Merge #23: request model's full output-token budget per turn

Thread the resolved model's catalog MaxOutput through to
provider.Request.MaxTokens so each turn requests the model's full output
capacity. Fixes Bedrock silently truncating long writes/edits at its
4096 default (stopReason=length). Other providers already defaulted to
MaxOutput on a zero request, so this is a no-op for them. Also surfaces
StopLength explicitly in the TUI instead of ending silently.

Co-authored-by: Raymond Gasper <raymondgasper@fastmail.com>

commit: c2c9a5ea2898399e02294749f70ecfb982082c02
author: patriceckhart · 4mo ago · 2026-06-09 16:38 UTC
merge of: 3d031dd + a373e82 · diff against the first parent
signature: unsigned
refs: tag: v0.2.23

5 files · +100 -0

M packages/agent/build.go +9 -0


@@ -32,6 +32,13 @@ type Resolved struct {
32	32		MaxSteps int
33	33		Sandbox *tools.Sandbox
34	34
	35	+	// MaxOutput is the resolved model's maximum output-token budget
	36	+	// (from the catalog). Passed to the agent so each turn requests
	37	+	// the model's full output capacity instead of the provider's
	38	+	// conservative default (e.g. Bedrock's 4096, which truncates
	39	+	// long writes/edits with stopReason=length).
	40	+	MaxOutput int
	41	+
35	42		// SkillTool is the on-demand skill loader registered with the
36	43		// agent's tool registry, or nil if no SKILL.md files were
37	44		// discovered. Exposed so the tui can list / preview skills.

@@ -501,6 +508,7 @@ func Resolve(args Args, requireCred bool) (Resolved, error) {
501	508		ToolSummary: summaries,
502	509		SystemPrompt: sys,
503	510		MaxSteps: max,
	511	+	MaxOutput: resolvedModel.MaxOutput,
504	512		Sandbox: sandbox,
505	513		SkillTool: skillTool,
506	514		systemAppend: append_,

@@ -768,6 +776,7 @@ func (r Resolved) UseSandbox(s tools.Sandbox) {
768	776		func (r Resolved) NewAgent() *core.Agent {
769	777		a := core.NewAgent(r.NewClient(), r.Model, r.SystemPrompt, r.ToolRegistry)
770	778		a.MaxSteps = r.MaxSteps
	779	+	a.MaxTokens = r.MaxOutput
771	780		a.Reasoning = r.Reasoning
772	781		return a
773	782		}


@@ -32,6 +32,13 @@ type Resolved struct {
32	MaxSteps int	32		MaxSteps int
33	Sandbox *tools.Sandbox	33		Sandbox *tools.Sandbox
34		34
		35	+	// MaxOutput is the resolved model's maximum output-token budget
		36	+	// (from the catalog). Passed to the agent so each turn requests
		37	+	// the model's full output capacity instead of the provider's
		38	+	// conservative default (e.g. Bedrock's 4096, which truncates
		39	+	// long writes/edits with stopReason=length).
		40	+	MaxOutput int
		41	+
35	// SkillTool is the on-demand skill loader registered with the	42		// SkillTool is the on-demand skill loader registered with the
36	// agent's tool registry, or nil if no SKILL.md files were	43		// agent's tool registry, or nil if no SKILL.md files were
37	// discovered. Exposed so the tui can list / preview skills.	44		// discovered. Exposed so the tui can list / preview skills.

@@ -501,6 +508,7 @@ func Resolve(args Args, requireCred bool) (Resolved, error) {
501	ToolSummary: summaries,	508		ToolSummary: summaries,
502	SystemPrompt: sys,	509		SystemPrompt: sys,
503	MaxSteps: max,	510		MaxSteps: max,
		511	+	MaxOutput: resolvedModel.MaxOutput,
504	Sandbox: sandbox,	512		Sandbox: sandbox,
505	SkillTool: skillTool,	513		SkillTool: skillTool,
506	systemAppend: append_,	514		systemAppend: append_,

@@ -768,6 +776,7 @@ func (r Resolved) UseSandbox(s tools.Sandbox) {
768	func (r Resolved) NewAgent() *core.Agent {	776		func (r Resolved) NewAgent() *core.Agent {
769	a := core.NewAgent(r.NewClient(), r.Model, r.SystemPrompt, r.ToolRegistry)	777		a := core.NewAgent(r.NewClient(), r.Model, r.SystemPrompt, r.ToolRegistry)
770	a.MaxSteps = r.MaxSteps	778		a.MaxSteps = r.MaxSteps
		779	+	a.MaxTokens = r.MaxOutput
771	a.Reasoning = r.Reasoning	780		a.Reasoning = r.Reasoning
772	return a	781		return a
773	}	782		}

