Merge pull request 'feat(run): single-loop max-steps salvage + enforce MaxToolCalls' (#25) from feat/single-loop-salvage-max-tool-calls into main

Two gaps surfaced by a live fan-out smoke in a downstream host (mort):

1) A single-loop run that exhausts its step budget (ErrMaxSteps / ErrToolLoop)
   returned empty Output + a hard error, discarding all the reasoning it had
   produced — unlike the phased path, which already salvages a partial
   transcript and continues. Add opt-in single-loop salvage
   (Defaults.SalvageSingleLoopMaxSteps, default off): on budget exhaustion with
   empty Output, reconstruct a best-effort answer from the step narration
   (reusing salvagePhaseTranscript) and downgrade the run to a successful
   partial result. Off by default so a structured-output host keeps its clean
   hard error; an interactive host opts in.

2) RunnableAgent carried no tool-call ceiling — MaxIterations bounds STEPS, but
   one step can dispatch several tool calls, so a host's "max tool calls" cap
   had nowhere to land. Add RunnableAgent.MaxToolCalls (<=0 = unlimited,
   backward-compatible): the step observer counts executed calls and, once the
   cap is reached, stepCeilingOption forces the loop to exit; the resulting
   ErrMaxSteps is relabeled to the new ErrMaxToolCalls sentinel. Enforced
   single-loop only for now (phased is a follow-up). A cap hit is budget
   exhaustion, so salvage recovers its partial reasoning too.

isPhaseBudgetExhaustion becomes the shared isBudgetExhaustion (now also matching
ErrMaxToolCalls) so the phased and single-loop paths never drift, and the
RunStateAccessor now reports the real tool-call cap (was hardcoded 0).

Additive + backward-compatible: new optional DTO field, new opt-in Defaults
flag, new exported sentinel; unlimited/off reproduce prior behavior exactly. No
new dependencies (go mod tidy clean). Tests cover salvage on/off/no-prose, cap
enforcement + sentinel, cap+salvage, unlimited no-op, and cap-overrides-critic.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01Q1eJh5NVR11z6RzyxLANMK

.gitea/workflows/adversarial-review.yml

@@ -42,7 +42,7 @@ jobs:
     # and cache the reusable-workflow ref, so a moved v1 tag keeps resolving to the
     # stale cached copy. A unique sha forces a cache miss → fresh fetch. Bump this
     # sha to adopt central swarm changes.
-    uses: steve/gadfly/.gitea/workflows/review-reusable.yml@5007597cf921dc3f0a83c708878facfe65fd8e8b
+    uses: steve/gadfly/.gitea/workflows/review-reusable.yml@8eb0265e759e8c37ade8eaa8cb4b3aa9c7a5a169
     # Least privilege: forward only the review secrets (not `secrets: inherit`,
     # which would expose every repo secret). GITEA_TOKEN is the automatic token.
     secrets:
@@ -53,3 +53,6 @@ jobs:
     with:
       # Consumer-specific allow-list; everything else is inherited.
       allowed_users: "steve,fizi,dazed"
+      # Gitea >= 1.27 does not propagate dispatch inputs into a called workflow's
+      # github.event — thread the PR number explicitly (empty on non-dispatch events).
+      pr_number: ${{ github.event.inputs.pr_number }}

go.mod

@@ -6,6 +6,7 @@ require (
 	gitea.stevedudenhoeffer.com/steve/majordomo v0.0.0-20260626223738-1fd7109a42f3
 	github.com/google/uuid v1.6.0
 	github.com/robfig/cron/v3 v3.0.1
+	github.com/santhosh-tekuri/jsonschema/v6 v6.0.2
 	golang.org/x/crypto v0.53.0
 	gopkg.in/yaml.v3 v3.0.1
 )

go.sum

@@ -13,6 +13,8 @@ github.com/client9/misspell v0.3.4/go.mod h1:qj6jICC3Q7zFZvVWo7KLAzC3yx5G7kyvSDk
 github.com/cncf/udpa/go v0.0.0-20191209042840-269d4d468f6f/go.mod h1:M8M6+tZqaGXZJjfX53e64911xZQV5JYwmTeXPW+k8Sc=
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/dlclark/regexp2 v1.11.0 h1:G/nrcoOa7ZXlpoa/91N3X7mM3r8eIlMBBJZvsz/mxKI=
+github.com/dlclark/regexp2 v1.11.0/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
 github.com/envoyproxy/go-control-plane v0.9.0/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4=
 github.com/envoyproxy/go-control-plane v0.9.1-0.20191026205805-5f8ba28d4473/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4=
 github.com/envoyproxy/go-control-plane v0.9.4/go.mod h1:6rpuAdCZL397s3pYoYcLgu1mIlRU8Am5FuJP05cCM98=
@@ -52,6 +54,8 @@ github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZN
 github.com/prometheus/client_model v0.0.0-20190812154241-14fe0d1b01d4/go.mod h1:xMI15A0UPsDsEKsMN9yxemIoYk6Tm2C1GtYGdfGttqA=
 github.com/robfig/cron/v3 v3.0.1 h1:WdRxkvbJztn8LMz/QEvLN5sBU+xKpSqwwUO1Pjr4qDs=
 github.com/robfig/cron/v3 v3.0.1/go.mod h1:eQICP3HwyT7UooqI/z+Ov+PtYAWygg1TEWWzGIFLtro=
+github.com/santhosh-tekuri/jsonschema/v6 v6.0.2 h1:KRzFb2m7YtdldCEkzs6KqmJw4nqEVZGK7IN2kJkjTuQ=
+github.com/santhosh-tekuri/jsonschema/v6 v6.0.2/go.mod h1:JXeL+ps8p7/KNMjDQk3TCwPpBy0wYklyWTfbkIzdIFU=
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
 github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
 github.com/stretchr/objx v0.5.0/go.mod h1:Yh+to48EsGEfYuaHDzXPcE3xhTkx73EhmCGUpEOglKo=

run/agent.go

@@ -37,6 +37,15 @@ type RunnableAgent struct {
 	MaxIterations int
 	MaxRuntime    time.Duration
 
