Untrack internal topology, add the Redmine time-tracking playbook

.gitignore

@@ -12,3 +12,10 @@ openclaw/workspace/.openclaw/
 services/linkedin/chromadb/
 services/linkedin/cookies/
 services/linkedin/posts.json
+
+# Runtime ecosystem map — real container names/ports; this repo is PUBLIC.
+# Bind-mounted from disk (see docker-compose.yml), so it never needs to be tracked.
+# Track ironclaw/workspace/TOOLS.md.example instead.
+ironclaw/workspace/TOOLS.md
+# Internal build roadmap — absolute local paths into private repos, auth routes, token names
+ironclaw/workspace/todo.md

AGENTS.md

@@ -17,7 +17,8 @@ clawbot/
 │       ├── .mcp.json       # MCP server connections (Drupal, Apify, etc.)
 │       ├── IDENTITY.md     # Agent name, persona, avatar
 │       ├── SOUL.md         # Agent values and behavioural guidelines
-│       ├── TOOLS.md        # Available tools and how to use them
+│       ├── TOOLS.md        # Available tools and how to use them — gitignored (real container
+│       │                   #   names); copy from TOOLS.md.example
 │       ├── USER.md         # Info about the user (fill in as needed)
 │       └── skills/         # Custom skill playbooks (SKILL.md per skill)
 ├── docker-compose.yml      # Main Compose file

ironclaw/workspace/TOOLS.md → ironclaw/workspace/TOOLS.md.example

@@ -1,15 +1,22 @@
 # TOOLS.md - Ecosystem Map
 
+> **Template.** Copy to `TOOLS.md` and replace the placeholder container names and ports with
+> your actual ones:  `cp TOOLS.md.example TOOLS.md`
+>
+> `TOOLS.md` is gitignored on purpose — this repo is public, and the real ecosystem map is
+> internal infrastructure detail. The workspace is bind-mounted, so the agent reads your local
+> copy directly; a `docker compose restart ironclaw` picks up changes.
+
 ## MCP Servers
 
 | Server | Container | Port | Tools | Use For |
 |---|---|---|---|---|
-| **`gateway`** | mcpserver-gateway | 8100 | 29 | **Unified entry point — all tools via single auth** |
-| **`site-files`** | clawbot-openclaw-1 (local) | — | ~10 | **Read/write Hugo source at `/site`** |
-| `ragchat` | ragchat-nginx-1 | 8080 | 21 | CMS content, RAG knowledge base, Drupal |
-| `seobot` | seo-tools-mcp | 3001 | 3 | Website SEO audits, crawl reports |
-| `osintbot` | osintbot-mcp | 8200 | — | LinkedIn (Apify)/OSINT research, prospect intel |
-| `sales` | sales-assistant-assistant-1 | 8000 | 3 | Product Q&A, knowledge base search |
+| **`gateway`** | mcp-gateway | 8100 | 29 | **Unified entry point — all tools via single auth** |
+| **`site-files`** | <agent-container> (local) | — | ~10 | **Read/write Hugo source at `/site`** |
+| `ragchat` | ragchat | 8080 | 21 | CMS content, RAG knowledge base, Drupal |
+| `seobot` | seo-mcp | 3001 | 3 | Website SEO audits, crawl reports |
+| `osintbot` | osint-mcp | 8200 | — | LinkedIn (Apify)/OSINT research, prospect intel |
+| `sales` | sales-api | 8000 | 3 | Product Q&A, knowledge base search |
 
 ### Gateway tool naming
 
@@ -36,15 +43,15 @@ Tools routed through the gateway use `{server}__{tool}` format, e.g.:
 
 ## Docker Network
 
-All containers on `nginx-proxy` bridge network.
-OpenClaw container: `clawbot-openclaw-1`.
+All containers on a shared bridge network (see `docker-compose.yml`).
+OpenClaw container: `<agent-container>`.
 Reach services by container name over the network.
 
