ollie-core/agent/loop.go

package agent

import (
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"strings"
	"sync"
	"time"

	"ollie/backend"
	"ollie/toolsrv"
)

const maxTransientRetries = 3

// retryBaseDelay is the base delay for rate-limit retries (5<<attempt seconds).
// Tests can override this to speed up retry tests.
var retryBaseDelay = 5 * time.Second

// streamDropBaseDelay is the base delay for stream-drop retries (2<<attempt seconds).
var streamDropBaseDelay = 2 * time.Second

// Consecutive tool-error thresholds. At the soft limit the model is nudged
// to try a different approach; at the hard limit the loop aborts.
const (
	consecutiveErrorSoftLimit = 5
	consecutiveErrorHardLimit = 10
)

// replanGate is the max consecutive tool rounds without a PLAN: block before
// a mandatory replan nudge is injected.
const replanGate = 8

// stallThreshold is the number of consecutive rounds where LastAction doesn't
// change before injecting a stall nudge.
const stallThreshold = 5

// planReinjectInterval is the number of tool-call rounds between periodic
// task state re-injection into the conversation.
const planReinjectInterval = 10

type toolExecutor func(ctx context.Context, name string, args json.RawMessage) (string, []backend.ContentBlock, error)

type agentConfig struct {
	Backend          backend.Backend
	Tools            []backend.Tool
	Exec             toolExecutor
	ClassifyTool     func(name string) bool                             // nil=treat all as serial; true=parallel-read-safe
	ClassifyTier     func(name string, args json.RawMessage) ResultTier // nil=TierHot for all; tools self-classify retention
	Output           EventHandler
	preamble         string // compiled system+agent prompt sent as the system role
	GenerationParams backend.GenerationParams
	PopInject        func() string // returns and clears pending inject, or ""
	AutoCompact      func(ctx context.Context)                                                              // called after each tool round; may compact in-place
	SaveSession      func()                                                                               // called after each state.update(); persists mid-turn progress
	// MaxSteps is the maximum number of tool-call rounds per turn.
	// When reached, a soft nudge is injected and the loop exits cleanly.
	// 0 means unlimited.
	MaxSteps int
	// ReadPlanStep returns the next unchecked step from the plan file, or "".
	// Used as the primary source for TaskState.PlanStep; falls back to
	// heuristic extraction from assistant text if nil.
	ReadPlanStep func() string
	// IncrToolCallCount increments and returns the session-wide tool call counter.
	IncrToolCallCount func() int64
	// ResultCache persists read-safe tool results across turns.
	// Cleared on write operations (file_write, file_edit).
	ResultCache *sync.Map
}

func run(ctx context.Context, cfg agentConfig, state state) error {
	var step int
	var consecutiveErrors int // rounds where every tool call returned an error
	var roundsWithoutPlan int // consecutive tool rounds without a PLAN: block
	var lastAction string     // previous round's LastAction for stall detection
	var stallRounds int       // consecutive rounds with unchanged LastAction
	var lastErrorSig string   // signature of previous round's error(s) for repeat detection
	var repeatErrorCount int  // consecutive rounds with identical error signature
	// resultCache stores outputs of read-safe tool calls keyed by name+args.
	// Only tools classified as parallel-safe (immutable reads) are cached.
	// sync.Map is required because execOne may run in concurrent goroutines.
	// Uses the cross-turn cache from agentConfig if available.
	resultCache := cfg.ResultCache
	if resultCache == nil {
		resultCache = &sync.Map{}
	}

	for {
		emit(cfg, Event{Role: "state", Content: "thinking"})
		// Proactive context gate: strip cold material before calling the backend.
		if budget := contextBudget(ctx, cfg.Backend); budget > 0 && state.estimateTokens() > budget {
			state.stripCold(ctx, cfg.Backend)
		}

		history := state.history()
		if cfg.preamble != "" {
			history = append([]backend.Message{{Role: "system", Content: cfg.preamble}}, history...)
		}
		if ts := state.taskState(); ts != nil {
			if msg := ts.render(); msg != "" && len(history) > 0 {
				// Insert after system prompt, before conversation.
				history = append([]backend.Message{history[0], {Role: "user", Content: msg}}, history[1:]...)
			}
		}

