Changelog: everlast-protocol

Every change to README.md, protocol/, scripts/, hooks/ and the plugin-level companions, newest first, each with the reason. Reasons cite findings in RESEARCH.md (R-), lessons in LEARNINGS.md (L-), and test runs in TESTS.md (T-). State in evergreen.json. Protocol: MAINTENANCE.md. Each skill keeps its own changelog.

Entry shape: ### C-YYYYMMDD-n · date · one-line summary, then because: (IDs or "user request"), files: (file and section), and a sentence on what changed. Cite section headings, not line numbers.

C-20261008-1 · 2026-10-08 · Plugin 0.6.3: everlast-capture's description names evergreen-wrapup for a full session wrap-up

• because: user request (evergreen-protocol C-20261008-1 adds evergreen-wrapup, which also matches 'wrap up')
• files: skills/everlast-capture/SKILL.md (description), skills/everlast-capture/CHANGELOG.md (C-20261008-1), .claude-plugin/plugin.json and plugin.json (0.6.3)
• everlast alone is unchanged: 'wrap up' still triggers everlast-capture. With evergreen installed, the description says a whole-session wrap-up (skills, scripts and other stores too) is evergreen-wrapup, which runs everlast-capture for the project's docs.

C-20261006-1 · 2026-10-06 · Plugin 0.6.2: ready for GitHub Copilot CLI and awesome-copilot (root plugin.json, skills link their own copies of the protocol docs)

• because: user request (list everlast in awesome-copilot; vally lint rejects SKILL.md links that leave the skill folder, and the intake reads only a root or .github/plugin/ manifest)
• files: plugin.json (new, Agent Plugins 1.0, keywords cut to 10), .claude-plugin/plugin.json (0.6.2), scripts/sync-skill-refs.py (new), skills//SKILL.md (links), skills//references/protocol/ (generated), .github/workflows/tests.yml (check step), AGENTS.md (Rules)
• Each skill links references/protocol/X.md, a generated copy of protocol/X.md; links in a copy to files the skill does not carry become GitHub URLs. protocol/ stays the one to edit; python scripts/sync-skill-refs.py rewrites the copies and --check fails CI when one drifts.

C-20261003-2 · 2026-10-03 · Plugin 0.6.1: ready for the Claude plugin directory (privacy and hooks disclosure, manifest links, pinned launcher text, worktree publish fix)

• because: user request (submit everlast to the Claude plugin directory; claude.com/docs/plugins/pre-submission-checklist, read 2026-10-03)
• files: README.md (new sections Hooks, git and files outside the project; Privacy), .claude-plugin/plugin.json (0.6.1; documentationUrl, supportUrl, privacyPolicyUrl), scripts/everlast.py (cmd_publish), protocol/PORTABILITY.md (Per tool), protocol/DOC-TYPES.md (HANDOFF row), skills/everlast-vault/SKILL.md (A new machine, step 4)
• The README now names every hook and what it runs, every commit, push and pull request the script can make and when, every file written outside the project, what leaves the machine and which credentials are used (only the user's own git and gh logins); hooks run only in Claude Code and Cowork. npx skills is pinned to 1.7.0 in the docs (tested with --help). publish treated a plugin clone checked out as a git worktree (a .git file, not a folder) as "not a git clone"; it now accepts both, and the self-test passes from a worktree (146 checks). The checklist's mechanical checks pass: claude plugin validate (one warning: the repository's own CLAUDE.md at the root is not loaded as plugin context, which is intended), 170 files, none over 256 KiB, no binaries, no system files, valid names, no export-ignore, export-subst or filter attributes.