+	// MaxToolCalls caps the TOTAL number of tool calls across the whole run,
+	// independent of the step ceiling (one step can carry several parallel tool
+	// calls). 0 (or negative) = unlimited — the backward-compatible default, so
+	// a run that never sets it is unaffected. Enforced at step granularity: the
+	// loop stops after the step in which the cumulative tool-call count reaches
+	// the cap, surfacing ErrMaxToolCalls. Single-loop runs only today (phased
+	// pipelines are a follow-up).
+	MaxToolCalls int
+
 	// LowLevelTools are tool-registry names the run may call directly.
 	// SkillPalette / SubAgentPalette name saved skills / sub-agents exposed as
 	// skill__<name> / agent__<name> delegation tools, resolved through

run/critic.go

@@ -110,18 +110,36 @@ func (b *criticBinding) recordToolStart(name, args string) {
 	}
 }
 
-// maxStepsOption returns the agent step-ceiling Option. With no critic it's a
-// fixed WithMaxSteps(base); with a critic it's a DYNAMIC WithMaxStepsFunc that
-// polls the handle each step (so the critic can raise a long run's budget),
-// falling back to base when the handle defers (MaxSteps() <= 0).
-func (b *criticBinding) maxStepsOption(base int) agent.Option {
-	if b == nil {
-		return agent.WithMaxSteps(base)
-	}
+// stepCeilingOption returns the agent step-ceiling Option, folding the run's
+// tool-call cap into the (optionally critic-driven) step ceiling. Priority:
+//
+//  1. tool-call cap hit (maxCalls > 0 && *toolCalls >= maxCalls): return the
+//     completed-step count so majordomo's `stepIdx < ceiling` check fails on the
+//     next iteration and the loop exits (the executor relabels the resulting
+//     ErrMaxSteps to ErrMaxToolCalls). Never returns <= 0 — majordomo would fall
+//     back to its own static ceiling — so a cap hit before any step completes
+//     clamps to 1.
+//  2. the critic's DYNAMIC ceiling, when a critic is bound and raises a long
+//     run's budget mid-flight (b.h.MaxSteps() > 0).
+//  3. the static base (MaxIterations).
+//
+// With maxCalls <= 0 and no critic this is exactly the old WithMaxSteps(base):
+// an unlimited-tool-call run is a true no-op. nil-safe on the binding (b == nil
+// skips the critic branch). toolCalls/steps are read live from the step
+// observer's counters (same goroutine as this poll — see executor.go).
+func (b *criticBinding) stepCeilingOption(base, maxCalls int, toolCalls, steps *int) agent.Option {
 	return agent.WithMaxStepsFunc(func() int {
+		if maxCalls > 0 && toolCalls != nil && *toolCalls >= maxCalls {
+			if steps != nil && *steps > 0 {
+				return *steps
+			}
+			return 1
+		}
+		if b != nil {
 			if n := b.h.MaxSteps(); n > 0 {
 				return n
 			}
+		}
 		return base
 	})
 }

run/executor.go

@@ -41,6 +41,17 @@ type Defaults struct {
 	// absolute max. Never shorter than the run's MaxRuntime. Non-critic runs ignore
 	// it (they keep the literal MaxRuntime kill).
 	CriticAbsoluteMax time.Duration
+
+	// SalvageSingleLoopMaxSteps opts a single-loop run into partial-transcript
+	// salvage on budget exhaustion (ErrMaxSteps / ErrToolLoop / ErrMaxToolCalls):
+	// instead of returning empty Output + a hard error and discarding the
+	// reasoning it produced, the executor reconstructs a best-effort answer from
+	// the step narration (like the phased path) and downgrades the run to a
+	// successful partial result. Default false (off): a structured-output host
+	// (e.g. a bounded review swarm) keeps the clean hard error, where partial
+	// prose masquerading as a valid answer would be worse than a loud failure. A
+	// host that wants the reasoning recovered (interactive research) sets it true.
+	SalvageSingleLoopMaxSteps bool
 }
 
 func (d Defaults) withFallbacks() Defaults {
@@ -159,6 +170,7 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 	if maxRuntime <= 0 {
 		maxRuntime = e.cfg.Defaults.MaxRuntime
 	}
+	maxCalls := ra.MaxToolCalls // <= 0 => unlimited (no tool-call ceiling)
 
 	// Budget gate (pre-run): a rejected run makes no model call.
 	if e.cfg.Ports.Budget != nil {
@@ -203,7 +215,7 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 		rec = e.cfg.Ports.Audit.StartRun(ctx, info)
 	}
 	if rec != nil {
-		stateAcc = NewRunStateAccessor(rec, maxIter, 0, started)
+		stateAcc = NewRunStateAccessor(rec, maxIter, maxCalls, started)
 		inv.RunState = stateAcc
 	}
 
@@ -372,6 +384,14 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 	// calls with their executed results PAIRWISE — a result without a matching
 	// call (or a call without a result) is skipped rather than recorded as an
 	// empty-name "ghost" step.
+	// Tool-call budget accounting (single-loop enforcement, RunnableAgent.
+	// MaxToolCalls). The step observer runs on majordomo's Run goroutine — the
+	// SAME goroutine that reads the step ceiling — so plain ints suffice (no
+	// atomics). stepCeilingOption reads these to force the loop to exit once the
+	// cap is reached; the caller then relabels the resulting ErrMaxSteps to
+	// ErrMaxToolCalls. maxCalls <= 0 (unlimited) leaves this inert.
+	var toolCallsSeen, stepsSeen int
+	var capHit bool
 	emitter := newStepEmitter(inv.OnStep)
 	stepObserver := func(s agent.Step) {
 		if stateAcc != nil {
@@ -398,6 +418,21 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 				rec.OnTool(call, r.Content)
 			}
 		}
+		// Tool-call budget: accumulate executed calls and note the crossing once.
+		// Only single-loop runs enforce the cap (phased pipelines are a follow-up),
+		// so gate the crossing detection on Phases to avoid a misleading audit
+		// event on a phased run that merely happens to carry a cap.
+		toolCallsSeen += n
+		stepsSeen = s.Index + 1
+		if len(ra.Phases) == 0 && maxCalls > 0 && toolCallsSeen >= maxCalls && !capHit {
+			capHit = true
+			if rec != nil {
+				rec.LogEvent("max_tool_calls_reached", map[string]any{
+					"max_tool_calls": maxCalls,
+					"tool_calls":     toolCallsSeen,
+				})
+			}
+		}
 	}
 
