attachSeedRemaining summed Remaining across lots in different units and then
dropped only the unit LABEL, leaving a bare number that reads as a quantity and
isn't one — 50 seeds plus 50 grams reported as "100". Worse for a model than for
a person, since it has nothing else to go on. The total is now omitted with the
label, and SeedLots reports how many lots there are, so "several lots, no single
total" stays distinguishable from "no seed at all".
Also folded TestToolboxScenario's inline setup into the newAgentTestService
helper the new tests use, rather than leaving two copies of the same six lines.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01H3zbym8Doka2d7D48maSgZ
"Change the garlic garden bed to instead be cucumbers this year" was nearly
executable already — describe_garden resolves "the garlic bed" to an object id
because it reports the NAMES of what's planted, clear_object empties it, and
fill_region replants it. Step three had no way to get its plantId. The toolbox
had no plant tool at all, so the agent could see the string "Garlic" come out of
describe_garden and had no route from "cucumbers" to an id. One missing tool
blocked the whole flagship interaction.
find_plant matches the actor's visible catalog and deliberately returns SEVERAL
candidates rather than one guess. "garlic" against a catalog holding both
"Garlic" and "German Red Garlic" is genuinely ambiguous, and a caller holding
the surrounding conversation is far better placed to disambiguate than a
fuzzy-match heuristic here. Results are ranked exact, then prefix, then
substring, then category, and ordered stably — a tool whose results reshuffle
between calls is one a model can't reason about across turns.
Each result also reports how much seed is left of that plant across the actor's
lots, so the agent can say "you only have enough for half that bed" instead of
confidently planting seed that doesn't exist. Omitted rather than zeroed when
there are no lots, and dropped when lots disagree about the unit, since summing
seeds and grams would be a lie.
create_plant lets a variety be named in conversation without leaving the chat.
It's user-scoped, not garden-scoped — someone with no editable garden can still
name a plant — which the tests assert deliberately rather than assume.
add_journal_entry arrived with #52: "note that the west bed has mildew" is
squarely the kind of thing you say out loud while walking around.
Also reviewed the descriptions on the existing seven tools, since they are the
model's only documentation. fill_region's region vocabulary now says north is
the top of the garden, gives a worked example of the replant sequence, and notes
that filling twice is safe.
The demo sequence is now a test: describe_garden → find_plant("cucumber") →
clear_object → fill_region, ending with a bed of cucumbers and no active garlic.
go.mod is deliberately untouched — majordomo stays out of the module until #56
drops the build tag, so the default build is unaffected.
Closes#55
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01H3zbym8Doka2d7D48maSgZ
steveandClaude Opus 4.8