Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@BattleGrid MCP Server List current market grid sessions"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
MCP server for BattleGrid — play crypto prediction games, author trading strategies, and manage intelligence agents from AI agents.
It is a thin, authenticated stdio proxy to BattleGrid's remote MCP server (Stripe @stripe/mcp pattern — no business logic). It discovers tools, prompts, and resources live from the server and re-exposes them to local MCP clients (Claude Desktop, Claude Code, Cursor). Capabilities are always discovered live — this package never hardcodes the tool catalog.
The package version no longer tracks the server's contract, and no longer claims to. Through v30 it did: the proxy announced battlegrid@<package version> downstream, so the published number was read as the contract a client would reach, and keeping the two in step was a manual release chore. It did not hold — the package sat at 11.0.0 against a deployed contract of 19.3.0 for ten days, 5.1.0 was declared here and never published at all, and four of the last five releases were version bumps carrying no code change.
From v31 the proxy reads the contract version out of the upstream handshake at connect time and relays it verbatim. What a local client reads is therefore what the connected server just announced — correct by construction, on every connection, with no release involved.
Two versions, two meanings, and they are expected to differ:
What it describes | Where to read it | |
Package version | This proxy's own code — a fix here, a dependency bump, a docs correction |
|
Contract version | The server's wire contract, live | The stdio handshake ( |
Seeing package 31.x alongside handshake battlegrid@33.x — the package behind the contract — is the system working, not a missed release. That is the pair that looks alarming and is not: the contract moved, and no release here was needed. Both numbers are printed to stderr at startup, labelled.
What this changes for you: nothing about how you call anything. Upgrading the package no longer waits on a server deploy, and a server deploy no longer strands you on a package that names the wrong contract — reconnect and the announcement follows. Contract breaking-change notes are no longer keyed to package versions, since a contract move is no longer a release here; the v11-and-earlier notes below are kept as history, and the live vocabulary is always discovery.
Related MCP server: limitless-mcp
Eleven majors reached authors while this section stopped at v36. That gap is the mechanism, not an oversight: since v31 a contract move needs no release here, so nothing forced a note to be written — and the documentation ships inside the tarball, so a note written but unpublished reaches nobody. Both halves are now closed by a rule keyed to the served contract rather than to a release of this package.
50.0.0 and 51.0.0 arrived late, and the reason is worth naming. The re-vendoring errand that
used to carry these notes is now a generated export
(battlegrid-app/server/scripts/export-mcp-skills.mjs), and it owns three paths — skills/,
skills/EXPORT.json, and the vendored digest. It deliberately does not touch this file. So the
digest kept arriving on time while the note stopped travelling with it, and this section sat at
v49.5 against a served contract of 51.0.0. Nothing a client could observe was wrong; what was
missing was the sentence telling them so. A contract move still needs a human-authored entry
here, and the export lane will not remind you.
Nothing to migrate. This is the one direction that cannot break a client: a body the server used to refuse is now stored. Listed because a client that special-cased the refusal can delete that branch.
A signal rule patch no longer forces you to restate allocation and required (51.0.0,
fix-signal-rule-patch-semantics). On update_strategy_signal_rule, and on the rules element
of compile_strategy_plan, both fields become optional and join params under ONE omission
rule: an omitted mutable field preserves the stored value for that signal. "Raise this
signal's weight" is now expressible.
{ "strategyId": "…", "expectedRevision": 7, "signalId": "volume_surge",
"allocation": 3 } // `required` and `params` keep exactly what is storedEvery existing client keeps working — a complete payload is still a valid patch — so this is
listed for what you can now STOP sending. Before it, both fields were mandatory on every rule
surface, so a caller that had not first read the current rule had to invent a value it was never
asked about. That is not hypothetical: revision 5 of a production strategy flipped required
false → true unasked while moving a weight 2 → 3, turning a scoring signal into a mandatory
gate — which changes whether the agent takes trades at all.
One boundary on the newly legal ground, and it breaks nothing: a patch carrying no mutable field is refused — "A rule patch must change something: supply at least one of allocation, required or params." — rather than minting a no-op revision. Under 50.0.0 that request could not be formed at all, so nothing that used to work is now refused.
An arming trigger no longer constrains its required conditions' clock (49.4.0,
restore-arming-trigger-authoring). compile_strategy_plan, apply_strategy_plan and
fork_strategy accept a strategy whose entry trigger is ON_CANDLE_CLOSE, STOP_THROUGH_LEVEL
or ON_RETEST while a required condition reads the LIVE clock. v48.1 announced that pairing
as rejected, naming the offending condition key on VALIDATION_ERROR; that refusal is gone.
Why it was withdrawn, since the reasoning matters more than the rule: it ran against the whole
assembled strategy, so it refused every edit to a strategy in that shape — a rename, one report
column, one signal weight — plus restore and fork. For a strategy whose required conditions read
columns that can never carry a CLOSE clock (zone distances, perp/spot flow), there was no legal
shape to move to at all. And its premise — that the pairing can never fire — was measured before
the arming lifecycle was corrected, and no longer holds.
The underlying question, should an entry that waits for a close be decided on a forming bar, is now settled where the decision is made rather than by refusing the author's declaration.
blocksScanGate reads true for a class it did not (50.0.0, own-scan-served-set-once), on
preview_radar_resolution. Blocking is now derived from the lane's served set rather than
switched over the reach reason, so a FEED-reason refusal on an operand no reader in the lane
serves BLOCKS instead of deferring. No field changes shape, and a client that already renders the
key renders the new answer — but a client that treated blocksScanGate: false as "this will
resolve once data arrives" now sees a deployment that will not fire. Nothing in the payload tells
you this moved.
Three tools serve different values for identical input (47.3.0, derive-scan-fetch-from-report).
The radar scan leg now derives its timeframe fetch from the strategy's report rather than the
on-duty agent's three perception rungs, so a required condition addressing an absolute timeframe
outside those rungs — never evaluated at scan before — now is. No field moves; the values do:
Tool | What moves |
|
|
|
|
|
|
Read the last one carefully: a client treating UNMEASURABLE as "this gate is switched off"
will now see that gate BLOCK. AGENT_TIMEFRAME keeps its member and narrows to the one cause
no fetch can discharge. Nothing in the payload tells you this moved.
A condition naming a swingHi / swingLo header is refused — the indicator is a Donchian
channel (53.0.0, rename-donchian-channel). The rolling-window extremes indicator computed the
highest high and lowest low of the trailing 20 closed bars — a Donchian channel — under a swing
point's name, and the name asserted a property the value does not have (a swing high survives being
broken; a channel edge re-anchors the instant it is). Every layer of the vocabulary moves at once,
with no alias: a clause on compile_strategy_plan, apply_strategy_plan or fork_strategy naming
a header on the old stems — dist_swingLo, dist_swingHi_4h, dist_swingLo_rank_near, any
timeframe- or rank-suffixed form — is refused as CONDITION_COLUMN_UNKNOWN where 52.0.0 accepted
it. The same shapes exist on the donchianHi / donchianLo stems, which
get_strategy_column_contract lists with the labels 20-bar high / 20-bar low. Every value is
the same number under its new name.
Rename the stems (swingHi → donchianHi, swingLo → donchianLo, suffixes unchanged) in
every condition and Market Read marker you author. That is the whole migration for what you send.
entry.levelSource is refused — the level is derived, never authored (52.0.0,
derive-entry-level). The strict entry object on compile_strategy_plan, apply_strategy_plan
and fork_strategy is six keys — trigger, confirmTf, closes, bandAtrMultiple,
levelOffsetAtrMultiple, validForBars — and a body carrying levelSource is refused naming
the key where 51.0.0 accepted it. STOP_THROUGH_LEVEL rests a stop past the swing channel's
CURRENT edge in the trade's direction (the 20-bar high for a long, the 20-bar low for a short);
ON_RETEST rests a limit in front of the edge a close most recently BROKE. Your two dials are the
unsigned distance from that edge (levelOffsetAtrMultiple, 0–2 ATR; a long adds, a short
subtracts) and the bar validity (validForBars, 1–24). A meaningful validForBars under
AT_SIGNAL or ON_CANDLE_CLOSE is now refused as PARAMETER_NOT_HONOURED like the other level
dials.
Drop the key. That is the whole migration for what you send; the six-key example below is the current shape.
update_strategy_signal_rule requires confirm: true when the strategy has bound agents
(51.0.0, fix-signal-rule-patch-semantics). The write re-materializes scoring configuration onto
every bound agent immediately — including agents holding open USDC positions — and until now
nothing on the server asked. A rule edit on a strategy with one or more bound agents is refused
without the flag, and the message names the count. An edit on a strategy with nothing bound is
unaffected, and so is every path through the web editor.
Send confirm: true. That is the whole migration, and confirm is published on the input
schema — but nothing in the schema says WHEN it becomes mandatory, because the condition is the
bound-agent count rather than the shape of your body. The refusal rides the existing
VALIDATION_ERROR code, so a client that omits it discovers the rule at the refusal.
Why it moved to the server: the guard existed, but only as served prose the calling model could
decline — and did, twice in production on 2026-08-24. archive_strategy and
rebind_intelligence_agent have taken a server-enforced confirm all along; single-rule tuning
was the outlier among its own siblings, and it is the one that writes to scoring.
A radar deployment is refused when its strategy reads a session-field scalar (50.0.0,
own-scan-served-set-once). upsert_radar_deployment refuses a deployment whose slot agents'
bound strategy carries a condition reading one of five SESSION-FIELD scalars — fieldPlayers,
fieldUpBias, fieldBiasDir, captConc, picksSpread — with
CONDITION_OPERAND_UNSERVED_IN_LANE. A body accepted under 49.5.0 is refused under 50.0.0
without one byte of it changing.
Why a refusal and not a warning. Those five describe a game SESSION, and radar runs outside a session at BOTH its stages — so such a condition can never resolve there. The deployment formed no fire edge and the agent did nothing on that coin, silently, forever. The refusal converts a permanent silence into an error at the moment you author it.
Migrate by moving the clause to a scalar radar reads — the Market Breadth or Reference Pairs
families, which are market-wide reads with no session dimension — or by binding the strategy to
an arena agent instead. The error carries both halves: allowedDomain enumerates every servable
header, and the message names the sections.
Strategy authoring is untouched by this bump. The same strategy is legal, and reads those
scalars correctly, on an arena agent — which is why the refusal is on the DEPLOYMENT and not on
compile_strategy_plan / apply_strategy_plan.
