Agent runtime: majordomo in-process, Ollama Cloud config, chat endpoint (#56)

Everything below the run loop already existed. This is the thing that runs a
model.

The build tag is gone, deliberately. internal/agent's doc comment promised two
separations — cmd/pansy not importing the package, and the tool wiring behind
//go:build majordomo — and both have been rewritten rather than left as a stale
aspiration. A tag that keeps the agent out of the binary only earns its keep if
you would ever ship a build without the agent, and the agent is the point;
keeping it meant an untagged CI that never compiled the code that matters.
majordomo is a real dependency now, resolved from the Gitea instance as a
pseudo-version with no replace directive, so the Docker build (which has no
sibling checkout) resolves it the same way this machine does. It is stdlib-first
and pure Go, so CGO_ENABLED=0 and the single static binary survive.

A TURN IS ONE CHANGE SET. That is the whole reason acting without a confirmation
prompt is defensible: "empty the garlic bed and plant cucumbers" is one object
edit and a dozen planting inserts, and it has to undo as one action rather than
thirteen. The scope is opened even for a turn that turns out to be a question,
because a change set with no revisions is never written — so asking costs
nothing and history isn't littered with empty entries.

The model spec goes to majordomo.Parse verbatim. That grammar, including
comma-separated failover chains, is majordomo's; re-implementing any of it here
would only mean two places to update when it grows. The key needs a bridge
though: majordomo's ollama-cloud preset reads OLLAMA_API_KEY while pansy (like
gadfly) is configured with OLLAMA_CLOUD_API_KEY, so the provider is registered
explicitly on a private registry rather than depending on ambient environment.

Runs are bounded by a step cap, a timeout and majordomo's loop guards. This is
loop safety, not cost control — pansy is a personal tool and spend caps are
explicitly not a v2 concern. A capped run does NOT fail: it kept whatever it
managed to do, that work is recorded and undoable, and the reply says it stopped
early rather than going silent.

The chat endpoint streams. A turn that clears a bed and replants it makes a
dozen tool calls over tens of seconds, and without streaming that is a long
silence followed by everything at once — which reads as a hang, and defeats a
design that rests on watching the canvas change as it happens.

Conversations persist per (user, garden). Client-held history would be lost on a
refresh, which is exactly when someone reloads to check whether the agent's
change landed. Only the user/assistant TEXT is stored, not the model's full
transcript: continuity needs what was said and what came back, and replaying a
stored tool call would replay a decision made against a garden that has since
moved on. It also keeps majordomo's message shape out of the schema.

An instance with no key starts, serves the app, and doesn't advertise the agent
— the routes aren't registered at all, the same shape as OIDC 404ing when
unconfigured. A configured-but-unresolvable model logs and disables the
assistant rather than refusing to boot: a garden planner that won't start
because of a chat feature is worse than one without chat.

Tool refusals reach the model as tool results it can explain, not 500s. The ACL
story only works if it can narrate the refusal.