M packages/agent/modes/interactive.go +11 -0


@@ -4656,6 +4656,17 @@ func (i *Interactive) handleEvent(ev core.AgentEvent) {
4656	4656		i.statusOK = "cancelled"
4657	4657		return
4658	4658		}
	4659	+	if e.Stop == provider.StopLength {
	4660	+	// The model hit its output-token cap mid-response, so the
	4661	+	// reply (often a long write/edit) is truncated. Surface it
	4662	+	// explicitly, otherwise the turn just ends and reads like
	4663	+	// the UI gave up. The agent already requests the model's
	4664	+	// full MaxOutput budget, so this means the response genuinely
	4665	+	// exceeded that ceiling; ask the user to continue.
	4666	+	i.statusErr = "response hit the model's output-token limit and was cut off, ask it to continue"
	4667	+	i.statusOK = ""
	4668	+	return
	4669	+	}
4659	4670		// Don't surface mid-loop stream errors as a red banner here.
4660	4671		// EvTurnEnd fires after every step in a multi-step tool loop,
4661	4672		// so a transient 503 / network blip would briefly paint a red


@@ -4656,6 +4656,17 @@ func (i *Interactive) handleEvent(ev core.AgentEvent) {
4656	i.statusOK = "cancelled"	4656		i.statusOK = "cancelled"
4657	return	4657		return
4658	}	4658		}
		4659	+	if e.Stop == provider.StopLength {
		4660	+	// The model hit its output-token cap mid-response, so the
		4661	+	// reply (often a long write/edit) is truncated. Surface it
		4662	+	// explicitly, otherwise the turn just ends and reads like
		4663	+	// the UI gave up. The agent already requests the model's
		4664	+	// full MaxOutput budget, so this means the response genuinely
		4665	+	// exceeded that ceiling; ask the user to continue.
		4666	+	i.statusErr = "response hit the model's output-token limit and was cut off, ask it to continue"
		4667	+	i.statusOK = ""
		4668	+	return
		4669	+	}
4659	// Don't surface mid-loop stream errors as a red banner here.	4670		// Don't surface mid-loop stream errors as a red banner here.
4660	// EvTurnEnd fires after every step in a multi-step tool loop,	4671		// EvTurnEnd fires after every step in a multi-step tool loop,
4661	// so a transient 503 / network blip would briefly paint a red	4672		// so a transient 503 / network blip would briefly paint a red

M packages/core/agent.go +9 -0


@@ -22,6 +22,14 @@ type Agent struct {
22	22		MaxSteps int
23	23		Reasoning string
24	24
	25	+	// MaxTokens caps the model's output tokens per turn. Zero leaves
	26	+	// the field unset on the provider request, letting each provider
	27	+	// apply its own default (which can be conservative, e.g. Bedrock
	28	+	// defaults to 4096, truncating long writes/edits). Hosts populate
	29	+	// this from the resolved model's MaxOutput so large single-turn
	30	+	// responses aren't silently cut off with stopReason=length.
	31	+	MaxTokens int
	32	+
25	33		// BeforeToolExecute, if set, is called immediately before each
26	34		// tool runs. Returning (allowed=false, reason) short-circuits
27	35		// the call with an error result containing reason. Optionally,

@@ -515,6 +523,7 @@ func (a *Agent) oneTurn(ctx context.Context, sink func(AgentEvent)) (provider.St
515	523		Messages: repairToolUseResultPairs(a.Messages()),
516	524		Tools: a.Tools.Specs(),
517	525		Reasoning: a.Reasoning,
	526	+	MaxTokens: a.MaxTokens,
518	527		}
519	528		stream, err := a.Client.Stream(ctx, req)
520	529		if err != nil {


@@ -22,6 +22,14 @@ type Agent struct {
22	MaxSteps int	22		MaxSteps int
23	Reasoning string	23		Reasoning string
24		24
		25	+	// MaxTokens caps the model's output tokens per turn. Zero leaves
		26	+	// the field unset on the provider request, letting each provider
		27	+	// apply its own default (which can be conservative, e.g. Bedrock
		28	+	// defaults to 4096, truncating long writes/edits). Hosts populate
		29	+	// this from the resolved model's MaxOutput so large single-turn
		30	+	// responses aren't silently cut off with stopReason=length.
		31	+	MaxTokens int
		32	+
25	// BeforeToolExecute, if set, is called immediately before each	33		// BeforeToolExecute, if set, is called immediately before each
26	// tool runs. Returning (allowed=false, reason) short-circuits	34		// tool runs. Returning (allowed=false, reason) short-circuits
27	// the call with an error result containing reason. Optionally,	35		// the call with an error result containing reason. Optionally,

