@battlegrid/mcp-server 32.0.97 → 32.0.100
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +38 -0
- package/dist/index.d.ts +1 -1
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +1 -1
- package/dist/index.js.map +1 -1
- package/package.json +1 -1
- package/skills/EXPORT.json +14 -14
- package/skills/battlegrid-agent-management/SKILL.md +79 -59
- package/skills/battlegrid-arena-play/SKILL.md +50 -28
- package/skills/battlegrid-officers/SKILL.md +17 -12
- package/skills/battlegrid-onboarding/SKILL.md +12 -10
- package/skills/battlegrid-radar-deployment/SKILL.md +112 -69
- package/skills/battlegrid-radar-deployment/references/arena-stage-examples.md +41 -1
- package/skills/battlegrid-radar-deployment/references/radar-stage-examples.md +49 -10
- package/skills/battlegrid-strategy-authoring/SKILL.md +72 -59
- package/skills/battlegrid-strategy-doctor/SKILL.md +20 -25
- package/skills/battlegrid-strategy-examples/SKILL.md +65 -35
- package/skills/battlegrid-strategy-examples/references/playbooks.md +1 -1
- package/skills/battlegrid-trade-analysis/SKILL.md +5 -1
- package/skills/battlegrid-trade-proposal/SKILL.md +6 -4
|
@@ -42,9 +42,10 @@ Explain a score, a payout or a captain choice from them, never from a remembered
|
|
|
42
42
|
2. **A stale or superseded agent grid submitted.** *Cue: the agent generated a grid and then
|
|
43
43
|
anything happened — a regeneration, a long pause, a change of mind.* The submit consumes
|
|
44
44
|
whatever generation is cached, not the one the player looked at. → step 4.
|
|
45
|
-
3. **
|
|
46
|
-
is $X".* Daily wager caps are platform config with per-user overrides
|
|
47
|
-
|
|
45
|
+
3. **A limit quoted from memory.** *Cue: the urge to say "you have N entries left today" or "your cap
|
|
46
|
+
is $X" without having just read it.* Daily wager caps are platform config with per-user overrides;
|
|
47
|
+
the account state serves today's values, and the pipeline is the authority on an attempt. → step 1
|
|
48
|
+
and the negatives below.
|
|
48
49
|
|
|
49
50
|
## Sequence
|
|
50
51
|
|
|
@@ -67,6 +68,20 @@ instead of `mcpWagerEnabled`.
|
|
|
67
68
|
is still refused at the fee. The wager pipeline re-checks at the fee itself and its refusal is
|
|
68
69
|
the authority in either direction.
|
|
69
70
|
|
|
71
|
+
**`dailyWagerCaps`** — today's MCP entries (`count`) and entry volume in USD (`volumeUsd`) against the
|
|
72
|
+
player's daily caps, each with `used`, `limit`, `remaining`, the server's `status` and `resetsAt`. The
|
|
73
|
+
caps bind every entry made through MCP — this conversation's included — for the player and their agents
|
|
74
|
+
alike; entries made in the web app are outside them.
|
|
75
|
+
|
|
76
|
+
- **`count.status` EXHAUSTED, or the session's entry fee above `volumeUsd.remaining`** ⇒ a **definite
|
|
77
|
+
refusal**, before any submission. Say which cap is spent and when it resets (`resetsAt`, a UTC
|
|
78
|
+
instant), and call no submission tool — a replacement of an agent's entry included.
|
|
79
|
+
- **LOW** ⇒ say how much remains — "this is your last entry today" — before the commit.
|
|
80
|
+
- **`null`** ⇒ the caps could not be read. Say so, quote no figure, and let the pipeline decide: it is
|
|
81
|
+
neither readiness nor a refusal.
|
|
82
|
+
- **Open** is never a promise. Another client can spend the last entry first; the pipeline's own
|
|
83
|
+
refusal is the authority for an attempt.
