Skip to content

CHANGELOG.md

Latest commit

 

History

History
321 lines (223 loc) · 36.5 KB

CHANGELOG.md

File metadata and controls

321 lines (223 loc) · 36.5 KB

Changelog

[Unreleased]

[1.10.0] — 2026-06-16

Skills

  • Skill 2 (Intent Detail Author) — sequential slot collection (blocking). Any intent with ≥2 caller-collectable slots now requires its validationPrompt to ask for them one at a time — one slot per numbered step in CollectionOrder, plus an IRON RULE forbidding bundled requests and requiring the bot to wait for each answer. Skill 2 will not mark the intent [detailed] until both hold. Slots populated from an upstream RT=2 response don't count toward the threshold; a single logical slot (e.g. full address) is one turn. Wired into §4.2 verify list, the Pattern V2 template, and the style-guide checklist.
  • Skill 2 (Intent Detail Author) — RT=2 live API verification (hard block, no waiver). Before authoring an RT=2 announcement, Skill 2 now curls the real endpoint with a user-supplied sample request. The intent cannot reach [detailed] unless the call returns HTTP 2xx AND every dotted path declared in 4.5.4 / referenced in the announcement is present in the live response JSON. Any failure (non-2xx, unreachable, unknown URL, missing path) blocks — there is no override. A redacted verification record (masked request, status, confirmed paths — never raw secrets/PII) is written to the new spec section 7.6. This replaces the prior "v1 trusts the declared shape" posture.
  • Skill 1 (Spec Designer): the §4.5.4 declared response shape is now documented as provisional pending Skill 2's live verification (was "v1 trusts the declared shape").
  • Skill 3 (JSON Assembler): new pre-flight Gate C — refuses to assemble any RT=2 intent lacking a section 7.6 verification record (backstop against hand-edited specs). Pre-flight is now three gates.

Documentation

  • New spec skeleton section 7.6 (RT=2 API verification log).
  • docs/skills/voicenter-bot-{intent-detail-author,spec-designer,json-assembler}/README.md mirrors updated.

Versioning

  • voicenter-bot-builder 1.9.1 → 1.10.0, marketplace metadata 1.9.1 → 1.10.0. voicenter-mcp and voicenter-api stay at 1.1.7 — neither plugin changed.

[1.9.1] — 2026-06-11

Skills

  • Skill 1 (Spec Designer): the Phase 1 Identifier is no longer prompted. It is now silently auto-derived from the Bot Name — ASCII names are snake_cased, non-ASCII (e.g. Hebrew) names are transliterated to Latin then snake_cased (יובל → yuval). Removes one AskUserQuestion from the interview. The separate Phase 3 intent-name reject-and-suggest prompt is unaffected.
  • Skill 3 (JSON Assembler) — token-budget gate raised. Check 8 (Compass rule 1) now goes advisory at 1,500–4,999 tok and blocking at ≥ 5,000 tok, with forced decomposition at ≥ 6,000 tok (was: advisory 1,500–2,499, blocking ≥ 2,500, decompose ≥ 4,000). This is a deliberate operator override: the Compass-measured degradation point (~2,500 tok) is unchanged and still surfaced as an advisory — only the pipeline's block threshold is relaxed to give authors working room. voice-prompt-doctrine.md §4 now separates the Compass measurement column from the Skill 3 enforcement column so the doc stays internally honest about the divergence.

Versioning

  • voicenter-bot-builder 1.9.0 → 1.9.1, marketplace metadata 1.9.0 → 1.9.1. voicenter-mcp and voicenter-api stay at 1.1.7 — neither plugin changed (versions are bumped per-plugin for what actually changed, not in blanket lockstep).

[1.9.0] — 2026-06-02

Skills

  • All 18 skills: added a language-mirroring directive — each skill now replies in the user's language (Hebrew→Hebrew, English→English) and follows mid-conversation switches. Affects conversational prose, questions, and AskUserQuestion labels only; emitted artifacts (identifiers, JSON keys, BCP-47 codes, API field names) are unchanged. The generated bot's runtime language behavior and the code-switch guardrail are not affected.
  • 4 interactive skills (voicenter-mcp setup, Skill 1 Spec Designer, Skill 2 Intent Detail Author, Skill 3 JSON Assembler): added a bilingual opening (Hebrew + English greeting on first contact). Skills 2, 3 and setup also gained a one-question-per-turn interview directive (single AskUserQuestion per message, never batch). Skill 1's existing §2.4 iron rule was upgraded with the same one-question-per-turn teeth rather than duplicated.

Versioning

  • voicenter-mcp 1.1.6 → 1.1.7, voicenter-api 1.1.6 → 1.1.7, voicenter-bot-builder 1.8.0 → 1.9.0, marketplace metadata 1.8.0 → 1.9.0.

