P3: meta + primitive tool group (think/now/cite + classify/extract/summarize)

Grow executus/tools into a real generic tool library:

- Register(reg): the always-available, zero-config tools — think, now (UTC
  unless a CurrentTimeProvider is wired), cite (inert unless a CitationStorage
  is wired). All nil-safe; a light host calls Register and is useful.
- RegisterMeta(reg, MetaDeps): the LLM-backed meta tools — classify,
  extract_entities, summarize — over the llmmeta helper. Budget defaults to the
  shipped in-memory per-run cap; Files optional; caps default.
- Seams moved (interface/type-only, no host coupling): research_providers.go
  (CurrentTimeProvider/CitationStorage/SearchBudget/PageExtractor/PDFFetcher/…)
  and file_storage.go (FileStorage + FileDomainMeta). Plus the in-memory budget
  default (research_defaults.go) and scope_validate.go.

calculate deferred (drags github.com/Krognol/go-wolfram + a module-path replace
— not worth it in the lean core for one tool). Core go.sum still free of
gorm/redis/discordgo/sqlite/wolfram.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

CLAUDE.md

@@ -58,11 +58,13 @@ CORE (majordomo + stdlib):
              structured output — no separate structured/ pkg)
   llmmeta/  shared meta-LLM helper over model/            [P1 ✓]
   compact/  context compactor (WithCompactor hook)       [P2 ✓]
-  tools/    generic tool library + Register entrypoint;  [P3 wip]
+  tools/    generic tool library: Register (think/now/    [P3 wip]
-            think moved; end-to-end "agent calls a tool"
+            cite, zero-config) + RegisterMeta (classify/
-            test green. Remaining: meta/web/net/store/
+            extract_entities/summarize); seams in
-            compose groups + their nil-safe Deps + default
+            research_providers.go/file_storage.go;
-            backends (the default.go grab-bag split)      [P3]
+            in-memory budget default. End-to-end "agent
+            calls a tool" test green. Remaining: web/net/
+            store/compose groups + their backends         [P3]
 
 BATTERIES (opt-in siblings, each nil-safe + a default):
   persona/   Agent noun + AgentStore seam + yml loader   [P4]

tools/cite.go

@@ -0,0 +1,128 @@
+// Package tools — v11 cite.
+//
+// Anti-hallucination forcing function. The convention: agents call
+// cite(claim, url) for every numbered reference in their final
+// answer. The tool verifies the URL appears in the run's
+// touched-URL set (populated by web_search results +
+// read_page/read_pdf/read_video). If yes → write to
+// skill_run_sources, return {ok: true}. If no → return
+// {ok: false, reason: "url_not_in_run_history"} and DO NOT write.
+//
+// Skills authored without this discipline don't lose anything;
+// skills WITH it produce more reliable citations. The webui
+// renders the skill_run_sources rows as a Sources panel on the
+// run trace page — invisible to skills that don't use cite().
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// citeParams is the LLM-facing param struct.
+type citeParams struct {
+	Claim string `json:"claim" description:"The claim or fact you are asserting (e.g. 'Mort was published in 1987')."`
+	URL   string `json:"url" description:"The URL that supports the claim. MUST be a URL the agent has previously read via read_page/read_pdf/read_video or seen as a web_search result."`
+}
+
+// citeResponse is the JSON envelope returned to the agent.
+//
+// On success: ok=true, the skill_run_sources row was written.
+// On failure: ok=false, reason=<one of the documented sentinels>.
+type citeResponse struct {
+	OK     bool   `json:"ok"`
+	Reason string `json:"reason,omitempty"`
+	Claim  string `json:"claim,omitempty"`
+	URL    string `json:"url,omitempty"`
+}
+
+// NewCite constructs the v11 cite tool. cs may be nil — handler
+// returns "not configured" at first call.
+//
+// The "anyone author / share-safe" permission shape matches every
+// other v11 research-class tool. Skills that adopt cite() get the
+// Sources panel automatically; skills that don't are unaffected.
+func NewCite(cs CitationStorage) tool.Tool {
+	return tool.NewGatedTool[citeParams](
+		"cite",
+		"Record a citation: a claim + the URL that supports it. The URL MUST be one the agent has actually fetched via read_page/read_pdf/read_video or seen as a web_search result — citing a URL the agent never visited is rejected with reason 'url_not_in_run_history'. Successful citations populate the run's Sources panel.",
+		tool.Permission{
+			AuthoringRequirement: tool.RequirementAnyone,
+			OperatesOn:           tool.ScopeGlobal,
+			SafeForShare:         true,
+			Categories:           []string{"citation"},
+		},
+		func(ctx context.Context, inv tool.Invocation, p citeParams) (string, error) {
+			if cs == nil {
+				return "", fmt.Errorf("cite: citation storage not configured")
+			}
+			claim := strings.TrimSpace(p.Claim)
+			urlStr := strings.TrimSpace(p.URL)
+			if claim == "" {
+				return marshalCite(citeResponse{
+					OK:     false,
+					Reason: "claim_empty",
+				}), nil
+			}
+			if urlStr == "" {
+				return marshalCite(citeResponse{
+					OK:     false,
+					Reason: "url_empty",
+				}), nil
+			}
+			if inv.RunID == "" {
+				// No run id → cite() can't verify history. Bail loud.
+				return marshalCite(citeResponse{
+					OK:     false,
+					Reason: "no_run_context",
+					Claim:  claim,
+					URL:    urlStr,
+				}), nil
+			}
+
+			touched, err := cs.GetTouchedURLs(ctx, inv.RunID)
+			if err != nil {
+				return marshalCite(citeResponse{
+					OK:     false,
+					Reason: fmt.Sprintf("touched_lookup_failed: %v", err),
+					Claim:  claim,
+					URL:    urlStr,
+				}), nil
+			}
+			if _, ok := touched[urlStr]; !ok {
+				return marshalCite(citeResponse{
+					OK:     false,
+					Reason: "url_not_in_run_history",
+					Claim:  claim,
+					URL:    urlStr,
+				}), nil
+			}
+
+			if err := cs.RecordCitation(ctx, inv.RunID, urlStr, claim); err != nil {
+				return marshalCite(citeResponse{
+					OK:     false,
+					Reason: fmt.Sprintf("record_failed: %v", err),
+					Claim:  claim,
+					URL:    urlStr,
+				}), nil
+			}
+			return marshalCite(citeResponse{
+				OK:    true,
+				Claim: claim,
+				URL:   urlStr,
+			}), nil
+		},
+	)
+}
+
+func marshalCite(r citeResponse) string {
+	b, err := json.Marshal(r)
+	if err != nil {
+		return fmt.Sprintf(`{"ok":false,"reason":"marshal_failed: %v"}`, err)
+	}
+	return string(b)
+}