@@ -515,6 +523,7 @@ func (a *Agent) oneTurn(ctx context.Context, sink func(AgentEvent)) (provider.St
515	Messages: repairToolUseResultPairs(a.Messages()),	523		Messages: repairToolUseResultPairs(a.Messages()),
516	Tools: a.Tools.Specs(),	524		Tools: a.Tools.Specs(),
517	Reasoning: a.Reasoning,	525		Reasoning: a.Reasoning,
		526	+	MaxTokens: a.MaxTokens,
518	}	527		}
519	stream, err := a.Client.Stream(ctx, req)	528		stream, err := a.Client.Stream(ctx, req)
520	if err != nil {	529		if err != nil {

M packages/core/agent_retry_test.go +35 -0


@@ -109,3 +109,38 @@ func TestAgentDropsPartialAssistantBeforeRetry(t *testing.T) {
109	109		t.Fatalf("final assistant text = %q; want recovered", got)
110	110		}
111	111		}
	112	+
	113	+	// captureClient records the last Request it received so tests can
	114	+	// assert what the agent put on the wire.
	115	+	type captureClient struct {
	116	+	lastReq provider.Request
	117	+	}
	118	+
	119	+	func (c *captureClient) Name() string { return "capture" }
	120	+
	121	+	func (c *captureClient) Stream(ctx context.Context, req provider.Request) (<-chan provider.Event, error) {
	122	+	c.lastReq = req
	123	+	out := make(chan provider.Event, 3)
	124	+	go func() {
	125	+	defer close(out)
	126	+	out <- provider.EventStart{Provider: "capture", Model: req.Model}
	127	+	out <- provider.EventDone{Stop: provider.StopEnd, Message: provider.Message{
	128	+	Role: provider.RoleAssistant,
	129	+	Content: []provider.Content{provider.TextBlock{Text: "ok"}},
	130	+	}}
	131	+	}()
	132	+	return out, nil
	133	+	}
	134	+
	135	+	func TestAgentPropagatesMaxTokens(t *testing.T) {
	136	+	client := &captureClient{}
	137	+	a := NewAgent(client, "fake-model", "system", Registry{})
	138	+	a.MaxTokens = 64000
	139	+
	140	+	if err := a.Prompt(context.Background(), "hello", nil, nil); err != nil {
	141	+	t.Fatalf("Prompt returned %v", err)
	142	+	}
	143	+	if client.lastReq.MaxTokens != 64000 {
	144	+	t.Fatalf("request MaxTokens = %d; want 64000 (Agent.MaxTokens not propagated)", client.lastReq.MaxTokens)
	145	+	}
	146	+	}


@@ -109,3 +109,38 @@ func TestAgentDropsPartialAssistantBeforeRetry(t *testing.T) {
109	t.Fatalf("final assistant text = %q; want recovered", got)	109		t.Fatalf("final assistant text = %q; want recovered", got)
110	}	110		}
111	}	111		}
		112	+
		113	+	// captureClient records the last Request it received so tests can
		114	+	// assert what the agent put on the wire.
		115	+	type captureClient struct {
		116	+	lastReq provider.Request
		117	+	}
		118	+
		119	+	func (c *captureClient) Name() string { return "capture" }
		120	+
		121	+	func (c *captureClient) Stream(ctx context.Context, req provider.Request) (<-chan provider.Event, error) {
		122	+	c.lastReq = req
		123	+	out := make(chan provider.Event, 3)
		124	+	go func() {
		125	+	defer close(out)
		126	+	out <- provider.EventStart{Provider: "capture", Model: req.Model}
		127	+	out <- provider.EventDone{Stop: provider.StopEnd, Message: provider.Message{
		128	+	Role: provider.RoleAssistant,
		129	+	Content: []provider.Content{provider.TextBlock{Text: "ok"}},
		130	+	}}
		131	+	}()
		132	+	return out, nil
		133	+	}
		134	+
		135	+	func TestAgentPropagatesMaxTokens(t *testing.T) {
		136	+	client := &captureClient{}
		137	+	a := NewAgent(client, "fake-model", "system", Registry{})
		138	+	a.MaxTokens = 64000
		139	+
		140	+	if err := a.Prompt(context.Background(), "hello", nil, nil); err != nil {
		141	+	t.Fatalf("Prompt returned %v", err)
		142	+	}
		143	+	if client.lastReq.MaxTokens != 64000 {
		144	+	t.Fatalf("request MaxTokens = %d; want 64000 (Agent.MaxTokens not propagated)", client.lastReq.MaxTokens)
		145	+	}
		146	+	}