C-20261003-1 · 2026-10-03 · Research refresh (minor) and three lint fixes in the plugin records

• because: R-20261003-1 to 3 (notes only); evergreen audit lint (Search plan without labelled track blocks; TESTS.md defining T-20260913-1 twice)
• files: RESEARCH.md (Search plan: four labelled tracks from the evergreen RESEARCH.md template, best and noisy sources; R-20261003-1 to 3), TESTS.md (the scaffold entry renumbered T-20260913-3; the eval run keeps T-20260913-1, which LEARNINGS.md and C-20260913-1 cite), LEARNINGS.md (the real L-001, the plugin-eval sandbox lesson, had sat inside the template's example comment, so lint saw it as undefined; moved out, comment kept below it)
• No change to README.md, protocol/, scripts/ or the skills: Claude Code 2.1.285 to 2.1.288, the Memory Trust Gap study and the tooling baselines all agree with what the plugin already does.

C-20260929-2 · 2026-09-29 · Plugin 0.6.0: hidden text never reaches a doc set or a model's context (hidden-text-check)

• because: R-20260929-2 (Claude Code 2.1.284 neutralises invisible characters and markup-imitating tags in MEMORY.md; arXiv 2607.14611; the owner approved the check 2026-09-29)
• files: scripts/everlast.py (hidden text section: neutralize, hidden_hits, clean; note, handoff, verify, log clean on write; lint and scan report; the SessionStart hook warns; agent-read output neutralised; search and entities JSON cleaned before serialising), scripts/test_everlast.py (check_hidden, 21 checks), protocol/PRIVACY.md (Hidden text), README.md (commands), skills/everlast-vault/SKILL.md (Privacy scan), .claude-plugin/plugin.json (0.6.0), RESEARCH.md (R-20260929-2 applied; Open questions)
• Zero-width, bidi-control, invisible-operator and Unicode tag characters are dropped and tags imitating an agent harness's markup are broken with a backslash, on every write and in everything the hook, search, recheck, entities, maintain and lint print; lint and scan name file and line; clean fixes files in place. Emoji joiners, subdivision flags, a leading byte order mark and look-alikes such as C# generics are left alone. The new checks fail against 0.5.1 and pass here (146 checks); a dry-run clean over the owner's vault and 20 registered doc sets (about 600 files) found nothing to change, and the SessionStart hook's time and output on a clean doc set are unchanged.

C-20260929-1 · 2026-09-29 · Research refresh: a local-clone install loads in place; installPath is not the test

• because: R-20260929-3 (and R-20260929-1, 2, 4 recorded as notes and open questions)
• files: protocol/PORTABILITY.md (Claude Code row), RESEARCH.md (Open questions; R-20260929-1 to 4); everlast-vault SKILL.md (A new machine, step 2), LEARNINGS.md (L-001), RESEARCH.md, CHANGELOG.md (C-20260929-1)
• The install docs told the reader to reinstall whenever installPath named the cache; a probe on the owner's PC and anthropics/claude-code#96223 show the CLI and VS Code load the clone regardless, so the check is now the loaded path (a skill's base directory or CLAUDE_PLUGIN_ROOT). The memory-injection check (R-20260929-2) is recommended, not built.

C-20260927-1 · 2026-09-27 · Plugin 0.5.1: the privacy scan catches GitHub fine-grained, npm and Google keys; install docs follow the vault and capture research refresh

• because: everlast-vault R-20260927-1, R-20260927-2, R-20260927-3, R-20260927-5; everlast-capture R-20260927-1 to 3 (skills C-20260927-1)
• files: scripts/everlast.py (PRIVACY_PATTERNS: github_pat_, npm_ plus 36, AIza plus 35), scripts/test_everlast.py (check_scan_key_shapes: the three shapes, a trailing hyphen, backticks, a link URL, and an npm config name that must not match; 123 checks), protocol/PORTABILITY.md (Per tool: Claude Code, Cowork, Cursor and Gemini CLI rows), protocol/PROTOCOL.md §6 (one sentence, no version change), templates/INSTALL-PROMPT.txt (step 4), README.md (Cross-tool line), .claude-plugin/plugin.json (0.5.1), skills/everlast-vault and skills/everlast-capture (SKILL.md, RESEARCH.md, CHANGELOG.md)
• Cowork loads plugin hooks and uploads a zip through Add > Upload plugin; Cursor and Gemini CLI read ~/.agents/skills, so export alone covers six tools; a local-clone install loads in place, so a source edit needs no reinstall. The new patterns found nothing new in the plugin's docs or the owner's vault.