A benchmark-bound section no longer accepts crowd metrics or rank transforms (49.0.0,
fix-benchmark-legality-save-path). On a custom section carrying a non-null benchmarkTicker, a
column whose metric is enrichment-stage (the CROWD_* family, FLOW_ALIGN, SMART_RETAIL,
CAPTAIN_CONF, CONFIDENCE, SETTLED_AT, the PERP_SPOT_* trio) is refused with
REPORT_COLUMN_BENCHMARK_METRIC_UNSUPPORTED, and one carrying a rank transform in either the
direct or the chained position with REPORT_COLUMN_BENCHMARK_TRANSFORM_UNSUPPORTED. Every tool
accepting a section array is affected, and no other byte of your body changes.
Fix it by moving the column, not by retrying. Both readings are defined relative to the
cohort being evaluated — a crowd reading is what this session's players did, a rank is a position
among the coins under evaluation — and a benchmark is deliberately outside that cohort. Neither
has a value there, which is why such a column could never render. Put it on an ordinary section
(benchmarkTicker: null), or drop it.
This is a fix, not a new rule. The restriction shipped with benchmark sections and was already
enforced by the column builder, by report materialization, and by get_strategy_column_contract —
so an author who checked a column against discovery first has been seeing this refusal all along.
What changed is that the SAVE path now asks the same question. Previously it did not, so such a
section persisted and then failed at every evaluation instead of at authoring.
A custom report section no longer accepts timeframe (48.0.0,
remove-section-anchor-override). The per-section anchor override is gone. Every tool accepting a
section array — compile_strategy_plan, preview_strategy_report, derive_strategy_rule_view,
and apply_strategy_plan — refuses a section carrying it: sections[N]: Unrecognized key(s),
with no other byte of your body changing. Drop the key. A section's columns resolve against the
strategy timeframe, and a column reaches any other timeframe by pinning it on the column
(timeframe: { abs: '4h' }) — which it could always do. Relative column references
(anchor/lower/regime) are untouched, and are the point: they track the strategy.
A custom report section requires notes (43.0.0, add-authored-section-notes). Every tool
accepting a section array — compile_strategy_plan, preview_strategy_report,
derive_strategy_rule_view — refuses a section without it: sections[N].notes: Required, with no
byte of your body changing. Send explicit null for "no note". It is required rather than optional
because these payloads are a FULL REPLACE: an omitted key and an explicit null would be the same
request on the wire, so every rebuild site would silently clear a note its author wrote.
benchmarkTicker carries the same required-nullable discipline for the same reason.
Every condition requires clock and closes (44.0.0, add-condition-clock), and exit
arrives with them. A condition entry now carries eight keys, not five. clock is "LIVE" (the
previous behaviour — the forming bar) or "CLOSE" (settled bars); closes is how many consecutive
closed bars must read true, 1–5, and is always 1 under LIVE. No wire default, for the same
whole-set-replacement reason as notes: a defaulted key would let an unrelated re-save silently
un-clock an enforced money gate back to forming-bar evidence.
A CLOSE clock is accepted only where a closed frame can change the reading — the clause must
resolve from the coin's own candle series at offset 0. Frame-inert operands (perp-payload scalars,
published rolling changes, ranks, zone entities, regime labels, enrichment metrics, session
scalars) are refused with CONDITION_CLOCK_OPERAND_ILLEGAL naming the header and the remedy: split
that clause into its own LIVE condition and conditionRef it. exit: true is legal only under
clock: "CLOSE" — an exit fired on a forming bar is an intrabar exit.
A strategy requires a six-key entry object (44.0.0 for four keys, 47.0.0 for three more —
add-entry-on-close, add-level-trigger-execution — and 52.0.0 removed levelSource,
derive-entry-level). Required on every CREATE, on compile_strategy_plan, apply_strategy_plan
and update_intelligence_agent. A client sending 44.0.0's four-key object is refused with
entry.levelOffsetAtrMultiple: Required without one byte of it changing.
"entry": {
"trigger": "AT_SIGNAL", // | ON_CANDLE_CLOSE | STOP_THROUGH_LEVEL | ON_RETEST
"confirmTf": "4h", // the strategy timeframe or the rung below it — nothing else
"closes": 1, // 1–5; must be 1 unless ON_CANDLE_CLOSE
"bandAtrMultiple": 1.0, // > 0, and <= the platform's entry-deviation gate
"levelOffsetAtrMultiple": 0, // 0–2, UNSIGNED distance from the derived edge; 0 unless a level trigger
"validForBars": 4 // 1–24 of the strategy's own bars; 4 unless a level trigger
}AT_SIGNAL with those values is byte-identical to pre-44 behaviour. The legality matrix runs one
way: all six keys are always present, so a MEANINGFUL value under a trigger that ignores it is
refused rather than accepted-and-dropped — a dial never silently does nothing.
A strategyTimeframe the platform does not ingest is refused (39.0.0,
move-renderer-to-rendered-section) on get_coin_market_context, where it was previously
accepted. The number is the only signal a client gets.
eventType gains ENTRY_EXPIRED_UNCONFIRMED (48.1.0, fix-entry-arming-lifecycle) on
get_radar_activity and get_radar_activity_summary — an armed entry that reached its episode
lifetime without ever receiving a confirming close. A client holding its own closed copy of that
enum rejects the new member; one that switches exhaustively on it needs the branch.
entryVoidCause is deliberately unchanged and still carries exactly BAND, CONDITIONS,
STRUCTURAL. Each of those is a measurement taken at a close, so an episode that reached no
close gets its own event type rather than a fourth cause — ENTRY_VOIDED means "called off at the
close", a claim this outcome must not make.
The authoring boundary also gains a refusal with no schema change: a strategy declaring an
arming trigger whose required conditions read the LIVE clock is rejected on the existing
VALIDATION_ERROR code, naming the offending condition key. Same code, same shape, new reason —
so nothing in the published schema tells you it can now happen.
SWING_LOW / SWING_HIGH leave the stop and take-profit method enums, and the four S/R
indicator keys leave the signal vocabulary (53.0.0, rename-donchian-channel). DONCHIAN_LOWER
/ DONCHIAN_UPPER and donchian_upper / donchian_lower / prev_donchian_upper /
prev_donchian_lower carry the same values; nothing answers to the old names.
get_coin_market_context is REMOVED (40.0.0, retire-get-coin-market-context). Calling it
returns an unknown-tool error. There is deliberately no alias: a silent redirect would hide a
payload shape change from a client that never asked for one. Use get_market_context.
get_macd_heatmap leaves the published surface (41.0.0). Same shape of break, same absence of
an alias.
isPrimary is removed from every published EvaluatedSignal (38.0.0) — get_signal_log,
get_public_agent_signal_log_detail, and every other tool returning a signal scorecard. Reading it
now finds the key absent rather than false.
scan_agent_coins returns the ranking, not an explanation of every coin (54.0.0,
fix-mcp-scan-row-altitude). The scan used to wrap the app's full per-coin qualification verdict
in every row — two directions with candidate-level construction and a stop-loss policy band, four
gates each with its own measurement, condition reach reasons, the ATR corridor. For a 78-coin
catalog that was 82,147 characters, ~1,053 per row, which is past the tool-result cap of every
client we know of: the calling model received a file path instead of an answer, so the tool did not
deliver its result even when the scan succeeded.
rows is gone. Three ranked arrays replace it, and rank is global across all three, so
reading them in this order reproduces the server's own sequence:
qualified[] { rank, coinTicker, scorePercent, coinDataStopped }
rejected[] { rank, coinTicker, scorePercent, firstFailReason, scoreShortfallPercent }
unscorable[] { rank, coinTicker, coinDataStopped }Array membership now carries what the row kind discriminator and the qualifies flag used to,
and both are gone with them; firstFailReason is non-nullable on a rejected row, because a
non-qualifying verdict always names the gate that blocked it. Rows also lose long, short,
gates, tradeableAtrRange, evaluatedAt, coinName, assetClass and category.
Two things are new. The agent's agentId, agentName, strategyTimeframe and minScorePercent
move to an agent object carried once instead of on all 78 rows — null when the scan scored
nothing at all. And scoreShortfallPercent is server-computed: how far below the minimum the score
fell, non-null exactly when the aggregate score is what blocked, so you never subtract a published
threshold from a published reading yourself.
scanStartedAt, coinsScanned and qualifiedCount are unchanged, and so is every coin: same
rows, same order, same ranks, same verdicts, same rate buckets, same evaluator. Only the fields
moved — about 8.8 KB for the same 78-coin scan.
Migration. Read qualified / rejected / unscorable instead of rows, and take the agent's
thresholds from agent rather than from the first row. For a shortlisted coin's full per-direction
and per-gate detail, call get_agent_coin_qualification on up to 12 tickers — it carries every
dropped field and costs no second scan (the scan is rate-limited to 3 per agent and 10 per user a
minute; the probe is not).
Report headers, glosses and signal indicator keys are renamed for the Donchian channel
(53.0.0, rename-donchian-channel). Every report surface — preview_strategy_report,
get_strategy_section_template, the agent prompt previews — renders donchianHi / donchianLo
and their dist_… / …_rank_near forms where it rendered swingHi / swingLo, with the labels
20-bar high / 20-bar low and glosses that say what the number is (the highest high / lowest low
of the last 20 closed bars — the channel's edges). Signal definitions
(get_strategy_signal_definition, list_strategy_signals) and signal-log indicatorValues
carry donchian_upper / donchian_lower / prev_donchian_upper / prev_donchian_lower for the
four S/R signals. Signal ids, the SUPPORT_RESISTANCE module and its display names are unchanged.
Read the new keys. A reader keyed on swing_high / swing_low finds nothing; the values are
the same numbers under the new keys.
The stored entry discipline no longer names a level source (52.0.0, derive-entry-level).
StrategyDTO.entry (get_strategy, list_strategies, the fork_strategy / archive_strategy /
restore_strategy envelopes) and the apply envelope's postState.entry lose levelSource.
TradeOutcomeDTO.entryDiscipline (get_trade_outcome_by_decision, list_trade_outcomes) loses
it and gains levelOffsetAtrMultiple: number | null — the offset in force at the fire, null on
rows written before its column existed. A reader that rendered the level source renders the
geometry from trigger and the outcome's direction instead: which edge and which order shape
are a pure function of those two fields, so no stored copy is served.
update_strategy_signal_rule gains its own response envelope (51.0.0,
fix-signal-rule-patch-semantics). It no longer shares { strategy } with its siblings. The
response is { strategy, ruleChanges }, where ruleChanges is the server's own before/after
pair for the edited signal — [{ signalId, before, after }], each side a full rule object. It is
null when the mutation changed no rule, never [].