[1.8.0] — 2026-06-01

Skills

  • Skill 1 (Spec Designer): new section-4 **Bot-intent role:** field (entry/global/chained, default chained). Approach-B role classification in close-out (propose from §2.4 opening targets + always-available intents, confirm in one batch, write explicit field). Section 6.2/6.4 now include auto global fan-out edges. Caller-silence (section 3) gains a structural silence failover intent (Skill 3 emits it as silence_behaviour.intent), defaulting to the transfer-to-human global; silence_ending_sentence then defaults to a transfer-to-representative line. Check 7 noted as auto-satisfied by fan-out when a global exists.
  • Skill 3 (JSON Assembler): parses the role field (§3.1). botIntents[] is now a selective registry — only entry (BotIntentTypeID 1) and global (2) intents; chained intents omitted (§4.3.3). 0-based SortOrder over the subset. Global fan-out (§4.3.4): an edge from every non-global intent to each global, deduped. §5 6.2 regeneration is fan-out-aware. Four new blocking cross-reference checks 11–14 (global-is-type-2, fan-out completeness, no-chained-in-botIntents, start-point-exists); the pass is now fourteen checks.
  • Silence-failover structural intent (design D8 revised). A production export (operator/משרד-התחבורה bot) showed the bot-level silence_behaviour carries a structural intent failover field (mirroring api_silence_behaviour.intent) — the initial "authored sentence, no structural field" decision was wrong. Skill 3 §4.2.5 now emits silence_behaviour.intent (resolved IntentId, first key, default transfer-to-human global, -999 sentinel if unknown), Doc 1 §6.B.3 documents it, and the invariant guard asserts both silence_behaviour.intent and the RT=2 api_silence_behaviour.intent/apiSilenceRelations pairing.

Documentation

  • Doc 1 (voicenter-bot-json-schema-audit-v1.md): §8.2 + G-10 rewritten — BotIntentTypeID is a discriminator (1=entry, 2=global); botIntents[] is a selective subset.
  • Doc 2 (voicenter-bot-skills-architecture-v1.md): botIntents emission note corrected.
  • validation-report.md §3.3 marked RESOLVED with the Brimag/Noa production evidence.
  • docs/skills/voicenter-bot-{spec-designer,json-assembler}/README.md mirrors updated.

Test artifacts

  • Golden outputs bot-ananit-2026-06-01.json and bot-noa-2026-06-01.json (hand-applied v1.8.0; Noa Hebrew re-authored from a lossy source — see banners).
  • validate-botintent-roles-v18.py invariant guard.

Plugin version bumps

  • marketplace.json metadata.version: 1.7.0 → 1.8.0
  • voicenter-bot-builder plugin: 1.7.0 → 1.8.0
  • voicenter-mcp and voicenter-api plugins: unchanged at 1.1.6

[1.7.0] — 2026-05-31

Changed (voicenter-bot-builder)

  • Skill 1 — Phase 1 interview: AI model config is no longer prompted. The interview previously asked the user to pick an AI model config via AskUserQuestion. It now silently defaults to the canonical model — Gemini Live (Voice driven 3.1) (AIModelConfigID=139, AIModelTypeId=18) — and writes it to spec section 1 without asking. A different model is used only if the user volunteers one (by catalog name or raw AIModelConfigID + AIModelTypeId). The <UNKNOWN: AI Model Config> deferral path is unchanged.
  • Skill 1 — Phase 1 interview: explicit agent-gender question added, with gender-filtered voice suggestions. Phase 1 now asks whether the agent should sound Female or Male (header "Agent voice") before the voice-name prompt, and offers only voices matching the chosen gender for the active model family. The skill is instructed to never infer gender from the bot name (unisex names previously caused male-only suggestions like Puck/Orus). Written to spec section 1 as **Agent Gender:** — selection-aid metadata only; not emitted to the JSON.
    • model-catalog.md: added a Gender column to both the Gemini and OpenAI voice tables (Gemini Female → Kore/Aoede/Leda/Zephyr, Male → Puck/Charon/Fenrir/Orus; OpenAI labelled per voice, alloy = Neutral).
    • spec-skeleton.md §1: added the Agent Gender field after Voice Name.
  • Skill 2 — Check 11 (RT=2 api_silence_behaviour completeness) now enforces the fallback intent. Check 11 previously claimed "six fields" but enumerated only the five silence_* fields, never the failover. It now requires all six components (3 language fields authored by Skill 2 + 3 structural fields owned by Skill 1: silence_duration, silence_loops, and the fallback intent) and halts/routes to Skill 1 patch mode if the fallback intent is missing or unresolved in section 4.
  • Skill 3 — RT=2 api_silence_behaviour.intent failover now explicitly specified and enforced. The inline failover pointer (Configuration.api_silence_behaviour.intent = resolved fallback IntentId) was never spelled out — Skill 3 only said "the six fields embedded inline," so the failover could be dropped or emitted as a string. Fixes:
    • New §4.4.1 documents the exact six-key api_silence_behaviour object; intent (resolved fallback IntentId, equal to apiSilenceRelations[].ApiSilenceIntentID) is marked mandatory — never omit, never emit as a string; -999 sentinel if the fallback intent is <UNKNOWN>.
    • Cross-reference Check 5 is now blocking on the inline intent being a present, non-null integer equal to ApiSilenceIntentID.
    • Cross-reference Check 6 deep-equality description now names all six api_silence_behaviour keys (including intent).

