Reduce MCP token usage

2026-05-29 13:16:05 +01:00
parent da46340a82
commit 51aac9f447
9 changed files with 453 additions and 137 deletions
--- a/internal/mcp/protocol.go
+++ b/internal/mcp/protocol.go
@@ -3,6 +3,8 @@ package mcp
 import (
 	"encoding/json"
 	"fmt"
+
+	"github.com/hjbdev/patterm/internal/scratchpad"
 )

 // MCP protocol surface. The patterm server originally exposed each
@@ -43,7 +45,7 @@ var serverInfo = map[string]any{
 //     up as sub-agents and won't be tied into the patterm lifecycle.
 //
 // Keep this short — clients vary in how much they surface to the LLM.
-const serverInstructions = "You are already running INSIDE patterm; the `patterm` MCP server is connected over the same stdio MCP transport you use for any other MCP server. Use the MCP tools you see in tools/list — do NOT (a) try to launch `patterm` or `patterm mcp-stdio` yourself, (b) poke the Unix socket through perl / nc / socat / curl, or (c) shell out to `claude` / `codex` / `opencode` to start a peer. Any of those bypasses caller-identity and the new agent will land as a stray top-level tab instead of a child under you. Start with `whoami` for your role and the full tool list, then `help('topics')` for orientation. `spawn_agent` is the only correct way to start a sub-agent; `spawn_process` is for non-LLM commands; `list_processes` / `get_process_output` inspect them; `send_input` / `send_message` drive them. Whatever you spawn is yours to `close_process` when done. When you `send_message` a sub-agent, its reply comes back into YOUR pane as `[sub-agent:<name>] …`, not into the sub-agent's output — to wait for it, use `timer_fire_when_idle_any([sub_agent])` and then read your own pane; do NOT `wait_for_pattern` on the sub-agent, that will deadlock until timeout."
+const serverInstructions = "You are inside patterm. Use these MCP tools; do not launch patterm or poke its Unix socket yourself. Use spawn_agent for sub-agents, close spawned panes when done, and use timer_fire_when_idle_* instead of wait_for_pattern to wait for send_message replies."

 // toolDescriptor is the shape returned by `tools/list`. inputSchema is
 // a JSON Schema object — we provide a minimal `{type: "object"}` schema
@@ -76,37 +78,41 @@ func objectSchema(properties map[string]any, required []string) map[string]any {
 }

 func stringProp(desc string) map[string]any {
-	return map[string]any{"type": "string", "description": desc}
+	_ = desc
+	return map[string]any{"type": "string"}
 }

 func numberProp(desc string) map[string]any {
-	return map[string]any{"type": "number", "description": desc}
+	_ = desc
+	return map[string]any{"type": "number"}
 }

 func integerProp(desc string) map[string]any {
-	return map[string]any{"type": "integer", "description": desc}
+	_ = desc
+	return map[string]any{"type": "integer"}
 }

 func booleanProp(desc string) map[string]any {
-	return map[string]any{"type": "boolean", "description": desc}
+	_ = desc
+	return map[string]any{"type": "boolean"}
 }

 func arrayOfStringsProp(desc string) map[string]any {
+	_ = desc
 	return map[string]any{
-		"type":        "array",
-		"description": desc,
-		"items":       map[string]any{"type": "string"},
+		"type":  "array",
+		"items": map[string]any{"type": "string"},
 	}
 }