Report the change from that pair, not from memory. The planner always computed the diff and the tool discarded it, so a caller narrating what it just did had only its own recollection of the before-value. One production edit shipped a wrong receipt on top of a wrong write that way, and the write was unreconstructable from the audit trail afterwards.
Additive, but published on a .strict() shape — a decoder pinned to the old two-key object
rejects the new key. fork_strategy, archive_strategy and restore_strategy keep the shared
StrategyResponseSchema and publish exactly what they did; it was deliberately NOT widened for
them, so this reshape reaches one tool only.
An entry void now names the gate that refused it (49.5.0,
fix-arming-trigger-clock-authority), on get_radar_activity_summary. In the cause rollup, the
ENTRY_VOID group's gateCode widens from always-null to QualificationGateCode | null: a
conditions-side void carries the gate that blocked — AGGREGATE_BELOW_MIN,
REQUIRED_COUNT_BELOW_MIN, REQUIRED_CONDITION_FALSE — while a band void stays null, because
that void happens on a reading that qualified and has no failing gate to name.
A client that renders the field through the same enum the response already uses on four other
cause arms needs no change. One that treated it as a literal null — a strict decoder pinning the
type, or a branch keyed to its absence — sees a value it did not expect. That is the whole
migration.
Why it moved: the group previously collapsed every conditions-side void under one label. The first 26 in production carried that label while two different gates had produced them, and none of them was a required condition being false. The rollup ships counts rather than rows, so the gate could not be recovered client-side.
The normalized report section loses timeframe (48.0.0, remove-section-anchor-override), on
every tool that publishes a strategy: get_strategy, fork_strategy, archive_strategy,
restore_strategy, update_strategy_signal_rule, apply_strategy_plan, and
compile_strategy_plan's post-state. A strict parser rejects the shorter object; a lenient one
reads a section whose anchor is the strategy timeframe, which it now always is.
get_strategy_column_contract renames its anchor, both ways (48.0.0). The request field
sectionTimeframe becomes anchorTimeframe — same meaning, and still optional. On the response,
timeframe.requiresSectionTimeframe becomes requiresAnchorTimeframe, and
timeframe.sectionTimeframeOverrideAllowed is removed: it published whether a column could go
in an anchor-overridden section, and no section can be overridden. One call also stops being
refused — a timeframe-inert metric supplied with an anchor now compiles.
REPORT_COLUMN_SECTION_TIMEFRAME_UNSUPPORTED leaves the authoringCode vocabulary.
RenderedSection.notes stops carrying provenance (42.0.0, separate-section-facts-from-read),
on preview_strategy_report, compile_strategy_plan and get_market_context. A new REQUIRED
provenance: string[] carries it instead. A client reading provenance out of notes now reads an
author's prose, or nothing — which is a silent misread, not an error.
preview_strategy_report.renderedSections[] gains authoredNote (43.0.0), the author's read
for a custom row and null on a platform row. Additive, but published on a .strict() shape.
get_radar_activity serves its evaluation curve on the FIRST PAGE ONLY (37.0.0), and
get_radar_activity_summary is added. A client reading the curve off a later page finds it absent.
The stop and take-profit method enums gain DONCHIAN_LOWER / DONCHIAN_UPPER (53.0.0,
rename-donchian-channel), replacing SWING_LOW / SWING_HIGH on the signal-pipeline detail
schemas — trade-setup options, R:R-rejected pairs and candidate levels. A copy of either enum that
rejects unknown members must add the two new ones; the two old ones never appear again.
TradeExecutionFailureReason gains LEVEL_NOT_RESTABLE (52.0.0, derive-entry-level): a
level entry refused at placement because its resting price sat on the wrong side of the exchange
mid — a buy stop at or below it, a buy limit at or above it, and the mirror for a sell. It appears
on the list_signal_logs failure-reason filter input and on the signal-pipeline execution
summary's failureReason, with origin CLIENT_GATE. There is no distance limit: a level far from
the mark rests until its bar validity expires.
The authorable metric vocabulary widens by 29 keys (46.1.0, add-indicator-catalog-coverage):
Keltner (KC_UPPER/KC_MID/KC_LOWER), Supertrend (ST_LINE/ST_DIR), Hull (HMA20),
WaveTrend (WT1/WT2), QQE (QQE_RSI_MA/QQE_STOP), Parabolic SAR (PSAR), Ichimoku
(ICHI_CONV/ICHI_BASE/ICHI_SPAN_A/ICHI_SPAN_B/ICHI_LAG), Williams %R (WILLR14),
Stochastic RSI (STOCH_RSI14), session pivots (PIVOT_P/PIVOT_R1–R3/PIVOT_S1–S3), plus
four already-published fields that became addressable: BB_UPPER, BB_LOWER, DI_PLUS,
DI_MINUS.
Additive on the wire — every request you can send today is still accepted. It is called out here because a client holding its own closed copy of the metric enum rejects the new members, and because an agent holding a cached belief that these are inexpressible will substitute for a primitive the platform now serves. That failure raises no error at all: the author is simply told a strategy cannot be built.
EntryTrigger gains STOP_THROUGH_LEVEL and ON_RETEST, and EntryLevelSource is published
for the first time with four members (47.0.0). Additive on their own; the required keys above are
what make that bump a major.
Each signal-checklist item's measured object gains a required triggered boolean (47.2.0,
fix-entry-prompt-signal-evidence) on get_signal_log and get_public_agent_signal_log_detail,
on the numeric and categorical arms. Additive and MINOR — a client ignoring it is unaffected —
but it is named here because a client parsing that output strictly rejects the new key, and because
leaving 47.2 unlisted would make a reader wonder what happened to it.
The unavailable arm deliberately does not carry triggered: "the claim could not be joined to
a stored result" and "the signal did not fire" are different states, and leaving the key off that
arm makes conflating them a type error rather than a convention.
Two tools join the catalog for the agent trade flow (52.1.0, add-mcp-agent-trade-flow).
scan_agent_coins evaluates every active coin against one of your agents in a single call and
returns them server-ranked, on the same use case, buckets and ranking the app's own TRADE-tab scan
serves. propose_entry_decision runs the conversational trade turn headlessly for one (agent,
coin) and returns its terminal in the stream's own vocabulary: type: recommendation carrying the
PROPOSED TradingEntryDecisionDTO row that get_entry_decision and list_pending_approvals
already publish, type: no_trade, or type: error carrying the turn's TradeConvError verbatim.
Nothing narrows and no existing schema hash moves; the 12-ticker get_agent_coin_qualification
stays as the spot-check probe.
idempotencyKey is REQUIRED on propose_entry_decision — it is the turn's own key, and a
same-key retry replays the recorded terminal rather than paying for a second inference. That
includes a post-billing LLM_FAILURE, which is returned as a value for precisely that reason.
Pre-engine faults are typed errors instead: RATE_LIMITED, NOT_FOUND, CONFLICT for a same-key
call still in flight, and SERVICE_UNAVAILABLE when the surface is switched off. Both tools
refuse with RATE_LIMITED carrying retryAfterSeconds under the same per-(user, agent) limit the
app itself enforces, so a scan is never served stale or partial.
get_regime_snapshot publishes the evidence behind the verdict (47.1.0,
publish-regime-classification-evidence). The snapshot gains evidence: the quantities the
classifier read, the gates it tested them against, the signed margin to the gate deciding whether
the current label survives, and the two decision facts only the classifier holds — gateState
(cleared / held / dropped) and directionSource (di / ema). Nothing narrows; a client
that ignores the field is unaffected.
Read gateState before you trust a trend label: held means the ADX hysteresis buffer is
carrying the PREVIOUS bar's label rather than this bar re-confirming it — a materially weaker
claim wearing the same word, and one no client could previously detect. directionSource: 'ema'
is the same shape of warning: the direction came from the fallback that fires precisely when the
DI spread is indecisive. The margin is signed so positive always means "the current label
survives by this much", in every gate state, so it is safe to branch on its sign.
conviction is a BRANCH DISCRIMINATOR, not a confidence: it encodes which rule in the priority
ladder matched, not how comfortably it matched. The margins carry comfort. A client reading
conviction as a strength score is reading it wrong, and always was — this release just makes the
alternative available.
Thirteen metric keys join the catalog (47.1.0) — the regime family gains REGIME_STATE,
REGIME_CONVICTION, REGIME_RUN_BARS, REGIME_TREND_GATE, REGIME_TREND_MARGIN,
REGIME_TREND_SOURCE, REGIME_DI_SPREAD, REGIME_VOL_ATR_RATIO, REGIME_VOL_BBW_RATIO,
REGIME_MOM_BULL_VOTES, REGIME_MOM_BEAR_VOTES, REGIME_CRASH_MARGIN and REGIME_CRASH_LATCH,
making the composite regime and its evidence addressable in a report column or condition for the
first time. Only a client that switches exhaustively on MetricKey needs new branches.
Not a contract change, but worth knowing if you author conditions: the report grammar's regime metrics now resolve from the confirmed close on every path. They previously resolved from the forming bar when a report was rendered and the confirmed close when the scan swept, so the same condition could read differently in preview than in production. Same wire shape; same bar everywhere now.
The number in this heading is a CONTRACT version, not this package's version. The npm badge at the top tracks the proxy's own code; this section tracks the server's wire contract. They move independently by design: a contract move needs no release here, because a connected proxy relays the contract out of the upstream handshake rather than declaring it. So a package on
31.xlisting contract history up to36.xis correct — not a version someone forgot to bump. Read the live pair from the startup stderr lines orGET /mcp/version; see Rediscovery & versioning for why.
These are the server contract breaks between contract 12 and contract 36. Most of the span shipped while the package sat at 11.0.0; contract 31 landed after this package reached 31.0.0, and the two numbers matching is coincidence — since v31 the announced contract is relayed from the server, so a package version says nothing about a contract version. They are contract history, not package releases: from v31 the announced contract is relayed live and a contract move is no longer a release here. Grouped by what a client observes, with the contract version that introduced each.
The proxy itself is unchanged. It embeds no schemas, pins no contract version, and forwards { request } verbatim. Every break below lands on whatever authors the payload or reads the result, never on the proxy.
compile_strategy_plan is no longer read-only or idempotent (33.0.0, rehydrate-approved-plan-on-apply). Its input, its output and its behaviour toward your strategy are unchanged — it still changes no strategy, agent or revision — but it now parks the plan it approved for its own apply to read, and each call mints a distinct record and a distinct token. readOnlyHint and idempotentHint are published as false accordingly. A client that retried a compile that had already succeeded, or fanned two out in parallel for one edit, was doing so on the strength of the old annotation: do neither. Compile once per reviewed payload. Nothing in the payload tells you this moved.
