b643e6ce64
* perf(mcp): demote ten unused catalog reads, reclaiming 3 763 tokens of budget The tools/list payload bench sat about 20 tokens under its 65 000 ceiling, so the next tool or field anyone added failed CI. That is not hypothetical: it already cost the agent-briefing field in #2079, which was built and then removed for want of roughly 85 tokens. The bench's own comments deferred the fix twice, in the same words both times: picking which read to demote needs prod usage data, not a guess inside an unrelated PR, so do the demotion as its own change and ratchet the ceiling back down. This is that change. Selection is from 60 days of mcp.tool_called: default-catalog READ tools with 25 or fewer calls, excluding anything named in RECOMMENDED_WORKFLOW_LOADOUTS. Ten qualified. Measured 65 046 -> 61 283, ceiling ratcheted to 61 600. Usage data is necessary but not sufficient, and three classes were kept in the catalog despite low counts: - gnubok_call_tool, which IS the search-only bridge. - the connect_* onboarding tools and gnubok_lookup_company. A fresh agent has not learned to search yet, which is exactly why #1936 put them in the catalog. - every arsredovisning, dispositioner, depreciation and accrual PROPOSAL tool. A 60-day window ending in August cannot see bokslut season: most Swedish companies close on 31 December and do the work January to June, so summer counts understate these to near zero. Demoting them would hide the year-end flow precisely when it is needed. Widget tools are a hard exclusion, and I found that by tripping over it: gnubok_receipt_matcher and gnubok_vat_review_widget were demoted in a first pass and their own suites failed. A widget is rendered from the _meta.ui it publishes in tools/list, so search-only leaves it callable but unrenderable. Both are back in the catalog, and a general invariant now holds the rule for every widget tool added later, verified by demoting one and watching it fail. Only READ tools are demotable at all: gnubok_call_tool refuses writes, so a search-only WRITE is uncallable on Claude.ai. Demoted reads stay callable through the bridge; they are no longer listed up front. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs: code-span the connect_* identifier in DECISIONS A bare asterisk opens an emphasis span in Markdown. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Accounted MCP server
JSON-RPC 2.0 server exposing the Accounted bookkeeping engine to MCP clients (Claude Desktop, Claude Code, etc.). Endpoint: /api/extensions/ext/mcp-server/mcp. Add ?tool_namespace=accounted for the Accounted tool names. Requests without it retain the legacy Gnubok namespace. OAuth and stdio bridges live alongside the API surface: see app/api/mcp-oauth/, packages/accounted-mcp/, and the compatibility package in packages/gnubok-mcp/.
Tool authoring contract
Enforced by tests in __tests__/: these are not style preferences, they're guard rails.
additionalProperties: falseon everyinputSchema. Guarded bystrict-schemas.test.ts. Forces clear rejections on hallucinated fields instead of silent ignores.- Descriptions ≤ 280 chars. Guarded by
output-schema.test.ts. NoArgs:/Returns:/Examples:prose: those belong in JSON Schema. Use agent-native hints ("Use to…", "Call X first", "HIGH risk"). - Staged-operation envelope for write tools:
outputSchema: STAGED_OPERATION_SCHEMA(server.ts). Fields:staged, risk_level, actor, message, preview, period_status?, next?. Thestaged: trueboolean is the explicit completion signal; agents must not infer completion from prose. Do NOT introduce a parallel{ success, shouldContinue, output }envelope. period_statusthreading: any tool that ties to a fiscal-period-bound date (categorize, mark paid, create voucher, correct/reverse entry, approve supplier invoice) passesdateForPeriodChecktostagePendingOperation. Response then includesperiod_status: { period_id, status: open|locked|closed, lock_date }so widgets and agents disable writes without round-trips.- Scope mapping: every new tool needs an entry in
lib/auth/api-keys.tsTOOL_SCOPE_MAP. Missing entries default to deny. - Tests for new write tools: add staging-gate coverage to
__tests__/voucher-tools.test.ts(or a sibling) plus executor coverage tolib/pending-operations/__tests__/voucher-executors.test.tsif the tool stages a newoperation_type.
Determinism / cache stability
Tool definitions (name, description, inputSchema, outputSchema, annotations) are declared as static object literals at module load: no timestamps, no UUIDs, no Date/Math.random in the definition layer. This makes the tools/list JSON payload byte-stable across requests, which lets agent-side prompt caches stay warm. Do not introduce per-request non-determinism into the definitions block. Anything time-bound or random belongs inside execute().
For internal Anthropic API usage (the SDK is called from lib/ai/provider.ts and lib/ai/services/anthropic-family.ts; features such as extensions/general/invoice-inbox/lib/extract-invoice-fields.ts go through getAiService() in lib/ai rather than the SDK directly): annotate stable prefixes with cache_control: { type: 'ephemeral' } and log usage.cache_read_input_tokens for hit-ratio observability. The 1h TTL from the agent-native API plan (item 10) requires the direct Anthropic API; Accounted's Bedrock path defaults to a shorter TTL.
Payload-size watchdog
payload-size.bench.test.ts enforces a tools/list JSON payload ceiling. If the test fires, the right answer is rarely "raise the ceiling". Instead, trim descriptions or set specialized wide tools to catalogVisibility: 'search'. Those tools remain discoverable with full schemas through gnubok_search_tools and callable through tools/call on the wire without bloating the default catalog. Claude.ai only calls tools present in tools/list, so a tool that a user or a skill must call directly stays in the default catalog.
Where things live
server.ts: the tools array + JSON-RPC dispatchertool-result.ts:withNext(),toToolError()response helpersresources/: read-onlyAccounted://URIs, registered inresources/index.ts:company/current,period/active,recent-activity,capabilities,attention,chart-of-accounts,settings/vat-treatments,booking-templates,ledger/context,reconciliation/summarywidgets/: inline HTML widgets (receipt-matcher, vat-review, pending-operations)prompts/: slash-command-style promptsskills/: domain-knowledge skill bodies served viagnubok_load_skillpublic-tools.ts: lazy authentication (issue #1814).ANONYMOUS_METHODS(initialize, ping, tools/prompts/resources listing) and the threePUBLIC_TOOLS(gnubok_search_tools,gnubok_list_skills,gnubok_load_skill) answer without credentials, rate-limited per truncated IP; every othertools/callgets a transport-level 401 +WWW-AuthenticatefromhandleMcpRequestinserver.ts, which is the challenge clients turn into their Connect prompttasks.ts: MCP Tasks extension (io.modelcontextprotocol/tasks): durable handles for long-running tool calls, rows inmcp_tasks(service-role writes only)origin-guard.ts: Origin-header validation on the Streamable HTTP endpoint (DNS-rebinding defence required by the MCP spec)staging-pii-guard.ts: refuses a plaintext personnummer in stagedpending_operationsparams/preview, so every staging tool inherits the encrypt-at-staging ruletool-namespace.ts:?tool_namespace=accountedhandling (accounted_*aliases for the canonicalgnubok_*ids)__tests__/: strictness guards + per-tool coverage