Closes #56

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01H3zbym8Doka2d7D48maSgZ

internal/api/agent.go

@@ -0,0 +1,173 @@
+package api
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"net/http"
+
+	mdagent "gitea.stevedudenhoeffer.com/steve/majordomo/agent"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"github.com/gin-gonic/gin"
+
+	"gitea.stevedudenhoeffer.com/steve/pansy/internal/agent"
+	"gitea.stevedudenhoeffer.com/steve/pansy/internal/domain"
+)
+
+// The garden assistant's chat surface (#56).
+//
+// Streaming, because a turn that clears a bed and replants it makes a dozen tool
+// calls over tens of seconds. Without streaming that is a long silence followed
+// by everything at once, which reads as a hang — and the whole design rests on
+// watching the canvas change as it happens.
+
+// chatRequest is the body of POST /agent/chat.
+type chatRequest struct {
+	GardenID int64  `json:"gardenId" binding:"required"`
+	Message  string `json:"message" binding:"required"`
+}
+
+// chatEvent is one server-sent event. Exactly one field is set.
+type chatEvent struct {
+	// Step reports a completed model round trip: which tools it called.
+	Step *stepEvent `json:"step,omitempty"`
+	// Done carries the finished turn.
+	Done *agent.Turn `json:"done,omitempty"`
+	// Error is a turn that failed, in words meant for a person.
+	Error string `json:"error,omitempty"`
+}
+
+type stepEvent struct {
+	Index int      `json:"index"`
+	Tools []string `json:"tools"`
+}
+
+func (h *handlers) agentChat(c *gin.Context) {
+	var req chatRequest
+	if err := c.ShouldBindJSON(&req); err != nil {
+		writeAPIError(c, http.StatusBadRequest, "INVALID_INPUT", "a gardenId and a message are required")
+		return
+	}
+	actor := mustActor(c)
+
+	history, err := h.svc.AgentHistory(c.Request.Context(), actor.ID, req.GardenID)
+	if err != nil {
+		writeServiceError(c, err)
+		return
+	}
+
+	// Headers before the first write, and a flush straight away: a proxy that
+	// buffers the response would reintroduce exactly the silence streaming is
+	// here to remove.
+	c.Header("Content-Type", "text/event-stream")
+	c.Header("Cache-Control", "no-cache")
+	c.Header("X-Accel-Buffering", "no")
+	c.Writer.Flush()
+
+	send := func(ev chatEvent) {
+		b, err := json.Marshal(ev)
+		if err != nil {
+			slog.Error("api: encode chat event", "error", err)
+			return
+		}
+		_, _ = fmt.Fprintf(c.Writer, "data: %s\n\n", b)
+		c.Writer.Flush()
+	}
+
+	turn, err := h.agent.Run(c.Request.Context(), actor.ID, req.GardenID, req.Message,
+		replayHistory(history),
+		func(s mdagent.Step) {
+			send(chatEvent{Step: &stepEvent{Index: s.Index, Tools: toolNames(s)}})
+		})
+	if err != nil {
+		// The stream is already open, so an error is an event rather than a
+		// status code — the client has committed to reading a stream by now.
+		send(chatEvent{Error: chatErrorMessage(err)})
+		return
+	}
+
+	// Record before announcing: if persistence fails, the user should see that
+	// their turn wasn't saved rather than a clean "done" followed by a thread
+	// that has forgotten it.
+	if _, err := h.svc.RecordAgentExchange(c.Request.Context(), actor.ID, req.GardenID,
+		req.Message, turn.Reply, turn.ChangeSetID); err != nil {
+		slog.Error("api: record agent exchange", "error", err, "garden", req.GardenID)
+	}
+	send(chatEvent{Done: turn})
+}
+
+// getAgentHistory returns the actor's thread for a garden.
+func (h *handlers) getAgentHistory(c *gin.Context) {
+	gardenID, ok := parseIDParam(c, "id")
+	if !ok {
+		return
+	}
+	msgs, err := h.svc.AgentHistory(c.Request.Context(), mustActor(c).ID, gardenID)
+	if err != nil {
+		writeServiceError(c, err)
+		return
+	}
+	c.JSON(http.StatusOK, gin.H{"messages": msgs})
+}
+
+// deleteAgentHistory is the "start over" escape hatch.
+func (h *handlers) deleteAgentHistory(c *gin.Context) {
+	gardenID, ok := parseIDParam(c, "id")
+	if !ok {
+		return
+	}
+	if err := h.svc.ClearAgentHistory(c.Request.Context(), mustActor(c).ID, gardenID); err != nil {
+		writeServiceError(c, err)
+		return
+	}
+	c.Status(http.StatusNoContent)
+}
+
+// replayHistory turns stored text into model messages.
+//
+// Only the text is replayed — no stored tool calls. Continuity needs what was
+// said and what came back; replaying a tool call would be replaying a decision
+// made against a garden that has since moved on.
+func replayHistory(msgs []domain.AgentMessage) []llm.Message {
+	out := make([]llm.Message, 0, len(msgs))
+	for _, m := range msgs {
+		if m.Role == domain.AgentRoleUser {
+			out = append(out, llm.UserText(m.Body))
+		} else {
+			out = append(out, llm.AssistantText(m.Body))
+		}
+	}
+	return out
+}
+
+func toolNames(s mdagent.Step) []string {
+	names := make([]string, 0, len(s.Results))
+	for _, r := range s.Results {
+		names = append(names, r.Name)
+	}
+	return names
+}
+
+// chatErrorMessage turns a failure into something worth reading. Permission
+// errors in particular get named: "the agent broke" and "you can only view this
+// garden" want very different reactions.
+func chatErrorMessage(err error) string {
+	switch {
+	case errors.Is(err, domain.ErrForbidden):
+		return "You can only view this garden, so I can't change anything in it."
+	case errors.Is(err, domain.ErrNotFound):
+		return "I can't find that garden."
+	case errors.Is(err, domain.ErrInvalidInput):
+		return "I didn't get a message to work from."
+	case errors.Is(err, io.EOF), errors.Is(err, context.Canceled):
+		return "The connection dropped partway through. Anything I'd already changed is on the canvas, and in History."
+	case errors.Is(err, context.DeadlineExceeded):
+		return "That took too long and I stopped. Anything I'd already changed is on the canvas, and in History."
+	default:
+		slog.Error("api: agent run failed", "error", err)
+		return "Something went wrong talking to the model. Anything I'd already changed is on the canvas, and in History."
+	}
+}