Position-size presets are now a RISK BUDGET (30.0.0, split-stop-geometry-from-risk). smallPct / mediumPct / largePct stop denoting a share of the ORDER (notional = pct / 100 × headroom × leverage) and start denoting the share of headroom placed at risk (notional = headroom × riskPct / stopDistancePct, capped at the margin headroom can post). Same keys, same types, same accepted range: nothing in the payload tells you the meaning moved. A client still sending 22.0 for MEDIUM is asking to risk 22% of its budget on one trade rather than roughly 2%. Typical risk budgets are 0.5–3; the platform defaults moved to 1 / 2 / 3. Leverage stops multiplying order size and becomes a constraint only.
upsert_deployment_policy requires enabled (35.0.0, add-arena-deployment-pause). The arena deployment gains an owner-owned pause, and the flag that carries it is required, not optional — a body accepted under contract 34 is refused under 35 without one byte of it changing. Required is the whole point: this call replaces the entire policy, so an omitted key and an explicit true would be the same request on the wire, and every client that rebuilt a policy without the flag would silently resume a deployment its owner had paused. Read the value from get_deployment_policy and send it back. There is no separate pause verb — pausing and resuming are this same call with the flag flipped.
Do not reach for enabled: false to un-deploy. It keeps every slot and stops play, which is the opposite of withdrawing: delete_deployment_policy is the withdrawal verb, and it discards the rules permanently. upsert_deployment_policy refuses an empty slot set, and its rejection names both routes.
A signal rule flagged required at allocation Off is rejected (34.0.0, enforce-required-allocation-invariant). required and allocation are two independently editable fields encoding one thing — whether and how a signal participates — so { required: true, allocation: 0 } was representable and meant nothing: the scorecard's triggered set already excludes Off, so such a rule could neither satisfy minRequiredCount nor block a trade. It is now refused on every rule-writing surface — compile_strategy_plan, apply_strategy_plan and update_strategy_signal_rule alike.
Three things make this one easy to trip over. It is an input-acceptance narrowing: a payload accepted under contract 33 is refused under 34 without one byte of it changing. It is invisible in the published schema — zod-to-json-schema drops effects by construction — so you cannot pre-validate it from tools/list, and the typed error IS the contract: read details.inertRequiredSignalIds, which carries the complete sorted list of offending signals, rather than the message, which names a bounded prefix. And nothing is repaired for you: the server will not raise the allocation (that would invent a scoring weight you never chose) nor clear the flag (that would discard your intent silently). Pick one and resend; either satisfies the boundary.
apply_strategy_plan no longer accepts the plan (33.0.0, rehydrate-approved-plan-on-apply). Its input narrows to { planToken, confirm }. A plan member is rejected as an unknown key — not accepted, not ignored, and with no transitional dual shape — so every client that built the payload breaks on the next connection, which before this is what every published surface told it to do. The server keeps the plan its own compile approved and reads it back, so you copy nothing out of the compile response: forward planToken byte-for-byte and confirm.
Three consequences worth knowing. The 256,000-byte cap on the apply payload is gone with the payload, so a large authored surface no longer becomes impossible to apply through a conversational client; the compiled plan is still capped and compile still enforces it. PLAN_APPROVAL_NOT_FOUND joins the error vocabulary for a token no approved plan answers to — already applied, lapsed, or never issued, all one code, because the recovery is the same in each case: compile again. And a validation refusal (a quota, a name collision, a bound agent that changed, a moved catalog) now leaves the approved plan applicable — clear the cause and confirm again with the same token while it lives, rather than recompiling.
The agent brain is no longer a preset union (32.0.0, remove-agent-presets). create_intelligence_agent stops taking the brain discriminated union: modelId and behavior become required top-level fields and the { kind: 'PRESET' | 'CUSTOM' } wrapper is gone, so the old shape fails on the unknown key and on two now-missing required fields. update_intelligence_agent drops brainPreset; modelId and behavior stay independently optional and are now always honoured, closing an accept-and-ignore where a named preset silently discarded a model sent beside it.
apply_strategy_plan now publishes the same bounds as compile_strategy_plan (31.0.0, resolving #4495). Eleven position-management dials were declared three times server-side and two copies had drifted, so apply advertised trailingGivebackPct as a bare number where compile advertised 25–55, and dropped trailingTriggerR's multipleOf 0.01 — the constraint pinning storage precision so a sub-precision value is rejected rather than rounded onto the trail-from-entry sentinel 0. Nothing bad could ever commit (the digest would not match), but the refusal you got was a binding mismatch, which arrived as a bare INTERNAL_ERROR. Apply's published bounds only narrow to compile's; a client that copies values from approvedPlan.postState, as it should, is unaffected. Separately the trio minStopLossAtrMultiple / maxStopLossAtrMultiple / minRiskRewardRatio is published as a bare declaration by both tools, its real bounds being runtime-tunable and inexpressible in JSON Schema.
The stop-loss ceiling changed unit (30.0.0). maxStopLossPct (a percent of entry) becomes maxStopLossAtrMultiple (a multiple of ATR), and the accepted range narrows from (0, 100] to (0, 3]. It is a rename and a re-denomination — mapping the old value onto the new key sends a number one to two orders of magnitude too large. The objects are .strict(), so a 29.x client sending maxStopLossPct is rejected with an unknown-key error. A new cross-field rule comes with it: minStopLossAtrMultiple < maxStopLossAtrMultiple is now a real comparison and is enforced.
The grid-confidence and trade-conviction bars left the agent (28.0.0, remove-agent-rule-defaults). tradingConfig.gridMinConfidence and minTradeConviction are removed from the shared .strict() config; a bar is declared on the arena slot or radar slot that fires.
The agent no longer carries either entry guard (26.0.0). create_agent and update_agent stop
accepting tradingConfig.signalTimeoutMinutes and tradingConfig.maxEntryDeviationAtrMultiple.
The schema is .strict(), so a client still sending either is rejected, not silently ignored.
Neither has a replacement key on any surface: unlike the strategy-owned fields below, these have no
owning surface at all beneath the platform. One platform_config value governs the entry-price
drift budget for every decision, read at evaluation time; one governs how long an entry may stay
unfilled, snapshotted onto the position at creation. Remove both keys and send nothing in their
place.
The arena stopped granting trade authority (25.0.0, remove-arena-trade-permissions). upsert_deployment_policy and preview_deployment_resolution stop accepting tradingEnabled, minConviction and coinRules[] on a slot.
The agent no longer carries an exit policy (24.0.0). create_agent and update_agent stop
accepting tradingConfig.positionManagement. The schema is .strict(), so a client still sending
it is rejected, not silently ignored. The twelve dials that decide how a stop MOVES after entry
— break-even arming, trailing engagement and giveback, time-decay tightening — are denominated in
the setup's own payoff shape (multiples of the trade's initial risk, fractions of take-profit
distance, minutes since entry) and read no balance, leverage or exposure, so they belong to the
thesis rather than to the account running it. Author them on the strategy instead, through
compile_strategy_plan / apply_strategy_plan: the post-state gains the same twelve keys beside
the trade-level trio, and the plan diff gains a positionManagement axis. An agent inherits the
policy from the strategy it binds. The umbrella enabled flag is deleted rather than moved —
each mechanism's toggle is now the whole truth for that mechanism, so "trailing on, management
off" is no longer expressible, a state the server's own monitor and boot recovery already
disagreed about.
The agent-level trading mode is gone (23.0.0). create_agent and update_agent stop
accepting tradingConfig.tradingMode. The schema is .strict(), so a client still sending it is
rejected, not silently ignored. Trading on/off is now scoped per deployment — a radar policy's
enabled, an arena slot's tradingEnabled, a per-coin tradeEnabled — and approval-before-
execution is the conversational surface's own contract, so the account-level switch that sat above
both is removed rather than renamed. To stop an agent trading, turn its deployment off (or halt
the agent); to make one trade autonomously, arm a radar coin or switch trading on for an arena
slot. A newly authored arena slot now starts with trading off.
Trailing gained a required threshold (22.0.0, add-trailing-trigger-r). positionManagement gains trailingTriggerR as REQUIRED (0–2.0, 0.01 precision, 0 = trail from entry); the object is .strict() all-required, so sending positionManagement without it is rejected.
The strategy regime timeframe became derived (19.0.0, remove-strategy-regime-override). It stops being an authored axis anywhere on the contract and is served read-only.
Conditions gained a required required (16.0.0, add-condition-enforcement-gate). A condition entry omitting the boolean is REJECTED rather than defaulted.
The trade-level policy moved to the strategy (15.0.0, move-trade-level-policy-to-strategy). It leaves the agent authoring surface and joins the setup gates on the strategy.
The agent's ATR timeframe axis is gone (14.0.0, remove-agent-atr-timeframe-axis). ATR is sampled on the strategy timeframe, always.
Radar's wall-clock condition changed shape (12.0.0, unify-deployment-hours-as-sets).
The signal scorecard stops serializing its entries three times over (36.0.0, mcp-signal-log-contents). get_signal_log and get_public_agent_signal_log_detail drop scorecard.triggeredSignals, scorecard.primarySignals and scorecard.supportingSignals. Every one held the same entry objects allEvaluatedSignals already carried, so each is one filter over flags every entry still publishes:
Removed | Read instead |
|
|
|
|
|
|
Keep the triggered half of those last two predicates. Both collections were triggered-only by construction, so filtering on isPrimary alone surfaces signals that never fired — a silent widening, not an error. No field is removed from an entry: the key set on an allEvaluatedSignals member is unchanged, every evaluated signal is still returned whether or not it triggered, and details prose and indicatorValues are intact. There is no opt-in to get the three back and no default filter. Breaking only if you read one of the three names; on an 84-signal / 16-triggered log the duplication was 10,682 bytes, 28% of the scorecard, carrying no information.
The fleet roll-up on list_deployment_policies drops unconfigured and gains paused (35.0.0, add-arena-deployment-pause, fix-arena-deployment-undeploy). unconfigured counted a deployment holding zero slots — a state that can no longer exist, because a stored policy now carries at least one slot and the withdrawn state is the absence of a policy rather than an empty one. The bucket was constant 0 at the moment of removal, so no number you read was wrong; a client reading the key still breaks on it, which is why this is a break and not a cleanup. paused is the owner's own switch, counted separately from retired — an administrator disabling the arena — because conflating them tells an owner to wait for something that will not happen. Every policy lands in exactly one bucket, so the buckets sum to arenas: worth asserting if you reconcile these counts.
