Specifies what counts as "tested" for this repo, the local test
commands the bot treats as gates, and how pragent reads the repo
config (`.pr-review.json` + static architecture.md from Nexus
raw-hosted).
Auto-loaded by Gitea on every new PR. Lists the surfaces and risks
pragent scores you on; tells the author to add the AI-REVIEW label
to trigger a bot review.
Wires the new ADDITIONAL_CONTEXT_URL feature on the central webhook service
(see gitea_admin/pragent e8ebc54) — fetches a static maintainer-curated brief
from the Nexus raw-hosted repo at review time. The doc distils the house
rules so the agent doesn't have to re-derive them from the diff each PR:
* chat messages are immutable after send (counts baked into buttons freeze)
* translatable components + singular-form rule for pt-BR sentences
* Geyser/Bedrock can't click — every click has a typed fallback
* vanilla stats only — <world>/players/stats/<uuid>.json, NOT <world>/stats
* reactions keep counting late (reacao-validade-minutos)
* Floodgate is optional runtime, never compile-time
Plus the canonical focus/instructions/languages config: java, balanced style,
claude-sonnet-5 cost target, JDK 25 build note.
Co-Authored-By: Claude <noreply@anthropic.com>
README: add ia to the modules table, the four ia permissions, and a full
IA section covering commands, profile switching, operator corrections,
memory/context, and the key + rate-limit model. Notes the safety boundary
(no tools, reply never executed, leading slashes stripped) and which config
values are baked vs live.
Co-Authored-By: Claude <noreply@anthropic.com>
Review of Task 11 (commit 5f616d3) flagged a blocking spec regression:
canalhandia.ia.privado was declared (default: true) but never checked --
ia(sender, args, isPrivate) only checked canalhandia.ia, so /iap silently
required canalhandia.ia and the privado perm did nothing.
Per operator decision, /ia and /iap are now separate gates, both default
op, so LuckPerms can grant them independently (operator + permitted only,
not everyone). /ia needs canalhandia.ia; /iap needs
canalhandia.ia.privado.
Also fixes three should-fix findings:
- Subcommand hijack is no longer greedy: perfil and feedback only hijack
when the second token is one they act on (a known profile key, or "ruim"),
so "/ia perfil do servidor" and "/ia feedback do mapa?" fall through and
are asked. corrigir stays greedy (a correction always reads the rest).
- AiProfile.isValid tells a real key from the PRECISO fallback, so /ia perfil
blah no longer silently switches to PRECISO.
Co-Authored-By: Claude <noreply@anthropic.com>
- Guard conversations.remember() on asker==null: a PlayerQuitEvent
forgets the history (carry-forward #6), and re-adding here after the
quit would resurrect it. lastAnswer stays regardless so /ia corrigir
can still correct the last answer.
- Saturate the retry token ceiling: max_tokens near Integer.MAX_VALUE
would overflow to a negative budget sent to the API.
- Tighten the constructor comment: aiUrl/aiWikiChars/aiMemoryExchanges/
aiMemoryMinutes are baked; everything else (aiProfile, aiModel,
aiMaxTokens, aiTemperature, aiInstructions, aiServerContext) is read
live, not just aiProfile.
- Drop unused import java.util.List; add trailing newline.
all() returns List.copyOf(entries), which iterates; add()/load()
structurally modify. Once Task 10 wires all() into runTaskAsynchronously
and Task 11 wires add() from the main-thread /ia corrigir command,
the race throws ConcurrentModificationException. Guarded by entries'
own monitor like Conversations: file parse and YAML save stay outside
the lock, only fast in-memory work is under it.
Guarded by a monitor: Ai calls this from runTaskAsynchronously and
history() prunes while it reads, so an unsynchronised map would throw
ConcurrentModification into a player's answer or corrupt itself on resize.
Expiry compares nanoTime differences rather than wall-clock instants, so
a zero window expires deterministically and an NTP step backwards cannot
leave entries stamped in the future and unexpirable. Player count is
capped LRU, since expiry only prunes players who ask again.
Fuzzing all 65,536 char values, not the 256 of the previous pass, shows
the JDK rejects 65,312 of them from a header value and every single one
echoes the value back. The previous deny-list covered 33. A key file
saved by Notepad or PowerShell Out-File carries a U+FEFF byte order
mark, which passed both filters and reached the quoting validator.
Ai.cleanKey and HttpFetcher.checkBearer now allow printable ASCII only,
an allow-list of the 94 characters a bearer token is made of, which
cannot drift out of what the JDK accepts. Demonstrated invariant:
cleanKey output ⊆ checkBearer accepts ⊆ JDK accepts, 0 violations.
Ai.call gains the same check and its failure log is now redacted; that
was the one path where the proof-of-concept leak surfaced.
MiniMax.message and answer now check JSON types before assuming them.
Gson throws unchecked on JSON that parses but has the wrong shape, and
both run outside post()'s try, so {"choices":["str"]} escaped to an
async Bukkit worker as a stack trace instead of the promised null.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3jSvSTrG4qsniLSr6TpC6
HttpRequest.Builder.header() quotes the offending header value back in
its IllegalArgumentException. Fuzzed on temurin-25: all 32 control
characters it rejects echo the value, so a key carrying any of them ends
up in whatever log catches the throw. A key file with a comment on line
two survives trim() and is enough to trigger it.
