Skip to content

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

Python

Preserve synthesis tool catalogs at the report boundary

Rach Pradhan5083ffeb3d1e
diff --git a/docs/adr/0300-workflow-synthesis-has-a-report-boundary.md b/docs/adr/0300-workflow-synthesis-has-a-report-boundary.md
index eb66d568..9c5478b7 100644
--- a/docs/adr/0300-workflow-synthesis-has-a-report-boundary.md
+++ b/docs/adr/0300-workflow-synthesis-has-a-report-boundary.md
@@ -12,8 +12,12 @@ failure while losing the previous phase's evidence from the workflow result.
Canonical `synthesize` phases and pipeline stages have a worker-local allowance
of eight evidence-bearing model requests, followed by one report-only request.
-The final request omits tools and explicitly asks for a qualified summary.
-Dispatch also refuses tools at that boundary even if a model emits one anyway.
+The final request explicitly asks for a qualified summary. On OpenAI chat and
+Responses wires it retains the stable tool catalog with `tool_choice: "none"`,
+so disabling execution does not also change the cacheable prefix. Auto-only
+routes and other wires still omit tools; pre-existing text-only restrictions
+are never widened. Dispatch refuses tools at the boundary even if a model
+emits one anyway.
The allowance stays on the worker across turn retries; it is not a renewed
per-request or per-turn counter. Ordinary workers and root turns are unchanged.
diff --git a/scripts/test-workflow-synthesis.py b/scripts/test-workflow-synthesis.py
index a4c50121..fa4798eb 100644
--- a/scripts/test-workflow-synthesis.py
+++ b/scripts/test-workflow-synthesis.py
@@ -68,7 +68,11 @@ class Model(fixture.ScriptedModel):
if name == 'GENERAL':
return {'text': 'GENERAL_DONE'}
assert n == 9, (name, n, 'synthesis exceeded its request allowance')
- assert not body.get('tools'), (name, 'report request still advertises tools')
+ previous = calls[-2]
+ assert body.get('tools') == previous['tools'], (name, 'report changed the tool catalog')
+ assert body.get('tool_choice') == 'none', (name, 'report permits tool use')
+ assert body['messages'][0] == previous['messages'][0], (name, 'report changed the system prefix')
+ assert body.get('prompt_cache_key') == previous.get('prompt_cache_key'), (name, 'report changed cache routing')
if name.startswith('BAD'):
return fixture.tool('bash', command='printf FORBIDDEN > forbidden-' + name + '.txt')
return {'text': 'SYNTHESIS_DONE_' + name}
diff --git a/src/agent_request_body.zig b/src/agent_request_body.zig
index a9b31412..9df5df14 100644
--- a/src/agent_request_body.zig
+++ b/src/agent_request_body.zig
@@ -161,7 +161,10 @@ pub fn buildBody(self: *Agent, tools_in: ?[]const u8, force_tool: bool, stream:
try s.objectField("tools");
// #261 follow-up: the rest need the root-schema repair too.
try @import("tool_call_repair.zig").writeChatTools(&s, self.scratchAlloc(), if (at.active(self.provider) and !self.sub) at.stripTail(self, self.scratchAlloc(), t) else t, self.provider.id, self.provider.model);
- if (force_tool or @import("meta_wire.zig").restricted(self.provider.id, self.provider.model)) {
+ if (self.synthesis.keepsCatalog(self)) {
+ try s.objectField("tool_choice");
+ try s.write("none");
+ } else if (force_tool or @import("meta_wire.zig").restricted(self.provider.id, self.provider.model)) {
try s.objectField("tool_choice"); // #751: Meta is auto-only
try s.write(@import("meta_wire.zig").toolChoice(force_tool, self.provider.id, self.provider.model));
}
diff --git a/src/agent_request_body_cache_test.zig b/src/agent_request_body_cache_test.zig
index 7d53b5a6..e0613474 100644
--- a/src/agent_request_body_cache_test.zig
+++ b/src/agent_request_body_cache_test.zig
@@ -44,3 +44,37 @@ test "GPT-5.6 and later Platform marks the stable prefix; Codex and older routes
try std.testing.expect(std.mem.indexOf(u8, old, "prompt_cache_breakpoint") == null);
try std.testing.expect(std.mem.indexOf(u8, old, "prompt_cache_options") == null);
}
+
+test "workflow synthesis report keeps chat and Responses cache prefixes" {
+ var arena_state = std.heap.ArenaAllocator.init(std.testing.allocator);
+ defer arena_state.deinit();
+ const a = arena_state.allocator();
+ const catalog = "[{\"type\":\"function\",\"function\":{\"name\":\"ping\",\"description\":\"fixture\",\"parameters\":{\"type\":\"object\"}}}]";
+ for ([_]@import("provider.zig").Provider.Kind{ .openai, .responses }) |kind| {
+ var agent = try bodies.testAgentFor(a, "openai", kind, "gpt-5.6");
+ agent.sub = true;
+ agent.label = "synthesize";
+ agent.synthesis = .{ .enabled = true, .requests = 7 };
+ try agent.synthesis.beforeRequest(&agent);
+ const before = try agent.buildBody(catalog, true, true, true);
+ defer std.testing.allocator.free(before);
+ try agent.synthesis.beforeRequest(&agent);
+ const after = try agent.buildBody(if (agent.synthesis.keepsCatalog(&agent)) catalog else null, true, true, true);
+ defer std.testing.allocator.free(after);
+ const first = try std.json.parseFromSliceLeaky(std.json.Value, a, before, .{});