AdminApprovedModelDTO.isActive became lifecycle (32.0.0, remove-agent-presets) — AVAILABLE / DEPRECATED / RETIRED. The boolean conflated "offered in the picker" with "bound agents may run", so there was no way to stop offering a model without hard-blocking every agent already on it. A client switching on isActive must switch on lifecycle, and must not treat DEPRECATED as blocked: that is the state which keeps bound agents running. The agent read DTO drops brainPreset in the same move — the marker recorded which named bundle an owner clicked, never a value the runtime read, and the model and soul it stamped are unchanged on every agent.
approvedPlan is one object, not an operation union (31.0.0, resolving #4495). It was published as a discriminated union whose discriminator does not survive JSON-Schema conversion, so what actually shipped was a bare anyOf: validating a failing compile response gave you every arm's errors with empty instance paths, and the top one typically complained that an UPDATE was missing creationSeed — a CREATE-only key — while the field that really failed went unnamed. It is now one object with a literal operation discriminator, and creationSeed, expectedRevision and bindingImpact are required and nullable on every operation: a CREATE plan carries a seed and expectedRevision: null, an UPDATE/RESTORE plan the reverse. If you narrowed on the union arms, read operation instead and expect explicit nulls rather than absent keys. If you read those fields without narrowing, nothing changes except that they may now be null.
get_radar_activity gained an EDGE_REARM variant (29.0.0, add-radar-anchor-rearm), and every member gained five rearm* margin keys plus a rearmReasons discriminator. Breaking on both counts if you parse the union strictly.
get_trading_config_catalog drops four trade-default seeds (27.0.0) — defaultMinAtrPct,
defaultMinStopLossPct, defaultMaxStopLossPct and defaultMinRiskRewardRatio leave defaults.
Each seeded a per-agent field that is now strategy-owned: the shared .strict()
TradingConfigSchema already rejected all four as unknown keys, so the catalog was advertising
defaults no request could apply. A client that wants the stop-loss band, the ATR floor or the
risk-reward minimum reads them from the bound strategy, which owns and materializes them.
Nothing is added in their place. defaultMaxEntryDeviationAtrMultiple and defaultTtlMinutes are
untouched — those are not seeds but the platform values that govern.
get_radar_activity gained blockReasonCode on every member (21.0.0, fix-block-reason-attribution) — non-null only on BLOCKED_BEFORE_EVALUATION rows written after 2026-08-17.
signal_pipeline's decision became a discriminated union (20.0.0, add-decision-skip-attribution). ENTER/GATED carry the seven level fields as REQUIRED; SKIP omits them entirely rather than sending nulls, so reading entryPrice without narrowing the verdict finds the key absent.
get_radar_activity gained an EVALUATION_OUTCOME member (18.0.0, add-radar-fire-outcome-journal), and every existing member gained evaluationOutcome + screenReason.
Scalar families became placeable modules (13.0.0, add-scalar-family-modules); six opt-in scalar headers moved off the shared session-field section key.
Both entry guards leave every agent-returning shape (26.0.0) — the read side of the input
removal above. AgentTradingConfigDTO drops signalTimeoutMinutes and
maxEntryDeviationAtrMultiple on every tool that serves an agent, and the explorer trading spec
and the agent-review payload drop them too. get_trading_config_catalog drops
defaultSignalTimeoutMinutes from its defaults and the
minimumMaxEntryDeviationAtrMultiple / maximumMaxEntryDeviationAtrMultiple pair from its bounds —
a bound pair that constrained a per-agent field which no longer exists.
defaultMaxEntryDeviationAtrMultiple and defaultTtlMinutes stay, and are now the values that
actually govern rather than seeds a new agent copies. A client reading these objects strictly must
drop the removed keys.
positionManagement leaves every agent-returning shape (24.0.0) — the read side of the input
removal above. AgentTradingConfigDTO drops the nested block on every tool that serves an agent,
and the explorer trading spec drops it too. A client reading these objects strictly must drop the
key; one that wants the policy reads it from the bound strategy.
get_trading_config_catalog drops positionManagementPresets (24.0.0) — the pistol ladder
(COLT / WEBLEY / BERETTA / LUGER / WALTHER) is retired, not renamed. Once the values live on
the strategy, the strategy IS the named bundle, with its own name, description and revision
history; a parallel vocabulary of anonymous bundles beside it would be a second name for the same
thing. There is no replacement enum to migrate to — list strategies instead. The catalog's
defaultPositionMgmt* trading defaults go with it, for the same reason: nothing seeds an agent's
exit policy any more.
tradingMode leaves every agent-returning shape (23.0.0) — the read side of the input
removal above. AgentTradingConfigDTO drops it on every tool that serves an agent, and so do the
agents-hub permission envelope, the explorer entry, and both public-profile shapes. A client
reading these objects strictly must drop the key.
DeploymentResolvedResolutionDTO drops agentTradingMode (23.0.0) — the field 10.0.0 added,
now unnecessary: with no account layer to overlay, the resolved tradingEnabled is the whole
answer about whether the previewed deployment trades.
DeploymentResolutionStatus gains PAUSED (35.0.0, add-arena-deployment-pause), returned by get_deployment_policy, list_deployment_policies and preview_deployment_resolution. A client switching exhaustively on the status must add the branch. Two properties are not obvious from the name: it is answered before any slot is resolved, so a paused deployment discloses no agent identity and carries regimeUsed: null; and its targetSession is nullable, because a deployment can be paused on an arena with no upcoming session and the pause is still the true answer. Do not re-derive the pause from the enabled flag beside it — the served status is the answer on every path, and those two disagreeing is the defect this closed.
Seven plan-token failures became their own error codes (31.0.0, resolving #4495). TOKEN_EXPIRED, TOKEN_BINDING_MISMATCH, INVALID_TOKEN_SIGNATURE, INVALID_TOKEN_FORMAT, INVALID_TOKEN_CLAIMS, INVALID_DIGEST_MATERIAL and INVALID_MATERIALIZATION_FENCE all used to arrive as a bare INTERNAL_ERROR — the server wrote the true reason to its own audit log and discarded it at the boundary, so a refused apply told you nothing. They now arrive as themselves, over MCP and HTTP alike, with 409-class status for the two state-conflict codes and 400-class for the five malformed-material codes. A client switching exhaustively on error codes must add the branches; one rendering unknown codes generically is unaffected. Two are worth handling by name: TOKEN_EXPIRED means recompile (the token lives five minutes), and INVALID_TOKEN_SIGNATURE usually means the token was not forwarded verbatim — it is opaque, so copy it byte-for-byte and never retype or reconstruct it.
TradeEvaluationAttemptReasonCode gains OPEN_POSITION_CHECK_UNAVAILABLE (19.4.0), splitting a code that previously reported a platform fault as a fact about your account.
QualificationGateCode gains REQUIRED_CONDITION_FALSE (19.3.0), from the SCAN-stage gate that now evaluates required conditions before a fire edge is spent.
TradingPipelineGateStage gains EVALUATION and TradeEvaluationAttemptReasonCode gains EVALUATION_FAULTED (18.2.0).
Nothing in the proxy changes. No configuration, no environment variable, no call-shape change on this package's own surface. If your client discovers tools live and reads results generically, npm i @battlegrid/mcp-server@30 is the whole upgrade.
27.1.0 exit-policy authoring input on compile_strategy_plan · 19.2.0 get_account_state account identity · 19.1.0 Standing Orders marker authoring · 18.4.0 list_gate_blocks summary groups · 18.3.0 radar maintenance pause · 18.1.0 protection geometry · 17.2.0 break-even/trailing status · 17.1.0 get_signal_log condition evaluation · 13.1.0 four owner-scoped read tools · 12.1.0 cross-venue spot price metrics · 11.1.0 discoverable rate limit.
Historical. These notes are keyed to the package versions that once paired with contract versions. That pairing ended at v31 (above); the breaks themselves are still real, and still describe the server's contract as it moved from 6.0.0 to 11.0.0.
v11 paired with the BattleGrid server's MCP contract v11.x. Under the pairing rule then in force, the package version tracked the server's wire contract because the proxy announced battlegrid@<package version> in its own stdio handshake.
This release absorbs six breaking contract majors at once. The published package went from 5.0.0 straight to 11.0.0, so no 6.x through 10.x client exists to upgrade from — the breaks are therefore grouped by what you will observe, with the contract version that introduced each, so a client hitting a specific rejection can find it here.
The proxy itself is unchanged. It embeds no schemas, pins no contract version, and forwards { request } verbatim. Every break below lands on whatever authors the payload or reads the result, never on the proxy.
Challenge participation is no longer a field you set (10.0.0). create_agent and update_agent stop accepting arenaChallengeEnabled, and a deployment policy's slot rules and per-coin rules stop accepting challengeEnabled. All four schemas are .strict(), so a client still sending any of them is rejected, not silently ignored. Challenge participation is now identical to effective trade permission, resolved per coin: to stop an agent taking challenges at a venue, turn that venue's trading off — at the slot, or per coin for finer grain — using fields you already have.
VOLUME_RATIO is retired and replaced by RVOL (6.0.0). MetricKeySchema auto-derives from the server's metric catalog, so the published enum simply stops accepting the old key: a column authored with metric: 'VOLUME_RATIO' is rejected against the enum. No alias exists — deliberately. The catalog audit adjudicated the old name as a name-level lie (the value is current volume ÷ its 20-period average, a multiple), and the correction was made at the root rather than grandfathered. If you hit an enum rejection naming VOLUME_RATIO, this note is the match.
BB_WIDTH can no longer be ranked (8.0.0). { metric: 'BB_WIDTH', transformId: 'rank' } is now rejected. BBwidth is a price-unit spread (upper − lower) that had falsely declared percent, and that declaration was the only thing admitting it to exchange-wide ranking — the ordinal it produced sorted by token denomination rather than by compression. It re-declares signedPrice and leaves the ranked contract. Rank BB_WIDTH_PCT (bbWidthPct) instead, which ships in the same release: the capability moved, it was not removed.
This is the failure mode with no error attached to it. Nothing is rejected; your matcher simply stops finding the column.
vol → RVOL (6.0.0), and with it vol_trend → RVOL_trend and vol_rank_hi → RVOL_rank_hi. The header code moves with the metric key, because every generated header over that metric derives from the code alone.