M packages/provider/amazon_bedrock_test.go +36 -0


@@ -163,6 +163,42 @@ func TestBedrockModelSupportsCaching(t *testing.T) {
163	163		}
164	164		}
165	165
	166	+	func TestBedrockBuildRequestMaxTokens(t *testing.T) {
	167	+	client := &bedrockClient{region: "us-east-1"}
	168	+
	169	+	// A non-zero MaxTokens must flow through to InferenceConfig so the
	170	+	// model gets its full output budget. This is the regression guard
	171	+	// for long writes/edits being truncated at Bedrock's 4096 default.
	172	+	req, err := client.buildRequest(Request{
	173	+	Model: "anthropic.claude-sonnet-4-5-20250929-v1:0",
	174	+	MaxTokens: 64000,
	175	+	Messages: []Message{
	176	+	{Role: RoleUser, Content: []Content{TextBlock{Text: "hello"}}},
	177	+	},
	178	+	})
	179	+	if err != nil {
	180	+	t.Fatal(err)
	181	+	}
	182	+	if req.InferenceConfig.MaxTokens != 64000 {
	183	+	t.Errorf("MaxTokens = %d, want 64000", req.InferenceConfig.MaxTokens)
	184	+	}
	185	+
	186	+	// Zero still falls back to the conservative provider default so an
	187	+	// unset budget never sends maxTokens:0 (which Bedrock rejects).
	188	+	reqZero, err := client.buildRequest(Request{
	189	+	Model: "anthropic.claude-sonnet-4-5-20250929-v1:0",
	190	+	Messages: []Message{
	191	+	{Role: RoleUser, Content: []Content{TextBlock{Text: "hello"}}},
	192	+	},
	193	+	})
	194	+	if err != nil {
	195	+	t.Fatal(err)
	196	+	}
	197	+	if reqZero.InferenceConfig.MaxTokens != 4096 {
	198	+	t.Errorf("zero MaxTokens default = %d, want 4096", reqZero.InferenceConfig.MaxTokens)
	199	+	}
	200	+	}
	201	+
166	202		func TestBedrockBuildRequestCachingClaudeModel(t *testing.T) {
167	203		// A Claude model (PriceCacheWrite > 0) should get cachePoint markers
168	204		// in the system array and on the last user message.


@@ -163,6 +163,42 @@ func TestBedrockModelSupportsCaching(t *testing.T) {
163	}	163		}
164	}	164		}
165		165
		166	+	func TestBedrockBuildRequestMaxTokens(t *testing.T) {
		167	+	client := &bedrockClient{region: "us-east-1"}
		168	+
		169	+	// A non-zero MaxTokens must flow through to InferenceConfig so the
		170	+	// model gets its full output budget. This is the regression guard
		171	+	// for long writes/edits being truncated at Bedrock's 4096 default.
		172	+	req, err := client.buildRequest(Request{
		173	+	Model: "anthropic.claude-sonnet-4-5-20250929-v1:0",
		174	+	MaxTokens: 64000,
		175	+	Messages: []Message{
		176	+	{Role: RoleUser, Content: []Content{TextBlock{Text: "hello"}}},
		177	+	},
		178	+	})
		179	+	if err != nil {
		180	+	t.Fatal(err)
		181	+	}
		182	+	if req.InferenceConfig.MaxTokens != 64000 {
		183	+	t.Errorf("MaxTokens = %d, want 64000", req.InferenceConfig.MaxTokens)
		184	+	}
		185	+
		186	+	// Zero still falls back to the conservative provider default so an
		187	+	// unset budget never sends maxTokens:0 (which Bedrock rejects).
		188	+	reqZero, err := client.buildRequest(Request{
		189	+	Model: "anthropic.claude-sonnet-4-5-20250929-v1:0",
		190	+	Messages: []Message{
		191	+	{Role: RoleUser, Content: []Content{TextBlock{Text: "hello"}}},
		192	+	},
		193	+	})
		194	+	if err != nil {
		195	+	t.Fatal(err)
		196	+	}
		197	+	if reqZero.InferenceConfig.MaxTokens != 4096 {
		198	+	t.Errorf("zero MaxTokens default = %d, want 4096", reqZero.InferenceConfig.MaxTokens)
		199	+	}
		200	+	}
		201	+
166	func TestBedrockBuildRequestCachingClaudeModel(t *testing.T) {	202		func TestBedrockBuildRequestCachingClaudeModel(t *testing.T) {
167	// A Claude model (PriceCacheWrite > 0) should get cachePoint markers	203		// A Claude model (PriceCacheWrite > 0) should get cachePoint markers
168	// in the system array and on the last user message.	204		// in the system array and on the last user message.