tools/classify.go

@@ -0,0 +1,319 @@
+// Package tools — v12 classify.
+//
+// Classification primitive: text + categories → labels + per-category
+// scores. Single-label mode (default) returns the top-1 category;
+// multi-label mode returns every category whose score crosses the
+// threshold.
+//
+// Why a dedicated tool (vs reusing extract_entities for one-of-N
+// classification): classification has a typed result (labels[] +
+// scores{}) that downstream agents consume programmatically. Folding
+// it into extract_entities would force every author to re-spec the
+// scoring schema.
+//
+// Score normalisation: the LLM's reply is normalised so each score
+// lands in [0, 1]. The single-label result returns scores for ALL
+// categories so the author can read the distribution; multi-label
+// returns labels[] of categories above 0.5.
+//
+// Test: classify_test.go covers single-label, multi-label, score
+// normalisation, > 20 categories rejected, unknown category in the
+// reply silently dropped.
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/llmmeta"
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// classifyMaxInputBytes is the input cap.
+const classifyMaxInputBytes = 16 * 1024
+
+// classifyMaxCategories is the hard cap on category count.
+const classifyMaxCategories = 20
+
+// classifyMultiLabelThreshold is the score threshold above which a
+// category appears in the labels[] array in multi-label mode.
+const classifyMultiLabelThreshold = 0.5
+
+// classifyFallbackMaxPerRun is the per-run cap when ClassifyConfig is
+// nil.
+const classifyFallbackMaxPerRun = 20
+
+// ClassifyConfig is the narrow per-deployment config surface.
+type ClassifyConfig interface {
+	MaxPerRun(ctx context.Context) int
+}
+
+// classifyArgs is the LLM-facing param struct.
+type classifyArgs struct {
+	Text       string   `json:"text" description:"The text to classify. Required. Capped at 16KB."`
+	Categories []string `json:"categories" description:"List of categories to score the text against. Required. Max 20."`
+	MultiLabel bool     `json:"multi_label,omitempty" description:"When true, returns every category scoring above 0.5. Default false → single-label (top-1) result."`
+}
+
+type classifyResult struct {
+	Labels    []string           `json:"labels,omitempty"`
+	Scores    map[string]float64 `json:"scores,omitempty"`
+	ModelUsed string             `json:"model_used,omitempty"`
+	RawReply  string             `json:"raw_reply,omitempty"`
+	Error     string             `json:"error,omitempty"`
+	BudgetMsg string             `json:"budget_message,omitempty"`
+}
+
+// NewClassify constructs the classify tool.
+func NewClassify(helper *llmmeta.Helper, cfg ClassifyConfig, budget SearchBudget) tool.Tool {
+	return tool.NewGatedTool[classifyArgs](
+		"classify",
+		"Classify text into one of N categories (or multiple via multi_label=true). Returns labels[] (top-1 by default) + scores{category: 0..1}. Counts against per-run and 7-day cost budgets.",
+		tool.Permission{
+			AuthoringRequirement: tool.RequirementAnyone,
+			OperatesOn:           tool.ScopeCaller,
+			SafeForShare:         true,
+			Categories:           []string{"llm-meta", "cost-bearing"},
+		},
+		func(ctx context.Context, inv tool.Invocation, args classifyArgs) (string, error) {
+			if helper == nil {
+				return "", fmt.Errorf("classify: not configured")
+			}
+			text := args.Text
+			if strings.TrimSpace(text) == "" {
+				return marshalClassifyResult(classifyResult{Error: "text is empty"}), nil
+			}
+			if len(args.Categories) == 0 {
+				return marshalClassifyResult(classifyResult{Error: "categories is empty"}), nil
+			}
+			if len(args.Categories) > classifyMaxCategories {
+				return marshalClassifyResult(classifyResult{
+					Error: fmt.Sprintf("too many categories (%d > %d)", len(args.Categories), classifyMaxCategories),
+				}), nil
+			}
+			// Trim + dedupe categories so the LLM sees a clean
+			// schema. Order is preserved for the prompt; the result
+			// map is order-agnostic.
+			categories := make([]string, 0, len(args.Categories))
+			seen := make(map[string]bool, len(args.Categories))
+			for _, c := range args.Categories {
+				c = strings.TrimSpace(c)
+				if c == "" || seen[c] {
+					continue
+				}
+				seen[c] = true
+				categories = append(categories, c)
+			}
+			if len(categories) == 0 {
+				return marshalClassifyResult(classifyResult{Error: "categories has no non-empty entries"}), nil
+			}
+
+			if len(text) > classifyMaxInputBytes {
+				text = text[:classifyMaxInputBytes]
+			}
+
+			// Per-run budget gate.
+			if budget == nil {
+				maxPerRun := classifyFallbackMaxPerRun
+				if cfg != nil {
+					maxPerRun = cfg.MaxPerRun(ctx)
+				}
+				budget = NewInMemorySearchBudget(map[string]int{
+					"classify": maxPerRun,
+				})
+			}
+			count, max, exceeded := budget.CheckAndIncrement(ctx, inv.RunID, "classify")
+			if exceeded {
+				return marshalClassifyResult(classifyResult{
+					Error:     "classify_budget_exceeded",
+					BudgetMsg: fmt.Sprintf("per-run classify budget exceeded (%d/%d). Ask an admin to raise skills.classify.max_per_run.", count, max),
+				}), nil
+			}
+
+			systemPrompt := "You classify text into a fixed set of categories. Return ONLY JSON. Score each category in [0,1] (1 = perfect fit). Sum of all scores does NOT need to be 1 — high overlap across categories is allowed."
+			userPrompt := buildClassifyPrompt(text, categories, args.MultiLabel)
+
+			res, callErr := helper.Call(ctx, llmmeta.CallSpec{
+				Tier:                 "fast",
+				SystemPrompt:         systemPrompt,
+				UserPrompt:           userPrompt,
+				MaxOutputTokens:      2048,
+				ResponseFormat:       "json",
+				RetryOnMalformedJSON: true,
+				ToolName:             "classify",
+				RunID:                inv.RunID,
+				SkillID:              inv.SkillID,
+				CallerID:             inv.CallerID,
+			})
+			if callErr != nil {
+				return "", callErr
+			}
+			if !res.Success {
+				kind := res.ErrorKind
+				if kind == "" {
+					kind = "llm_unavailable"
+				}
+				return marshalClassifyResult(classifyResult{Error: kind}), nil
+			}
+			if res.ErrorKind == llmmeta.ErrorKindMalformedJSON || res.Parsed == nil {
+				return marshalClassifyResult(classifyResult{
+					Error:     "classification_failed",
+					RawReply:  res.Text,
+					ModelUsed: res.ModelUsed,
+				}), nil
+			}
+
+			parsedMap, ok := res.Parsed.(map[string]any)
+			if !ok {
+				return marshalClassifyResult(classifyResult{
+					Error:     "classification_failed_not_object",
+					RawReply:  res.Text,
+					ModelUsed: res.ModelUsed,
+				}), nil
+			}
+
+			scores := normaliseClassifyScores(parsedMap, categories)
+			labels := selectClassifyLabels(scores, categories, args.MultiLabel)
+
+			return marshalClassifyResult(classifyResult{
+				Labels:    labels,
+				Scores:    scores,
+				ModelUsed: res.ModelUsed,
+			}), nil
+		},
+	)
+}
+
+// buildClassifyPrompt composes the user message.
+func buildClassifyPrompt(text string, categories []string, multiLabel bool) string {
+	var sb strings.Builder
+	sb.WriteString("Classify the text below.\n\nCategories:\n")
+	for _, c := range categories {
+		sb.WriteString("- ")
+		sb.WriteString(c)
+		sb.WriteString("\n")
+	}
+	sb.WriteString("\nText:\n")
+	sb.WriteString(text)
+	sb.WriteString("\n\nReturn ONLY a JSON object: {\"scores\": {\"<category>\": <0..1 float>, ...}}.")
+	if multiLabel {
+		sb.WriteString(" The same text may score high in MULTIPLE categories — score each independently.")
+	} else {
+		sb.WriteString(" Score each category; the highest-scoring one will be the chosen label.")
+	}
+	return sb.String()
+}
+
+// normaliseClassifyScores extracts the scores map from the LLM's
+// reply and clamps each value into [0, 1]. Categories absent from the
+// reply default to 0.
+//
+// Why we accept either {"scores": {...}} or {...}: some models reply
+// with the inner object directly, dropping the wrapping key. Both
+// shapes are valid as long as the keys match the requested category
+// names.
+func normaliseClassifyScores(parsed map[string]any, categories []string) map[string]float64 {
+	scoresIn, ok := parsed["scores"].(map[string]any)
+	if !ok {
+		// Accept the bare-map shape too.
+		scoresIn = parsed
+	}
+	out := make(map[string]float64, len(categories))
+	for _, c := range categories {
+		v, has := scoresIn[c]
+		if !has {
+			out[c] = 0
+			continue
+		}
+		f, ok := coerceClassifyScore(v)
+		if !ok {
+			out[c] = 0
+			continue
+		}
+		// Clamp into [0, 1].
+		if f < 0 {
+			f = 0
+		}
+		if f > 1 {
+			f = 1
+		}
+		out[c] = f
+	}
+	return out
+}
+
+// coerceClassifyScore reads a JSON value as a float in [0, 1]. Accepts
+// floats, ints, and percent-strings ("85%" → 0.85).
+func coerceClassifyScore(raw any) (float64, bool) {
+	switch v := raw.(type) {
+	case float64:
+		return v, true
+	case int:
+		return float64(v), true
+	case int64:
+		return float64(v), true
+	case string:
+		s := strings.TrimSuffix(strings.TrimSpace(v), "%")
+		var f float64
+		if _, err := fmt.Sscanf(s, "%f", &f); err == nil {
+			if strings.HasSuffix(strings.TrimSpace(v), "%") {
+				f = f / 100.0
+			}
+			return f, true
+		}
+	}
+	return 0, false
+}
+
+// selectClassifyLabels picks the labels to surface. Single-label mode
+// returns the highest-scoring category. Multi-label returns every
+// category above the threshold (sorted by score desc for stable
+// rendering).
+func selectClassifyLabels(scores map[string]float64, categories []string, multiLabel bool) []string {
+	if multiLabel {
+		var labels []string
+		for _, c := range categories {
+			if scores[c] >= classifyMultiLabelThreshold {
+				labels = append(labels, c)
+			}
+		}
+		// Sort labels by score desc, then category-list order for ties.
+		sortClassifyLabelsByScore(labels, scores)
+		return labels
+	}
+	// Single-label: top-1.
+	bestCat := ""
+	bestScore := -1.0
+	for _, c := range categories {
+		if scores[c] > bestScore {
+			bestScore = scores[c]
+			bestCat = c
+		}
+	}
+	if bestCat == "" {
+		return nil
+	}
+	return []string{bestCat}
+}
+
+// sortClassifyLabelsByScore sorts labels desc by score. Stable on
+// ties (preserves category-list order).
+func sortClassifyLabelsByScore(labels []string, scores map[string]float64) {
+	for i := 1; i < len(labels); i++ {
+		j := i
+		for j > 0 && scores[labels[j]] > scores[labels[j-1]] {
+			labels[j], labels[j-1] = labels[j-1], labels[j]
+			j--
+		}
+	}
+}
+
+func marshalClassifyResult(r classifyResult) string {
+	b, err := json.Marshal(r)
+	if err != nil {
+		return fmt.Sprintf(`{"error":"marshal_failed: %v"}`, err)
+	}
+	return string(b)
+}

