Repository navigation
fix: pricing needs a comps source; chat sales go to mark_sold, not the paste tool (0.16.4) - #31
Conversation
…e paste tool (0.16.4) Price the sell pile on an agent without the ebay plugin (a Merchant archetype agent) started a turn with no comps tool: at best everything came back unpriced, at worst it burned turns on web searches. start_pricing now checks sdk.config().plugins_enabled and, without ebay, closes the pricing at once with every piece unpriced and the reason (Set price by hand). An unknown host keeps the old behavior. The tool-selection eval showed both gateway models picking the turn-only inventory_sale_candidates first for a chat "sold the X for $Y". Its description now scopes it to the view's paste-a-sale turns and sends chat sales to inventory_list / inventory_get and inventory_mark_sold. Re-run: fast 81.1% -> 86.8% (sale 0/5 -> 3/5), smart 90.6% (sale 1/5 -> 3/5), turn-only picks 0 on both; the rest are read-before-acting. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
QA panel review — PASS
code-review-structural · head a2167b90124d · formal
Low-risk PR: adds a comps-availability guard to start_pricing and re-scopes the inventory_sale_candidates docstring. Production logic is clean across all lanes. The single surviving item is a stale docs row in docs/tools.md that still describes the tool with its pre-PR scope — minor, confirmed by the verifier. No panel disagreement; no coverage gaps of concern.
Findings
| Severity | Location | Finding | Verified | |
|---|---|---|---|---|
| 🟡 | minor | docs/tools.md |
The human-facing tool reference for inventory_sale_candidates was not updated to reflect the PR's re-scoping, so it still presents the tool as a general 'find … | confirmed |
findings JSON (machine-readable)
[
{
"file": "docs/tools.md",
"line": 0,
"severity": "minor",
"category": "conventions",
"claim": "The human-facing tool reference for inventory_sale_candidates was not updated to reflect the PR's re-scoping, so it still presents the tool as a general 'find items matching what sold' tool and omits the new paste-turn-only scope and the chat-sales \u2192 inventory_mark_sold routing.",
"evidence": "docs/tools.md row (unchanged by the PR): `| inventory_sale_candidates | Find items matching what sold. Items sold in the last 180 days come back flagged with that sale, so a repeat paste is caught |` \u2014 while tools.py's new description now opens `\"\"\"ONLY for a paste-a-sale turn the Inventory view starts (its prompt names a paste id). When the operator tells you about a sale in chat, find the item with inventory_list or inventory_get and record it with inventory_mark_sold instead.`",
"verdict": "confirmed",
"note": "Read docs/tools.md at head SHA: the 'Turn tools: pasted sales' table still has `| inventory_sale_candidates | Find items matching what sold. Items sold in the last 180 days come back flagged with that sale, so a repeat paste is caught |`. The PR diff modifies only tools.py's docstring, not this file. The new scope and chat-sales routing are absent from the docs row."
}
]1 finding(s) excluded from the verdict by in-diff confinement (file not among this PR's changed paths):
docs/tools.md(minor) — The human-facing tool reference for inventory_sale_candidates was not updated to reflect the PR's re-scoping, so it still presents the tool as a general 'find i
There was a problem hiding this comment.
Promoting the PASS verdict for head a2167b90124d: all checks terminal-green, zero unresolved review threads. (approve-on-green)
Open findings carried by this approval — non-blocking, but they did not go away:
- minor
docs/tools.md:0— The human-facing tool reference for inventory_sale_candidates was not updated to reflect the PR's re-scoping, so it still presents the tool as a general 'find items matching what sold' tool and omits
Approving a WARN does not resolve its findings (issue #22).
… use mark_sold Review on #31. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
QA panel review — PASS
code-review-structural · head 7acb32c3db2a · formal
No findings were raised, so there was nothing to verify. The panel's synthesized result was an empty array and the verifier confirmed nothing-to-verify. One note: the prior round-1 minor finding on docs/tools.md (tool reference not updated for the re-scoping) is still marked open in the ledger, even though the prior round's brief claimed it was resolved in this PR — the author may want to confirm whether that doc update actually landed. No verification gaps.
No findings — the review came back clean.
findings JSON (machine-readable)
[]Addressed since the last round (1 minor/nit): docs/tools.md:0 — no longer raised, and the PR changed the code around them since they were raised.
Two fixes found while making Inventory first class.
Pricing needs a comps source. On an agent without the ebay plugin, such as one created from the new Merchant archetype, Price the sell pile started a turn with no comps tool. At best every piece came back unpriced; at worst the agent burned turns on web searches.
start_pricingnow checkssdk.config().plugins_enabled. Withoutebayit closes the pricing at once, with every piece unpriced and the reason ("set a price by hand with Set price"), and starts no agent turn. A host that can't say keeps the old behavior.Chat sales go to
inventory_mark_sold. The tool-selection eval (#30) showed both gateway models picking the turn-onlyinventory_sale_candidatesfirst for a chat "sold the X for $Y". Its description now scopes it to the view's paste-a-sale turns.protolabs/fastprotolabs/smartThe remaining misses are all read-before-acting (
inventory_getorinventory_listfirst), which costs a step and writes nothing.Tests: the no-comps path (no turn started, every piece unpriced with the reason), comps-present and unknown-host paths, and a pin on the description. 484 passed; ruff clean.
docs/upgrading.mdhas notes for 0.16.3 and 0.16.4.🤖 Generated with Claude Code