dsh-agi-federated

Author: txcdragon
Version: 1.5.0
Platform: DeepSeek Harness Web (Host + webServer index tap)
Persistence: D:\DeepSeek Harness\agi-federated (created automatically)

A federated AGI layer for DeepSeek Harness that combines bounded memory, deterministic context compaction, staged skill review, an evidence gate for factual claims, live token/cost observability, zero-trust audit, and advisory JEPA/CUV/mp-rcp seams.

Bundled plugins

This repository ships more than one DSH bundle. The root is dsh-agi-federated; nested bundles live under plugins/.

Path	Bundle	What it is
/	dsh-agi-federated	The federated AGI layer described below.
plugins/dsh-quota-nexus	dsh-quota-nexus v1.12.0	Multi-provider quota pool: provider registry UI (OpenAI-compatible, Anthropic Messages incl. custom base_url relays, Gemini), O(1) lease scheduling with an 8-account budget, split 429/5xx breakers with credential fuse, write-behind usage ledger, and 0s SSE quota telemetry with 80/90/95% alerts. Adds a left-sidebar panel entry beside Settings.

Publishing local edits to GitHub (no git needed on this machine):

node scripts/publish.mjs --all
node scripts/publish.mjs --plugin dsh-quota-nexus

What it does

• Hot Memory — a bounded, byte-stable prompt snapshot per session.
• Clue Store + BM25 — regex-distilled summaries, keywords, seed tokens, and anchors; no transcript storage and no LLM calls.
• Project Memory Bank — local-first, project-scoped knowledge pages shared by sessions in the same working directory, with BM25 recall, tags, graph links, and explicit remember/search/get/reflect tools.
• Surface checkpoint compaction — old messages are removed from the model-visible surface with a constant-length [AGI_CHECKPOINT] node; no summarization model call.
• Staged skills — writes go to staged/; promotion is deferred to idle/session-end review.
• Evidence gate — every write/edit claim is cross-checked against a per-session index of tool parameter and output anchors (400 records), with a five-step false-positive ladder.
• Usage monitor — dual-channel token accounting (live estimate + official usage settlement), provider balance refresh every second, monthly-budget breaker for spendOnly providers, and a floating status window.
• Context guard — occupancy percentage with 45/60/80% warning levels, backed by the last settled request breakdown.
• Model tracking — the active provider/model follows agent/request and request-context changes, with a recent model-switch history in the panel.
• Official cost — DeepSeek Flash pricing from the API price table (off-peak/peak), with the last settled request breakdown shown in the panel.
• Zero trust — workspace credential required for federated mutations; every deny/ask is appended to a hash-chained audit log.
• Federation seams — JEPA world model, CUV value/conflict scorer, and mp-rcp quorum arbiter are mounted as advisory, default-allow seams.

Left sidebar

After DSH is restarted and the page is refreshed, the plugin adds an AGI icon to the left sidebar. Clicking it opens the installed-state panel: host/model, Hot Memory, evidence index, skills, audit chain, federation seams, repository link, and the pending market-PR milestone. The panel polls the same status route as the floating window.

Tools

Tool	Purpose
agi_status	Model, session tokens, costs, provider balances, breaker countdown, evidence/skill counts, audit sequence.
agi_memory_snapshot	Exact bounded Hot Memory + BM25 injection text.
agi_clue_search	BM25 long-tail search over distilled clues.
agi_memory_remember	Save a project/global knowledge page (redacted, bounded, deduplicated).
agi_memory_search	BM25 search over project/global memory pages with graph links.
agi_memory_get	Read one page by id or best query match.
agi_memory_reflect	Deterministic duplicate merge; no model call.
agi_evidence_query	Query the evidence index by keyword or anchor.
agi_skill_stage	Stage a skill; never promotes immediately.
agi_skill_list	List staged/promoted/rejected skill metadata.
agi_skill_review	Schedule the background review.
agi_federation_ping	Probe JEPA/CUV/mp-rcp decisions.

Install

# Recommended: install the bundle from this directory
dsh plugin --profile web add "D:\Deepseek宸ヤ綔鍖篭dsh-agi-federated"
dsh --profile web

Or use the DSH Plugin Manager's Install bundle with target:

D:\Deepseek宸ヤ綔鍖篭dsh-agi-federated

For zero-trust skill writes, set a workspace credential before starting DSH:

$env:DSH_AGI_WORKSPACE_TOKEN = "<your-workspace-token>"

After installation, refresh the Web page. The floating window is served from the plugin client route and reads the plugin status route.

Persistence layout

D:\DeepSeek Harness\agi-federated\
  memory.json          # bounded Hot Memory + distilled clues only
  evidence.json        # anchors and summaries, never raw tool output
  usage.json           # token/cost aggregates and monthly buckets
  checkpoints.json     # surface checkpoint metadata
  audit.jsonl          # hash-chained deny/ask records
  skills\
    staged\            # candidates awaiting idle review
    promoted\          # reviewed skills
    rejected\          # failed deterministic review

Design document

The complete architecture is in ARCHITECTURE.md: component boundaries, pure snapshot rules, BM25 formula, surface replacement protocol, evidence ladder, usage parsers, failure modes, threat model, and verification matrix. The selective-adoption record for the refresh, PUA, white-box AGI, and evidence-first research proposals is in EVOLUTION.md.