tools/extract_entities.go

@@ -0,0 +1,342 @@
+// Package tools — v12 extract_entities.
+//
+// Structured-output workhorse: text + field schema → typed JSON
+// object. The author specifies which fields they want and what
+// types; the tool builds an appropriate prompt, asks for JSON, and
+// validates + coerces the response back into the requested types.
+//
+// Why a structured-output tool (vs forcing the agent to write its
+// own prompt): every agentic skill that needs to "pull X, Y, Z out
+// of unstructured text" otherwise re-invents the same prompt-
+// engineering pattern. extract_entities centralises it so authors
+// just describe the schema.
+//
+// Type coercion: an LLM responding with "42" when an int field was
+// requested is normal noise. The tool coerces strings to
+// int/float/bool when possible; coercion failures land the field in
+// missing_fields rather than the entities map.
+//
+// Test: extract_entities_test.go covers happy path, missing optional
+// field, missing required field surfaces in missing_fields, malformed
+// JSON retry, second-attempt failure, type coercion (string→int,
+// string→bool), unknown field type rejected at args validation.
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strconv"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/llmmeta"
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// extractEntitiesMaxInputBytes is the hard input cap.
+const extractEntitiesMaxInputBytes = 32 * 1024
+
+// extractEntitiesFallbackMaxPerRun is the per-run cap when
+// ExtractEntitiesConfig is nil.
+const extractEntitiesFallbackMaxPerRun = 10
+
+// ExtractEntitiesConfig is the narrow per-deployment config surface
+// extract_entities reads at execute time.
+type ExtractEntitiesConfig interface {
+	MaxPerRun(ctx context.Context) int
+}
+
+// extractField is one row in the schema the agent supplies. The four
+// supported types match the JSON-shape primitives we can validate +
+// coerce reliably.
+//
+// Why an enum-shaped Type field (vs free-form): we need to know how
+// to validate the LLM's reply. Free-form ("integer", "Number",
+// "boolean") would invite typos that silently miss the validation.
+type extractField struct {
+	Name        string `json:"name" description:"Field name to populate (e.g. 'author', 'year_published'). Becomes a key in the returned entities object."`
+	Description string `json:"description" description:"Short description of what to extract (e.g. 'the book author', 'the year the article was published'). Helps the model find the right value."`
+	Type        string `json:"type" description:"One of: 'string', 'int', 'float', 'bool', 'list_of_strings'. Determines how the LLM's reply is validated and coerced."`
+	Required    bool   `json:"required,omitempty" description:"When true, a missing/uncoercible value lands in missing_fields rather than skipping silently."`
+}
+
+// extractEntitiesArgs is the LLM-facing param struct.
+type extractEntitiesArgs struct {
+	Text   string         `json:"text" description:"The text to extract from. Required. Capped at 32KB."`
+	Fields []extractField `json:"fields" description:"Schema describing what to extract. Each field has name, description, type, and optional required flag."`
+}
+
+type extractEntitiesResult struct {
+	Entities      map[string]any `json:"entities,omitempty"`
+	MissingFields []string       `json:"missing_fields,omitempty"`
+	ModelUsed     string         `json:"model_used,omitempty"`
+	RawReply      string         `json:"raw_reply,omitempty"`
+	Error         string         `json:"error,omitempty"`
+	BudgetMsg     string         `json:"budget_message,omitempty"`
+}
+
+// validExtractTypes is the closed set of Type strings the tool
+// accepts. Anything else is rejected at args validation.
+var validExtractTypes = map[string]bool{
+	"string":          true,
+	"int":             true,
+	"float":           true,
+	"bool":            true,
+	"list_of_strings": true,
+}
+
+// NewExtractEntities constructs the extract_entities tool.
+func NewExtractEntities(helper *llmmeta.Helper, cfg ExtractEntitiesConfig, budget SearchBudget) tool.Tool {
+	return tool.NewGatedTool[extractEntitiesArgs](
+		"extract_entities",
+		"Extract structured fields from unstructured text via a fast LLM. Caller supplies a schema (each field has name + description + type + required); tool returns an entities object with values matching the requested types. Types: string, int, float, bool, list_of_strings. Counts against per-run and 7-day cost budgets.",
+		tool.Permission{
+			AuthoringRequirement: tool.RequirementAnyone,
+			OperatesOn:           tool.ScopeCaller,
+			SafeForShare:         true,
+			Categories:           []string{"llm-meta", "cost-bearing"},
+		},
+		func(ctx context.Context, inv tool.Invocation, args extractEntitiesArgs) (string, error) {
+			if helper == nil {
+				return "", fmt.Errorf("extract_entities: not configured")
+			}
+			text := args.Text
+			if strings.TrimSpace(text) == "" {
+				return marshalExtractEntities(extractEntitiesResult{Error: "text is empty"}), nil
+			}
+			if len(args.Fields) == 0 {
+				return marshalExtractEntities(extractEntitiesResult{Error: "fields is empty"}), nil
+			}
+			// Validate each field's Type before paying for an LLM
+			// call.
+			for _, f := range args.Fields {
+				if strings.TrimSpace(f.Name) == "" {
+					return marshalExtractEntities(extractEntitiesResult{Error: "field with empty name"}), nil
+				}
+				if !validExtractTypes[strings.ToLower(strings.TrimSpace(f.Type))] {
+					return marshalExtractEntities(extractEntitiesResult{
+						Error: fmt.Sprintf("field %q has unsupported type %q (allowed: string|int|float|bool|list_of_strings)", f.Name, f.Type),
+					}), nil
+				}
+			}
+
+			if len(text) > extractEntitiesMaxInputBytes {
+				text = text[:extractEntitiesMaxInputBytes]
+			}
+
+			// Per-run budget gate.
+			if budget == nil {
+				maxPerRun := extractEntitiesFallbackMaxPerRun
+				if cfg != nil {
+					maxPerRun = cfg.MaxPerRun(ctx)
+				}
+				budget = NewInMemorySearchBudget(map[string]int{
+					"extract_entities": maxPerRun,
+				})
+			}
+			count, max, exceeded := budget.CheckAndIncrement(ctx, inv.RunID, "extract_entities")
+			if exceeded {
+				return marshalExtractEntities(extractEntitiesResult{
+					Error:     "extract_entities_budget_exceeded",
+					BudgetMsg: fmt.Sprintf("per-run extract_entities budget exceeded (%d/%d). Ask an admin to raise skills.extract_entities.max_per_run.", count, max),
+				}), nil
+			}
+
+			systemPrompt := "You extract structured data from unstructured text. Return ONLY valid JSON with the requested keys. If a value is not present in the text, omit the key. Do NOT invent values."
+			userPrompt := buildExtractPrompt(text, args.Fields)
+
+			res, callErr := helper.Call(ctx, llmmeta.CallSpec{
+				Tier:                 "fast",
+				SystemPrompt:         systemPrompt,
+				UserPrompt:           userPrompt,
+				MaxOutputTokens:      4096,
+				ResponseFormat:       "json",
+				RetryOnMalformedJSON: true,
+				ToolName:             "extract_entities",
+				RunID:                inv.RunID,
+				SkillID:              inv.SkillID,
+				CallerID:             inv.CallerID,
+			})
+			if callErr != nil {
+				return "", callErr
+			}
+			if !res.Success {
+				kind := res.ErrorKind
+				if kind == "" {
+					kind = "llm_unavailable"
+				}
+				return marshalExtractEntities(extractEntitiesResult{Error: kind}), nil
+			}
+
+			// Second-failure malformed JSON (success=true but parsed
+			// is nil and ErrorKind=malformed_json). Surface the raw
+			// reply so the agent can salvage.
+			if res.ErrorKind == llmmeta.ErrorKindMalformedJSON || res.Parsed == nil {
+				return marshalExtractEntities(extractEntitiesResult{
+					Error:     "extraction_failed",
+					RawReply:  res.Text,
+					ModelUsed: res.ModelUsed,
+				}), nil
+			}
+
+			parsedMap, ok := res.Parsed.(map[string]any)
+			if !ok {
+				return marshalExtractEntities(extractEntitiesResult{
+					Error:     "extraction_failed_not_object",
+					RawReply:  res.Text,
+					ModelUsed: res.ModelUsed,
+				}), nil
+			}
+
+			entities, missing := coerceExtractedEntities(parsedMap, args.Fields)
+			return marshalExtractEntities(extractEntitiesResult{
+				Entities:      entities,
+				MissingFields: missing,
+				ModelUsed:     res.ModelUsed,
+			}), nil
+		},
+	)
+}
+
+// buildExtractPrompt composes the user message describing the schema
+// + source text.
+func buildExtractPrompt(text string, fields []extractField) string {
+	var sb strings.Builder
+	sb.WriteString("Extract the following fields from the text below. Return a JSON object with the field names as keys.\n\nFields:\n")
+	for _, f := range fields {
+		fmt.Fprintf(&sb, "- %s (%s): %s", f.Name, f.Type, f.Description)
+		if f.Required {
+			sb.WriteString(" [required]")
+		}
+		sb.WriteString("\n")
+	}
+	sb.WriteString("\nText:\n")
+	sb.WriteString(text)
+	return sb.String()
+}
+
+// coerceExtractedEntities walks the LLM's response, validating + (when
+// possible) coercing each value to the requested type. Required fields
+// missing or uncoercible land in missing[]; optional fields silently
+// drop.
+func coerceExtractedEntities(parsed map[string]any, fields []extractField) (map[string]any, []string) {
+	entities := make(map[string]any, len(fields))
+	var missing []string
+	for _, f := range fields {
+		raw, present := parsed[f.Name]
+		if !present || raw == nil {
+			if f.Required {
+				missing = append(missing, f.Name)
+			}
+			continue
+		}
+		value, ok := coerceFieldValue(raw, f.Type)
+		if !ok {
+			if f.Required {
+				missing = append(missing, f.Name)
+			}
+			continue
+		}
+		entities[f.Name] = value
+	}
+	return entities, missing
+}
+
+// coerceFieldValue attempts to convert raw to the requested type.
+// Returns (value, true) on success or (nil, false) on failure.
+//
+// Why coerce (vs strict reject): LLMs frequently reply with strings
+// that contain numbers ("42") or pseudo-booleans ("yes"). Strict
+// rejection would force every author to clean the response themselves.
+// Coercion is conservative — string "42" → int 42 succeeds; string
+// "forty-two" → int 42 fails (the agent never asked for word-form
+// parsing).
+func coerceFieldValue(raw any, fieldType string) (any, bool) {
+	switch strings.ToLower(strings.TrimSpace(fieldType)) {
+	case "string":
+		switch v := raw.(type) {
+		case string:
+			return v, true
+		case float64:
+			return strconv.FormatFloat(v, 'f', -1, 64), true
+		case bool:
+			return strconv.FormatBool(v), true
+		}
+		return nil, false
+
+	case "int":
+		switch v := raw.(type) {
+		case float64:
+			// JSON numbers are float64 by default.
+			if v == float64(int64(v)) {
+				return int64(v), true
+			}
+			return nil, false
+		case string:
+			if n, err := strconv.ParseInt(strings.TrimSpace(v), 10, 64); err == nil {
+				return n, true
+			}
+			// Try float-string-with-zero-fractional ("42.0").
+			if f, err := strconv.ParseFloat(strings.TrimSpace(v), 64); err == nil && f == float64(int64(f)) {
+				return int64(f), true
+			}
+		}
+		return nil, false
+
+	case "float":
+		switch v := raw.(type) {
+		case float64:
+			return v, true
+		case string:
+			if f, err := strconv.ParseFloat(strings.TrimSpace(v), 64); err == nil {
+				return f, true
+			}
+		}
+		return nil, false
+
+	case "bool":
+		switch v := raw.(type) {
+		case bool:
+			return v, true
+		case string:
+			s := strings.ToLower(strings.TrimSpace(v))
+			switch s {
+			case "true", "yes", "1", "y":
+				return true, true
+			case "false", "no", "0", "n":
+				return false, true
+			}
+		case float64:
+			return v != 0, true
+		}
+		return nil, false
+
+	case "list_of_strings":
+		switch v := raw.(type) {
+		case []any:
+			out := make([]string, 0, len(v))
+			for _, e := range v {
+				if s, ok := e.(string); ok {
+					out = append(out, s)
+				} else {
+					// Mixed-type lists fail the type contract.
+					return nil, false
+				}
+			}
+			return out, true
+		case string:
+			// Single-string can be lifted into a one-element list.
+			return []string{v}, true
+		}
+		return nil, false
+	}
+	return nil, false
+}
+
+func marshalExtractEntities(r extractEntitiesResult) string {
+	b, err := json.Marshal(r)
+	if err != nil {
+		return fmt.Sprintf(`{"error":"marshal_failed: %v"}`, err)
+	}
+	return string(b)
+}