 ## Key URLs
 
 - OpenClaw dashboard: http://127.0.0.1:18790
-- **MCP Gateway: http://mcpserver-gateway:8100** (internal) / http://127.0.0.1:8100 (host)
-- RagChat: http://ragchat-nginx-1:8080
-- Sales assistant: http://sales-assistant-assistant-1:8000
-- SEO crawler MCP: http://seo-tools-mcp:3001
-- OSINT MCP: http://osintbot-mcp:8200
+- **MCP Gateway: http://mcp-gateway:8100** (internal) / http://127.0.0.1:8100 (host)
+- RagChat: http://ragchat:8080
+- Sales assistant: http://sales-api:8000
+- SEO crawler MCP: http://seo-mcp:3001
+- OSINT MCP: http://osint-mcp:8200

ironclaw/workspace/skills/redmine-time-tracking/SKILL.md

@@ -0,0 +1,161 @@
+---
+name: redmine-time-tracking
+description: >-
+  Log the time an agent spends on a project to the self-hosted Redmine (CRM).
+  Resolve the working directory to a Redmine project, find or create an issue for
+  the task, then file a time entry under the "AI Agent" activity. Run automatically
+  after ~15+ minutes of substantive work on a project, or when asked to "log my
+  time" / "track this work" / "record hours in Redmine".
+---
+
+# Redmine Time Tracking
+
+Log agent work time to Redmine. Data model: **Project → Issue → Time entry**.
+
+> **ClawBot copy.** The canonical version lives in Luděk's monorepo at
+> `ai/skills/redmine-time-tracking/SKILL.md`; keep the two in sync when the procedure
+> changes. This copy is deliberately narrower — `clawbot` is a public repository, so the
+> stack's endpoints and the admin procedure stay in the canonical copy only. You don't need
+> them: you reach Redmine through the gateway's `redmine__*` tools.
+
+## Environment
+
+- **Tools:** `redmine__<tool>` through the MCP gateway.
+- **Session start time:** the `WORK SESSION` line you appended to `MEMORY.md`.
+- **Clock:** `date -u +%Y-%m-%dT%H:%M:%SZ`; if you have no shell, estimate conservatively
+  from the conversation.
+
+## When to run
+
+- **Automatically**, after roughly 15 minutes or more of real effort on one project
+  (research, coding, audits, multi-step tasks). This is internal bookkeeping on Luděk's
+  own infrastructure — you do **not** need to ask permission first.
+- On explicit request ("log my time", "track this").
+- On an end-of-day flush: process every open work session that hasn't been logged.
+
+Skip trivial work (a one-line fix, a quick question) — it's not worth an issue.
+
+## Procedure
+
+### 1. Resolve the Redmine project
+
+Map the working directory to a Redmine project **identifier**:
+
+| Subproject / area | Redmine identifier |
+|---|---|
+| `customers/plymovent`, `plymovent` | `plymovent` |
+| `ragchat` | `ragchat` |
+| `osintbot` | `osintbot` |
+| `sales-assistant` | `sales-assistant` |
+| `mcpserver`, mcpserver.cz site | `mcpserver-cz` |
+| `seo-tools` | `seo-tools` |
+| `geo` | `geo` |
+| `startupfactory` | `startupfactory` |
+| `drupal-lama` | `drupal-lama` |
+| `ludekkvapil` / `/site` (ludekkvapil.cz) | `ludekkvapil-cz` |
+| `clawbot` — you, your workspace and your skills | `clawbot` |
+| `customers/drupal-cz`, `drupalcz` (drupal.cz) | `drupal-cz` |
+| `ai`, `redmine`, `docs`, `scripts`, `ops-proxy`, `observability`, `reception`, infra, monorepo root | `lkv` |
+
+Call `list_redmine_projects` and resolve the mapped identifier to its **numeric `id`** —
+`create_redmine_issue` and the time-entry tools want the integer project id, not the
+identifier string. If you can't tell which project the work belongs to, **ask** — don't
+guess, don't invent a project.
+
+### 2. Compute hours
+
+`elapsed = now − session start`, in hours, **rounded to the nearest 0.25, minimum 0.25**.
+Count actual working time, not wall-clock time you were idle.
+
+### 3. Resolve the "AI Agent" activity id
+
+Call `list_time_entry_activities`, find the entry named **"AI Agent"**, keep its `id`.
+If it's missing, the one-time host-side setup hasn't been run — tell Luděk and stop.
+
+### 4. Find or create the issue (one issue per task)
+
+- Call `search_redmine_issues` (`query` = a few words from the task) or `list_redmine_issues`
+  scoped to the project for an **open** issue that already matches this task (you may be