Documentation (voicenter-bot-builder)

  • All four SKILL.md changes above mirrored into the paired docs/skills/voicenter-bot-*/README.md.
  • voice-agent-llm v1.0.3 runtime behavior documented in all three bot-builder skills. No schema, validation rule, or plugin-version change — Skill 2's Check 10 still requires authored announcement for RT=2.
    • Empty announcement (or legacy apiResponseAnnouncement) at runtime is now substituted by the service with the sentinel [START THE CONVERSATION] as an LLM instruction — bot opens from persona; the literal string is not spoken aloud. Documented as a production safety net, not an authoring relaxation.
    • Voice-active text fields (validationPrompt, announcement, fail_output, function_output, post-execution intentInstructions) are now sanitized server-side before TTS. The existing Compass rule 8 authoring rule (write plain conversational prose; no Markdown/URLs in these fields) still applies.
    • Spec Designer SKILL.md: three remaining references to apiResponseAnnouncement updated to announcement (was apiResponseAnnouncement pre-v1.5.0).
  • Internal voice-agent service traceability (informational only, no skill change): Mastra library bumped 1.04 → 1.36.0; mastra-voicenter bumped 2.0.3 → 2.1.0.

Plugin version bumps

  • marketplace.json metadata.version: 1.6.0 → 1.7.0
  • voicenter-bot-builder plugin: 1.6.0 → 1.7.0
  • voicenter-mcp and voicenter-api plugins: unchanged at 1.1.6

[1.6.0] — 2026-05-24

Note: this release was developed under the working version name "v1.5.0" (visible in inline references throughout the changelog entry below and in the SKILL.md / docs README v1.5.0 correction notes). The release was renumbered to 1.6.0 to avoid collision with the prior 1.5.0 Compass doctrine release (May 14). Skill 3 SKILL.md and docs may continue to use "v1.5.0" as the working label for these changes; the platform-facing release version is 1.6.0.

Changed (voicenter-bot-builder)

  • BREAKING (wire format): Skill 3 emission restructured to match the production Voicenter export shape for Gemini 3.1 Voice driven bots. Bots emitted by pre-1.5.0 Skill 3 will import successfully but won't round-trip cleanly through the platform's export UI. Re-emit any in-progress bots after upgrading.
  • Skill 1 — Phase 1 interview: added three new bot-identity questions: Created by, Max call duration, Record agent calls.
  • Skill 1 — spec-skeleton.md §4: added optional Max turns / Max turns sentence per-intent fields.
  • Skill 2 — per-intent authoring: field renamed and shape changes for apiResponseAnnouncement (→ announcement), function_output (→ object { "default": ... }), response_success (→ object { "instructions": ... }).
  • Skill 3 — emission shape (largest change set):
    • Top-level wrapper field order: intentList moves to position #4.
    • AiModelConfig (top-level catalog reference) restructured: new fields (ApiKey, AIModel, IsActive, AccountId, ModifiedBy, CreatedDate, ModifiedDate, nested AIModelConfig carrying only the model string). Removed: Description, BaseUrl, Type, AIModelTypeId, full created payload (lives in the version-level only).
    • ActiveVersionInfo.AIModelConfig (version-level runtime config): added max_duration, recordAgentCalls. Removed: tools: [], instructions: "".
    • created payload reduced to lean shape: realtimeInputConfig.automaticActivityDetection.disabled: "true" + voice config only (version-level); model string only (top-level catalog reference). Dropped: temperature, topP, topK, responseModalities, proactivity, thinkingConfig, systemInstruction, tools.
    • intents[] entry: 17-field skeleton (intent-root IsActive and AccountId restored from production observation, removed incorrectly in the v1.4.1 schema correction). IsSilenceIntent now integer 0/1. IntentSources shape includes SourceName and IntentSourceID (was [{ SourceID: 1 }]). Optional IntentConfig.max_turns / max_turns_sentence per-intent.
    • IntentParameters[] entry: audit fields added (Schema: null, CreatedBy, ModifiedBy: " ", CreatedDate, ModifiedDate). Type fields now integers (IsRequired: 0/1). OptionList: null for non-ENUM (was []). DefaultValue: "" for unset strings (was null). ParameterType fully nested with frozen type-catalog metadata.
    • botIntents[]: BotId/IntentId lowercase-d casing. DTMFList: [] always. BotVersionId added. SortOrder 0-based. ConditionGroupList populated by default.
    • intentRelations[]: Order 0-based. IntentRelatedID is a unique row PK (placeholder range -2000+), no longer a NextIntentID mirror. ConditionGroupList populated by default.
    • intentCategories[]: no BotID. Added IsActive, AccountId, Description. PriorityId: 1 (was 2).
    • apiSilenceRelations[].Configuration: full mirror of parent IntentResponces.Configuration (was just the six silence_* fields).
    • RT=2 Configuration: announcement (was apiResponseAnnouncement). function_output → object { "default": ... }. response_success → object { "instructions": ... }. IntentLoadingAnnouncement (capital I) dropped.
    • RT=3 Configuration: response_success → object.
  • Skill 3 — §6.2 cross-reference checks: Check 6 now validates full Configuration deep equality between RT=2 intents and their apiSilenceRelations[] Configuration (was just silence_* six-field). Check 10 (Compass rule 12 model-config doctrine) inverted: instead of positively asserting present fields, now catches regressions to dropped fields under generationConfig.
  • Skill 3 — Appendix A: Quirk row 2 (the intentLoadingAnnouncement / IntentLoadingAnnouncement casing-bug pair) marked REMOVED in v1.5.0. Quirks 16–19 added (nested AIModelConfig, recordAgentCalls as string, realtimeInputConfig.automaticActivityDetection.disabled as string, IntentParameters[].ModifiedBy single-space literal).
  • Skill 3 — new ID placeholder ranges: IntentRelatedID (-2000+), IntentConditionGroupID (-3000+), IntentSourceID (-4000+).