 	// Shared agent options used by BOTH the single-loop path and every phase: the
@@ -480,7 +515,7 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 		}
 		opts := append([]agent.Option{
 			agent.WithToolbox(toolbox),
-			critic.maxStepsOption(maxIter),
+			critic.stepCeilingOption(maxIter, maxCalls, &toolCallsSeen, &stepsSeen),
 			agent.WithStepObserver(obs),
 		}, sharedOpts...)
 		ag := agent.New(model, e.systemPrompt(ra), opts...)
@@ -492,6 +527,28 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 			runRes, runErr = runAgent(runCtx, ag, input, inv.Images, agent.WithSteer(steer))
 		}
 
+		// A tool-call-cap exit surfaces from majordomo as ErrMaxSteps (we drove
+		// the step ceiling down to force the loop to stop). Relabel it so status,
+		// salvage, and audit report the real cause. A capped step always carries
+		// tool calls (a final-answer turn has zero), so capHit + ErrMaxSteps
+		// pinpoints a cap-triggered exit.
+		if capHit && errors.Is(runErr, agent.ErrMaxSteps) {
+			runErr = fmt.Errorf("%w (max %d)", ErrMaxToolCalls, maxCalls)
+		}
+
+		// ResultSchema (optional): validate the final answer and let the model
+		// self-repair on the same conversation before the host sees it. Runs
+		// BEFORE FinalGuard so the delivery check inspects the answer that will
+		// actually be returned. Fail-open on compile problems; best-effort on
+		// persistent violations (the host's fallback validator is authoritative).
+		if runErr == nil && runRes != nil && len(inv.ResultSchema) > 0 {
+			e.applyResultSchema(runCtx, resultSchemaDeps{
+				inv: inv, ra: ra, runRes: runRes, rec: rec,
+				model: model, toolbox: toolbox, sharedOpts: sharedOpts,
+				obs: obs, steer: steer, noteCheckpoint: noteCheckpoint,
+			})
+		}
+
 		// FinalGuard (optional): the run finished successfully — before the result
 		// is finalized, let the host inspect it for undelivered work (the
 		// announce-then-stop failure: an artifact rendered into run storage, a
@@ -557,12 +614,42 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 				}
 			}
 		}
+
+		// Single-loop budget-exhaustion salvage (opt-in). A loop that hit its
+		// step/tool-call budget without composing a final answer returns empty
+		// Output + a hard error, discarding the reasoning it DID produce. Mirror
+		// the phased path: reconstruct a best-effort answer from the step
+		// narration and downgrade the hard error to a successful partial result.
+		// This runs only after a budget error — ResultSchema + FinalGuard above
+		// are gated on runErr == nil, so neither fired. Gated on the host default
+		// so a structured-output host keeps its clean hard error.
+		if e.cfg.Defaults.SalvageSingleLoopMaxSteps &&
+			isBudgetExhaustion(runErr) && runRes != nil &&
+			strings.TrimSpace(runRes.Output) == "" {
+			if salvaged := salvagePhaseTranscript(runRes); salvaged != "" {
+				runRes.Output = salvaged +
+					"\n\n(Note: reached its step budget before composing a final answer; the above is its partial reasoning.)"
+				if rec != nil {
+					rec.LogEvent("single_loop_salvaged", map[string]any{"error": runErr.Error()})
+				}
+				runErr = nil // downgrade: status becomes "ok" and Output is delivered
+			}
+		}
 	} else {
 		// Multi-phase pipeline: each phase runs its own prompt/tier/tools/step-cap
 		// sequentially, threading outputs through {{.<PhaseName>}} templates. The
 		// shared step observer (audit/steps/critic) is wired per phase by the phase
 		// runner; checkpointing is phase-boundary granular (completed phases are
 		// recorded so a resumed run skips them).
+		//
+		// ResultSchema is single-loop-only (FinalGuard's scope) — make the drop
+		// observable rather than silent.
+		if len(inv.ResultSchema) > 0 {
+			slog.Warn("run: result_schema ignored on multi-phase run", "run_id", inv.RunID)
+			if rec != nil {
+				rec.LogEvent("result_schema_skipped", map[string]any{"error": "multi-phase runs are not schema-validated"})
+			}
+		}
 		runRes, runErr = e.runPhases(runCtx, ra, phaseDeps{
 			baseModel:    model,
 			baseToolbox:  toolbox,

run/phases.go

@@ -156,7 +156,7 @@ func (e *Executor) runPhases(runCtx context.Context, ra RunnableAgent, deps phas
 					deps.rec.LogEvent("phase_failed_optional", map[string]any{"phase": phase.Name, "error": err.Error()})
 				}
 
-			case isPhaseBudgetExhaustion(err) && (!isLast || trimmed != ""):
+			case isBudgetExhaustion(err) && (!isLast || trimmed != ""):
 				// Soft stop: the phase ran out of its step/tool budget before
 				// composing a final answer. Not fatal — it did real work (runOnePhase
 				// salvaged its partial transcript into output), and aborting would
@@ -248,7 +248,7 @@ func (e *Executor) runOnePhase(runCtx context.Context, ra RunnableAgent, deps ph
 	}
 	// Budget/guard exhaustion leaves a usable partial transcript but an empty
 	// final answer; salvage the narrated work so the pipeline can carry it forward.
-	if runErr != nil && isPhaseBudgetExhaustion(runErr) {
+	if runErr != nil && isBudgetExhaustion(runErr) {
 		if salvaged := salvagePhaseTranscript(res); salvaged != "" {
 			output = salvaged
 		}
@@ -278,11 +278,15 @@ func (e *Executor) phaseModel(ctx context.Context, deps phaseDeps, ra RunnableAg
 	return modelCtx, m
 }
 
-// isPhaseBudgetExhaustion reports whether err is a soft budget/guard stop (the
-// loop hit its step cap or tripped a tool-error guard) — which leaves a usable
-// partial transcript — as opposed to a hard error (cancellation, model failure).
-func isPhaseBudgetExhaustion(err error) bool {
-	return errors.Is(err, agent.ErrMaxSteps) || errors.Is(err, agent.ErrToolLoop)
+// isBudgetExhaustion reports whether err is a soft budget/guard stop (the loop
+// hit its step cap, its tool-call cap, or tripped a tool-error guard) — which
+// leaves a usable partial transcript — as opposed to a hard error (cancellation,
+// model failure). Shared by the phased pipeline and the single-loop salvage in
+// executor.go, so the two paths never drift on what counts as exhaustion.
+func isBudgetExhaustion(err error) bool {
+	return errors.Is(err, agent.ErrMaxSteps) ||
+		errors.Is(err, agent.ErrToolLoop) ||
+		errors.Is(err, ErrMaxToolCalls)
 }
 