		// Stream the assistant's response, retrying on rate limits, transient
		// backend errors (5xx, network), and mid-stream drops. One stable ID is
		// shared by all chunks and the completed assistant message.
		responseID := NewResponseID()
		var content strings.Builder
		var reasoning strings.Builder
		var toolCalls []backend.ToolCall
		var stopReason string

		for attempt := range maxTransientRetries + 1 {
			content.Reset()
			reasoning.Reset()
			toolCalls = nil

			ch, err := cfg.Backend.ChatStream(ctx, history, cfg.Tools, cfg.GenerationParams)
			if err != nil {
				if ctx.Err() != nil {
					return ctx.Err()
				}
				wait, retryable := transientWait(err, attempt)
				if !retryable || attempt >= maxTransientRetries {
					return fmt.Errorf("step %d: %w", step, err)
				}
				var rlErr *backend.RateLimitError
				if errors.As(err, &rlErr) {
					emit(cfg, Event{Role: "limitretry"})
				}
				if err := retryCountdown(ctx, cfg, wait); err != nil {
					return fmt.Errorf("step %d: %w", step, err)
				}
				continue
			}

			var done bool
			var hadReasoning bool
			for ev := range ch {
				if ev.Reasoning != "" {
					if !hadReasoning {
						emit(cfg, Event{Role: "reasoning", Content: "<think>\n"})
						hadReasoning = true
					}
					reasoning.WriteString(ev.Reasoning)
					emit(cfg, Event{Role: "reasoning", Content: ev.Reasoning})
				}
				if ev.Content != "" {
					if hadReasoning {
						emit(cfg, Event{Role: "reasoning", Content: "\n</think>\n"})
						hadReasoning = false
					}
					content.WriteString(ev.Content)
					emit(cfg, Event{Role: "assistant", Content: ev.Content, ResponseID: responseID})
				}
				toolCalls = append(toolCalls, ev.ToolCalls...)
				if ev.Done {
					if hadReasoning {
						emit(cfg, Event{Role: "reasoning", Content: "\n</think>\n"})
						hadReasoning = false
					}
					stopReason = ev.StopReason
					done = true
					if ev.Usage.InputTokens > 0 || ev.Usage.OutputTokens > 0 {
						emit(cfg, Event{
							Role:    "usage",
							Content: fmt.Sprintf("%d %d 0 %g %d %d", ev.Usage.InputTokens, ev.Usage.OutputTokens, ev.Usage.CostUSD, ev.Usage.CachedInputTokens, ev.Usage.CacheCreationTokens),
						})
					} else {
						// Backend didn't report usage; estimate from content.
						inChars := 0
						for _, m := range history {
							inChars += len(m.Content)
							for _, tc := range m.ToolCalls {
								inChars += len(tc.Name) + len(tc.Arguments)
							}
						}
						emit(cfg, Event{
							Role:    "usage",
							Content: fmt.Sprintf("%d %d 1 0", inChars/4, content.Len()/4),
						})
					}
					break
				}
			}

			if done {
				break
			}

			if ctx.Err() != nil {
				// User pause: record partial state and return.
				if hadReasoning {
					emit(cfg, Event{Role: "reasoning", Content: "\n</think>\n"})
				}
				msg := backend.Message{ID: responseID, Role: "assistant", Content: content.String(), Reasoning: reasoning.String(), ToolCalls: toolCalls}
				var results []toolResult
				for _, tc := range toolCalls {
					results = append(results, toolResult{
						ToolCallID: tc.ID,
						Name:       tc.Name,
						Content:    `{"status":"cancelled","error":"stream interrupted"}`,
						IsError:    true,
					})
				}
				state.update(msg, results)
				if cfg.SaveSession != nil {
					cfg.SaveSession()
				}
				return ctx.Err()
			}

			// Pure stream drop — retry if attempts remain.
			if attempt >= maxTransientRetries {
				return fmt.Errorf("step %d: stream dropped (no more retries)", step)
			}
			if hadReasoning {
				emit(cfg, Event{Role: "reasoning", Content: "\n</think>\n"})
			}
			wait := streamDropBaseDelay << attempt
			if err := retryCountdown(ctx, cfg, wait); err != nil {
				return fmt.Errorf("step %d: %w", step, err)
			}
		}

		switch stopReason {
		case "stop", "tool_calls", "length", "error", "":
			// normal
		default:
			return fmt.Errorf("step %d: %s", step, stopReason)
		}

