m4bwav/everlast-protocol
Nothing an AI session learned is lost when you change model, session, tool or vendor. Two tiers of plain-markdown knowledge: a per-project doc set (solutions with dead ends, decisions with reasons, plans, a 50-line HANDOFF, an index) committed with the repo or kept in an excluded folder when the repo cannot hold it, and a user tier in a private vault repository (profile, environments, cross-project lessons). A privacy split keeps people, politics and credentials out of any shared repo. Hooks orient each session, nudge capture once, and at session end sync the vault and push the project's docs (a pull request whenever the push is not certain). The vault asks once for a private remote or creates one. Every entry carries a recheck date: the index, the session-start line and a BM25 search over titles, aliases and bodies flag what is due, and recheck and verify check a stored fix against the current code before an agent acts on it. Typed links, a deterministic maintain command, and a reproducible benchmark. Installs in Claude Code, Cowork, Copilot and anything that reads Agent Skills. Follows the Evergreen Protocol in pointer mode.
Write down what a work session learned so no later agent, model or tool relearns it: dead ends with the fix and its verified command, design decisions with reasons and rejected options, living plans, a short HANDOFF for the next session, and lessons about the user or their machines, routed to the project's doc set, its private sidecar (people, politics, credentials, internal names) or the vault's user tier. Use whenever the user says 'document what we learned', 'write that down', 'handoff', 'wrap up', 'before you finish', 'record this decision', 'note the dead end', 'update the plan', 'remember that for next time', or a hook line starting [everlast] appears; also unprompted at the end of a task that hit a dead end, verified a non-obvious command, made a design choice or taught you something about the user. Also for 'refresh everlast-capture' and 'is everlast-capture stale'. A full session wrap-up that also improves skills, scripts and other stores is evergreen-wrapup when the evergreen plugin is installed; it calls this skill for the project docs. Setup is everlast-setup; reading is everlast-resume; the vault is everlast-vault; skill lessons go to that skill's LEARNINGS.md.
Start a task from what earlier sessions, models and tools already learned instead of re-exploring: read ai-docs/INDEX.md and its private index, open only the matching solutions, decisions and plans (or search them), check a stale fix before acting on it, continue or resume work from HANDOFF.md, read the user tier on a new machine or environment, run maintain on stale docs. Use whenever a non-trivial task starts in a repo with ai-docs/ or a vault registration, above all before fixing an error the user says is back after an earlier fix, and on 'pick up where we left off', 'what do we already know about X', 'did we try this before', 'search the notes first', 'what was decided about', 'is that fix still good', 'we fixed this before, it fails again', 'same problem as last time', 'what do you know about me / this machine', 'prune / consolidate the docs'. Also for 'refresh everlast-resume' and 'is everlast-resume stale'. Writing entries is everlast-capture; setup is everlast-setup; the vault is everlast-vault.
Give a project the Everlast Protocol's knowledge layer so nothing a session learns is lost: an ai-docs/ doc set (generated INDEX, 50-line HANDOFF, append-only log, decisions/ solutions/ plans/ notes/), registered in the private vault in one of two modes, committed with the repo or kept in an excluded folder junctioned from the vault when the repo must not hold it, plus a private sidecar for notes naming people, politics or credentials, an eight-line AGENTS.md pointer, and the lint. Use whenever the user asks to 'set up everlast', 'set up ai-docs', 'add agent docs to this repo', 'register this project', 'keep the notes out of the repo', 'make the docs private', 'stop the AI from relearning things', or when everlast-capture finds no doc set in a repo. Also for 'refresh everlast-setup' and 'is everlast-setup stale'. Writing entries is everlast-capture; reading them at task start is everlast-resume; the vault itself, a new machine, or export to other tools is everlast-vault.
Run the everlast vault, the git repository holding what AI agents learned for and about the user across machines, models and tools: create it, back it up to a private remote (named, or created with gh), check and sync it (status, commit, push the learnings, pull), run the privacy check before notes reach a shared repository (people, politics, credentials, internal names), bring everlast to a new machine or agent product (Claude Code, Cowork, Copilot, Codex), and list the registered projects. Use whenever the user says 'set up everlast here', 'install everlast on a new machine', 'back up the vault', 'create a repo for the notes', 'sync the vault', 'is the vault behind', 'scan for private info', 'export the skills to Copilot', 'pack the plugin for Cowork', 'where is the vault', or the session-start line says it is missing or behind. Also for 'refresh everlast-vault' and 'is everlast-vault stale'. Setting up one codebase is everlast-setup; writing entries is everlast-capture; reading is everlast-resume.