 // toolCatalog is the full list advertised via tools/list. Descriptions
 // are intentionally short — clients are expected to fetch help() for
 // detail. Schemas mirror the param structs in tools.go.
-func toolCatalog() []toolDescriptor {
-	return []toolDescriptor{
+func toolCatalog(role CallerRole) []toolDescriptor {
+	tools := []toolDescriptor{
 		{
 			Name:        "spawn_agent",
-			Description: "Spawn a sub-agent from an agent preset and optionally seed it with initial instructions. This is the ONLY correct way to start a sub-agent under you — do not shell out to `claude` / `codex` / `opencode` and do not poke patterm's Unix socket via perl / nc / socat. Either bypasses caller identity and the new agent lands as a stray top-level tab instead of your child. Caller owns lifecycle: when the sub-agent's work is done (it reports back via send_message, or you no longer need it), call close_process on its process_id to free the pane and tear down the PTY. See help('spawning') and help('lifecycle').",
+			Description: "Spawn a sub-agent from an agent preset.",
 			InputSchema: objectSchema(map[string]any{
 				"agent":              stringProp("Preset name (e.g. \"claude\", \"codex\")."),
 				"agent_instructions": stringProp("Initial prompt typed into the agent after it's ready."),
@@ -115,14 +121,14 @@ func toolCatalog() []toolDescriptor {
 		},
 		{
 			Name:        "spawn_process",
-			Description: "Spawn a process: a terminal, a process preset, or a freeform argv command. Caller owns lifecycle: when the process is no longer needed, call close_process to remove its entry (live children are SIGKILL'd first). See help('lifecycle').",
+			Description: "Spawn a terminal, process preset, or argv command.",
 			InputSchema: objectSchema(map[string]any{
 				"kind":        stringProp("\"terminal\" or \"command\"."),
 				"preset":      stringProp("Process preset name (mutually exclusive with argv)."),
-				"argv":        map[string]any{"type": "array", "items": map[string]any{"type": "string"}, "description": "Argv vector for freeform commands."},
+				"argv":        map[string]any{"type": "array", "items": map[string]any{"type": "string"}},
 				"name":        stringProp("Display name for the pane."),
 				"working_dir": stringProp("Working directory for the spawned process."),
-				"env":         map[string]any{"type": "object", "additionalProperties": map[string]any{"type": "string"}, "description": "Extra environment variables."},
+				"env":         map[string]any{"type": "object", "additionalProperties": map[string]any{"type": "string"}},
 				"shell":       booleanProp("Run argv through sh -lc."),
 			}, nil),
 		},
@@ -188,7 +194,9 @@ func toolCatalog() []toolDescriptor {
 		{
 			Name:        "get_project_status",
 			Description: "One-shot orientation: project, caller, processes, scratchpads.",
-			InputSchema: objectSchema(nil, nil),
+			InputSchema: objectSchema(map[string]any{
+				"include_tools": booleanProp("Include available_tools in caller metadata."),
+			}, nil),
 		},
 		{
 			Name:        "get_process_output",
@@ -197,6 +205,7 @@ func toolCatalog() []toolDescriptor {
 				"process_id":   stringProp("Target process id."),
 				"mode":         stringProp("\"grid\" (default) or \"stream\"."),
 				"since_offset": integerProp("Watermark offset from a previous call."),
+				"max_bytes":    integerProp("Maximum content bytes to return."),
 			}, []string{"process_id"}),
 		},
 		{
@@ -205,6 +214,7 @@ func toolCatalog() []toolDescriptor {
 			InputSchema: objectSchema(map[string]any{
 				"process_id":   stringProp("Target process id."),
 				"since_offset": integerProp("Byte offset from a previous call."),
+				"max_bytes":    integerProp("Maximum content bytes to return."),
 			}, []string{"process_id"}),
 		},
 		{
@@ -214,12 +224,13 @@ func toolCatalog() []toolDescriptor {
 				"process_id": stringProp("Target process id."),
 				"pattern":    stringProp("Regex pattern."),
 				"kind":       stringProp("\"rendered\" (default) or \"raw\"."),
-				"limit":      integerProp("Max matches (default 20)."),
+				"limit":      integerProp("Max matches (default 10)."),
+				"max_bytes":  integerProp("Max bytes per returned match line."),
 			}, []string{"process_id", "pattern"}),
 		},
 		{
 			Name:        "wait_for_pattern",
-			Description: "Block until pattern appears in the TARGET process's own output, or timeout elapses. Use this for waiting on text the target itself will emit (a shell prompt, a build's \"tests passed\" line, etc.). Anti-pattern: do NOT use this to wait for a sub-agent's reply to send_message — replies are routed into the CALLER's pane tagged `[sub-agent:<name>]`, not into the sub-agent's output, so this call will spin to timeout. For sub-agent coordination use `timer_fire_when_idle_any` and then read your own pane.",
+			Description: "Block until pattern appears in the target process output.",
 			InputSchema: objectSchema(map[string]any{
 				"process_id":      stringProp("Target process id."),
 				"pattern":         stringProp("Regex pattern."),