C-20260926-4 · 2026-09-26 · Plugin 0.5.0: entities, time-range search, evidence with proof counts, merge candidates

• because: the owner's request (bring over the cheap, no-dependency ideas from Hindsight, vectorize-io/hindsight); comparison in ai-docs/notes/2026-09-26-hindsight-comparison.md
• files: scripts/everlast.py (note --entities/--evidence, canonical_entities, proof_count, search --since/--until/--entity and filter-only listing, entities subcommand, verify increments proof_count, merge_pairs in maintain), scripts/test_everlast.py (check_050, 10 checks), protocol/PROTOCOL.md §8, README (Using the script), skills/everlast-capture and everlast-resume (one bullet each)
• Stays stdlib and plain markdown: entities are a frontmatter list, not an extracted graph; the time filter reads date and verified; merging stays a judgment call. Embeddings and an external benchmark remain the open items.

C-20260926-3 · 2026-09-26 · Plugin 0.4.3: the four skill descriptions tidied with skill-tidy (all under the 1,024-character spec limit, at most 12 quoted phrases)

• because: the owner's request (clean up the descriptions without losing function or trigger coverage); skill-tidy lint ST005 (capture 1,196, resume 1,420, vault 1,313 characters), ST007 and ST013 (resume and vault, 17 quoted phrases each), ST011 (capture, setup); T-20260926-2
• files: skills/everlast-capture, everlast-resume, everlast-vault, everlast-setup (SKILL.md description; CHANGELOG.md C-20260926-1 each), .claude-plugin/plugin.json (0.4.3), TESTS.md (T-20260926-2)
• One rewrite per skill, each passed tidy.py check (no lint error, no lost trigger, closest sibling no higher) before tidy.py apply: capture 1,015 characters (closest sibling 0.27 to 0.22), resume 1,018 (0.31 to 0.30), vault 1,015 (0.36 to 0.28), setup 985 (0.36 unchanged). Phrases that shared an intent were merged; the rest moved into the prose. Some hosts drop a description over 1,024 characters, so three of the four skills could have been invisible there.

C-20260926-2 · 2026-09-26 · Plugin 0.4.2: the vault lookup honours the repo a command names; recheck triggers on recurring errors; graders that measure the property

• because: T-20260926-1, L-007, L-009, L-010, L-011, R-20260926-2
• files: scripts/everlast.py (REPO_HINT, vault_path, main), scripts/test_everlast.py (note --private from another directory), skills/everlast-resume/SKILL.md (description: recurring-error phrasing and the stale-fix clause), evals/capture/graders (content.md removed; dead-ends.md, fix.md, verified-by.md, index-updated.md), evals/recheck/graders/check-before-edit.md (tool_order replaced by a trace regex), evals/privacy/graders/private-file.md (any private entry kind), scripts/eval-wsl.sh, scripts/eval_summary.py, scripts/eval_trace.py (new), TESTS.md (harness notes, T-20260926-1), LEARNINGS.md (L-008 to L-011), .claude-plugin/plugin.json (0.4.2)
• A private note written from outside the repo went to an unset-up ~/everlast-vault; the resume skill fired one time in three on "we fixed this before, it fails again"; three graders failed correct runs (an elided trace, an Edit-only order check, a solutions-only path). The suite is 5/5 at 3/3 with the plugin, every case ahead of the baseline except resume, which the model now does unaided.

C-20260926-1 · 2026-09-26 · Research refresh: vendor project memory, handoff comparables, how to tune the failing triggers

• because: R-20260926-1, R-20260926-2
• files: RESEARCH.md (Current understanding: vendor memory list, packaging and plugin eval; Open questions; R-20260926-1, R-20260926-2), README.md (comparable tools), TESTS.md (harness notes)
• Cursor and Claude Code both regained a project-scoped memory store, still product-local; the most-installed handoff skill writes to a temp directory; the recheck undertrigger must be tuned with plugin eval, not skill-creator's trigger rate, and plugin eval now needs git 2.31.