Added (voicenter-bot-builder)

  • references/test-artifacts/test-prod-bot-transport-planner.json: production export of the user-supplied "סוכן תכנון מסלול - משרד התחבורה" v0.0.38 Gemini 3.1 Voice driven bot, serving as the third reference fixture and the ground-truth round-trip target for v1.5.0+ emission.

Documentation

  • references/docs/voicenter-bot-json-schema-audit-v1.md: ~12 subsections rewritten (§4, §5, §6.A, §6.B, §6.B.2, §8.2, §8.3, §8.4, §8.6, §9.0, §9.1, §10, §11.2, §11.3, §11.5, §16). The doc remains the canonical wire-format contract; this update aligns it to the production Gemini 3.1 Voice driven export.
  • docs/skills/voicenter-bot-*/README.md: all three plugin docs READMEs mirror the corresponding SKILL.md changes.

Internal

  • references/test-artifacts/test-emitted-json-yuval.json and test-emitted-json-refua.json regenerated to v1.5.0 shape. Placeholder IDs preserved.

Plugin version bumps

  • marketplace.json metadata.version: 1.5.0 → 1.6.0 (renumbered to avoid collision with the May 14 1.5.0 Compass doctrine release)
  • voicenter-bot-builder plugin: 1.3.0 → 1.6.0 (production wire-format alignment)
  • voicenter-mcp and voicenter-api plugins: unchanged at 1.1.6

[1.5.0] - 2026-05-14

Added

  • Compass doctrine integration in voicenter-bot-builder. New shared reference plugins/voicenter-bot-builder/references/voice-prompt-doctrine.md distilled from the Gemini Live 3.1 voice agent engineering guideline. The reference catalogs 13 enforceable rules and is loaded by all three skills.
    • Skill 1 (Agent Spec Designer) gains 5 new self-validation checks (11–15) covering rules 3 (English operational), 4 (intent description in English), 5 (recency-slot language-lock guardrail), 6 (contradictory pacing/length), 7 (generic-policy boilerplate), plus a rule-11 mirror on rewritten fields. New Appendix D documents the mapping.
    • Skill 2 (Intent Detail Author) gains 4 new per-intent iron rules covering rules 8 (TTS-safe formatting; blocking on markdown/URLs), 9 (date math in prompt), 10 (few-shot example cap), 11 (Hebrew-utterance isolation; blocking). The conversation-routines-style-guide.md gets a TTS-safety addendum.
    • Skill 3 (JSON Assembler) gains 3 new cross-reference checks (8, 9, 10) plus a DOCTRINE SENTINELS banner section (rule 13). Check 8 (token budget) is advisory at 1,500–2,499 tok and blocking at ≥ 2,500. Check 10 (model-config doctrine) is blocking on any mismatch. All three are gated on AiModelConfig.created.model = "models/gemini-3.1-flash-live-preview".
    • Token-counting uses a char-based estimate (Latin 1/4 tok, Hebrew/Arabic/CJK 1/1.5 tok) — ±15% accuracy, sufficient for the doctrine thresholds.