 // maxSalvageBytes bounds a salvaged partial transcript so a long phase's narrated

run/ports.go

@@ -16,6 +16,14 @@ import (
 // kill via KillCause(); the executor wraps that reason with this sentinel.
 var ErrCriticKill = errors.New("run: critic killed the run")
 
+// ErrMaxToolCalls is the cause when a run reaches its RunnableAgent.MaxToolCalls
+// budget before composing a final answer. It surfaces from majordomo as
+// ErrMaxSteps (the kernel forces the step ceiling down to stop the loop) and is
+// relabeled to this sentinel so audit + status report the real cause. Treated
+// as budget exhaustion by isBudgetExhaustion, so the single-loop salvage path
+// recovers a partial answer from it when enabled.
+var ErrMaxToolCalls = errors.New("run: max tool calls reached without a final answer")
+
 // Ports are the host seams the run executor consumes. Every field is nil-safe:
 // a light host passes the zero Ports and gets a bounded, in-memory run with no
 // persistence, audit, budget, critic, delegation, or delivery — which is

run/result_schema.go

@@ -0,0 +1,295 @@
+package run
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io/fs"
+	"log/slog"
+	"sort"
+	"strings"
+
+	"github.com/santhosh-tekuri/jsonschema/v6"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/agent"
+	llm "gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// result_schema.go — kernel-native structured output for single-loop runs.
+//
+// When tool.Invocation.ResultSchema is set, the executor validates the loop's
+// final answer against the schema and, on violation, sends the loop back for
+// bounded corrective rounds ON THE SAME CONVERSATION — the model repairs its
+// own answer with full task context, which beats any post-hoc host-side
+// reformat (the host can't know what the model meant). Mirrors the FinalGuard
+// extra-round mechanics (same model, same observers, small step cap), with one
+// deliberate contrast: FinalGuard discards the extra round's text, while a
+// schema round's text IS the new candidate answer.
+//
+// Fail-open posture: a schema that doesn't compile logs a WARN and skips
+// validation entirely — the host's pre-dispatch compile is the authoritative
+// gate, and the kernel must not brick runs on a host bug. Persistent
+// violations keep the last candidate output; hosts layer their own fallback
+// validator/envelope on top.
+
+const (
+	// resultSchemaMaxBytes caps the schema the kernel will compile.
+	resultSchemaMaxBytes = 64 << 10
+	// resultSchemaMaxRounds caps corrective rounds per run.
+	resultSchemaMaxRounds = 2
+	// resultSchemaRoundMaxSteps caps each corrective round. Producing a
+	// corrected JSON document needs no tools; 2 steps tolerates one stray
+	// tool call before the reply.
+	resultSchemaRoundMaxSteps = 2
+	// resultSchemaExtractCap bounds the output size scanned for a JSON
+	// document (the balanced-span scan is quadratic in the worst case).
+	resultSchemaExtractCap = 64 << 10
+)
+
+// denyLoader fails every $ref resolution. jsonschema/v6's default compiler
+// resolves file:// refs from local disk (and can be configured for http) —
+// a caller-supplied schema must never reach either.
+type denyLoader struct{}
+
+func (denyLoader) Load(url string) (any, error) {
+	return nil, fmt.Errorf("external $ref %q refused: remote/file schema loading is disabled: %w", url, fs.ErrNotExist)
+}
+
+// compileResultSchema compiles a raw schema with external loading disabled.
+func compileResultSchema(raw json.RawMessage) (*jsonschema.Schema, error) {
+	if len(raw) > resultSchemaMaxBytes {
+		return nil, fmt.Errorf("schema is %d bytes (cap %d)", len(raw), resultSchemaMaxBytes)
+	}
+	doc, err := jsonschema.UnmarshalJSON(strings.NewReader(string(raw)))
+	if err != nil {
+		return nil, fmt.Errorf("schema is not valid JSON: %w", err)
+	}
+	c := jsonschema.NewCompiler()
+	c.UseLoader(denyLoader{})
+	if err := c.AddResource("inline://result-schema", doc); err != nil {
+		return nil, err
+	}
+	return c.Compile("inline://result-schema")
+}
+
+// extractResultJSON finds the answer's JSON document: the whole string, a
+// fenced block, or the first parseable balanced {...}/[...] span scanning
+// left to right. Input is capped at resultSchemaExtractCap.
+func extractResultJSON(s string) (json.RawMessage, bool) {
+	s = strings.TrimSpace(s)
+	if len(s) > resultSchemaExtractCap {
+		s = s[:resultSchemaExtractCap]
+	}
+	if json.Valid([]byte(s)) {
+		return json.RawMessage(s), true
+	}
+	// Fenced block: ```json ... ``` or bare ``` ... ```.
+	for _, fence := range []string{"```json", "```"} {
+		if i := strings.Index(s, fence); i >= 0 {
+			rest := s[i+len(fence):]
+			if j := strings.Index(rest, "```"); j >= 0 {
+				candidate := strings.TrimSpace(rest[:j])
+				if json.Valid([]byte(candidate)) && candidate != "" {
+					return json.RawMessage(candidate), true
+				}
+			}
+		}
+	}
+	// First parseable balanced span.
+	for i := 0; i < len(s); i++ {
+		if s[i] != '{' && s[i] != '[' {
+			continue
+		}
+		if end, ok := balancedSpanEnd(s, i); ok {
+			candidate := s[i : end+1]
+			if json.Valid([]byte(candidate)) {
+				return json.RawMessage(candidate), true
+			}
+		}
+	}
+	return nil, false
+}
+
+// balancedSpanEnd returns the index of the bracket closing the span opened
+// at start, honoring JSON string/escape rules.
+func balancedSpanEnd(s string, start int) (int, bool) {
+	depth := 0