tools/file_storage.go

@@ -0,0 +1,49 @@
+// file_storage.go declares the narrow FileStorage interface that the
+// four v4 file tools (file_save, file_get, file_list, file_delete)
+// need at execute time.
+//
+// Why a narrow interface (vs importing pkg/logic/skills directly): same
+// cycle constraint as kv_storage.go — pkg/logic/skills imports
+// pkg/skilltools, so we mirror the FileMeta shape here and let
+// pkg/logic/mort.go adapt at wiring time.
+//
+// FileDomainMeta is field-for-field with skills.FileMeta; the production
+// adapter is a struct copy.
+package tools
+
+import (
+	"context"
+	"errors"
+	"time"
+)
+
+// FileStorage is the narrow surface file tools need from the skills
+// package. Production wiring (mort.go) bridges *skills.System.Storage().
+// nil-safe: tools constructed against a nil FileStorage surface "not
+// configured" at the first call.
+type FileStorage interface {
+	FileSave(ctx context.Context, meta FileDomainMeta, content []byte) (string, error)
+	FileGet(ctx context.Context, fileID string) (*FileDomainMeta, []byte, error)
+	FileList(ctx context.Context, skillID, scope string) ([]FileDomainMeta, error)
+	FileDelete(ctx context.Context, fileID string) error
+	FileUsageBytes(ctx context.Context, skillID string) (int64, error)
+}
+
+// FileDomainMeta mirrors skills.FileMeta. Field-for-field; the
+// production adapter is a struct copy.
+type FileDomainMeta struct {
+	ID          string // UUID, the public file_id
+	SkillID     string
+	Scope       string
+	Name        string
+	ContentHash string // SHA256 hex
+	MimeType    string
+	SizeBytes   int64
+	CreatedAt   time.Time
+}
+
+// ErrFileNotFound mirrors skills.ErrFileNotFound. The production
+// adapter returns this sentinel when wrapping a skills.ErrFileNotFound;
+// tools detect it with errors.Is to surface a "not_found" string to the
+// LLM rather than a generic error.
+var ErrFileNotFound = errors.New("file: not found")

tools/now.go

@@ -0,0 +1,101 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+	"time"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// nowParams is the LLM-facing param struct for current_time / now.
+//
+// Why optional `timezone`: most agent prompts know the user's local
+// timezone (it's in the chatbot's system prompt) but the agent has
+// no way to override on a per-call basis. An explicit arg lets a
+// research skill ask "what time is it in NYC for the user reading
+// this report?" without needing access to a member-config lookup
+// tool.
+type nowParams struct {
+	Timezone string `json:"timezone,omitempty" description:"Optional IANA timezone name (e.g. 'America/Chicago', 'Europe/London'). Defaults to the calling user's configured timezone, falling back to UTC."`
+}
+
+// nowResponse is the JSON envelope returned to the agent.
+//
+// Why a structured shape: the v1 tool returned a markdown blob.
+// Agents that needed just the year had to substring-parse, which
+// fails on locale variations. JSON lets the agent pick the field
+// it cares about.
+type nowResponse struct {
+	NowISO   string `json:"now_iso"`
+	NowHuman string `json:"now_human"`
+	Timezone string `json:"timezone"`
+	Weekday  string `json:"weekday"`
+	Year     int    `json:"year"`
+	Month    int    `json:"month"`
+	Day      int    `json:"day"`
+	Hour     int    `json:"hour"`
+	Minute   int    `json:"minute"`
+	Second   int    `json:"second"`
+	Warning  string `json:"warning,omitempty"`
+}
+
+// NewNow constructs the v11 current_time / now tool. The provider
+// supplies the calling member's configured timezone (per-user
+// localisation). nil falls back to UTC.
+//
+// V11 keeps the registered tool name "now" for back-compat with the
+// existing tool catalog tests AND adds the same tool surface under
+// the agent-facing description "current time". The design spec
+// called the tool "current_time" but the v1 registry already used
+// "now" — switching the registry name would break stored skills'
+// `tools` lists. Same name, expanded behaviour.
+func NewNow(provider CurrentTimeProvider) tool.Tool {
+	return tool.NewGatedTool[nowParams](
+		"now",
+		"Return the current time. Optional 'timezone' (IANA name e.g. 'America/Chicago'); defaults to the calling user's configured timezone or UTC. Returns ISO + human-readable formats plus structured year/month/day/weekday for time-relative reasoning. Use this BEFORE assuming a year — the agent's knowledge cut-off may differ from real time.",
+		tool.Permission{
+			AuthoringRequirement: tool.RequirementAnyone,
+			OperatesOn:           tool.ScopeGlobal,
+			SafeForShare:         true,
+			Categories:           []string{"utility"},
+		},
+		func(ctx context.Context, inv tool.Invocation, p nowParams) (string, error) {
+			tzName := strings.TrimSpace(p.Timezone)
+			warning := ""
+			if tzName == "" && provider != nil {
+				tzName = provider.UserTimezone(ctx, inv.CallerID)
+			}
+			if tzName == "" {
+				tzName = "UTC"
+			}
+			loc, err := time.LoadLocation(tzName)
+			if err != nil {
+				warning = fmt.Sprintf("unknown timezone %q; falling back to UTC", tzName)
+				tzName = "UTC"
+				loc = time.UTC
+			}
+			t := time.Now().In(loc)
+			out := nowResponse{
+				NowISO:   t.Format(time.RFC3339),
+				NowHuman: t.Format("Monday, January 2, 2006 at 3:04 PM MST"),
+				Timezone: tzName,
+				Weekday:  t.Weekday().String(),
+				Year:     t.Year(),
+				Month:    int(t.Month()),
+				Day:      t.Day(),
+				Hour:     t.Hour(),
+				Minute:   t.Minute(),
+				Second:   t.Second(),
+				Warning:  warning,
+			}
+			b, mErr := json.Marshal(out)
+			if mErr != nil {
+				return "", fmt.Errorf("now: marshal: %w", mErr)
+			}
+			return string(b), nil
+		},
+	)
+}

