brightwayai/cortex

v4.28.0MIT

Cortex. A host-portable Markdown second brain shared by Claude, ChatGPT Work, and Codex.

GitHub Open Website

Changelog

All notable changes to the Cortex Plugin are documented here.

Format follows Keep a Changelog. Versions match plugin.json.

[4.28.0] — Nightly-listen loud-failure hardening (2026-09-18)

Fixed

Root-caused why nightly-listen could "succeed" every night without ever reaching <config-root>: schedules registered by ops's /register-schedules without requires_local_device/folder binding can fire on a device that has no path to the user's config root, and the scheduler still reports success.

Added

/listen Step 0 is now a hard gate on config-root reachability: stat <config-root> then write+delete a probe file at memory/staged/queues/.listen-probe. On failure, write a receipt (ops's Step 5 schema) with status: "failed", error_code: "config_root_unreachable", and exit non-zero instead of degrading gracefully — graceful degradation is preserved only for individual connector failures.
/listen Step 1.5a preflight now reads <config-root>/briefs/.artifact-runtime.json (written by briefing's /brief Step 3.0) to locate a hosted artifact's db capability and read back briefs/<target_date> before mining, instead of relying on undocumented shared state.
/listen Step 1.5j writes a scheduled-run receipt with counts (closed, carried, snoozed, suppressed, reflection_written) and state_source: "none" when no brief state was found for the target date — a zero-item run now says why.
/listen now treats an existing <date>.closures.json (e.g. one written by briefing's /brief Step 0D0 fallback) as authoritative and never overwrites it.
commands/start-nucleus.md Step 6 now states the device/folder-binding requirement for nightly-listen up front and points to /register-schedules --verify for confirming it.

[4.27.0] — /listen mines the daily brief; /end-day and /morning stop depending on it (2026-09-17)

Added

/listen Step 1.5 — "Mine yesterday's brief." Reads briefs/<date>.state.json (tolerating tasks_checked-only and priority-only entries), maps ids back to titles/nodes via the brief markdown/seed, and stages explicit-mark proposals (done/delegate/skip/not_important/reprioritized/annotations/calendar notes/outreach actions) using the same semantics /end-day Step 2c always used. Runs an always-on inference pass for unmarked items against the freshly pulled archive ("likely done" with cited evidence, never auto-closed). Writes briefs/<date>.closures.json (the machine-readable handoff /brief now reads), the new cross-plugin snooze ledger briefs/.snooze-ledger.json (mirrored into growth's relationships/snoozes.json for outreach items), and carries the artifact's own "Today's Reflection" into that day's ## Reflection section without overwriting an existing one. Idempotent against --remine and against a same-day /end-day run (via a new briefs/<date>.state.processed marker).
Hosted claude.ai artifact-capabilities discovery path in /brief Step 3.0 (documented in briefing/commands/brief.md, mirrored here since /listen shares the same preflight): a shared-state read-back capability, not localStorage, for the case where the brief renders as a hosted artifact with no MCP bridge at all.

Changed

/morning Step 2 now walks brief-derived proposals first, under "From yesterday's brief" — explicit marks get a bulk-accept, inferred closures still walk one at a time.
/morning Step 4.6 checks for an artifact-authored reflection before asking cold; offers edit instead of re-asking when one exists.
/end-day Step 2c now writes (and checks) a briefs/<date>.state.processed marker so a same-day /listen --remine doesn't duplicate write-backs.
references/surfacing-prefs-template.md action taxonomy extended with reprioritized, calendar-note annotations, and the snooze-ledger fields (return_on, skip_count, last_detail).
Codex/OpenAI manifests now carry v4.27.0, and both generated Agent Skill discovery trees describe /listen's brief-state mining plus /morning's brief-derived proposal review.

[4.26.0] — Eval suite: onboarding + morning natural-language cases (2026-09-15)

Added

Four new claude plugin eval cases in evals/: start-nucleus-natural-language ("start nucleus"), lets-get-started-natural-language ("let's get started") — both check /start-nucleus fires without an explicit command — and run-my-morning-natural-language / good-morning-natural-language, extending the existing morning-natural-language case with two more phrasings that must fire /morning. All llm-graded, matching the existing suite's frontmatter/grader pattern.

[4.25.0] — Phase 3 consistency fixes (2026-09-15)

Fixed

scripts/configure_cortex.py now writes the legacy pointer ~/Documents/.claude-plugin-config-root alongside the primary ~/.cortex/config-root on first-time setup (previously only the primary pointer was written, silently starving hosts/plugins that still resolve the legacy path first-run). Also creates memory/me/ and plugins/ directories if missing.
commands/setup-identity.md Step 0 now documents the legacy-pointer mirroring and directory creation performed by the configurator.

[4.24.0] — `/setup-voice` moved to Comms Desk (2026-09-15)

Removed

commands/setup-voice.md and its skills/setup-voice/ (+ .agents/skills/setup-voice/) wrapper moved to the comms plugin (Comms Desk), which becomes a third Nucleus starter plugin alongside cortex and ops. The one-time voice-capture interview is now owned and shipped by Comms Desk. Removed setup-voice from scripts/generate_codex_skills.py's command list and spec table.

Changed

<config-root>/memory/me/voice.md remains a Cortex-owned canonical file location/contract — only the interview command moved, not the data file.
/start-nucleus Step 2 now offers comms:setup-voice instead of a local command, and degrades with an explicit "voice setup isn't available without Comms Desk" message (matching the existing "not installed" degraded-mode convention) if Comms Desk isn't installed. Step 5 (renamed "Ops and Comms Desk starter verification") and the frontmatter description now name Comms Desk alongside Cortex and Ops as the minimum supported bundle.
Updated all internal references (README.md, references/gitignore-template.md, references/autonomy.md, commands/migrate-scopes-v2.md, commands/setup-identity.md) to attribute /setup-voice to Comms Desk rather than Cortex.

[4.23.0] — `/review` and `/timeline` moved to briefing (2026-09-15)

Removed

/review (weekly synthesis digest) and /timeline (chronological activity view) moved to the briefing plugin (Today's Brief), along with their skills. Cortex no longer owns any periodic "what's going on" synthesis surface — that's now consolidated in briefing per the "Today's Brief is every what's-going-on surface at any cadence" container rule. /diagnose and /status stay in ops.

Changed

/end-week Step 3 now invokes briefing:review instead of an in-plugin /review; the step is skipped with a message (matching the existing "not installed" degraded-mode convention) if briefing isn't installed.
Updated all internal references (README.md, CLAUDE.md, SECURITY.md, claude-code/INSTRUCTIONS.md, references/autonomy.md, references/log-chronicle.md, references/reflections-template.md, commands/end-day.md) to point at briefing's /review//timeline instead of cortex's own.
Removed review/timeline from the Codex/Claude adapter generators (scripts/generate_codex_skills.py, scripts/lib/command_sync.py) and deleted their generated .agents/skills/ and .claude/commands/ mirrors.

[4.22.0] — local dir rename + cross-plugin reference sync (2026-09-15)

Changed

Renamed from claude-cortex local dir convention updated; no plugin ID change (cortex). Cross-plugin references updated for the 2026-09-15 Nucleus plugin rename (core-ops->ops, daily-brief->briefing, relationships->growth, delivery->clients, voice->comms, time-tracking->admin, news-curator->research, weekly-alignment->alignment).

[4.21.4] — operating-model architecture hardening (2026-09-15)

Added

Versioned autonomy-policy acknowledgment with actor and host provenance.
Shared-memory proposal and merge contract for safe future multi-writer use.
Stable actor provenance on new knowledge entries and deterministic exclusion of private/proposal state from index and hot-cache generation.

Changed

/start-nucleus now makes an honest ≤15-minute foundation promise and defers specialist interviews.
Identity and voice setup use the vendor-neutral config-root resolver.

[4.21.3] — Codex adapter synchronization (2026-09-15)

Fixed

Synchronized the portable and Codex manifests with the current Cortex version.
Added generated Agent Skill coverage for /migrate-scopes-v2.
Preserved disable-model-invocation gates in both portable and Codex-discovery skill wrappers so regeneration cannot re-enable ritual or destructive workflows.

[4.21.2] — claude plugin eval suite (2026-09-15)

Nucleus Operating Model Refactor Phase 4 step 4.2.

Added

evals/morning-natural-language/ — eval case testing the morning skill's natural-language trigger ("good morning" without typing /morning), with an llm grader checking the response acts on the natural-language request directly rather than asking the user to type the explicit command. Run with claude plugin eval . --case morning-natural-language; --ablation with-without (the default when the plugin resolves) reports the delta between installed and not — a delta near zero means the skill's description isn't matching natural phrasing and needs work.
.gitignore — excludes evals/results/ (per-run output, not checked in).

[4.21.1] — Fix: /morning should stay model-invocable (2026-09-15)

Fixed

Reverted skills/morning/SKILL.md's disable-model-invocation: true from the Phase 3.7 skill audit — /morning is the required daily touch and is specifically designed to fire on "good morning"/"what's on my plate" phrasing (see /start-nucleus's own "Try these" example). Caught while writing Phase 4 step 4.2's eval suite for it. 28 model-invocable skills now (was 27; still under the ≤30 target).

[4.21.0] — /start-nucleus gains autonomy acknowledgment, catalog fixed (2026-09-15)

Nucleus Operating Model Refactor Phase 4 step 4.1.

Added

/start-nucleus Step 2.5 — autonomy policy acknowledgment. Reads the ALWAYS/ASK FIRST/NEVER tiers from memory/CLAUDE.md, presents them plainly, lets the user accept defaults or customize, writes a .autonomy-acknowledged marker. Skipped silently on a first-ever run before memory exists (fires on the next re-run instead).

Changed

Step 5's per-plugin setup table and menu example rewritten for the current 9-plugin catalog — was still listing lead-engine, referral-engine, client-status, project-setup as separate installs (all retired/merged earlier in this refactor). Now lists core-ops, daily-brief, relationships (natively covers Apollo/signals + referral cooling), delivery (covers project setup, status, and deliverable QA), news-curator, time-tracking, voice, weekly-alignment.
Step 8's closing summary and "Try these" examples updated from nucleus-router//route to the chief-of-staff agent (/cos in core-ops).

Not done

The brief's literal 4.1 also asked for team/<slug>/ scope creation during onboarding. Skipped per Zach's decision — Phase 2 already chose the simpler memory/me/ scope with no physical org//team/ wrapper, since there's no second teammate yet. Revisit when one actually onboards.
Connector check is not a new step — /diagnose (invoked at Step 6) already covers it.

[4.20.1] — Skill auto-invocation audit (2026-09-15)

Nucleus Operating Model Refactor Phase 3 step 3.7. Ritual and side-effecting skills marked disable-model-invocation: true so they only run on explicit invocation, not loose natural-language matching — the model can still be asked to run them by name. Read-mostly, low-stakes, or high-frequency conversational skills are left auto-invocable. Marketplace-wide this brings model-invocable skills from ~81 to 27, under the ≤30 target audited with /skill-doctor.

Changed

Marked disable-model-invocation: true on: cleanup, end-day, end-week, forget, indexer, listen, log-writer, merge-research-draft, migrate-staged-substrates, morning, observe, rehearse, reindex, relink-memory, research-gaps, setup-identity, setup-obsidian, setup-sources, setup-voice, start-nucleus, start-workstream, sync-linked-entities.

[4.20.0] — /note and /learn folded into /remember (2026-09-15)

Nucleus Operating Model Refactor Phase 3 step 3.5.

Added

/remember --quick [node] [content] — one-line changelog capture, no extraction. Same behavior as the former /note.
/remember --knowledge <type> [node] [content] — single typed knowledge entry, no full-conversation extraction. Same behavior as the former /learn, now against the canonical seven-type taxonomy (Insight / Lesson / Model / Gotcha / Recipe / Correction / Decision) instead of /learn's stale six-type table.
Fixed a pre-existing bug while folding /learn in: it still resolved memory at the hardcoded ~/Documents/Claude/memory/ path, predating the <config-root> convention entirely. Now resolves via references/core-contract.md §1 like every other command.

Changed

commands/note.md, commands/learn.md — reduced to thin redirects documenting the equivalent /remember invocation. Kept as separate registered commands (not deleted) so existing muscle memory, scheduled tasks, and router intent mappings that reference /note//learn/$note/ $learn keep working. skills/note/, skills/learn/ already delegate to their command file as canonical — no change needed there.
README.md, CLAUDE.md, claude-code/INSTRUCTIONS.md command tables updated to describe /note//learn as aliases.

[4.19.0] — transcript-reviewer + conversation-miner + activity-miner merged into note-taker (2026-09-15)

Nucleus Operating Model Refactor Phase 3 step 3.6.

Changed

New agents/note-taker.md replaces the three mining agents, mode-dispatched (mode: transcript / mode: conversation / mode: activity). All three ran at the same pipeline step against the same node inventory — merging them is a packaging change, not a behavior change; each mode's procedure, triage gate, and return shape is preserved verbatim. A shared routing algorithm and shared learnings_delta proposal shape are now factored out once instead of duplicated three times.
/listen, /end-day, /end-week, /setup-sources updated to invoke note-taker with an explicit mode instead of the three separate agent names.
adapters/codex/agents/note-taker.toml replaces adapters/codex/agents/{transcript-reviewer,activity-miner}.toml. mode: conversation remains unavailable on Codex (no connector.session_history.read implementation), same as before.
Cross-referencing docs updated: references/note-sources.md, references/archive-layout.md, references/capability-matrix.md, docs/CODEX_SETUP.md, docs/PORTABILITY_READINESS.md, adapters/codex/README.md, AGENTS.md, README.md.

Removed

agents/transcript-reviewer.md, agents/conversation-miner.md, agents/activity-miner.md — content fully preserved in note-taker.md.

