Skip to content

Instantly share code, notes, and snippets.

View c0wb0yp0tat0's full-sized avatar
🎯
testing... updates soon

Ryan Stone c0wb0yp0tat0

🎯
testing... updates soon
View GitHub Profile
@c0wb0yp0tat0
c0wb0yp0tat0 / receipt.md
Created November 3, 2026 23:49
Full-suite receipt β€” PRs #139606 / #140008 (moonbeam pick-up)

Full-suite receipt β€” Copilot pick-up lane (moonbeam)

Run date: 2026-11-03 (UTC). Host: pve (Linux 7.0.14-16-pve, x86_64), 12 cores, ~31 GiB RAM. Python: 3.14.9 (hermes test venv, activated via scripts/run_tests.sh). Hermès version: v0.22.6+canary (git describe on PR head: v0.22.6+canary.20261103T064339Z-130-gd1bdb4582c).

Head verification (git ls-remote / checkout)

Fork: c0wb0yp0tat0/hermes-agent.

@c0wb0yp0tat0
c0wb0yp0tat0 / 001-conventions.md
Created October 16, 2026 03:49
Anchor + chain model, updated: the lean chain live + the 810-doc experimental chain β€” the model at two scales, with the write-time gate (#135039) as its guard.

001 β€” Conventions (how the model works, and why)

Append-only. Each entry: the rule, then the history that earned it.

The anchor rule

  • MEMORY.md = anchor only. Behavior-critical, every-session facts in ≀1 line; never nuance, never history. Anything longer moves out: nuance to its owning segment, identity to USER.md, a repeatable procedure to a skill. (Earned the hard way: an anchor that accumulated rules and pointers stopped being readable as an index; the cap held, but the signal diluted.)
@c0wb0yp0tat0
c0wb0yp0tat0 / README.md
Last active October 20, 2026 23:37
Chain-doc stress receipts for #116673: scale-invariant delimiter fragmentation, batch type-coercion leak, falsifier 5/5, failure-class alignment (gate_refused vs write_gated), skill-bridge pattern

Chain-doc memory tool β€” stress receipts for #116673

For @Finn763 β€” this is the stress-test report for the chain-doc memory tool, offered off-thread so the PR stays clean. Use it, adapt it, or set it aside β€” any part, or none. If anything reads off, the comment section here is open.

This gist is the offering, standing alone β€” a complete, self-contained stress-test report for the chain-doc memory tool, offered off-thread (no PR comment carries its context; readers need nothing else). Everything verified (scale tables, the batch-coercion leaks, five falsifier receipts, the optional skill-bridge pattern) is here, each tracing to a file and a line.

@c0wb0yp0tat0
c0wb0yp0tat0 / 001-conventions.md
Last active November 1, 2026 20:07
Memory chain: four append-only segments behind a MEMORY.md anchor β€” a working example of the anchor + chain model.

001 β€” Conventions (how we work, and why)

Append-only. Each entry: the rule, then the history that earned it.

  • Approved plans run END-TO-END. No gates invented mid-plan; only genuinely new decisions get gated. (09-17: earned the hard way β€” the thesis session-2 incident; a gate that didn't exist in the plan stalled an approved run and burned trust.)
  • Status = reality, never intent. Any published status (home page, reports) reflects what IS, not what's planned. (09-18: the "dashboard said in progress / you asked me anyway" trust incident β€” the pulse had been a hand-written snapshot. Fixed by machine-derived NOW + stamped LATEST.)
  • Tracker altitude. The home page is the network's OVERALL tracker; every project gets its own folder/page. Never grow project detail inside the tracker. (09-18, user-mandated.)
  • Doc waves: audit β†’ update β†’ audit. After work lands: fresh-eyes audit (read-only) β†’ apply findings β†’ second audit confirms. The last audit is the proof.
  • **Planner β†’ builder β†’