Spark

SparkLang roadmap

Prioritized against the adoption bar. Status: done / next / won't (near-term product focus). Details: ADOPTION_BAR.md.

Done (this era)

  • Public MIT repo, SparkLang branding, About disambiguation
  • Changelog + release process doc
  • embed / retrieve first-class ops + dry fixtures + spark-rag-http
  • Dry-run escape hatch: shell / run allowlist fixtures
  • Live --allow-shell: ./spark-shell argv execve (echo|true|false)
  • Host embed: Python + JS (js/sparklang) + C (host/c/sparklang.h); ./spark --embed JSON handshake
  • Per-ask [accounting] line (dry zeros; live wall-clock + usage)
  • Live run-level [accounting-run] rollup (spark-ask-http --rollup)
  • Streaming ask: ./spark-ask-http --stream SSE + ask stream language form (live); dry prints stream=1
  • Live extract: schema validate + retry-on-miss (--retries, --stub-file offline); forks ./spark-ask-http
  • Positioning: not tied to a single AI gateway; dry-run first
  • http get / http post + timeout + dry fixture files + live ./spark-http with bearer / header + retries / backoff
  • SPARK_BC Builder factory (init): --compile Spark → .sparkbc with TRAIN/STEP/TRAIN_STATUS; bootstrap --run-bc dry; Spark-created init safetensors; published dumps + sha256. See SPARK_BC Builder. Dry ≠ trained.

Next

Track What
Extract done (live + --retries / stub)
Ask done (--stream / ask stream)
Eval done — expect equal / contains; make test-expect
Escape done — live --allow-shell; JS / C host FFI
Accounting done — live wall-clock + run-level rollup
Builder STEP→weights done (dry weights.safetensors); later owner train/eval growth — not claimed today
LSP done — tools/spark_lsp + extension v0.2; LSP.md
Receptionist Dry path done (examples/receptionist.spark + expect); live transfer/hold/hangup still goal
Releases Cut GitHub Release tags from CHANGELOG (see RELEASE.md); Pages = human CF dashboard (no wrangler)

Done (rename)

Track What
Rename Public sparklang-dev/sparklang live. Site sparklang.dev. CLI remains spark.

Won't (near-term focus)

Drop Why
MITM / pcap / browser automation as hero Distracts from AI workflow language
Homegrown Electron/PyQt IDE as hero chrome Prefer LSP for editor story; IDE tree kept (hard-delete plan revoked 2026-09-02)
Model-build as training product Blueprint / eval sugar only
Production voice/telephony claim Gaps remain (barge-in, EOU, SM, adapters) — gated demo only
Single-gateway-as-identity Gateway optional; any OpenAI-compatible URL

Voice gaps

If not shipping soon, keep voice off the production claim list:

  • Barge-in, EOU / silence detection, latency budget
  • Call state machine: transfer / hold / hangup / VM / DTMF / STT retry
  • Named STT / TTS / telephony adapters as stable contracts
  • Concurrency model for multi-call

Dry listen / speak / gated PSTN remain demos. See VOICE.md.