		// Parse text-based tool calls: models that don't support the function
		// calling API emit tool invocations as plain text (e.g.
		// "file_read: args=[...]" (text-based tool call syntax)). Parse these into proper ToolCall structs.
		if len(toolCalls) == 0 && len(cfg.Tools) > 0 {
			if parsed := parseTextToolCalls(content.String(), cfg.Tools); len(parsed) > 0 {
				toolCalls = parsed
			}
		}

		// Execute tool calls, running consecutive parallel-read-safe tools concurrently.
		msg := backend.Message{ID: responseID, Role: "assistant", Content: content.String(), Reasoning: reasoning.String(), ToolCalls: toolCalls}
		results := make([]toolResult, 0, len(toolCalls))
		interrupted := false

		cancelledResult := func(tc backend.ToolCall) toolResult {
			return toolResult{
				ToolCallID: tc.ID,
				Name:       tc.Name,
				Content:    `{"status":"cancelled","error":"interrupted"}`,
				IsError:    true,
			}
		}

		// execOne runs a single tool call end-to-end and reports whether the
		// context was cancelled during execution.
		execOne := func(tc backend.ToolCall) (toolResult, bool) {
			if tc.Name == "" {
				return toolResult{ToolCallID: tc.ID, Name: tc.Name, Content: "error: empty tool name", IsError: true}, false
			}
			if ctx.Err() != nil {
				cr := cancelledResult(tc)
				emit(cfg, Event{Role: "tool", Name: tc.Name, Content: cr.Content})
				return cr, true
			}
			emit(cfg, Event{Role: "call", Name: tc.Name, Content: string(tc.Arguments)})
			readSafe := cfg.ClassifyTool != nil && cfg.ClassifyTool(tc.Name)
			if readSafe {
				key := tc.Name + "\x00" + string(tc.Arguments)
				if v, ok := resultCache.Load(key); ok {
					cached := v.(string)
					emit(cfg, Event{Role: "tool", Name: tc.Name, Content: cached})
					return toolResult{ToolCallID: tc.ID, Name: tc.Name, Content: cached}, false
				}
			}
			var result string
			var resultBlocks []backend.ContentBlock
			var isErr bool
			streamed := false
			if cfg.Exec != nil {
				streamBytes := 0
				streamCtx := toolsrv.WithOutputStream(ctx, func(data string) {
					if streamBytes >= defaultToolResultMaxBytes {
						return // already at ceiling, drop further chunks
					}
					if !streamed {
						emit(cfg, Event{Role: "tool", Name: tc.Name})
						streamed = true
					}
					streamBytes += len(data)
					if streamBytes > defaultToolResultMaxBytes {
						// Emit only the portion within the ceiling.
						excess := streamBytes - defaultToolResultMaxBytes
						data = data[:len(data)-excess]
					}
					emit(cfg, Event{Role: "tool", Name: tc.Name, Content: data})
				})
				out, blocks, err := cfg.Exec(streamCtx, tc.Name, tc.Arguments)
				if err != nil {
					isErr = true
					if ctx.Err() != nil {
						result = "error: tool execution interrupted by user"
						if cfg.PopInject != nil {
							if injected := cfg.PopInject(); injected != "" {
								result += "\n\n<system-user-interruption>\n" + injected + "\n</system-user-interruption>"
							}
						}
						emit(cfg, Event{Role: "tool", Name: tc.Name, Content: result})
						return toolResult{ToolCallID: tc.ID, Name: tc.Name, Content: result, IsError: true}, true
					}
					result = fmt.Sprintf("error: %v", err)
				} else {
					result = out
					resultBlocks = blocks
				}
			} else {
				result = "error: no tool executor configured"
				isErr = true
			}
			// Accumulate suffix text (user-interruptions,
			// truncation hints) that must be emitted after streaming completes.
			var suffix string
			if cfg.PopInject != nil {
				if injected := cfg.PopInject(); injected != "" {
					suffix += "\n\n<system-user-interruption>\n" + injected + "\n</system-user-interruption>"
				}
			}
			result += suffix
			// Safety ceiling: unconditionally cap all tool results at 128KB.
			if len(result) > defaultToolResultMaxBytes {
				orig := len(result)
				result = strings.ToValidUTF8(result[:defaultToolResultMaxBytes], "")
				suffix = fmt.Sprintf("\n\n[HARD LIMIT: %s output truncated — %d of %d bytes shown. This is a safety ceiling, not a semantic boundary.]",
					tc.Name, defaultToolResultMaxBytes, orig)
				result += suffix
			}
			if readSafe && !isErr {
				resultCache.Store(tc.Name+"\x00"+string(tc.Arguments), result)
			}
			// Invalidate cache on write operations that may change file contents.
			if !readSafe && (tc.Name == "file_write" || tc.Name == "file_edit") {
				*resultCache = sync.Map{}
			}
			if !streamed {
				emit(cfg, Event{Role: "tool", Name: tc.Name, Content: result})
			} else if suffix != "" {
				// Emit suffixes that were appended after streaming completed
				// so they appear at the end of the chat output.
				emit(cfg, Event{Role: "tool", Name: tc.Name, Content: suffix})
			}
			tier := TierHot
			if !isErr && cfg.ClassifyTier != nil {
				tier = cfg.ClassifyTier(tc.Name, tc.Arguments)
			}
			return toolResult{ToolCallID: tc.ID, Name: tc.Name, Content: result, ContentBlocks: resultBlocks, IsError: isErr, Tier: tier}, false
		}