+	inStr := false
+	esc := false
+	for i := start; i < len(s); i++ {
+		c := s[i]
+		if inStr {
+			switch {
+			case esc:
+				esc = false
+			case c == '\\':
+				esc = true
+			case c == '"':
+				inStr = false
+			}
+			continue
+		}
+		switch c {
+		case '"':
+			inStr = true
+		case '{', '[':
+			depth++
+		case '}', ']':
+			depth--
+			if depth == 0 {
+				return i, true
+			}
+		}
+	}
+	return 0, false
+}
+
+// validateResultJSON returns flattened error paths, nil when the doc conforms.
+func validateResultJSON(sch *jsonschema.Schema, doc json.RawMessage) []string {
+	inst, err := jsonschema.UnmarshalJSON(strings.NewReader(string(doc)))
+	if err != nil {
+		return []string{"document is not valid JSON: " + err.Error()}
+	}
+	err = sch.Validate(inst)
+	if err == nil {
+		return nil
+	}
+	var ve *jsonschema.ValidationError
+	if !errors.As(err, &ve) {
+		return []string{err.Error()}
+	}
+	var out []string
+	flattenValidationError(ve, &out)
+	sort.Strings(out)
+	if len(out) == 0 {
+		out = []string{ve.Error()}
+	}
+	return out
+}
+
+// flattenValidationError collects leaf causes as "/path: message" lines.
+func flattenValidationError(ve *jsonschema.ValidationError, out *[]string) {
+	if len(ve.Causes) == 0 {
+		loc := "/" + strings.Join(ve.InstanceLocation, "/")
+		if len(ve.InstanceLocation) == 0 {
+			loc = "/"
+		}
+		*out = append(*out, fmt.Sprintf("%s: %v", loc, ve.ErrorKind))
+		return
+	}
+	for _, c := range ve.Causes {
+		flattenValidationError(c, out)
+	}
+}
+
+// resultSchemaDeps bundles the per-run state applyResultSchema needs from
+// the executor's single-loop path.
+type resultSchemaDeps struct {
+	inv            tool.Invocation
+	ra             RunnableAgent
+	runRes         *agent.Result
+	rec            RunRecorder
+	model          llm.Model
+	toolbox        *llm.Toolbox
+	sharedOpts     []agent.Option
+	obs            func(agent.Step)
+	steer          func() []llm.Message
+	noteCheckpoint func(llm.Message)
+}
+
+// applyResultSchema runs the validate + corrective-round loop, mutating
+// d.runRes in place. Panic-isolated like safeFinalNudge: a bug in the
+// validator (or a pathological schema) must never convert a successful
+// run into a panic error.
+func (e *Executor) applyResultSchema(runCtx context.Context, d resultSchemaDeps) {
+	defer func() {
+		if r := recover(); r != nil {
+			slog.Error("run: result_schema validation panicked; keeping run result",
+				"run_id", d.inv.RunID, "panic", r)
+		}
+	}()
+	sch, cerr := compileResultSchema(d.inv.ResultSchema)
+	if cerr != nil {
+		slog.Warn("run: result_schema failed to compile; skipping kernel validation",
+			"run_id", d.inv.RunID, "error", cerr)
+		if d.rec != nil {
+			d.rec.LogEvent("result_schema_skipped", map[string]any{"error": cerr.Error()})
+		}
+		return
+	}
+	for round := 0; ; round++ {
+		var errs []string
+		doc, ok := extractResultJSON(d.runRes.Output)
+		if ok {
+			errs = validateResultJSON(sch, doc)
+			if errs == nil {
+				// Conforming: the bare validated document IS the answer.
+				d.runRes.Output = string(doc)
+				if d.rec != nil {
+					d.rec.LogEvent("result_schema_valid", map[string]any{"rounds": round})
+				}
+				return
+			}
+		} else {
+			errs = []string{"no JSON document found in output"}
+		}
+		if round >= resultSchemaMaxRounds {
+			// Best effort exhausted — keep the last candidate; the host's
+			// own validator decides what to do with it.
+			if d.rec != nil {
+				d.rec.LogEvent("result_schema_unmet", map[string]any{
+					"rounds": round, "validation_errors": errs,
+				})
+			}
+			return
+		}
+		correction := resultSchemaCorrection(d.inv.ResultSchema, errs)
+		if d.rec != nil {
+			d.rec.LogEvent("result_schema_round", map[string]any{
+				"round": round + 1, "validation_errors": errs,
+			})
+		}
+		if d.noteCheckpoint != nil {
+			d.noteCheckpoint(llm.UserText(correction))
+		}
+		roundOpts := append([]agent.Option{
+			agent.WithToolbox(d.toolbox),
+			agent.WithMaxSteps(resultSchemaRoundMaxSteps),
+			agent.WithStepObserver(d.obs),
+		}, d.sharedOpts...)
+		roundAg := agent.New(d.model, e.systemPrompt(d.ra), roundOpts...)
+		roundRes, roundErr := roundAg.Run(runCtx, correction,
+			agent.WithSteer(d.steer), agent.WithHistory(d.runRes.Messages))
+		if roundRes != nil {
+			roundUsage := roundRes.Usage
+			d.runRes.Usage.Add(roundUsage)
+		}
+		if roundErr != nil || roundRes == nil {
+			slog.Warn("run: result_schema corrective round failed; keeping last output",
+				"run_id", d.inv.RunID, "error", roundErr)
+			return
+		}
+		// The round's text is the NEW candidate (contrast FinalGuard, whose
+		// extra-round text is discarded meta-commentary) — but a BLANK round
+		// reply must not clobber the prior non-empty candidate.
+		if strings.TrimSpace(roundRes.Output) != "" {
+			d.runRes.Output = roundRes.Output
+		}
+		d.runRes.Messages = roundRes.Messages
+	}
+}
+
+// resultSchemaCorrection builds the corrective user turn for one round.
+func resultSchemaCorrection(raw json.RawMessage, errs []string) string {
+	return fmt.Sprintf(`Your final answer must be a SINGLE JSON document conforming to this JSON Schema (no prose, no code fences):
+%s
+
+Your previous answer failed validation:
+- %s
+
+Reply now with ONLY the corrected JSON document. Preserve your answer's content; do not invent data; use null or empty values for anything missing.`,
+		string(raw), strings.Join(errs, "\n- "))
+}

