# Changelog --- ## 2026-Jul-01 (v3.10.1) ### Added — Notification workflows (zero-SMTP, GitHub-native delivery) - **Source repo:** `.github/workflows/notify-collaborators.yml` — on every push to `main`, posts a diff summary as a comment on a tracking issue. Anyone subscribed to that issue gets a GitHub notification + email. One-time setup: create the tracking issue, pin it, tag collaborators. - **Mirror repo:** `scripts/mirror-workflows/release-announce.yml` — staged into the mirror's `.github/workflows/` by `release.yml` on every tag. Fires on `release: published` and posts the release body as a new GitHub Discussion under an "Announcements" category. Users opt in via Watch -> Custom -> Releases on the mirror, OR Subscribe on the Announcements category. Delivers via GitHub's notification system; no SMTP secrets needed. ### Changed — Stripped all `DUG_Malampaya_Prime_2022` references Removed project-name mentions and reference-output tables from `index.html` (Seismic Tools panel and dug_projdb panel), `skills/dug_projdb/SKILL.md` (intro born-from line + section 7 reference values, sections renumbered), `docs/Seismic-Tools.md` (Reference Output block, version-history parenthetical), `CHANGELOG.md` (Validated-against sentence). Provenance now stays in commit history rather than user-facing docs. ### Fixed — Duplicate `
` panel in `index.html` A hand-added panel duplicated the one already produced by `scripts/build-index-html.py`. Removed the hand-added duplicate; kept the build-script-generated panel. --- ## 2026-Jun-29 (v3.10.0) · New `seismic` family — `dug_projdb` skill ### Added — `dug_projdb` skill (Tier-1, family `seismic`) First non-DOCX tools family in Coeus. Reverse-engineers a **DUG Insight `project.dugprj`** (undocumented SQLite database, schema v16.x) into a multi-sheet xlsx inventory: - **Volumes** — every type-13 product with file path, 3D/2D classification, trace_count, samples, revision history, derived-workflow counts. - **Volume_Processes** — one row per (volume, derived process). ProcessType label + **full parameter spec** from `AttributeRevision.attributeType=528` (BandpassFilter `filterType=Highpass`, VolumeMaths formulas, VelocityConversion `velocityVolumeType=vintt`, etc.). - **Process_Catalog** — deduplicated catalogue of distinct derived processes. - **Horizons** (gridded + markers) — productType 12 + 8. - **Polygons** (interpreted + shapefiles) — productType 18 + 37. - **ProductType_Map** — inferred category mapping with row counts. - **UUID_Index** — every UUID in parameter strings resolved to a human-readable product name via two's-complement int64 → 128-bit reconstruction of `productUUIDMSB/LSB`. **Pipeline (three scripts in `scripts/`, run in order, each editing the xlsx in place):** 1. `01_build_inventory.py` — workbook scaffold + main sheets. 2. `02_extract_parameters.py` — enrich Volume_Processes with attrType=528 params; add Process_Catalog. 3. `03_resolve_uuids.py` — replace UUIDs in params with `«name» (kind/pidN)`; add UUID_Index. All three accept `--db ` and `--out `. Dependencies: `openpyxl` only. ### Added — `seismic` family + `Seismic Tools` website tab - Introduced the third Coeus intent family: **`seismic`** (upstream O&G subsurface project artefacts). Sits alongside `decision`, `tools`, and `meta`. - `coeus-router` updated: Step 1 family table now includes seismic; new Step 2c (`Skill Within seismic`); new tie-breaker rule for `dug_projdb` vs `project-lifecycle`. - `skills/_shared/SKILL_REGISTRY.md` — new row for `dug_projdb` (`seismic`, Tier 1, ~2,137 tokens, well under the 3,000-token cap). Families section expanded to three. ### Changed — Website navigation - **Tools** tab renamed to **Office Tools** to make the family split explicit. - New **Seismic Tools** tab added — landing page at `docs/Seismic-Tools.md` with one-skill bundle (more seismic skills planned). - `scripts/build-index-html.py` — updated SECTIONS, MAIN_TABS, SUB_PARENT, SUB_LABEL, MD2TAB to wire both tabs and their sub-pages (`seismic-dug-projdb`). - `index.html` regenerated. ### Updated - `README.md` — Skills-at-a-Glance and Slash Commands tables include `dug_projdb`. - `docs/Tools.md` — title is now "Office Tools — OOXML & Project Lifecycle"; cross-link to Seismic Tools. - `.claude-plugin/plugin.json` — version 3.9.1 → 3.10.0; description updated to mention the seismic family. ### Honest notes — kept as in-skill caveats - DUG schema is **undocumented**. Every productType→category mapping is empirical inference, verified against the reference project. The skill flags this prominently in the workbook README sheet and the ProductType_Map sheet. - Types 29 (1,275) and 30 (2,393) remain unidentified in the reference project. - Some processes are baked into volume names ("spectral shaping", "curvature_max") rather than as separate Volume_Processes rows. - 2 of 42 UUIDs in the reference run are unresolved — they're horizon-property IDs, not Products. Kept verbatim rather than silently dropped. --- ## 2026-Jun-29 (v3.9.1) · EP-Council Skill v1.7 — Trap T10 Duration Drift ### Added — T10 Duration Drift (Shell) New 11th trap in the Council Trap Screen, catching the *subtractive* portfolio drift T4 misses. Fires when: - Distributions are funded before/instead of the best organic reinvestment - R/P (reserve life) shortens materially faster than production - M&A becomes the primary reserve-replacement mechanism without a working organic engine (acquired duration ≠ created duration) Sits alongside T4 (Complexity Creep) under Shell — T4 catches *additive* drift, T10 catches *subtractive* drift. A major can pass T4 and still fail T10. ### Changed — Shell council member lens extended Shell's lens now explicitly covers duration vs distribution discipline (reserve life, organic renewal, harvest-vs-compounder framing) alongside the existing portfolio-complexity lens. Primary question and red-flag list extended to cover both traps. ### Updated — All T0–T9 / "10 traps" references to T0–T10 / "11 traps" - `skills/ep-council/SKILL.md` (v1.6 → v1.7, footer, three trap-screen call-sites, trap table) - `docs/EP-Council-Trap-Screen.md` (title, intro, trap table, walkthrough — plus new T10 explainer section mirroring the T9 pattern) - `docs/EP-Council.md`, `docs/EP-Council-Council-Members.md`, `docs/EP-Council-Walkthrough.md`, `docs/LLM-Council.md` - `README.md` (trap pipeline, trap table, footer string) - `index.html` (Overview Skill Reference table + EP-Council panel + Council Members panel + Trap Screen panel, including new T10 explainer matching the T9 pattern) ### Skipped (per maintainer note) Separating the "harvest-vehicle vs compounder" multiple question into its own trap. Reason: the harvest-vs-compounder framing is currently covered inside T10's lens; a separate trap would only be warranted by a non-Shell case study. --- ## 2026-Jun-26 (v3.9.0 follow-up) ### Added — One-line installer (`scripts/install-coeus.ps1`) Closes the three-surface install pain (Code + Cowork + Desktop chat) with a single PowerShell command. The three Claude surfaces don't share a plugin runtime, so the script handles each separately in one pass: - **Claude Code (CLI)** — prints the 3 `/plugin` slash commands to paste inside a `claude` session. - **Cowork** — downloads every `.skill` from the latest mirror release into `Downloads\Coeus-skills\` and opens the folder in Explorer for drag-into-chat. - **Claude Desktop chat** — downloads every `.paste.md` bundle into the same folder for copy-paste-into-chat. Usage: `iwr https://raw.githubusercontent.com/keithceh/Coeus-plugin/main/install-coeus.ps1 | iex` The release workflow copies the installer to the mirror root on every tag so the raw URL stays stable. ### Fixed — Release skill swap clobbered `check_skills()` Releases v3.9.0-pre runs failed with `caveman-local-backup: frontmatter name='caveman' != dir`. The `Bundle coeus.plugin` step backs up the local skills before swapping in the upstream copies, but the backup paths (`skills/caveman-local-backup`, `skills/prompt-master-local-backup`) lived inside `skills/`. `check_skills()` scans every `skills/*` dir, so the backups were treated as broken skills (name mismatch, missing from registry, over token cap). Moved backups to `/tmp/coeus-backup/` outside `skills/`. ### Fixed — `.claude-plugin/plugin.json` version drift caught by tag assertion First v3.9.0 tag failed the `Assert plugin.json version matches release tag` guard because `plugin.json` was still on `3.8.0`. Bumped to `3.9.0`. The guard itself is working as intended — this was a process miss, not a CI bug. ### Updated — README Installation rewritten "for dummies" - New **TL;DR** at the top: one PowerShell line installs across all three surfaces. - New explainer: why a single click isn't structurally possible (no shared plugin runtime). - New **Option 1 — One-line installer** with step-by-step instructions (open PowerShell, paste, follow on-screen prompts) and the SSH-host-key gotcha pre-fixed inline. - Existing manual paths renumbered (3→4, 4→5, 5→6, 6→7); Option 2 retained as "manual marketplace install" for Code-only users who want to skip the script. - Install-order table moved below the first two options so dummies see the easy path before the dependency graph. --- ## 2026-Jun-24 (v3.8.0) · 2026-Jun-24 ### Fixed — CI release workflow hardening (3 commits before v3.9.0 tag) First push of v3.9.0 surfaced three CI failures, each fixed independently before the tag could publish: - **`release.yml` step name unquoted colon-space** — `Bundle coeus.plugin (single source of truth: scripts/build-plugin.py)` parsed as a YAML mapping. Quoted the name. Same regression class as v3.3.1's `caveman.md` description fix. - **`release.yml` missing PyYAML** — `build-plugin.py`'s `check_skills()` does `import yaml`; release runner had no setup-python or pip install. Added `setup-python@v6` + `pip install pyyaml` before the bundle step. - **`release.yml` inline upstream swap missing normalization** — `Bundle coeus.plugin` swaps in upstream `caveman` + `prompt-master`, then runs `check_skills()`. Upstream `caveman` ships as `name: caveman-protocol`; invariant #3 (name matches dir) hard-failed. Added inline `sed` normalization that mirrors `sync-upstream-skills.yml`'s step (pin `name:`, re-inject `argument-hint:`). Closes the "release-time pull bypasses the weekly normalization" gap. ### Bumped — GitHub Actions to silence Node 20 deprecation - `actions/checkout@v4` → `@v5` (test.yml, sync-upstream-skills.yml — release.yml was already on v5) - `actions/setup-python@v5` → `@v6` (release.yml, test.yml) ### Added — Paste-prompt install path (tier-2 B2 — web-chat fallback) claude.ai web chat does not support plugin uploads. For web-chat users, generate a self-contained paste bundle per skill: - **`scripts/build-skill-paste.py`** (new) — takes a skill name; strips the SKILL.md frontmatter; finds every `[..](../_shared/X.md)` reference and inlines the content; follows one hop of nested shared refs; writes `dist/coeus-.paste.md` with a paste-prelude that tells the model to treat the embedded protocol as load-bearing instructions. - Built for 6 owned skills at release time: llm-council (19.5 KB), the-architect (16.8 KB), ep-council (14.8 KB), morpheus (12.2 KB), plugin-creator (14.6 KB), project-lifecycle (37.9 KB). - **`release.yml`** extended — builds all 6 paste bundles after the per-skill `.skill` build and ships them as release assets alongside `coeus.plugin`. - **README Option 5** documents the use: download the bundle, paste as first chat message, ask question in next message. Loses slash commands + auto-fire + cross-skill chaining; preserves the protocol. ### Added — Org / Enterprise managed install path (tier-3 C1) The most durable Anthropic-supported install path — survives uninstall attempts ([#45323](https://github.com/anthropics/claude-code/issues/45323)) and bypasses all Personal-upload state bugs. **README Option 4** documents the admin setup: managed `claude-config.json` with `extraKnownMarketplaces` referencing `keithceh/Coeus-plugin` + `enabledPlugins.coeus@coeus: "required"` (auto-install, cannot uninstall) or `"installed by default"` (auto-install, can opt out). Member experience: zero install steps; `[org-managed]` marker in `/plugin list`. ### Updated — README install options (5 → 6, reordered by durability + reach) 1. **Marketplace** (recommended, persists across restart) 2. Personal-upload (legacy fallback — Anthropic state bugs) 3. Individual `.skill` files from Releases 4. **Org / Enterprise managed install** (new — most durable) 5. **Paste-prompt for web chat** (new — only path that works on claude.ai) 6. Build locally (maintainers / forks) Now covers every Claude surface — Cowork / Code / Desktop / claude.ai web chat / Team-Enterprise org — with the durability-appropriate path documented for each. ### Added — Marketplace install path (option B from the architect diagnostic) Closes the Anthropic-side Personal-upload friction (no-restart-persistence, dirty-disable-state bugs #40600 / #28554 / #63624) without making the source repo public. - **`scripts/marketplace.json.template`** (new) — canonical `marketplace.json` shipped INTO the public mirror as `.claude-plugin/marketplace.json`. References the public mirror repo `keithceh/Coeus-plugin` with `source: github` + `repo` fields per the Anthropic schema. - **`.github/workflows/release.yml`** extended with a `Publish bundle to public mirror` step. On every tag, the workflow extracts `dist/coeus.plugin`, drops the marketplace template + a public-facing README, force-pushes to `keithceh/Coeus-plugin` (history-less, source-free). Skipped with a clear bootstrap message if `COEUS_MIRROR_PAT` secret is not set. - **README install section reordered** — marketplace is now **Option 1 (recommended)**; Personal-upload demoted to **Option 2 (legacy fallback)** with a callout linking to the relevant Anthropic bug numbers; Releases (`.skill` files) is Option 3; Build-locally is Option 4. - **Bootstrap (one-time, manual):** user creates empty public `keithceh/Coeus-plugin` on GitHub, generates PAT with `repo` scope, adds repo secret `COEUS_MIRROR_PAT`. After that, every `git tag vX.Y.Z && git push --tags` syncs the mirror. ### Why this persists The bug class the marketplace path closes is **on Anthropic's side**; the repo can't fix it directly. Marketplace install is the path Anthropic actually maintains — persists across restart, survives uninstall+reinstall cleanly, auto-updates. The bundle is history-less and re-pushed on every tag, so the mirror state is always a deterministic function of the latest release. The mirror repo has no source — only the runtime — so BSL 1.1 is preserved. ### Added — `docs/Project-Lifecycle.md` + own landing-page tab (5th main tab) - `docs/Project-Lifecycle.md` — full reference page mirroring the EP-Council pattern: what it does, install, triggers table (with sub-function mapping), pipeline diagram, the five sub-functions, the three core files (handover / artefacts_index / telemetry log), cadence options table, telemetry §8 spec, session-close protocol, mid-session commands, when-to-use vs alternatives, version history (v1.0 → v1.3.0). - `index.html` landing page extended from 4 to **5 main tabs**: Overview / EP-Council / LLM-Council / Tools / **Project-Lifecycle**. Tab carries one sub-panel (the project-lifecycle SKILL.md) plus the new `Project-Lifecycle` wiki-style cross-link target. ### Fixed — `lifecycle_close.md` drift from v1.3.0 single-log model - Operator checklist line updated: `Outputs/_telemetry/log.md` **appended** with session-close entry (never a new timestamped file). Was stale from v1.2.1 per-update design. - Step-7 comment updated: telemetry history lives in `Outputs/_telemetry/log.md`, not in per-update notes. - Checklist header bumped to v1.3. ### Verified — stress test all impacted skills (3 invariants) - check_skills PASS: 12 skills, 0 warnings. - Orphan-skill injection → exit 1. - 1100-char description injection → exit 1. - All recovered to CHECK PASS after cleanup. - Full harness: 162/162 (no plugin) and 182/182 (with plugin). - Plugin rebuilt clean (UTF-8-flagged, POSIX-permed, every skill present). ### Added — `project-lifecycle` v1.3.0 (three core files, create-once-update-in-place) Followed the v1.2 → v1.2.1 council pass. Replaces per-update timestamped files and the `_telemetry/index.md` rollup with a single append-only `_telemetry/log.md`. - **Three core files** created once at Sub-K (kickoff), updated in place thereafter: `_LLM_Handover.md`, `Outputs/artefacts_index.md`, `Outputs/_telemetry/log.md`. No timestamped duplicates anywhere. - **`_telemetry/log.md`** is append-only — every cadence-tick / audit / close appends a new timestamped section at the bottom. Prior sections are immutable history. Header tracks `Last appended` + `Total updates`. - **`_telemetry/index.md` rollup file deleted** — an append-only chronological log is its own index. Closes FM-L2-03 by construction. - **Cadence default = `end-of-session`** (silent, no 6-option menu at kickoff; user overrides anytime). Closes FM-L2-01 kickoff friction. - SKILL.md token estimate: ~2,935 → ~1,853 (post-refactor). ### Added — `project-lifecycle` v1.2: housekeeping-first close + artefacts index + cadence The Sub-D (Session Close) protocol now runs **housekeeping before handover writes**, so the artefacts index never references files about to be deleted. - **`_shared/lifecycle_artefacts.md`** (new, 220 lines) — full spec for §4a Inputs / §4b Outputs in the handover, the `Outputs/artefacts_index.md` single-source-of-truth index (clickable, three buckets), per-update telemetry notes (`Outputs/_telemetry/_.md`), and the `Outputs/_telemetry/index.md` rollup. Closes FM-L2-03 (telemetry orphan). - **`_shared/lifecycle_frequency.md`** (new, 98 lines) — handover update-cadence options (per-action / ultradian / end-of-session / daily / milestone / custom) with evidence basis. Referenced by handover §9. - **`_shared/lifecycle_close.md`** rewritten as v1.2 — 11-step close protocol with HOUSEKEEPING FIRST as step 1; updated 13-item operator checklist. - **`_shared/lifecycle_handover.md`** expanded — new §4a Inputs / §4b Outputs / §9 Update Cadence sections; §8 Skill Telemetry now points at the per-update notes for full history. ### Changed — `coeus-router` trigger coverage Added 3 near-match phrasings for the "I don't know which skill" intent: `"i'm not sure which coeus"`, `"i am not sure which coeus"`, `"not sure which coeus"`. Closes the synonym gap caught in the trigger-eval round 2. ### Updated — `SKILL_REGISTRY.md` - `project-lifecycle` token estimate refreshed: ~2,935 → ~1,781 (reflects the R1+R2+R3+R4 reductions and v1.2 reorg). Delegates row updated to list all six lifecycle `_shared/` references. ### Verified - check_skills() PASS: 12 skills, all invariants OK (0 warnings). - 3-round stress test of check_skills passes/fails as designed: orphan skill → exit 1; over-1024-char description → exit 1; `_*/SKILL.md` → warning. - Full harness 182/182 PASS (with plugin). Plugin rebuilt clean. ### Fixed — Cowork validator rejection: `description` > 1024 chars Cowork's SKILL.md validator caps the `description` frontmatter field at **1024 characters**. Three skills exceeded: - `ooxml-repair` (1048) — condensed Covers list (kept all OOXML check categories) and Do-NOT-use rewrites - `project-lifecycle` (1405) — trimmed redundant near-synonym triggers, condensed prose - `the-architect` (1313) — condensed ROUTE descriptions, prose tightened Trims preserved every `Trigger on:` phrase (essential for routing). After: 762 / 1048 / 968 chars respectively. **New invariant in `check_skills()`** (`scripts/build-plugin.py`): hard fail if any description > 1024 chars. **New harness assertion** (`scripts/coeus_full_test.py` section A): per-skill char-cap check (12 new assertions, harness goes 162 → 182 with plugin). Stress-tested: injecting a 1736-char description fails build with `exit=1` and message `morpheus: description 1736 chars > cap 1024`. Restored, `CHECK PASS`. This bug class — Cowork-side validator constraints — is now caught at commit time, not at upload time. ### Fixed — CI failure on Linux runners (`FileNotFoundError: 'powershell'`) - **`scripts/coeus_full_test.py`** — the hook `.ps1` syntax check hard-coded `'powershell'` as the subprocess binary, which crashed `FileNotFoundError` on `ubuntu-latest` GitHub Actions runners. Replaced with `shutil.which('powershell') or shutil.which('pwsh')`; if neither is present, the test reports `syntax skipped (no powershell on PATH)` and continues. Verified locally on both paths: with powershell on PATH (syntax check runs) and with powershell hidden (syntax check skipped) — both 170/170 PASS. ### Refactored — R4 token reduction round (1,101 tokens saved, 4 more skills slimmed) Architect council ROUTE A (autonomous, user pre-approved) ran against the 8 skills above 50% of the 3000-token cap. Verdict: ship 4 quality-safe, runtime-neutral extractions; reject 4 candidates that would compromise quality (always-loaded sections) or runtime (per-invocation file load). **Shipped:** - **`_shared/lifecycle_resume.md`** + **`_shared/lifecycle_close.md`** (new) — Sub-A (Session Resume) + Sub-D (Session Close) bodies extracted from `project-lifecycle`. Both fire conditionally so reference loads with the sub-function. **Saved 687 tokens (80% → 57%).