tools/research_defaults.go

@@ -0,0 +1,97 @@
+package tools
+
+import (
+	"context"
+	"sync"
+)
+
+// DefaultResearchConfig returns a ResearchConfig pinned to the v11
+// design defaults. Production wiring overrides via a convar-aware
+// adapter; tests use the defaults directly.
+func DefaultResearchConfig() ResearchConfig {
+	return defaultResearchConfig{}
+}
+
+type defaultResearchConfig struct{}
+
+func (defaultResearchConfig) MaxInlineBytes(_ context.Context) int            { return 12 * 1024 }
+func (defaultResearchConfig) PDFMaxPages(_ context.Context) int               { return 50 }
+func (defaultResearchConfig) WebSearchEnabled(_ context.Context) bool         { return true }
+func (defaultResearchConfig) WebSearchMaxPerRun(_ context.Context) int        { return 10 }
+func (defaultResearchConfig) ReadPageMaxPerRun(_ context.Context) int         { return 10 }
+func (defaultResearchConfig) VideoMaxPerRun(_ context.Context) int            { return 5 }
+func (defaultResearchConfig) VerifyURLMaxPerRun(_ context.Context) int        { return 20 }
+func (defaultResearchConfig) ReadPDFMaxPerRun(_ context.Context) int          { return 5 }
+func (defaultResearchConfig) HTTPGetMaxPerRun(_ context.Context) int          { return 20 }
+func (defaultResearchConfig) HTTPPostMaxPerRun(_ context.Context) int         { return 20 }
+func (defaultResearchConfig) WebSearchAugmentThreshold(_ context.Context) int { return 5 }
+
+// InMemorySearchBudget is the package-default SearchBudget — a
+// simple per-(run,kind) counter held in a map. NOT
+// production-correct because the map persists across the process
+// lifetime; production wiring MUST plug a per-run reset.
+//
+// Why a default at all: tests want a working SearchBudget without
+// rolling their own. Documenting the production-correctness gap
+// here keeps the production adapter (in mort.go) honest.
+type InMemorySearchBudget struct {
+	cap map[string]int // by kind; "" means "use Default"
+
+	mu     sync.Mutex
+	counts map[string]int // key = runID+"|"+kind
+}
+
+// NewInMemorySearchBudget constructs a default SearchBudget. Pass a
+// per-kind cap map (e.g. {"web_search": 10, "read_page": 10}); kinds
+// missing from the map fall back to maxPerKindDefault.
+func NewInMemorySearchBudget(caps map[string]int) *InMemorySearchBudget {
+	if caps == nil {
+		caps = map[string]int{}
+	}
+	return &InMemorySearchBudget{
+		cap:    caps,
+		counts: make(map[string]int),
+	}
+}
+
+// CheckAndIncrement implements SearchBudget. Returns the count AFTER
+// incrementing on success; the counter is NOT incremented when the
+// call would exceed the cap (so a "search_budget_exceeded" rejection
+// doesn't burn budget on retry).
+func (b *InMemorySearchBudget) CheckAndIncrement(_ context.Context, runID, kind string) (int, int, bool) {
+	max := b.cap[kind]
+	if max <= 0 {
+		max = 10 // safe default
+	}
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	key := runID + "|" + kind
+	cur := b.counts[key]
+	if cur >= max {
+		return cur, max, true
+	}
+	b.counts[key] = cur + 1
+	return cur + 1, max, false
+}
+
+// ResetRun is a test helper: clears the counters for a single run
+// across all kinds. Production wiring uses its own per-run lifecycle
+// (the executor's RunFinalizer interface).
+func (b *InMemorySearchBudget) ResetRun(runID string) {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	prefix := runID + "|"
+	for k := range b.counts {
+		if len(k) > len(prefix) && k[:len(prefix)] == prefix {
+			delete(b.counts, k)
+		}
+	}
+}
+
+// StaticTimeProvider is the package-default CurrentTimeProvider —
+// returns "" for every member (the tool then falls back to UTC).
+// Tests that need a specific timezone wire a one-line struct.
+type StaticTimeProvider struct{}
+
+// UserTimezone implements CurrentTimeProvider with a flat fallback to "".
+func (StaticTimeProvider) UserTimezone(_ context.Context, _ string) string { return "" }