[4.18.6] — Autonomy docs updated for nucleus-router retirement (2026-09-15)

Changed

references/autonomy.md — references to nucleus-router updated to the new chief-of-staff agent (/cos in core-ops), which replaces it (Nucleus Operating Model Refactor Phase 3 step 3.1). No behavior change to the autonomy-slider mechanism itself.

[Unreleased]

[4.18.5] — writing-style renamed to voice (2026-09-15)

Changed

References to the writing-style plugin updated to voice in commands/setup-voice.md and commands/start-nucleus.md (Nucleus Operating Model Refactor Phase 3 step 3.4). /setup-voice itself stays in cortex.

[4.18.4] — /morning runs the Friday rehearsal batch (2026-09-15)

Nucleus Operating Model Refactor, Phase 2 step 2.5. Same rationale as v4.18.3: /end-week (which used to be the only place /rehearse ran weekly) is a separate optional ritual most users don't run every week.

Added

/morning Step 4.65 — on Fridays, runs /rehearse --batch 5 automatically. Idempotent alongside /end-week Step 3.5 if the user still runs both the same day.

Changed

Drained the 87-entry rehearse backlog once via cluster review with Zach (83 confirmed, 4 archived as superseded/dead-plugin references).

[4.18.3] — /morning commits + pushes memory-as-git (2026-09-15)

Closes a real gap: /end-day Step 5.8's commit-and-push logic never ran automatically because /end-day is now optional (v4.16+) and /morning — the required daily touch — never committed or pushed anything.

Added

/morning Step 4.7 — same commit + optional push logic as /end-day Step 5.8, condensed. Idempotent alongside /end-day (empty commits are a no-op, so running both on the same day doesn't double-push).

Changed

/end-day Step 5.8's push line now reads the remote name from memory_as_git.remote_name (default origin) instead of hardcoding origin — needed since a user's remote may be named anything (e.g. nucleus-memory).

[4.18.2] — Follow-up path/privacy fixes (2026-09-15)

Caught in a second sweep after v4.18.0/4.18.1 shipped.

Fixed

Missed memory/user.md / memory/reflections.md / memory/surfacing-prefs.md references (bare, without the <config-root> prefix pattern the first pass searched for) in references/hot-cache.md, commands/setup-identity.md, commands/setup-voice.md, commands/end-day.md, references/memory-index.md, docs/CODEX_SETUP.md — updated to memory/me/....
references/node-taxonomy.md — two more real BrightWay entity names used as examples (bizdev/jennifer-ives, bizdev/six-red-marbles), genericized.
.claude/commands/remember.md + commands/remember.md — real bizdev name ("trinity-education-group") in a worked example, genericized.
commands/relink-memory.md — real client name ("new-leaders") in a worked example, genericized.

[4.18.1] — /migrate-scopes-v2 command (2026-09-15)

Added

commands/migrate-scopes-v2.md + references/migrations.md entry — the idempotent, marker-gated command that performs the v4.18.0 file move (identity.md/voice.md/user.md/reflections.md/surfacing-prefs.md/ style-eval-guide.md → memory/me/) for any install that hasn't run it yet. Handles the identity.md/voice.md root-vs-memory/ duplicate cleanly if present. Deliberately does not create memory/org/.

[4.18.0] — identity.md/voice.md move to memory/me/ (2026-09-15)

Nucleus Operating Model Refactor, Phase 2 step 2.4. identity.md and voice.md are personal facts (this user's name/role/tools, this user's writing voice), not org-shared BrightWay facts — they belong in the private me/ scope alongside user.md, not at the shared config-root root. This is a breaking config-layout change: any script or note referencing <config-root>/identity.md directly needs the new path.

Changed

commands/setup-identity.md, commands/setup-voice.md, commands/setup-obsidian.md, commands/start-nucleus.md, commands/end-day.md — all path references updated to <config-root>/memory/me/identity.md / <config-root>/memory/me/voice.md.
Every other plugin in the marketplace that reads identity/voice was updated in the same pass (tracked per-plugin in their own changelogs) so the whole marketplace moves together — a partial migration would have broken any plugin still reading the old path.
The actual file move (existing identity.md/voice.md content into <config-root>/memory/me/) happens via the new /migrate-scopes-v2 migration (see that command's own changelog entry), not this doc change alone — read-path and file-location updates ship together in this session so there's no broken intermediate state.

[4.17.1] — Genericize hardcoded proprietary examples (2026-09-15)

Nucleus Operating Model Refactor privacy audit, prompted by Zach: confirmed plugin source is MIT-licensed/public, so any real client/personal data hardcoded as an "example" in command docs ships to every Nucleus user.

Fixed

references/node-taxonomy.md, commands/sync-linked-entities.md — example slugs/entities (client/holt-riptoes, person/rob-buelow, bizdev:vector-solutions, bizdev/jacquie-moen) replaced with fictional placeholders (client/aurora-labs, person/alex-rivera, bizdev:nova-solutions, bizdev/morgan-reyes).
commands/start-nucleus.md — example onboarding checklist no longer hardcodes "Zach Wagner, BrightWay AI"; uses "Jane Doe, Acme Consulting."
commands/end-day.md — example HubSpot batch-write table no longer names real deals/contacts ("Barker & Scott," "Common Cause," "Javier (Globant)," "Gaggle diagnostic") — replaced with fictional equivalents.

[4.17.0] — /recall load boundary + cap enforcement (2026-09-15)

Nucleus Operating Model Refactor, Phase 2 steps 2.2/2.3. Companion to the summary-layer restructure (v4.16.0's node changes) — this is where the token savings actually get enforced.

Added

scripts/lib/cap_check.py + cortex_cli.py check-caps — read-only, deterministic scan for cap violations: warns at 150 / fails at 200 lines for a node's ## Current state section, warns at 200 for always-loaded files (hot.md, CLAUDE.md, user.md). Wired into /reindex Step 2.5 and /cleanup Section M.

Changed

/recall <node> now defaults to loading only frontmatter + ## Current state + ## Open loops on restructured nodes — stops at the ---/## Record divider and never opens a sibling .record.md file. /recall <node> --full loads everything, same as the old default. Nodes not yet restructured (no ## Current state) behave exactly as before.

[4.16.0] — /end-day demoted to optional; /morning seeds today's brief (2026-09-15)

Nucleus Operating Model Refactor, Phase 1 step 1.5. The nightly /listen loop covers passive overnight mining, but it never replicated /end-day's deliberate tomorrow-priority curation — that gap is closed here before /end-day stops being a daily requirement.

Added

/morning Step 4.6 — ported from /end-day Steps 4.1/4.5/4.6: asks the three reflection questions (biggest thing done yesterday / what blocked you / one thing today has to move), writes a ## Reflection section to today's brief markdown, then walks candidate priorities/outreach individually and writes accepted items to <config-root>/briefs/<today_local>.seed.json for /brief to read. Skips cleanly if daily-brief isn't installed.

Changed

/end-day is now disable-model-invocation: true and described as an optional 2-minute close — the required daily touch is /morning, not /end-day. Still useful for full-mode transcript/inbox/Slack mining, the longitudinal reflections.md entry, and the HubSpot catch-all log; must be invoked explicitly.

[4.15.0] — ChatGPT Work and organization distribution (2026-09-14)

Added

Portable ChatGPT Work MCP bridge with bounded recall/search, confirmed note and section writes, deterministic reindex/hot-cache tools, and a confirmation-gated first-run memory-location setup.
Vendor-neutral scripts/configure_cortex.py onboarding command. It writes ~/.cortex/config-root, initializes only missing starter files through the shared Cortex CLI, and preserves existing memory.
Repository marketplace and compatibility manifests for direct local installs and ChatGPT workspace GitHub imports.
ChatGPT Work and organization-distribution documentation, including clean install, per-user storage, permissions, and update procedures.
Release audit, safe archive builder, and an end-to-end MCP smoke test that operate only on the shareable repository surface and temporary memory.

Changed

Shared Agent Skills now wrap the canonical workflows for Claude, ChatGPT Work, and Codex without forking behavior.
Plugin version and portable metadata are synchronized across Claude and OpenAI manifests.

Security

Local Slack/user cache files and macOS metadata are excluded from Git.
First-run configuration refuses unsafe roots and refuses to replace a different pointer unless the user explicitly confirms the switch.
All tests and distribution smoke checks use temporary homes and fixture memory; no real Cortex data is read or written. The suite now contains 156 fixture-only unit/integration tests.

[4.14.0] — Portability and stabilization refactor (2026-09-14)

Host-neutral core: a canonical storage/workflow contract, a capability matrix, and real deterministic code replacing several previously prose-only "the model does this carefully every time" claims. Establishes the foundation for a future Codex adapter without adding any Codex-specific workflow files.

Added

references/core-contract.md — single canonical contract: config-root resolution, memory layout, node schema, the 7-type knowledge taxonomy, wikilinks, timestamp/append-vs-replace rules, staging/archive, locking requirements, concept-drift, privacy tiers, backward-compat guarantees, and the capability-name list.
references/capability-matrix.md — logical capabilities (filesystem.*, web.search, connector.*, subagent.delegate, etc.) with per-host (Claude Code / Cowork) implementation and degrade behavior.
AGENTS.md — durable entrypoint for a future Codex session.
docs/PORTABILITY_READINESS.md — Codex mapping table, known gaps, and risks requiring real-memory testing (explicitly not done here).
scripts/cortex_cli.py + scripts/lib/{locking,atomic_write,node_paths, sections,decay,index_generator,hot_cache_generator,command_sync, repo_checks}.py — real, fixture-tested code for cross-process locking (with automatic stale-lock reclaim), atomic writes, node-path traversal guards, section-aware Markdown edits, decay classification, and deterministic memory/index.md / memory/hot.md generation.
scripts/generate_claude_commands.py — regenerates .claude/commands/*.md from the canonical commands/*.md files; CI now fails if they drift.
131 unit/integration tests under tests/, including tests/test_integration_cli_subprocess.py, which exercises cortex_cli.py as a real subprocess through full command-like sequences (a /remember- style write chain, /forget --archive, /rehearse's demote flow, and the paired lock-acquire/lock-release around an external critical section like /end-day's git commit step).

Changed

commands/remember.md, README.md, skills/learn/SKILL.md — reverted a stray four-type knowledge-taxonomy consolidation back to the canonical seven types (Insight/Lesson/Model/Gotcha/Recipe/Correction/Decision).
/remember and /note now write through cortex_cli.py (real locking + atomic writes) instead of ad-hoc file edits.
/forget, /cleanup, /rehearse, /relink-memory, /sync-linked-entities, and /end-day's index/hot-cache/git-commit-lock steps also now write through cortex_cli.py.
/reindex and the indexer skill now call the real deterministic generator instead of re-deriving the algorithm from prose each time.
All 9 .claude/commands/*.md files regenerated from canonical commands/*.md (they had silently drifted — missing features like /search's subagent-delegation step and /remember's commit triage).
Canonical lock filename standardized on .lock (was inconsistently .write-lock in some prose/gitignore templates that predated any real lock implementation).
agents/*.md role files annotated with capability-matrix references instead of naming Claude tools/MCP servers directly in role prose; added a missing connector.crm.read (and new connector.slack.read / connector.drive.read) capability.
Bare Haiku/Sonnet model-tier references across commands/*.md, agents/*.md, and references/*.md reframed as "low-cost/fast-tier" / "full-tier" with the Claude-specific choice kept as an explicit adapter annotation.
SECURITY.md rewritten to match actual behavior (previously claimed no network access and no code execution — both now false given /listen, /research-gaps, and cortex_cli.py).
CONTRIBUTING.md updated to describe the generator-based .claude/commands/ workflow instead of manual duplication.
scripts/check_repo.py expanded: skill name/directory agreement, broken internal file references, taxonomy-drift regression guard, command/skill coverage with documented exceptions, plugin.json/CHANGELOG.md version agreement, .claude/commands/ freshness, and the full test suite — all as part of one python3 scripts/check_repo.py entry point.

Compatibility

Existing ~/Documents/Claude default, the legacy ~/Documents/.claude-plugin-config-root pointer, and all existing node identifiers continue to resolve unchanged. A new optional CORTEX_CONFIG_ROOT env var and ~/.cortex/config-root pointer file are added above the legacy pointer in precedence, for sharing memory with a future Codex host.
No real user memory was read, migrated, or tested against in this refactor — all tests use tempfile.TemporaryDirectory() fixtures.

[4.13.2] — `/end-day` brief-state paste path (2026-07-07)

Companion to daily-brief v0.6.1 (state-mirror fix). Root cause of "end-day never sees my brief actions": the v0.6.0 artifact auto-mirror silently never worked (malformed tool name, no mcp_tools allowlist declaration, no connected filesystem MCP server — see daily-brief 0.6.1 changelog), so Step 2c.0 always fell through to the multi-select gate, forcing the user to re-state every action by hand.

Added — Step 2c.0p paste path

New source in the Step 2c.0 read chain, between widget context and the multi-select gate: ask the user to click 🔄 Sync for end-day in the brief artifact and paste the copied blob. /end-day validates the JSON and writes it verbatim to <config-root>/briefs/<today_local>.state.json, then proceeds exactly as if the state-mirror file had existed. One click + one paste instead of re-stating the day item by item. Step 4.0's chain reference updated to match.

Changed

Step 2c.0 source 1 documents the real mirror mechanism (daily-brief ≥0.6.1: filesystem MCP tool resolved at render time + mcp_tools allowlist; manual paste otherwise) and adds a staleness check on last_interaction_at.
Step 2c.0a fallback gate now fires only after the paste path is skipped or fails.

[4.13.1] — Daily-Brief Improvement Spec companion: brief-state fallback gate + doc fixes (2026-07-03)

Companion to daily-brief v0.6.0 (Daily-Brief Improvement Spec, 2026-07-02).