		isParallelSafe := func(name string) bool {
			return name != "" && cfg.ClassifyTool != nil && cfg.ClassifyTool(name)
		}

		for i := 0; i < len(toolCalls) && !interrupted; {
			if ctx.Err() != nil {
				for _, remaining := range toolCalls[i:] {
					cr := cancelledResult(remaining)
					emit(cfg, Event{Role: "tool", Name: remaining.Name, Content: cr.Content})
					results = append(results, cr)
				}
				interrupted = true
				break
			}

			// Collect a run of consecutive parallel-safe calls.
			j := i + 1
			if isParallelSafe(toolCalls[i].Name) {
				for j < len(toolCalls) && isParallelSafe(toolCalls[j].Name) {
					j++
				}
			}
			batch := toolCalls[i:j]

			fillCancelled := func(from int) {
				for _, remaining := range toolCalls[from:] {
					cr := cancelledResult(remaining)
					emit(cfg, Event{Role: "tool", Name: remaining.Name, Content: cr.Content})
					results = append(results, cr)
				}
			}

			if len(batch) == 1 {
				tr, wasInt := execOne(batch[0])
				results = append(results, tr)
				if wasInt {
					fillCancelled(j)
					interrupted = true
				}
			} else {
				// Fan out the batch concurrently; preserve submission order in results.
				// Deduplicate identical calls so only one executes per unique key.
				type inflightResult struct {
					tr     toolResult
					wasInt bool
				}
				inflight := make(map[string]int) // key -> index of first occurrence
				batchResults := make([]toolResult, len(batch))
				batchInt := make([]bool, len(batch))
				var wg sync.WaitGroup
				uniqueResults := make([]inflightResult, len(batch))
				for k, tc := range batch {
					key := tc.Name + "\x00" + string(tc.Arguments)
					if first, dup := inflight[key]; dup {
						// Will copy result from first occurrence after wg.Wait.
						batchResults[k] = toolResult{} // placeholder
						_ = first                      // used below
						continue
					}
					inflight[key] = k
					wg.Add(1)
					go func(k int, tc backend.ToolCall) {
						defer wg.Done()
						uniqueResults[k].tr, uniqueResults[k].wasInt = execOne(tc)
					}(k, tc)
				}
				wg.Wait()
				// Fill results: unique calls get their own result, duplicates copy from first.
				for k, tc := range batch {
					key := tc.Name + "\x00" + string(tc.Arguments)
					first := inflight[key]
					if k == first {
						batchResults[k] = uniqueResults[k].tr
						batchInt[k] = uniqueResults[k].wasInt
					} else {
						batchResults[k] = toolResult{
							ToolCallID: tc.ID,
							Name:       tc.Name,
							Content:    uniqueResults[first].tr.Content,
							IsError:    uniqueResults[first].tr.IsError,
						}
						batchInt[k] = uniqueResults[first].wasInt
						emit(cfg, Event{Role: "call", Name: tc.Name, Content: string(tc.Arguments)})
						emit(cfg, Event{Role: "tool", Name: tc.Name, Content: batchResults[k].Content})
					}
				}
				// Add all batch results — model needs a result for every tool call.
				results = append(results, batchResults...)
				for _, wasInt := range batchInt {
					if wasInt {
						fillCancelled(j)
						interrupted = true
						break
					}
				}
			}
			i = j
		}

		state.update(msg, results)
		if cfg.SaveSession != nil {
			cfg.SaveSession()
		}