tools/research_providers.go

@@ -0,0 +1,332 @@
+// Package tools — research provider plumbing for v11.
+//
+// This file declares the narrow interfaces v11's research tools
+// (web_search, read_page, read_video, read_pdf, verify_url, etc.) need
+// at execute time. Production wiring lives in pkg/logic/mort.go and
+// closes over the searcher chain, the extractor / chromedp client, the
+// PDF extractor, and the yt-dlp wrapper.
+//
+// Why narrow interfaces (vs importing pkg/logic/searcher / extractor
+// directly): the same cycle-break pattern used by KVStorage, FileStorage,
+// HTTPConfigProvider — keeps pkg/skilltools/tools free of the wiring
+// layer so tests can stub each dependency. Each provider is nil-safe:
+// the tool surfaces "not configured" at first call rather than failing
+// at registration.
+//
+// Test: each tool under pkg/skilltools/tools/ wired against these
+// interfaces has its own *_test.go using the in-package fakes in
+// research_providers_fakes_test.go.
+package tools
+
+import (
+	"context"
+	"errors"
+	"time"
+)
+
+// PageCache is the narrow surface read_page (and read_pdf) consult to
+// avoid re-fetching the same URL within the cache's TTL. Production
+// wiring bridges this interface to the legacy *cache.Cache held by
+// pkg/logic/query.System so a `.query foo.com` and a
+// `.skill query foo.com` for the same URL share one cache slot.
+//
+// Why a narrow interface (vs importing the cache package directly):
+// same cycle-break pattern as KVStorage / FileStorage / CitationStorage
+// — keeps pkg/skilltools/tools free of the wiring layer. The legacy
+// cache slot key is `sha256(url)`; the production adapter is
+// responsible for hashing so this interface stays clean (raw URL in/out)
+// and skill-tool authors never need to know the slot shape.
+//
+// nil-safe: a tool constructed with a nil PageCache simply skips the
+// cache layer (always treat Get as a miss; Set is a no-op).
+//
+// Test: tests pass a fake PageCache that records Get/Set calls and
+// returns canned hits. See page_cache_test.go for the read_page hit /
+// miss scenarios.
+type PageCache interface {
+	// Get returns the cached body for urlStr and true on hit, or
+	// (nil, false) on miss. Implementations MUST treat any backing-
+	// store error as a miss (best-effort, never fail the caller).
+	Get(ctx context.Context, urlStr string) ([]byte, bool)
+
+	// Set writes body under the slot for urlStr with the supplied TTL.
+	// Implementations MUST swallow backing-store errors (best-effort
+	// caching is correct: a write failure should not propagate to the
+	// agent loop).
+	Set(ctx context.Context, urlStr string, body []byte, ttl time.Duration)
+}
+
+// PageCacheTTL is the default TTL applied by tools that consult a
+// PageCache. Mirrors the legacy `query.pageCacheTTL` constant
+// (1 hour) so a `.query`-warmed slot reads back from a `.skill query`
+// (and vice versa) within the same window.
+//
+// Tools that want a different TTL pass an explicit value to
+// PageCache.Set; this constant is the project default the v11 / v-research
+// tools all use.
+const PageCacheTTL = 1 * time.Hour
+
+// PageExtractor is the narrow surface read_page needs at execute
+// time. The production adapter wraps mort's existing extractor
+// (Ollama web_fetch first, chromedp fallback on JS-heavy pages).
+//
+// nil-safe: a tool constructed with a nil PageExtractor surfaces
+// "not configured" at first call.
+//
+// Why: read_page used to be a thin io.ReadAll over the URL — it
+// missed JS rendering, didn't honour the v6 page cache, and could
+// not surface the underlying provider name. v11 routes through this
+// interface so the production wiring (mort.go) can plug in the
+// existing query-side extractor without exposing query.Agent.
+type PageExtractor interface {
+	// ExtractPage fetches and extracts readable text from urlStr.
+	// Returns the extracted body, a final URL (after any redirects
+	// the extractor followed), the provider name ("ollama" |
+	// "chromedp" | "ytdlp"), and an error.
+	//
+	// The returned body is the FULL extracted text — callers apply
+	// the v10 byte-vs-reference cap before surfacing to the agent.
+	//
+	// bypassCache=true skips any page cache and forces a fresh
+	// extraction. Default false.
+	ExtractPage(ctx context.Context, urlStr string, bypassCache bool) (text string, finalURL string, provider string, err error)
+}
+
+// VideoTranscriber is the narrow surface read_video needs at
+// execute time. Production wiring wraps internal/ytdlp.
+//
+// nil-safe: tool surfaces "not configured" at first call.
+//
+// Why a separate interface from PageExtractor: video is a different
+// shape (transcript + metadata) and a different binary (yt-dlp).
+// Keeping them distinct lets tests stub each independently.
+type VideoTranscriber interface {
+	// ExtractVideoTranscript returns the transcript text and the
+	// best-effort metadata (title, duration in seconds, channel).
+	// Implementations MUST return a non-empty transcript or an
+	// error — empty-transcript success is interpreted by the tool
+	// as a "transcript_unavailable" failure.
+	ExtractVideoTranscript(ctx context.Context, urlStr string) (transcript string, meta VideoMeta, err error)
+}
+
+// VideoMeta is best-effort metadata returned alongside a video
+// transcript. Any field may be empty/zero if the implementation
+// could not extract it.
+type VideoMeta struct {
+	Title           string
+	Channel         string
+	DurationSeconds int
+}
+
+// PDFFetcher is the narrow surface read_pdf needs at execute time.
+// Production wiring uses an HTTP-aware fetcher that HEAD-validates
+// content-type before downloading the body.
+//
+// nil-safe: tool surfaces "not configured" at first call.
+//
+// Why: a tool that just embedded PDF extraction would couple
+// fetching + parsing. Splitting the fetch (allowlist + SSRF +
+// HEAD check) from the extract (page-level parsing) keeps each
+// step testable and lets the same fetcher serve verify_url one
+// day if we want a PDF-aware fast path.
+type PDFFetcher interface {
+	// FetchPDF downloads the PDF at urlStr (after HEAD-validating
+	// content-type) and returns the raw bytes plus the final URL.
+	// HEAD-validation rejects a URL whose Content-Type is not a
+	// PDF mime AND whose path does not end in .pdf.
+	FetchPDF(ctx context.Context, urlStr string) (body []byte, finalURL string, err error)
+}
+
+// PDFExtractor parses PDF bytes into plain text + page count.
+// Production wires internal.ExtractPDFText.
+//
+// Why split from PDFFetcher: tests want to vary the fetch (mock
+// server returning bytes) without rebuilding the extractor.
+type PDFExtractor interface {
+	// ExtractPDFText returns the concatenated plain-text content
+	// of the PDF along with the page count. The caller applies any
+	// per-page cap and the v10 byte-vs-reference cap on the result.
+	ExtractPDFText(ctx context.Context, body []byte, maxPages int) (text string, pageCount int, truncated bool, err error)
+}
+
+// HEADChecker is the narrow surface verify_url needs at execute
+// time. Production wiring uses the same SSRF-pinned transport as
+// http_get so the security envelope is consistent.
+//
+// Why a separate interface (vs reusing HTTPConfigProvider+doHTTP):
+// verify_url's contract is simpler — HEAD only, no body bytes
+// returned, and the agent only cares about reachable / status /
+// final URL / content-type. A bespoke surface lets the production
+// adapter optimise for that path (no body buffer, no body close).
+type HEADChecker interface {
+	// HEAD performs a HEAD request against urlStr (with SSRF +
+	// allowlist enforcement) and returns the final URL after any
+	// redirects, the HTTP status code, and the Content-Type header.
+	// Returns reachable=false with a non-nil err for transport
+	// failures (DNS, TCP, allowlist rejection); reachable=true with
+	// any HTTP status (including 4xx/5xx) is the success shape —
+	// the agent decides whether the URL is "real".
+	HEAD(ctx context.Context, urlStr string) (finalURL string, status int, contentType string, reachable bool, err error)
+}
+
+// CitationStorage is the narrow surface cite() needs at execute
+// time. Production wires *skills.System.Storage(); tests stub.
+//
+// nil-safe: tool surfaces "not configured" at first call.
+//
+// Why a narrow interface (vs importing pkg/logic/skills): same
+// cycle constraint as KVStorage / FileStorage. Production adapter
+// in mort.go bridges to skills.Storage's RecordCitation /
+// ListCitations methods AND a separate URL-history tracker.
+//
+// Two responsibilities, deliberately separate:
+//
+//  1. RecordCitation writes a row into skill_run_sources — this is
+//     the user-visible citations table for the Sources panel and
+//     CSV export. ONLY rows the agent successfully cited via
+//     cite() land here.
+//  2. RecordURLTouch / GetTouchedURLs maintains a per-run set of
+//     URLs the agent has interacted with (web_search results,
+//     read_page input, read_pdf input, read_video input). cite()
+//     reads this set to reject claims for URLs the agent never
+//     touched. This set lives in a different table or scope from
+//     the citations table — it's working state, not a record.
+type CitationStorage interface {
+	// RecordCitation appends one (run_id, url, claim, cited_at)
+	// row to the citations table (skill_run_sources). cited_at is
+	// set by the storage layer to time.Now() when zero. The caller
+	// has already verified the URL is in the touched-URL set
+	// (via GetTouchedURLs); this method is the persistence step.
+	RecordCitation(ctx context.Context, runID, url, claim string) error
+
+	// RecordURLTouch records that the agent has interacted with
+	// `url` during `runID`. Called by web_search (per result),
+	// read_page, read_pdf, and read_video. Idempotent — repeat
+	// calls for the same (run_id, url) are no-ops at the storage
+	// layer.
+	RecordURLTouch(ctx context.Context, runID, url string) error
+
+	// GetTouchedURLs returns the set of URLs the run has
+	// interacted with. Used by cite() to verify that a claim's
+	// URL is one the agent actually visited. Empty for a fresh
+	// run — cite() then rejects every claim with
+	// "url_not_in_run_history".
+	GetTouchedURLs(ctx context.Context, runID string) (map[string]struct{}, error)
+
+	// ListCitations returns all citations recorded for the run, in
+	// insertion order. Powers the /skills/{id}/runs/{run_id}
+	// Sources panel.
+	ListCitations(ctx context.Context, runID string) ([]CitationRow, error)
+}
+
+// CitationRow mirrors the skill_run_sources row shape. Fields
+// match the spec: run_id is implicit in the query, url + claim are
+// what the agent submitted, cited_at is the wall-clock timestamp
+// at insert.
+type CitationRow struct {
+	URL     string
+	Claim   string
+	CitedAt int64 // unix-seconds; storage adapter normalises from time.Time
+}
+
+// CurrentTimeProvider exposes a "now" + per-user timezone lookup.
+// Production wiring closes over the bot's member-config getter.
+//
+// nil-safe: a tool constructed with a nil provider falls back to
+// server-time + UTC (current behaviour of NewNow before v11).
+type CurrentTimeProvider interface {
+	// UserTimezone returns the IANA timezone name configured for
+	// the given Discord member ID, or "" when the member has no
+	// timezone configured. Empty fallback is "UTC".
+	UserTimezone(ctx context.Context, memberID string) string
+}
+
+// SearchBudget is the narrow surface web_search reads at execute
+// time to honour skills.web_search.max_per_run.
+//
+// Production wiring closes over a per-run counter held by the
+// executor. nil-safe: tool falls back to a built-in package
+// counter (process-wide, NOT per-run) — useful for tests but NOT
+// production-correct because budget bleeds across runs. The
+// production adapter MUST be wired.
+type SearchBudget interface {
+	// CheckAndIncrement returns the current count AFTER incrementing
+	// for the given runID, the configured max, and an error when
+	// the call would exceed the cap. The handler returns a clean
+	// "search_budget_exceeded" string on exceed (not an error so
+	// the agent can react).
+	CheckAndIncrement(ctx context.Context, runID, kind string) (count, max int, exceeded bool)
+}
+
+// ResearchConfig is the narrow surface that read_page / read_video /
+// read_pdf / verify_url read at execute time for per-tool budget caps
+// and inline-vs-file_id thresholds. Production wiring closes over
+// the relevant convars.
+//
+// nil-safe: tools fall back to package defaults.
+type ResearchConfig interface {
+	// MaxInlineBytes returns the cap above which extracted text is
+	// persisted as a file_id under run-scope (v10 byte-vs-reference
+	// principle). Default 12 KiB.
+	MaxInlineBytes(ctx context.Context) int
+
+	// PDFMaxPages returns the cap on pages extracted from a PDF
+	// before truncation. Default 50.
+	PDFMaxPages(ctx context.Context) int
+
+	// WebSearchEnabled is the master switch for web_search.
+	WebSearchEnabled(ctx context.Context) bool
+
+	// WebSearchMaxPerRun is the per-run search cap.
+	WebSearchMaxPerRun(ctx context.Context) int
+
+	// ReadPageMaxPerRun is the per-run page-read cap.
+	ReadPageMaxPerRun(ctx context.Context) int
+
+	// VideoMaxPerRun is the per-run video-read cap.
+	VideoMaxPerRun(ctx context.Context) int
+
+	// VerifyURLMaxPerRun is the per-run HEAD-check cap.
+	VerifyURLMaxPerRun(ctx context.Context) int
+
+	// ReadPDFMaxPerRun is the per-run PDF-read cap.
+	ReadPDFMaxPerRun(ctx context.Context) int
+
+	// HTTPGetMaxPerRun (v15.2) is the per-run http_get cap. The agent
+	// otherwise can retry-storm through random URLs and bloat its own
+	// context with each tool result. Default 20.
+	HTTPGetMaxPerRun(ctx context.Context) int
+
+	// HTTPPostMaxPerRun (v15.2) is the per-run http_post cap. Default 20.
+	HTTPPostMaxPerRun(ctx context.Context) int
+
+	// WebSearchAugmentThreshold is the minimum number of primary
+	// (Ollama) results required to skip the secondary (DDG/Brave)
+	// search. When the primary backend returns fewer than this many
+	// results, the augmented searcher also queries the secondary and
+	// merges both result sets. Default 5.
+	WebSearchAugmentThreshold(ctx context.Context) int
+
+	// ReplyChainDepthMax is unused here; placeholder shape for
+	// future per-tool caps. Kept off this interface — callers reach
+	// into the convar reader directly when they need it.
+}
+
+// ErrPageExtractionFailed is the sentinel returned by a PageExtractor
+// when both Ollama and chromedp paths produce empty content.
+var ErrPageExtractionFailed = errors.New("page extraction failed: empty content")
+
+// ErrVideoTranscriptUnavailable is the sentinel returned by a
+// VideoTranscriber when no captions / transcript could be obtained.
+var ErrVideoTranscriptUnavailable = errors.New("video transcript unavailable")
+
+// ErrPDFNotPDF is the sentinel returned by a PDFFetcher when the
+// HEAD response indicates a non-PDF content-type AND the URL path
+// has no .pdf extension. Surfaces a clean "url_is_not_a_pdf"
+// rejection rather than a generic transport error.
+var ErrPDFNotPDF = errors.New("url does not serve a PDF")
+
+// ErrPDFEncrypted is returned by a PDFExtractor when the PDF refuses
+// extraction because it is password-protected. Surfaces a clean
+// "pdf_encrypted" rejection.
+var ErrPDFEncrypted = errors.New("pdf is encrypted")