Fixed — `/end-day` can now actually read brief state (D2)

Step 2c specified read_widget_context(artifact_id="todays-brief"), but Cowork exposes no widget-context handle for persisted artifacts, so brief clicks were invisible to the close and completed items resurfaced in the next brief. Step 2c.0 now reads through a fallback chain:

State-mirror file <config-root>/briefs/<today>.state.json (written by the v0.6.0 artifact on every action) — canonical read path.
Widget context (legacy fallback).
Explicit fallback gate (Step 2c.0a, D2c) — when state is unreadable, /end-day asks one multi-select seeded with today's brief task list ("Which items did you complete / delegate / kill?") and routes answers through the normal Step 2c write-backs, instead of silently skipping. Makes the 7/2 manual reconciliation a spec'd step.

Added — reprioritize write-back (Step 2c.1a, D3)

tasks entries with reprioritized: true edit the priority tag on the source-node action ([P0]/[P1]/[P2]) and carry the change into Step 4.5's tomorrow-priority ordering. Readers tolerate entries with a priority but no action.

Fixed — Step 0.7 cost-card `$`-placeholder corruption (P1)

Literal $0.05-style figures in the Step 0.7 example were mangled by command-args interpolation (rendered ≈ --full.05) when /end-day was invoked with arguments. Examples now use the USD 0.05 form, with a note to keep it.

Changed — brief-contract references bumped to v0.6.0

Step 2c.0 / 4.0 / 5 reference the v0.6.0 blob shape, the state-mirror read path, and the no-native-dialogs sandbox invariant.

[4.13.0] — End-Day Routine Improvement Spec: brief mining, cost gate, taxonomy consolidation, reflections store (2026-06-08)

Implements Part B + the cross-cutting pieces of the End-Day Routine Improvement Spec. Coordinated with daily-brief v0.5.0.

Added — Step 0.7 source review & cost gate (B.1)

Before reading anything, /end-day shows one lightweight batch card listing the sources it will review, each with a checkbox (defaults from config) and an approximate token→$ cost shown before running, with a running total. "Today's Brief responses" is required, always first, never a checkbox. Unchecked sources are skipped and logged so the close summary is honest. Respects the autonomy slider (auto skips the card).

Added — Step 2c mine the brief artifact (headline fix; runs in BOTH modes)

The brief was the one source /end-day never mined. Step 2c reads the todays-brief widget context (v0.5.0 blob: tasks / annotations / outreach_actions) and writes back:

task actions — done→COMPLETE source action + "biggest thing done" candidate; delegate→reassign + delegatee task; skip→defer + increment per-task skip counter (.brief-skip-counts.json); not_important→suppress.
outreach actions — sent→log touch + value-add; booked→advance stage + prep task; nudge→follow-up touch; let_go/dead→mark dead; categories roll into outreach analytics.
suppression learning (2c.3) — not_important items + the repeat-ignore rule (≥3 surfacings, 0 actions) write into <config-root>/memory/surfacing-prefs.md Do-not-resurface, which /brief reads to filter future briefs. Creates surfacing-prefs.md from references/surfacing-prefs-template.md if missing.

Added — Step 2.9 learnings-first narrative + memories AND forgettings (B.2)

Before walking proposals, /end-day says what today was about in plain language first, then presents memories and forgettings side by side — forgettings (stale-fact demotions, suppressions, superseded beliefs) are first-class. Each proposal shows type, confidence, target node, source citation.

Added — Steps 4.5 / 4.6 propose tomorrow's priorities & outreach (B.5 / B.6)

After memory settles, /end-day walks tomorrow's candidate priorities and outreach individually (keep/edit/drop) with a batch shortcut for unchanged carryovers, respecting surfacing-prefs.md. Then asks once: "Any other priorities or outreach for tomorrow?" Output is written to briefs/<tomorrow>.seed.json, which /brief reads when Step 5 pre-stages tomorrow's brief.

Added — Step 5.8 offers to initialize memory-as-git when absent

/end-day Step 5.8 previously skipped silently (one-line suggestion) when <config-root>/memory/.git/ didn't exist. It now offers to initialize once (autonomy-aware: auto skips, suggest/confirm prompt y / not now / never→writes memory_as_git.enabled:false), running the same init path as /setup-identity Step 3.6. Closes the memory-as-git proposal's migration build-plan item; pairs with the new core-ops /diagnose Step 1D health line. Memory-as-git is now fully wired into the routine (/setup-identity init · /end-day 5.8 commit · /morning 0.5 diff review · /diagnose health).

Added — Step 4.7 optional HubSpot logging

After "anything else," /end-day asks once: "Anything to log in HubSpot before we close?" (skipped silently if HubSpot MCP isn't connected). Interprets the user's answer into concrete CRM writes — notes/activities, tasks, deal-stage moves, new contacts/companies/deals — presents one batch confirmation table, then writes via the HubSpot MCP and cross-links each object to its cortex person/bizdev/client node. Never auto-writes in a fire-and-forget run (10s → "no").

Added — Step 4.2 longitudinal reflections store (B.7)

Reflection answers are appended (newest-first) to a rolling <config-root>/memory/reflections.md in addition to the per-day brief ## Reflection. Created from references/reflections-template.md. Decays slowly; excluded from the v4.4 decay sweep. Step 4.0 pre-fill updated to read the v0.5.0 brief shape.

Changed — knowledge taxonomy consolidated to four types (B.3)

/remember now writes Insight · Decision · Gotcha · Correction. Model→INSIGHT [mental-model], Lesson→INSIGHT, Recipe→INSIGHT [recipe]. Node-file Knowledge sections, entry formats, concept-drift detection, and the Step 5 confirm list all updated. Legacy MODEL/LESSON/RECIPE entries remain readable. Absence of migration is a legible default.

Added — reference templates

references/surfacing-prefs-template.md and references/reflections-template.md.

[4.12.3] — Followup: deferred items from v4.12.2 (2026-05-28)

Completes the deferred-list from v4.12.2 CHANGELOG. Coordinated with daily-brief v0.4.2 + relationships v0.2.3.

Added — `/cleanup` Section L.0 first-run baseline-stamp

DASHBOARD provenance retrofit was going to be a multi-run grind for existing users (15 "missing provenance" lines per /cleanup run × N runs to clear backlog). v4.12.3 adds an L.0 prompt at first /cleanup run after v4.12.3 install: offers (b)aseline-stamp to mark all existing lines as  (defers them for 60 days), (w)alk per normal Section L behavior, or (s)kip for this run. Marker at staged/skip-logs/dashboard-baseline-acknowledged prevents re-prompt.

Same pattern applied to Section K (graph-isolation) — first-run defer-90-days option for legacy isolated nodes. Marker at staged/skip-logs/section-k-baseline-acknowledged.

Added — `/cleanup` skip-log enum constraints

Section L now uses a constrained enum for reason-code in dashboard-prune.md: not-stale | intentional-archive | pending-review | unclear. Same fix pattern applied to staged/skip-logs/sync-linked.md (was free-form reason). Closes the PII-leakage class of bug surfaced for the schema-validation override log.

Changed — `/sync-linked-entities` mtime check refinement

v4.12.0 used a 60-second mtime gate that false-positive'd for Obsidian users (auto-save every 2-3s). v4.12.3 splits behavior by invocation source:

User-invoked (explicit slug arg OR natural-language skill trigger) → skip the mtime gate entirely; trust the user is done editing.
Auto-fired (chained from /network-rebalance, /end-day, or scheduled task) → tightened to 10-second mtime gate with a warn-and-continue prompt rather than hard-exit.
Linked nodes read as-is (stale risk accepted; documented).

Changed — Stale-references sweep (continuation from v4.12.2)

Five more files now use current plugin names instead of retired weekly-outreach + bizdev-outreach:

commands/end-week.md — description + Step 5 (Monday outreach pre-stage) + behavior rule about HOLDING PATTERN
commands/start-nucleus.md — Step 0 status check + Step 2 voice description + Step 5 plugin table + Step 5 menu example + Step 7 schedule description + Step 8 closing summary

(commands/setup-voice.md already handled in v4.12.2 followup commit c962d0f.)

Acceptance

/cleanup Section L.0 baseline-stamp + Section K baseline-stamp prompts spec'd.
Skip-log enum constraints on dashboard-prune.md + sync-linked.md.
/sync-linked-entities source-based mtime check refinement.
No live-plugin references to weekly-outreach / bizdev-outreach in setup-identity, setup-voice, start-nucleus, end-week. Only migration-context references remain.
plugin.json bumped to 4.12.3.

Findings still deferred (not in v4.12.3)

Migration step in /setup-relationships for legacy user-context files — moved to relationships v0.2.3 (separate plugin release).
Automated test fixtures for plugin specs — meta-recommendation, separate effort.

[4.12.2] — Coordinated patch: security, race-conditions, backward-compat (2026-05-28)

Same-day coordinated release fixing 13 CRITICAL + 16 WARNING + 7 INFO findings surfaced by two independent post-ship review passes. Coordinated with daily-brief v0.4.1 + relationships v0.2.2 (shared state contracts).

Security/privacy

git config user.email default changed to nucleus-memory@localhost (was local@brightwayai — brand-leaking AND not RFC-valid). Real user email opt-in only via memory_as_git.commit_author_email in cortex.user-context.md.
Memory .gitignore split into two variants (references/memory-gitignore-template.md): local-only (default, tracks triage-log + dismissed-proposals locally for grep) and remote-safe (auto-selected when memory_as_git.remote set; excludes PII-dense files). Both variants verified against git check-ignore.
Parent .gitignore template extended to exclude relationships/events.jsonl, snoozes.json, inbox/, today.json. Prevents accidental config-root-level git tracking from exfiltrating relationship action history.
/morning Step 0.5 PII advisory before diff render — "Diff includes raw memory content; sharing this session will expose names/summaries."

Race-conditions / failure modes

Memory write-lock pattern (new): /end-day Step 5.8.0, /listen Step 0.8, /morning Step 0.5.0, /remember Step 3.0, /cleanup Step 4 all acquire <config-root>/memory/.write-lock (10-min stale TTL) before mutations. Coordinates concurrent commands. .write-lock is gitignored.
Listen-in-progress marker (new): /listen writes <config-root>/memory/staged/queues/listen-in-progress at Step 0.7; /morning Step 0.5.0 Check 3 detects it and pauses diff review until /listen completes.
/end-day Step 5.8 post-commit verification: git status --porcelain check after commit warns on dirty state (catches hook failures / interrupted runs).
/end-day Step 5.8 pre-commit recovery: detects non-empty index from a prior failed run; prompts user to commit-or-skip before adding today's changes.
/morning Step 0.5 dirty-tree recovery: if git status --porcelain shows dirty, surfaces a (c)ommit-now / (i)nspect / (s)kip prompt.
/morning Step 0.5 HEAD~1 precheck: git rev-parse --verify HEAD~1 gates the diff. First-day-of-memory-git users see "Memory-as-git is fresh — first diff appears after next /end-day" instead of a git error.

Backward-compat

.gitignore bug-detect changed to fingerprint-precise (v4.12.1's regex was too broad — false-positived on legitimate user-added inline comments). v4.12.2 detects only the exact v4.12.0 buggy template: requires all three specific inline-comment lines (log.md + operations chronicle, hot.md + 7-day rolling cache, index.md + auto-maintained catalog).
Pre-v4.12 install path: /end-day Step 5.8.1 detects when memory_as_git.remote is set after init and promotes the gitignore from local-only to remote-safe variant (running git rm --cached on triage-log + dismissed-proposals).
v4.12.0 → remote-pushed remediation: /setup-identity Step 4b detects when user has pre-v4.12.2 commits with cache files on a remote; surfaces git filter-repo remediation instructions one-time, marks acknowledged.
Symlink check on memory/: Step 3.6 step 2 warns before init if <config-root>/memory/ is a symlink (iCloud sync of .git/ corrupts across machines).

Other (Step 4 read implementation, contracts cleanup)

/end-day Step 4.0 NEW — reads brief artifact state: calls mcp__cowork__read_widget_context(artifact_id="todays-brief") to load tasks_checked + annotations from daily-brief v0.4.1's canonical JSON-blob localStorage shape. Pre-fills reflection prompts with task titles + candidate blockers + candidate priorities. Sanitizes content (paraphrases, never quotes verbatim) to keep sensitive client material out of the committed memory trail.
Sub-step numbering refactor: /end-day Step 5.8 split into 5.8.0 (write-lock), 5.8.1 (.gitignore validation), 5.8.2 (pre-commit recovery), 5.8.3 (compose+commit), 5.8.4 (post-commit verify), 5.8.5 (lock release).
Stale references swept: removed weekly-outreach + plan-tomorrow mentions from /setup-identity Step 4 (user-facing close-out line) + /setup-voice description.
Cross-doc step number alignment: /setup-identity Step 3.7 references "/setup-voice Step 3.5" correctly (was "Step 3.7").

Acceptance

Findings still deferred

/cleanup Section L baseline-stamp prompt + Section K defer-90 option (handles DASHBOARD provenance retrofit flood + Section K legacy-page flood) — pending v4.13.
/sync-linked-entities mtime check refinement (replace 60s gate with auto-fired-vs-user-invoked distinction) — pending v4.13.
Migration step in /setup-relationships for legacy weekly-outreach/bizdev-outreach user-context files — pending relationships v0.3.
Stale-references sweep across remaining files (end-week, start-nucleus, plan-tomorrow, process-brief table, setup-brief, setup-plan) — pending follow-up.

[4.12.1] — Fix memory/.gitignore inline-comment bug (2026-05-28)

Same-day patch to v4.12.0.

Bug

references/memory-gitignore-template.md v4.12.0 emitted patterns with inline # comments:

log.md                     # operations chronicle...
hot.md                     # 7-day rolling cache...
index.md                   # auto-maintained catalog...