@@ -238,18 +249,19 @@ func toolCatalog() []toolDescriptor {
 			Name:        "send_input",
 			Description: "Type text, paste a block, or fire a named key into a process. Optional tail-after-send.",
 			InputSchema: objectSchema(map[string]any{
-				"process_id": stringProp("Target process id."),
-				"kind":       stringProp("\"text\", \"paste\", or \"key\"."),
-				"text":       stringProp("Text payload for kind=text/paste."),
-				"key":        stringProp("Named key for kind=key (e.g. \"enter\", \"escape\")."),
-				"submit":     booleanProp("Whether to append a submit keystroke."),
-				"wait_ms":    integerProp("After sending, wait this many ms before tailing."),
-				"tail_mode":  stringProp("\"none\" (default), \"stream\", or \"grid\"."),
+				"process_id":     stringProp("Target process id."),
+				"kind":           stringProp("\"text\", \"paste\", or \"key\"."),
+				"text":           stringProp("Text payload for kind=text/paste."),
+				"key":            stringProp("Named key for kind=key (e.g. \"enter\", \"escape\")."),
+				"submit":         booleanProp("Whether to append a submit keystroke."),
+				"wait_ms":        integerProp("After sending, wait this many ms before tailing."),
+				"tail_mode":      stringProp("\"none\" (default), \"stream\", or \"grid\"."),
+				"tail_max_bytes": integerProp("Maximum bytes in returned tail."),
 			}, []string{"process_id", "kind"}),
 		},
 		{
 			Name:        "send_message",
-			Description: "Deliver a text message to another process as orchestrator-owned input. Fire-and-forget: returns immediately, without waiting for the recipient to read or act. If the recipient replies via send_message, that reply arrives in YOUR pane tagged `[sub-agent:<name>]` (child→parent) or `[orchestrator]` (parent→child) — NOT in the recipient's output. To wait for a sub-agent's reply, schedule `timer_fire_when_idle_any([sub_agent_id], body=…)` and then read your own pane when the timer fires. Do not `wait_for_pattern` on the recipient for a reply; it will deadlock.",
+			Description: "Send a tagged message to a parent or child process.",
 			InputSchema: objectSchema(map[string]any{
 				"target_process_id": stringProp("Recipient process id."),
 				"message":           stringProp("Message body."),
@@ -283,7 +295,7 @@ func toolCatalog() []toolDescriptor {
 		},
 		{
 			Name:        "timer_fire_when_idle_any",
-			Description: "Canonical way to wait for a sub-agent to finish working: send_message the sub-agent, then schedule this with watched=[sub_agent_id]; when it fires, the reply is already sitting in your own pane tagged `[sub-agent:<name>]`. Schedules a timer that fires when any watched process enters idle (already-idle entries excluded), or when max_wait_seconds elapses.",
+			Description: "Fire when any watched process becomes idle.",
 			InputSchema: objectSchema(map[string]any{
 				"watched":          arrayOfStringsProp("Process ids to watch."),
 				"body":             stringProp("Message delivered verbatim to the owning agent when the timer fires."),
@@ -294,7 +306,7 @@ func toolCatalog() []toolDescriptor {
 		},
 		{
 			Name:        "timer_fire_when_idle_all",
-			Description: "Canonical way to wait for several sub-agents to finish working in parallel: send_message each one, then schedule this with watched=[…ids]; when it fires, each reply is in your own pane tagged `[sub-agent:<name>]`. Schedules a timer that fires when all watched processes are idle (already-idle entries count as satisfied), or when max_wait_seconds elapses.",
+			Description: "Fire when all watched processes are idle.",
 			InputSchema: objectSchema(map[string]any{
 				"watched":          arrayOfStringsProp("Process ids to watch."),
 				"body":             stringProp("Message delivered verbatim to the owning agent when the timer fires."),
@@ -338,7 +350,9 @@ func toolCatalog() []toolDescriptor {
 			Name:        "scratchpad_read",
 			Description: "Read a scratchpad entry, returning content and revision.",
 			InputSchema: objectSchema(map[string]any{
-				"name": stringProp("Scratchpad name."),
+				"name":      stringProp("Scratchpad name."),
+				"offset":    integerProp("Byte offset to start reading."),
+				"max_bytes": integerProp("Maximum content bytes to return."),
 			}, []string{"name"}),
 		},
 		{
@@ -367,8 +381,10 @@ func toolCatalog() []toolDescriptor {
 		},
 		{
 			Name:        "whoami",
-			Description: "Return the caller's identity, role, parent, project metadata, and available tools.",
-			InputSchema: objectSchema(nil, nil),
+			Description: "Return caller identity, role, parent, and project metadata.",
+			InputSchema: objectSchema(map[string]any{
+				"include_tools": booleanProp("Include full available tool list."),
+			}, nil),
 		},
 		{
 			Name:        "help",
@@ -378,6 +394,16 @@ func toolCatalog() []toolDescriptor {
 			}, nil),
 		},
 	}
+	if role != RoleSubAgent {
+		return tools
+	}
+	filtered := tools[:0]
+	for _, tool := range tools {
+		if tool.Name != "spawn_agent" {
+			filtered = append(filtered, tool)
+		}
+	}
+	return filtered
 }