|
|
84
|
+
|
|
70
85
|
**In a conversation a web Commander surface hosts, the switch renders beneath the account card just
|
|
71
86
|
shown, or beneath the refusal — offer it there.** Say the arena needs Agent Wagers and point at the
|
|
72
87
|
control rather than at a settings path they would have to leave the conversation for. Every other
|
|
@@ -120,7 +135,7 @@ ask. The commit requires a `RATIONALE` axis — `reasoning`, `confidenceScore`,
|
|
|
120
135
|
- Each entry's reasoning cites **the market context you read in this conversation**, for that
|
|
121
136
|
coin. Boilerplate reasoning attached to a real wager is worse than none — it looks like
|
|
122
137
|
evidence.
|
|
123
|
-
- `confidenceScore` is your honest read, and you state it to the player
|
|
138
|
+
- `confidenceScore` is your honest read, and you state it to the player before the commit.
|
|
124
139
|
- **The `RATIONALE` is public.** Once the session locks, every player sees it beside the grid — the summary,
|
|
125
140
|
the confidence, the model and each pick's reasoning — as they see an agent's. Write it for that reader.
|
|
126
141
|
|
|
@@ -133,29 +148,35 @@ Present the generation as a card: the agent, its confidence against the threshol
|
|
|
133
148
|
(`confidenceThreshold: null` means no deployment declares a bar here, so the grid is **ungated** —
|
|
134
149
|
say that rather than inventing a bar), whether it meets the threshold, and the per-coin picks.
|
|
135
150
|
|
|
136
|
-
### 4. Read again,
|
|
151
|
+
### 4. Read again, state it, then commit promptly on the player's approval
|
|
137
152
|
|
|
138
153
|
`get_market_grid_entry_draft` again after staging: show the player the diff, any diagnostic, and the
|
|
139
|
-
`impact` — `ENTER` with the `entryFee` the commit charges, or `REPLACE` for none. One
|
|
140
|
-
naming **the session, that fee, and the confidence** — then commit on
|
|
154
|
+
`impact` — `ENTER` with the `entryFee` the commit charges, or `REPLACE` for none. One statement
|
|
155
|
+
naming **the session, that fee, and the confidence** — then commit, **only on the player's
|
|
156
|
+
approval**. A client that prompts the player to approve tool calls collects it: call right after the
|
|
157
|
+
statement and ask no separate confirming question — the client asks the player, or applies the
|
|
158
|
+
standing setting they chose, and the call runs only if approved. A client that does not asks in its
|
|
159
|
+
reply and waits for the player's word. Words typed while an approval prompt is open are not
|
|
160
|
+
approval: that call does not run, so answer what they said and call again only when they ask.
|
|
141
161
|
|
|
142
162
|
- Direct → `commit_market_grid_entry_draft({ sessionId, draftVersion, expectedRevision })`,
|
|
143
163
|
naming exactly the `draftVersion` and `committedRevision` the last read returned (`null` with no
|
|
144
|
-
entry). A refusal naming `get_draft` means the draft or the entry moved: read again and
|
|
145
|
-
again — never retry with a bumped number.
|
|
164
|
+
entry). A refusal naming `get_draft` means the draft or the entry moved: read again and state it
|
|
165
|
+
again before a fresh call — never retry with a bumped number.
|
|
146
166
|
- Agent → `submit_agent_grid`, which consumes the cached generation for that (session, agent).
|
|
147
167
|
|
|
148
|
-
**A regeneration voids any prior
|
|
149
|
-
whatever is currently cached — so a grid regenerated after the player
|
|
150
|
-
under
|
|
151
|
-
|
|
168
|
+
**A regeneration voids any prior approval.** The submit takes no generation token — it reads
|
|
169
|
+
whatever is currently cached — so a grid regenerated after the player approved would be submitted
|
|
170
|
+
under an approval given for different picks. State the new generation, every time, and submit it
|
|
171
|
+
promptly, its call waiting on its own approval.
|
|
152
172
|
|
|
153
173
|
**If the submit is refused with `CONFLICT`**, the player's entry changed after the grid was generated —
|
|
154
174
|
replaced, cancelled, re-placed, or placed where there was none. No fee moved. The recovery is
|
|
155
|
-
`generate_agent_grid` again, a fresh
|
|
175
|
+
`generate_agent_grid` again, a fresh statement of the new grid, then `submit_agent_grid` on the
|
|
176
|
+
player's approval.