** - **`docs/Ooxml-Repair-MAE1R1-Reference.md`** (new) — project-specific Known-Good State table + Session 4/5/6 Repair History moved out of `ooxml-repair` SKILL.md (it's example data, not protocol). **Saved 193 tokens (69% → 62%).** - **`_shared/architect_route_c.md`** (new) — ROUTE C deep-explore pipeline extracted from `the-architect`. Rare path — only fires on `--explore` or ambiguous problem. **Saved 154 tokens (77% → 72%).** - **`_shared/plugin_build_recipes.md`** (new) — Phase 4 Bash + PowerShell build-script templates extracted from `plugin-creator`. Fires once per plugin scaffold. **Saved 67 tokens (75% → 73%).** **Rejected (council red-team verdict):** - `llm-council` (61%) — artifact templates fire every council run; extraction adds per-invocation latency (priority 3 violation). - `morpheus` (72%) — tool-routing table referenced every run; same constraint. - `ep-council` (69%) — already lean for scope; in-skill trap table is the routing index (full detail already in `docs/EP-Council-Trap-Screen.md`). - `ooxml-fields` (52%) — barely over 50%; cosmetic-only reduction not worth churn. **Conclusion:** the <50% target is architecturally infeasible for 4 skills given the quality + runtime constraints. The 3000-token hard cap and 2700-token authoring margin remain the load-bearing budgets. Sub-50% is cosmetic and was not pursued where it would degrade quality or add per-run latency. R4 closes the at-margin problem and is the right floor. ### Refactored — R1+R2+R3 token reduction (1,855 tokens freed across 4 skills) Three architect-approved opt-in/cross-cutting sections extracted from at-margin skills into `_shared/`: - **R1 — `_shared/phase5_docx_recipe.md`** (new) — full Phase 5 (12-section consolidation order, emoji-substitution table, 10-item QA checklist, python-docx → pandoc → markdown-fallback generator preference, delivery format). `llm-council` SKILL.md keeps only the user-offer prompt + pointer. **Saved 1,085 tokens** (98% → 61% of cap). - **R2 — `_shared/lifecycle_handover.md` + `_shared/lifecycle_audit.md`** (new, 2 files) — full Sub-B handover template (8 sections + update protocol) and full Sub-C audit protocol (file-listing recipe, 4-category classifier, report template, pre-delete confirmation including the `mcp__cowork__allow_cowork_file_delete` step). `project-lifecycle` Sub-B/Sub-C keep only when/protocol-pointer rows. **Saved 584 tokens** (99% → 80% of cap). - **R3 — `_shared/decision_skill_guardrails.md`** (new) — 9 universal hard rules common to council/decision skills (phase gates, mandatory artifacts, uncertainty flagging, no fabricated citations/capabilities, models-are-simulations, generic personas, surface dissent, caveman-not-applied-here). `llm-council`, `ep-council`, `the-architect` now keep only skill-specific guardrail additions. **Saved 186 tokens** across 3 skills. - Quality trade-off: each `_shared/` file loads only when its phase or sub-function fires (Phase 5 is opt-in / rare; Sub-B and Sub-C fire on specific triggers; guardrails are referenced rather than always-loaded). Net-neutral for code-bearing paths; net-positive for DRY consistency. ### Added — Tier C4 cross-link extended to all 8 judgment-producing skills The Tier C4 pointer (`> Shared rules: …_shared/uncertainty_rules.md`) now appears in ALL 8 judgment-producing skills (added to `llm-council` and `project-lifecycle` after R1+R2 freed cap headroom). Previously deferred on those two skills per the FM-CI-01 author-with-2700-margin rule. ### Added — `check_skills()` invariant enforcement in build script (Council ROUTE A approved) - **`scripts/build-plugin.py` extended with `check_skills()`**, called from `__main__` before `build()`. Four invariants from the council Final_Plan (`C:/Claude/Claude-Work/Projects/AI_Tools/Docx skills/Final_Plan.md`): - **#1 Registry parity** (hard fail) — every `skills//` has a row in `SKILL_REGISTRY.md`. Closes FM-03. - **#2 Token cap** (hard fail, vendored exempt) — every non-vendored `SKILL.md` ≤ 3,000 tokens by `wc -w × 1.33` heuristic. Closes FM-01. - **#3 Name match** (hard fail) — every `SKILL.md` frontmatter `name:` matches its directory name. Closes FM-05. - **#4 Router coverage** (warning, → hard fail in v3.9) — every routable skill appears in `coeus-router`'s match table. Closes FM-02/FM-05. - **`--check-only` flag** for lint-without-build (also wired into `docs/SKILL_ARCHITECTURE.md` as the pre-commit invocation). - **FM-CI-04 mitigation** (Premortem_Report.md): warn if any `_*` directory contains a `SKILL.md` (treated as "convention only, not a skill" by the skip rule; a SKILL.md inside is almost certainly misplaced). - **Stress-tested both directions**: - Inject orphan skill (no registry entry) → exit 1, message names the orphan. - Inject `_test_misplaced/SKILL.md` → warning fires. - Remove the injection → CHECK PASS, exit 0. ### Added — Remote-environment install diagnostic - **`scripts/check-install.py`** — autodetects canonical vs remote install layouts, returns exit 0/1/2, prints path-resolution workaround if the path-mismatch bug from `D:/Downloads/Coeus_Plugin_Error.md` is detected. Catches stale `caveman-protocol/` directories (pre-v3.0.1 install marker). - **`docs/Skill-Install-Diagnostics.md`** — full diagnostic walkthrough: symptom, severity, path resolution map (broken indexed path → real path), how to run the diagnostic, exit-code semantics, permanent fixes, how to confirm you are NOT affected. - **`Coeus_LLM_HANDOVER.md`** — added entry pointing future LLM sessions at the diagnostic + workaround. Clarified this is NOT a repo bug; the canonical install produced by `scripts/build-plugin.py` is correct. The bug is in remote-environment install indexes that disagree with the actual filesystem. ### Changed — project-lifecycle trigger coverage - **`project-lifecycle` triggers broadened** to catch near-match handover-update phrasings that fell through `"update the handover"` literal-match. Added: `"update the handover note"`, `"update handover"`, `"update llm handover"`, `"update coeus llm handover"`. Symptom caught during a session diagnosis where the user said `"update coeus llm handover"` and project-lifecycle did not auto-fire because its trigger string did not contain that exact substring. ### Added — Skill-invocation honesty rule (handover §) - **New convention in `Coeus_LLM_HANDOVER.md`**: do not list a skill as "used implicitly", "de facto applied", or "pattern referenced" in session reports unless the skill's trigger fired and its content loaded. Either the skill loaded (list it) or it did not (do not). Caught a diagnosis where four skills (morpheus, project-lifecycle, plugin-creator, the-architect) were claimed as "implicit" contributors when only ponytail had actually loaded. Conflating borrowed-idea with skill-fired inflates credit and obscures real routing gaps. ### Audited — Skills not invoked this session After-the-fact analysis of all non-invoked skills concluded: - **project-lifecycle would have been useful** had it fired (Sub-B Handover Update directly matches the handover-edit work) — fixed by trigger broadening above. - caveman would have curbed verbose prose — correctly silent (user did not request). - All others (morpheus, plugin-creator, the-architect, llm-council, ep-council, coeus-router, ooxml-*) correctly did not fire — task did not match their domain. ### Added — Tier B1 enforcement: GitHub Actions test workflow - **`.github/workflows/test.yml`** — runs `scripts/coeus_full_test.py --plugin` on every push/PR to `main`. Closes Tier B1 from `Coeus_Architecture_Improvements.md` ("CI check: registry parity — fail build if any `skills/*/` directory is missing from `SKILL_REGISTRY.md`"). The harness already covered registry-parity, frontmatter conformance, XML-token regression class (v3.5.1), cross-skill trigger collisions, dependency resolution, embedded-Python syntax, hook syntax + Guard-2, and full plugin-bundle integrity. Wiring it into Actions converts every guarantee from convention to enforcement. - **`scripts/coeus_full_test.py` made self-locating** — ROOT now resolves from `$COEUS_ROOT`, then `$GITHUB_WORKSPACE`, then the script's parent-parent directory. No CI-side sed patching, no Windows-path hardcode for non-Windows runners. Verified portable: 169/169 PASS when invoked from any cwd. ### Verified post-CI-wire (3 rounds) - R1 baseline: 169/169 PASS against current registered state (registry + router + arch docs already in sync per §14 of `Coeus_v3.8.0_LLM_Handover.md`). - R2 after self-locating: 169/169 from the repo root. - R3 cross-cwd portability: 169/169 invoked from `/tmp`. ### Added — Tools bundle (4 new skills) Four standalone skills delivered together as one umbrella for long-running DOCX-centric project work. Sourced from `C:/Claude/Claude-Work/Projects/Technical_Reports/Outputs/LLM_Handover_Skill_Creation.md` (the architect-approved spec from the MAE-1R1 Well Completion Report project, Sessions 1–15). All four conform to the Coeus standard (front-loaded `Trigger on:` line, `argument-hint:`, no XML tokens in description, 0 cross-skill trigger collisions verified). - **`ooxml-repair`** — diagnose and repair "Word found unreadable content" errors. Covers ZIP integrity, XML well-formedness, w14:paraId validity, orphaned bookmarkEnd tags, orphaned comment references (the most common root cause), duplicate style IDs, numbering chain integrity, relationship target existence, revision ID uniqueness, footnote integrity, mc:Requires namespace declarations. Includes the safe ZIP rebuild recipe that preserves `flag_bits` and `create_version`. - **`ooxml-fields`** — SEQ field + REF field + caption numbering management via direct XML manipulation. Locates