feat(routing): project_create MCP tool — gitea-first new-project pipeline (#10)

Adds the project_create tool to the routing pod that automates the
"new project" bootstrap end-to-end from claude.ai. Gitea-first
architecture: GitHub receives the repo only via push-mirror, never
via a direct GitHub API call from this server.

Four sequential calls to the gitea-mcp server (configured via
GITEA_MCP_URL):

  1. create_project_from_template — Gitea repo from
     template-go-{agent,web} per the 'stack' arg
  2. repo_mirror_push (action=add) — push-mirror to
     github.com/<GITHUB_OWNER>/<name>.git, interval 8h, sync_on_commit
  3. file_write_branch — k3s/staging/<name>/namespace.yaml committed
     on a staging/<name> branch in the infra repo
  4. issue_create — experiment brief (hypothesis + description + stack
     + provisioning log) on the new repo, returns the issue_url

Returns gitea_url, github_url, issue_url, next_steps. The next_steps
string is the exact shell sequence the operator runs locally to
clone, scaffold via local-dev 'task new-project', and push.

Idempotency: create_project_from_template + repo_mirror_push +
file_write_branch all return JSON-RPC code -32003 (Conflict) when
their target already exists; the orchestrator swallows the conflict
and continues. Re-running on an existing repo restates the brief in
a fresh issue.

Error handling: on any non-conflict downstream failure the response
returns {reached: ["<step>",...], failed_step: "<step>"} alongside
a JSON-RPC error. No rollback — partial state stays so the operator
can resume manually.

New env vars (all optional except GITEA_MCP_URL):
  GITEA_MCP_URL    enables the tool
  GITEA_MCP_TOKEN  bearer auth for gitea-mcp
  GITEA_OWNER      default mathias
  GITHUB_OWNER     default mathiasb
  INFRA_REPO       default infra
  GITHUB_PAT       repo scope, used as mirror remote_password; never logged

Without GITEA_MCP_URL set, the tool is not registered and the
routing pod starts normally (degrades open).

internal/mcpclient/: new minimal JSON-RPC tools/call client with
bearer auth, used by project_create. Unwraps MCP's
content[0].text envelope and surfaces typed errors via mcpclient.Error.

Tests: table-driven against an httptest fake gitea-mcp covering happy
path (4-step success + correct PATCH-style arg shapes), idempotent
repo-exists, mirror failure (partial-success response with reached=
[create_repo] + failed_step=mirror), infra-commit failure (reached up
to mirror + failed_step=infra_commit), and validation errors.

Closes #10

.aider.conventions.md

@@ -36,9 +36,18 @@ These rules apply to every task across every project, regardless of harness.
 4. **Goal-driven execution.** Define clear success criteria up front for every task.
    Loop — implement, verify, refine — until those criteria are met. Don't claim
    completion without evidence (tests pass, command output, observed behavior).
-5. **Branch-per-task for multi-agent repos.** When another agent may be active on
-   the same repo, create a branch (`agent/<description>`), commit there, and open a
-   PR. Do not merge without explicit instruction from Mathias.
+5. **Trunk-Based Development — commit directly to main.** Every commit is one
+   logical change (one tool, one fix, one test) with passing tests. Main is always
+   deployable. Never create long-lived feature branches.
+
+   **Exception — parallel agents on same repo:** If another agent is known to be
+   actively working on the same repo simultaneously, create a short-lived branch
+   (`agent/<description>`), finish the task, and merge to main within the same
+   session. Do not leave agent branches open between sessions.
+
+   **Exception — external contributor or client four-eyes requirement:** Use
+   PR flow only when a human reviewer outside the project is required. Document
+   the reason in PROJECT.md.
 
 ## Default stack
 
@@ -49,9 +58,10 @@ These rules apply to every task across every project, regardless of harness.
 | Build | Task (taskfile.dev) | Make | — |
 | Containers | Docker Compose (dev), k3s (prod) | — | — |
 | DB | PostgreSQL + sqlc | SQLite | — |
-| Search | Qdrant (vector), BM25 | — | — |
+| Search | pgvector (vector), BM25 | Qdrant (when >1M vectors or hybrid retrieval) | — |
 | Logging | slog (structured) | — | — |
 | Testing | Table-driven, testify | — | — |
+| Agents (Go) | google.golang.org/adk + pkg/litellm adapter | — | — |
 
 Exploratory: Rust, Zig — I'll tell you when I want these.
 
@@ -61,9 +71,12 @@ Exploratory: Rust, Zig — I'll tell you when I want these.
 - **Errors**: `fmt.Errorf("operation: %w", err)` — never naked, never log-and-return
 - **Naming**: stdlib conventions, no stuttering
 - **Architecture**: prefer stdlib over frameworks, constructor injection, env-var config parsed into typed structs
-- **Git**: conventional commits (`feat:`, `fix:`, `chore:`), one concern per PR, PR describes *why* not *what*
+- **Git**: conventional commits (`feat:`, `fix:`, `chore:`), commit directly to main,
+  one logical change per commit, CI is the quality gate
+- **Never**: long-lived feature branches, PRs for solo work, direct push without
+  passing `task check` locally first
 - **Security**: no secrets in code, govulncheck before adding deps, SOPS for encrypted config
-- **Dependencies**: prefer stdlib. testify, slog, templ, sqlc are pre-approved; anything else needs justification in the commit message
+- **Dependencies**: prefer stdlib. testify, slog, templ, sqlc, google.golang.org/adk (agent projects only) are pre-approved; anything else needs justification in the commit message
 
