mit einem Klick
murmur
murmur enthält 33 gesammelte Skills von murmur-io, mit Repository-Berufsabdeckung und Skill-Detailseiten auf SkillsMP.
Skills in diesem Repository
Run a change through the full Murmur harness — isolated worktree, writer, deterministic checks, independent adversarial + risk reviews, hash-bound PASS attestation, guarded commit. Use PROACTIVELY when a change is risky, multi-step, or you want the earned safety net (lock/crypto/egress/protocol, or anything you want independently verified). Skip it for docs, chores, and low-risk edits — those commit normally.
Cut a signed, notarized macOS release of Murmur (Tauri 2 + Angular 22). The exact, proven step-by-step runbook — preflight gates → version bump (+ Cargo.lock sync) → QueaT commit → PR-merge to the `murmur` trunk (never direct-push) → rustup targets → stop dev → universal build → Developer-ID sign BY IDENTITY HASH (the Polish-ń cert gotcha) → DMG → notarize → staple/spctl → gh release create + upload. Use whenever the user wants to ship, release, cut a version, build a distributable .app/.dmg, sign/notarize, or publish a GitHub release of Murmur. Supersedes the stale docs/RELEASE-CHECKLIST.md.
Cut a signed, notarized macOS release of Murmur (Tauri 2 + Angular 22). The exact, proven step-by-step runbook — preflight gates → version bump (+ Cargo.lock sync) → QueaT commit → PR-merge to the `murmur` trunk (never direct-push) → rustup targets → stop dev → universal build → Developer-ID sign BY IDENTITY HASH (the Polish-ń cert gotcha) → DMG → notarize → staple/spctl → gh release create + upload. Use whenever the user wants to ship, release, cut a version, build a distributable .app/.dmg, sign/notarize, or publish a GitHub release of Murmur. Supersedes the stale docs/RELEASE-CHECKLIST.md.
The agentic feature-shipping discipline for Murmur (Tauri 2.11 Rust core + Angular 22 zoneless). Scope → (Workflow) plan → build backend and/or FE under the project conventions → ADVERSARIAL verify (an adversarial-verifier pass, plus a lock-security review when the change touches the lock/crypto/visibility path) → gates green (cargo test --lib, ng lint, ng build, scripts/ci.sh) → QueaT commit → PR to the `murmur` trunk. Use whenever the user wants to build, implement, add, or ship a feature/fix in Murmur end-to-end. Encodes the verify-before-trust discipline that caught 7 bugs.
Run and iterate Murmur locally (Tauri 2.11 + Angular 22 zoneless). The exact dev-run recipe — MURMUR_DEV_DEK/MURMUR_DEV_KEK to avoid keychain prompts, source ~/.cargo/env, ng on :1420 + MCP on :8765, the `cargo test --lib`-not-`clippy --all-targets` inner loop, owner-aware clean relaunch, and reading the boot/abort log. Use whenever the user wants to start/run/serve Murmur in dev, debug a launch/abort, relaunch cleanly, or run the Rust test loop.
The agentic feature-shipping discipline for Murmur (Tauri 2.11 Rust core + Angular 22 zoneless). Scope → (Workflow) plan → build backend and/or FE under the project conventions → ADVERSARIAL verify (an adversarial-verifier pass, plus a lock-security review when the change touches the lock/crypto/visibility path) → gates green (cargo test --lib, ng lint, ng build, scripts/ci.sh) → QueaT commit → PR to the `murmur` trunk. Use whenever the user wants to build, implement, add, or ship a feature/fix in Murmur end-to-end. Encodes the verify-before-trust discipline that caught 7 bugs.
Run and iterate Murmur locally (Tauri 2.11 + Angular 22 zoneless). The exact dev-run recipe — MURMUR_DEV_DEK/MURMUR_DEV_KEK to avoid keychain prompts, source ~/.cargo/env, ng on :1420 + MCP on :8765, the `cargo test --lib`-not-`clippy --all-targets` inner loop, owner-aware clean relaunch, and reading the boot/abort log. Use whenever the user wants to start/run/serve Murmur in dev, debug a launch/abort, relaunch cleanly, or run the Rust test loop.
Murmur app-wide performance map + a runnable measurement recipe across the Rust->IPC->Angular seam — FE change-detection anti-patterns, Tauri event-vs-Channel + payload windowing, OOM/RAM-pressure guards, deferred startup, the missing [profile.release], and the profiling toolchain. Use when the user reports jank/lag/heat/OOM, before a release, when adding a feature that touches a hot path, or on any 'why is X slow'. Not for model tokens/sec or quantization (that is ondevice-model-perf).