|
|
156
177
|
|
|
157
178
|
**If the submit is refused for a missing pending generation**, the ten-minute window lapsed. The
|
|
158
|
-
recovery is `generate_agent_grid` again and a fresh
|
|
179
|
+
recovery is `generate_agent_grid` again and a fresh statement. **Never** convert the stale picks into
|
|
159
180
|
a hand-staged draft committed under the player's own name — that enters an agent's reasoning as the
|
|
160
181
|
player's, and the attribution is part of what was scored.
|
|
161
182
|
|
|
@@ -188,7 +209,7 @@ arrives as the attached scope, and the session it names is the one in scope.
|
|
|
188
209
|
- **Stage, do not enter.** When the player asks you to pick or fill in their grid, read the draft and
|
|
189
210
|
stage the cells into it (step 3). Each staged pick appears in their builder at once, labelled as
|
|
190
211
|
yours; nothing is entered and no fee or deposit is needed. Their own **PLACE ENTRY** places it;
|
|
191
|
-
offer to place it for them only through step 4's
|
|
212
|
+
offer to place it for them only through step 4's statement and the draft commit, on their approval.
|
|
192
213
|
- **Their picks are in the draft.** A pick they make in the builder lands there as they make it, so
|
|
193
214
|
read the draft before answering what they have picked or staging over it. A pick their builder
|
|
194
215
|
rewrote after you staged it is their answer — stage it again only when they ask.
|
|
@@ -196,8 +217,8 @@ arrives as the attached scope, and the session it names is the one in scope.
|
|
|
196
217
|
`generate_agent_grid` for the session in scope (the billed-inference warning in step 3 still
|
|
197
218
|
applies). The result reaches them as a card with **LOAD INTO GRID**, which puts the picks into their
|
|
198
219
|
builder as a starting point. A loaded grid they have not changed is the agent's: if they want it
|
|
199
|
-
entered as it stands, run step 4, then `submit_agent_grid` on their
|
|
200
|
-
|
|
220
|
+
entered as it stands, run step 4, then `submit_agent_grid` on their approval as step 4 describes.
|
|
221
|
+
Once they change a pick, the grid is theirs.
|
|
201
222
|
- **A newer generation for the same agent replaces the older card.** The older one reads SUPERSEDED;
|
|
202
223
|
only the newest is what `submit_agent_grid` would enter.
|
|
203
224
|
|
|
@@ -208,24 +229,25 @@ Each is typed and distinguishable. Read the code, do not paraphrase:
|
|
|
208
229
|
- **No wallet consent** — the shared wager pipeline's own refusal, carrying its enablement copy.
|
|
209
230
|
Surface that path verbatim. This is the same refusal an external MCP client gets; there is no
|
|
210
231
|
chat-side difference and no chat-side override.
|
|
211
|
-
- **A daily cap** — a typed rate-limit refusal. Say the cap is spent and
|
|
212
|
-
|
|
213
|
-
can't play right now".
|
|
232
|
+
- **A daily cap** — a typed rate-limit refusal. Say the cap is spent and when it resets — the
|
|
233
|
+
`resetsAt` of a fresh `get_account_state`. Distinguishable from the run's op budget and from
|
|
234
|
+
consent — do not merge them into "you can't play right now".
|
|
214
235
|
- **Funds, session state, an already-submitted grid** — ordinary typed refusals. Report what the
|
|
215
236
|
server said.
|
|
216
237
|
|
|
217
238
|
## Negatives
|
|
218
239
|
|
|
219
|
-
- **Never
|
|
220
|
-
the
|
|
221
|
-
|
|
240
|
+
- **Never quote a limit from memory** — not the daily count, not the volume cap, not the reset.
|
|
241
|
+
Quote the served `dailyWagerCaps` of the read you just made: the server holds the real value
|
|
242
|
+
including per-user overrides, and a number repeated from anywhere else will eventually be wrong on
|
|
243
|
+
a surface where being wrong costs money.
|
|
222
244
|
- **Never use `random_submit_market_grid`.** It shuffles the pool and wagers on the result with no
|
|
223
245
|
reasoning at all. It exists for other clients; it is not something to offer a player who asked
|
|
224
246
|
you to think.