		// Auto-update TaskState from what actually happened this round.
		if ts := state.taskState(); ts != nil {
			inferTaskStateUpdate(ts, msg, results, cfg.ReadPlanStep)
			state.updateTaskState(*ts)
		}

		if interrupted {
			return ctx.Err()
		}

		if len(toolCalls) == 0 {
			break
		}

		// Track consecutive rounds where every tool call errored.
		allErrors := len(results) > 0
		for _, r := range results {
			if !r.IsError {
				allErrors = false
				break
			}
		}
		if allErrors {
			consecutiveErrors++
		} else {
			consecutiveErrors = 0
		}
		if consecutiveErrors >= consecutiveErrorHardLimit {
			emit(cfg, Event{Role: "error", Content: fmt.Sprintf("%d consecutive tool errors — aborting", consecutiveErrors)})
			return fmt.Errorf("step %d: %d consecutive tool errors", step, consecutiveErrors)
		}
		if consecutiveErrors == consecutiveErrorSoftLimit {
			// Nudge the model to try a different approach and keep the plan current.
			state.update(backend.Message{
				Role:    "user",
				Content: "[system: your last several tool calls all failed. Try a different approach, or ask the user for help. Keep your plan current — update s/$OLLIE_SESSION_ID/plan to reflect where you are and what remains.]",
			}, nil)
		}

		// Repeated identical error detection: if the same error occurs 3 times
		// in a row, the model is stuck on a format/argument issue. Inject the
		// error content as a corrective nudge immediately rather than waiting
		// for the generic soft limit.
		if allErrors && len(results) > 0 {
			var errSig strings.Builder
			for _, r := range results {
				if r.IsError {
					errSig.WriteString(r.Content)
					errSig.WriteByte('\n')
				}
			}
			sig := errSig.String()
			if sig == lastErrorSig {
				repeatErrorCount++
			} else {
				repeatErrorCount = 1
				lastErrorSig = sig
			}
			if repeatErrorCount == 3 {
				state.update(backend.Message{
					Role:    "user",
					Content: "[system: you have made the same malformed tool call 3 times in a row. Read the error message carefully and fix the arguments. The error was: " + strings.TrimSpace(results[0].Content) + "]",
				}, nil)
			}
		} else {
			repeatErrorCount = 0
			lastErrorSig = ""
		}

		// Replan gate: if the model has been calling tools without emitting a
		// PLAN: block, force it to replan before continuing.
		if strings.Contains(msg.Content, "PLAN:") {
			roundsWithoutPlan = 0
		} else {
			roundsWithoutPlan++
		}
		if roundsWithoutPlan >= replanGate {
			state.update(backend.Message{
				Role:    "user",
				Content: "[system: you have executed " + fmt.Sprintf("%d", replanGate) + " tool rounds without replanning. Stop and write a PLAN: block showing your current checklist before calling any more toolsrv. Update s/$OLLIE_SESSION_ID/plan.]",
			}, nil)
			roundsWithoutPlan = 0
		}

		// Stall detection: if LastAction hasn't changed across rounds, the
		// model is likely stuck in a loop.
		if ts := state.taskState(); ts != nil {
			if ts.LastAction == lastAction && lastAction != "" {
				stallRounds++
			} else {
				stallRounds = 0
			}
			lastAction = ts.LastAction
			if stallRounds >= stallThreshold {
				state.update(backend.Message{
					Role:    "user",
					Content: "[system: stall detected — you have repeated the same action for " + fmt.Sprintf("%d", stallRounds) + " rounds. Step back, reassess your approach, and try something different. Update your plan.]",
				}, nil)
				stallRounds = 0
			}
		}