Run, understand, and debug Murmur's CI — the local `scripts/ci.sh` gate and the GitHub Actions macOS PR-gate that wraps it. The map of every gate step (remote enforcement → control-plane audits/evals → swiftc → clippy → cargo test → cargo audit → cargo deny → cargo build → ng lint → ng build → E2E), how to reproduce a red CI run locally with the exact command CI uses, and every repo-specific gotcha (heavy always-compiled ML build, MISTRALRS_METAL_PRECOMPILE, clippy-in-loop timeout, macOS-only steps, the MURMUR_CI_SKIP_E2E toggle). Use whenever the user wants to run/fix/understand CI, triage a red pipeline, or check why the gate failed.
The seam-decision checklist for Murmur — how a new AI capability, model, provider, tool, connector, or surface should fit the existing seams (the SummarizerProvider trait + capability methods, the ONE egress-consent-redaction-ledger seam, the agentic-loop envelope, the gated tool ACI, routing shadow-cutover, SQLite-canonical readers, the eval gate). Use proactively when a task is "add/change an AI capability and the question is HOW it fits" — it sits between /research (whether to build) and /ship-feature (the mechanical build). Outputs a decision-ready spec naming the seam location, the trait/enum shape, the gate placement, the loop bounds, the eval-delta plan, and which implementer(s) + verifier(s) to dispatch.
The MAP of Murmur's shipped retrieval + agent-memory stack by file:symbol (hybrid fusion, FTS5/vec0 schema, embedding models, the reranker seam, bitemporal facts, consolidation, the entity graph, the eval harness) plus the MEASURED gaps and the load-bearing invariants. Use for any task touching retrieval quality, chunking, embeddings, reranking, fact/memory storage, consolidation, entity graph, or RAG eval. Not for embedder runtime cost on Metal (ondevice-model-perf).
Murmur app-wide performance map + a runnable measurement recipe across the Rust->IPC->Angular seam — FE change-detection anti-patterns, Tauri event-vs-Channel + payload windowing, OOM/RAM-pressure guards, deferred startup, the missing [profile.release], and the profiling toolchain. Use when the user reports jank/lag/heat/OOM, before a release, when adding a feature that touches a hot path, or on any 'why is X slow'. Not for model tokens/sec or quantization (that is ondevice-model-perf).
Run, understand, and debug Murmur's CI — the local `scripts/ci.sh` gate and the GitHub Actions macOS PR-gate that wraps it. The map of every gate step (remote enforcement → control-plane audits/evals → swiftc → clippy → cargo test → cargo audit → cargo deny → cargo build → ng lint → ng build → E2E), how to reproduce a red CI run locally with the exact command CI uses, and every repo-specific gotcha (heavy always-compiled ML build, MISTRALRS_METAL_PRECOMPILE, clippy-in-loop timeout, macOS-only steps, the MURMUR_CI_SKIP_E2E toggle). Use whenever the user wants to run/fix/understand CI, triage a red pipeline, or check why the gate failed.
The seam-decision checklist for Murmur — how a new AI capability, model, provider, tool, connector, or surface should fit the existing seams (the SummarizerProvider trait + capability methods, the ONE egress-consent-redaction-ledger seam, the agentic-loop envelope, the gated tool ACI, routing shadow-cutover, SQLite-canonical readers, the eval gate). Use proactively when a task is "add/change an AI capability and the question is HOW it fits" — it sits between /research (whether to build) and /ship-feature (the mechanical build). Outputs a decision-ready spec naming the seam location, the trait/enum shape, the gate placement, the loop bounds, the eval-delta plan, and which implementer(s) + verifier(s) to dispatch.
The MAP of Murmur's shipped retrieval + agent-memory stack by file:symbol (hybrid fusion, FTS5/vec0 schema, embedding models, the reranker seam, bitemporal facts, consolidation, the entity graph, the eval harness) plus the MEASURED gaps and the load-bearing invariants. Use for any task touching retrieval quality, chunking, embeddings, reranking, fact/memory storage, consolidation, entity graph, or RAG eval. Not for embedder runtime cost on Metal (ondevice-model-perf).
The discipline for adding or changing a check in Murmur's CI gate the right way — put it in scripts/ci.sh (the single source of truth), in the correct order, tool-absent-safe, without slowing the iterate loop, and (for a guardrail hook) with a RED-before-GREEN selftest assertion; then confirm GitHub Actions still just wraps ci.sh. Use whenever the user wants to add a lint/test/security/format/build check to CI, tighten the gate, or wire a new tool into the pipeline.