Notes

  • Rules 1, 2, 12 apply only when AIModelConfigID=139 (Gemini Live 3.1). Rules 3, 4, 5, 6, 8, 10, 11 apply to any active voice channel. Rules 7, 9, 13 apply universally.
  • Non-goal: this release does not retroactively fix existing bot artifacts (e.g., bot-noa-2026-05-12.json). To apply the doctrine to an existing bot, re-run Skill 1 → 2 → 3 on its spec.
  • Non-goal: platform-side concerns the Compass flags as out-of-prompt (PII redaction, prompt-injection classifier, rate limiting, recording consent) are not enforced by the bot-builder. The doctrine reference §3 notes which concerns belong to which plane.

Plugin version bumps

  • marketplace.json metadata: 1.4.2 → 1.5.0
  • voicenter-bot-builder plugin: 1.2.2 → 1.3.0 (Compass doctrine integration across all three skills)
  • voicenter-mcp plugin: 1.1.6 (unchanged)
  • voicenter-api plugin: 1.1.6 (unchanged)

[1.4.2] - 2026-05-11

Changed

Cache-refresh bump across all three plugins to force /reload-plugins to resync SKILL.md content on existing installs. No behavior or surface-area change since 1.4.1.

Plugin version bumps (lockstep per CLAUDE.md)

  • marketplace.json metadata: 1.4.1 → 1.4.2
  • voicenter-mcp plugin: 1.1.5 → 1.1.6 (no content change; bumped for cache refresh)
  • voicenter-api plugin: 1.1.5 → 1.1.6 (no content change; bumped for cache refresh)
  • voicenter-bot-builder plugin: 1.2.1 → 1.2.2 (no content change; bumped for cache refresh)

[1.4.1] - 2026-05-11

Fixed (voicenter-bot-builder 1.2.0 → 1.2.1) — Skill 3 IntentResponces.IsActive structural correction

Skill 3 (voicenter-bot-json-assembler) now emits the per-intent active flag inside IntentResponces (as the middle key between ResponseTypeId and Configuration) and no longer emits IsActive or IsDeleted at the intent root. The corrected shape matches the platform-validated bot JSON (docs/json-bag/good.json intent -10). The ImportBotFromJSON procedure reads IntentResponces.IsActive for the per-intent active flag; the prior intent-root location was silently ignored, so the bot's runtime active state was unchanged by the fix — this is a wire-format correctness fix, not a behavior change.

  • SKILL.md §4.3.1 — removed the two intent-root rows (IsActive: 1, IsDeleted: 0); added an inline note pointing readers to §4.4 for the corrected location.
  • SKILL.md §4.4 — added IsActive: 1 row to all four RT-specific tables (RT=1, RT=2, RT=3, RT=4) immediately below the ResponseTypeId row. Added an invariant-shape header note documenting that every IntentResponces has the same three-key outer shape regardless of RT.
  • SKILL.md Appendix A — added quirk #15 (IntentResponces.IsActive: 1 emission rule + anti-quirk note explicitly forbidding intent-root IsActive/IsDeleted). Preamble updated from "14 quirks" to "15 quirks". Skill 3's §4.5 quirk-preservation verification pass now covers the new quirk.
  • Companion docs (docs/skills/voicenter-bot-json-assembler/README.md) — per-RT keys preamble and the quirk-preservation walk paragraph mirror the SKILL.md changes. (Drive-by fix: IntentScripts: {} corrected to IntentScripts: [] to match the v1.2.1 SKILL.md Appendix A quirk #8 amendment.)
  • Schema audit (references/docs/voicenter-bot-json-schema-audit-v1.md) — §9.0 renamed "16-Field Skeleton" → "14-Field Skeleton" (intent-root IsActive/IsDeleted rows removed); §9.2 IntentResponces tree updated from two fields to three with IsActive between ResponseTypeId and Configuration. Inline "Schema correction (2026-05-11)" addenda explain the rationale.

Test artifacts

references/test-artifacts/test-emitted-json-{yuval,refua}.json predate this fix and may show the pre-v1.4.1 shape. Regeneration is deferred — these files are reference samples, not consumed by any runtime. The next genuine Skill 3 invocation against either spec will produce the corrected shape.

Plugin version bumps (lockstep per CLAUDE.md)

  • marketplace.json metadata: 1.4.0 → 1.4.1
  • voicenter-mcp plugin: 1.1.4 → 1.1.5 (no content change; bumped for cache refresh)
  • voicenter-api plugin: 1.1.4 → 1.1.5 (no content change; bumped for cache refresh)
  • voicenter-bot-builder plugin: 1.2.0 → 1.2.1 (Skill 3 wire-format correction)

[1.4.0] - 2026-05-03

Added (voicenter-bot-builder 1.1.0 → 1.2.0) — Skill 1 intent flow diagram + refinement loop

Skill 1 now generates a Mermaid flowchart TD of the bot's intent graph as the final structural artifact, embedded in the spec under new section 6.6, and offers a refinement loop before final emission. Same diagram regenerates after every patch, so the user can see the structural impact visually before finalizing.