C-20260923-13 · 2026-09-23 · Plugin 0.4.1: the lint's link check skips inline code and fenced blocks

• because: L-006 (the first 0.4.0 lint of a real user tier reported the format example [title](path) inside inline code as a dead link)
• files: scripts/everlast.py (CODE_RE, without_code, the dead-link loop in lint), scripts/test_everlast.py (typed Related lint block), .claude-plugin/plugin.json (0.4.1)
• Links are found in the body with fenced blocks and inline code spans blanked, so a documented example is not a finding and does not inflate the SessionStart maintain count. Backticked paths are still checked by the separate path check, which is what backticks are for.

C-20260923-12 · 2026-09-23 · note keeps a body's own H1; the lint checks backticked paths in active entries only

• because: T-20260923-3; L-005
• files: scripts/everlast.py (cmd_note: no second # title when the body starts with an H1; lint_root: the dead-path check skips superseded, done and abandoned entries), scripts/test_everlast.py (two checks)
• History may name files that are gone; flagging them pushed an agent to reword a correct path in a new entry. Links between documents are still checked in every entry.

C-20260923-11 · 2026-09-23 · Eval cases for check before use (recheck) and search; resume trigger cases for the new phrasing

• because: the owner's request (prove the agent checks a stale fix before editing); R-20260923-2; T-20260923-2
• files: evals/recheck/ (case.yaml, prompt.md, scaffold.sh, graders: ran-recheck, check-before-edit, no-stale-fix, recorded, not-blind, skill-fired), evals/search/ (case.yaml, prompt.md, scaffold.sh, graders: ran-search, answer, skill-fired), skills/everlast-resume/evals/evals.json (trigger-3, trigger-4, decoy-3, action-2)
• The recheck case seeds a stale solution whose cited file moved on after it was verified; graders pass only on a recheck or Verified-by call in the trace before the first Edit. Both new cases need Bash, which claude plugin eval grants only under an OS sandbox (none on native Windows), so their first proof is recorded as described in T-20260923-2.

C-20260923-10 · 2026-09-23 · Plugin 0.4.0, protocol 1.5: README, research and version

• because: the owner's request ("update and improve everlast with everything learned"); R-20260923-1 to R-20260923-8; T-20260923-1
• files: README (intro bullets, What you get, Releases and contribution: comparable tools, Using the script, Design notes, new Benchmark subsection), RESEARCH.md (Current understanding edited in place, Open questions: external score plan, R-20260923-1..8), protocol/PROTOCOL.md header (1.5, 2026-09-23), .claude-plugin/plugin.json (0.4.0 and description; VERSION in scripts/everlast.py reads it)
• R-20260917-1 keeps its SkillsBench v1 numbers as logged; R-20260923-6 carries v4's.

C-20260923-9 · 2026-09-23 · CLAUDE.md and excluded mode's CLAUDE.local.md import AGENTS.md with an @AGENTS.md line

• because: everlast-capture R-20260922-3 (Claude Code 2.1.277+ reads AGENTS.md itself only when no CLAUDE.md or CLAUDE.local.md exists; a prose pointer loads nothing); the contradiction flagged on everlast-setup 2026-09-22
• files: templates/CLAUDE.md.snippet (why the import stays first; excluded-mode note), protocol/PORTABILITY.md (The always-on pointer per tool), skills/everlast-setup (SKILL.md Step 5, its C-20260923-1)
• No script writes these files; the skill does, following Step 5.