+  resuming work). If found, use its `id`.
+- Otherwise `create_redmine_issue`:
+  - `project_id`: the **numeric id** from step 1
+  - `subject`: a short, specific task title (e.g. *"Fix CLS on blog list pages"*, not *"work"*)
+  - `description`: 1–3 sentences — what and why
+  - `fields`: an object with `tracker_id` (from `list_project_trackers` for this project —
+    use the one that matches the project name), `status_id` for **"In progress"** (look it
+    up with `list_redmine_issue_statuses`), and `assigned_to_id` = your own user id
+    (`get_current_user` → `mcp-agent`).
+
+`status_id` and `parent_issue_id` are usually **dropped silently** — see
+[Permission limits of `mcp-agent`](#permission-limits-of-mcp-agent). The issue is still
+created; it just lands in the project's default status and unparented. Don't retry, and
+don't treat it as a failure of the whole step.
+
+### 5. File the time entry
+
+`manage_time_entry`:
+  - `action`: `"create"`
+  - `issue_id`: from step 4
+  - `hours`: from step 2
+  - `activity_id`: the "AI Agent" id from step 3
+  - `spent_on`: today (`YYYY-MM-DD`)
+  - `comments`: one factual sentence on what you did, suffixed with your agent tag —
+    `… (ClawBot)`. Max 255 chars, no filler.
+  - Do **not** pass `user_id` — it defaults to the current MCP user (`mcp-agent`).
+
+### 6. Close the task if it's done
+
+If the task is complete and produced its deliverable, `update_redmine_issue` to set the
+status to **"Done"** and `done_ratio` 100 (pass these in `fields`). Leave it "In progress"
+if there's follow-up work.
+
+Then **read the issue back** and check the status actually changed — the update reports
+success even when the status was ignored. If it didn't land, say so in your reply and leave
+the status alone; don't loop on it.
+
+### 7. Record that it's logged
+
+Change that `WORK SESSION` line's trailing state in `MEMORY.md` to `logged <time-entry-id>`.
+
+## Verify
+
+`list_time_entries` scoped to the project + today — confirm the entry shows user
+`mcp-agent`, activity `AI Agent`, and the agent-tag comment suffix.
+
+Human-readable check, for Luděk: the project's **Time entries** tab, and its **Report**
+grouped by Activity — "AI Agent" hours are a separate line from human "Development" /
+"Design" hours.
+
+## Permission limits of `mcp-agent`
+
+The `mcp-agent` account holds the role **"Agent"**, which can create and edit issues and
+file time entries, but **not** manage subtasks or move issues through the workflow. Redmine
+does not reject the disallowed attributes — it drops them and still answers `200`/`204`, so
+a write can look like it succeeded while nothing changed. Confirmed on 2026-09-12 for:
+
+| Attribute | What happens |
+|---|---|
+| `status_id` (on create and update) | ignored; issue keeps the project's default status |
+| `parent_issue_id` | ignored; the issue stays top-level |
+
+What works: `subject`, `description`, `tracker_id`, `assigned_to_id`, `done_ratio`, notes,
+and all time-entry writes.
+
+So: **after any status or parent write, read the issue back**. If it didn't take, report it
+in your reply rather than retrying or escalating on your own.
+
+Anything the Agent role can't do — nesting subtasks, forcing a status, re-running the
+setup — needs admin access you don't have. Say what you couldn't do and stop; the host-side
+copy of this playbook carries the admin procedure. Lifting the limitation permanently is
+Luděk's call: Administration → Roles → Agent → *Manage subtasks*, plus the workflow
+transitions for the Agent role per tracker.
+
+## Notes
+
+- Never log to a project that isn't in the map above.
+- If issue creation fails with a tracker or permission error, the setup script needs to be
+  (re-)run for that project — report it, don't retry blindly.
+- The MCP server authenticates as `mcp-agent` (legacy mode) — you send no credentials.
+- The server wraps free-text fields in returned JSON in `<insecure-content-…>` markers as a
+  prompt-injection guard. That's a display wrapper on reads only — the stored value is
+  clean; don't include those markers when you write text back.