  • Mermaid diagram (spec section 6.6) — Skill 1 §3.6.1. One node per intent in section 4 (label: <identifier><br/>RT=<n> · slots: <count>, plus ⚑ if hard-intent). Node shapes encode response type: stadium for RT=1 transfer, rounded rectangle for RT=2 API, default rectangle for RT=3 conversational, subroutine shape for RT=4 outbound dial. One labeled edge per transition (success / fallback / escalation). If section 4.7 (advanced overrides) declares dtmf_list: for a transition, digits append to the edge label. Skill 3 ignores section 6.6 — it's for human comprehension only, not the import contract.
  • Refinement loop at greenfield close-out — Skill 1 §3.6 step 5. After section 6 + 6.6 are generated and soft-cap warnings surface, Skill 1 renders the diagram and prompts via AskUserQuestion (header: "Diagram review", 4 options: "Looks good — finalize (Recommended)" / "Adjust an intent" / "Adjust a transition" / "Adjust persona / opening behavior"). Any "Adjust" pick routes back to the relevant phase, applies the change, regenerates section 6 (including 6.6), re-runs the self-validation checklist, and re-prompts. Capped at 5 iterations to prevent endless cycles — beyond 5, Skill 1 logs the iteration count to section 7.3 and proceeds.
  • Patch-mode regeneration — Skill 1 §4.6 + §4.7. Section 6.6 regenerates after every applied patch, alongside the cascade summary, and the same refinement loop is offered before final emission.

Changed

  • Skill 1 output contract updated to list section 6.6 as a greenfield/patch artifact (and to clarify it's not consumed by Skill 3 or the import proc).
  • Docs lockstep: docs/skills/voicenter-bot-spec-designer/README.md mirrors the diagram + refinement-loop additions, with a new "Intent flow diagram + refinement loop" section under Output contract.

Plugin version bumps (lockstep per CLAUDE.md)

  • marketplace.json metadata: 1.3.0 → 1.4.0
  • voicenter-mcp plugin: 1.1.3 → 1.1.4 (no content change; bumped for cache refresh)
  • voicenter-api plugin: 1.1.3 → 1.1.4 (no content change; bumped for cache refresh)
  • voicenter-bot-builder plugin: 1.1.0 → 1.2.0 (Skill 1 diagram + refinement loop)

[1.3.0] - 2026-05-03

Added (voicenter-bot-builder 1.0.1 → 1.1.0) — Skill 1 interactive UX + optional advanced features

Skill 1 (voicenter-bot-spec-designer) now uses live MCP lookup for Voicenter platform resources and AskUserQuestion (interactive menu inputs) for every closed-set choice in the interview, instead of free-text capture. Skill 1 also gains an opt-in path for the two runtime-supported features (ConditionGroupList, DTMFList) that were previously inaccessible from the build pipeline.

  • Live resource lookup via voicenter-mcp.list_resources (recommended default). Customer Account ID (Phase 1) and RT=1 Layer ID (Phase 4) are now fetched live with entityFilter: ["Accounts"] / ["Layers"] and presented as id+name tables, then prompted via AskUserQuestion. New SKILL.md §2.4.A documents a 3-tier fallback that is never silently skipped: (1) plugin not installed → offer install + auth via AskUserQuestion; (2) plugin installed but unauthenticated → offer authenticate via AskUserQuestion; (3) user declines or retry fails → fall back to text-only mode and <UNKNOWN: …> markers, logged once to spec section 7.3 with the reason; the user is not re-prompted in the same session.
  • AskUserQuestion for every closed-set choice (SKILL.md §2.4.B). New iron rule: if the user can answer with one of a fixed set of strings, route through interactive inputs. Covers runtime/mode detection, channel scope, voice/model catalog picks, caller-silence yes/no, identifier ASCII confirmation, every Phase 2 "Accept draft / Edit" prompt, Deep Research pause/skip, Response Type (RT=1/2/3/4), per-slot ParameterTypeId + IsRequired, RT=2 Method (POST/GET) + fallback intent reference (from existing intent set), RT=4 dial source + record + rarity-warning confirmation, account / layer selection from live MCP lists, patch-mode cascade confirm, every self-validation iron-rule re-prompt, and the new MCP install/auth/skip prompts. Free-text capture is reserved for genuinely open-ended fields (names, descriptions, free-form text content, integer/numeric values).
  • Optional advanced features (default: skip — not required) — new SKILL.md §3.5.5 adds an opt-in capture path for ConditionGroupList (conditional branching on BotIntent / IntentRelated) and DTMFList (DTMF keypad routing). After Phase 4 captures the structural intent set, Skill 1 prompts once via AskUserQuestion with "Skip — accept defaults (Recommended)" as the default. Skip path writes nothing; Skill 3 falls back to existing safe defaults (ConditionGroupList: [], DTMFList omitted), and the ImportBotFromJSON proc skips both arrays cleanly via NULL-guards in CreateConditionGroups and the IntentRelatedDTMF insert. Opt-in path captures into a new freeform spec section 4.7 Advanced overrides; Skill 3 (§4.3.3 / §4.3.4) lifts condition_groups: and dtmf_list: blocks verbatim into the corresponding botIntents[] / intentRelations[] entries. Skill 1 does not validate §4.7 contents — pass-through to Skill 3.
  • RT=3 schema cross-reference clarification — Skill 1 §3.4.3 RT prompt now includes a parenthetical noting the DB seed name for ResponseTypeId=3 is "Message" / "Update Bot Configuration" but the operational use is conversational data-collection. Cosmetic only; no behavior change.