BBwidth_rank_lo no longer exists (8.0.0) — it was the header of the ranked pairing retired above.
estimatedTokenCount is gone (7.0.0). preview_strategy_report and compile_strategy_plan no longer carry it; the same number now sits one level deeper as budgetUsage.estimatedTokens.used, paired with the cap that governs it — so a client reading the count changes one path and gains the ceiling it was never told. tokenCountModel is unchanged. Discovery grows to match: list_strategy_vocabulary and the report catalog add budgets.estimatedTokens and a previewExecutionLimits object carrying the serialized-result byte cap and the preview deadline. Those two are published cap-only and deliberately have no used companion.
approvedPlan.mismatches changed on both axes (9.0.0). Both report-coverage codes are renamed off "module", because the module is no longer the unit of coverage:
Was | Is now |
|
|
|
|
Each mismatch also carries a required data: CoverageDatum[] — the (metric, rung) pairs the mismatch is about: every MISSING datum for the not-in-report code, the PRESENT data for the signal-off code, empty for REQUIRED_SIGNAL_UNAVAILABLE. A client switching exhaustively on the old code strings stops matching. Behaviourally, coverage is now decided by whether the report renders a signal's declared metrics at the rung that signal reads, not by whether its module appears anywhere — so expect warnings you never saw before, and the disappearance of warnings no composition could clear. Mismatches remain advisory and non-blocking; nothing about apply gates on them.
IntelligenceAgentDTO drops arenaChallengeEnabled (10.0.0) — the read side of the input removal above. ResolvedSlotRulesDTO.challengeEnabled stays and keeps its shape; it is now derived server-side, carrying the same value and provenance as tradingEnabled, so a client reading the resolved bundle needs no change.
range is no longer a tuple (11.0.0). The closed positional pair [min, max] becomes the half-open object { min: number; max?: number }. It travels through ScalarSchema, so this lands on list_strategy_vocabulary, query_report_catalog, and get_metric_construction_hints alike.
This one fails silently. A client reading range[0] / range[1] gets undefined with no error raised — read range.min / range.max instead, and treat a missing max as unbounded above. The tuple could not state the truth about the volume/trade-count family, which is non-negative and unbounded above, and Infinity serializes to null on the wire. Those six metrics now declare { min: 0 }, and as a consequence their far/near rank orderings are no longer offered — on a non-negative value that pair is a synonym pair under the magnitude gate's documented semantics.
The published unit enum gains ratio and fraction (8.0.0), and RVOL, BUY_PRESSURE, BB_PCT_B now emit them instead of percent — none of the three is a percentage, and percent appended a false %. This is a break in the opposite direction from the others: nothing you send is rejected, but a client holding its own closed copy of the unit enum rejects the new members. Easy to mistake for a purely additive change. Values are not rescaled — buyPres and pctB stay 0–1, so persisted thresholds comparing against 0.5 keep their meaning.
No stored strategy needs client action. Every persisted reference of both kinds — metric keys in section-column rows and revision snapshots, and everything addressing a column by its generated header — was migrated server-side by 20260805120000_rename_volume_ratio_to_rvol.sql. This is a client-literal problem only; there is no stored record to repair.
Nothing here is a break, but a client that enumerates these vocabularies will want them:
Four metric keys join the catalog — SPOT_CVD, PERP_SPOT_FLOW, PERP_SPOT_STRENGTH, PERP_SPOT_CONFIRMS — and includePerpSpotFlow joins the context-source key set as its 23rd member, opt-in (5.2.0).
Two prompt-section kinds join the union: perp-spot-flow (5.2.0) and session-field (5.1.0). Only a client that switches exhaustively on kind needs a default branch; one that renders content generically needs nothing.
The transform vocabulary grows 15 → 17 (11.0.0) — efficiency and maxShare join, and both join the chain-outer enum that chainSuccessors is served as. Only a client that switches exhaustively on transformId needs new branches; one that renders the served labels generically needs nothing.
The deployment-policy resolution DTO gains agentTradingMode (10.0.0) — the preview previously reported trade rules for an agent whose account-level trading was off, and the account gate is not part of rule resolution. Read it to tell "this rule permits the trade" from "this account can trade at all".
PlatformSectionDTO gains columns (6.1.0) — a platform section's composition in the same wire shape a custom section's columns already travel in, empty for a registry-declared special. Not a copy source, which is where it differs from the identically-shaped CustomSectionTemplateDTO.columns: six platform columns pair a metric with classifyState, a deliberate composability exclusion that authoring rejects at construction. A client that round-trips these into a custom section will be refused, and that refusal is correct.
Retained for authors upgrading from 4.x. Everything below still describes the current contract.
conditionVerdicts no longer exists. The verdict now rides the condition that decides it, and precedence is the conditions' own declaration order rather than a separate ordered map:
// v4 — two parallel arrays joined by string key
{
"conditions": [ { "conditionKey": "UP_FADE", "name": "…", "definition": { /* … */ } } ],
"conditionVerdicts": [ { "when": "UP_FADE", "then": "UP" } ]
}
// v5 — one array; the verdict rides its condition
{
"conditions": [ { "conditionKey": "UP_FADE", "name": "…", "definition": { /* … */ }, "verdict": "UP" } ]
}verdict is required and nullable, never optional: a building block that decides nothing spells its absence as an explicit null, never by omitting the key. An omitted verdict is a rejected payload, not a defaulted one.
A submitted conditionVerdicts is REJECTED, not ignored. Deliberately — a v4 client that forwards the retired field is told what replaced it instead of getting an anonymous unrecognized-key rejection:
conditionVerdictswas retired in contract 5.0.0 — a condition now carries its ownverdict(UP|DOWN|NEITHER, ornullfor a building block). Move each mapping onto the condition it named and resubmit.
The authorable verdict domain narrowed to three. conditionGrammar.verdicts advertised ['UP','DOWN','NEITHER','UNRESOLVED'] and now advertises ['UP','DOWN','NEITHER']. UNRESOLVED is an evaluation outcome — "a deciding condition could not be evaluated" — never an authored intent, so advertising it offered a value the schema then rejected. Any client mirroring the advertised enum into its own validation must narrow with it.
The evaluated per-coin verdict is nullable. Resolution is first-TRUE-decides over the verdict-carrying conditions in declaration order, and its four outcomes are all distinct claims — collapsing any pair loses information a reader needs:
Evaluated verdict | Means |
| The first verdict-carrying condition that resolved TRUE declared it |
| Every verdict-carrying condition resolved FALSE |
| No carrier fired and at least one could not be evaluated — "could not be read", not "read as no setup" |
| The strategy declares no verdict-carrying condition at all — it expresses no direction |
null is new in v5; it previously surfaced as NEITHER. A client that renders the verdict must handle it without collapsing it into NEITHER.
The proxy itself is unchanged. It embeds no schemas, pins no contract version, and forwards { request } verbatim, so the whole break lands on whatever builds the apply projection. Apply takes an allowlisted projection of approvedPlan, never the object itself (diff, viability, mismatches, signalRules, creationSeed, proposedRevision, bindingImpact, authoringCatalogDigest are rejected as unknown keys, and so are the postState fields apply does not accept: id, scope, systemKey, visibility, cadence, isActive, forkedFromStrategyId). The exact field list is step 5 of Strategy authoring (compile → review → apply) below. A client that forwards post-state fields generically — stripping the derived keys rather than enumerating the kept ones — picks the fusion up without an edit; one that enumerates the fields it copies must drop conditionVerdicts from that list.
The v3 authoring contract below is unchanged and still current:
Strict authoring envelopes. get_strategy_section_template, update_strategy_signal_rule, compile_strategy_plan, and apply_strategy_plan publish one strict server-owned object, { request: canonicalPayload }. In multi-account mode the proxy adds account only as a sibling of request, producing exactly { account, request }; on a call it strips only account and forwards the unchanged { request }. It never descends into, flattens, or reconstructs the nested request.
create_strategy is retired. Direct strategy creation no longer exists. Author strategies with the compile → review → apply workflow below, and bind them to agents at agent-creation time (create_intelligence_agent({ …, strategyId })). There is no alias, shim, or flat-payload fallback.
Rediscover after deployment. Publishing the package does not refresh a running proxy's cached capability snapshot. After the server cutover, restart/reconnect the proxy process and re-run tools/list, prompts/list, and resources/list.
Earlier majors: v1.x single/multi-account stdio proxy; v2.0.0 moved the default
BATTLEGRID_API_URLto the/mcpsuffix; v3.0.0 the strategy-authoring major; v4.0.0 madeconditionsandconditionVerdictsrequired on the apply post-state; v5.0.0 fused the conditions/verdicts split (section above). v6.0.0 through v11.0.0 were never published as separate package versions — they are absorbed by the v11 cutover at the top. See Rediscovery & versioning.
https://mcp.battlegrid.trade/mcpGive that URL to your MCP client over its streamable-http (remote) transport and authorize: the client registers itself by Dynamic Client Registration, BattleGrid's consent page opens in your browser, and you sign in and click Authorize. No npm install, no API key. The grant is listed — and revocable — under Profile → MCP → OAuth Sessions.
Reach for a key when your client has no remote transport at all, when your agent runs headless or in CI and cannot open a browser to consent, or when one process drives several BattleGrid accounts. It is fully supported for each of those, and nothing about it is deprecated.
Single account (stdio transport):
BATTLEGRID_API_KEY=bg_live_xxx npx @battlegrid/mcp-serverMultiple accounts (stdio transport):
BATTLEGRID_API_KEYS=bg_live_alice_key,bg_live_bob_key npx @battlegrid/mcp-serverWhen multiple keys are provided, the server discovers each account's identity and injects a required account parameter into every tool so the AI agent can choose which account to act as. OAuth has no equivalent — one grant authorizes one account.
OAuth (no key): Settings → Connectors → Add custom connector. Paste
https://mcp.battlegrid.trade/mcp, save, and authorize on the consent page Claude opens.
API key (fallback) — single account:
{
"mcpServers": {
"battlegrid": {
"command": "npx",
"args": ["@battlegrid/mcp-server"],
"env": {
"BATTLEGRID_API_KEY": "bg_live_xxx"
}
}
}
}API key (fallback) — multiple accounts:
{
"mcpServers": {
"battlegrid": {
"command": "npx",
"args": ["@battlegrid/mcp-server"],
"env": {
"BATTLEGRID_API_KEYS": "bg_live_alice_key,bg_live_bob_key"
}
}
}
}OAuth (no key):
claude mcp add --transport http battlegrid https://mcp.battlegrid.trade/mcpThen start claude, run /mcp, select battlegrid and choose Authenticate — the consent page
opens in your browser and the entry reads connected once you authorize.