		// Soft step-budget guardrail: when MaxSteps is set and the budget is
		// exhausted, nudge the model to wrap up and exit cleanly. This is not
		// a hard abort — the model gets one final turn without tools to emit
		// a summary or hand-off message.
		if cfg.MaxSteps > 0 && step >= cfg.MaxSteps-1 {
			emit(cfg, Event{Role: "maxsteps", Content: fmt.Sprintf("%d", step+1)})
			state.update(backend.Message{
				Role:    "user",
				Content: fmt.Sprintf("[system: step budget exhausted (%d/%d steps used). Stop calling toolsrv. Summarize what you have done and what remains, then stop.]", step+1, cfg.MaxSteps),
			}, nil)
			break
		}

		// Periodic task state re-injection: every N tool rounds, surface the
		// structured task state back into the conversation to keep the model on track.
		if step > 0 && step%planReinjectInterval == 0 {
			if ts := state.taskState(); ts != nil {
				if msg := ts.render(); msg != "" {
					state.update(backend.Message{
						Role:    "user",
						Content: "[system: review your current task state and continue. Update it if your approach has changed.]\n\n" + msg,
					}, nil)
				}
			}
		}

		if cfg.AutoCompact != nil {
			cfg.AutoCompact(ctx)
		}

		step++
	}

	return nil
}

func emit(cfg agentConfig, msg Event) {
	if cfg.Output != nil {
		cfg.Output(msg)
	}
}

// contextBudget returns the token threshold (50% of context length) above which
// cold material should be proactively stripped. Returns 0 if unknown.
func contextBudget(ctx context.Context, b backend.Backend) int {
	if b == nil {
		return 0
	}
	ctxLen := b.ContextLength(ctx)
	if ctxLen <= 0 {
		return 0
	}
	return ctxLen / 2
}

// inferTaskStateUpdate updates TaskState fields from the actual tool round
// without requiring model cooperation. Keeps the state current even if the
// model ignores re-injection nudges.
func inferTaskStateUpdate(ts *TaskState, msg backend.Message, results []toolResult, readPlanStep func() string) {
	// LastAction: summarize what tools ran and whether they succeeded.
	if len(msg.ToolCalls) > 0 {
		names := make([]string, 0, len(msg.ToolCalls))
		for _, tc := range msg.ToolCalls {
			names = append(names, tc.Name)
		}
		errCount := 0
		for _, r := range results {
			if r.IsError {
				errCount++
			}
		}
		action := strings.Join(names, ", ")
		if errCount > 0 {
			action += fmt.Sprintf(" (%d/%d failed)", errCount, len(results))
		}
		ts.LastAction = action
	}

	// PlanStep: prefer the plan file (ground truth), fall back to text heuristic.
	if readPlanStep != nil {
		if step := readPlanStep(); step != "" {
			ts.PlanStep = step
			return
		}
	}
	// Fallback: extract from PLAN: block if present in assistant text.
	if idx := strings.Index(msg.Content, "PLAN:"); idx >= 0 {
		lines := strings.Split(msg.Content[idx:], "\n")
		for _, line := range lines[1:] {
			trimmed := strings.TrimSpace(line)
			if strings.HasPrefix(trimmed, "- [ ]") {
				ts.PlanStep = strings.TrimSpace(trimmed[5:])
				break
			}
			if trimmed == "" || (!strings.HasPrefix(trimmed, "- [") && trimmed != "") {
				break
			}
		}
	}
}

// parseTextToolCalls extracts tool calls from assistant text when the model
// emits them as plain text instead of using the function calling API.
// Recognizes formats:
//   - tool_name: {"key": ...}           (JSON object)
//   - tool_name: key=[...], key2=...    (shorthand: key=value pairs → JSON object)
//
// Returns nil if no valid tool calls are found.
func parseTextToolCalls(text string, tools []backend.Tool) []backend.ToolCall {
	toolNames := make(map[string]bool, len(tools))
	for _, t := range tools {
		toolNames[t.Name] = true
	}

	var calls []backend.ToolCall
	for _, name := range []string{"shell"} {
		if !toolNames[name] {
			continue
		}
		prefix := name + ":"
		idx := strings.Index(text, prefix)
		if idx < 0 {
			continue
		}
		rest := strings.TrimSpace(text[idx+len(prefix):])
		if rest == "" {
			continue
		}

