PathOfAIBuilding v1.1.5

V1.1.5 — context recovery

The bridge continues when a model validly requests additional data but wraps the block in explanatory prose.

• Recovery
  One valid <context_request> block is now accepted with prose before or after it; the prose is discarded and the question is repeated with the requested context.
• Safety
  Any <actions> marker alongside the request still blocks the response before the build can change.
• Boundary
  Duplicate blocks, invalid JSON, and scopes outside the catalog remain rejected.
• Test
  Regressions cover prose-wrapped context, duplicate blocks, and mixed actions.

V1.1.4 — canonical gem actions

A minion recommendation no longer loses its whole batch because the model pluralized a gem that PoB stores in the singular.

• Context
  Questions about skills, minions, summons, and summoners receive PoB's canonical gem catalog.
• Contract
  The AI must use exact state.reference.gemNames strings in add_skill and set_main_skill.
• Compatibility
  The bridge normalizes a singular/plural variation only when one canonical gem name is possible, before preflight.
• Test
  Regressions cover a level-30 minion starter and Summon Raging Spirits normalized to Summon Raging Spirit.

V1.1.3 — resilient AI chat

The session remains usable when transport or model output leaves the expected path.

• Transport
  The bridge detects when a request or connection test cannot start and returns an immediate error, allowing a new attempt.
• Actions
  <actions> tags accept case and whitespace, while incomplete, invalid, duplicate, object-shaped, or trailing JSON is rejected with a visible message.
• Stream
  An animated reply is completed before a later fast reply replaces the stream cursor, so no earlier response remains blank.
• Coverage
  Regression cases cover startup failure, response-processing exceptions, malformed action tags and JSON, pending proposals, and immediate follow-ups.

V1.1.2 — unapplied proposal correction

A correction request does not assume the previous proposal took effect. The serialized PoB state remains the source of truth.

• Context
  The bridge explicitly tells the AI that the previous proposal was not applied and requires a fresh <actions> block for corrections or replacements.
• UI
  Actions from the new reply replace the earlier proposal; Apply represents only the current batch again.
• Safety
  The build does not change during the conversation: every batch still requires explicit player Apply.
• Test
  Regression coverage corrects a level-98 proposal to level 97 without touching the build before Apply.

V1.1.1 — generic provider options

AI configuration no longer depends on model-specific conditionals: users enter extra fields for their OpenAI-compatible endpoint directly in PoB.

• Config
  AI Setup now exposes Extra request options (JSON), persisted in ai_config.json.
• Bridge
  Validated options are sent for the connection test and every chat request; there is no MiniMax- or provider-specific branch.
• Safety
  model, messages, and stream remain bridge-managed to preserve the chat contract.
• MiniMax
  For M3 direct answers, use {"thinking":{"type":"disabled"}} in the new field.

V1.1.0 — 3.29 passive-tree support

Compatibility update for the current PoE 1 passive tree, integrated from the Path of Building Community dev branch.

• Tree
  Added TreeData/3_29 and TreeData/3_29_ruthless, including tree.lua, sprites, textures, and passive-tree assets.
• Default
  GameVersions.lua registers standard and Ruthless 3.29; latestTreeVersion now resolves to 3.29 for new builds.
• Import
  Import/export mappings now recognise 3.29 and its Ruthless variant.
• Tree UI
  Passive-tree sprite loading accepts the coordinate shape used by the 3.29 data; the UI also recognises the Luminary ascendancy while importing characters.
• Sync
  Synced upstream commits after V1.0, including calculation, comparison, import, and SimpleGraphic runtime fixes.
• AI
  The AI tab, local configuration, preflight, and V1.0 guards were preserved through the upstream merge.

V1.0.0 — AI integration

Technical history for the first public release. The sections below describe the architecture, validation, and limits that remain in force in V1.1.

Integration and interface

• Added
  Classes/AIChatTab.lua as the new Build.viewMode == "AI", registered next to the Compare tab in Modules/Build.lua.
• Added
  Global AI Setup/AI Ready button in Modules/Main.lua, opening AIConfigPanel through PoB's native popup system.
• Added
  Chat messages with user, assistant, and system roles; multiline input; Enter to send; and three quick prompts.
• Changed
  Requests run in a Lua subprocess through LaunchSubScript, keeping PoB's drawing thread responsive.
• Added
  The HTTP callback receives the complete answer, then AIChatTab:UpdateStream() reveals it progressively in the UI. This is not API token streaming.
• Note
  History is session-only: AIChatTab:Load() does not restore messages and Save() does not write chat into the build.
• Guard
  AIBridge.pending and AIChatTab.pending restrict the integration to one concurrent request.

Build serialisation and context

• Added
  AIBridge:SerializeBuild() produces a versioned JSON payload (version = 1) with meta, stats, items, skills, and tree.
• Added
  Real calcsTab.mainOutput values: DPS, EHP, life, ES, mana, ward, armour, evasion, resistances, regeneration, leech, block, suppression, attributes, DoT, five Max Hits, and minion stats when present.
• Added
  Equipped items are sent per slot with name, rarity, base, and Item:BuildRaw(); skills include group, slot, level, quality, enabled state, support flag, and skillPart.
• Added
  The tree serialises allocated IDs, notables, keystones, sockets, jewels, masteries, and points. Optional candidates are capped at the 50 closest reachable notables/keystones by node.path.
• Changed
  ClassifyQuestion() routes context by intent. Optional scopes are gems, uniques, tree, and config; large catalogues are not sent on every request.
• Added
  Structured escalation through <context_request>. The model may request one context expansion; a response cannot combine a context request with <actions>.
• Changed
  Recent history is capped by HISTORY_CHAR_BUDGET = 12000; older message pairs are removed before request assembly.
• Guard
  State fingerprint: SHA-1(tostring(build) .. "\0" .. build:SaveDB()). Changes during a response invalidate the request, caches, and Apply.