C-20260923-8 · 2026-09-23 · A benchmark: bench/ and scripts/bench_everlast.py, held as a floor by the self-test

• because: the owner's request (no benchmark score); R-20260923-3; T-20260923-1
• files: bench/README.md, bench/queries.json (41 typed queries), bench/fixture/ai-docs/ (44 synthetic entries, their INDEX.md, a HANDOFF.md and log.md), scripts/bench_everlast.py, scripts/test_everlast.py (BENCH_FLOOR_R3 0.90)
• Compares an index scan with search on Recall@1, Recall@3, MRR by query type and two staleness measures; the query types are checked before scoring. Written by the same team as the method: a regression floor, not evidence against RAG.

C-20260923-7 · 2026-09-23 · SessionStart line ends with recheck due: N (titles) · maintain: M when something is due

• because: the owner's request (upkeep that does not depend on the agent); R-20260923-2, R-20260923-8
• files: scripts/everlast.py (cmd_hook_run SessionStart), adapters/copilot/hooks.json unchanged (it runs the same script)
• Read-only (no git call, nothing written), at most three titles, silent when nothing is due. No new hook event. About 0.2 s on the 44-entry fixture and 1.3 s on 528 near-duplicate entries (T-20260923-1).

C-20260923-6 · 2026-09-23 · everlast.py maintain: an upkeep report; --apply archives and relinks, nothing else

• because: the owner's request (upkeep depends on the agent); R-20260923-8
• files: scripts/everlast.py (maintenance, duplicate_pairs, archive_entries, rewrite_links, rewrite_frontmatter_paths, cmd_maintain; entries(root, archive=True); the index lists archive/ entries under an Archive heading; lint findings carry a category), protocol/PROTOCOL.md §3 (Upkeep), protocol/DOC-TYPES.md (Prune pass), skills/everlast-resume (Step 5)
• The report: entries due for a recheck, done/abandoned/superseded entries older than 90 days, active entries more than a window past stale_after (proposed only), near-duplicate titles (difflib 0.85 within a kind, or 0.7 with the same tags), open contradictions, dead links and other lint findings. --apply moves the first group to archive/ in the same layout, rewrites every relative link to and inside them (and frontmatter supersedes paths), rebuilds the index, logs prune; it never merges, deletes or edits content.

C-20260923-5 · 2026-09-23 · Typed Related: links and per-fact stamps, checked by the lint

• because: the owner's request (coarse tracking of time and links); R-20260923-7; the Evergreen Protocol's parallel update (same grammar)
• files: scripts/everlast.py (related_links, LINK_RE, split_outside_links; lint: every relative link resolves, a supersedes target is superseded, unknown labels, open contradicts pairs, frontmatter supersedes targets, (verified YYYY-MM-DD) stamps older than the window; note --supersedes writes Related: supersedes [title](path) when the body has no Related line; search flags (stale facts)), protocol/PROTOCOL.md §8, protocol/DOC-TYPES.md (rules, Check before use), templates/AGENTS.md.snippet
• Labels: supersedes, superseded by, contradicts, builds on, see also; an unlabelled link counts as see also, so every existing Related line stays valid.

C-20260923-4 · 2026-09-23 · everlast.py search: BM25 over title, aliases, tags, summary and body; aliases frontmatter

• because: the owner's request (no meaning-based search); R-20260923-3; decision "Search is lexical (BM25 plus aliases) first" in ai-docs/decisions
• files: scripts/everlast.py (tokenize, stem, bm25_corpus, bm25_scores, search_entries, search_roots, cmd_search; note --aliases a,b and --alias "text"; the frontmatter parser reads quoted values and the writer quotes what YAML would misread, set_fields edits keys in place), protocol/PROTOCOL.md §8 and principle 4, protocol/DOC-TYPES.md (Finding entries, frontmatter, rules), skills/everlast-resume (Step 2), skills/everlast-capture (Step 5)
• Weights title and aliases x3, tags and summary x2, body x1; light stemming; identifiers and dotted names count whole and in parts; a superseded entry x0.5, done or abandoned x0.8, recheck due x0.9, so the current entry ranks first. Scope: the project root, --private, --user, or --all registered roots.