Author and maintain GitHub Actions workflows for THIS repo (Murmur — Tauri 2 Rust + Angular 22, macOS-first, heavy always-compiled ML build). The best-practice patterns tuned to Murmur — macos-14 runner, Swatinem/rust-cache + npm cache + whisper-model cache, SHA-pinned actions resolved live, least-privilege permissions, concurrency cancel, secrets by name, and the ci.sh-wraps-not-duplicates rule. Includes the release-workflow blueprint (CD) that stays a design, since notarization belongs to release-engineer. Use whenever the user wants to write/change a GitHub Actions workflow, tune CI runtime/caching, pin actions, or design a cloud release pipeline for Murmur.
The discipline for adding or changing a check in Murmur's CI gate the right way — put it in scripts/ci.sh (the single source of truth), in the correct order, tool-absent-safe, without slowing the iterate loop, and (for a guardrail hook) with a RED-before-GREEN selftest assertion; then confirm GitHub Actions still just wraps ci.sh. Use whenever the user wants to add a lint/test/security/format/build check to CI, tighten the gate, or wire a new tool into the pipeline.
Run a MULTI-PR program (audit remediation, God-file split, a batch refactor, a fix-all-findings sweep) on Murmur without freezing the Mac. The RAM-safe SEQUENTIAL, CI-gated cadence + the pre-push checklist + worktree hygiene + verifier-recovery discipline — proven driving a 12-PR program clean. Use whenever the work is "many PRs, one theme" and you'd otherwise be tempted to fan out parallel builders (which freezes this 64GB Mac). NOT for a single feature (that's /ship-feature); this is the orchestration LAYER that runs many ship-feature-shaped PRs in series.
Query Murmur's local read-only MCP server for gated, indexed knowledge — semantic + full-text search over meetings and notes, entity dossiers, structured transcripts, open-commitment rollups, and typed note-folder database queries. Prefer this over grepping raw vault files. Use when you need to find, retrieve, or cross-reference meeting/note content that Murmur has indexed. Read-only, loopback-only, excludes sealed content.
How to coexist with Murmur's notes in an Obsidian vault — note anatomy, front-matter keys, managed sections you must never edit, wikilink discipline, entity stubs, where files live, the locked-folder disappeared-file rule, and the add-freely/prefer-append coexistence contract. Use when reading, editing, or creating notes in a vault that Murmur also writes to.
Deep on-device model-performance context for Murmur on Apple Silicon Metal — the bandwidth-bound mental model, the quant/KV-cache ladder, whisper flash-attn/audio_ctx/profile discipline, the four crash/leak war stories, residency/RAM math, and the measurement bar. Use whenever a task is about MODEL-level performance / thermals / RAM (whisper.cpp / mistralrs / candle / parakeet) — tuning a quant or KV-cache, the Mac running hot, note-gen/ASR speed, the brain OOMing, the thermal governor, or a speculative-decode / Core ML ANE spike. NOT app IPC/UI/startup (use app-perf-audit) and NOT retrieval quality (use retrieval-memory-brain — this covers only the embedder/reranker/NER runtime cost).
Obszerny, ugruntowany research w kontekście naszej apki Murmur/brain2 — co możemy dodać, ulepszyć, jak to zrobić technicznie i jak wypadamy vs konkurencja. Rozbija temat na kąty, odpala równolegle subagentów `murmur-researcher` (web + kod), syntetyzuje i zapisuje raport do docs/research/. Użyj, gdy użytkownik chce zbadać pomysł/feature/usprawnienie/podejście techniczne dla apki, albo napisze /research.
Use before cutting a Murmur release (or whenever the user wants the public-facing copy checked/refreshed) to verify the landing page (`landing/index.html`) and the GitHub repo description still match what the app actually does now, and update them. Diffs the change surface since the last SHIPPED release, cross-references every public claim against the real code, flags stale/wrong/missing copy with evidence, then proposes edits and — on approval — lands them via a PR to `murmur` (the page auto-deploys) plus `gh repo edit`. Text only; screenshots are out of scope. Use whenever the user says the landing/description is out of date, wants marketing copy synced to reality, or runs this as a release preflight.
