This is a public slice of the JARVIS agent stack referenced in pewdiepie-archdaemon/odysseus#NNNN (the announcement Discussion that points here).
JARVIS is a personal coding-agent built on top of Claude. It has been running against the author's own workflow for about a year. This repo is the substrate the agent uses, not the agent itself: cron prompts, memory primitives, hooks, the advice-mining queues.
Everything in here is what gets read at boot or fires on tool events. Personal entries (engagement notes, partner-specific receipts, private channel content) stay local.
The canonical bodies for the cron-driven Odysseus loops.
odysseus-discovery.md— discovery loop canonical. C1 self-perpetuation, C2 dispatch, C2.5 advice mining (the part this announcement is about), C3 primitives extraction.odysseus-daily-cadence.md— daily-cadence loop canonical. Same 1/2/2.5/3 structure, different cap rules._advice-actionable-vibeswap.md— advice surfaced by C2.5 that applies to the VibeSwap codebase (React+Vite+Tailwind frontend, Solidity contracts, Python oracle)._advice-actionable-jarvis.md— advice surfaced by C2.5 that applies to JARVIS itself (this repo)._advice-mined-log.md— rolling log of generic-good-design advice that did not directly map to either substrate._primitives-pending.md— candidates for promotion to memory primitives, awaiting triage.
The discrete memory primitives bundled here are the ones load-bearing for the public posture and the Odysseus mission loop.
feedback_odysseus-as-advisory-substrate.md— the rule that frames Odysseus threads as a public audit of our shared AI-cosplay substrate.feedback_odysseus-daily-discussion-campaign.md— campaign rules for the daily-cadence loop.feedback_code-comment-why-only.md— origin: a style nit @vdmkenny left after merging PR #3363 in the Odysseus repo. Folded into the substrate the same week.feedback_spec-vs-deployed-severity-calibration.md— separates "spec-only" from "deployed" when classifying a finding's severity. Cross-cuts a lot of partner-review work.primitive_time-logic-anti-hallucination-gate.md— verify-then-assert for any temporal claim. Sibling of an entity-cross-reference primitive not bundled here.primitive_authorship-via-conditions-and-context.md— when conditions and context are authored by a human and execution is by an agent, attribution flows back through the conditions.project_odysseus-mission-loop.md— the 5-step mission loop. Inbound advice-mining, outbound public substrate, embedded principles, contribution-graph attribution, guided entry over time.
Selected hooks. They run from a Claude Code (Anthropic CLI) configuration that fires on tool-use events.
em-dash-augmentation-gate.py— scans partner-facing drafts for em-dashes and surfaces a scrub-warning. Generic partner patterns shipped; fork users plug in their own handles.hiero-gate.py— enforces operator-dense, short-line format for memory primitives. Blocks writes that read as prose.atomic-reflection-gate.py— fires on tool errors and timeouts; surfaces a pause-and-extract reminder so the agent does not paper over root causes.memory-preprocessor.py— at SessionStart, expands a hierarchical memory index by loading sub-indexes relevant to the active task. Keeps the always-loaded core small while making situation-matched memory available.corpus-cache-warmer.py— at SessionStart, loads the BPE-pre-encoded token cache (_system/corpus.tokens.jsonl), detects sha256 drift, auto-refreshes if stale, emits per-file token cost + boot-budget visibility into boot context. Shipped 2026-06-10 as L4 of the compression-layer stack.code-mode-nudge.py— PostToolUse on Bash/Read/Grep/Glob. Counts identical-tool calls in a sliding window; nudges toward one Python script over N round-trips when threshold trips. Shipped 2026-06-10 from the Anthropic Nov-2025 customer trace finding (150k → 2k tokens via code-mode orchestration)._terse_ops.py— L3 module: APL/J/K-style 1-char hot-path ops for hook-internal scoring math. Importable from any hook. First consumer: AA#4recent_research_count.
profile_hiero_tokens.py— measures cl100k_base token cost for HIERO Unicode operators vs ASCII alternatives. Built the empirical case for the HIERO++ swap dictionary (42% reduction on the operator set).pretokenize_corpus.py— builds the L4 Z-token cache. Reads MEMORY.md + sub-indexes + WWWD priority cache, encodes via cl100k_base BPE, writes_system/corpus.tokens.jsonlwith sha256-incremental updates.measure_savings.py— honest accounting of compression deltas per layer. Distinguishes operator-set savings from real-prose savings; distinguishes infrastructure cost from realized win.render_memory_graph.py— generates a clickable HTML graph of the memory corpus with type filters and search. DOMcreateElement+textContent(XSS-immune by construction).
odysseus_discovery.py— scores recent Odysseus repo activity for mechanism-design / coordination / governance fit. Cron-fired Claude session reads the JSON, picks the top candidate above threshold, runs a convergence-proof gate, drafts and dispatches.
Each memory primitive ends with a composes-with section that names parent and sibling primitives. Each hook references the primitives it enforces. Each cron-prompt references both. The dependency graph is mechanical to derive once an extractor is wired up; that is v2 of this repo.
This is downstream of work in two other repos:
- WGlynn/VibeSwap — the omnichain DEX and mechanism-design substrate. Specifically:
contracts/incentives/FractalShapley.sol,contracts/incentives/MicroGameFactory.sol,cogproof/src/shapley-dag/shapley.js,docs/architecture/patterns-existing/fractalized-shapley-attribution.md. The contribution-graph shape used by JARVIS for advice-attribution is the same shape proven on the VibeSwap side. - Usd8-fi/Usd8-fi-usd8-cover-score — USD8's cover-score algorithm, where I contributed the Shapley-value mathematical foundation (PRs #2, #3, #4, spring 2026).
MIT.