Spaces:
Running
Running
Commit History
-
fix(postmeet): overwrite declared data slots in populated examples
k0880kf223cdc -
fix(postmeet): hydrate every declared source field from template schema
k0880k0fc3cae -
fix(postmeet): retain YouTube source fields for template hydration
k0880kfd79ed3 -
feat(postmeet): add cheap GPU resident contract probe
k0880k01a78ab -
fix(postmeet): speak current Gradio v2 GPU protocol
k0880k3e4f662 -
fix(postmeet): target the real signalmesh-gpu Space
k0880k80db70d -
fix(postmeet): restore checkpointed GPU-first ASR
k0880kd94bb73 -
fix(postmeet): restore private signalmesh_gpu client
k0880kf44d560 -
fix(postmeet): restore semantic multi-source media resolver
k0880k5881a55
-
feat(media): resolve arbitrary URLs into semantic media objects
k0880kd38b3da
-
fix(bighelp): return valid worker HTTP envelopes
k0880ka087d1d -
feat(bighelp): add direct authenticated runtime worker
k0880k43b3295 -
fix(postmeet): ship the ASR engine, keep orphan motions, unbreak /openapi.json
ffmpeg + faster-whisper in the image A Space builds this Dockerfile — python:3.11-slim plus LaTeX — so being on HuggingFace supplies the hardware and the model hub, not the binaries. Without these, capability.health reported "unavailable" and every import was refused. faster-whisper runs the same weights on CTranslate2: no torch, int8 on CPU, measured 18.3s against the reference CLI's 59.1s on the same five minutes. Weights are baked at build so the first import does not wait on a download, and HF_HOME points somewhere a non-root Space user can actually write. Default model is small: base transcribes the vote but garbles who moved what. Orphan motions were being dropped find_outcomes only attached a motion to an agenda item that already existed, so a motion made in a meeting where no agenda heading was ever detected vanished — the record showed nothing happened, which is the one thing it must never do. Motions now stand as their own item when nothing owns them,…5bc60a6 -
feat(capability): typed .fq driver contract, Studio adapter, evidence-based closure
Wraps PostMeet as a resident capability and gives Studio a generic way to operate any typed .fq driver pack, with closure state computed from evidence rather than asserted by whoever wrote the document. Capability contract (core/capability_fq.py) Seven roles carried by a typed .fq: manifest, health, inventory, execute, status, evidence, receipt. Registration refuses a pack that advertises a role with no runtime behind it, and health is graded by the contract from reported checks — a driver cannot mark itself healthy. Studio adapter (core/studio_fq.py, /studio/fq) Renders any capability from describe() alone: identity, health, closure, actions, live jobs, evidence, receipts. Job phases and progress fields come from the manifest, so no product-specific branch exists — asserted by a test that greps the source for product names. Accepted is never success; only capability.receipt can report verified. Driver auth renders as status only, never as a secret. Promotion aut…249db15 -
feat(meeting-lens): preset, stage runner, brand extract, validate, hydrate
- meeting_lens_full preset registered in presets.manifest.json - 4 tools: brand_extract, meeting_validate, meeting_hydrate, meeting_stage_runner - 2 schemas: meeting_lens_payload, sample_payload - 8 ML specialist agents (ML1-ML8) appended to matrix_agents.json - docs: MEETING_LENS_PLAN.md, MEETING_LENS_BAKE_IN.md9e527b3
-
feat: export full git history for maintenance bootstrap
k0880k91b4eb6 -
feat(fq): add local agents in/out CLI
k0880k98ed28c
-
generalize nexus_ingest into site_ingest — nothing here was nexus-specific
Only the URL list and the demo questions were about Nexus; the fetch, chunk, broadcast and ask path works for any server-rendered site. Takes a namespace plus URLs and derives frequency names from the paths: python3 tools/site_ingest.py acme https://acme.com/ https://acme.com/docs -> acme.index, acme.docs The nexus set stays as the default demo. Kept as a script rather than reusing what's in the mesh: source_hydrator targets data URLs (it rewrites Sheets to CSV) and has no add route, feed_hydrator is RSS-only, and MPort's _snapshot_site mirrors a site instead of broadcasting it as signals. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019afygHW8NAJBJsaU5Kz3Yo366c409 -
nexus demo: chunk pages so the whole doc is reachable, and quote the hit
_do_tune_in renders each signal at [:400], so a 4.9k-char page exposed only its first 400 chars — "what is the data layer" missed because Data Layer sits deeper in the page. Pages now ship as ~360-char signals on one frequency (nothing gets truncated away), and ask() quotes the frequency whose text actually mentions the question instead of printing the head of the blob. All four demo questions now answer from the docs, including auth (OAuth 2.0, API keys, SSO) and the data layer. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019afygHW8NAJBJsaU5Kz3Yo37fa732 -
tune_in: treat '.' as a namespace separator
_tokens split on _, - and whitespace, so a dotted frequency name was a single token and only an exact-string keyword could ever tune to it. 75 of the 88 live frequencies are dotted — the 71 nicoli_harness.* tools among them were unreachable by keyword: "ollama" did not match nicoli_harness.ollama. Splitting on '.' as well makes the namespace addressable at every level while the subset/coverage rule keeps it from over-matching (hf_blog still does not answer for aws_blog). nexus_ingest.py moves to the keyless /ui/broadcast and /ui/tune_in path, so the demo needs no SignalMesh key at all. Verified: 8/8 match cases, tests/ 30 passed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019afygHW8NAJBJsaU5Kz3Yod1f500a -
fix(chat): only force code output when the turn asks for a build
The native-chat system prompt applied its code-first rules unconditionally, so any message — "test", a greeting, a question — had to produce code. The model grabbed the nearest topic in history and wrote a file for it. detect_completion_gaps then read the conversational-reply-without-code as incomplete and the Seek chain pushed up to 4 more rounds of it. is_build_request() gates both the prompt block and Seek. Non-build turns get rules that forbid unsolicited code and forbid resuming a previous coding task. Adds tools/nexus_ingest.py: ingests nexusaifirst.com public docs into nexus.* frequencies and answers questions from them in one hop (`ask` subcommand). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019afygHW8NAJBJsaU5Kz3Yo59a2511
-
Deploy HF runtime overlay (PR#11 @1a2ca3c): MPort, /api/submesh/*, hydration, Nicoli, security, persistence
Verified signature34c6191
-
feat: reposition as dev-environment hub; sync all gap-closing components
- Replace chat-UI landing with developer reference page (quickstart, API reference, AGENTS.md drop-in guide, cost model, live mesh status, tab code snippets in curl/python/js, copy buttons) - Sync all components from GitHub: sec_omega, manifest_parser, manifest_gc, agentified_tool (dual-import fix), autonomous_pumps (all 19 adys dirs + AGENTS.md scrape), matrix_agents (think() override), signalmesh_api (startup wiring + manifest endpoints + guarded_broadcast) - Add AGENTS.md node manifest standard - Add tests/ with 26 passing tests Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>78e921e
-
Initial Release: Signal Mesh Spatially Indexed Orchestrator
dba8b8e