.gitignore does NOT support inline comments. Each line is either a comment (starts with #) or a pattern — never both. Git parsed those lines as literal filenames, matching no actual file. hot.md, index.md, log.md, .state.json, and the person-mention counters were silently NOT excluded.

Result for any user who ran /setup-identity Step 3.6 on v4.12.0: memory git repo committed those high-churn cache files on every /end-day, drowning the daily diff review in cache churn — defeating the whole point of memory-as-git.

Fix

references/memory-gitignore-template.md rewritten: every comment on its own line above the pattern.
/setup-identity Step 4a (new) — defensive validation that detects the inline-comment bug via regex (^[^#]\S+\s+#), rewrites the .gitignore from the corrected template, and runs git rm --cached on the affected cache files. Runs even when .git/ already exists. Idempotent.
/end-day Step 5.8 — embedded the same validation as a belt-and-suspenders check. Catches users who ran /setup-identity on v4.12.0 and never re-run it; daily commit ritual self-heals.

Affected users

Anyone who installed cortex v4.12.0 AND ran /setup-identity between v4.12.0 and v4.12.1. Same-day patch — realistically a small population.

Self-healing on v4.12.1 install: either re-run /setup-identity (explicit; repair message surfaces) OR wait for next /end-day (silent repair before commit).

Credit: bug reported by the user during same-session post-ship verification.

[4.12.0] — Memory-as-git + dogfooding-driven hygiene (2026-05-28)

Addresses five observations surfaced during the inaugural /network-rebalance walk on 2026-05-28 plus the workstream/nucleus-improvements observations from 2026-05-21. Together these close the largest cross-artifact drift gaps in the cortex substrate.

Added — memory-as-git (substrate-level versioning)

Per nucleus/docs/proposals/memory-as-git.md. <config-root>/memory/ becomes a git repository. Each /end-day commits the day's memory changes as one reviewable unit; /morning surfaces overnight diff as a review surface.

New references/memory-gitignore-template.md — defines the memory-level .gitignore that excludes staged/, hot.md, index.md, log.md, .state.json, and deprecated pre-v4.8.1 dotfiles. Keeps the diff focused on knowledge changes.
/setup-identity Step 3.6 — first-run init prompt. Default: local-only, no remote. Writes memory/.gitignore from template, runs git init + initial commit. Optional remote prompt for off-machine backup (private GitHub / self-hosted).
/end-day Step 5.8 — auto-commits memory after Step 5.7 log. Empty commits are no-ops. Commit message includes day's source summary (which commands touched memory today). Optional push if remote configured.
/morning Step 0.5 — surfaces git diff HEAD~1..HEAD as the review surface BEFORE walking the /listen draft. Three render modes: full diff / file-level skim / skip.
cortex.user-context.md schema additions: memory_as_git.enabled, memory_as_git.remote, memory_as_git.push_on_close, memory_as_git.morning_diff.
Migration: existing installs gain init on next /end-day (if memory_as_git.enabled is true) or via /setup-identity re-run. Pre-v4.12 memory that already had git history (rare) is preserved.

Three privacy levels: local-only (default), private GitHub, self-hosted. None of them upload to BrightWay servers — memory stays the user's data.

Added — `/sync-linked-entities` command + skill

New cortex command that walks a source node's ## Linked Entities section and surfaces drift candidates in linked nodes. Addresses the cross-artifact drift gap: when a person's status changes, the bizdev/client/workstream nodes referencing them need updates but had no built-in mechanism to propagate.

commands/sync-linked-entities.md — read-only against linked nodes; surfaces 5 drift-check categories (status contradiction, stale summary, orphaned open threads, frontmatter intent/tier mismatch, provenance freshness). User accepts/rejects per candidate.
skills/sync-linked-entities/SKILL.md — natural-language entrypoint ("check the linked nodes," "any drift after that change").
Skip-log respected; 30-day suppression for (s)kip candidates.
Cap 25 candidates per invocation; sorted by severity.

Added — DASHBOARD line provenance (`/remember` Step 3 + `/cleanup` Section L)

Every line written or modified in DASHBOARD now carries  provenance. Drift detection becomes routine.

/remember Step 3.6 — provenance comment appended on every DASHBOARD line write/update.
/cleanup Section L (NEW) — DASHBOARD staleness scan via provenance. Per-line thresholds (auto-mining 7d / user-driven 30d / manual 60d). Surfaces stale lines, missing provenance, and orphaned references (lines pointing at archived/renamed nodes).
HTML-comment syntax renders invisibly in Markdown previews and Obsidian — humans see clean lines.

Changed — `/end-day` Step 5 artifact consistency (workstream/nucleus-improvements observation)

Pre-staging tomorrow's brief now ALWAYS calls mcp__cowork__update_artifact with id todays-brief. Never creates a new artifact and never produces a markdown-only fallback when Cowork is available. The brief is always the same persistent surface.

Canonical 6-section artifact format documented inline in Step 5 (sticky header, timeline strip, meetings card, priority tasks w/ checkboxes + progress bar, bizdev outreach queue, yesterday's reflection).
localStorage key brief-YYYY-MM-DD rotates with the date.
Reference implementation lives in daily-brief v0.4.0 (separate plugin release).
Markdown snapshot at <config-root>/briefs/<date>.md is still the canonical text record.

Changed — `/setup-voice` Step 3.5 + `/setup-identity` Step 3.7: graph completion (workstream/nucleus-improvements observation)

Both commands now upsert wikilinks to <config-root>/memory/user.md as a final step. Closes the orphan-graph problem: voice.md and identity.md were canonical root-level files referenced by every drafting plugin but had no inbound graph edges.

/setup-voice appends [[voice]] — writing voice descriptors... to a ## Canonical Files section in user.md (creates section if missing).
/setup-identity appends [[identity]] — user profile... symmetrically.
Idempotent (skips if wikilink already present).
Best-effort write — if user.md doesn't exist (cortex memory not bootstrapped yet), skips silently and waits for /remember's first run.

Storage layout additions

<config-root>/memory/.git/ — git metadata for memory-as-git. Excluded from cortex's general node-walking; tools that operate on memory should ignore .git/ explicitly.

<config-root>/memory/.gitignore — memory-level gitignore from references/memory-gitignore-template.md.

Acceptance criteria

Not in this release

Multi-command undo for individual /remember runs — day-granularity is the right tradeoff; per-command commit would flood the log.
Real-time multi-device sync — git pull/push gives manual cross-machine, not CRDTs. Out of scope.
/cleanup Section L for non-DASHBOARD files — provenance scan currently DASHBOARD-only; extending to person pages / client nodes would require provenance written there too. Defer to v4.13 if needed.

[4.11.0] — Structural discipline: taxonomy + write-time orphan check + cleanup section K (2026-05-20)

Closes the loop on the user's question — "will the structure be improved for future users so they don't end up with random nodes?" v4.10.x was retroactive (relink existing memory). v4.11.0 is proactive — prevents new memory from drifting back into the disconnected pattern.

Added — prescriptive node taxonomy

New references/node-taxonomy.md — explicit rules for what node type to use when. Covers 8 types: user / client / person / company / topic / workstream / bizdev / domain-root. Plus reserved subdirs infra/ and archive/.
Decision rules: 8 questions in priority order ("Is this about a specific person? a paying engagement? a prospect? an initiative? a company? a subject area? infrastructure? a persistent area of YOUR work?") — first match wins.
Anti-patterns documented: don't conflate prospect with client; don't dump knowledge into a mega-node; don't create ad-hoc top-level directories; don't create nodes without wikilink connections.
Naming conventions: kebab-case; firstname-lastname for persons with company-hint disambiguation; descriptive slugs for workstreams.

Added — `/remember` Step 1 consults the taxonomy

Step 1 ("Detect the target node") rewritten to apply the taxonomy's decision rules before creating a new node. Old colon-prefix conventions (client:acme, bizdev:stripe) marked deprecated; new content follows the prescriptive taxonomy directly.
Default fallback when nothing matches cleanly: topic/<slug>. No more ad-hoc top-level directories.

Added — `/remember` Step 3.4 orphan warning for new nodes

After Step 3 creates a NEW node file, check the outbound wikilink count.
If outbound < 2 AND content mentions named entities (capitalized "First Last" patterns, capitalized companies, recognized topics) NOT wikilinked, surface a one-shot prompt: "Heads up: has wikilinks but mentions <entity 1>, <entity 2>, <entity 3>. Convert?" — yes / no / skip-always-this-session.
Autonomy-aware: auto mode converts silently using the /relink-memory heuristic; suggest (default) surfaces the prompt; confirm adds emphasis.
Skips: already-existing node updates (only fires on first-write of new nodes); staged/ content (drafts); explicit --no-orphan-check flag; truly solo content (no named entities).

Added — `/cleanup` Section K — structurally-isolated nodes

Scans memory for nodes with zero outbound AND zero inbound wikilinks (excludes system files like DASHBOARD, CLAUDE.md, index.md, hot.md, log.md, user.md).
Per-isolated-node prompt: relink (try to convert plain-text mentions) / archive (move to memory/archive/) / merge (into another node) / keep (legitimate solo note; suppress 90 days) / skip.
Cap 10 per /cleanup run to prevent fatigue.
Different from existing Section G (orphan-detection by dashboard-presence and recency) — Section K checks graph connectivity specifically.

Why this matters

v4.10.0 fixed wikilink emission rules. v4.10.1 fixed DASHBOARD, first-name expansion, and person cross-linking. But all of those were retroactive — they fix accumulated mess in existing memory.

v4.11.0 is proactive: the taxonomy prevents ad-hoc node placement, the orphan warning prevents disconnected node creation, and Section K periodically audits structural health. Together they ensure new users don't drift into the same disconnected mess that required /relink-memory to clean up.

This matters across Nucleus, not just the graph view:

memory-librarian quality depends on consistent node structure
mining accuracy depends on canonical paths the agents can route to
daily-brief context loading depends on [[person/]] and [[client/]] patterns
cross-plugin reads (lead-engine, weekly-outreach, client-status) all assume the taxonomy

How users apply it

New users: automatically. Fresh installs start with the taxonomy; new content follows it from the first /remember.
Existing users: run /relink-memory --rerun to apply v4.10.1's retroactive fixes (DASHBOARD regen, first-name expansion, person cross-links), then /cleanup to surface any remaining isolated nodes via Section K. After that, structural discipline holds going forward.

[4.10.1] — Wikilink density: DASHBOARD hub + first-name expansion + person cross-links (2026-05-20)

Patch to v4.10.0. Real-user dogfooding showed three remaining issues after the initial wikilink fix:

DASHBOARD.md had zero outbound wikilinks even though it's the conceptual central hub. Pre-v4.10.1 dashboards used ### node-id section headers (markdown), not [[node-id]] wikilinks. In Obsidian's graph view, DASHBOARD appeared as an isolated island while every active node radiated from somewhere else.
First-name-only references missed by relink. Files mentioning "Erica Hruby" once and "Erica" three more times had only the full-name match converted. Subsequent bare-first-name occurrences stayed as plain text.
Person pages didn't cross-link to other persons. Graduated pages linked to clients and brightway-profile but not to colleagues mentioned in the same source contexts. Result: graph looked like spoke-clusters around each client, not a network fabric between people.

Fixed — DASHBOARD wikilink emission

commands/remember.md Dashboard File Format template updated. Every [node-id] reference in the active-nodes table, P0 list, Waiting On, Recent Knowledge, Stale Threads, Dormant Nodes, Isolated Notes, and Active People rows now uses [[wikilinks]]. Section headers stay as ### markdown but the references inside use wikilinks.
commands/relink-memory.md adds new Step 2.5 — detects pre-v4.10.1-style DASHBOARDs (< 5 wikilinks, ≥ 3 ### node-id section headers in Active Nodes) and regenerates them with wikilinks. Idempotent.

Fixed — first-name expansion in relink heuristic

commands/relink-memory.md Step 3 updated. After matching a full-name occurrence ("Erica Hruby") in a file, scan the same file for bare first-name occurrences ("Erica") and wikilink them. Constraints: first name must be unambiguous in this file (no two persons share it), word-boundary regex required, minimum 3-character match, contextually consistent.

Fixed — person-to-person cross-linking

commands/relink-memory.md Step 7 updated. During person-page synthesis, scan source nodes for OTHER persons mentioned in the same contexts (PEOPLE blocks, Recent Interactions, Open Threads, Changelog lines). For each other person with an existing page, add to ## Linked entities under Other people: [[person/<slug>]]. Limit 10 cross-links per page.
Reciprocal back-linking: when adding "Mary Kate" to Erica's Linked Entities, also add "Erica" to Mary Kate's Linked Entities. Avoids one-way edges. Idempotent.
New cross-linking pass on /relink-memory --rerun — back-fills person-to-person edges on existing person pages that were graduated before v4.10.1.

How to apply

Users on v4.10.0 should run /relink-memory --rerun after upgrading to v4.10.1. The rerun:

Detects + regenerates pre-v4.10.1 DASHBOARD with wikilinks.
Re-scans memory with first-name expansion enabled.
Cross-links existing person pages.

Expected outcome on a 30-node memory with ~12 person pages: DASHBOARD becomes the central hub with ~10-20 outbound edges. First-name references convert (~20-50 additional wikilinks). Person pages add ~3-8 cross-links each (~50-100 additional edges).

Combined with v4.10.0, total graph edges should be 5-10× the pre-v4.10 baseline.

[4.10.0] — Wikilink density: schema discipline + /relink-memory backfill (2026-05-20)

Real-user diagnostic: a memory scan against a user's <config-root>/memory/ showed 28 wikilinks across 30 node files (~1 per file). Obsidian graph view was mostly disconnected. Root cause: cortex schemas used bare-bracket [Name] placeholders, not wikilink [[name]] syntax; mining agents emitted plain-text entity references; person-page graduation never fired in practice.

This release fixes all four parts. See nucleus/docs/proposals/wikilink-density.md for the full diagnosis.

