mirror of
https://github.com/Heretek-AI/awesome-opencode.git
synced 2026-11-09 17:14:47 -10:00
John SmithandClaude Opus 5.5
3a6d104fcf
- kind/topics: keyword first pass, then reviewed and corrected by hand for all 513 entries - compat: detected from npm packages, package.json and source (scripts/utils/compat.js). Plugins: 134 v2, 125 v1, 12 unverified; 13 entries the review loop marked V2 pin @opencode-ai/plugin 1.x, and ~20 loop V2 verdicts based only on the absence of V1 markers are V1 plugin functions in source - npm and path filled where verifiable (package repository matches the entry); 9 bogus npm values from install prose were not kept - 29 repo URLs updated to the renamed/transferred repository; xberg-io entries point at their live repos instead of the archived monorepo; lavi uses repo + path instead of a /tree/ URL - removed projects/opencode-sessions.yaml (same repo as plugins/opencode-sessions.yaml) - install: renamed to installation:, mojibake fixed, inline HTML links removed from two descriptions Co-Authored-By: Claude Opus 5.5 <[email protected]>
17 lines
1.7 KiB
YAML
17 lines
1.7 KiB
YAML
name: staffetta
|
|
repo: https://github.com/RaffaeleSpezia/staffetta
|
|
tagline: "Model-decided context handoff: close a milestone, write memory to disk, restart with a clean session"
|
|
description: "A relay-race alternative to compaction for long-running agents. Instead of a threshold rewriting the conversation in place, the agent itself decides when a milestone is closed, writes what the next session needs to disk, and hands over — the session ends and a fresh one picks up the baton. Memory lives in files (MEMORY.md overwritten, JOURNAL.md append-only, AGENTS.md for rules), so it is readable, diffable, correctable and survives the session. Four tools the agent calls itself: checkpoint (verifies by checksum that memory actually changed, rejects journal entries with no numbers in them, detects tampering with older entries), ask (pauses for a human decision, requires the agent to state what it would do without an answer), resume (refuses until the decision is written down), and loop (reopens the session and nothing else — it ends when the model declares --done or --blocked, not when a counter infers it). Measured on a local KAT-Coder V2.5 via llama.cpp: 7 milestones closed autonomously, 7 of 7 checkpoints invoked unprompted, 587 tests green, ~5,500 tokens of context per round instead of a compacted session restarting from 142,000 — and it does not grow across rounds. Pure bash, no dependencies; adapters/opencode/loop.sh is the opencode adapter and everything under core/ works with any agent that can run a shell command."
|
|
kind: tool
|
|
topics:
|
|
- context
|
|
- memory
|
|
compat: n/a
|
|
tags:
|
|
- context-management
|
|
- memory
|
|
- long-running-agents
|
|
- autonomous-agents
|
|
- handoff
|
|
- bash
|