 // handleProtocolMethod handles MCP protocol-level methods. Returns
@@ -416,7 +442,14 @@ func (s *Server) handleProtocolMethod(callerID, method string, params json.RawMe
 		return map[string]any{}, true, 0, "", nil

 	case "tools/list":
-		return map[string]any{"tools": toolCatalog()}, true, 0, "", nil
+		role := RoleOrchestrator
+		s.mu.Lock()
+		host := s.host
+		s.mu.Unlock()
+		if host != nil {
+			role = host.CallerRole(callerID)
+		}
+		return map[string]any{"tools": toolCatalog(role)}, true, 0, "", nil

 	case "tools/call":
 		var p struct {
@@ -472,25 +505,12 @@ func (s *Server) handleProtocolMethod(callerID, method string, params json.RawMe
 	return nil, false, 0, "", nil
 }

-// wrapToolResult turns a structured tool result into an MCP tools/call
-// response. Plain strings (e.g. "ok") become text content; structured
-// values are JSON-encoded into a single text block and also exposed
-// under structuredContent so capable clients can read the shape.
+// wrapToolResult turns a tool result into an MCP tools/call response.
+// Structured values are exposed once under structuredContent; content
+// carries only a short model-readable summary to avoid duplicating
+// large JSON payloads into the transcript.
 func wrapToolResult(result any) map[string]any {
-	var text string
-	switch v := result.(type) {
-	case nil:
-		text = "ok"
-	case string:
-		text = v
-	default:
-		b, err := json.Marshal(v)
-		if err != nil {
-			text = fmt.Sprintf("%v", v)
-		} else {
-			text = string(b)
-		}
-	}
+	text := summarizeToolResult(result)
 	out := map[string]any{
 		"content": []map[string]any{{"type": "text", "text": text}},
 		"isError": false,
@@ -505,3 +525,70 @@ func wrapToolResult(result any) map[string]any {
 	}
 	return out
 }
+
+func summarizeToolResult(result any) string {
+	switch v := result.(type) {
+	case nil:
+		return "ok"
+	case string:
+		return v
+	case ProcessInfo:
+		return fmt.Sprintf("%s %s %s", v.ID, v.Kind, v.Status)
+	case []ProcessInfo:
+		return fmt.Sprintf("%d processes", len(v))
+	case ProcessStatus:
+		return fmt.Sprintf("%s %s %s", v.ID, v.Kind, v.Status)
+	case ProjectStatus:
+		return fmt.Sprintf("%d processes, %d scratchpads", len(v.Processes), len(v.Scratchpads))
+	case ProcessOutput:
+		return outputSummary(v.Mode, v.ContentBytes, v.Truncated, v.NewOffset)
+	case RawOutput:
+		return outputSummary("raw", v.ContentBytes, v.Truncated, v.NewOffset)
+	case SearchResult:
+		if v.Truncated {
+			return fmt.Sprintf("%d matches (truncated)", len(v.Matches))
+		}
+		return fmt.Sprintf("%d matches", len(v.Matches))
+	case SendInputResult:
+		if v.Tail != nil {
+			return "ok; tail included"
+		}
+		return "ok"
+	case TimerHandle:
+		return "timer " + v.ID
+	case TimerFireWhenIdleResponse:
+		if v.ID != "" {
+			return fmt.Sprintf("%s timer %s", v.Status, v.ID)
+		}
+		return v.Status
+	case []TimerInfo:
+		return fmt.Sprintf("%d timers", len(v))
+	case []scratchpad.Entry:
+		return fmt.Sprintf("%d scratchpads", len(v))
+	case ScratchpadReadResult:
+		if v.Truncated {
+			return fmt.Sprintf("%d/%d bytes from offset %d", v.ContentBytes, v.TotalBytes, v.Offset)
+		}
+		return fmt.Sprintf("%d bytes", v.ContentBytes)
+	case WhoAmI:
+		if v.ProcessID == "" {
+			return string(v.Role)
+		}
+		return fmt.Sprintf("%s %s", v.ProcessID, v.Role)
+	case HelpResponse:
+		return fmt.Sprintf("help: %s", v.Topic)
+	default:
+		return "ok"
+	}
+}
+
+func outputSummary(mode string, bytes int, truncated bool, offset int64) string {
+	s := fmt.Sprintf("%s output: %d bytes", mode, bytes)
+	if offset > 0 {
+		s += fmt.Sprintf(", offset %d", offset)
+	}
+	if truncated {
+		s += " (truncated)"
+	}
+	return s
+}