		var argsJSON string
		if rest[0] == '{' {
			// Direct JSON object
			argsJSON = extractJSONObject(rest)
			if argsJSON == "" {
				argsJSON = extractJSONObject(relaxJSON(rest))
			}
		} else {
			// Shorthand format: key=[...], key2=value
			argsJSON = shorthandToJSON(rest)
		}
		if argsJSON == "" {
			continue
		}
		var check json.RawMessage
		if json.Unmarshal([]byte(argsJSON), &check) != nil {
			continue
		}
		calls = append(calls, backend.ToolCall{
			ID:        fmt.Sprintf("text-%s-%d", name, len(calls)),
			Name:      name,
			Arguments: json.RawMessage(argsJSON),
		})
	}
	return calls
}

// shorthandToJSON converts "key=[...], key2=value" format to a JSON object.
// Handles: steps=[{code: "date"}], timeout=30
func shorthandToJSON(s string) string {
	// Find the first key=value where value starts with [ or {
	eqIdx := strings.IndexByte(s, '=')
	if eqIdx < 0 {
		return ""
	}
	key := strings.TrimSpace(s[:eqIdx])
	if key == "" {
		return ""
	}
	valStr := s[eqIdx+1:]

	// Extract the value (balanced brackets/braces)
	val := extractBalanced(strings.TrimSpace(valStr))
	if val == "" {
		return ""
	}

	// Relax the value's JSON (unquoted keys → quoted)
	relaxed := relaxJSON(val)

	// Build the JSON object: {"key": relaxed_value}
	keyJSON, _ := json.Marshal(key)
	result := "{" + string(keyJSON) + ":" + relaxed + "}"

	// Check if there are more key=value pairs after the value
	afterVal := strings.TrimSpace(valStr[len(val):])
	if len(afterVal) > 0 && afterVal[0] == ',' {
		// Parse additional key=value pairs
		extra := parseExtraKV(afterVal[1:])
		if extra != "" {
			// Merge: strip trailing } from result, append extra
			result = result[:len(result)-1] + "," + extra + "}"
		}
	}

	return result
}

// extractBalanced extracts a balanced [...] or {...} from the start of s.
func extractBalanced(s string) string {
	if len(s) == 0 {
		return ""
	}
	open := s[0]
	var close byte
	switch open {
	case '[':
		close = ']'
	case '{':
		close = '}'
	default:
		// Simple value (number, string) — take until comma or newline
		end := strings.IndexAny(s, ",\n")
		if end < 0 {
			return strings.TrimSpace(s)
		}
		return strings.TrimSpace(s[:end])
	}

	depth := 0
	inStr := false
	escaped := false
	for i := 0; i < len(s); i++ {
		c := s[i]
		if escaped {
			escaped = false
			continue
		}
		if c == '\\' && inStr {
			escaped = true
			continue
		}
		if c == '"' {
			inStr = !inStr
			continue
		}
		if inStr {
			continue
		}
		if c == open {
			depth++
		} else if c == close {
			depth--
			if depth == 0 {
				return s[:i+1]
			}
		}
	}
	return ""
}

// parseExtraKV parses "key=value, key2=value2" into JSON fields (without outer braces).
func parseExtraKV(s string) string {
	s = strings.TrimSpace(s)
	if s == "" {
		return ""
	}
	var parts []string
	for s != "" {
		eqIdx := strings.IndexByte(s, '=')
		if eqIdx < 0 {
			break
		}
		key := strings.TrimSpace(s[:eqIdx])
		s = strings.TrimSpace(s[eqIdx+1:])
		val := extractBalanced(s)
		if val == "" {
			break
		}
		s = strings.TrimSpace(s[len(val):])
		if len(s) > 0 && s[0] == ',' {
			s = s[1:]
		}
		keyJSON, _ := json.Marshal(key)
		parts = append(parts, string(keyJSON)+":"+relaxJSON(val))
	}
	return strings.Join(parts, ",")
}

