feat(mcp): restore merchant_summary tool#1843
Closed
canalesb93 wants to merge 8 commits into
Closed
Conversation
…actions Agents rarely set the optional fields param, so query_transactions returned all ~22 fields per row by default — most unused. Make the MCP default a compact projection (core,category); pass fields=all for the full struct. This is MCP-only: the REST API shares service.ParseFields and keeps returning full objects when ?fields= is omitted. Also hoist iso_currency_code to a single top-level field when every returned row shares one currency (the common single-currency household case), with a safe per-row fallback the moment a response mixes currencies. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…transaction_rules Extend the field-projection mechanism (previously transaction-only) to two heavy read tools. fields.go's parser is generalized into parseFieldsWith so each entity supplies its own valid-field/alias maps; ParseFields stays a stable wrapper so the REST API is untouched. - list_series: lean "overview" default (identity + renewal prediction) drops the verbose detection_signals blob; fields=all or get_series for full detail. - list_transaction_rules: lean "summary" default drops the conditions/actions trees; fields=all restores the full definition. Both are MCP-tool-only; the breadbox://rules resource still returns full, and the tool<->resource parity test now asserts that invariant via fields=all. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a safety ceiling on any single MCP tool response, enforced at the jsonResult chokepoint on the post-compaction wire bytes. Over the cap, the tool returns an actionable RESPONSE_TOO_LARGE error (narrow filters / lower limit / paginate / use fields) instead of dumping a context-blowing payload — the pattern GitHub's MCP server uses. Default 100 KB (~25K tokens). Override via BREADBOX_MCP_MAX_RESPONSE_BYTES (env, consistent with config precedence); set to 0 to disable. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…esponse cap Keep the canonical MCP tools reference in sync with the token-optimization changes: query_transactions / list_series / list_transaction_rules lean defaults + fields=all, top-level currency hoisting, and the per-response byte cap with its BREADBOX_MCP_MAX_RESPONSE_BYTES override. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Adds query_transaction_rules — the filterable/sortable analogue of query_transactions for the rule set — and fixes the agent sidecar so its Bash/Grep analysis tools actually work inside the container. query_transaction_rules: filter by category_slug / enabled / trigger / creator_type / search / min_hit_count / only_unused; sort by priority (default) | hit_count | last_hit_at | created_at | name. Lean-by-default summary projection + response cap, shared with list_transaction_rules. Cursor pagination is emitted only for the default priority sort; an explicit sort_by returns a single top-N page. Lets a reviewer audit coverage and prune dead rules without dumping the whole roster (which overflowed the tool-result cap in prod). Sidecar shell: the Alpine runtime image shipped no bash and no ripgrep and left SHELL unset, so an agent that tried to grep a large saved tool-result hit "No suitable shell found" / vendored-rg ENOENT and was left unable to do file-level analysis. Install bash + ripgrep, set SHELL=/bin/bash and USE_BUILTIN_RIPGREP=0 (the SDK's bundled rg isn't embedded in the Bun --compile output), with defensive defaults in the sidecar itself for non-Docker (per-user worktree) installs. Also fixes the routine-review prompt's reference to the phantom merchant_summary tool (no merchant grouping exists; the small queue is already in context) and points rule-coverage analysis at the new tool. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Six agent prompts (merchant-analysis, strategy-spending-report, strategy-rule-foundation, strategy-anomaly-detection, strategy-large-charge-sentinel, review-depth-thorough) and docs/mcp-tools-reference.md all reference a merchant_summary MCP tool that wasn't registered. The service method (GetMerchantSummary) and the REST endpoint (/transactions/merchants) already exist — only the MCP wiring was missing — so every prompt that told an agent to "use merchant_summary" hit an unknown tool and the agent improvised. Re-register the tool: per-merchant aggregation (count, total, avg, first/last date) over a window, with min_count / spending_only / category_slug / account_id / user_id / amount-range / search (fuzzy) filters. It's the merchant-grouped companion to transaction_summary (which groups by category/time, not merchant), backed by the same service method the REST endpoint uses. The admin tool-label map already listed merchant_summary; docs already documented it. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
canalesb93
added
stacked
Owner
Author
|
Closing — the decision changed: we want merchant_summary removed and scrubbed, not restored. A better merchant-data offering is planned for the future, so the phantom references are being cleaned up instead (separate PR off main). See follow-up. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follow-up to #1841. Restores the
merchant_summaryMCP tool.While auditing the Routine Reviewer's prompts I found
merchant_summaryreferenced as a tool in six agent prompts and indocs/mcp-tools-reference.md, but it was not registered in the MCP server — agents told to "use merchant_summary" hit an unknown tool and improvised. The building blocks already existed:GetMerchantSummary(internal/service/transaction_summary.go) — production-grade, fully parameterized.GET /transactions/merchantsalready exposes it.docs/mcp-tools-reference.mdalready documents it, and the admin tool-label map already lists it.Only the MCP wiring was missing. This re-registers it.
What it does
Per-merchant aggregation over a window (default last 90 days) — the merchant-grouped companion to
transaction_summary(which groups by category/time, not merchant). Each row:merchant,transaction_count,total_amount,avg_amount,first_date,last_date,iso_currency_code(never summed across currencies; capped at 500).Filters:
min_count(2 = recurring, 3+ = subscriptions),spending_only,category_slug,account_id,user_id,min_amount/max_amount,search(+search_modeincl. fuzzy for bank-feed name variants),exclude_search.This makes the six prompts that already reference it correct, no prompt edits needed.
Tests
min_countexclusion.go build/go vetclean; full unit suite +mcp/adminintegration green.Stack
Stacked on #1841 (
agents/rule-query-and-agent-shell) since both touchserver.go/server_test.go/tools.go/ the docs reference — basing here keeps this diff to just the merchant change. Once #1841 squash-merges, GitHub retargets this tomaincleanly.🤖 Generated with Claude Code