fields, reads/corrects cached values, detects hardcoded captions, audits sequence integrity (gaps/duplicates), checks REF target bookmarks. - **`docx-inventory`** — extract a complete figure and table inventory from a DOCX to a two-sheet xlsx (Figures, Tables) with per-row issue flagging (no-SEQ-field, duplicates, gaps). - **`project-lifecycle`** — three sub-functions in one cluster skill: **Resume** (orient a new LLM session on an existing project), **Handover** (create/update the 7-section handover template), **Audit** (classify project files as Active / Superseded / Temp / Unknown). Sub-functions share user-intent domain and combined token count fits within the 3,000-token cluster limit. ### Added — Umbrella reference page + landing-page tab - `docs/Tools.md` — umbrella reference page describing what each Tool does, when to use which, real workflow chains, and how they avoid duplicating existing skills (`docx`, `xlsx`, `memory-management`). - `index.html` landing page now has a fourth main tab **Tools** with four sub-panels (one per skill), back-bar navigation, wiki-style cross-tab links (`Tools-OOXML-Repair`, `Tools-OOXML-Fields`, `Tools-DOCX-Inventory`, `Tools-Project-Lifecycle`). Pattern matches the existing EP-Council / LLM-Council umbrellas. ### Verified - Conformance: **75/75 PASS** (was 50/50 — +25 from the four new skills, with the harness updated to skip `_shared/` resource dirs). - Cross-skill trigger collision audit: **0 collisions** across all 11 skills' quoted trigger phrases. - XML-token-in-description scan: **0 hits** (passes Desktop validator that rejected v3.5.0's `` token). - Plugin builds clean with the UTF-8-flagged Python builder (Cowork-safe). - New full-harness `coeus_full_test.py` ran 3 rounds: R1 caught 7 issues (SKILL_REGISTRY missing 4 new entries; vendored-skill required-field over-strictness), R2 after fixes 145/145 PASS, R3 with rebuilt plugin 165/165 PASS. Harness covers: per-skill frontmatter / no-BOM / no-XML-tokens / no-emoji-in-frontmatter / python-block syntax / front-loaded Trigger line (owned only); cross-skill trigger collisions + dependency resolution + SKILL_REGISTRY completeness; hook bash+PS syntax + functional Guard-2 (refuses on repo, preserves .git); plugin bundle POSIX paths + UTF-8 flag + create_system + manifest schema + every skill present + every SKILL.md re-parses from the extracted bundle. - SKILL_REGISTRY updated with the 4 new Tools-bundle skills. --- ## 2026-Jun-24 (v3.8.0) [pre-Tools] · 2026-Jun-24 ### Added — Skill Architecture v1.0 (Clustered-Modular with Delegation) - **`docs/SKILL_ARCHITECTURE.md`** — formal spec for the three-tier model (nano-skill / cluster / router), 3,000-token hard cap, `skills/_shared/` convention, registry requirement, and the seven hard rules. - **`docs/Coeus-Architecture.md`** — user-facing reference page (same format as `LLM-Council.md` and `EP-Council.md`; linked from the new top-level tab in the landing page). - **`skills/_shared/`** — convention dir for cross-skill rules: - `uncertainty_rules.md` — single source of truth for confidence markers (`Likely:`, `[UNVERIFIED]`, hedged voices) and the four hard uncertainty rules (no fabricated citations, no fabricated capabilities, surface minority positions, don't carry false certainty forward). - `output_formats.md` — common artifact, header, table, path, and length conventions. - `SKILL_REGISTRY.md` — canonical registry of every shipped skill. Same- commit-entry rule enforces the FM-03 (orphan-skill) mitigation from the Architect premortem. - **`skills/coeus-router/SKILL.md`** — Tier-3 meta-skill. Routes user requests to the correct Tier-1/2 Coeus skill using a deterministic match table. Contains zero domain logic — emits only a `ROUTE → / WHY → / RUN →` block and invokes the target skill. Triggers: `/coeus:router`, `/coeus-router`, `"which coeus skill"`, `"route this"`, `"pick the right coeus skill"`. Cross-trigger collision check vs all existing skills: zero collisions. ### Rationale (from The Architect council run, 2026-06-24) The Architect ran a full 4-phase council (ROUTE A — prompt-master → caveman → council) on the question "monolithic vs modular Coeus skill architecture?". Outcome: clustered-modular with delegation pattern, 5:1 council vote against the monolith. Full record in `Architect_Final_Plan_Skill_Architecture.md`; 6-month premortem with six failure modes (FM-01 cluster creep, FM-02 dead delegation path, FM-03 orphan skill, FM-04 shared-rule divergence, FM-05 trigger war, FM-06 handover gap) in `Architect_Premortem_Report_Skill_Architecture.md`. v3.8.0 ships the architecture as four concrete artefacts plus this changelog entry; the mitigation for each failure mode is captured in the architecture doc. ### Compatibility All four artefacts follow the Claude Code plugin spec (`/skills//SKILL.md` with frontmatter, slash commands auto-registered from `name:`), so they load identically in **Claude Chat, Cowork, and Claude Code**. `_shared/` has no `SKILL.md` and is therefore invisible to skill auto-discovery; it is read on demand by Tier-1/2 skills that reference it. ### Compliance - All eight skill files validated against the architecture: 7 of 8 under the 3,000-token cap. `prompt-master` (5,408 tokens) is the documented vendored-upstream exception — `skills/_shared/SKILL_REGISTRY.md` notes the rationale and the upstream-fork upgrade path. - All eight `SKILL.md` `name:` fields match their directory names. - All eight skills registered in `SKILL_REGISTRY.md`. - Cross-trigger collision audit: zero collisions. --- ## 2026-Jun-24 (v3.7.0) · 2026-Jun-24 ### Changed (auto-trigger reliability pass) - **Description front-load** for all five owned skills (`ep-council`, `llm-council`, `morpheus`, `the-architect`, `plugin-creator`): the `Trigger on:` line is now the **first line** of the `description:` block, before any prose. Surfaces that truncate descriptions for intent matching see the triggers first. Vendored skills (`caveman`, `prompt-master`) untouched to avoid re-opening the upstream-sync wipe class (see v3.6.2). - **New triggers added** (≤2 per skill, picked to be unambiguous and collision-free): - `ep-council`: `"farm-in"`, `"FID"` - `llm-council`: `"i'm torn between"`, `"weigh the options"` - `morpheus`: `"refine my prompt for"`, `"compress this prompt"` (promoted from body to trigger line) - `the-architect`: `"full pipeline"`, `"architect this"`, `"stress-test and plan this properly"`, `"run the full council with prompt engineering"` (promoted from body to trigger line) - Cross-skill collision audit: 0 collisions across all owned-skill trigger phrases. - **Dropped `morpheus` and `architect` as bare-word triggers on `ep-council`** (v8 behaviour). The cross-fires caused false positives in non-E&P chats. E&P operators should now use `/coeus:ep-council`, `"board review"`, `"investible"`, `"farm-in"`, `"FID"`, or any documented natural-language trigger. The bare words now route to their canonical skills. ### Added - **Opt-in SessionStart available-skills banner** (`COEUS_STARTUP_BANNER=1` env var). When set, both `cleanup-stale-install.sh` and `.ps1` print a single ASCII line to stderr listing every skill in `skills/`. Default OFF — no behaviour change for users who don't opt in. Uses plain `[Console]::Error.WriteLine` / `>&2` to avoid re-hitting the v3.3.1 Write-Error exit-code class. ### Fixed (this-session incident) - **Cleanup hook `cleanup-stale-install.{sh,ps1}` now refuses to run inside a git repo.** Added Guard 2 (`.git/` or `.github/` present → exit 0, no touch). The original Guard 1 (`.claude-plugin/plugin.json` exists) was too weak: source repos have that file too, so during a hook smoke test the source tree's `.git`, `.github`, `dist`, `index.html`, and `vercel.json` got purged as "stale". Repo recovered from origin (commit `97e69fa`) and a `/tmp` clone; working-tree edits preserved. Guard 2 closes the entire mistake class. ### Changed - **`release.yml` now delegates the `coeus.plugin` bundle step to `python3 scripts/build-plugin.py`** — the same builder used locally. Guarantees byte-for-byte parity between `dist/coeus.plugin` built on a developer machine and the asset attached to a GitHub Release, eliminating the toolchain drift between `zip -r` (release) and the local Python builder. The release still swaps in upstream-fresh `caveman` + `prompt-master` before invoking the builder, so release-time upstream freshness is preserved. ### Added - **Morpheus v1.1.0 — Step 3 auto-execute.** When all pre-flight checks pass without flags, no pause condition fires, and the target model is the current runtime (Claude / Claude Code), Morpheus now runs the compressed prompt immediately in the same conversation. If the target is an external tool (ChatGPT, Gemini, Midjourney, etc.) or the user opts out ("just give me the prompt"), Step 3 is skipped and only the Morpheus Output block is delivered. Step 3 is also skipped if caveman auto-disabled and the user only wanted the engineered prompt back. ### Updated - **`index.html` landing page condensed to 3 top-level tabs** — Overview, EP-Council, LLM-Council. The five companion panels (EP Members, EP Trap Screen, EP Walkthrough, LLM Members, LLM Walkthrough) remain in the document and continue to receive cross-tab link routing via `MD2TAB`, but are reached only via in-content links from their parent tab. Each sub-panel shows a `← Back to ` banner that returns the reader to the tab they came from (origin stack tracked in JS). - README, `docs/` references, and Coeus_LLM_HANDOVER updated to describe the 3-tab navigation model and the new Morpheus auto-execute behaviour. - `.claude-plugin/plugin.json` → `3.7.0`. --- ## 2026-Jun-21 (v3.6.2) · 2026-Jun-21 17:00 ### Added - **Three new reference docs for `llm-council`** in `docs/`, mirroring the structure of the EP-Council documentation set: - **`docs/LLM-Council.md`** — main reference: what it does, install, triggers, pipeline diagram, six-voice table, faction allocation, phase gates, output artifacts, mid-session commands, mandatory flags & guardrails, when-to-use-this-vs-architect-vs-EP-Council, version history. - **`docs/LLM-Council-Members.md`** — full profile of each of the six simulated voices: lens, primary question, characteristic failure mode, when to listen, when to discount, default factional bias. Includes faction mechanics and calibration notes. - **`docs/LLM-Council-Walkthrough.md`** — complete worked example (warehouse consolidation) showing every phase, every gate, every artifact, the optional Phase 5 production .docx flow, and common pitfalls. - **`scripts/build-index-html.py` extended** — `index.html` landing page now exposes 8 tabs (was 5): Overview, EP-Council + 3 EP companions, LLM-Council + 2 LLM companions. Cross-tab link map (`MD2TAB`) updated so in-doc markdown links to the new files switch tabs instead of navigating away. ### Updated - `.claude-plugin/plugin.json` → `3.6.2`. - README and Coeus_LLM_HANDOVER updated to point at the new LLM-Council docs. - Archive mirror refreshed. ### Fixed - **Regression** — the weekly `sync-upstream-skills` workflow (commit `d0ccf0e`, post-v3.4.0) wiped the Coeus-added `argument-hint:` from `skills/caveman/SKILL.md`. The workflow's name-normalisation step pinned `name:` but didn't preserve `argument-hint:`, so each sync silently stripped it. Conformance check caught it during this release. Restored the field and extended the workflow with an idempotent re-injection step: each matrix entry now carries an `argument_hint:` value, and the normalisation block re-inserts it after the `name:` line if upstream omits it. Prevents recurrence on every future sync. --- ## 2026-Jun-21 (v3.6.1) · 2026-Jun-21 16:00 ### Fixed - **Claude Cowork upload rejected `coeus.plugin` with `zip file contains path with invalid characters`.