|
|
225
|
-
- **
|
|
226
|
-
a permission check, do not add a wager gate of your own, and do not
|
|
227
|
-
|
|
228
|
-
|
|
247
|
+
- **The player's approval of the call is the client's, never authorization.** Do not describe it as
|
|
248
|
+
a permission check, do not add a wager gate of your own, and do not ask a separate confirming
|
|
249
|
+
question where the client already prompts for the call. Words typed while an approval prompt is
|
|
250
|
+
open never approve the call.
|
|
229
251
|
- **Never re-commit over an unknown outcome.** An interrupted commit may have moved the fee. A retry
|
|
230
252
|
of the same `draftVersion` returns the first commit's receipt; otherwise run
|
|
231
253
|
`check_market_grid_submission` and `get_market_grid_entry_draft` first and report what actually
|
|
@@ -12,17 +12,18 @@ setting an officer up, running it, reading where it stands, and acting on what i
|
|
|
12
12
|
|
|
13
13
|
Every officer run is billed to the player's LLM credit, and an applied proposal changes a live strategy, agent or
|
|
14
14
|
Radar deployment. So a setup goes through its own draft, a run is started only on the player's word, and Apply is
|
|
15
|
-
|
|
15
|
+
stated first and runs only on the player's approval.
|
|
16
16
|
|
|
17
17
|
## The four failures this flow exists to prevent
|
|
18
18
|
|
|
19
19
|
1. **Writing a review yourself.** *Cue: "what would my officer say about…", wanting to record a finding.* You read
|
|
20
20
|
an officer's reviews and give your own reading in the conversation; only the officer records a review. → §2.
|
|
21
|
-
2. **Applying without
|
|
22
|
-
target. Show the review, then
|
|
21
|
+
2. **Applying without approval.** *Cue: "apply it", "make it live".* Apply publishes the proposal to a live
|
|
22
|
+
target. Show the review and name what Apply publishes, then call it on the player's approval. → §4.
|
|
23
23
|
3. **Computing what is served.** *Cue: working out the next run from the schedule, counting open reviews.* Every
|
|
24
24
|
status figure is on the read. → §3.
|
|
25
|
-
4. **Committing a setup unasked.** *Cue: a staged officer that looks right.* The commit
|
|
25
|
+
4. **Committing a setup unasked.** *Cue: a staged officer that looks right.* The commit runs on the player's
|
|
26
|
+
approval, never your judgement. → §1.
|
|
26
27
|
|
|
27
28
|
## 1. Set up and update an officer through its draft
|
|
28
29
|
|
|
@@ -43,15 +44,18 @@ confirmed before it is called.
|
|
|
43
44
|
`list_intelligence_agents`) and `levers`; `MODEL` names a `modelId` from `list_approved_models`. `INSTRUCTIONS`
|
|
44
45
|
are the player's brief in their own words: at most 2,000 characters, naming no tool and no id — the commit
|
|
45
46
|
refuses either.
|
|
46
|
-
- **Show, then commit on the player's
|
|
47
|
+
- **Show, then commit on the player's approval.** Read `get_officer_draft` and show the diff, the diagnostics and the
|
|
47
48
|
impact — `nextRunAt` is when it first runs, and a create's `admission` says how many officers the player holds
|
|
48
49
|
against their account plan's officer limit and how many more it admits (`officersRemaining`). Every unarchived
|
|
49
50
|
officer counts, paused ones included; a create past the limit is refused as `PLAN_LIMIT_REACHED`, and archiving
|
|
50
|
-
an officer frees its place.
|
|
51
|
-
`committedRevision` the read returned,
|
|
51
|
+
an officer frees its place. Then commit with `commit_officer_draft` naming exactly the `draftVersion` and
|
|
52
|
+
`committedRevision` the read returned, only on the player's approval: a client that prompts the player to
|
|
53
|
+
approve tool calls collects it, so call right after showing the read and ask no separate confirming question; a
|
|
54
|
+
client that does not asks in its reply and waits. Words typed while an approval prompt is open are not
|
|
55
|
+
approval: that call does not run, so answer what they said and call again only when they ask. A cadence of `OFF` keeps the setup and
|
|
52
56
|
runs nothing on a schedule.