+ const report = try std.json.parseFromSliceLeaky(std.json.Value, a, after, .{});
+ try std.testing.expectEqualStrings(
+ try std.json.Stringify.valueAlloc(a, first.object.get("tools").?, .{}),
+ try std.json.Stringify.valueAlloc(a, report.object.get("tools").?, .{}),
+ );
+ try std.testing.expectEqualStrings(first.object.get("prompt_cache_key").?.string, report.object.get("prompt_cache_key").?.string);
+ const first_prefix = if (kind == .responses) first.object.get("instructions").? else first.object.get("messages").?.array.items[0];
+ const report_prefix = if (kind == .responses) report.object.get("instructions").? else report.object.get("messages").?.array.items[0];
+ try std.testing.expectEqualStrings(
+ try std.json.Stringify.valueAlloc(a, first_prefix, .{}),
+ try std.json.Stringify.valueAlloc(a, report_prefix, .{}),
+ );
+ try std.testing.expectEqualStrings("required", first.object.get("tool_choice").?.string);
+ try std.testing.expectEqualStrings("none", report.object.get("tool_choice").?.string);
+ }
+}
diff --git a/src/agent_request_body_responses.zig b/src/agent_request_body_responses.zig
index 851147f4..f359267e 100644
--- a/src/agent_request_body_responses.zig
+++ b/src/agent_request_body_responses.zig
@@ -83,7 +83,7 @@ pub fn write(self: *Agent, s: *std.json.Stringify, tools: ?[]const u8, force_too
try s.objectField("tools");
try serde.writeOpenAITools(s, self.scratchAlloc(), async_payload); // #261 follow-up
try s.objectField("tool_choice");
- try s.write(if (force_tool) "required" else "auto");
+ try s.write(if (self.synthesis.keepsCatalog(self)) "none" else if (force_tool) "required" else "auto");
try s.objectField("parallel_tool_calls");
try s.write(true);
}
diff --git a/src/agent_turn.zig b/src/agent_turn.zig
index 3c7990cd..8df6379e 100644
--- a/src/agent_turn.zig
+++ b/src/agent_turn.zig
@@ -82,7 +82,7 @@ pub fn run(self: *Agent, root_turn_prepared: *std.atomic.Value(bool)) anyerror![
}
if (self.sub) try self.synthesis.beforeRequest(self);
const hist_len = self.messages.items.len;
- const root = self.request(if (self.text_only) null else self.toolsJson()) catch |err| {
+ const root = self.request(if (self.text_only and !self.synthesis.keepsCatalog(self)) null else self.toolsJson()) catch |err| {
if (try @import("agent_length_cutoff.zig").recoverRunaway(self, err, hist_len)) continue;
return review_deadline.finish(try @import("agent_model_loop.zig").finishError(self, err));
};
diff --git a/src/workflow_synthesis.zig b/src/workflow_synthesis.zig
index b6022f53..7a3c15fc 100644
--- a/src/workflow_synthesis.zig
+++ b/src/workflow_synthesis.zig
@@ -22,6 +22,15 @@ pub const State = struct {
return self.enabled and self.requests > evidence_requests;
}
+ pub fn keepsCatalog(self: State, agent: *const Agent) bool {
+ if (!self.reporting() or (self.prior_text_only orelse true)) return false;
+ return switch (agent.provider.kind) {
+ .responses => true,
+ .openai => !@import("meta_wire.zig").restricted(agent.provider.id, agent.provider.model),
+ else => false,
+ };
+ }
+
pub fn beforeRequest(self: *State, agent: *Agent) !void {
if (!self.enabled) return;
// Kept on the worker, not the turn stack: transport retries cannot
@@ -81,13 +90,14 @@ test "workflow synthesis restores the original tool restriction before retention
agent.synthesis = .{ .enabled = true, .requests = 8 };
try agent.synthesis.beforeRequest(&agent);
try std.testing.expect(agent.text_only);
+ try std.testing.expectEqual(!prior, agent.synthesis.keepsCatalog(&agent));
agent.synthesis.finish(&agent);
try std.testing.expectEqual(prior, agent.text_only);
try std.testing.expectError(error.WorkflowSynthesisLimit, agent.synthesis.beforeRequest(&agent));
}
}
-test "workflow synthesis gate refuses tools even when a model ignores tool withdrawal" {
+test "workflow synthesis gate refuses tools even when a model ignores report-only choice" {
var arena = std.heap.ArenaAllocator.init(std.testing.allocator);
defer arena.deinit();
var agent = fixture(arena.allocator());
@@ -98,6 +108,21 @@ test "workflow synthesis gate refuses tools even when a model ignores tool withd
try std.testing.expectEqualStrings(tool_refusal, denied.text);
}
+test "workflow synthesis withdraws catalogs on unsupported report-only wires" {
+ var arena = std.heap.ArenaAllocator.init(std.testing.allocator);
+ defer arena.deinit();
+ var agent = fixture(arena.allocator());
+ agent.synthesis = .{ .enabled = true, .requests = 8 };
+ try agent.synthesis.beforeRequest(&agent);
+ for ([_]@import("provider.zig").Provider.Kind{ .anthropic, .interactions }) |kind| {
+ agent.provider.kind = kind;
+ try std.testing.expect(!agent.synthesis.keepsCatalog(&agent));
+ }
+ agent.provider.kind = .openai;
+ agent.provider.id = "meta";
+ try std.testing.expect(!agent.synthesis.keepsCatalog(&agent));
+}
+
test "workflow synthesis does not cap ordinary workers or discard earlier findings" {
var arena = std.heap.ArenaAllocator.init(std.testing.allocator);
defer arena.deinit();