Deploy, redeploy, monitor, debug, and manage the murmur-server backend (the zero-knowledge E2EE sharing relay) on Railway. The complete, hard-won flow — project/service IDs, the token model, the GraphQL-API-not-CLI approach, deploy/logs/vars/domain/restore, and every gotcha that bit during the first deploy (env-freeze on redeploy, the Postgres name-collision DNS trap, raw-postgres password-init, the Dockerfile fixes). Use whenever the user wants to deploy/redeploy/check/fix/scale the Murmur server, read its logs, change its env, or manage its Railway hosting.
Tryb kreatywnego śnienia o Murmurze — Codex nasiąka apką, wędruje po necie, kombinuje i WYMYŚLA co możemy dodać, nie jako suchy research tylko swobodne myślenie generatywne oparte o głębokie zrozumienie apki i o to, co akurat przyjdzie do głowy. Rozchodzi się szeroko (dużo dzikich pomysłów), skupia na iskrach, buduje klikalny prototyp HTML+JS „na czuja" i daje userowi do akceptu; zaakceptowane deleguje do /ship-feature. Użyj, gdy user chce pomarzyć / pobawić się pomysłami / „wymyśl coś fajnego" dla apki, albo napisze /dreaming.
Promote mature Murmur learnings journal entries into an agent's binding Recurring patterns. Use when the user asks to curate lessons, promote repeated learnings, or names an agent learnings file under `.codex/learnings/`.
Record one operator-observed lesson in Murmur's compounding learnings journal. Use when the user asks to record a learning or gives input like "<agent>: <lesson>" for `.codex/learnings/<agent>.md`.
Deploy, redeploy, monitor, debug, and manage the murmur-server backend (the zero-knowledge E2EE sharing relay) on Railway. The complete, hard-won flow — project/service IDs, the token model, the GraphQL-API-not-CLI approach, deploy/logs/vars/domain/restore, and every gotcha that bit during the first deploy (env-freeze on redeploy, the Postgres name-collision DNS trap, raw-postgres password-init, the Dockerfile fixes). Use whenever the user wants to deploy/redeploy/check/fix/scale the Murmur server, read its logs, change its env, or manage its Railway hosting.
Deep on-device model-performance context for Murmur on Apple Silicon Metal — the bandwidth-bound mental model, the quant/KV-cache ladder, whisper flash-attn/audio_ctx/profile discipline, the four crash/leak war stories, residency/RAM math, and the measurement bar. Use whenever a task is about MODEL-level performance / thermals / RAM (whisper.cpp / mistralrs / candle / parakeet) — tuning a quant or KV-cache, the Mac running hot, note-gen/ASR speed, the brain OOMing, the thermal governor, or a speculative-decode / Core ML ANE spike. NOT app IPC/UI/startup (use app-perf-audit) and NOT retrieval quality (use retrieval-memory-brain — this covers only the embedder/reranker/NER runtime cost).
Obszerny, ugruntowany research w kontekście naszej apki Murmur/brain2 — co możemy dodać, ulepszyć, jak to zrobić technicznie i jak wypadamy vs konkurencja. Rozbija temat na kąty, odpala równolegle subagentów `murmur-researcher` (web + kod), syntetyzuje i zapisuje raport do docs/research/. Użyj, gdy użytkownik chce zbadać pomysł/feature/usprawnienie/podejście techniczne dla apki, albo napisze /research.
Tryb kreatywnego śnienia o Murmurze — Claude nasiąka apką, wędruje po necie, kombinuje i WYMYŚLA co możemy dodać, nie jako suchy research tylko swobodne myślenie generatywne oparte o głębokie zrozumienie apki i o to, co akurat przyjdzie do głowy. Rozchodzi się szeroko (dużo dzikich pomysłów), skupia na iskrach, buduje klikalny prototyp HTML+JS „na czuja" i daje userowi do akceptu; zaakceptowane deleguje do /ship-feature. Użyj, gdy user chce pomarzyć / pobawić się pomysłami / „wymyśl coś fajnego" dla apki, albo napisze /dreaming.
Use before cutting a Murmur release (or whenever the user wants the public-facing copy checked/refreshed) to verify the landing page (`landing/index.html`) and the GitHub repo description still match what the app actually does now, and update them. Diffs the change surface since the last SHIPPED release, cross-references every public claim against the real code, flags stale/wrong/missing copy with evidence, then proposes edits and — on approval — lands them via a PR to `murmur` (the page auto-deploys) plus `gh repo edit`. Text only; screenshots are out of scope. Use whenever the user says the landing/description is out of date, wants marketing copy synced to reality, or runs this as a release preflight.