Ai.cleanKey now takes the first non-blank line and drops control
characters, so a malformed key never forms. HttpFetcher.checkBearer
rejects one anyway before the request is built, with a message naming
only the position, so no future caller has to remember to redact. Its
reject set is a strict superset of the JDK's.
Also renames MiniMax.Msg to MiniMax.Turn: the package already has a
top-level Msg, the chat-formatting helper, which the record shadowed
inside MiniMax.java.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3jSvSTrG4qsniLSr6TpC6
Term selection is a forced tool call rather than free text: measured 5/5
against 2/7 for a free-text extraction call, because a tool argument is
structured output and survives the hidden reasoning eating the budget.
With tool_choice auto the model skipped the search on exactly the
questions it was most likely to get wrong.
Failures are reported through an optional warn consumer, as Wiki does,
and the endpoint is injectable for MiniMax's regional hosts.
Warnings are redacted of the key: a key with an embedded newline makes
the JDK throw invalid header value: "Bearer sk-...", quoting the whole
value back, and that lands in the generic call-failure path.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3jSvSTrG4qsniLSr6TpC6
The substring fallback in materialFor returned plausible recipes for
items the player did not ask about. "Redstone Repeater" tied on length
and resolved to REDSTONE, handing someone asking about repeaters the
recipe for redstone dust; "Book and Quill" gave BOOK, "Minecart with
Chest" gave MINECART, "Rabbit's Foot" gave RABBIT, "Chestplate" gave
CHEST. That is the exact failure this feature exists to remove, and it
is worse than no answer because the model states it confidently.
Dropped it. Exact match plus a hand-checked alias table, else null. No
suffix rule either: "axe" is a suffix of "pickaxe", so tool and armour
families would fail the same way. Beds and wool now return null, which
is correct.
brewing() matched keys as bare substrings and hijacked real questions:
"salto" inside basalto, "cura" inside curar, "forca" inside reforcar.
Since describe() answers brewing first, each took over the whole answer.
Now gated on the question mentioning a potion, and matched on word
boundaries.
Table corrections: "Fogo do Dragao" is not an item, it is Bafo do Dragao
giving a lingering potion; Frasco de Agua, not Garrafa de Agua, which is
the empty bottle; Pe de Coelho; Fatia de Melancia Reluzente. Added the
in-game item names players actually type (Agilidade, Dano) and stripped
hyphens so Mestre-Tartaruga reaches the table.
describeShapeless no longer emits a dangling "Sem formato: ".
describeChoice could not be covered after all: constructing a
MaterialChoice initialises org.bukkit.Registry, which needs a server,
and the class is sealed so it cannot be faked. Verified in Task 13.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3jSvSTrG4qsniLSr6TpC6
Crafting recipes come from Bukkit.recipeIterator(), which is authoritative
for this exact version. Player questions are Portuguese and Material names
are English, so the subject of the question is searched on the pt wiki and
prop=langlinks&lllang=en gives the English title, which uppercases onto the
enum constant.
Matching material names against the question directly does not work.
Measured over twenty realistic pt-BR questions it resolved 1 of 20, and that
one ("tridente" containing "trident") by coincidence rather than
translation. Through langlinks the same twenty resolve 17 of 20.
Brewing is not exposed by Bukkit at all: there is no brewing Recipe type,
PotionBrewer has no getter or iterator, and vanilla brewing is hardcoded in
PotionBrewing rather than registered as a recipe. Potions therefore come
from a hand-written pt-BR table. This is what actually answers the
fire-resistance question that motivated the feature; recipeIterator() alone
never could have.
Ingredients are read through getChoiceMap/getChoiceList. The deprecated
getIngredientMap/getIngredientList collapse a choice to one arbitrary stack,
printing "oak planks" where the recipe accepts any plank.
The design doc claimed recipeIterator() closed the potion case. It did not;
corrected to record what is true.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3jSvSTrG4qsniLSr6TpC6
Every failure mode returned null, which the caller cannot tell apart
from a term the wiki has no article for. A 403 would revert /ia to the
confidently wrong answers grounding exists to stop, against a clean log.
Also clamps maxChars so a config of 0 cannot switch grounding off for
good, restores the interrupt flag on disable, and distinguishes a
malformed response from an outage in the log.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
hasForeignScript detects by alphabet, so Latin-script leakage such as
the observed French "contiennent" is not caught. Say so in the javadoc
and admit the gap in the design doc rather than implying coverage.
failIfNoTests catches a misplaced or misnamed test class, not a
disabled one: an @Disabled class still reports as skipped and the
build stays green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3jSvSTrG4qsniLSr6TpC6
Moves sanitise out of the Bukkit-bound Ai class so it can be unit
tested, and adds hasForeignScript to catch the CJK words the model
intermittently drops into Portuguese answers.
Colour-code stripping now removes the code character too: replacing
only the section sign left "§c" reading as a stray "c" in chat.
Surefire now fails on an empty suite, so a misplaced or disabled test
class cannot pass as a green build.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3jSvSTrG4qsniLSr6TpC6
pragent-bot
Marcos Paulo
masi