run/result_schema_test.go

@@ -0,0 +1,208 @@
+package run
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"testing"
+
+	llm "gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/fake"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+func resultSchemaExecutor(t *testing.T, m llm.Model, rec RunRecorder) *Executor {
+	t.Helper()
+	ports := Ports{}
+	if rec != nil {
+		ports.Audit = auditFunc{start: func(RunInfo) RunRecorder { return rec }}
+	}
+	return New(Config{
+		Registry: tool.NewRegistry(),
+		Models:   func(ctx context.Context, _ string) (context.Context, llm.Model, error) { return ctx, m, nil },
+		Ports:    ports,
+	})
+}
+
+var personSchema = json.RawMessage(`{"type":"object","required":["name"],"properties":{"name":{"type":"string"}},"additionalProperties":false}`)
+
+// TestResultSchema_ConformingAnswerPassesThrough — a first-try conforming
+// answer (fence-wrapped, even) is replaced by the bare JSON with no extra
+// model rounds.
+func TestResultSchema_ConformingAnswerPassesThrough(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("Here you go:\n```json\n{\"name\": \"mort\"}\n```"))
+	m, _ := fp.Model("m")
+
+	ex := resultSchemaExecutor(t, m, nil)
+	res := ex.Run(context.Background(),
+		RunnableAgent{Name: "g", ModelTier: "m"},
+		tool.Invocation{RunID: "r-rs1", CallerID: "c", ResultSchema: personSchema}, "who are you")
+
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	if res.Output != `{"name": "mort"}` {
+		t.Fatalf("output = %q, want the bare validated JSON", res.Output)
+	}
+	if n := fp.CallCount("m"); n != 1 {
+		t.Fatalf("model calls = %d, want 1 (no corrective round)", n)
+	}
+}
+
+// TestResultSchema_CorrectiveRoundRepairs — a non-conforming first answer
+// triggers one corrective round whose text becomes the run output.
+func TestResultSchema_CorrectiveRoundRepairs(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m",
+		fake.Reply(`{"name": 42}`),      // wrong type
+		fake.Reply(`{"name": "fixed"}`), // corrective round conforms
+	)
+	m, _ := fp.Model("m")
+	rec := &eventRecorder{}
+
+	ex := resultSchemaExecutor(t, m, rec)
+	res := ex.Run(context.Background(),
+		RunnableAgent{Name: "g", ModelTier: "m"},
+		tool.Invocation{RunID: "r-rs2", CallerID: "c", ResultSchema: personSchema}, "go")
+
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	if res.Output != `{"name": "fixed"}` {
+		t.Fatalf("output = %q, want the corrected JSON", res.Output)
+	}
+	if n := fp.CallCount("m"); n != 2 {
+		t.Fatalf("model calls = %d, want 2", n)
+	}
+	if !recHas(rec, "result_schema_round") || !recHas(rec, "result_schema_valid") {
+		t.Fatalf("expected round + valid audit events, got %v", rec.events)
+	}
+}
+
+// TestResultSchema_PersistentViolationKeepsLastOutput — rounds exhaust, the
+// last candidate stands, and result_schema_unmet is recorded (the host's own
+// validator is the fallback authority).
+func TestResultSchema_PersistentViolationKeepsLastOutput(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m",
+		fake.Reply("not json at all"),
+		fake.Reply("still prose"),
+		fake.Reply("prose forever"),
+	)
+	m, _ := fp.Model("m")
+	rec := &eventRecorder{}
+
+	ex := resultSchemaExecutor(t, m, rec)
+	res := ex.Run(context.Background(),
+		RunnableAgent{Name: "g", ModelTier: "m"},
+		tool.Invocation{RunID: "r-rs3", CallerID: "c", ResultSchema: personSchema}, "go")
+
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	if res.Output != "prose forever" {
+		t.Fatalf("output = %q, want the last candidate kept", res.Output)
+	}
+	if n := fp.CallCount("m"); n != 1+resultSchemaMaxRounds {
+		t.Fatalf("model calls = %d, want %d", n, 1+resultSchemaMaxRounds)
+	}
+	if !recHas(rec, "result_schema_unmet") {
+		t.Fatalf("expected result_schema_unmet, got %v", rec.events)
+	}
+}
+
+// TestResultSchema_NoSchemaUnchanged — anchor: without a schema the output
+// and call count are untouched.
+func TestResultSchema_NoSchemaUnchanged(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("plain prose answer"))
+	m, _ := fp.Model("m")
+
+	ex := resultSchemaExecutor(t, m, nil)
+	res := ex.Run(context.Background(),
+		RunnableAgent{Name: "g", ModelTier: "m"},
+		tool.Invocation{RunID: "r-rs4", CallerID: "c"}, "go")
+
+	if res.Err != nil || res.Output != "plain prose answer" {
+		t.Fatalf("output = %q err = %v", res.Output, res.Err)
+	}
+	if n := fp.CallCount("m"); n != 1 {
+		t.Fatalf("model calls = %d, want 1", n)
+	}
+}
+
+// TestResultSchema_CompileFailureFailsOpen — an uncompilable schema skips
+// validation (WARN + audit event), never bricks the run.
+func TestResultSchema_CompileFailureFailsOpen(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("prose"))
+	m, _ := fp.Model("m")
+	rec := &eventRecorder{}
+
+	ex := resultSchemaExecutor(t, m, rec)
+	res := ex.Run(context.Background(),
+		RunnableAgent{Name: "g", ModelTier: "m"},
+		tool.Invocation{RunID: "r-rs5", CallerID: "c",
+			ResultSchema: json.RawMessage(`{"type": "not-a-type"}`)}, "go")
+
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	if res.Output != "prose" {
+		t.Fatalf("output = %q, want untouched", res.Output)
+	}
+	if !recHas(rec, "result_schema_skipped") {
+		t.Fatalf("expected result_schema_skipped, got %v", rec.events)
+	}
+}
+
+// TestCompileResultSchema_ExternalRefsFailClosed — http(s) AND file $refs
+// must not resolve (jsonschema/v6's default loader reads local disk).
+func TestCompileResultSchema_ExternalRefsFailClosed(t *testing.T) {
+	for _, ref := range []string{
+		"https://example.com/schema.json",
+		"file:///etc/passwd",
+	} {
+		raw := json.RawMessage(`{"$ref": "` + ref + `"}`)
+		if _, err := compileResultSchema(raw); err == nil {
+			t.Errorf("$ref %q compiled — external loading must fail closed", ref)
+		} else if !strings.Contains(err.Error(), "refused") && !strings.Contains(err.Error(), "not exist") {
+			// Any error is acceptable as long as it FAILS; the refused
+			// wording just confirms the deny loader was consulted.
+			t.Logf("$ref %q failed with: %v", ref, err)
+		}
+	}
+}
+
+// TestExtractResultJSON_Shapes — extraction tolerance.
+func TestExtractResultJSON_Shapes(t *testing.T) {
+	cases := []struct {
+		in   string
+		want string
+		ok   bool
+	}{
+		{`{"a":1}`, `{"a":1}`, true},
+		{"prefix text {\"a\": {\"b\": \"}\"}} suffix", `{"a": {"b": "}"}}`, true},
+		{"```json\n[1,2]\n```", `[1,2]`, true},
+		{"no json here", "", false},
+		{"broken { \"a\": ", "", false},
+	}
+	for _, c := range cases {
+		got, ok := extractResultJSON(c.in)
+		if ok != c.ok || (ok && string(got) != c.want) {
+			t.Errorf("extract(%q) = (%q, %v), want (%q, %v)", c.in, got, ok, c.want, c.ok)
+		}
+	}
+}
+
+// recHas reports whether the recorder saw an event type.
+func recHas(r *eventRecorder, name string) bool {
+	for _, e := range r.events {
+		if e == name {
+			return true
+		}
+	}
+	return false
+}