Changed

  • Skill 1 anti-list updated: live MCP lookup is now in scope (was previously listed as out-of-scope with the model catalog); ConditionGroupList / DTMFList are documented as opt-in only via §4.7.
  • Skill 3 §4.3.3 + §4.3.4 (botIntents[] and intentRelations[]): ConditionGroupList and DTMFList rows now read from spec §4.7 if present, fall back to the existing default-skip behavior if absent.
  • Docs lockstep: docs/skills/voicenter-bot-spec-designer/README.md and docs/skills/voicenter-bot-json-assembler/README.md updated with the new tool conventions, the §3.5.5 opt-in summary, and the §4.7 pass-through behavior.

Plugin version bumps (lockstep per CLAUDE.md)

  • marketplace.json metadata: 1.2.1 → 1.3.0
  • voicenter-mcp plugin: 1.1.2 → 1.1.3 (no content change; bumped for cache refresh)
  • voicenter-api plugin: 1.1.2 → 1.1.3 (no content change; bumped for cache refresh)
  • voicenter-bot-builder plugin: 1.0.1 → 1.1.0 (Skill 1 interactive UX + optional §3.5.5; Skill 3 §4.7 pass-through)

[1.2.1] - 2026-05-03

Fixed (voicenter-bot-builder 1.0.0 → 1.0.1) — Skill 3 alignment with ImportBotFromJSON stored procedure

The wire-format JSON Skill 3 emits is now consumable by the platform's ImportBotFromJSON MySQL procedure without manual editing. Five hard-blocking and one fragile gap closed.

  • G1 — AiModelConfig.AccountId: 0 added to top-level AiModelConfig. Routes the procedure to its "reuse existing default config" branch instead of falling through to an INSERT that fails on AIModel and AIModelConfig NOT NULL columns. (Skill 3 §4.2.3.)
  • G2 — intentCategories[].PriorityId: 2 (Medium) emitted explicitly. Was previously absent; column is TINYINT NOT NULL and the proc passes the extracted value, so omission caused a NULL INSERT failure. (Skill 3 §4.3.5.)
  • G3 — botIntents[].IntentId / BotIntentId lowercase d (was capital ID). MySQL JSON paths are case-sensitive; the proc reads $.IntentId and the prior emission resolved NULL, breaking the BotIntent INSERT. Capital ID is preserved on intentRelations[] (matches the proc's read there) — deliberate asymmetry. (Skill 3 §4.3.3.)
  • G4 — botIntents[].SortOrder added (1-based ordinal). Required NOT NULL column previously omitted. (Skill 3 §4.3.3.)
  • G5 — intentRelations[] deduplication by (OriginIntentID, NextIntentID). The DB unique key forbids duplicates; previously, a spec listing the same target twice (e.g., success path AND fallback both → transfer_to_human) emitted two rows and broke the second INSERT. Skill 3 now keeps the lowest-Order survivor and notes the collapse in the banner. (Skill 3 §4.3.4.)
  • G6 — IntentScripts: [] (was {}). The proc iterates with JSON_LENGTH + integer indexing; the object form would index [0] on a populated {} and break. Doc 1 §16 quirk #8 amended. (Skill 3 §4.3.1, Appendix A row 8.)
  • G7 — IntentSources per intent, derived from spec section 1 Channels Active mapped through the DB Sources static table (1=VOICE, 2=CHAT, 3=WEB). (Skill 3 §4.3.1.)