// extractJSONObject finds the first balanced {...} in s.
func extractJSONObject(s string) string {
	start := strings.IndexByte(s, '{')
	if start < 0 {
		return ""
	}
	depth := 0
	inStr := false
	escaped := false
	for i := start; i < len(s); i++ {
		c := s[i]
		if escaped {
			escaped = false
			continue
		}
		if c == '\\' && inStr {
			escaped = true
			continue
		}
		if c == '"' {
			inStr = !inStr
			continue
		}
		if inStr {
			continue
		}
		if c == '{' {
			depth++
		} else if c == '}' {
			depth--
			if depth == 0 {
				return s[start : i+1]
			}
		}
	}
	return ""
}

// relaxJSON converts common relaxed-JSON patterns to valid JSON:
// - unquoted keys (word: → "word":)
// - single-quoted strings → double-quoted
// This is a best-effort heuristic for model output.
func relaxJSON(s string) string {
	var out strings.Builder
	out.Grow(len(s))
	i := 0
	for i < len(s) {
		c := s[i]
		// Single-quoted string → double-quoted
		if c == '\'' {
			out.WriteByte('"')
			i++
			for i < len(s) && s[i] != '\'' {
				if s[i] == '"' {
					out.WriteString(`\"`)
				} else if s[i] == '\\' && i+1 < len(s) {
					out.WriteByte(s[i])
					i++
					out.WriteByte(s[i])
				} else {
					out.WriteByte(s[i])
				}
				i++
			}
			out.WriteByte('"')
			if i < len(s) {
				i++ // skip closing '
			}
			continue
		}
		// Unquoted key before colon: word followed by optional whitespace then ':'
		if c >= 'a' && c <= 'z' || c >= 'A' && c <= 'Z' || c == '_' {
			j := i
			for j < len(s) && (s[j] >= 'a' && s[j] <= 'z' || s[j] >= 'A' && s[j] <= 'Z' || s[j] >= '0' && s[j] <= '9' || s[j] == '_') {
				j++
			}
			// Check if followed by optional whitespace then ':'
			k := j
			for k < len(s) && (s[k] == ' ' || s[k] == '\t') {
				k++
			}
			if k < len(s) && s[k] == ':' {
				// It's an unquoted key — quote it
				out.WriteByte('"')
				out.WriteString(s[i:j])
				out.WriteByte('"')
				i = j
				continue
			}
			// Not a key, just copy the word
			out.WriteString(s[i:j])
			i = j
			continue
		}
		out.WriteByte(c)
		i++
	}
	return out.String()
}

// classifyError returns a short string identifying the error type for the
// turnError hook payload.
func classifyError(err error) string {
	var rlErr *backend.RateLimitError
	if errors.As(err, &rlErr) {
		return "rate_limit"
	}
	var tuErr *backend.ToolUnsupportedError
	if errors.As(err, &tuErr) {
		return "tool_unsupported"
	}
	var coErr *backend.ContextOverflowError
	if errors.As(err, &coErr) {
		return "context_overflow"
	}
	var tErr *backend.TransientError
	if errors.As(err, &tErr) {
		return "transient"
	}
	return "unknown"
}

// transientWait returns the retry wait for a retryable error and whether it is
// retryable. Rate limits use longer waits; transient/network errors use shorter.
func transientWait(err error, attempt int) (time.Duration, bool) {
	var rlErr *backend.RateLimitError
	if errors.As(err, &rlErr) {
		wait := rlErr.RetryAfter
		if wait == 0 {
			wait = retryBaseDelay << attempt
		}
		return wait, true
	}
	var tErr *backend.TransientError
	if errors.As(err, &tErr) {
		return time.Duration(2<<attempt) * time.Second, true
	}
	return 0, false
}

func retryCountdown(ctx context.Context, cfg agentConfig, wait time.Duration) error {
	deadline := time.Now().Add(wait)
	for {
		remaining := time.Until(deadline)
		if remaining <= 0 {
			return nil
		}
		secs := int(remaining.Seconds()) + 1
		emit(cfg, Event{Role: "retry", Content: fmt.Sprintf("%d", secs)})
		select {
		case <-ctx.Done():
			return ctx.Err()
		case <-time.After(min(remaining, time.Second)):
		}
	}
}

// defaultToolResultMaxBytes caps tool result content sent back to the model.
const defaultToolResultMaxBytes = 131072

// toolClassifier reports whether a named tool is safe to run concurrently
// with other read-class toolsrv. nil means treat all tools as serial.
type toolClassifier func(name string) bool