run/salvage_test.go

@@ -0,0 +1,250 @@
+package run_test
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+	"sync"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/agent"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/fake"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/run"
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// salvageNoopTool is an always-succeeds tool injected via Invocation.ExtraTools
+// so a scripted tool-call loop advances (and counts a real tool call) without
+// tripping the tool-error guard.
+func salvageNoopTool() []llm.Tool {
+	return []llm.Tool{{
+		Name:        "noop",
+		Description: "no-op",
+		Handler:     func(context.Context, json.RawMessage) (any, error) { return "ok", nil },
+	}}
+}
+
+// narratingLoop returns a fake provider whose every reply NARRATES a step and
+// calls noop with DISTINCT args (so the same-call-repeat guard never trips), so
+// the loop never finalizes and runs until its step / tool-call budget. Each
+// step's narration is recoverable by the salvage path.
+func narratingLoop() *fake.Provider {
+	var mu sync.Mutex
+	var n int
+	return fake.New("fake", fake.WithDefault(func(_ string, _ llm.Request) fake.Step {
+		mu.Lock()
+		n++
+		step := n
+		mu.Unlock()
+		return fake.ReplyWith(llm.Response{
+			Parts: []llm.Part{llm.Text(fmt.Sprintf("reasoning %d: partial finding %d", step, step))},
+			ToolCalls: []llm.ToolCall{{
+				ID:        fmt.Sprintf("c%d", step),
+				Name:      "noop",
+				Arguments: []byte(fmt.Sprintf(`{"n":%d}`, step)),
+			}},
+		})
+	}))
+}
+
+// silentLoop is like narratingLoop but writes NO prose — so there is nothing for
+// the salvage path to recover.
+func silentLoop() *fake.Provider {
+	var mu sync.Mutex
+	var n int
+	return fake.New("fake", fake.WithDefault(func(_ string, _ llm.Request) fake.Step {
+		mu.Lock()
+		n++
+		step := n
+		mu.Unlock()
+		return fake.ReplyWith(llm.Response{
+			ToolCalls: []llm.ToolCall{{
+				ID:        fmt.Sprintf("c%d", step),
+				Name:      "noop",
+				Arguments: []byte(fmt.Sprintf(`{"n":%d}`, step)),
+			}},
+		})
+	}))
+}
+
+func modelsFor(m llm.Model) run.ModelResolver {
+	return func(ctx context.Context, _ string) (context.Context, llm.Model, error) { return ctx, m, nil }
+}
+
+// TestSalvageSingleLoop_OnRecoversPartialReasoning: with the salvage default ON,
+// a single-loop run that exhausts its step budget without a final answer is
+// downgraded to a successful PARTIAL result carrying its step narration, instead
+// of returning empty output + a hard ErrMaxSteps.
+func TestSalvageSingleLoop_OnRecoversPartialReasoning(t *testing.T) {
+	fp := narratingLoop()
+	m, _ := fp.Model("m")
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   modelsFor(m),
+		Defaults: run.Defaults{SalvageSingleLoopMaxSteps: true},
+	})
+	res := ex.Run(context.Background(),
+		run.RunnableAgent{Name: "researcher", ModelTier: "m", MaxIterations: 3},
+		tool.Invocation{RunID: "r", ExtraTools: salvageNoopTool()},
+		"do research")
+
+	if res.Err != nil {
+		t.Fatalf("salvage ON: budget exhaustion should downgrade to success; got err %v", res.Err)
+	}
+	if !strings.Contains(res.Output, "partial finding") {
+		t.Errorf("salvaged output should contain the step narration; got %q", res.Output)
+	}
+	if !strings.Contains(res.Output, "step budget") {
+		t.Errorf("salvaged output should carry the partial-answer note; got %q", res.Output)
+	}
+}
+
+// TestSalvageSingleLoop_OffKeepsHardError: with salvage OFF (the zero-value
+// default), the same run returns empty output + a hard ErrMaxSteps — the
+// structured-output host's clean failure.
+func TestSalvageSingleLoop_OffKeepsHardError(t *testing.T) {
+	fp := narratingLoop()
+	m, _ := fp.Model("m")
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   modelsFor(m),
+		// Defaults zero => SalvageSingleLoopMaxSteps false.
+	})
+	res := ex.Run(context.Background(),
+		run.RunnableAgent{Name: "x", ModelTier: "m", MaxIterations: 3},
+		tool.Invocation{RunID: "r", ExtraTools: salvageNoopTool()},
+		"go")
+
+	if !errors.Is(res.Err, agent.ErrMaxSteps) {
+		t.Fatalf("salvage OFF: expected ErrMaxSteps, got %v", res.Err)
+	}
+	if strings.TrimSpace(res.Output) != "" {
+		t.Errorf("salvage OFF: output should be empty; got %q", res.Output)
+	}
+}
+
+// TestSalvageSingleLoop_NoProseStillErrors: salvage recovers nothing when the
+// loop wrote no prose, so a hard error stands even with salvage ON.
+func TestSalvageSingleLoop_NoProseStillErrors(t *testing.T) {
+	fp := silentLoop()
+	m, _ := fp.Model("m")
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   modelsFor(m),
+		Defaults: run.Defaults{SalvageSingleLoopMaxSteps: true},
+	})
+	res := ex.Run(context.Background(),
+		run.RunnableAgent{Name: "x", ModelTier: "m", MaxIterations: 3},
+		tool.Invocation{RunID: "r", ExtraTools: salvageNoopTool()},
+		"go")
+
+	if !errors.Is(res.Err, agent.ErrMaxSteps) {
+		t.Fatalf("no prose to salvage: ErrMaxSteps should stand, got %v", res.Err)
+	}
+	if strings.TrimSpace(res.Output) != "" {
+		t.Errorf("output should be empty; got %q", res.Output)
+	}
+}
+