Added — canonical wikilink rule in CLAUDE.md

New section in cortex CLAUDE.md formalizes the wikilink convention:

Every entity reference in memory uses [[<type>/<slug>]] syntax (e.g., [[person/sarah-chen]], [[client/acme]], [[workstream/q3-outbound]]).
If a node file exists for the entity → emit the wikilink. If not → emit bare name + increment memory/.person-mention-counts.json for graduation tracking.
Applies across PEOPLE entries, knowledge entries, changelogs, open threads, DECISION Affected: fields, workstream Linked entities:, DASHBOARD, index.md.
Bare-bracket notation in cortex docs ([Name], [role], [context]) is template placeholder syntax, not output syntax. Don't confuse template brackets with wikilink brackets — they're visually similar but semantically opposite.

Updated — schema documents now use wikilinks

commands/remember.md Step 2 PEOPLE template + Step 3 People Index template both updated to [[person/<slug>]] ([role]) — [context]. Also in: [[<other-node-id>]].
Person-page schema's ## Linked entities section template updated to wikilink form.
Wikilink rule explicitly cross-referenced in mining-agent specs (transcript-reviewer, conversation-miner, activity-miner, memory-librarian inherit via cortex CLAUDE.md context).

Added — `/end-day` Step 3.7: passive person-page graduation

After Step 3 commits land, scan <config-root>/memory/.person-mention-counts.json for names with ≥ 3 mentions across ≥ 2 nodes that don't have a person page yet. Surface a one-line graduation prompt (capped at 3 per /end-day to prevent fatigue). On accept: synthesize a person page from all source nodes + relink the source nodes to use [[person/<slug>]]. On "never": suppress this name from future prompts.

Added — `/relink-memory` retroactive command

New commands/relink-memory.md + skills/relink-memory/SKILL.md. The load-bearing fix for existing memory that predates v4.10.

What it does:

Builds an entity registry by walking memory/ (every existing node file → its target wikilink + display-name variants from # Title lines, slug expansion, aliases).
Scans every node file for plain-text mentions of known entities NOT already inside [[...]]. Counts and contextualizes each.
Identifies person-name strings without person pages that meet the graduation threshold (≥ 3 mentions across ≥ 2 nodes OR ≥ 5 total).
Surfaces a single proposal with conversion + graduation counts.
User picks: accept-all / links-only / select (per-entity gate) / graduate-only / cancel.
On accept: edits file contents (plain-text → wikilink), synthesizes person pages from source contexts, refreshes index.md and hot.md.

Idempotent per the v4.8.1 migration pattern. Gated by memory/.migration-wikilink-relink-done marker. --rerun forces re-scan.

Conflict handling: name collisions (two persons → same slug) get the CLAUDE.md disambiguation rule. Partial matches require word-boundary + ≥ 3-character match. Case-insensitive proper-noun match preserves original casing in wikilink display.

Token cost: ~$0.10-0.50 per run depending on graduation count.

Added — migrations.md entry

references/migrations.md now lists wikilink-relink (v4.10+) as an active migration with command, marker, and rerun behavior documented.

Added — router intent rows (in nucleus-router v0.2.1)

Router routes "relink my memory", "fix my wikilinks", "my graph is disconnected", "back-fill the wikilinks", "graduate the people in my memory" → /relink-memory.

Why this matters

Before v4.10, memory worked but the graph view didn't. With the wikilink rule canonical + mining emitting wikilinks + /end-day graduating people automatically + /relink-memory back-filling existing memory, the Obsidian graph view becomes a real navigable network of your operating world. Expected outcome on a 30-node memory: edges go from ~28 to ~150-400 after one /relink-memory --accept-all run.

Downstream distribution

Schema + mining + graduation changes: ship as cortex code; every user updating to v4.10+ gets correct behavior automatically.
Retroactive /relink-memory command: ships as code, but the user has to RUN it once against their own existing memory (each install's <config-root>/memory/ is different). New users with empty memory don't need to run it (auto-detects empty state).

[4.9.0] — Workstream nodes + DECISION knowledge type (2026-05-20)

Part of the Chief-of-Staff evolution (see nucleus/docs/proposals/chief-of-staff-evolution.md). Two new cortex primitives close real navigability gaps in the second brain.

Added — `workstream/` node type

A workstream is an ongoing initiative pipeline that spans multiple projects, people, and topics — distinct from projects (specific engagements), topics (subject areas), and domains (persistent areas of work).

Examples: "Q3 outbound campaign," "BrightWay 2026 product strategy," "ops platform evaluation."

New references/workstream-schema.md — formal schema with sections for current state, pinned context, recent activity, open loops, linked entities, and workstream-scoped knowledge.
New commands/start-workstream.md + skills/start-workstream/SKILL.md — interview-driven workstream creation. Triggers: "start a workstream X", "I'm beginning work on X", "track X as a workstream".
Workstreams default to decay_profile: slow (1.5× modifier on freshness thresholds) — they stay fresh longer than topic nodes.
memory-librarian agent gains a new query path: workstream-shape queries ("what's happening with X initiative") route to memory/workstream/<slug>.md first.
indexer skill + references/memory-index.md updated — new "Workstreams" section in memory/index.md, grouped by status front-matter (active first, paused after, completed at bottom).
hot.md regeneration walks workstream nodes and surfaces active ones (last_active within 7d) in the "What I worked on" section.
Storage layout in CLAUDE.md updated to show memory/workstream/<slug>.md.

Added — `DECISION` knowledge entry type

Decisions become first-class memory entries alongside INSIGHT / MODEL / GOTCHA / LESSON / RECIPE / CORRECTION.

Why: decisions are forward-looking commitments, not retrospective lessons. They're worth surfacing when context changes (cheaper to re-evaluate a decision than rediscover it).

New DECISION type with required fields: what, when, why, affected entities, Revisit when trigger, status (active / superseded / revisit-now).
/remember Step 2 extraction updated — recognizes decision cues ("we decided", "I'm going with", "settled on", "going forward we'll", "the call is", "made the call to"). Prompts for the Revisit-when field when captured.
DECISIONs decay slowly (1.5× modifier, same as RECIPE and workstream nodes). They supersede via concept-drift detection — a later DECISION on the same topic moves the earlier one to Demoted knowledge with ↳ superseded by: ....
New /cleanup section J — DECISION revisit-trigger scan. Haiku-tier classifier checks whether each DECISION's Revisit when trigger appears to have fired in recent activity; surfaces flagged DECISIONs for user re-evaluation. Cap 10 per /cleanup run to prevent fatigue.
memory-librarian ranks DECISIONs higher for strategy queries ("why are we using X", "what was our approach to Y").
hot.md "Recent decisions" section is now properly typed — pulls DECISION entries from last 7 days, plus any with Status: revisit-now regardless of age. Never truncated even at word cap.
references/decay-model.md updated — DECISION joins GOTCHA and RECIPE in the slow-decay tier.

Schema documentation

references/workstream-schema.md — workstream node format
commands/remember.md Step 2 — DECISION extraction template + example
references/decay-model.md — DECISION decay rules
CLAUDE.md knowledge-taxonomy table — DECISION added with description

Migration

No migration required. Existing nodes continue to work; new workstream nodes are opt-in via /start-workstream. DECISION entries can be added to any node going forward; existing knowledge stays untyped.
/end-day Pre-chain C from v4.8.1 (staged-substrates migration) is unaffected.

Why this matters

The user identified two real navigability gaps in the second brain:

Operators run ongoing initiatives that don't fit project/topic/domain. Workstreams capture them with the right shape (current state, pinned context, open loops, linked entities).
Decisions weren't first-class — they got buried as changelog lines or retrospective lessons. With the DECISION type, every choice has a place to live with traceable reasoning + a re-evaluation trigger.

Together with the router evolution (nucleus-router v0.2.0), this is the Chief-of-Staff layer becoming real.

[4.8.1] — Staged substrates reorg + migrations pattern doc (2026-05-20)

Cleanup pass 1 (per nucleus/docs/proposals/cleanup-pass-1.md). Pure clarity refactor — no behavioral changes for users post-migration.

Changed — unified `memory/staged/` tree

Pre-v4.8.1, in-flight memory state lived in 8 different dotfile paths under memory/:

.commit-drafts/, .research-drafts/, .heartbeat-drafts/ (drafts)
.reindex-queue, .rehearse-queue.md (queues)
.rehearse-skip-log.md, .research-skip-log.md, .morning-reject-log.md (skip-logs)

These now consolidate under <config-root>/memory/staged/:

memory/staged/
├── commit-drafts/
├── research-drafts/
├── heartbeat-drafts/
├── queues/
│   ├── reindex
│   └── rehearse.md
└── skip-logs/
    ├── rehearse.md
    ├── research.md
    └── morning-reject.md

All 22 cortex files that previously referenced the old paths updated. CLAUDE.md storage layout updated.

Added — migration command + pattern doc

New commands/migrate-staged-substrates.md — one-time migration that detects pre-v4.8.1 dotfiles and moves them to the new layout. Idempotent; gated by <config-root>/memory/.migration-staged-reorg-done marker.
New references/migrations.md — names the marker-gated one-time migration pattern, lists active migrations (scope-migration, decay-config-init, gitignore-privacy-defaults, staged-substrates-reorg, hot-cache-first-generation), provides template for adding new ones.
/end-day gains Pre-chain C that invokes /migrate-staged-substrates if the marker is missing — ensures every upgrading user migrates within 24h.

Why this matters

Eight scattered dotfiles → one named tree. Easier to read, easier to gitignore (single memory/staged/ entry replaces 8 individual exclusions).
Future "staged state" features (e.g., /sweep heartbeat drafts in cortex v4.9+) drop into staged/heartbeat-drafts/ without inventing new dotfiles.
The migration pattern is now documented and named — future one-time data transforms follow a consistent shape.

Migration safety

/migrate-staged-substrates does conflict-checks; if both old and new paths exist (rare), it skips and surfaces a manual-resolve prompt rather than overwriting.
Old .commit-drafts/ / .research-drafts/ paths remain in the gitignore template as legacy exclusions (defensive — in case any leftover files exist after migration).
No content changes; pure path renames. Existing drafts continue to be reviewable by /morning and /merge-research-draft.

Files updated in this pass

20 cortex command + skill + reference files had path references rewritten via sed
CLAUDE.md (storage layout + slash-command table)
References: archive-layout.md, hot-cache.md, log-chronicle.md, memory-index.md, gitignore-template.md
New: commands/migrate-staged-substrates.md, references/migrations.md

Coordinated with nucleus repo

nucleus/docs/proposals/cleanup-pass-1.md (this cleanup's parent spec).
nucleus/docs/proposals/memory-as-git.md (next-up: vault-as-git pattern).
nucleus/docs/proposals/sweep-heartbeat.md (uses new staged/heartbeat-drafts/ path).
nucleus/docs/contracts.md (lists every cross-plugin file-format dependency with both old and new paths).
nucleus README "Start here" callout added.

[4.8.0] — /start-nucleus onboarding walker + /observe pruned (2026-05-16)

Added — `/start-nucleus` foundational onboarding walker

New commands/start-nucleus.md + skills/start-nucleus/SKILL.md — the "I just installed Nucleus, now what" command. Chains the essential setups in order:
1. /setup-identity (if identity.md missing)
2. /setup-voice (skip if user opts out of drafting)
3. /setup-sources (skip if user has no note adapters)
4. /setup-obsidian (recommended; user opts in)
5. Per-plugin /setup-* for each detected installed plugin (daily-brief, lead-engine, weekly-outreach, referral-engine, news-curator, client-status, project-setup, time-tracking, writing-style, core-ops, weekly-alignment, bizdev-outreach)
6. /diagnose (if core-ops installed)
7. /register-schedules (optional, if core-ops installed)
Idempotent. Detects completed setups via marker files; silently skips done steps. Re-running picks up where you left off. --reset (with per-file confirmation) walks from scratch.
Every step has a skip. Nothing is mandatory beyond /setup-identity (and even that surfaces a warning rather than failing).
Honors autonomy slider. autonomy: /start-nucleus: auto collapses the menu to "going through all setups now."
Logs to <config-root>/memory/log.md via the log-writer skill at completion.
Router (v0.1.4) added intent rows mapping "start nucleus", "let's get started", "set me up", "onboard me", "first time setup", "let's begin", "configure everything" → /start-nucleus.

Removed — `/observe` slash command

commands/observe.md deleted. The full spec consolidated into skills/observe/SKILL.md (which was always the canonical home — the command file's own description said "this is NOT a user-facing command — it is a background behavior").
Passive observation continues unchanged via the always-on skill. No functional change.
Router intent table updated (v0.1.4) to remove the "/observe" row. Trigger phrases like "passively observe X" / "watch for X" now route to no command — observation is always on, no command needed.
Migration: if you typed /observe in a memory entry or script, it will no longer resolve. The behavior it described is still active; just no slash-command surface.

Why this matters

/start-nucleus is the productization unblock. Before this, a new user had to know about and run 5-9 setup commands in the right order. After this, they say "start nucleus" and the AI walks them. Critical for sharing Nucleus with operators who aren't power users.
Pruning /observe removes a phantom command. The doc said it wasn't user-facing; now the catalog matches the doc.

[4.7.2] — Wiring + privacy: autonomy gates, log centralization, gitignore defaults (2026-05-16)

Added — `.gitignore` privacy defaults

New references/gitignore-template.md — the canonical privacy gitignore + cloud-sync notes for <config-root>/. Documents what's protected (archive/, .commit-drafts/, .research-drafts/, queue markers, plugin runtime state, Obsidian local-only state) and what's still versioned (active node files, briefs, hot.md, log.md, identity/voice/VAULT).
/setup-identity Step 3.5 (new) — writes .gitignore with privacy defaults if missing. Surfaces a one-line cloud-sync warning if <config-root>/ is inside iCloud / Dropbox / OneDrive / Google Drive (since .gitignore doesn't apply to cloud sync).
/listen Step 0.5 (new) — defensive check before pulling raw substrate. If .gitignore doesn't exist, creates it. If it exists but is missing archive/ or memory/.commit-drafts/, appends them under a "Added by cortex /listen first-run" section. Idempotent.
Cloud-sync caveat documented in the template — iCloud lacks selective sync; recommended pattern is symlinking archive/ to ~/.cache/nucleus/archive/ (out of the synced directory) before the first /listen run.