** Root cause: PowerShell's `Compress-Archive` (and even Python's `zipfile.writestr` for ASCII filenames) does not set the **0x800 UTF-8 filename flag** in zip entries. Cowork's strict validator interprets unflagged paths as CP437/legacy and rejects the archive. Desktop's validator was lenient and accepted them, masking the bug. - **New cross-platform builder `scripts/build-plugin.py`** is now the single source of truth for the bundle. It: - Subclasses `ZipInfo` (`Utf8ZipInfo`) to override `_encodeFilenameFlags` and force `0x800` on every entry — CPython otherwise resets `flag_bits` to 0 in `_open_to_write` before writing the local header. - Writes POSIX permissions (0644 for files, 0755 for `*.sh` / `*.ps1` / `*.py`) with `create_system = 3` (UNIX). - Uses forward-slash POSIX paths exclusively (`Path.as_posix()`). - Self-verifies the produced archive against the strict ruleset before exiting. - **`scripts/build-plugin.ps1` and `scripts/build-plugin.sh` are now thin wrappers** that delegate to `build-plugin.py`. Reason: previously each platform produced subtly different zips; the Python backend guarantees identical bytes regardless of host OS. ### Tested Three independent rounds of build+verify confirm compliance with both Claude Cowork (strict UTF-8 flag + POSIX) and Claude Desktop (lenient): - **R1:** Python builder → 35/35 entries UTF-8-flagged, POSIX-permissioned, safe-named. - **R2:** Rebuild + extract-back round-trip → 0 utf8/posix/perm violations, 0 content mismatches. - **R3:** PowerShell wrapper end-to-end + `unzip -tq` integrity scan → no errors detected. ### Updated - `.claude-plugin/plugin.json` → `3.6.1`. - `README.md` build instructions updated to call out the Python builder. --- ## 2026-Jun-21 (v3.6.0) · 2026-Jun-21 15:00 ### Added - **`llm-council` Phase 5 — Production-Ready Document (opt-in)**. After Phase 4 delivers `Final_Plan.md` + `Premortem_Report.md`, the council now offers the user a single consolidated `.docx` that incorporates every artifact (Strategic Roadmap, six-model synthesis, tri-team rounds, full Final Plan, full Premortem Report, ranked mitigation framework, assumptions, decision-gate appendix). The skill prompts exactly once with a yes/no gate; only on explicit "yes" does it proceed. - **Canonical emoji substitution table** in Phase 5.3 — every emoji used anywhere in council output (severity dots, pass/fail marks, team colours, etc.) has a plain-language substitute (`[CRITICAL]`, `[PASS]`, `Blue Team`, etc.). A final regex sweep (`[\U0001F300-\U0001FAFF☀-➿️]`) aborts delivery if any emoji slips through. Reason: Word renders Unicode emojis inconsistently across locked-down corporate desktops; substitution guarantees portability. - **10-item pre-delivery QA checklist** (failable). Covers emoji clearance, section completeness, no `TBD`/placeholder tokens, table integrity, link resolution, heading-level contiguity, page/header fields, word/section counts, source-artifact match, ownership field. Any failure blocks delivery. - **Generator preference order**: `python-docx` → `pandoc` → fallback (deliver consolidated emoji-free markdown + one-line pandoc command for the user to run locally). ### Updated - `skills/llm-council/SKILL.md` → v1.1.0. Added Phase 5 (6 steps), added 3 new behavioural guardrails (Phase 5 is opt-in only; never issue .docx with failed QA; never ship emojis in the production .docx). - `.claude-plugin/plugin.json` → `3.6.0`. ### Note for downstream skills - `the-architect` chains llm-council in its full-pipeline route, so Phase 5 surfaces automatically there too — no separate change needed in `the-architect/SKILL.md`. --- ## 2026-Jun-21 (v3.5.1) · 2026-Jun-21 14:00 ### Fixed - **`skills/plugin-creator/SKILL.md` failed Claude Desktop install** with `Skill 'skills/plugin-creator': SKILL.md description cannot contain XML tags`. The description's `skills//SKILL.md` token was parsed by the desktop validator as an XML tag. Reworded to `per-skill SKILL.md auto-registration` — angle-bracket-free. Body references using `` are not affected; the validator only inspects the frontmatter `description:` field. Plugin now uploads cleanly via **Add Plugin**. ### Updated - `.claude-plugin/plugin.json` → `3.5.1`. - `dist/coeus.plugin` rebuilt. --- ## 2026-Jun-21 (v3.5.0) · 2026-Jun-21 12:00 ### Added - **New skill `plugin-creator`** (the "Autobot Transformer"). Packages a raw idea, a single prompt, a workflow note, or an existing folder of skills into a spec-compliant Claude Code / Cowork plugin bundle ready for desktop "Add Plugin" install. Encodes the full anthropics/knowledge-work-plugins standard plus the lessons learned across the Coeus v3.0 → v3.4.0 conformance pass: canonical 4-key manifest, `skills//SKILL.md` auto-registration (no `commands/` wrappers), optional SessionStart cleanup hook, reproducible build scripts (Bash + PowerShell), UTF-8/LF/no-BOM hygiene, and a Phase-5 conformance check that can fail before delivery. - **Triggers:** `/plugin-creator`, `/plugin-quick`, `"plugin this"`, `"make plugin"`, `"autobot transform"`, `"package this as a plugin"`, `"turn this into a plugin"`, `"scaffold a plugin"`. - **Slash command:** `/coeus:plugin-creator`. - **Argument hint:** `[idea, prompt, workflow, or existing folder to package as a Claude plugin]`. - Six phases: Intake & Classification → Canonical Layout → Optional Cleanup Hook → Build Scripts → Conformance Check → Documentation. Includes a migration patterns table for converting existing non-spec plugins. ### Fixed - **Stale `caveman-protocol` namespace error documented in remote-environment trace** (`coeus_error.md`). Verified that the current Coeus repo (v3.4.0+) already uses `caveman` as the folder + frontmatter `name:`; the error originated from an older install in the user's `/mnt/skills/user/` environment. No code change required in the repo — the fix is delivered by re-installing the current `coeus.plugin`, whose SessionStart cleanup hook (introduced in v3.2.0) purges the stale `caveman-protocol/` directory on first session. ### Updated - **`.claude-plugin/plugin.json`** — bumped to `3.5.0`; description now mentions `plugin-creator`. - **`README.md`** — Skills Reference and Slash Commands tables updated with the new `plugin-creator` row. - **`Coeus_LLM_HANDOVER.md`** — skills table extended; archive mirror at `//192.168.0.119/Big_Data_II/Claude/Archives` updated. ### Conformance - 45/45 PASS against `anthropics/knowledge-work-plugins` standard with the new skill included (folder name matches frontmatter `name:`; description carries `Trigger on:` line; `argument-hint:` present; YAML parses). --- ## 2026-Jun-20 (v3.4.0) · 2026-Jun-20 13:00 ### Removed - **Root-level `plugin.json` mirror.** The canonical manifest lives at `.claude-plugin/plugin.json`. A duplicate at the repo root forced two-source-of-truth maintenance with no benefit. Deleted. - **`commands/` directory** (6 wrapper `.md` files). Slash commands now auto-register from each skill's `SKILL.md` frontmatter (`name` + `argument-hint`) — matching the convention in `anthropics/knowledge-work-plugins`. The `argument-hint` values previously held in the wrappers were migrated into each skill's frontmatter. - **`wiki/`** (5 files duplicating `docs/EP-Council*.md`). Wiki content belongs in the GitHub Wiki feature, not in the plugin tree. The four `docs/EP-Council*.md` are the canonical copies. - **`skill_downloader.py`**. Obsolete since v3.2.0 when the official install path became `coeus.plugin` via Add Plugin. Anyone wanting individual `.skill` files now downloads them as Release assets. - **`.github/workflows/sync-upstream-caveman.yml`**. Retired since v2.1.0 (superseded by `sync-upstream-skills.yml`); the dead file lingered in the repo. ### Changed - **`.claude-plugin/plugin.json` slimmed to the official 4-key schema**: `{name, version, description, author{name, url}}`. Removed `slug`, `homepage` (folded into `author.url`), `license`, `skills[]`, `hooks`, `tracks_upstream`, `packaging`. Skills are auto-discovered from `skills/`. The displaced metadata is preserved in `docs/PLUGIN_MANIFEST_EXTENSIONS.md`. - **`author` field converted to object** `{name, url}` from string, matching the official convention. - **`hooks/cleanup-stale-install.{sh,ps1}` whitelist trimmed** — removed `commands`, `wiki`, `plugin.json`, `skill_downloader.py` from the canonical whitelist (so they get *purged* on upgrade from a v3.3.x install, completing the migration automatically). - **`scripts/build-plugin.{sh,ps1}` updated** to bundle the new canonical layout (drops `commands/`, `plugin.json`, `skill_downloader.py`; adds `docs/`). - **`.github/workflows/release.yml`** — bundle step updated to match; tag-vs-version assertion now checks only the canonical manifest (no reference copy to keep in sync). - **Every SKILL.md gained `argument-hint:`** in frontmatter (caveman, ep-council, llm-council, morpheus, prompt-master, the-architect). Previously this field lived in the deleted `commands/*.md` wrappers. ### Added - **`docs/CONFORMANCE_REPORT.md`** — full discrepancy matrix vs `anthropics/knowledge-work-plugins`. 