+// TestMaxToolCalls_ForcesExitAndSurfacesSentinel: a run capped at 2 tool calls
+// stops after the 2nd (well before its high step ceiling) and surfaces the
+// distinct ErrMaxToolCalls sentinel.
+func TestMaxToolCalls_ForcesExitAndSurfacesSentinel(t *testing.T) {
+	fp := narratingLoop()
+	m, _ := fp.Model("m")
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   modelsFor(m),
+		// salvage OFF so the raw sentinel is observable.
+	})
+	res := ex.Run(context.Background(),
+		run.RunnableAgent{Name: "x", ModelTier: "m", MaxIterations: 50, MaxToolCalls: 2},
+		tool.Invocation{RunID: "r", ExtraTools: salvageNoopTool()},
+		"go")
+
+	if !errors.Is(res.Err, run.ErrMaxToolCalls) {
+		t.Fatalf("expected ErrMaxToolCalls, got %v", res.Err)
+	}
+	// It must NOT read as a plain step-budget exhaustion.
+	if !strings.Contains(res.Err.Error(), "max tool calls") {
+		t.Errorf("error should name the tool-call cap; got %q", res.Err.Error())
+	}
+}
+
+// TestMaxToolCalls_SalvagedWhenEnabled: a tool-call-cap exit is treated as budget
+// exhaustion, so with salvage ON it recovers the partial reasoning too.
+func TestMaxToolCalls_SalvagedWhenEnabled(t *testing.T) {
+	fp := narratingLoop()
+	m, _ := fp.Model("m")
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   modelsFor(m),
+		Defaults: run.Defaults{SalvageSingleLoopMaxSteps: true},
+	})
+	res := ex.Run(context.Background(),
+		run.RunnableAgent{Name: "x", ModelTier: "m", MaxIterations: 50, MaxToolCalls: 2},
+		tool.Invocation{RunID: "r", ExtraTools: salvageNoopTool()},
+		"go")
+
+	if res.Err != nil {
+		t.Fatalf("cap + salvage: expected downgraded success, got %v", res.Err)
+	}
+	if !strings.Contains(res.Output, "partial finding") {
+		t.Errorf("expected salvaged narration, got %q", res.Output)
+	}
+}
+
+// TestMaxToolCalls_ZeroIsUnbounded: MaxToolCalls <= 0 imposes no ceiling — a run
+// that makes several tool calls then finalizes completes normally (regression:
+// unlimited must be a true no-op).
+func TestMaxToolCalls_ZeroIsUnbounded(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m",
+		fake.ReplyWith(llm.Response{ToolCalls: []llm.ToolCall{{ID: "c1", Name: "noop", Arguments: []byte(`{"n":1}`)}}}),
+		fake.ReplyWith(llm.Response{ToolCalls: []llm.ToolCall{{ID: "c2", Name: "noop", Arguments: []byte(`{"n":2}`)}}}),
+		fake.ReplyWith(llm.Response{ToolCalls: []llm.ToolCall{{ID: "c3", Name: "noop", Arguments: []byte(`{"n":3}`)}}}),
+		fake.Reply("final answer after 3 tools"),
+	)
+	m, _ := fp.Model("m")
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   modelsFor(m),
+	})
+	res := ex.Run(context.Background(),
+		run.RunnableAgent{Name: "x", ModelTier: "m", MaxIterations: 10, MaxToolCalls: 0},
+		tool.Invocation{RunID: "r", ExtraTools: salvageNoopTool()},
+		"go")
+
+	if res.Err != nil {
+		t.Fatalf("unbounded run should complete: %v", res.Err)
+	}
+	if res.Output != "final answer after 3 tools" {
+		t.Errorf("output = %q, want the final answer", res.Output)
+	}
+}
+
+// TestMaxToolCalls_CapOverridesCriticRaise: the tool-call cap is a HARD budget —
+// it stops the run even when a critic is willing to raise the step ceiling far
+// past it. Also proves stepCeilingOption composes with a critic.
+func TestMaxToolCalls_CapOverridesCriticRaise(t *testing.T) {
+	h := &fakeCriticHandle{maxSteps: 50}
+	fp := narratingLoop()
+	m, _ := fp.Model("m")
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   modelsFor(m),
+		Ports:    run.Ports{Critic: &fakeCritic{h: h}},
+	})
+	res := ex.Run(context.Background(),
+		run.RunnableAgent{Name: "x", ModelTier: "m", MaxIterations: 1, MaxToolCalls: 2, Critic: run.CriticConfig{Enabled: true}},
+		tool.Invocation{RunID: "r", ExtraTools: salvageNoopTool()},
+		"go")
+
+	if !errors.Is(res.Err, run.ErrMaxToolCalls) {
+		t.Fatalf("the tool-call cap must override the critic's raised ceiling; got %v", res.Err)
+	}
+}

tool/registry.go

@@ -19,6 +19,7 @@ package tool
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"sync"
 	"time"
@@ -238,6 +239,20 @@ type Invocation struct {
 	// persisted agent row. Skill runs ignore this field.
 	SystemPromptPrepend string
 
+	// ResultSchema, when non-empty, is a JSON Schema (draft 2020-12) the
+	// run's final answer must conform to. The executor validates the
+	// single-loop final output against it and, on violation, runs up to
+	// two bounded corrective rounds on the same conversation so the
+	// model self-repairs with full task context; a conforming answer is
+	// replaced by the bare validated JSON document. The kernel is
+	// FAIL-OPEN on schema problems (an uncompilable schema logs a WARN
+	// and skips validation — the host's own pre-dispatch compile is the
+	// authoritative gate) and best-effort on persistent violations (the
+	// last output stands; hosts keep their own fallback validator /
+	// error envelope). Multi-phase runs are not validated, matching
+	// FinalGuard's single-loop-only scope.
+	ResultSchema json.RawMessage
+
 	// SuppressDelivery, when true, instructs the skill executor to
 	// SKIP its OutputTarget Delivery (Deliver / DeliverError) entirely.
 	// The run still produces an output string (returned from Run) and