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

DESIGN.md

@@ -23,6 +23,7 @@ SQLite, centimeters everywhere (display-side imperial conversion only), `version
 - **garden_objects** — one polymorphic table: `kind` (`bed`|`grow_bag`|`container`|`in_ground`|`tree`|`path`|`structure`), name, `shape` (`rect`|`circle`; `polygon` + `points` JSON reserved in schema, not in the v1 editor), center `x_cm`,`y_cm` in garden space, `width_cm`,`height_cm` (circle: width=diameter), `rotation_deg`, `z_index`, `plantable` bool, nullable `color` override, `props` JSON for kind-specific extras (bed height, tree species…), notes.
 - **seed_lots** — one purchase: vendor, source URL, sku, lot code, purchase date, packed-for year, quantity + unit, cost, germination %. `owner_id` is on the LOT, not inherited from the plant, so you can record buying generic built-in Garlic from Johnny's and keep it private. `plantings.seed_lot_id` (ON DELETE SET NULL) attributes a plop to a purchase; **`remaining` is derived** (quantity − summed effective count of active plantings), never stored — a decremented column drifts the moment a planting is edited behind its back and can't be audited afterwards. Lots are never shared along with a garden.
 - **plants** — catalog, plus `source_url`/`vendor` provenance for the variety itself. `owner_id NULL` = built-in (~30 seeded via migration, read-only, clone to customize). name, category (`vegetable`|`herb`|`flower`|`fruit`|`tree_shrub`|`cover`), `spacing_cm` (mature spacing), `color` hex, `icon` (emoji string — zero assets in v1), nullable `days_to_maturity`, notes. Users see built-ins + their own.
+- **agent_conversations** / **agent_messages** — the assistant's chat, one thread per (user, garden). Only the user/assistant TEXT of each turn 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. `agent_messages.change_set_id` is the handle the chat panel uses to offer Undo on the turn that caused it.
 - **journal_entries** — the grow log: `garden_id` (NOT NULL), nullable `object_id`/`planting_id` to narrow the target, author, body, `observed_at`. `notes` on a garden/object/plant says what the thing IS and a new note overwrites the old; a journal entry says what HAPPENED and when, and entries accumulate. `garden_id` is set even for a bed- or plop-level entry so permission checks reuse `requireGardenRole` unchanged rather than inventing a second ACL path. `observed_at` is distinct from `created_at` — you write up Saturday's observations on Sunday. Deliberately **not** in the revision history: entries are already append-shaped and individually versioned.
 - **change_sets** / **revisions** — the undo substrate. A change set groups the row-level revisions one logical operation produced (`source` ui/agent/api, `summary`, nullable `agent_run_id`, nullable `reverts_id`); a revision is `(entity_type, entity_id, op, before, after)` with full JSON row snapshots. The unit of undo is the **operation, not the row**. Revert is itself a change set, so an undo can be undone; it is version-guarded per entity and reports conflicts rather than clobbering a later edit. Garden *deletion* is deliberately not revertible (the cascade is invisible to the service, and would take the history with it).
 - **plantings** ("plops") — object_id, plant_id, center `x_cm`,`y_cm` **in the parent object's local frame** (moving/rotating a bed moves its plants for free), `radius_cm` (a plop is a circular patch), nullable `count` (NULL = derived: `max(1, round(π·r² / spacing_cm²))`), nullable label, `planted_at`/`removed_at` dates. "Clear bed" sets `removed_at`; v1 shows rows where `removed_at IS NULL`. Year-over-year history and a future plan/scenario layer build on these dates without schema surgery.
@@ -64,6 +65,10 @@ POST /gardens/:id/objects     PATCH,DELETE /objects/:id
 POST /objects/:id/plantings   PATCH,DELETE /plantings/:id
 GET,POST /plants              PATCH,DELETE /plants/:id            (own plants only)
 GET,POST /seed-lots           GET,PATCH,DELETE /seed-lots/:id     (own lots only; private)
+GET,POST /gardens/:id/journal PATCH,DELETE /journal/:id           (editor writes; author edits own)
+GET  /gardens/:id/journal/counts  ← entries per object, for the "has notes" indicator
+POST /agent/chat              ← SSE: step events, then the finished turn (editor only)
+GET,DELETE /gardens/:id/agent/history                             (the actor's own thread)
 GET,POST /gardens/:id/shares  PATCH,DELETE /gardens/:id/shares/:userId    (invite by email)
 ```
 
@@ -77,6 +82,7 @@ GET,POST /gardens/:id/shares  PATCH,DELETE /gardens/:id/shares/:userId    (invit
 cmd/pansy/main.go        config load → store.Open → migrate → service → api → ListenAndServe
 internal/config/         PANSY_PORT, PANSY_DB, PANSY_BASE_URL, PANSY_REGISTRATION,
                          PANSY_OIDC_* (issuer/client_id/secret/button label), PANSY_LOCAL_AUTH,
+                         PANSY_AGENT_MODEL / OLLAMA_CLOUD_API_KEY / PANSY_AGENT_ENABLED,
                          trusted proxies
 internal/store/          sqlite.go (open/pragmas), migrate.go, migrations/*.sql (embedded),
                          queries per entity (users.go, gardens.go, objects.go, plantings.go,
@@ -89,7 +95,8 @@ internal/service/        ★ THE SEAM. All permission checks + business ops. aut
 internal/api/            thin gin handlers (decode → service → encode), session middleware,
                          ginserver setup, routes.go, oidc.go, spa.go (embed fallback)
 internal/webdist/        dist.go: //go:embed all:dist   (web build copied in by Makefile)
-internal/agent/          later: llm.DefineTool[Args] wrappers over the SAME service methods
+internal/agent/          tools.go (llm.DefineTool wrappers over the SAME service methods),
+                         runtime.go (model resolution, run loop, one change set per turn)
 web/                     Vite app
 Makefile                 (cd web && npm run build) → copy dist → CGO_ENABLED=0 go build
 ```
@@ -123,7 +130,8 @@ React 19 + TypeScript + Vite + Tailwind 4 (`@tailwindcss/vite`), `@tanstack/reac
 6. **Plops** — focus-zoom, place/move/resize, derived counts, semantic zoom. *The core vision lands here.*
 7. **Sharing** — invite by email, roles, viewer read-only mode.
 8. **Polish** — imperial toggle, mobile ergonomics, clear-bed, keyboard nudging.
-9. **Agent seam** — `ops.go` bulk ops + `internal/agent` DefineTool wrappers; optional mort-style API-key agent endpoint. The agent harness itself lives in majordomo/executus, outside pansy.
+9. **Agent seam** — `ops.go` bulk ops + `internal/agent` DefineTool wrappers.
+10. **Garden assistant** — majordomo in-process, Ollama Cloud, streaming chat. Each turn runs inside ONE change set (`source='agent'`), so a turn that clears a bed and replants it undoes as one action; that is what makes acting without a confirmation prompt defensible. Bounded by a step cap and a timeout — loop safety, not spend control. The `majordomo` build tag is gone: a tag that keeps the agent out of the binary only earns its keep if you'd ship a build without it, and the agent is the point.
 
 ## Deliberate v1 limits