16 axes audited; lists each fix taken and why. - **`docs/COEUS_EXTENSIONS.md`** — documents the three deliberate divergences from the official spec (hooks/, Coeus_LLM_HANDOVER, upstream-tracked skills) with rationale for each. - **`docs/PLUGIN_MANIFEST_EXTENSIONS.md`** — preserves the metadata displaced from the slimmed `plugin.json` (upstream-tracking provenance, packaging instructions, skill enumeration). - Conformance test script (in maintainer's local toolchain, not bundled): 45 assertions covering manifest schema, deprecated paths, build artefacts, skill frontmatter, encoding, and hook resolvability. All pass. ### Why - Coeus had drifted from the official `anthropics/knowledge-work-plugins` conventions over six version bumps. The accumulated extras (root manifest mirror, commands wrappers, wiki copy, downloader script, retired workflow, oversize manifest) didn't add functionality — they added maintenance surface and made Coeus visibly non-conforming to the reference standard. v3.4.0 brings the layout back in line with one decisive pass while preserving the three genuine Coeus extensions (cleanup hook for desktop-app overlay-install, LLM handover doc, vendored-with-sync upstream skills) under explicit documentation. ### Notes - **Upgrade path**: on the first session after an in-place upgrade from v3.3.x, the SessionStart cleanup hook will purge the now-stale `commands/`, `wiki/`, `plugin.json`, `skill_downloader.py`, and any old `.coeus-cleaned-*` marker, then write `.coeus-cleaned-3.4.0`. No manual cleanup required. - **Slash commands still resolve** as `/coeus:` — Claude Code auto-registers them from each skill's `name` field. The `argument-hint` carried over to the skill frontmatter so Cowork's UI still shows the argument prompt. - The four `docs/EP-Council*.md` files remain at `docs/` for reference; they were already there independent of the deleted `wiki/` copies. --- ## 2026-Jun-20 (v3.3.1) · 2026-Jun-20 11:00 ### Fixed - **`commands/caveman.md` — YAML frontmatter parse error.** The `description:` value contained `Levels: lite, full (default), ultra, wenyan.` — an unquoted scalar with a colon-space that strict YAML parsers interpreted as the start of a nested mapping, producing `mapping values are not allowed here`. Wrapped the description in single quotes so colons inside are literal. Claude Code's command loader may be lenient enough to accept the original, but any strict-YAML tooling (linters, validators, future spec test suites) would reject it. See [docs/TEST_REPORT_v3.3.0.md](docs/TEST_REPORT_v3.3.0.md) Bug #1 for the full trace and audit (only `commands/caveman.md` was affected; the other 5 commands have no colon-space in their descriptions, and all 5 SKILL.md files using `description: >` folded block scalars are immune by construction). - **`hooks/cleanup-stale-install.ps1` — safety-guard exit code.** When the hook is invoked outside a real plugin install dir (`.claude-plugin/plugin.json` missing), it correctly refused to delete anything — but `Write-Error` set `$LASTEXITCODE = 1` and (under the script's `$ErrorActionPreference = 'Stop'`) became terminating, never reaching `exit 0`. Claude Cowork's hook runner would then flag the hook as broken on every session start launched outside a Coeus install, surfacing a noisy stack trace. Replaced `Write-Error` with `[Console]::Error.WriteLine(...)` which mirrors the bash counterpart's `echo >&2` exactly — writes to stderr without raising, exit code stays 0. See [docs/TEST_REPORT_v3.3.0.md](docs/TEST_REPORT_v3.3.0.md) Bug #3. ### Added - **`docs/TEST_REPORT_v3.3.0.md`** — full rigorous-test report for v3.3.0: stage map, all 46 assertion results, the three bugs found (2 fixed in this patch, 1 documented-but-not-fixed because the upstream sync would revert it), defensive tests (safety guard + idempotency), and a self-critique section listing five additional risks that didn't fail any check. ### Changed - Both `plugin.json` files: `version` `3.3.0` → `3.3.1`. - **Cleanup-hook marker GC.** Both `hooks/cleanup-stale-install.{sh,ps1}` now treat any `.coeus-cleaned-*` marker other than the current-version one as stale and delete it. Previously, the install dir accumulated one marker per upgrade (`.coeus-cleaned-3.2.0`, `.coeus-cleaned-3.3.0`, …); now it holds at most one. Verified end-to-end: injected two old markers (`3.2.0`, `3.3.0`) into a fresh v3.3.1 install, ran the hook → both removed, current `.coeus-cleaned-3.3.1` written. ### Notes - No skill content changed; this is a patch for plugin-loading hygiene and hook robustness only. - The cleanup hook in v3.3.1 will purge a stale `.coeus-cleaned-3.3.0` marker on first run after upgrade (markers are not whitelisted), then write a fresh `.coeus-cleaned-3.3.1` marker — same idempotency contract as before. --- ## 2026-Jun-19 (v3.3.0) · 2026-Jun-19 19:30 ### Changed - **EP-Council triggers** — `skills/ep-council/SKILL.md` (v1.5 → v1.6): natural-language triggers `"morpheus"` and `"architect"` removed; `"board review"` and `"investible"` added. The triggers now lean on the language EP investment teams actually use at the deal-screen stage and stop colliding with the `/coeus:morpheus` and `/coeus:the-architect` slash commands. README + wiki trigger tables updated to match. - **Release version is now enforced as the source of truth.** `.github/workflows/release.yml` gained two new pre-flight steps: 1. **Resolve release tag** — reads the tag from `release` events or `workflow_dispatch` input. 2. **Assert plugin.json version matches release tag** — compares the tag (with optional leading `v` stripped) against the `version` field of *both* `.claude-plugin/plugin.json` and `plugin.json`. **Fails the workflow** if either disagrees with a clear "Bump both plugin.json files to X.Y.Z before re-running the release" message. This is the fix for "release not recognising actual release version" — the bundle that ships can no longer disagree with what the Release page advertises. - **`.github/workflows/update-badge.yml` rewritten.** Now: - Writes a third gist file `coeus-release-count.json` for a new "release count" shields.io endpoint badge (paginates `/releases?per_page=100` until exhausted, with a 50-page guard against runaway pagination). - All release-event values (`tag_name`, `published_at`, repo, etc.) are passed via env vars and `set -eu` is enabled, eliminating shell-injection surface (best-practice — collaborator-controlled in a private repo, but cheap discipline). - Drops `actions/checkout` (the workflow only hits the API; no checkout needed). - Reduced `permissions:` to `contents: read`. - Adds the GitHub API version header (`X-GitHub-Api-Version: 2022-11-28`). - **`README.md` + `wiki/Home.md`** — added the release-count badge under the title; updated badge-block prose; documented the version-vs-tag assertion; bumped the EP-Council trigger table. - **`Coeus_LLM_HANDOVER.md`** (renamed from `HANDOVER.md`) — version bump, archive-mirror pointer, refs to itself updated. The repo copy is canonical; a snapshot lives at `//192.168.0.119/Big_Data_II/Claude/Archives/Coeus_LLM_HANDOVER.md`. ### Added - **Release-count badge** — third gist-driven shields.io endpoint badge, refreshed by `update-badge.yml` on every Release. Useful at a glance ("how active is this plugin"); cheaper than scraping a GitHub UI page. ### Security - **`skill_downloader.py`** — defence-in-depth against CVE-2007-4559 (tar-slip / Zip Slip). Both extraction sites (`build_skill_from_tarball`, `package_external_skill`) now: - **Skip non-regular tar members** (`if not member.isreg(): continue`) so symlinks, hardlinks, device nodes, and directories never become entries in the produced `.skill` ZIP. - **Explicitly reject** any path that starts with `/` or contains a `..` segment, in addition to the existing dotfile filter. End users extract these `.skill` ZIPs into their Claude Desktop install dir, so a malicious upstream tarball must not be able to write outside that dir. - **`.github/workflows/release.yml`** — the new tag/version assertion step passes values via `env:` rather than embedding `${{ }}` expressions directly into the `run:` script. Same hardening as `update-badge.yml`. ### Renamed - `HANDOVER.md` → `Coeus_LLM_HANDOVER.md`. All packaging manifests, hooks (`hooks/cleanup-stale-install.{sh,ps1}` whitelist), build scripts, and the release workflow updated. On in-place upgrade from v3.2.0, the SessionStart cleanup hook will treat the stale `HANDOVER.md` in the install dir as not-whitelisted and remove it automatically — no manual cleanup required. ### Notes - The cleanup-hook whitelist now ends with `Coeus_LLM_HANDOVER.md` instead of `HANDOVER.md`. Future renames must propagate to: `hooks/cleanup-stale-install.sh`, `hooks/cleanup-stale-install.ps1`, both `plugin.json` files (packaging commands), `.github/workflows/release.yml`, and `scripts/build-plugin.{sh,ps1}`. - The release-count badge will read `0 total` until the next published Release triggers `update-badge.yml`. Either tag-and-release `v3.3.0` (best) or run the workflow manually via `workflow_dispatch` to backfill against the latest existing Release. --- ## 2026-Jun-19 (v3.2.0) · 2026-Jun-19 18:30 ### Added - **Clean-install hook.** New `hooks/hooks.json` registers a `SessionStart` hook (`hooks/cleanup-stale-install.sh`, plus a PowerShell mirror) that runs once per version on the first session after install. The hook deletes any file or directory in the plugin install root that isn't part of the canonical v3.x layout — so when a user re-uploads an updated `coeus.plugin` via **Claude Desktop → Settings → Capabilities → Customize → Add Plugin**, stale leftovers from prior versions are purged automatically. Idempotent (writes `.coeus-cleaned-` marker), safe-by-default (refuses to run if `.claude-plugin/plugin.json` is missing). - **`scripts/build-plugin.sh` + `scripts/build-plugin.ps1`.** Reproducible local builds that produce `dist/coeus.plugin` containing ONLY the canonical layout — so manual builds match the release-workflow output. PowerShell mirror provided for Windows. - **`.gitignore`.