feat(hermes): EverOS MemoryProvider integration, ops-mcp, DX harness and superpowers docs

.algo-profile/README.md

@@ -0,0 +1,10 @@
+# Algorithm Profile - EverOS
+
+## Optimization
+
+- [Content-Addressed Manifest Delta Sync](optimization/content-addressed-manifest-delta-sync.md) - O(N) first run, O(Δ) incremental, used in docs/superpowers/specs/2026-05-16-hermes-supergrok-nixos-auth-plane-design.md
+- [Bounded Top-K Heap](optimization/bounded-top-k-heap.md) - O(M log K), used in docs/superpowers/specs/2026-05-16-hermes-supergrok-nixos-auth-plane-design.md
+
+## Structures
+
+- [LRU Retrieval Cache](structures/lru-retrieval-cache.md) - O(1) average lookup, used in docs/superpowers/specs/2026-05-16-hermes-supergrok-nixos-auth-plane-design.md

.algo-profile/optimization/bounded-top-k-heap.md

@@ -0,0 +1,20 @@
+---
+algorithm: Bounded Top-K Heap
+category: optimization
+complexity_time: O(M log K)
+complexity_space: O(K)
+used_in: docs/superpowers/specs/2026-05-16-hermes-supergrok-nixos-auth-plane-design.md
+date: 2026-05-16
+---
+
+## Why This Was Chosen
+
+When multiple candidate sources are merged, the system only needs the best K snippets, not a full sort of the entire candidate pool. A bounded min-heap keeps the strongest candidates while avoiding the extra cost of sorting low-value items that will never be shown to Hermes.
+
+## Implementation Notes
+
+Use this only at the merge boundary where candidate sets from collection search, local cache, or memory providers are combined. If the source already returns a stable top-k list, the heap can be skipped; otherwise keep the heap small and enforce K as a hard cap.
+
+## Reference
+
+[Heap / Priority Queue](https://github.com/trekhleb/javascript-algorithms)

.algo-profile/optimization/content-addressed-manifest-delta-sync.md

@@ -0,0 +1,20 @@
+---
+algorithm: Content-Addressed Manifest Delta Sync
+category: optimization
+complexity_time: O(N) first run, O(Δ) incremental
+complexity_space: O(N)
+used_in: docs/superpowers/specs/2026-05-16-hermes-supergrok-nixos-auth-plane-design.md
+date: 2026-05-16
+---
+
+## Why This Was Chosen
+
+The xAI knowledge bundle is repo-shaped data, so a manifest keyed by content hash lets the host distinguish unchanged files from changed ones without re-uploading everything. That keeps the initial build linear while making steady-state refreshes proportional to the actual delta instead of the full corpus.
+
+## Implementation Notes
+
+The manifest should store path, content hash, upload state, and a stable bundle hash so a successful upload can become the new baseline atomically. Removed files should be tombstoned rather than silently forgotten, which keeps reconciliation explicit on the next sync run.
+
+## Reference
+
+[Big-O Reference](https://github.com/trekhleb/javascript-algorithms)

.algo-profile/structures/lru-retrieval-cache.md

@@ -0,0 +1,20 @@
+---
+algorithm: LRU Retrieval Cache
+category: structures
+complexity_time: O(1)
+complexity_space: O(capacity)
+used_in: docs/superpowers/specs/2026-05-16-hermes-supergrok-nixos-auth-plane-design.md
+date: 2026-05-16
+---
+
+## Why This Was Chosen
+
+Repeated Hermes queries against the same collection bundle should reuse prior retrieval results instead of hitting the collection on every turn. An LRU-style cache gives bounded memory with constant-time average lookup and eviction, which matches the read-heavy, hot-query pattern of the remote NixOS lane.
+
+## Implementation Notes
+
+Cache keys should include collection id, bundle hash, normalized query hash, top_k, and filter serialization so different auth contexts do not collide. A TTL layer should sit on top of the LRU policy so stale entries disappear even if the bundle hash does not change.
+
+## Reference
+
+[Data Structures Reference](https://github.com/trekhleb/javascript-algorithms)

