expectEqualStrings("gpt-6-luna", moved.pin.?.model); // every worker gets it
try std.testing.expectEqual(policy.Source.vision_ask, moved.sourceFor(false));
try std.testing.expectEqual(policy.Source.vision_ask, moved.sourceFor(true)); // even with a genome override
@@ -351,8 +351,8 @@ test "#380 phase seat: a whole phase moves at once, and only an automatic one" {
test "#380 trace: the routing line carries source=vision-ask in #372's own format" {
var buf: [320]u8 = undefined;
try std.testing.expectEqualStrings(
- "shape=adhoc role=- tier=small resolved_model=gpt-5.6-luna source=vision-ask policy_or_genome_id=-",
- policy.formatDecision(&buf, .{ .tier = .small, .resolved_model = "gpt-5.6-luna", .source = .vision_ask }),
+ "shape=adhoc role=- tier=small resolved_model=gpt-6-luna source=vision-ask policy_or_genome_id=-",
+ policy.formatDecision(&buf, .{ .tier = .small, .resolved_model = "gpt-6-luna", .source = .vision_ask }),
);
// The spelling is hyphenated like every other source, and distinct.
try std.testing.expectEqualStrings("vision-ask", policy.Source.vision_ask.label());