Added — autonomy wiring at command level (Karpathy Iron Man slider, take 2)

/forget Step 3 — consults autonomy mode (default confirm). auto skips the gate; confirm per-effect prompts; suggest single yes/no.
/cleanup Step 3 — consults mode (default suggest). auto executes all proposed actions inline; confirm walks per-item.
/setup-obsidian Step 2 — consults mode (default confirm). auto skips the plan-and-confirm prompt.
references/autonomy.md updated with "Where autonomy is currently wired (v4.7.2)" section documenting the three levels (router, command gates, future) and listing which commands are/aren't wired and why.

Added — log centralization

New skills/log-writer/SKILL.md — programmatic primitive invoked by other cortex commands at their "Log to chronicle" steps. Takes op_name + summary (and optional body, timestamp); writes one formatted entry to <config-root>/memory/log.md. Single source of truth for the log format.
All 9 commands with Log steps (/listen, /morning, /end-day, /end-week, /reindex, /research-gaps, /merge-research-draft, /cleanup, /rehearse) updated to invoke log-writer with structured inputs instead of inlining the format string. If the chronicle format ever changes, only log-writer needs updating.

Why this matters

Privacy defaults shipped. First-run /listen no longer risks committing raw email / Slack / transcript content to git. Cloud-sync gap is documented for users to address.
Autonomy slider actually moves the needle. Three high-traffic cortex commands now respect user-set autonomy preferences at their internal gates. auto mode for /cleanup lets scheduled cleanups run unattended; auto mode for /forget is hands-off for users who trust the operation.
Log format is now single-source-of-truth. Future format changes won't require touching 9 command files.

[4.7.1] — Karpathy patterns: memory/log.md chronicle + autonomy slider (2026-05-16)

Added — `memory/log.md` chronicle

New references/log-chronicle.md — formal spec for the unified append-only operations log at <config-root>/memory/log.md. Grep-friendly date prefixes (## [YYYY-MM-DD HH:MM] <op> | <summary>). Karpathy LLM-wiki log.md pattern.
New ## Log step appended to: /listen (Step 7.5), /morning (Step 4.5), /end-day (Step 5.7), /end-week (Step 5.7), /reindex (Step 5.5), /research-gaps (Step 4.5), /merge-research-draft (Step 4.5), /cleanup (Step 4.7), /rehearse (Step 4.5). Each appends one line at completion. No silent commands log; the log is for audit-worthy operations only.
CLAUDE.md storage-layout section updated to mention memory/log.md and link to the spec.
Optional retention controls via <config-root>/plugins/cortex.user-context.md (log_chronicle.max_entries, log_chronicle.archive_yearly).

Added — per-command autonomy slider

New references/autonomy.md — Karpathy Software 3.0 "Iron Man suit with autonomy sliders" pattern. Three modes per command: auto (no confirmation), suggest (default; suggest+confirm), confirm (extra-strict; confirm each material step).
Default settings tuned to operational risk: /recall, /search, /timeline, /reindex, /listen, /note, /diagnose → auto. /forget, /morning, /merge-research-draft, /lead-draft, /track-time, /generate-invoices, /setup-* → confirm. Everything else → suggest.
User override at <config-root>/plugins/cortex.user-context.md under autonomy: section. Per-command opt-in; users tune as trust develops.
nucleus-router skill (v0.1.3) reads the autonomy section before suggesting confirmations. auto mode runs the command directly with a one-line "Running /X now" note. confirm mode adds emphasis to the suggestion.

Why this matters

The log chronicle gives a grep-friendly answer to "what did I do on date X" across every audit-worthy operation. Closes the last LLM-wiki pattern gap.
The autonomy slider lets trust develop unevenly across capabilities. Commands you trust (/note, /recall) stop interrupting. Commands with blast radius (/lead-draft, /generate-invoices) stay gated.

[4.7.0] — Overnight learning: /listen + /morning + hot.md (2026-05-16)

Why this exists

Cortex was event-driven: passive observation during sessions, auto-commit at session close, mining triggered by /end-day. Nothing learned while the user slept. v4.7 adds the missing layer — a nightly ingest pipeline that pulls yesterday's substrate into an immutable archive, mines it autonomously, and stages proposals for a 2-minute morning review.

Inspired by Karpathy's LLM-wiki separation of raw/ (immutable source room) and wiki/ (LLM-owned knowledge), and by his hot.md rolling-buffer pattern for fast warm-start.

Added — `/listen` nightly ingest pipeline

New commands/listen.md + skills/listen/SKILL.md — unattended (no user gates).
Pulls yesterday's calendar (Calendar MCP), inbox metadata (Gmail MCP), Slack mentions + authored messages, Drive activity in watched folders, and transcripts via the configured note-source adapters (Granola / Gemini / Fireflies / Otter / generic-Drive).
Writes immutable substrate to <config-root>/archive/YYYY-MM-DD/ per references/archive-layout.md. Each day has calendar.md, inbox.md, slack.md, drive.md, transcripts/<id>-<slug>.md, and _index.md (counts + source health + errors). Privacy-conservative defaults: metadata-only for inbox / Slack; full bodies opt-in.
Runs transcript-reviewer, conversation-miner, activity-miner against the archive read-only. Stages all proposals to a single <config-root>/memory/.commit-drafts/YYYY-MM-DD.md file. Active memory is never modified.
Modes: default (yesterday), --date YYYY-MM-DD, --backfill N, --remine YYYY-MM-DD, --rewrite YYYY-MM-DD (rare; force re-pull).
Weekly retention tail (Sundays only) compresses archive directories older than 30 days into monthly tarballs. No silent deletion of >180-day tarballs.
Recommended cron registration via core-ops /register-schedules: 0 23 * * * (11pm local) or 0 5 * * 1-5 (5am weekdays).

Added — `/morning` interactive merge + brief handoff

New commands/morning.md + skills/morning/SKILL.md — the JARVIS morning routine.
Reads the latest .commit-drafts/ file. Renders an overnight summary (counts, sources status, proposal totals).
Walks proposals interactively: each gets accept / reject / edit / defer / skip-remaining. skim mode renders one-liners and offers accept-all-high, accept-all, select, reject-all shortcuts.
v4.4 drift detection runs on every knowledge-entry accept; conflicts prompt supersede / keep-both / skip.
Idempotent — re-running on a partially-walked draft picks up where the last session left off. Merged / rejected sections marked ~~struck out~~.
After merge, refreshes memory/index.md and memory/hot.md, archives the fully-resolved draft to .commit-drafts/archive/<date>-merged.md.
Optional handoff to /brief after merge: "Run /brief to start the day? (y/N)".

Added — `memory/hot.md` rolling 7-day context cache

New references/hot-cache.md — formal spec.
Sections: what I worked on (last 7 days), active people, active threads, recent commitments (to others / from others), recent reflections, recent decisions. Verbatim citations from existing nodes — no synthesis, no judgment, zero LLM cost.
Capped at 3000 words; truncates intelligently when over.
Configurable via <config-root>/plugins/cortex.user-context.md (hot_cache.enabled, window_days, word_cap, refresh_on).
Read first by /recall auto-fire at conversation start — every session opens warm.
Maintained by /listen (nightly), /morning (after merge), /end-day Step 5.6.
Graceful fallback: if hot.md is missing or disabled, /recall reverts to v4.6 behavior.

Changed — `/recall` Step 0 + `/end-day` Step 5.6

/recall auto-fire now loads memory/hot.md before memory/user.md (when present and enabled). Adds a pending-overnight-draft check that surfaces a one-line "Run /morning" hint if .commit-drafts/ has unmerged content.
/end-day quick chain gains Step 5.6 — refresh hot.md after index refresh, before close.

Why this matters

The system actually learns while the user sleeps. Wake up to a single 2-minute draft instead of a blank canvas.
Reproducible mining. If a mining proposal was wrong, re-run /listen --remine against the same archive with different settings. The Karpathy immutable-raw/ pattern.
Warm starts every session. /recall no longer cold-loads context from scratch. The hot cache makes the AI feel like it already knows what's been going on.

Coordinated with daily-brief v0.3.0+ and `/end-day` v4.6.0

/morning can chain directly into /brief after merge.
/end-day's ## Reflection writes feed hot.md via the cache refresh in Step 5.6.

[4.6.0] — /end-day cleanup: quick-default + --full opt-in (2026-05-16)

Why this exists

Real-user feedback: the v4.5 8-step /end-day chain felt like overhead on days without transcripts, inbox volume, or things to triage. The user gates fired even when nothing needed gating. This release reshapes /end-day so the default chain is the actionable spine and heavy work is opt-in.

Changed — `/end-day` defaults to quick mode

Quick mode (default). Runs Step 3 (cortex auto-commit with cheap-tier triage), Step 4 (reflective prompts), Step 5 (pre-stage tomorrow's brief), Step 5.5 (refresh memory index), Step 6 (close). 30s – 3 min.
Full mode (/end-day --full). Adds Step 1 (inbox triage), Step 2 (transcript review), Step 2a (mining of non-transcript sources), Step 2b (unified review gate) before the quick chain. The full chain from v4.5. 10-15 min.
Auto-offer prompt. In quick mode, before Step 3, run a fast pre-check counting today's transcripts and today's unread inbox where the user is To/Cc. If transcripts ≥ 2 OR inbox ≥ 5, surface a one-line prompt offering the full close. Default no after ~5s. If both are low, no prompt at all — no friction for nothing.
Step 4 reflection write target changed. Reflection answers now append to today's brief markdown as a ## Reflection section (was Section 7 in daily-brief v0.2.x; daily-brief v0.3.0 removed Section 7). Idempotent re-runs replace existing reflection content rather than duplicating. Tomorrow's /brief Section 6 reads from this section.
Step 5 inbox-triage handoff is now conditional. Full mode passes Step 1 results to tomorrow's brief; quick mode lets the brief query Gmail itself in the morning. No shared-state requirement in quick mode.

Why this matters

Most days are quiet days. The user gates on Steps 1 and 2 fired even when there was nothing to gate, training users to skim through empty motions. After v4.6 the chain auto-shrinks to the actionable spine and asks for more only when the data justifies it. Skipped-by-default steps stay available via --full for the days that need them.

Coordinated with daily-brief v0.3.0

This release lands alongside daily-brief v0.3.0, which removes Section 7 (end-of-day prompts) from /brief. The ## Reflection section written by /end-day Step 4 is the new contract between the two plugins — daily-brief reads it; cortex writes it.

[4.5.0] — Legibility upgrade: memory index + research-gaps + Obsidian (2026-05-16)

Why this exists

v4.3 (mining) made memory grow with the user. v4.4 (decay) made it forget gracefully. v4.5 makes memory legible and self-improving: a single auto-maintained catalog file gives a one-glance overview of the whole second brain, a new /research-gaps command actively finds and fills weak spots with web-sourced research (user-gated), and a new /setup-obsidian command makes <config-root>/ a graph-viewable, mobile-readable Obsidian vault.

Part of the "Nucleus as JARVIS" initiative (Phase 1 finish + Phase 2 human UI). See nucleus repo docs/proposals/cortex-v4.5-legibility.md and docs/proposals/obsidian-as-ui.md for the design rationale.

Added — `memory/index.md` (auto-maintained catalog)

New skills/indexer/SKILL.md — deterministic, zero-LLM file walker. Reads every memory node, extracts descriptor + latest [confirmed:...] date, classifies decay state, renders a grouped catalog (user / clients / people / companies / topics / domain / bizdev / system). Writes <config-root>/memory/index.md wholesale.
New commands/reindex.md — explicit invocation. Runs in seconds, no model calls, no node writes.
New references/memory-index.md — formal spec covering walk rules, classification formula, output template, and refresh triggers.
/end-day Step 5.5 — auto-runs the indexer after Step 5 brief pre-stage; clears .reindex-queue marker.
/cleanup Step 4.5 — auto-runs the indexer if any Step 4 action touched memory.
/remember Step 3.5 — appends to <config-root>/memory/.reindex-queue after writes (does NOT run indexer synchronously; defers to the next /end-day).
Storage-layout update in CLAUDE.md to include index.md, .reindex-queue, and .research-drafts/.

Added — `/research-gaps` autonomous gap-fill

New commands/research-gaps.md + skills/research-gaps/SKILL.md — active-maintenance loop complementing the v4.4 decay model. Scans <config-root>/memory/ for seven gap types (thin entity, stale fact in active rotation, contradiction within a node, orphan, under-cited high-confidence claim, decision gap, sparse domain) per references/gap-detection-rules.md. Renders ranked gap list; user picks per-gap actions (research-now / skip / mark-ok / archive-node / ask-me).
New agents/gap-researcher.md — subagent with WebSearch, WebFetch, file I/O scoped to <config-root>/memory/.research-drafts/ only. Enforces ≥2-independent-sources rule per claim. Enforces private-individual privacy rule (verifiable professional facts only — no home location, family details, real-estate, social media beyond official professional profiles, speculation). Returns confidence-aware findings (high / medium / low). Cap of 25K WebFetch tokens per run; default 5-gap cap.
New commands/merge-research-draft.md — interactive walk-through of the most recent .research-drafts/ file. Each finding: accept / reject / edit / defer / skip-remaining. Accepts apply REPLACE / MERGE / ADD / ARCHIVE per the proposed action and stamp [confirmed:today]. Rejects log to .research-skip-log.md for 90-day suppression. Drafts archive to .research-drafts/archive/ once fully resolved.
New references/gap-detection-rules.md — formal scanner rules with priority, scan procedure, and false-positive caveats.
/end-week Step 5.5 — optional invocation of /research-gaps in chained mode (top-5 by priority, no per-gap prompts unless user chooses "select").