tools/scope_validate.go

@@ -0,0 +1,113 @@
+// scope_validate.go centralises the storage-scope authorisation check
+// shared by every v4 KV and file tool. It enforces:
+//
+//   - "skill"           — always allowed (the skill's shared, cross-caller area).
+//   - "user:<callerID>" — allowed if it matches inv.CallerID (or admin).
+//   - "user:<other>"    — allowed only for admin callers.
+//   - "run:<runID>"     — allowed if it matches inv.RunID (or admin).
+//   - "run:<other>"     — allowed only for admin callers.
+//   - "root_run:<id>"   — allowed if it matches inv.RootRunID (or admin):
+//     the dispatch tree's SHARED scratchpad, readable
+//     and writable by every run under one root
+//     (parallel sibling workers coordinate here).
+//   - any other shape   — rejected with a descriptive error.
+//
+// Why a single helper (vs inline checks in each tool): the parsing rules
+// must match exactly across kv_get/set/list/delete and file_save/get/
+// list/delete. Centralising them means one place to fix when the
+// vocabulary evolves and one place for the test matrix.
+//
+// Why the isAdmin parameter: the v4 Invocation does NOT carry an
+// admin flag — production tools always pass isAdmin=false. The
+// parameter exists for tests (which exercise the admin paths) and for a
+// future Invocation extension that adds an admin signal without
+// breaking this helper's signature.
+package tools
+
+import (
+	"fmt"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// ValidateScope rejects scope strings the caller is not authorised to
+// access. See file-level doc for the exact ruleset.
+//
+// Why isAdmin is parameterised: tests pass true to verify admin paths;
+// production tools currently always pass false because Invocation
+// doesn't carry admin status. The gate is "you can access your own
+// scope only" until a future extension threads an admin signal through
+// the executor.
+func ValidateScope(inv tool.Invocation, scope string, isAdmin bool) error {
+	if scope == "skill" {
+		return nil
+	}
+	if rest, ok := strings.CutPrefix(scope, "user:"); ok {
+		if rest == "" {
+			return fmt.Errorf("scope: empty user id after 'user:'")
+		}
+		if rest == inv.CallerID {
+			return nil
+		}
+		if isAdmin {
+			return nil
+		}
+		return fmt.Errorf("scope %q: cannot access another user's storage", scope)
+	}
+	if rest, ok := strings.CutPrefix(scope, "root_run:"); ok {
+		if rest == "" {
+			return fmt.Errorf("scope: empty run id after 'root_run:'")
+		}
+		// The dispatch tree's shared scratchpad. Every run in one tree
+		// carries the same RootRunID (stamped by both executors from the
+		// dispatchguard chain), so siblings spawned in parallel — even
+		// ephemeral workers with distinct agent IDs — validate against
+		// the same scope string. Storage-side, root_run scopes live in
+		// the shared RootRunKVPartition; this check is the isolation
+		// boundary between trees.
+		if rest == inv.RootRunID && inv.RootRunID != "" {
+			return nil
+		}
+		if isAdmin {
+			return nil
+		}
+		return fmt.Errorf("scope %q: cannot access another dispatch tree's storage", scope)
+	}
+	if rest, ok := strings.CutPrefix(scope, "run:"); ok {
+		if rest == "" {
+			return fmt.Errorf("scope: empty run id after 'run:'")
+		}
+		if rest == inv.RunID {
+			return nil
+		}
+		// V10: when this run is a reply continuation, the agent may
+		// access the PARENT run's scope. The parent's run-scope KV is
+		// the natural carrier for "ask user a question, save state,
+		// resume on reply" — without this access, every continuation
+		// would have to re-derive state from parent_output alone.
+		// Note: the parent's run-scope is subject to the v4
+		// auto-purge (24h after parent finished). Long-delayed replies
+		// will see an empty scope.
+		if inv.Continuation != nil && rest == inv.Continuation.ParentRunID {
+			return nil
+		}
+		// V14: when this run is invoked via skill_invoke /
+		// skill_invoke_parallel from a parent skill, the agent may
+		// access the PARENT run's scope. This is the natural carrier
+		// for the "scout fans out, parent reads consolidated state"
+		// pattern that deepresearch uses — research-scout writes its
+		// touched-URL list under run:<parent_run_id> and the parent
+		// reads it back during the investigate phase. Without this
+		// access, every parent/child handoff would have to be
+		// serialised through tool-result strings.
+		if inv.ParentRunID != "" && rest == inv.ParentRunID {
+			return nil
+		}
+		if isAdmin {
+			return nil
+		}
+		return fmt.Errorf("scope %q: cannot access another run's storage", scope)
+	}
+	return fmt.Errorf("scope %q: unknown shape; expected 'skill', 'user:<id>', 'run:<id>', or 'root_run:<id>'", scope)
+}