C-20260923-3 · 2026-09-23 · Protocol 1.5: check before use (stale_after, recheck, verify, the index flag)

• because: the owner's request (the check-before-use info for everlast solutions); R-20260923-2, R-20260923-4
• files: scripts/everlast.py (stale_due, is_stale, stale_stamps, find_entry, cited_paths, git_changes_since, cmd_recheck, cmd_verify, add_to_section; note --stale-after DATE|never; index_text puts (recheck due) in the status slot; INDEX header; EVERLAST_TODAY for tests; lint staleness by kind, --stale-days now an explicit override), everlast.config.json (stale_after_days), protocol/PROTOCOL.md (principle 6, §3 Start and During), protocol/DOC-TYPES.md (evidence rows, frontmatter, Check before use), protocol/PORTABILITY.md (the check by hand), templates/AGENTS.md.snippet, AGENTS.md (the everlast block), skills/everlast-resume (Step 3), skills/everlast-capture (Step 5)
• Windows by kind: solution 90 days, decision 180, note 120, plan 30. An entry without stale_after falls back to its window, so existing doc sets join unedited; --stale-after never writes the literal never (omitting the field would make the entry fall back to the window). recheck is read-only; verify renews, verify --failed records what broke and marks the entry due now. Behaviour change for existing users: lint now reports entries past their kind's window (a plan after 30 days) instead of after a flat 120 days.

C-20260923-2 · 2026-09-23 · A private machine path removed from the install prompt

• because: the repository rule (no machine paths beyond everlast.config.json); found in this release's privacy scrub
• files: templates/INSTALL-PROMPT.txt (step 1)
• The example marketplace folder named the owner's own directory; it now reads "an existing local marketplace folder if this machine has one". The git history still carries the old string.

C-20260923-1 · 2026-09-23 · Vault paths expand environment variables, so the shipped Windows default works

• because: L-004
• files: scripts/everlast.py (vault_path: os.path.expandvars for EVERLAST_VAULT and the config value), scripts/test_everlast.py
• The public everlast.config.json names %USERPROFILE%\everlast-vault on Windows, which was never expanded, so the vault resolved to a literal %USERPROFILE% folder under the current directory.

C-20260922-1 · 2026-09-22 · Machine and project labels generalised in the published files

• because: the repository rule (no personal data, machine names only in a fork's state files); found while refreshing everlast-capture
• files: TESTS.md (T-20260913-1, T-20260913-2 env), ai-docs/plans/2026-09-13-everlast-rollout.md (log line), skills/everlast-capture/LEARNINGS.md (a Scope and an Evidence line), skills/everlast-capture/TESTS.md, skills/everlast-resume/TESTS.md, skills/everlast-setup/TESTS.md (T- env labels), skills/everlast-setup/evergreen.json (tests env)
• The development machine's nickname became owner-pc (the label the public state files already use) and a private project and skill name became generic descriptions. The git history still carries the old strings.

C-20260920-1 · 2026-09-20 · No console windows from the detached SessionEnd sync on Windows (fork 0.3.4; official 0.4.0)

• because: L-003 (the owner saw a burst of console windows open and close when closing VS Code; each was a git.exe spawned by the detached vault or project sync)
• files: scripts/everlast.py (NO_WINDOW constant beside run(); run(), status_lines() and both --detach Popen calls pass it)
• CREATE_NO_WINDOW (0x08000000) on every subprocess call, the same flag evergreen_sync.py already uses. Made in the owner's fork first and brought here unchanged in code. General fix, worth a pull request to the official repository.

C-20260918-5 · 2026-09-18 · project register --mode excluded merges an existing ai-docs/ into the scaffolded store instead of nesting it (fork 0.3.3; official 0.4.0)

• because: L-20260918-1 in everlast-setup (registering a project whose ai-docs/ already had plans/ and decisions/ moved them to <store>/plans/plans/ because the store had been scaffolded first)
• files: scripts/everlast.py (cmd_project_register, the move loop)
• When both the source and the store hold a folder of the same name, the folder's files are moved into the store's folder and the empty source folder removed; files and folders with no counterpart move as before. General fix, worth a pull request to the official repository.