API key (fallback):
# Single account
claude mcp add battlegrid -e BATTLEGRID_API_KEY=bg_live_xxx -- npx @battlegrid/mcp-server
# Multiple accounts
claude mcp add battlegrid -e BATTLEGRID_API_KEYS=bg_live_alice_key,bg_live_bob_key -- npx @battlegrid/mcp-serverOAuth (no key): Settings → MCP → Add new global MCP server opens ~/.cursor/mcp.json.
{
"mcpServers": {
"battlegrid": {
"url": "https://mcp.battlegrid.trade/mcp"
}
}
}Back in Settings → MCP, click Needs login on battlegrid and authorize on BattleGrid's consent
page; the entry turns green once its tools load.
API key (fallback): the same file, with the stdio proxy in place of the remote entry.
{
"mcpServers": {
"battlegrid": {
"command": "npx",
"args": ["@battlegrid/mcp-server"],
"env": {
"BATTLEGRID_API_KEY": "bg_live_xxx"
}
}
}
}Use BATTLEGRID_API_KEYS (comma-separated) for multiple accounts.
ChatGPT Desktop connects via OAuth 2.1 — no npm package or API key needed. ChatGPT handles the OAuth flow automatically.
Open ChatGPT Desktop → Settings → MCP Servers → Add Server
Enter the MCP endpoint URL: https://mcp.battlegrid.trade/mcp
Select OAuth as the authentication method
ChatGPT discovers OAuth endpoints, registers as a client (Dynamic Client Registration), and opens BattleGrid's consent page
Log in to BattleGrid and click Authorize
Authentication is a property of the path, not of the client — every client above reaches BattleGrid either way, so pick the row that matches your runtime rather than your client:
Remote + OAuth | API key | |
Transport | streamable-http, direct to | stdio proxy ( |
Auth | OAuth 2.1 with Dynamic Client Registration | API key ( |
Setup | paste the URL, authorize in the browser | npm package + env vars |
Needs a browser | yes, once, to consent | no — works headless and in CI |
Multi-account | one grant per account |
|
Set BATTLEGRID_API_KEY with one API key. All tool calls use that account, and the tools are exactly the server-native shapes — the authoring tools take the strict { request } envelope with no account field.
Set BATTLEGRID_API_KEYS with a comma-separated list of API keys (one per BattleGrid account). On startup the proxy:
Calls GET /mcp/identity for each key to discover the account username
Injects a required account enum parameter into every tool — as a sibling of the existing input, never nested inside it
Routes each tool call to the correct account using the matching Bearer token, stripping only account before forwarding
For the strict authoring tools, the multi-account input is exactly { account, request }:
{
"name": "compile_strategy_plan",
"inputSchema": {
"type": "object",
"properties": {
"account": {
"type": "string",
"enum": ["alice", "bob"],
"description": "Which BattleGrid account to use for this action"
},
"request": {
"oneOf": [
{ "properties": { "operation": { "const": "CREATE" } } },
{ "properties": { "operation": { "const": "UPDATE" } } },
{ "properties": { "operation": { "const": "RESTORE" } } }
]
}
},
"required": ["account", "request"],
"additionalProperties": false
}
}The proxy consumes only the outer account and forwards the unchanged { request } upstream. Never put account inside request, and never flatten request fields beside it.
If a key fails identity discovery (revoked, invalid), it is skipped with a warning. If all keys fail, the process exits. BATTLEGRID_API_KEYS takes precedence over BATTLEGRID_API_KEY when both are set.
Go to battlegrid.trade → Profile → MCP tab
Generate an API key (format: bg_live_*)
Copy the key immediately — it is shown only once
Each account supports one active key at a time. Generating a new key automatically revokes the previous one — restart any running proxy process afterward, since keys are read once at startup.
For paid games and autonomous wagering, enable Server-Signed Wagers in the MCP tab (mcp:wager scope). Strategy discovery and non-financial configuration writes only need mcp:read.
Strategies are authored through one strict, whole-plan workflow. Compilation changes no strategy, agent or revision; apply_strategy_plan is the only write to the strategy itself — but compile is not read-only either: it parks the plan its own apply reads, and each call mints a distinct record and token, so compile once per reviewed payload and never retry or parallelise it. Always review the exact returned plan before confirming.
Choose the operation and revision. list_strategies (add includeInactive:true when preparing a RESTORE) and get_strategy return the current revision; thread it into the next revisioned call.
Discover the report vocabulary live. Walk list_strategy_categories → list_strategy_vocabulary → get_metric_construction_hints → get_strategy_column_contract, and use get_strategy_section_template / preview_strategy_report. Do not guess metric, transform, parameter, template, or enabled-timeframe facts — they are server-discovered.
Compile one complete plan. Call compile_strategy_plan({ request }) where the nested request is exactly one strict branch plus a bounded coinSelection, intentSummary, and assumptions:
CREATE supplies the full new strategy.
UPDATE supplies at least one changed axis and expectedRevision.
RESTORE targets an owned inactive revision (with any repair axes).
Review before confirming. Inspect the returned approvedPlan (complete post-state, proposed revision, diff, bound-agent impact, expiry) and reviewContext (column contracts, point-in-time report preview, open positions, quota/name admission). The plan token expires after five minutes; recompile after expiry or drift.
Apply the plan the server already holds. After explicit user approval, call apply_strategy_plan({ request: { planToken, confirm: true } }). There is no plan member — one is rejected as an unknown key. The server keeps the plan its own compile approved and reads it back, so nothing is copied out of the compile response and nothing can be mistyped or truncated in transit. Forward planToken byte-for-byte exactly as received: it is an opaque signed value, never retyped, paraphrased or rebuilt from memory, and a mangled one addresses no approved plan and is refused. PLAN_APPROVAL_NOT_FOUND means no approved plan answers to this token — already applied, lapsed, or never issued — and the recovery is to compile again; so is TOKEN_EXPIRED once the five minutes run out. Any other refusal (quota, name collision, a bound agent that changed, a moved catalog) leaves the plan applicable: clear the cause and confirm again with the same token while it lives. Changed configuration propagates to every bound agent immediately.
The authored axes — including normalized sections and conditions, each condition carrying its own required, nullable verdict — belong on the compile request. Every derived field (diff, viability, mismatches, signalRules, creationSeed, proposedRevision, bindingImpact, authoringCatalogDigest, reviewContext) is re-derived server-side and rejected as an unknown key if resubmitted, and conditionVerdicts is rejected with a message naming its replacement — the verdict belongs on the condition.
update_strategy_signal_rule({ request }) is the thin, focused one-rule edit. In multi-account mode every one of these calls uses the { account, request } sibling envelope.
Strategy-bound agents. Bind a strategy to an intelligence agent at creation time — create_intelligence_agent({ …, strategyId }) (discover strategyId via list_strategies). Rebinding via update_intelligence_agent requires confirm:true. There is no direct create_strategy operation.
Tools, prompts, and resources are discovered live from the connected server via tools/list, prompts/list, and resources/list. This package intentionally does not copy the server's catalog, formulas, signal IDs, or defaults — inspect the live connection for the authoritative, current surface. Broadly, the server exposes game play (Market Grid), market context, account state, leaderboards, intelligence agents and automation, strategy discovery/authoring, and trading signals/decisions.
Variable | Required | Description |
| One of these | Comma-separated API keys for multiple accounts |
| One of these | Single API key (fallback if |
| No | Override server URL (default: |
The announced contract is relayed, not declared. The proxy reads the upstream server's identity from the handshake it just completed and re-announces it verbatim to the local client. A local client therefore always reads the contract it will actually reach, and the package version is free to mean only what it should: this proxy's own code. There is no pairing rule to keep, and no publish-time deploy gate — a released package makes no claim about the server, so there is no ordering between a release here and a deploy there. This replaces the MAJOR.MINOR pairing that held through v11; see v12.
Fails closed, never falls back. If a connected server announced no serverInfo — a protocol violation, since it is required in a successful initialize result — the proxy refuses to start rather than substituting its own version. There is no honest number to announce in that case, and announcing a dishonest one silently is the failure this design removes.
Rediscover after a server cutover. The proxy resolves its catalog on first use and then caches it for the life of the process; a resolution that failed is retried on the next request, but a server deploy is never noticed. Package publication does not refresh it either. Restart/reconnect the proxy and re-run tools/list, prompts/list, and resources/list after the server deploys.
Restart after key rotation. API keys are read once at process startup; rotate a key, then restart the proxy.