tools/summarize.go

@@ -0,0 +1,243 @@
+// Package tools — v12 summarize.
+//
+// One fast-tier LLM call: text in → concise text summary out. Either
+// `text` or `file_id` (mutually exclusive) supplies the source. Per-run
+// budget enforced via the existing v11 SearchBudget surface (kind=
+// "summarize"); per-skill cost accounting via the meta-LLM helper's
+// ledger (skill_llm_meta_calls).
+//
+// Why a dedicated tool (vs reusing summary_summarise): summary_
+// summarise wraps the URL-summary pipeline used by /summary; it's
+// over-coupled to a specific extraction flow. v12's summarize is the
+// "given any text, give me a summary" primitive that downstream tools
+// (read_page → summarize, extract → summarize) can compose freely.
+//
+// File-id input path: when the caller supplies file_id, we dereference
+// via FileStorage. Cross-skill check rejects stolen IDs (matching
+// file_get's pattern). Scope check denies user:bob's file from alice's
+// invocation.
+//
+// Test: summarize_test.go covers happy path (mock helper), file_id
+// input, oversize input truncation, budget exceeded, focus-arg
+// pass-through, cross-skill file_id rejection, and the
+// missing-both-args validation.
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/llmmeta"
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// summarizeMaxInputBytes is the hard input cap. Inputs longer than
+// this are truncated with a `truncated=true` flag in the response so
+// the agent knows the summary covers a prefix.
+const summarizeMaxInputBytes = 32 * 1024
+
+// summarizeDefaultMaxWords is the default max_words when the caller
+// doesn't supply one. Capped further by skills.summarize.max_words.
+const summarizeDefaultMaxWords = 200
+
+// summarizeFallbackMaxWords is the cap used when SummarizeConfig is nil.
+const summarizeFallbackMaxWords = 1000
+
+// summarizeFallbackMaxPerRun is the per-run cap used when SummarizeConfig
+// is nil.
+const summarizeFallbackMaxPerRun = 10
+
+// SummarizeConfig is the narrow per-run + per-deployment config surface
+// summarize reads at execute time. Production wires a closure over the
+// `skills.summarize.*` convars; nil falls back to package defaults.
+type SummarizeConfig interface {
+	MaxPerRun(ctx context.Context) int
+	MaxWords(ctx context.Context) int
+}
+
+// summarizeArgs is the LLM-facing param struct.
+//
+// Why two source fields (text + file_id) with exactly-one validation:
+// the agent often produces large content via read_page / read_pdf and
+// stores it as a file_id (per the v10 byte-vs-reference principle);
+// forcing it to round-trip through a string would defeat the file_id
+// pattern. Inline `text` is the simpler path for short snippets.
+type summarizeArgs struct {
+	Text     string `json:"text,omitempty" description:"The text to summarise. Either 'text' OR 'file_id' is required (not both). Capped at 32KB; longer inputs truncate with truncated=true in the result."`
+	FileID   string `json:"file_id,omitempty" description:"Alternative to 'text': summarise the contents of a saved file (from read_page/read_pdf/file_save). Must belong to this skill."`
+	MaxWords int    `json:"max_words,omitempty" description:"Maximum word count for the summary. Default 200, capped at skills.summarize.max_words (default 1000)."`
+	Focus    string `json:"focus,omitempty" description:"Optional: what aspect to emphasise (e.g. 'security implications', 'cost analysis', 'main characters')."`
+}
+
+type summarizeResult struct {
+	Summary   string `json:"summary"`
+	WordCount int    `json:"word_count"`
+	ModelUsed string `json:"model_used"`
+	Truncated bool   `json:"truncated,omitempty"`
+	BudgetMsg string `json:"budget_message,omitempty"`
+	Error     string `json:"error,omitempty"`
+}
+
+// NewSummarize constructs the summarize tool. helper / cfg / budget /
+// fileStorage may all be nil; the handler surfaces clean errors at
+// first call.
+func NewSummarize(helper *llmmeta.Helper, cfg SummarizeConfig, budget SearchBudget, fileStorage FileStorage) tool.Tool {
+	return tool.NewGatedTool[summarizeArgs](
+		"summarize",
+		"Produce a concise summary of input text using a fast LLM. Pass either 'text' or 'file_id' (one of them is required). Optional 'focus' steers the summary; 'max_words' caps length (default 200). Counts against per-run and 7-day cost budgets.",
+		tool.Permission{
+			AuthoringRequirement: tool.RequirementAnyone,
+			OperatesOn:           tool.ScopeCaller,
+			SafeForShare:         true,
+			Categories:           []string{"llm-meta", "cost-bearing"},
+		},
+		func(ctx context.Context, inv tool.Invocation, args summarizeArgs) (string, error) {
+			if helper == nil {
+				return "", fmt.Errorf("summarize: not configured")
+			}
+			text, truncated, err := loadSummarizeInput(ctx, inv, args, fileStorage)
+			if err != nil {
+				return marshalSummarizeResult(summarizeResult{Error: err.Error()}), nil
+			}
+
+			// Per-run budget BEFORE the LLM call so a runaway loop is
+			// bounded.
+			if budget == nil {
+				maxPerRun := summarizeFallbackMaxPerRun
+				if cfg != nil {
+					maxPerRun = cfg.MaxPerRun(ctx)
+				}
+				budget = NewInMemorySearchBudget(map[string]int{
+					"summarize": maxPerRun,
+				})
+			}
+			count, max, exceeded := budget.CheckAndIncrement(ctx, inv.RunID, "summarize")
+			if exceeded {
+				return marshalSummarizeResult(summarizeResult{
+					Error:     "summarize_budget_exceeded",
+					BudgetMsg: fmt.Sprintf("per-run summarize budget exceeded (%d/%d). Work with the summaries you already have, or ask an admin to raise skills.summarize.max_per_run.", count, max),
+				}), nil
+			}
+
+			maxWords := args.MaxWords
+			if maxWords <= 0 {
+				maxWords = summarizeDefaultMaxWords
+			}
+			cap := summarizeFallbackMaxWords
+			if cfg != nil {
+				cap = cfg.MaxWords(ctx)
+			}
+			if maxWords > cap {
+				maxWords = cap
+			}
+
+			systemPrompt := "You produce concise, accurate summaries. Honor the requested word count. Do NOT invent facts."
+			userPrompt := buildSummarizePrompt(text, maxWords, args.Focus)
+
+			res, callErr := helper.Call(ctx, llmmeta.CallSpec{
+				Tier:            "fast",
+				SystemPrompt:    systemPrompt,
+				UserPrompt:      userPrompt,
+				MaxOutputTokens: maxWords * 8, // ~8 tokens per word upper bound
+				ResponseFormat:  "text",
+				ToolName:        "summarize",
+				RunID:           inv.RunID,
+				SkillID:         inv.SkillID,
+				CallerID:        inv.CallerID,
+			})
+			if callErr != nil {
+				return "", callErr
+			}
+			if !res.Success || res.Text == "" {
+				kind := res.ErrorKind
+				if kind == "" {
+					kind = "llm_unavailable"
+				}
+				return marshalSummarizeResult(summarizeResult{Error: kind}), nil
+			}
+			summary := strings.TrimSpace(res.Text)
+			return marshalSummarizeResult(summarizeResult{
+				Summary:   summary,
+				WordCount: countWords(summary),
+				ModelUsed: res.ModelUsed,
+				Truncated: truncated,
+			}), nil
+		},
+	)
+}
+
+// loadSummarizeInput resolves the input text from either args.Text or
+// args.FileID. Exactly one MUST be supplied; both empty AND both
+// populated are rejected.
+func loadSummarizeInput(ctx context.Context, inv tool.Invocation, args summarizeArgs, fileStorage FileStorage) (string, bool, error) {
+	hasText := strings.TrimSpace(args.Text) != ""
+	hasFile := strings.TrimSpace(args.FileID) != ""
+	if hasText == hasFile {
+		// Both empty OR both populated.
+		if !hasText {
+			return "", false, fmt.Errorf("summarize: one of 'text' or 'file_id' is required")
+		}
+		return "", false, fmt.Errorf("summarize: 'text' and 'file_id' are mutually exclusive — pass one")
+	}
+	if hasText {
+		return capInput(args.Text)
+	}
+	if fileStorage == nil {
+		return "", false, fmt.Errorf("summarize: file_id input requires file storage to be configured")
+	}
+	meta, content, err := fileStorage.FileGet(ctx, args.FileID)
+	if err != nil {
+		if errors.Is(err, ErrFileNotFound) {
+			return "", false, fmt.Errorf("summarize: file_id not found")
+		}
+		return "", false, fmt.Errorf("summarize: file fetch: %w", err)
+	}
+	if meta.SkillID != inv.SkillID {
+		return "", false, fmt.Errorf("summarize: file does not belong to this skill")
+	}
+	if err := ValidateScope(inv, meta.Scope, false); err != nil {
+		return "", false, fmt.Errorf("summarize: %w", err)
+	}
+	return capInput(string(content))
+}
+
+// capInput truncates input to the hard byte cap, returning the
+// (possibly truncated) text and a flag indicating truncation occurred.
+func capInput(text string) (string, bool, error) {
+	if len(text) <= summarizeMaxInputBytes {
+		return text, false, nil
+	}
+	return text[:summarizeMaxInputBytes], true, nil
+}
+
+// buildSummarizePrompt composes the user message handed to the LLM.
+func buildSummarizePrompt(text string, maxWords int, focus string) string {
+	var sb strings.Builder
+	fmt.Fprintf(&sb, "Summarise the following text in at most %d words.", maxWords)
+	if focus = strings.TrimSpace(focus); focus != "" {
+		fmt.Fprintf(&sb, " Emphasise: %s.", focus)
+	}
+	sb.WriteString("\n\n")
+	sb.WriteString(text)
+	return sb.String()
+}
+
+// countWords returns a rough word count via whitespace splitting.
+// Good enough for the response's word_count column; the agent might
+// see slight discrepancies vs the LLM's internal counter, which is
+// acceptable.
+func countWords(text string) int {
+	return len(strings.Fields(text))
+}
+
+// marshalSummarizeResult serialises a summarizeResult to JSON.
+func marshalSummarizeResult(r summarizeResult) string {
+	b, err := json.Marshal(r)
+	if err != nil {
+		return fmt.Sprintf(`{"error":"marshal_failed: %v"}`, err)
+	}
+	return string(b)
+}

tools/tools.go

@@ -3,28 +3,87 @@
 // A host registers the tools it wants against a tool.Registry, then runs an
 // agent whose RunnableAgent.LowLevelTools name them. Tools split two ways:
 //
-//   - Always-available, zero-dependency tools (think, ...) need no host backend
+//   - Always-available, zero-configuration tools register via Register (think,
-//     and register via Register. A light host (gadfly) can call Register and be
+//     now, cite) — all nil-safe, so a light host (gadfly) calls Register and is
 //     immediately useful.
-//   - Backed tools (web search, file/kv storage, summarize, ...) take a nil-safe
+//   - Backed tools take a nil-safe Deps describing their host backend and
-//     Deps describing their host backend; they register via grouped registrars
+//     register via grouped registrars (RegisterMeta, and RegisterWeb/Store/…
-//     (RegisterWeb, RegisterStore, ...) as those land.
+//     as they land). Each Deps ships sensible defaults so "some setup" is small.
 //
-// Every tool ships with the same three-stage permission model as mort's, and a
+// A host adds its own domain tools against the SAME registry.
-// host adds its own domain tools against the SAME registry.
 package tools
 
-import "gitea.stevedudenhoeffer.com/steve/executus/tool"
+import (
+	"context"
+	"errors"
 
-// Register adds the always-available, zero-dependency generic tools to reg
+	"gitea.stevedudenhoeffer.com/steve/executus/llmmeta"
-// (currently: think). Returns the first registration error, if any.
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// Register adds the always-available, zero-configuration generic tools:
+//
+//   - think  — record a thought to the run trace (no external effect)
+//   - now    — current time (UTC unless a CurrentTimeProvider is wired)
+//   - cite   — record a source citation (inert unless a CitationStorage is wired)
+//
+// All are nil-safe. Returns the first registration error.
 func Register(reg tool.Registry) error {
-	for _, t := range []tool.Tool{
+	return registerAll(reg,
 		NewThink(),
-	} {
+		NewNow(nil),
+		NewCite(nil),
+	)
+}
+
+// MetaDeps wires the LLM-backed meta tools (classify, extract_entities,
+// summarize). Helper is required. Budget defaults to an in-memory per-run cap;
+// Files is optional (summarize's file_id input is inert without it); MaxPerRun
+// and MaxWords default when non-positive.
+type MetaDeps struct {
+	Helper    *llmmeta.Helper
+	Budget    SearchBudget
+	Files     FileStorage
+	MaxPerRun int // per-run cap for each meta tool; default 10
+	MaxWords  int // summarize length cap; default 200
+}
+
+// RegisterMeta adds classify, extract_entities, and summarize. It requires a
+// configured llmmeta.Helper (the fast-tier meta-LLM caller); everything else
+// defaults.
+func RegisterMeta(reg tool.Registry, d MetaDeps) error {
+	if d.Helper == nil {
+		return errors.New("tools: MetaDeps.Helper is required for the meta tools")
+	}
+	if d.Budget == nil {
+		d.Budget = NewInMemorySearchBudget(nil)
+	}
+	if d.MaxPerRun <= 0 {
+		d.MaxPerRun = 10
+	}
+	if d.MaxWords <= 0 {
+		d.MaxWords = 200
+	}
+	cfg := fixedMetaConfig{maxPerRun: d.MaxPerRun, maxWords: d.MaxWords}
+	return registerAll(reg,
+		NewClassify(d.Helper, cfg, d.Budget),
+		NewExtractEntities(d.Helper, cfg, d.Budget),
+		NewSummarize(d.Helper, cfg, d.Budget, d.Files),
+	)
+}
+
+func registerAll(reg tool.Registry, ts ...tool.Tool) error {
+	for _, t := range ts {
 		if err := reg.Register(t); err != nil {
 			return err
 		}
 	}
 	return nil
 }
+
+// fixedMetaConfig satisfies ClassifyConfig / ExtractEntitiesConfig /
+// SummarizeConfig with static caps read from MetaDeps.
+type fixedMetaConfig struct{ maxPerRun, maxWords int }
+
+func (c fixedMetaConfig) MaxPerRun(context.Context) int { return c.maxPerRun }
+func (c fixedMetaConfig) MaxWords(context.Context) int  { return c.maxWords }