C-20260918-4 · 2026-09-18 · Protocol 1.4: the index first, with a summary clause per entry and no back-link requirement; plugin 0.3.2

• because: the owner's request (indexes that link most docs as a net positive for agents and people, without costing agent performance; back-links only where useful); R-20260918-1
• files: protocol/PROTOCOL.md §8, protocol/DOC-TYPES.md (layer table, frontmatter summary, rules), templates/AGENTS.md.snippet, scripts/everlast.py (note --summary, build_index shows the summary after the link), README §Using the script, .claude-plugin/plugin.json (0.3.2)
• 1.2 asked every entry to link back to its index; 1.4 drops that (the index-to-entry edge is the one graph editors show, and no study shows an agent using the reverse) and instead asks for a one-line summary so an index line says when to open the entry rather than being a bare path. Existing entries without a summary keep their old line; the field is optional.

C-20260918-3 · 2026-09-18 · Plugin 0.3.1: commit subjects list added files first

• because: the owner's stated reason for the remote ("I can see what it's adding in the commits"); the first real vault commit under 0.3.0 led with ~.gitignore, -.sync.log, ... and the new note fell into the body
• files: scripts/everlast.py (change_summary sorts + before ~ before -), .claude-plugin/plugin.json (0.3.1)
• The body still lists every file in that order.

C-20260918-2 · 2026-09-18 · Protocol 1.3, plugin 0.3.0: the vault asks once for a private remote (or creates one with gh); mode repo docs are committed and pushed at session end, with a pull request whenever the push is not certain; commit subjects list the files

• because: the owner's request (ask for a repository or permission to create a private one, back the notes up, and make the commits show what everlast adds; then, after a mode repo project turned out to have three unpushed commits carrying its docs, "really it should just be pushing commits, if it's not sure then a pull request instead"); L-002
• files: scripts/everlast.py (vault remote [<url> | --create [name]] with a visibility re-check; project sync <repo> [--if-changed] [--detach] [--pr] [--dry-run] with project_git_state, docs_pull_request in a temporary worktree, commit_paths that commits only the doc root; project register --sync push|pr|off kept across re-registration; change_summary for every commit subject, vault included; status_lines and author_email helpers; SessionStart reports a vault without a remote and a project's uncommitted or unpushed docs; SessionEnd runs project sync --if-changed --detach; vault init prints the back-up question), scripts/test_everlast.py (twenty-six new checks against local bare remotes: docs-only commits, empty remote, fast-forward push, behind-upstream pull request branch, sync off, re-register, vault remote), protocol/PROTOCOL.md §2 (backup paragraphs, the sure/not-sure rule) and §3, protocol/PORTABILITY.md, README §What you get, §Install and §Using the script, templates/INSTALL-PROMPT.txt step 2, skills everlast-vault (§Back it up), everlast-setup (§Step 3 sync choice), everlast-capture (§Prod mechanism), .claude-plugin/plugin.json (0.3.0)
• Sure means: upstream set, not behind after a fetch, every unpushed commit the user's own, push accepted. Otherwise the doc root's current state goes on an everlast/docs-<host>-<stamp> branch off the remote's default branch and a pull request is opened from it (gh when the remote is GitHub, else the compare URL), so no code travels and the user's working tree, branch and index are untouched. An empty remote is pushed to directly. Existing registrations without a sync key behave as push. Also fixed in passing: shareable_changes lost the first unstaged file of publish because run() strips the porcelain output (L-002).

C-20260918-1 · 2026-09-18 · Protocol 1.2: documents link each other (index per folder, Related lines, relative markdown links, no wikilinks); plugin 0.2.1

• because: the owner's request (make AI-written doc sets navigable for a human in Obsidian as well as for agents; a set of markdown files should be a graph, reachable from an index, not a pile)
• files: protocol/PROTOCOL.md §8 (new paragraph; version line), protocol/DOC-TYPES.md §Rules that keep it small (paths rule narrowed to files outside the doc set; new "linked, not just filed" rule), templates/AGENTS.md.snippet (one bullet), .claude-plugin/plugin.json (0.2.1)
• The generated INDEX.md already made every entry reachable in two hops; the new rule adds the edges between entries and states why relative markdown links are the only form that works on GitHub, in VS Code and in graph editors alike. No script change: the lint still checks backticked paths; a link check on Related: lines is a candidate for a later release.

