docs(wiki): note retrospection learnings diffuse across the grid (agent namespace)
Co-Authored-By: Blumi <ankur.nairit@gmail.com>
docs(wiki): document retrospection (daily transcript replay → memory)
Memory & Knowledge: a Retrospection bullet + the conflict-resolution update under
"Learning from outcomes"; Configuration: memory.retrospect + retrospect_hours.
Co-Authored-By: Blumi <ankur.nairit@gmail.com>
docs(wiki): code graph is 5 languages + CLI graph subcommands + TUI/blugo/gateway surfaces
Memory & Knowledge: languages now Rust/Python/Go/JS/TS (was "Currently Rust"),
the new `blumi knowledge impact|callers|implementers` subcommands, and a note
that the same queries are reachable from the TUI hot-spots, the gateway
/api/knowledge/graph endpoint, and the blugo Code-tab Impact sheet.
Co-Authored-By: Blumi <ankur.nairit@gmail.com>
docs(wiki): document the structural code graph, code_graph tool, RPL fan-in + rollover
- Memory & Knowledge: a "Structural code graph (opt-in)" section (graph.mode +
the code-graph feature + typed edges) and the code_graph agent tool
(callers/callees/impact/implementers).
- Configuration: knowledge.graph { mode, resolve_imports, max_edges_per_symbol,
rpl_impact } and llm.wake_on_rollover.
- RPL-Judgement: a note that file fan-in folds into the blast radius.
Co-Authored-By: Blumi <ankur.nairit@gmail.com>
docs(wiki): GEO/AEO pass across all pages — answer-first intros, FAQs, question headings
Optimize every wiki page for AI answer engines (GEO/AEO) and classic SEO,
additively — all facts, commands, config keys, defaults, and links preserved:
- Answer-first, self-contained intros so each page stands alone if an AI surfaces
it with no other context (each restates what blumi is in one clause).
- TL;DR / "Key facts" blocks of atomic, extractable facts near the top.
- Question-style headings ("What is X?", "How do I Y?") matching natural queries.
- A FAQ (Q&A) section on each substantive page — the most AI-citable format.
- Consistent entity naming + acronyms expanded on first use (RAG, SEDM, RPL,
BYOK, mDNS, FCM, STT/TTS, …).
- Fixed stale version (0.3.0 -> 0.5.0) on Home.
Co-Authored-By: Blumi <ankur.nairit@gmail.com>
docs(wiki): dedicated RPL-Judgement deep-dive page (the science of the judgement)
New SEO-friendly page documenting Raskolnikov's Psychological Loop in depth,
grounded in the actual code: the blast-radius risk model (severity weights),
the adversarial Porfiry LLM-as-a-judge + fail-open parsing, the five phases,
the Error-Delta regret signal feeding memory fitness, the Dostoevsky mapping
(inspired by Crime and Punishment), config reference, an FAQ, and an honest
"implementation status & roadmap" (live MVP vs the reserved multi-branch ToT /
worktree-sandbox knobs). Wired into the sidebar, Home nav, and cross-linked
from Self-Management + Memory & Knowledge.
Co-Authored-By: Blumi <ankur.nairit@gmail.com>
docs(wiki): v0.5.0 — memory learning-from-outcomes + RPL-Judgement
Memory & Knowledge: new "Learning from outcomes" section (probationary
failure→fix, value-based fitness + eviction, structure-aware/hub-suppressed
recall, off-gateway curation, conflict + consensus-diffusion groundwork);
heal.verify default flipped to true in the config example.
Self-Management: new "RPL-Judgement" section (the five-phase Raskolnikov loop +
config); updated self-healing to reflect probationary, verify-on-by-default
learning.
Co-Authored-By: Blumi <ankur.nairit@gmail.com>
Document cost-aware routing, always-on discovery, white-box memory
- Configuration: add `router` + `always_on` to the config reference (cross-refs).
- CLI-Usage: new TUI commands /route + /discoveries.
- Self-Management: full "Cost-aware routing" + "Always-on discovery" sections
(config blocks, safety, how to drive them).
- Memory-and-Knowledge: a "White-box memory (view/pin/edit/delete)" subsection
+ the /api/memory/* endpoints.
Co-Authored-By: Blumi <ankur.nairit@gmail.com>
Document GPU acceleration, self-healing & evolution, new TUI commands
Refresh the help/wiki for the [Unreleased] features:
- CLI-Usage: new slash commands /open-workspace, /heal, /accel (+ /grid and
/autocontinue) and a `blumi accel` management row.
- Memory-and-Knowledge: a GPU acceleration section (Apple CoreML default, CUDA
opt-in, local-server/Ollama path, grid-embed offload) + acceleration/heal
config; update embeddings.backend="grid" to the real strongest-peer transport.
- Self-Management: a Self-healing & evolution section (reflex recovery, failure
to fix learning, evolution autonomy, cross-step verify, /heal surfacing).
- Grid: a GPU-aware embedding offload section (POST /api/grid/embed).
- Installation: GPU/embeddings build options (BLUMI_CUDA, gpu-cuda, lean default).
- Troubleshooting: GPU/accelerator + self-healing entries.
- Configuration: acceleration/heal/embeddings/memory/knowledge cross-refs.
Co-Authored-By: Blumi <ankur.nairit@gmail.com>
Add Memory & Knowledge page (semantic memory, SEDM diffusion, code KB)
Co-Authored-By: Blumi <ankur.nairit@gmail.com>