** Now ignores build artefacts (`*.skill`, `coeus.plugin`, `dist/`), `__pycache__`, editor cruft. ### Removed - **Committed root-level `.skill` baselines** (`caveman.skill`, `ep-council.skill`, `llm-council.skill`, `morpheus.skill`, `prompt-master.skill`, `the-architect.skill`). These were stale build outputs that bloated the repo and confused the desktop-app install (they ended up extracted into the plugin install dir on every upload, even though `coeus.plugin` v3.1.0 already excluded them from the bundle). They're now generated fresh by `release.yml` on every Release and uploaded as Release assets, or built locally with `scripts/build-plugin.*`. ### Changed - **`.claude-plugin/plugin.json` + `plugin.json`** — `version` `3.1.0` → `3.2.0`; added `"hooks": "hooks/hooks.json"`; packaging commands updated to include `hooks/`, `scripts/`, `CHANGELOG.md`, `HANDOVER.md`, and the root `plugin.json` reference copy. - **`.github/workflows/release.yml`** — `coeus.plugin` bundle step now includes `hooks/`, `scripts/`, `CHANGELOG.md`, `HANDOVER.md`, and the root `plugin.json`. ### Why - Before v3.2.0, "Add Plugin" via the desktop app did an overlay-extract: it wrote the new ZIP's files into the existing install dir but did not remove files that had been deleted between versions. Re-uploading a fresh `coeus.plugin` would still leave behind the old root-level `.skill` files (and, hypothetically, any retired skill folder from a future restructure). The new SessionStart hook makes the upload behave like a replace-install: anything outside the canonical whitelist is purged on the next session start. This means **the installed version always matches the uploaded ZIP** — no stale duplicates, no orphaned skills, no two-versions-of-the-same-skill bugs. ### Notes - The cleanup hook's whitelist is hard-coded in `hooks/cleanup-stale-install.sh` (and `.ps1`). Any future top-level addition to the canonical layout must also be added there, otherwise the hook will delete it on the next session start. The whitelist currently covers: `.claude-plugin`, `commands`, `skills`, `hooks`, `scripts`, `assets`, `docs`, `wiki`, `plugin.json`, `skill_downloader.py`, `README.md`, `CHANGELOG.md`, `HANDOVER.md`, `CONTRIBUTING.md`, `CLA.md`, `LICENSE`, `.gitattributes`, `.gitignore`. - Removing the `.coeus-cleaned-` marker forces a re-clean on the next session start — useful for debugging. --- ## 2026-Jun-19 (v3.1.0) · 2026-Jun-19 17:30 ### Changed - **Repo restructured to the Claude Code plugin spec.** All six skill directories moved from the repo root into `skills/`: - `caveman/` → `skills/caveman/` - `ep-council/` → `skills/ep-council/` - `llm-council/` → `skills/llm-council/` - `morpheus/` → `skills/morpheus/` - `prompt-master/` → `skills/prompt-master/` (with `references/` preserved) - `the-architect/` → `skills/the-architect/` Moves done via `git mv` so history is preserved. - **`.claude-plugin/plugin.json` + `plugin.json`** — `version` `3.0.1` → `3.1.0`; `skills` array entries now use `"skills/"` paths (e.g. `"skills/the-architect"`); packaging commands collapsed to `skills/` instead of six per-skill paths. - **`.github/workflows/release.yml`** — `coeus.plugin` bundle step now zips `skills/` as a single tree; per-skill `.skill` artefacts are now built from `skills//` (zipped from inside `skills/` so the archive layout stays `/SKILL.md`); upstream replace step swaps `skills/caveman/` and `skills/prompt-master/` in place. - **`.github/workflows/sync-upstream-skills.yml`** — `local_dir` for both upstream-tracked skills updated to `skills/caveman` and `skills/prompt-master`. - **README.md + wiki/Home.md** — Repo Structure tree updated; manual packaging commands (bash + PowerShell) updated; plugin-spec note clarified. ### Why - Coeus shipped with skills at the repo root (`//SKILL.md`). The Claude Code plugin spec expects skills under `/skills//SKILL.md`. On install, Claude Cowork registered the plugin but skipped skill auto-discovery, so even the slash-command wrappers (`/coeus:the-architect`, `/coeus:morpheus`, …) failed to resolve with `Unknown command: /coeus:the-architect`. Moving the six dirs into `skills/` brings the layout to spec; the slash commands resolve and the skills register cleanly on any machine. ### Added - **`HANDOVER.md`** — LLM handover note covering current architecture, install paths, packaging steps, and the rationale for the v3.1.0 restructure. ### Notes - The committed `*.skill` zip baselines at the repo root (`caveman.skill`, `the-architect.skill`, etc.) are still single-skill packages; their internal layout is unchanged (`/SKILL.md`). Only `coeus.plugin` was affected by the restructure. - `skill_downloader.py` was already layout-agnostic (locates `SKILL.md` by suffix match in the upstream tarball), so no code change was needed. --- ## 2026-Jun-19 (v3.0.1) · 2026-Jun-19 12:45 ### Changed - **Version → 3.0.1** in both `.claude-plugin/plugin.json` and `plugin.json`. The distributed zip was labelled `Coeus-3.0.1` while the manifest still read `2.1.0`; the manifest is the source of truth, so the installed plugin reported 2.1.0. Now aligned. - **Skill `name` fields normalised to their kebab folder names** (official skill-guide convention; makes `/coeus:` predictable): `LLM-council` → `llm-council`, `caveman-protocol` → `caveman`. The other four already matched. Command wrappers and dependency references updated to the new names. - **`.github/workflows/sync-upstream-skills.yml`** — added a post-sync **name-normalisation** step so upstream-synced skills (`caveman`, `prompt-master`) keep their Coeus folder-matching `name:` instead of reverting to the upstream value each weekly sync. - **README.md + wiki/Home.md** — skill-name references (`caveman`, `llm-council`) and version updated. ### Why - All six skills loaded on install, but `/coeus:caveman` and `/coeus:llm-council` could not resolve because those skills’ `name:` fields were `caveman-protocol` / `LLM-council`, not the folder/command name. Aligning the names removes that mismatch. (Whether a given surface exposes plugin slash commands at all is separate; in chat surfaces, invoke skills by intent.) --- ## 2026-Jun-18 (v2.1.0) · 2026-Jun-18 12:45 ### Added - **`commands/` directory (NEW) — six slash-command wrappers.** Adds `/coeus:llm-council`, `/coeus:morpheus`, `/coeus:the-architect`, `/coeus:ep-council`, `/coeus:caveman`, `/coeus:prompt-master`. Root cause of the prior `"/coeus:the-architect → Unknown command"` (and `/coeus:morpheus`) failures: the plugin shipped Skills only, with no `commands/` dir — typeable slash commands are sourced from `commands/*.md`. Each wrapper is a thin loader (`description` + `argument-hint` frontmatter) mirroring the bigdata-com command format. `caveman.md` loads the `caveman` skill; `llm-council.md` loads the `llm-council` skill. - **`prompt-master.skill`** — Regenerated baseline package added at repo root (previously absent), so all six skills now ship a committed `.skill`. ### Changed - **`.claude-plugin/plugin.json` + `plugin.json`** — `version` `1.1.0` → `2.1.0`; packaging commands now include `commands/` so the wrappers ship in manually-built `coeus.plugin` archives. - **`.github/workflows/release.yml`** — `coeus.plugin` bundle step now includes `commands/`, so released plugins register the slash commands. - **`skill_downloader.py`** — Added `upstream_path` to `EXTERNAL_SKILLS` (`prompt-master` = `""`, `caveman-protocol` = `"skills/caveman"`); `package_external_skill()` now keeps only that subtree and strips the prefix so `SKILL.md` lands at `caveman-protocol/SKILL.md` (previously the whole upstream repo was nested under the skill). prompt-master description `v1.6+` → `v1.7+`. - **`skill_downloader.py` — single-skill download reworked to release-first.** Each selected Coeus skill is now pulled from the latest GitHub Release's `{skill}.skill` asset, falling back to a **layout-agnostic** build from the `main` tarball (locates the skill dir by suffix match). Added a baked-in `STATIC_SKILLS` fallback registry (menu browses with no auth) and a `--source release|main` flag; OAuth (`repo` scope) is requested lazily, only when a skill is actually fetched. - **`README.md` + `wiki/Home.md`** — Added a **Slash Commands** section documenting the `/coeus:` commands; added `commands/` to Repo Structure; manual packaging commands updated to include `commands/`; documented the release-first Skill Downloader and its `--source release|main` flag. ### Fixed - **`ep-council/SKILL.md`** — added missing `version: 1.5` to frontmatter. - **`skill_downloader.py` Option 3 (Coeus private-repo skills) was broken** — `list_skills()` / `package_skill()` assumed a `skills/{skill}/` layout that does not exist (skills live at the repo root), so Coeus-skill downloads 404'd / found nothing. Replaced with the release-first + layout-agnostic resolver above; the `--external` upstream path is retained. ### Notes - **`ep-council` `"morpheus"` / `"architect"` natural-language triggers are intentional** (bare words → EP-Council; `/coeus:*` slash commands → the pipeline skills). Left as designed. --- ## 2026-Jun-18 (session 9) · 2026-Jun-18 08:57 ### Verified / Docs - **Verification pass — no source changes.** `Inputs/` empty; the session-8 bundle was re-confirmed intact (no merge required). Re-confirmed both access constraints: git checkout (`...OneDrive\Documents\GitHub\Coeus`) and network share (`\\192.168.0.119\...