Added — `/setup-obsidian` (config-only Obsidian vault scaffolding)

New commands/setup-obsidian.md + skills/setup-obsidian/SKILL.md — writes <config-root>/.obsidian/ workspace config (app.json, core-plugins.json, daily-notes.json) and a VAULT.md home page with Dataview-powered active-entity tables.
New references/obsidian-config-templates/ — bundled defaults for the four files written above.
Idempotent and non-destructive — existing .obsidian/ files and existing VAULT.md are preserved on re-run. Force-reset available via /setup-obsidian --reset with per-file confirmation.
Daily-brief integration is config-only. daily-notes.json points the daily-notes folder at <config-root>/briefs/ — daily-brief's existing markdown snapshots become Obsidian daily notes with zero plugin code change.
No external installs. The command does not install Obsidian or community plugins. It lists recommended community plugins (Dataview, Tasks, Calendar, Periodic Notes) for the user to install through Obsidian's UI.

Added — schema and trigger updates

CLAUDE.md storage-layout, auto-fire triggers, and slash-command tables updated for /reindex, /research-gaps, /merge-research-draft, /setup-obsidian.

Why this matters

Obsidian becomes a first-class human UI immediately. index.md + VAULT.md + the existing wikilink convention give graph view, mobile access, and daily-note integration with zero plugin code.
Non-cortex agents can read memory cold via index.md without loading memory-librarian — a one-file entry point for any future Nucleus plugin or one-off agent.
Active gap-filling complements passive decay. v4.4 forgets; v4.5 asks "what's missing?" and proposes user-gated answers.
Productization-ready. A new user can run /setup-identity → /setup-voice → /setup-obsidian and end up with a graph view of their own second brain in under five minutes.

[4.4.0] — Forgetting / decay layer (2026-05-12)

Why this exists

v4.3's mining layer ships the "perpetually learning" half of the second-brain vision. v4.4 ships the other half: forgetting, decay, and consolidation. Without it, cortex grows but never lets go — beliefs that get superseded silently linger, and "what do we know about X" returns mixed signal because the system can't tell fresh insight from stale.

Added — Decay model

Knowledge entries pass through four states based on the age of [confirmed:...]: Fresh → Stale → Dormant → Cold. State drives surfacing behavior but never deletes content.
<config-root>/memory/.decay-config.md auto-created on first v4.4 run with documented defaults: threshold_fresh: 60, threshold_dormant: 180, threshold_cold: 365 (days). Per-type modifiers (GOTCHA and RECIPE decay 1.5× slower; CORRECTION is immune). Per-node decay_profile: fast | normal | slow front-matter override stacks with type modifiers multiplicatively.
references/decay-model.md: full spec covering read-time decay (via /recall), event-time decay (via /cleanup and /rehearse), state transitions, threshold computation, and rationale.

Added — Decay-aware `/recall`

Every entry surfaced by /recall is now state-classified inline. Stale entries render with [stale-confidence]; Dormant with [dormant — last confirmed Nd ago]; Cold with stronger flag.
Recall-time triage offer at the end of explicit /recall runs (not auto-recall or contextual): re-confirm all / select per-entry / demote all dormant+cold / skip. Per-entry actions: confirm (update tag) / demote (move to ## Demoted knowledge) / edit-then-confirm / skip.
Demoted entries render after active sections in project-view and topic-view recall, de-emphasized and explicitly labeled.

Added — Concept-drift detection in `/remember`

Before writing a new INSIGHT / MODEL / GOTCHA / LESSON entry, a Haiku-tier classifier compares it against existing same-type entries on the same node (scoped to most-recent-20 by [confirmed:...] for cost).
If the classifier flags supersedes / contradicts / refines → FULL mode prompts the user: supersede (move old to Demoted knowledge, write new in its place) / keep both / edit relationship / skip new entry.
SILENT mode (auto-commit) never auto-supersedes — writes alongside and flags in changelog for review at next /recall or /rehearse. Silently demoting a held belief is too destructive for an unattended path.
On supersede: old entry moves to the node's ## Demoted knowledge section with metadata trail (↳ demoted <today> by supersede, ↳ superseded by: <new entry first 60 chars>). Tags preserved on the moved entry.
RECIPE excluded from the check (additive, not competing). CORRECTION already encodes supersede explicitly.

Added — `/rehearse` command and skill

New commands/rehearse.md and skills/rehearse/SKILL.md — active retention loop.
Selects 3-5 aging entries past their freshness threshold but not yet cold. Walks the user through each: confirm / update / demote / archive / skip.
Selection algorithm: composite score by age × type-weight, diversified across nodes. Entries from <config-root>/memory/.rehearse-queue.md (deferred by /cleanup) get priority.
Skip-log at <config-root>/memory/.rehearse-skip-log.md suppresses skipped entries for 30 days so they don't immediately resurface.
Default cadence: weekly via /end-week (new Step 3.5). Can run on demand any time.
Cost: zero model calls in steady state — date arithmetic + file edits.

Added — Decay-aware `memory-librarian`

Freshness multiplier on relevance ranking: Fresh 1.0, Stale 0.85, Dormant 0.6, Cold 0.3. Older entries sink, never hidden.
Skips ## Demoted knowledge sections by default. Reads them only when the parent skill explicitly requests historical context (e.g., "what did we used to think about X").
Surfaces aging in Confidence when > 30% of Source Entries are Dormant or Cold ("Most relevant memory on this is aging — consider /rehearse or fresh capture").

Added — `/cleanup` deepening

Section H expanded for person-page maintenance: cooling / dormant / cold-archive states drive concrete archive proposals. On accept, moves files from memory/person/<slug>.md to memory/person/archive/<slug>.md. /recall person:<slug> continues to find archived pages but flags them.
New section I — Dormant knowledge entries: scans all active nodes for entries past threshold_dormant. Per-entry suggestion: rehearse (defer to /rehearse via the queue file) / demote / archive. Cold entries surface separately with stronger flags. CORRECTIONs are excluded (immune to decay).

Added — `/end-week` chain integration

New Step 3.5 — Rehearse: invokes /rehearse between /review (Step 3) and reflective prompts (Step 4). Exits cleanly if the candidate pool is empty.

Added — CLAUDE.md schema updates

New Decay model section pointing at references/decay-model.md
New Demoted knowledge convention section documenting the per-node ## Demoted knowledge format
/rehearse added to auto-fire trigger table and slash command list

Why this matters

Now the second-brain learns AND forgets. /remember adds, /recall flags aging, /rehearse consolidates, /cleanup audits, drift detection prevents silent belief-flipping. The system carries less stale weight over time without ever silently deleting content — every transition is user-gated or logged, and demoted entries stay readable for as long as the node exists.

The combination of v4.3 (mining) + v4.4 (decay) is the bidirectional learning the user asked for: "perpetually updating and learning (and in some cases forgetting or moving things to the back of my memory similarly to how the brain is always learning)."

[4.3.0] — `/end-day` mining layer (2026-05-12)

Added — Mining layer (the main change)

/end-day now mines beyond the current chat session. Three read-only agents run in parallel at Step 2a after the existing transcript-review (Step 2):
- transcript-reviewer (expanded — was commitments-only): now returns TWO output streams — commitments_delta (unchanged) AND learnings_delta (decisions, insights, gotchas, models, relationship context, blockers, recipes, corrections). Reads notes from all configured sources via the new adapter pattern, not just Granola. Cross-source dedup merges the same meeting captured by multiple providers (e.g., Granola + Gemini) before extraction.
- conversation-miner (new): mines the user's other Cowork sessions in the time window. Excludes the current session, sessions already committed via /remember, sessions under ~4k tokens, and sessions tagged [no-mine]. Groups same-topic sessions before extraction so one insight surfacing in 3 sessions = 1 reinforced proposal, not 3 duplicates.
- activity-miner (new): mines CRM events (deal stage changes, lifecycle changes, task closures with outcome), sent email (user's own outbound decisions only), and calendar event metadata. Privacy guardrails: paraphrase always, hard-skip [CONFIDENTIAL] threads, no verbatim quotes of counterparties. Scoped to events, not extraction from CRM note bodies.
code-miner deferred to v4.4+ per user decision (low marginal value for consulting/ops-focused users).

Added — Unified review gate (Step 2b)

Merges all mining proposals across the three agents, dedupes against existing node content, groups by target_node, sorts by confidence DESC, update_type.
One review gate per /end-day run instead of three. Modes: accept-all / select-per-node / edit-each / high-confidence-only / skip-all.
Cross-refs auto-expand inline — when a proposal's cross_ref links to another proposal, the linked content renders as a "cross-ref →" line beneath it so the user reviews both in context.
New-node creation is explicit. Proposals with node_type: new require a confirmation step (with inline Scope-section interview) before any content is accepted into them. Prevents silent taxonomy bloat.
Dismissed proposals are logged at <config-root>/memory/dismissed-proposals.log (append-only, keyed by source-ref + content hash). Miners read this log at the head of every run and skip matching items within 7 days so dismissals don't re-surface tomorrow. After 7 days, re-surfacing is permitted.
30s gate timeout with skip-all default (longer than the 10s elsewhere because the gate has more density and the user may actually be reviewing).

Added — Source-agnostic note-source adapters

agents/lib/note-source-adapters.md: prompt-only adapters for granola, gemini (drive-folder and gmail-search methods), fireflies, otter, notion, drive-folder (generic), gmail-label (generic), custom (free-form). Each adapter documents config schema, tools used, fetch(time_window) logic, and health_check().
Adding a new provider is one new section in this file — no code change required.
<config-root>/plugins/cortex.note-sources.md: configured source list (markdown wrapper with a fenced YAML block). Per-source fields: id, provider, label, enabled, scope (global or project:<node-id>), config. Per-project scope lets a user run Fireflies only on one client and Granola everywhere else.

Added — Node-routing model

Scope section convention on domain nodes. Four fields: Topics:, Aliases:, What goes here:, What does NOT go here:. Used by miners to route extracted content correctly.
One-time migration in /end-day Pre-chain B: detects domain-shaped nodes (root-level .md files plus first-level nodes in non-reserved subdirectories), synthesizes Scope drafts from existing content, prompts user to accept/edit/skip per node. Open-ended "create new domain nodes?" step after. Marker file <config-root>/memory/.scope-migration-done prevents re-prompting. Re-run via /end-day --rerun-scope-migration.
Generic detection — no hardcoded node names anywhere. Migration scans whatever the user has.
references/node-routing.md and references/note-sources.md: docs covering the routing algorithm and source config schema.

Added — `/setup-sources` command

New commands/setup-sources.md and skills/setup-sources/SKILL.md — standalone interview to configure note sources.
Walks provider menu (Granola / Gemini / Fireflies / Otter / Notion / Drive-folder / Gmail-label / custom) and per-provider config questions.
Mandatory health-check on every new/updated source — adapter's health_check() runs during setup; failures surface verbatim with options to fix/disable/remove. Loud failure at setup, not silent at /end-day.
Supports --health-check-only and --remove <source-id> flags for maintenance.

Added — v4.4 decay substrate

[confirmed:YYYY-MM-DD] [recalled:YYYY-MM-DD] tags on every knowledge entry. confirmed: updates when an entry is re-affirmed or referenced as evidence. recalled: updates when /recall surfaces the entry to the user. Pre-v4.3 entries are treated as if both default to the original commit date.
/recall updates the recalled: tag on every entry it renders. Backfills tags on pre-v4.3 entries as it touches them.
memory-librarian returns tag values in Source Entries so callers can update recalled: after consuming the list. Agent stays strictly read-only.
v4.3 maintains the tags but does not yet decay or demote. v4.4 reads them.

Changed

/remember Step 3 C: knowledge-entry formats updated to include the [confirmed:...] [recalled:...] tag suffix
/end-day Step 3: now accepts pre-routed accepted-proposals as a second input alongside the current-session content. Accepted proposals bypass the Haiku triage (the user just accepted them).
skills/end-day/SKILL.md description rewritten to reflect the mining layer

Why this matters

The earlier /end-day only auto-committed the current Cowork chat session. Most of the user's day happens elsewhere — Granola meetings, other Cowork sessions, CRM activity, sent email. None of that flowed into the right cortex nodes. The mining layer closes that gap with three read-only agents that route extracted content through a single unified review gate, scoped to actually-changed events and source-agnostic notes via the adapter pattern.

The [confirmed/recalled] tags ship as substrate for v4.4's forgetting/decay layer (coming next).

v4.3 → v4.4 sequencing

v4.4 will be the forgetting half: decay weights on retrieval, concept-drift detection on contradicting entries, auto-archive of dormant person pages, periodic "still true?" rehearsal prompts. Builds on the substrate that lands in v4.3.

[4.2.0] — Second-brain v2 Phases 3-6 (2026-05-12)

Added — Person pages (Phase 3)

memory/person/<firstname-lastname>.md — new directory and schema for canonical per-contact pages. Identity / Relationship / Recent interactions / Open threads / Notes / Linked entities. Schema documented in CLAUDE.md.
Usage-graduated, not mention-graduated. Six triggers create a page (contact-researcher dossier, 3+ recalls, project-setup primary contact, time-log billing role, explicit [ENTITY:person] tag, 3+ calendar meetings in 30 days). Casual mentions stay in project-node PEOPLE indices.
memory/.person-recall-counter.json — JSON map tracking recall count per slug. After 3 recalls without a page, /recall offers to graduate.
/recall person:<slug> — new query form renders the full person page when it exists; falls back to legacy cross-project profile when it doesn't.
/remember Step 3 D.1 person-page graduation logic — detects triggers, creates pages with pre-filled content from the conversation, or appends Recent interactions to existing pages. Never overwrites Identity / Notes / Linked entities without explicit confirmation.
memory-librarian agent — now checks memory/person/<slug>.md first for person-shaped queries; falls back to project-node ## People sections only if no page exists.
DASHBOARD.md — new ## Active People section (top 10 graduated pages, sorted by Last updated).