Changed

  • model-catalog.md populated with seven real default AIModelConfig rows (AccountId=0) drawn from database/Tables/StaticData/AIModelConfig.Data.sql: Gemini Live (139/18), Gemini 2.5 (52/10), Gemini Voice Driven (136/16), Gemini 3.1 LLM Driven (142/21), GPT-4 Realtime (1/1), GPT-5 Realtime (91/13), GPT Realtime Mini (132/15). Replaces the prior <TODO> placeholders.
  • Voice catalog expanded from the 2-row Puck/Orus list to the full provider inventories — 10 OpenAI voices (Alloy/Ash/Ballad/Coral/Echo/Sage/Shimmer/Verse/Cedar/Marin) and 8 Gemini voices (Puck/Charon/Kore/Fenrir/Aoede/Leda/Orus/Zephyr).
  • Skill 3 Appendix D — Static reference data added as the single source of truth for every static integer ID Skill 3 emits (BotStatusId, BotVersionStatusId, BotIntentTypeID, IntentCategoryId/PriorityId, ResponseTypeId, SourceID, ParameterTypeId, IntentRelatedTypeID, IntentScriptType, default AIModelConfig rows). Mirrors database/Tables/StaticData/*.Data.sql; must be re-verified when those files change.

Plugin version bumps (lockstep per CLAUDE.md)

  • marketplace.json metadata: 1.2.0 → 1.2.1
  • voicenter-mcp plugin: 1.1.1 → 1.1.2 (no content change; bumped for cache refresh)
  • voicenter-api plugin: 1.1.1 → 1.1.2 (no content change; bumped for cache refresh)
  • voicenter-bot-builder plugin: 1.0.0 → 1.0.1 (Skill 3 + model-catalog content changes)

[1.2.0] - 2026-05-02

Added

  • voicenter-bot-builder — new third plugin (v1.0.0) that ships a 3-skill bot-authoring pipeline:
    • voicenter-bot-spec-designer (Skill 1) — interview-driven structural design; produces agent-spec.md
    • voicenter-bot-intent-detail-author (Skill 2) — per-intent language content (Conversation Routines style)
    • voicenter-bot-json-assembler (Skill 3) — mechanical projection to Bot JSON wire format with §15.4 cross-reference pass and fail-loud sentinels
  • docs/plugins/voicenter-bot-builder.md and per-skill long-form references under docs/skills/voicenter-bot-*/
  • "Bot authoring (build-time)" entry in docs/architecture.md taxonomy + dedicated build-time pipeline section

Fixed (Skill suite v1 patches surfaced by Conv 6 end-to-end test)

  • Patch 1 — Identifier field. Added **Identifier:** to spec section 1 so Skill 3 produces useful filenames for non-ASCII bot names. Pre-fix: Hebrew bot names produced bot-bot-<date>.json. Post-fix: bot-yuval-<date>.json / bot-refua-<date>.json.
  • Patch 2 — RT-specific bold sub-labels. spec-skeleton.md formalized section 4 RT-specific sub-labels (**URL:**, **Method:**, **Headers:**, **Body:**, **API silence behavior:**, **Layer:**); Skill 3 §3.1 strict-template parser enumeration extended; Skill 3 §3.3 deviation table added.
  • RT=4 production-shape rewrite. spec-skeleton.md, Skill 1 §3.5.1, Skill 3 §3.1, and Skill 3 §4.4 RT=4 emission table updated to match real production Configuration shape — dual modes (parameter / static), three phone slots, selectdial_option, response_success.instructions, optional announcement / loading announcement / post-execution instructions.

[1.1.1] - 2026-04-26

Fixed

  • Bump to force plugin cache refresh — 1.1.0 update was not re-syncing SKILL.md files

[1.1.0] - 2026-04-26

Fixed

  • Skills now register correctly on /reload-plugins (was reporting 0 skills loaded)
    • Added explicit name: field to all 15 SKILL.md frontmatter entries
    • Removed redundant "skills": "./skills/" from plugin.json (default discovery handles it)
  • voicenter-mcp MCP server config now includes required "type": "http" field
  • Optimized all 14 SKILL.md files for clearer Claude Code invocation

Changed

  • Conformed plugin.json files to the official Claude Code plugin manifest schema
  • Removed unsupported icon field from plugin and marketplace configs
  • Removed nested V2 marketplace duplicate

Added

  • LICENSE file (MIT)
  • CHANGELOG.md
  • .gitignore for local Claude settings

[1.0.0] - 2025-04-04

Added

  • Initial marketplace release with 2 plugins
  • voicenter-mcp — Live API access via OAuth MCP server at mcp01.voicenter.co
  • voicenter-api — 14 API integration skills:
    • Push APIs: VoiceBot, Pop-Up Screen, CDR Notification, External Layer
    • Outgoing APIs: Click2Call, Call Log, Blacklist, Mute Recording, Extension List, Real-Time, Productive Dialer, Login/Logout, Lead Tracker, Active Calls