`) are both unreachable from the Cowork session. - **`Coeus_HANDOVER.md`** — Advanced to session 9: added a verification block, the staged `robocopy /MOV` command to move the note to the network-share archive, and an explicit status line (bundle staged ✅ · repo push pending ⏳ · note move pending ⏳). --- ## 2026-Jun-18 (session 8) · 2026-06-18 08:03 ### Added - **`prompt-master/`** — Prompt-master is now **bundled and auto-installed exactly like caveman** (previously a manual prerequisite). Committed `prompt-master/SKILL.md` (v1.7.0) + `references/patterns.md` + `references/templates.md`, mirrored from [nidhinjs/prompt-master@main](https://github.com/nidhinjs/prompt-master). - **`.github/workflows/sync-upstream-skills.yml`** — New unified **directory-level** upstream sync (matrix: caveman + prompt-master). Replaces the single-file caveman sync with a method that mirrors the entire upstream skill directory (so multi-file skills like prompt-master stay complete). Coeus wrapper docs (README/LICENSE/CONTRIBUTING/CLA) are preserved. Runs weekly (Mon 09:00 UTC) and on demand; `max-parallel: 1` + `git pull --rebase` to avoid push races. - **`.github/upstream-prompt-master-sha.txt`** — Sync tracking file seeded to upstream commit `d15eabbe` so the first scheduled run is a no-op. - **`.github/workflows/update-badge.yml`** — Un-retired. Drives the README release + release-date badges from a public Gist on every published release (see Fixed below). ### Changed - **`.github/workflows/sync-upstream-caveman.yml`** — **Retired** (tombstone). Superseded by `sync-upstream-skills.yml`, which directory-syncs caveman with the same better method, as requested. No scheduled trigger remains. - **`.github/workflows/release.yml`** — `coeus.plugin` now bundles **all six skills**: prompt-master is fetched from upstream `main` at release time and swapped into the bundle (same mechanism already used for caveman) and added to the plugin ZIP file list and release-summary table. - **`.claude-plugin/plugin.json`** and **`plugin.json`** (reference copy): - Added `"prompt-master"` to the `skills` array (now six skills). - Replaced the `prerequisites` block with a `tracks_upstream` block documenting both caveman and prompt-master (source, branch, upstream path, sync + release workflows). prompt-master `latest_version` → `1.7.0`. - Packaging commands updated to include `prompt-master/`. - `version` bumped `1.0.0` → `1.1.0` (new bundled skill). - Top-level `description` updated to note prompt-master is bundled and auto-synced. - **`README.md`**: - **Fixed broken release badge** — replaced the native `shields.io/github/v/release/keithceh/Coeus` badge (which renders "no releases / repo not found" because Coeus is a **private** repo and shields.io cannot read private repos) with two **shields.io endpoint badges** backed by a public Gist: a **release** version badge and a **released** date badge in `YYYY-MM-DD` format. Both auto-update on every release via `update-badge.yml`. - Notes, Prerequisites, Install Order, Options 1–4, and Repo Structure updated to reflect prompt-master as a bundled, auto-synced skill (no separate install when using the plugin). prompt-master version references `v1.6+` → `v1.7+`. Added the two upstream-SHA tracking files and the new/retired workflows to Repo Structure. - **`wiki/Home.md`** — Prerequisites (both skills bundled + auto-synced, replacing the "Cowork marketplace" wording), Installation options, and Repo Structure brought in line with the README; prompt-master version updated; badge note added. ### Fixed - **README header badge error "no releases or repo not found"** — root cause identified as repo privacy (native shields.io cannot read private repos). Resolved via the Gist-endpoint badge approach above, which works while the repo stays private and auto-updates with each release. ### Setup required (one-time, by maintainer) - Add repo secret **`GIST_TOKEN`** (PAT with `gist` scope, owned by `keithceh`) so `update-badge.yml` can write the badge Gist `9d9181a0691bac7053e355ade3e9bd01`. See `APPLY_THESE_CHANGES.md`. --- ## 2026-Jun-17 (session 7) ### Fixed - **Plugin install error: "No skills found in specified directories"** — Claude treats `skills` entries as directory paths and looks for `SKILL.md` inside them. Skills array changed from explicit file paths (`"llm-council/SKILL.md"`) to bare directory names (`"llm-council"`). - Both `.claude-plugin/plugin.json` (canonical) and root `plugin.json` (reference copy) updated. - **`README.md`** — Option 4 schema requirements note updated with correct format. - **`wiki/Home.md`** — Plugin schema requirements note updated; Repo Structure section updated to include `.claude-plugin/`, `plugin.json`, `CHANGELOG.md`, `caveman/`, `skill_downloader.py`, and GitHub Actions workflows (was missing these entries). --- ## 2026-Jun-17 (session 6) ### Fixed - **Plugin install errors — schema compliance:** - `name` field changed from `"Coeus"` to `"coeus"` — Claude rejects names with uppercase letters - `skills` array changed from array-of-objects (with `name/path/description/triggers` keys) to flat array-of-strings (`"skillname/SKILL.md"`) — Claude reads skill metadata from each SKILL.md directly, not from plugin.json - Both `.claude-plugin/plugin.json` (canonical) and root `plugin.json` (reference copy) updated - Errors fixed: *"Plugin name can only contain lowercase letters, numbers, and hyphens"* and *"skills list items must be strings at index N, got dict"* - **`README.md`** — Option 4 packaging note expanded with full plugin.json schema requirements (name format, skills array format) - **`wiki/Home.md`** — Installation section rewritten to match current install options (plugin, individual skills, manual clone); schema requirements noted --- ## 2026-Jun-17 (session 5) ### Fixed - **Plugin install error: "Zip must contain a .claude-plugin/plugin.json file"** — Claude requires the plugin manifest at `.claude-plugin/plugin.json` inside the ZIP, not at the root. - **`.claude-plugin/plugin.json`** — New canonical manifest location. Content identical to root `plugin.json`; packaging commands updated to reference `.claude-plugin/` path. - **`plugin.json`** (root) — Updated packaging note to clarify this is a human-readable reference copy only; `.claude-plugin/plugin.json` is canonical. - **`.github/workflows/release.yml`** — `zip` command updated: replaced `plugin.json` with `.claude-plugin/` so the plugin bundle contains the manifest at the correct path. - **`README.md`** — Repo Structure updated to show `.claude-plugin/` directory; Option 4 manual packaging instructions updated with correct ZIP commands and a warning note about the `.claude-plugin/` path requirement. --- ## 2026-Jun-17 (session 4) ### Changed - **`wiki/Home.md`** — Removed KC-modified caveman additions block (§3 Workspace Memory Map, §4 Secure Micro-Subagent Roles, §5 Hardened Auto-Clarity Circuit Breaker). Replaced with pointer to JuliusBrussee/caveman@main and note that Coeus auto-syncs weekly. - **`.github/PULL_REQUEST_TEMPLATE.md`** — Removed checklist items and Affected Sections entries referencing §3/§4/§5 (KC-specific additions that no longer exist in Coeus). Retained §3 entry renumbered to reflect upstream's native circuit breaker section. - **`README.md`** — Added native GitHub release badge (`shields.io/github/v/release/keithceh/Coeus`); removed Gist-backed "Last Commit" badge. - **`.github/workflows/update-badge.yml`** — Retired; overwritten with notice. Replaced by native GitHub releases badge (no workflow required). --- ## 2026-Jun-17 (session 3) ### Changed - **`caveman/SKILL.md`** — KC-modified version superseded. Caveman now tracks [JuliusBrussee/caveman@main](https://github.com/JuliusBrussee/caveman) directly. Future improvements follow upstream. - **`.github/workflows/sync-upstream-caveman.yml`** — Changed from opening a draft PR for manual review to directly committing upstream `SKILL.md` to `main` on each weekly sync. No manual merge step required. - **`.github/workflows/release.yml`** — `caveman.skill` now built from upstream fetch rather than local `caveman/` dir. Removed `caveman-upstream` naming distinction. Simplified release artifacts: `coeus.plugin`, 5 Coeus `.skill` files, `prompt-master.skill`. - **`plugin.json`** — Caveman description updated to reflect upstream tracking. - **`README.md`** — Removed all KC-modified references for caveman. Updated caveman section, install order table, Option 1 (Plugin), Option 2 (Releases), Option 3 (Skill Downloader). Removed `--external caveman-protocol` from downloader docs. --- ## 2026-Jun-17 (session 2) ### Added - **`.github/workflows/release.yml`** — CI_BUNDLE release workflow. On every GitHub release (or manual `workflow_dispatch`), automatically fetches latest `main` from `nidhinjs/prompt-master` and `JuliusBrussee/caveman`, packages each as a `.skill` ZIP, bundles all Coeus skills + prerequisites into `coeus.plugin`, and attaches all artifacts to the release. Eliminates the need for manual prerequisite install steps. ### Changed - **`README.md`** — Installation section updated to reflect auto-bundled prerequisites: - Option 1 (Plugin Install) now describes downloading `coeus.plugin` from Releases (batteries included); documents the `bundled-prerequisites/` directory inside the plugin ZIP; explains how the CI workflow works. - Option 2 (GitHub Releases) now lists `prompt-master.skill` and `caveman-upstream.skill` as auto-built release artifacts. - Repo Structure updated to show all three workflow files. - **`CHANGELOG.md`** — This entry. All notable changes to Coeus are documented here. --- ## 2026-Jun-17 12:37 ### Added - **`plugin.json`** — Plugin manifest converting Coeus into a Claude plugin. Lists all 5 skills (llm-council, morpheus, the-architect, ep-council, caveman) with trigger keywords, inter-skill dependencies, and packaging commands for building `coeus.plugin`. - **`CHANGELOG.md`** — This file. Full change history. - **`skill_downloader.py` — `--external` mode** — New flag to download external prerequisite skills from public GitHub repos without OAuth. Supports `prompt-master` (nidhinjs/prompt-master, main branch) and `caveman-protocol` (JuliusBrussee/caveman, upstream original). Usage: `python3 skill_downloader.py --external [skill-name]`. ### Changed - **`README.md`** — Major update: - Installation section restructured: Plugin install (Option 1) added as primary method; former options renumbered (Options 2–4); Skill Downloader section expanded to document `--external` flag for prerequisites. - Prerequisites section expanded with dedicated "Installing prompt-master (latest version)" subsection covering CLI (Option A) and manual ZIP (Option B) install paths. - "Issues and Planned Improvements" section replaced with "Changelog" (link to this file) and "Notes" (resolved/ongoing context). - Repo Structure updated to show `plugin.json`, `CHANGELOG.md`, `caveman/`, and `skill_downloader.py`. ### Notes - prompt-master (nidhinjs/prompt-master) has no formal releases or tags; latest version (v1.6+) is always on `main` branch. Install instructions now reflect this. - Caveman bundled in Coeus is the KC-modified version. The `--external caveman-protocol` flag downloads the upstream original from JuliusBrussee/caveman if needed. --- ## Prior History Coeus is a rolling mono-repo. Changes before 2026-Jun-17 are tracked via Git commit history at https://github.com/keithceh/Coeus/commits/main.