Added — Cheap-tier commit triage (Phase 4)

/remember Step 0 — Haiku-class classifier decides commit: true | false plus affected node list before any Sonnet synthesis runs. Trivial conversations cost ~$0.001 (classifier only); substantive conversations skip the classifier overhead and proceed normally.
memory/triage-log.md — append-only audit log of triage decisions. User reviews weekly for false-negatives.
skills/observe/SKILL.md — auto-commit flush now always runs /remember Step 0. CORRECTION-type observations bypass the classifier and always commit to user.md.
Bypass rules — explicit /remember <node-id> invocations and conversations with [ENTITY:person] tags skip Step 0 (user is asserting commit-worthiness).

Changed — `/end-day` orchestration chain (Phase 5)

Rewrote commands/end-day.md as a five-step chain with user gates:
1. Inbox triage for tomorrow (delegates to inbox-triage plugin if installed; falls back to lighter Gmail search otherwise)
2. Transcript-reviewer agent surfaces uncaptured commitments → per-item user gate to convert to CRM task / cortex P0 / skip
3. Cortex auto-commit with Phase 4 cheap-tier triage
4. Three reflective prompts (biggest done / blockers / one thing to move tomorrow) — answers also append to today's brief markdown snapshot
5. Pre-stage tomorrow's brief artifact via daily-brief plugin (skipped if not installed)
Default-wait on gates. If no user response within ~10s, conservative default is chosen. The chain never blocks.

Added — Guardrails (Phase 6)

/cleanup orphan / isolated-note detection — section G in the cleanup audit. Flags nodes with no incoming/outgoing links and last updated > 30 days ago. Per-orphan suggestion: archive / merge into candidate / keep standalone. Surfaces in DASHBOARD.md's new ## Isolated Notes section.
/cleanup person-page maintenance — section H. Flags dormant (no contact 12+ months), stale-interaction (entries > 90 days that haven't been archived to the page's archive section), and premature-graduation pages.
/recall duplicate-topic surfacing — Haiku-tier semantic match on topic queries. If an existing node summary is plausibly about the same topic, surface "Note: <node> may already cover this. Read that first?" before the regular topic answer.

Why this matters

Phases 3-6 of SECOND-BRAIN-V2-SPEC. Brings cortex from project-state memory to a relationship-and-cost-aware knowledge graph. Person pages make "who is this in 5 seconds before a meeting" possible. Cheap-tier triage keeps per-commit cost predictable as commit volume grows. /end-day ties the day's work into a single ritual. Guardrails prevent silent drift into noise.

Phase 2 (inbox-triage as a separate plugin) was deliberately skipped — daily-brief's built-in Gmail fallback for section 2 is sufficient for now.

[4.1.3] — Platform-agnostic Step 0 (2026-05-12)

Changed

/setup-identity and /setup-voice Step 0 instructions are now platform-agnostic. Every request_cowork_directory(...) call is wrapped in a conditional: "In Cowork, call request_cowork_directory(...). In Claude Code (or any environment with direct filesystem access), no mount is needed." This lets the same plugin source serve both Cowork and Claude Code without two divergent code paths.

Why this matters

Phase 0 of SECOND-BRAIN-V2-SPEC. Removes the implicit Cowork-only assumption that was forcing Claude Code users to debug mount calls that don't exist in their runtime. Future plugins (daily-brief, end-day, etc.) inherit this convention.

Fixed

.claude/commands/remember.md now includes v4 features: silent mode, user observation extraction, user node writes, and dashboard template (parity with commands/remember.md).

[4.1.0] — Config-root awareness

Added

/setup-identity and /setup-voice now honor ~/Documents/.claude-plugin-config-root, a single-line text pointer file at the user-level home that records the user-chosen plugin config root (set by any marketplace plugin's first-time setup, including these two commands). When the pointer exists, identity and voice files are written to <config-root>/identity.md and <config-root>/voice.md respectively. When the pointer does not exist, both commands either fall back to a pre-existing legacy default at ~/Documents/Claude/ or prompt the user to pick a config root.
Step 0 added to both commands: resolve the canonical file path before any read or write. Documented variables <identity-path> and <voice-path> for downstream references inside each command body.

Why this matters

The other plugins in this marketplace previously tried to write per-plugin config to their own folder (read-only under Cowork's mount), which failed silently. The refactor across all plugins centralizes user-writable config under a user-chosen folder. Cortex was already writing to a writable user-level path, but adopting the same pointer means cortex's identity and voice files live alongside the other plugins' per-plugin configs when the user picks a non-default config root — and the convention is generic enough that any user (not just the original maintainer) can install or fork this marketplace.

[4.0.0] — Always-On Learning

Added

Passive observation engine (commands/observe.md, skills/observe/SKILL.md) — silently learns user preferences, corrections, domain knowledge, and relationship context during every conversation. Never interrupts. Adapts in real-time. Flushes observations to memory at conversation end.
User profile node (user.md) — persistent model of the user: communication preferences, working style, corrections, domain expertise, relationships, tool preferences. Carries across all projects and both platforms.
Auto-recall at conversation start — loads user profile silently, checks for overdue P0s and stale threads, surfaces attention items in <=8 lines.
Auto-commit at conversation end — detects farewell signals, silently commits decisions, knowledge, and observations. Skips trivial conversations.
Contextual recall mid-conversation — when user mentions a known project/person/topic, surfaces 1-3 relevant knowledge entries naturally (no recall block).
Silent mode for /remember — auto-triggered commits produce no output unless creating a new node.
Claude Code full support (claude-code/INSTRUCTIONS.md, claude-code/hooks.json) — drop-in CLAUDE.md instructions and optional hooks for auto-recall/auto-commit in Claude Code.
Per-project config (.cortex.json, cortex.config.md) — control capture aggressiveness (aggressive/normal/minimal), toggle auto behaviors, set default node, override memory path per project.
Dashboard template — DASHBOARD.md now has a defined format (table-based Active Nodes, P0 list, Waiting On, Recent Knowledge, Stale Threads, Dormant Nodes).

Changed

skills/remember/SKILL.md — now auto-fires on farewell signals (conversation end), runs silent extraction, flushes user observations.
skills/recall/SKILL.md — now auto-fires on conversation start and contextual mid-conversation mentions.
skills/search/SKILL.md — disambiguated trigger phrases from /recall (search is cross-project; recall is single-node context).
skills/timeline/SKILL.md — removed "weekly review" trigger (routes to /review instead).
commands/remember.md — added user observation extraction (Section D), user node write step, confidence gating, silent mode, dashboard template.
commands/recall.md — added Step 0 (always load user profile), auto-recall attention summary, contextual recall section.
commands/cleanup.md — aligned staleness thresholds to 4-tier system (7/14/30 days) matching recall.md.
All 9 command files — platform-aware directory access (Cowork request_cowork_directory vs Claude Code direct filesystem).
CLAUDE.md — updated to v4 with always-on behaviors, user profile, per-project config.
plugin.json — v4.0.0, updated description, added platforms field.

[3.0.1]

Added

MIT LICENSE file (matches plugin.json license field).
CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md.
CHANGELOG.md as the canonical version history (README changelog remains a summary).
GitHub issue templates (bug report, feature request) and pull request template.
docs/ARCHITECTURE.md describing command/skill flow and memory layout.
CI workflow Validate — JSON checks for plugin.json, frontmatter checks for command/skill markdown (and .claude/commands when present).
scripts/check_repo.py used by CI.

[3.0.0]

Added

File-based storage — memory now persists to ~/Documents/Claude/memory/ as markdown files.
Two-tier structure: DASHBOARD.md for fast orientation + individual node files for detail.
Node-to-file mapping: client:acme-corp → memory/client/acme-corp.md.
Directories created dynamically from node prefixes — any prefix is valid.
Every write operation updates both the node file and the dashboard.
Archive support: /forget --archive moves files to memory/archive/.
/cleanup now audits actual files on disk, detects orphaned entries and missing files.
All 9 commands updated with explicit storage instructions.

[2.2.0]

Changed

Business-operator friendly — all examples, taxonomy, and language updated for general business use (not just technical projects).
Node conventions now lead with client work, bizdev, strategy, and ops.
Added strategy/planning node type (strategy:q2-growth, strategy:pricing).
All knowledge examples rewritten for business contexts (proposals, procurement, pricing, onboarding, retention).

[2.1.0]

Added

Knowledge-first redesign — memory now captures insights, lessons, mental models, gotchas, recipes, and corrected beliefs as first-class entry types.
/learn — quick knowledge capture without full session extraction.
/note — one-liner capture for quick facts.
/review — synthesized weekly digest with "Learned" as a prominent section.

Changed

/recall now surfaces knowledge entries prominently (not buried under project state).
/recall [topic] — topic-based knowledge recall across all projects.
/search redesigned to prioritize knowledge entries for "how/what/why" queries.
/remember extraction expanded to cover both project state and knowledge categories.
Knowledge entries preserved longer than logs during memory consolidation.

[2.0.0]

Added

/search, /forget, /timeline, /cleanup commands.
Priority system, staleness tracking, blocker tracking, people index.
Genericized taxonomy, cross-project signals.

[1.0.0]

Added

Initial release with /remember and /recall.

Changelog

[4.28.0] — Nightly-listen loud-failure hardening (2026-09-18)

Fixed

Added

[4.27.0] — /listen mines the daily brief; /end-day and /morning stop depending on it (2026-09-17)

Added

Changed

[4.26.0] — Eval suite: onboarding + morning natural-language cases (2026-09-15)

Added

[4.25.0] — Phase 3 consistency fixes (2026-09-15)

Fixed

[4.24.0] — /setup-voice moved to Comms Desk (2026-09-15)

Removed

Changed

[4.23.0] — /review and /timeline moved to briefing (2026-09-15)

Removed

Changed

[4.22.0] — local dir rename + cross-plugin reference sync (2026-09-15)

Changed

[4.21.4] — operating-model architecture hardening (2026-09-15)

Added

Changed

[4.21.3] — Codex adapter synchronization (2026-09-15)

Fixed

[4.21.2] — claude plugin eval suite (2026-09-15)

Added

[4.21.1] — Fix: /morning should stay model-invocable (2026-09-15)

Fixed

[4.21.0] — /start-nucleus gains autonomy acknowledgment, catalog fixed (2026-09-15)

Added

Changed

Not done

[4.20.1] — Skill auto-invocation audit (2026-09-15)

Changed

[4.20.0] — /note and /learn folded into /remember (2026-09-15)

Added

Changed

[4.19.0] — transcript-reviewer + conversation-miner + activity-miner merged into note-taker (2026-09-15)

Changed

Removed

[4.18.6] — Autonomy docs updated for nucleus-router retirement (2026-09-15)

Changed

[Unreleased]

[4.18.5] — writing-style renamed to voice (2026-09-15)

Changed

[4.18.4] — /morning runs the Friday rehearsal batch (2026-09-15)

Added

Changed

[4.18.3] — /morning commits + pushes memory-as-git (2026-09-15)

Added

Changed

[4.18.2] — Follow-up path/privacy fixes (2026-09-15)

Fixed

[4.18.1] — /migrate-scopes-v2 command (2026-09-15)

Added

[4.18.0] — identity.md/voice.md move to memory/me/ (2026-09-15)

Changed

[4.17.1] — Genericize hardcoded proprietary examples (2026-09-15)

Fixed

[4.17.0] — /recall load boundary + cap enforcement (2026-09-15)

Added

Changed

[4.16.0] — /end-day demoted to optional; /morning seeds today's brief (2026-09-15)

Added

Changed

[4.15.0] — ChatGPT Work and organization distribution (2026-09-14)

Added

Changed

Security

[4.14.0] — Portability and stabilization refactor (2026-09-14)

Added

Changed

Compatibility

[4.13.2] — /end-day brief-state paste path (2026-07-07)

Added — Step 2c.0p paste path

Changed

[4.13.1] — Daily-Brief Improvement Spec companion: brief-state fallback gate + doc fixes (2026-07-03)

Fixed — /end-day can now actually read brief state (D2)

Added — reprioritize write-back (Step 2c.1a, D3)

Fixed — Step 0.7 cost-card $-placeholder corruption (P1)

[4.24.0] — `/setup-voice` moved to Comms Desk (2026-09-15)

[4.23.0] — `/review` and `/timeline` moved to briefing (2026-09-15)

[4.13.2] — `/end-day` brief-state paste path (2026-07-07)

Fixed — `/end-day` can now actually read brief state (D2)

Fixed — Step 0.7 cost-card `$`-placeholder corruption (P1)

Added — `/cleanup` Section L.0 first-run baseline-stamp

Added — `/cleanup` skip-log enum constraints

Changed — `/sync-linked-entities` mtime check refinement

Added — `/sync-linked-entities` command + skill

Added — DASHBOARD line provenance (`/remember` Step 3 + `/cleanup` Section L)

Changed — `/end-day` Step 5 artifact consistency (workstream/nucleus-improvements observation)

Changed — `/setup-voice` Step 3.5 + `/setup-identity` Step 3.7: graph completion (workstream/nucleus-improvements observation)

Added — `/remember` Step 1 consults the taxonomy

Added — `/remember` Step 3.4 orphan warning for new nodes

Added — `/cleanup` Section K — structurally-isolated nodes

Added — `/end-day` Step 3.7: passive person-page graduation

Added — `/relink-memory` retroactive command

Added — `workstream/` node type

Added — `DECISION` knowledge entry type

Changed — unified `memory/staged/` tree