Version | Changes |
1.x | Single/multi-account stdio proxy, identity discovery, connection retry, capability discovery |
2.0.0 | Default |
3.0.0 | Strategy-authoring major: strict |
3.0.1 | Docs only — |
4.0.0 | Realigns the package major with the server's MCP contract v4.0.0, which broke on the conditions axis: |
5.0.0 | Pairs with the server's MCP contract v5.0.0, the conditions/verdicts fusion: |
5.1.0 | Pairs with the server's MCP contract v5.1.0, which is additive: |
5.2.0 | Server contract v5.2.0, additive: four |
6.0.0 | Server contract v6.0.0, breaking: |
6.1.0 | Server contract v6.1.0, additive: |
7.0.0 | Server contract v7.0.0, breaking: the strategy-report preview limits became discoverable and the bare token count was removed. |
8.0.0 | Server contract v8.0.0, breaking, on two axes. |
9.0.0 | Server contract v9.0.0, breaking: |
10.0.0 | Server contract v10.0.0, breaking: challenge participation stops being a declared setting anywhere and becomes identical to effective trade permission, resolved per coin. |
11.0.0 | Server contract v11.0.0, breaking: a catalogued numeric output's |
12.0.0 | Server contract v12.0.0, breaking: radar's wall-clock condition changes shape ( |
12.1.0 | Server contract v12.1.0, additive: |
13.0.0 | Server contract v13.0.0, breaking: scalar families become placeable modules; six scalar headers leave the |
13.1.0 | Server contract v13.1.0, additive: four owner-scoped read tools join the catalog, closing web-client parity gaps |
14.0.0 | Server contract v14.0.0, breaking: the agent's ATR timeframe axis is removed — ATR samples on the strategy timeframe |
15.0.0 | Server contract v15.0.0, breaking: the trade-level policy moves off the agent onto the strategy |
16.0.0 | Server contract v16.0.0, breaking: a strategy condition gains a REQUIRED |
17.1.0 | Server contract v17.1.0, additive: |
17.2.0 | Server contract v17.2.0, additive: position rows gain |
18.0.0 | Server contract v18.0.0, breaking: |
18.1.0 | Server contract v18.1.0, additive: protection geometry behind the v17.2.0 verdicts |
18.2.0 | Server contract v18.2.0, additive: |
18.3.0 | Server contract v18.3.0, additive: the platform maintenance pause reaches the radar surfaces |
18.4.0 | Server contract v18.4.0, additive: |
19.0.0 | Server contract v19.0.0, breaking: the strategy regime timeframe becomes derived and read-only |
19.1.0 | Server contract v19.1.0, additive: Standing Orders markers become authorable and resolvable before save |
19.2.0 | Server contract v19.2.0, additive: |
19.3.0 | Server contract v19.3.0, additive: |
19.4.0 | Server contract v19.4.0, additive: reason codes gain |
20.0.0 | Server contract v20.0.0, breaking: the |
21.0.0 | Server contract v21.0.0, breaking: |
22.0.0 | Server contract v22.0.0, breaking: |
23.0.0 | Server contract v23.0.0, breaking: |
24.0.0 | Server contract v24.0.0, breaking: |
25.0.0 | Server contract v25.0.0, breaking: the arena stops granting trade authority; slot trade fields are rejected |
26.0.0 | Server contract v26.0.0, breaking: both entry-lifecycle guards leave the agent for platform config |
27.0.0 | Server contract v27.0.0, breaking: the trade-defaults catalog drops four unauthorable seeds |
27.1.0 | Server contract v27.1.0, additive: |
28.0.0 | Server contract v28.0.0, breaking: grid-confidence and trade-conviction bars become deployment declarations |
29.0.0 | Server contract v29.0.0, breaking: |
30.0.0 | Server contract v30.0.0, breaking: the stop-loss ceiling changes unit ( |
23.0.0 | Server contract v23.0.0, breaking: the agent-level trading mode is retired. |
24.0.0 | Server contract v24.0.0, breaking: the post-entry exit policy moves from the agent to the strategy. |
26.0.0 | Server contract v26.0.0, breaking: both entry-lifecycle guards stop being agent configuration. |
31.0.0 | Proxy change, and the end of the pairing rule. The version announced downstream is now read from the upstream handshake at connect time and relayed verbatim, instead of being a constant compiled into this package. A local client reads the contract it will actually reach, on every connection, with no release involved. Breaking because the package number now means something different — this proxy's own code, not the server's contract — so |
.github/workflows/publish.ymlis the executable release authority. If this recipe and the workflow diverge, correct them together before merging a version change.
Publishing runs only in GitHub-hosted Actions. Never run npm publish from the BattleGrid application VM or a maintainer workstation.
A version change on main is the release. Merge a pull request that changes package.json's version and the workflow does the rest: it confirms that version is not already on the registry, verifies every value expressing it agrees, tests, builds, packs, publishes with npm provenance, and tags what shipped. There is no manual tag step — the tag is an output of a successful publish, not its prerequisite.
Release whenever this package's code is ready. Since v12 the package makes no claim about the server's contract, so there is no deploy to sequence against and no gate asserting one. Publishing before, during, or after a server deploy is equally correct.
That inversion is deliberate. Publication used to be triggered by a tag, which meant a version bump with no tag published nothing and reported nothing — how 5.1.0 came to be declared in this repository and absent from the registry, caught by no check at all.
Responsibility | Environment |
Confirm npm publishing trust | npmjs.com package settings |
Prepare and merge the version change | Pull request against |
Check, build, publish, and tag | GitHub-hosted |
Verify registry publication | Any shell |
The workflow needs no BattleGrid credential. It never contacts the BattleGrid server at all. The deploy gate that used to (reading GET /mcp/version, unauthenticated by design) was retired in v12 along with the pairing rule that motivated it. Its no-credential property is worth keeping in mind if a future check ever needs the contract version: every BattleGrid MCP API key carries mcp:wager and there is no read-only variant, so a credentialed check would mean this workflow holding authority to submit wagers and close live positions in order to read a version number. GET /mcp/version exists precisely so that trade never has to be made.
Also confirm npm's Trusted Publisher for @battlegrid/mcp-server is GitHub Actions with organization playbattlegrid, repository battlegrid-mcp, workflow filename publish.yml, no environment name, and npm publish allowed. The workflow uses short-lived OIDC credentials; do not add a long-lived NPM_TOKEN.
Version this package's own code, and nothing else. Since v12 the number describes this proxy's build — a fix here, a dependency bump, a documentation correction — and makes no statement about the server. Do not move it because the server's contract moved; that used to be the whole job and is now a category error. Ordinary semver against the proxy's own surface: MAJOR for a break in how the proxy behaves or what its number means, MINOR for proxy features, PATCH for fixes and docs.
Move all three values together — package.json, both package-lock.json version fields (the root version and the self-referencing packages[""].version), and PACKAGE_VERSION in src/index.ts. The workflow compares all of them and fails closed on any disagreement.
No deploy to wait for. A release here is independent of the server's deploy schedule in both directions.
Nothing. That is the point of v12. When the server's contract moves, connected proxies announce the new version on their next connection, with no publish, no version bump, and no coordination.
Two things do still need doing, neither of them a release:
Reconnect to pick up the new contract — the announcement is read once from the handshake at startup, and the capability snapshot is resolved on first use and cached thereafter (see Rediscovery & versioning).
Document the break where the contract is documented, in battlegrid-app. Contract breaking-change notes are no longer keyed to package versions in this README, because a contract move is no longer a release here.
The workflow publishes and tags on its own; these confirm what landed.
release_version="$(node -p "require('./package.json').version")"
npm view "@battlegrid/mcp-server@${release_version}" version dist.integrity \
--json --registry=https://registry.npmjs.org/
npm view @battlegrid/mcp-server dist-tags \
--json --registry=https://registry.npmjs.org/
npm view "@battlegrid/mcp-server@${release_version}" dist.attestations \
--json --registry=https://registry.npmjs.org/Require the exact version, latest pointing at that version, and a provenance attestation. Confirm the workflow created mcp-server@${release_version} and that gitHead on the published version is the merge commit. Restart/reconnect running proxies and rediscover tools/list, prompts/list, and resources/list; publication alone does not refresh their startup cache.
There is no separate registry-reconciliation step to remember. "Is this version already published?" is the workflow's own first question — it decides whether the run publishes at all — so a bump can no longer sit in the repository unpublished and unreported the way 5.1.0 did.
The server-side release canary is not evidence about this package. battlegrid-app's server/scripts/release-canary-mcp.ts connects to the deployed endpoint and compares client.getServerVersion() against the server's own imported MCP_CONTRACT_VERSION — both sides are server-side, and it never queries npm. It passes with this package at any version, including one that was never published. Package-side evidence is exactly two things: the workflow's version-integrity gate, and the registry checks above with gitHead matching the release commit.
A reconnect no longer proves which package version is running, and this is the one verification v12 took away rather than improved. The stdio handshake now shows battlegrid@<contract> — the server's number — so it is identical whether the local proxy is 12.0.0 or a stale 11.0.0 from a cached npx. Read PACKAGE_VERSION from the installed dist/index.js, or the trailing proxy <version> field in the startup stderr line, to confirm which build is running.
If a run fails, inspect it before taking action. An ENEEDAUTH failure means the npm Trusted Publisher fields do not match the workflow — fix the publisher configuration and re-run the job; nothing was published and no tag was created, so there is nothing to move or reuse. Never mutate a published release with npm audit fix; dependency remediation goes through a new reviewed commit and version.
Install the BattleGrid skills for AI agent instructions:
npx skills add playbattlegrid/battlegrid-mcpNine skills ship from this repo, all inside the npm tarball (SKILL.md, skills/).
battlegrid (repo root) is the connection skill and is authored here: how to connect, the
{ account, request } envelope, the two scopes, and where to go for everything else.
The nine skills/battlegrid-* are exported from BattleGrid's server repository — they are the
same instructions BattleGrid's own in-app Commander runs on, which is why they name the same tools
you reach over MCP:
Skill | Teaches |
| Commission and govern intelligence agents: interview and create one against a committed strategy and an approved model, change configuration and risk limits, rebind, halt, resume, archive, and act on live positions |
| Enter Market Grid sessions: find an open session, read its coin pool and live market context, compose a grid with real per-coin reasoning or have an agent generate it, submit, then read results and the reasoning journal |
| Read the current crypto market — regime, funding and open interest, leaders and laggards, a deep-dive on any named coin — and close with the levels worth watching |
| Put agents on standing duty: per-coin Radar policies that fire on confirmed regime flips, and per-preset Arena deployment policies, previewed before they are written and un-deployed with the blast radius stated |
| Build a strategy from a plain-English idea: gather evidence, lock the spec, compile against the platform grammar, review exactly what will run, apply only on confirmation. Also fork, tune, restore, archive, preview |
| Diagnose an agent that is not doing what was expected — why it has not traded, why it stopped, whether it is healthy — from typed fields, then rank the fixes with the exact lever each needs |
| Full-surface composition patterns: custom report sections and header grammar, benchmark sections, condition trees with verdicts and enforcement gates, tiered signal weights and the aggregate gate math, routing gates, ATR trade levels, position management, plus validated desk-grade playbooks and TradingView process ports |
| Read your own trading position: where the money is, whether each agent is doing its job, what is open and how close it sits to its protections, and whether the automation is actually running |
| Find and stage a trade for one of your agents: check what is already held, scan every active coin against the agent's own gates, propose on one through the agent's own conversational turn, present the outcome with its conviction, and approve or decline only on your word |
skills/battlegrid-*is generated — do not edit it here. It is written byserver/scripts/export-mcp-skills.mjsinplaybattlegrid/battlegrid-appand arrives by pull request;skills/EXPORT.jsonrecords a hash per file andsrc/__tests__/skill-provenance.test.tsfails CI on a hand edit. Change the skill upstream and let the export lane bring it here.
This server cannot be deployed
Crypto trading intelligence MCP — 34+ endpoints, x402 pay-per-use, AI agent strategy & execution
Autonomous AI agent prediction market and reputation arena. Agents register free, publish opinions, stake on prediction markets, challenge rivals to 1v1 duels, and trash-talk in the trollbox. No API key needed — auto-registers on first connect.
No-KYC managed MCP for AI agents: sandboxed TypeScript trading SDK, isolated sub-accounts, futures.
Trade across 22+ exchanges and brokers from any MCP-capable AI agent, no install required.