API and HTTP transport

• Added
  OpenAI-compatible API client. The configured endpoint has trailing slashes removed before appending /chat/completions.
• Added
  Main payload: model, messages array, max_tokens = 2048, and temperature = 0.3.
• Added
  Separate connection test with a minimal prompt, max_tokens = 16, and temperature = 0.
• Changed
  launch:DownloadPage() now accepts a POST body and timeout; it uses lcurl.safe, HTTP compression, redirects, and an asynchronous callback.
• Changed
  Total timeout is configurable from 1 to 600 seconds, default 120 seconds; connection timeout is min(timeout, 10).
• Guard
  A response requires HTTP 200, a non-empty body, and JSON with choices[1].message.content. Transport and shape errors are returned to the panel.

Simulations and shortlists

• Gems
  ComputeGemShortlist() tests non-legacy supports not already linked to the main group at level 20/quality 0, using ProcessSocketGroup() + BuildOutput().
• Guard
  Every temporary support is removed even when simulation throws. The original output is rebuilt at the end; only positive gains enter the top 10.
• Uniques
  ComputeUniqueShortlist() preselects up to 40 candidates per item type: 30 by skill/defence affinity and 10 deterministic samples from the tail.
• Changed
  Uniques are calculated without equipping them by calling GetMiscCalculator() with repSlotName/repItem. Ranking preserves up to three objectives per slot: balanced, DPS, and EHP.
• Added
  Gem and unique results are cached against the exact build fingerprint and invalidated on any state change or Apply.
• Added
  Gem, unique, or passive mentions in one answer feed a detailed lookup into the next question.
• Scope
  Shortlists are calculated samples, not exhaustive proof across every game combination.

Action protocol — 18 types

Actions are emitted as JSON inside <actions>[...]</actions>. Array order is preserved.

• set_levelinteger from 1 to 100
• set_classclass by name
• set_ascendancyprimary ascendancy by name
• set_secondary_ascendancyleague secondary ascendancy
• set_banditNone, Oak, Kraityn, or Alira
• set_pantheonmajor and/or minor
• add_skillnew socket group with gems
• remove_skillgroup by label or gem name
• set_main_skillmain group/skill
• set_skill_partpositive part index
• equip_itemraw item text and optional slot
• equip_jewelraw jewel and optional nodeId
• alloc_nodenode ID or name; uses PoB's path
• dealloc_nodenode ID or name
• set_masteryeffect index or text substring
• apply_tattoonode + exact tattoo name
• remove_tattoonode ID or name
• set_configexisting key + boolean/number

Validation, preflight, and Apply

• Validate
  ParseActions() uses dkjson; the batch must be a non-empty dense array with no gaps, one of the 18 types, and a valid shape for each action.
• Snapshot
  Before preview, build:SaveDB() captures the canonical build XML.
• Preflight
  new("CompareEntry", snapshotXml, "AI action preflight") creates an isolated clone. Every action runs in order on that clone, allowing dependencies within the batch.
• Changed
  A complete rebuild calls wipeGlobalCache(), increments outputRevision, runs BuildOutput(), and refreshes the stat list.
• Diff
  Preview captures DPS (Combined/Total/Minion fallback), EHP, physical/fire/cold/lightning/chaos Max Hit, passive points, and ascendancy points before and after.
• Guard
  The fingerprint is checked on response, before snapshot, after any rebuild, and after preflight. A stale preview cannot be applied.
• Apply
  Only the Apply button executes the batch on the real build; it is followed by a rebuild, another metrics capture, and cache invalidation.
• Caveat
  Preflight is all-or-nothing. Real Apply is sequential: an unexpected divergence after a valid preflight may produce partial = true; the report exposes appliedCount and the error. V1 has no automatic rollback.

Configuration and security

• Added
  Modules/AIConfig.lua stores endpoint, API key, model, timeout, and config version in userPath/ai_config.json, outside the repository and build file.
• Storage
  The key is masked in the EditControl but persisted as plain text in local JSON. On compatible systems the app attempts chmod 600; V1 has no keychain integration.
• Guard
  The endpoint must start with https://. Certificate verification follows PoB's global connection setting; enabling global noSSL also affects AI requests.
• Added
  The key appears only in Authorization: Bearer. The payload contains the question, recent history, and selected build context.
• Guard
  GetSanitizedConfig() replaces the key with [SET]/[EMPTY]. Debug logs contain sizes, timings, scopes, and counts, not the request body or key.
• Network
  The project has no backend, proxy, telemetry, or analytics. PoB connects directly to the user-configured endpoint.
• Release
  ai_config.json, logs, settings, and saves are excluded from Git and release archives.

Release and verification

• Package
  Portable Windows ZIP; extract it and run Path of Building.exe. V1 does not include an installer.
• Linux
  The same binary was validated under Wine. There is no native Linux build.
• CI
  387 upstream + AI integration tests passed on the release commit.
• Audit
  1,323 ZIP entries audited: no key, local config, log, personal save, duplicate path, unsafe path, or symlink.
• Integrity
  SHA256SUMS.txt is published with the archive for integrity verification.
• Changed
  manifest.xml and the updater point to kaiqueramos/PathOfAIBuilding, keeping the release channel separate from upstream.