.claude/hooks/commit-boundary-check.sh

@@ -0,0 +1,64 @@
+#!/bin/bash
+# commit-boundary-check.sh
+#
+# PreToolUse hook for EverOS. Reads the Claude Code hook payload from stdin,
+# self-filters to git-commit / gh-pr-create invocations, and warns when a
+# staged change set crosses multiple top-level directories.
+#
+# Rationale: PR #31 (Raven v2 closure) accidentally bundled 6 independent lanes
+# (raven, hermes use-case, skillhub, upstream-return, ci, chores) into one
+# 27-commit PR. EverOS convention from that retrospective: one component,
+# one PR. This hook is a soft nudge, not a block — cross-cutting work (lint
+# sweeps, dependency bumps, .gitignore policy) still proceeds.
+
+set -eu
+
+# Read the hook payload (we don't strictly need to parse it; we just want to
+# self-filter and inspect git state). Discard the JSON.
+cat >/dev/null 2>&1 || true
+
+# The actual command Claude Code is about to run is exposed via tool input.
+# Hook payload format varies across CC versions; we keep this hook scope-safe
+# by running unconditionally and only acting when staged changes exist.
+
+# Find the repo root from cwd so the hook works from worktrees too.
+repo_root=$(git rev-parse --show-toplevel 2>/dev/null || true)
+if [ -z "$repo_root" ]; then
+  exit 0
+fi
+
+cd "$repo_root"
+
+# What is staged for the next commit?
+staged=$(git diff --cached --name-only 2>/dev/null || true)
+if [ -z "$staged" ]; then
+  exit 0
+fi
+
+# Extract the first path segment of each staged file. Filter out hidden
+# top-level dirs (.github, .gitignore, .claude) and the root README so a
+# legitimate root-doc fix doesn't trip the warning by itself.
+top_dirs=$(echo "$staged" \
+  | awk -F/ 'NF>1 {print $1} NF==1 {print "_root_"}' \
+  | sort -u \
+  | grep -Ev '^(\.github|\.claude|_root_)$' || true)
+
+count=$(echo "$top_dirs" | grep -c . || true)
+
+if [ "${count:-0}" -ge 2 ]; then
+  cat >&2 <<EOF
+⚠ commit-boundary-check: this staged change set crosses $count top-level directories:
+$(echo "$top_dirs" | sed 's/^/  - /')
+
+EverOS convention (post-PR-#31 retrospective): one component, one PR.
+If these directories are part of the same logical change (e.g., a feature
+that genuinely spans methods/ and use-cases/), proceed.
+Otherwise, consider:
+  git reset HEAD <files-from-other-lanes>
+and committing the lanes separately.
+
+This is a warning, not a block. Re-run the commit command to proceed.
+EOF
+fi
+
+exit 0

.claude/settings.json

@@ -0,0 +1,16 @@
+{
+  "$schema": "https://json.schemastore.org/claude-code-settings.json",
+  "hooks": {
+    "PreToolUse": [
+      {
+        "matcher": "Bash",
+        "hooks": [
+          {
+            "type": "command",
+            "command": ".claude/hooks/commit-boundary-check.sh"
+          }
+        ]
+      }
+    ]
+  }
+}

.claude/skills/everos-prompts-sync/SKILL.md

@@ -0,0 +1,97 @@
+---
+name: everos-prompts-sync
+description: Use when editing EverCore prompts under methods/EverCore/src/memory_layer/prompts/en or .../zh, or before opening a PR that touches that tree. Verifies EN/ZH file-name mirror + export-symbol parity, surfaces missing zh files and divergent __all__ lists, and falls back to the existing src/devops_scripts/i18n/i18n_tool.py for code-comment translation drift.
+---
+
+# everos-prompts-sync
+
+Keeps `methods/EverCore/src/memory_layer/prompts/{en,zh}/` in lockstep at the
+**file-name + symbol-export** layer. Does **not** judge translation quality —
+that's a content review.
+
+This skill encodes a recurring EverCore failure mode: a contributor adds a new
+prompt constant to `en/` and forgets to add a matching entry under `zh/`,
+which silently breaks imports the moment a tenant uses the ZH locale.
+
+## When to invoke
+
+- A diff under `methods/EverCore/src/memory_layer/prompts/` is in flight.
+- A new prompt constant is being added to `en/<file>.py`.
+- A PR is about to be opened and the prompt tree has any change at all.
+
+If neither side of `prompts/` changed, skip — this skill has nothing to say.
+
+## Procedure
+
+1. **Confirm scope.** From repo root:
+
+   ```bash
+   cd methods/EverCore/src/memory_layer/prompts
+   ```
+
+2. **File-name mirror.** Both directories must have identical file lists
+   (excluding `__pycache__`):
+
+   ```bash
+   diff <(ls en/ | grep -v __pycache__) <(ls zh/ | grep -v __pycache__)
+   ```
+
+   Any difference is a bug. The fix is **always** to add the missing file to
+   the side that lacks it. The new file can be a translation OR an explicit
+   re-export from the other side (the existing convention — see
+   `zh/agent_prompts.py` for the re-export pattern).
+
+3. **Export-symbol parity.** For each file pair `en/X.py` and `zh/X.py`,
+   their public exports must be the same set:
+
+   ```bash
+   python -c "
+   import ast, sys, pathlib
+   for f in pathlib.Path('en').glob('*.py'):
+       if f.name == '__init__.py': continue
+       z = pathlib.Path('zh') / f.name
+       if not z.exists(): print(f'MISSING zh: {f.name}'); continue
+       def syms(p):
+           tree = ast.parse(p.read_text())
+           return {t.id for n in tree.body if isinstance(n, ast.Assign)
+                   for t in n.targets if isinstance(t, ast.Name) and t.id.isupper()}
+       en_syms, zh_syms = syms(f), syms(z)
+       if en_syms != zh_syms:
+           missing_in_zh = en_syms - zh_syms
+           missing_in_en = zh_syms - en_syms
+           if missing_in_zh: print(f'{f.name}: zh missing {missing_in_zh}')
+           if missing_in_en: print(f'{f.name}: en missing {missing_in_en}')
+   "
+   ```
+
+   Re-exports count: `zh/agent_prompts.py` doing
+   `from ...en.agent_prompts import FOO, BAR` exposes `FOO` and `BAR` as
+   ZH symbols — that satisfies parity even though the strings live on the EN
+   side only. The AST scan above catches direct top-level assignments;
+   re-exports need either a `__all__` list or a wider AST walk if you want to
+   be exhaustive.
+
+4. **Report.** Output one of:
+   - `PASS: EN/ZH prompt parity OK` (no further action)
+   - `FAIL: <list of mismatches>` (fix before merge)
+
+5. **Adjacent tooling.** This skill does **not** translate Chinese code
+   comments to English. That's `src/devops_scripts/i18n/i18n_tool.py`, which
+   is already wired into `make lint`. Use that for code-comment drift, this
+   skill for prompt-constant drift.
+
+## What this skill explicitly does NOT do
+
+- Translate prompts from EN to ZH or vice versa. That's a human/LLM content
+  task, not a parity check.
+- Validate template variables (`{messages_json}`, `{new_count}`, etc.) match
+  between EN and ZH versions. That's a deeper content check worth a separate
+  skill if it turns out to be a recurring failure mode.
+- Block commits. This is informational. Wire it into a hook only after the
+  false-positive rate is known to be near zero.
+
+## Recurrence threshold
+
+If the parity check has surfaced the same root cause **three times** (e.g.,
+"forgot to add zh re-export when adding a new EN prompt constant"), upgrade
+this skill into a pre-commit hook under `.claude/hooks/`.