C-20260917-2 · 2026-09-17 · pull from the official repository, SessionStart reports a clone that is behind, publish pushes to the official remote

• because: the owner's request ("point the local copy to the repo as the official version to check for updates and to pull request learnings back to"); a private fork's origin is not the official repository, so publish pushing to origin could never open the pull request it announced
• files: scripts/everlast.py (official_remote, behind_official, cmd_pull; SessionStart line; publish pushes to the official remote), README §Using the script, protocol/PROTOCOL.md §7, AGENTS.md (fork rules), scripts/test_everlast.py
• A plain clone has origin = official; a private fork adds upstream = official and keeps origin private. Both work the same way from the script's point of view.

C-20260917-1 · 2026-09-17 · Protocol 1.1, plugin 0.2.0: public release at https://github.com/m4bwav/everlast; consent-gated contribution; cross-platform scripts with CI on three operating systems; inline <private> marker and provenance fields

• because: the owner's request (make everlast public the way evergreen was: research, latest protocol, cross-platform scripts, versions, tests, releases; ask consent before any automatic pull request); R-20260917-1 to R-20260917-3; the Evergreen Protocol 1.6 sections 8 and 10
• files: scripts/everlast.py (contribute yes|no|status stored in the user's config dir with DO_NOT_TRACK and CI as no; publish gated on it, carrying only LEARNINGS/RESEARCH/CHANGELOG/TESTS files as a draft pull request; SessionStart nudges until decided; strip_private_blocks on repo-safe writes; --agent/--model provenance; portability fixes from the audit), scripts/test_everlast.py (gate cases, private-block case), scripts/everlast-hook.sh (interpreter probing), protocol/PROTOCOL.md §5 and §7, protocol/PORTABILITY.md (Codex path, SessionEnd timeout, autoMemoryDirectory), README (releases, contribution, comparable tools, python3 note), templates/INSTALL-PROMPT.txt (the consent question), .github/workflows (tests matrix, release on tag), everlast.config.json (placeholder vault path in the public tree)
• The public repository started from a scrubbed single commit (no hostnames, no owner paths); the owner's private clone is a fork of it that pulls routinely and contributes general lessons only.

C-20260913-1 · 2026-09-13 · Plugin created: protocol 1.0, four skills, everlast.py, hooks, vault, cross-tool packaging

• because: user request (an evergreen-style plugin so nothing learned across sessions, models and agent products is lost; user-level and per-project doc sets; an excluded-folder mode for repos that must not hold the docs; a repo-safe versus private split; encryption and other backups deferred); R-20260913-1 to R-20260913-8
• files: protocol/PROTOCOL.md, PRIVACY.md, DOC-TYPES.md, PORTABILITY.md; scripts/everlast.py (aidocs.py plus vault, registry, modes, tiers, privacy gate and scan, export, pack, sync, hooks), scripts/everlast-hook.sh, scripts/test_everlast.py; hooks/hooks.json; adapters/copilot/hooks.json; templates/; evals/; .claude-plugin/plugin.json and marketplace.json; everlast.config.json; skills/everlast-setup, everlast-capture, everlast-resume (absorbed from the ai-docs-* skills with their histories), everlast-vault (new)
• Decisions recorded in ai-docs/decisions/: absorb the ai-docs skills rather than depend on them; a separate private vault repository rather than data inside the plugin repo; excluded mode as a junction from the vault plus .git/info/exclude, with a lint re-check because Claude Code rewrites that file; privacy by rules first with a one-line question when borderline; evergreen pointer mode for every unit; encryption deferred.