|
|
53
|
-
- **Archive** with `archive_officer` only
|
|
54
|
-
reviews stay in the inbox.
|
|
57
|
+
- **Archive** with `archive_officer` only when the player asks, saying first that an archived officer runs no
|
|
58
|
+
more and its open reviews stay in the inbox; the call runs on their approval, as a commit's does.
|
|
55
59
|
- In the app, the player sets an officer up in its own console, where this conversation may be docked beside the
|
|
56
60
|
form — stage into the officer it is scoped to.
|
|
57
61
|
|
|
@@ -90,9 +94,10 @@ them as served, through their display names.
|
|
|
90
94
|
served; `applyAvailable` says Apply is offered, `heldAxes` names every axis the target's draft holds and who wrote
|
|
91
95
|
it, and `baseMoved` says the target was committed since the review. `awaitingOwner` says the review waits on the
|
|
92
96
|
player; a review that proposes no change (lever `NONE`) is for their information and never waits.
|
|
93
|
-
- **Apply** with `apply_officer_proposal` only after
|
|
94
|
-
|
|
95
|
-
|
|
97
|
+
- **Apply** with `apply_officer_proposal` only after naming the review, its target and its axis, and only on the
|
|
98
|
+
player's approval of the call: a client that prompts for tool approval collects it, so ask no separate
|
|
99
|
+
confirming question; one that does not asks in its reply and waits. Apply publishes the proposal to the live
|
|
100
|
+
target; it needs `mcp:wager`.
|
|
96
101
|
- **Push** with `push_officer_proposal` and **Dismiss** with `dismiss_officer_review` on the player's explicit word,
|
|
97
102
|
naming the review. A push places the proposal in the target's draft for the player to save from the editor; a
|
|
98
103
|
dismissal closes the review and touches no draft.
|
|
@@ -14,20 +14,22 @@ step the server says is next, and keep going until the account is battle-ready.
|
|
|
14
14
|
Call `get_onboarding_requirements` before you say anything about setup, and again after every
|
|
15
15
|
step you complete. Read its `status`, its `next` and the `remedy` on the step `next` names, as
|
|
16
16
|
served. Never judge readiness from anything else — a deposit you saw, an agent you created, a
|
|
17
|
-
policy you read — never
|
|
17
|
+
policy you read — never list or count the steps, and never decide an order: the server already
|
|
18
18
|
chose it, and the app's setup bar names the same step from the same read.
|
|
19
19
|
|
|
20
20
|
## Each turn
|
|
21
21
|
|
|
22
|
-
1. **Read the ladder.** If `status` is battle-ready, say the account is set up
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
to
|
|
30
|
-
|
|
22
|
+
1. **Read the ladder.** If `status` is battle-ready, say in one sentence that the account is set up
|
|
23
|
+
and end your turn. Ask nothing, and do not offer an optional step unless the player asks for one.
|
|
24
|
+
2. **Take the step `next` names.** Name the step by its served `title`, then say what it gets the
|
|
25
|
+
player in the step's served sentence, and act on its `remedy` by who performs it (`actor`):
|
|
26
|
+
- **`COMMANDER` — you do it, with the player's go-ahead.** Offer it with `ask_user` and end your
|
|
27
|
+
turn on that question. Call nothing for the step, not even the owning skill's reads, until the
|
|
28
|
+
player says yes; then turn to the skill that owns the act and follow that skill's own
|
|
29
|
+
procedure to the end: the `strategy-authoring` skill builds a strategy; the `agent-management`
|
|
30
|
+
skill creates an agent or resumes a paused one; the `radar-deployment` skill puts an agent on
|
|
31
|
+
Radar or Arena duty. A resume is the act to try, not a promise: if the owning skill reports
|
|
32
|
+
the stop is still breached, offer its exits and do not repeat the resume.
|
|
31
33
|
- **`PLAYER` — only the player can do it, in the app.** It is a deposit, a signature, a
|
|
32
34
|
payment or an API key. Say what it is and why it is next. On the web, the control that does
|
|
33
35
|
it sits directly under your reply; elsewhere, tell the player where in the app it lives.
|