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,6 +1,6 @@
 ## Summary
 
-<!-- What changed, and why? -->
+<!-- What changed, why, and what evidence shows it is correct? -->
 
 ## Area
 
@@ -15,7 +15,7 @@
 
 ## Verification
 
-<!-- List commands run, manual checks, screenshots, or reasons verification was not needed. -->
+<!-- List exact commands, checks, screenshots, or a clear reason verification was not needed. -->
 
 ```text
 
@@ -27,6 +27,7 @@
 - [ ] I updated docs, examples, or setup notes when behavior changed.
 - [ ] I added or updated tests when the change affects behavior.
 - [ ] I did not commit secrets, `.env` files, dependency folders, or generated output.
+- [ ] I listed the exact evidence, checks, or blocker for this change.
 - [ ] Active relative links in Markdown files resolve.
 
 ## Notes for Reviewers

.github/workflows/docs.yml

@@ -94,6 +94,7 @@ jobs:
               Path("methods/README.md"),
               Path("benchmarks/README.md"),
               Path("AGENTS.md"),
+              Path(".github/PULL_REQUEST_TEMPLATE.md"),
               Path(".github/CONTRIBUTING.md"),
               Path(".github/CODE_OF_CONDUCT.md"),
               Path(".github/SECURITY.md"),

.gitignore

@@ -206,7 +206,10 @@ src/memory_layer/memory_extractor/profile_memory_extractor keep llm merge.py
 #LLM related
 AGENTS.mk
 docs/api_docs/profile_extraction_fields.md
-.claude/
+# Claude Code: session-local artifacts only; skills/hooks/commands stay versioned
+.claude/settings.local.json
+.claude/worktrees/
+.claude/verify/
 .cursor/*
 
 #tmp_data
@@ -226,4 +229,22 @@ use-cases/**/package-lock.json
 
 # Local playwright + goal state traces
 .playwright-mcp/
-.goal/
+.goal/
+
+# Hermes EverOS Raven SkillHub dogfood artifacts (runtime / local traces)
+# These are intentionally generated during dogfood smoke tests but should not pollute git
+.kilo/
+use-cases/hermes-everos-memory/raven/.local-runs/
+
+# Windburn system artifacts (runtime, not temporary)
+.kilo/raven-aliases.zsh
+
+# Multica Ultimate Workbench runtime
+.automations/
+.automations/**/*.json
+
+# Codex session artifacts (already covered but be explicit)
+.codex/.codex-global-state.json
+.codex/.codex-global-state.json.bak
+.sisyphus/
+.gstack/

benchmarks/EverMemBench/.gitignore

@@ -26,9 +26,8 @@ htmlcov/
 .cursor/
 .DS_Store
 
-# Claude Code
+# Claude Code session artifacts (CLAUDE.md is versioned, .claude/ is session-local)
 .claude/
-CLAUDE.md
 
 # Logs
 *.log