 ## Infrastructure
 
@@ -71,7 +84,7 @@ Three machines on Tailscale:
 
 | Machine | Role | Key specs |
 |---------|------|-----------|
-| koala | GPU inference, heavy compute | RTX 5070, runs llama-swap, Qdrant |
+| koala | GPU inference, heavy compute | RTX 5070, runs k3s + llama-swap + shared postgres18/pgvector |
 | iguana | Services, builds | M2 Ultra Mac |
 | flamingo | Daily driver, edge | Mac mini, ~/dev is here |
 
@@ -103,18 +116,64 @@ See `~/dev/PROJECT_SUMMARY.md` for detailed descriptions of each project.
 - **koala-ai-stack** (`AGENTS/`) — local AI server infrastructure management
 - **klimatkollen** (`XT/`) — Swedish municipal climate data platform
 
-## Knowledge base
+## Knowledge base — actively use it
 
-When available, agents can query the shared knowledge base:
+A persistent brain (BM25 search + LLM-synthesised Q&A) survives across sessions,
+hosts, and harnesses. It holds 100+ hard-won entries: infra incident postmortems,
+Go pitfalls, framework gotchas, design principles, ADRs. **It is not optional
+reference material — query it actively, not just when explicitly told.**
 
-- **MCP**: `mcp://hyperguild.<TAILNET>.ts.net:3100/knowledge`
-- **HTTP**: `http://hyperguild.<TAILNET>.ts.net:3100/api/v1/search`
+### When to query (treat as a reflex)
 
-<!-- TODO: replace <TAILNET> placeholder with the real Tailscale tailnet
-     name once hyperguild is deployed. Until then, agents that try to
-     reach the knowledge service on a host where it isn't running will
-     get DNS NXDOMAIN, which is the desired fail-loudly behavior. -->
-- **Scoping**: defaults to `public` collection; client projects filter to `{client}` + `public`
+- **Before** starting a non-trivial task — search for prior art with the symptom
+  AND the system component ("how did we solve X in Y?"). 5 seconds beats 5 hours.
+- **When debugging** — search for the error string, the stack frame, the affected
+  service. Past you may have already paid this tax.
+- **Before adopting** a pattern, library, framework, or model name — check if it
+  was tried and rejected, or what the integration footguns are.
+- **When making architectural decisions** — search for the domain + "ADR" or
+  "decision" to find prior reasoning before re-deriving it.
+- **When a recommendation feels novel** — challenge yourself: "has this been
+  documented?" The brain often has it.
+
+### When to write
+
+After you discover something that **future-you would forget** and that **isn't
+recoverable from the code, git log, or PR description alone**:
+
+- Bugs whose root cause is non-obvious and generalisable beyond this project.
+- Framework / library / model-name quirks that bit you and would bite anyone.
+- Design principles validated under fire (e.g. "every `_get` needs a `_list`").
+- Postmortems for incidents: what broke, why, how diagnosed, what to do next time.
+
+DON'T write project status, sprint progress, PR summaries, or "what I did this
+session" — those rot fast and the originals are in git/gitea anyway. Brain
+entries that age well are about *why*, *how to avoid*, and *what to do when*.
+
+### How to access (per harness)
+
+| Harness | Query | Write |
+|---------|-------|-------|
+| **Claude Code, Claude Desktop** | `brain_query` (BM25), `brain_answer` (LLM-synth + sources) MCP tools | `brain_write` MCP tool |
+| **Crush, Pi, Antigravity, other MCP-capable** | same MCP server: `ingestion-brain` (via the `mcp__*_brain__*` namespace once authenticated) | same |
+| **Anything HTTP-only (curl, scripts)** | `POST https://brain-mcp.d-ma.be/query` with `{"query":"..."}` (auth via `BRAIN_MCP_TOKEN`) | `POST .../write` with `{"content":"...","filename":"..."}` |
+| **Browser / human inspection** | `https://gitea.d-ma.be/mathias/hyperguild` → `knowledge/` and `wiki/` markdown files |
+
+- **Scoping**: defaults to `public` collection; client projects filter to `{client}` + `public`.
+- **Routing**: brain_answer's LLM uses berget.ai as primary, iguana ollama as
+  fallback. Both are configurable in the `supervisor/ingestion-deployment.yaml`
+  on the koala k3s cluster; don't hardcode local-only model names into the
+  berget URL (see knowledge entry on namespace mismatches).
+
+### Quick reflex checks
+
+If you find yourself about to say any of these out loud, you owe yourself a brain query first:
+
+- "I think the issue might be..."
+- "Let me try X and see..."
+- "I'll just write a script to..."
+- "This is probably a new bug..."
+- "Has anyone done this before?" — *yes, probably, go check.*
 
 ## Client work rules
 

.context/system-prompt.txt

@@ -41,9 +41,18 @@ These rules apply to every task across every project, regardless of harness.
 4. **Goal-driven execution.** Define clear success criteria up front for every task.
    Loop — implement, verify, refine — until those criteria are met. Don't claim
    completion without evidence (tests pass, command output, observed behavior).
-5. **Branch-per-task for multi-agent repos.** When another agent may be active on
-   the same repo, create a branch (`agent/<description>`), commit there, and open a
-   PR. Do not merge without explicit instruction from Mathias.
+5. **Trunk-Based Development — commit directly to main.** Every commit is one
+   logical change (one tool, one fix, one test) with passing tests. Main is always
+   deployable. Never create long-lived feature branches.
+
+   **Exception — parallel agents on same repo:** If another agent is known to be
+   actively working on the same repo simultaneously, create a short-lived branch
+   (`agent/<description>`), finish the task, and merge to main within the same
+   session. Do not leave agent branches open between sessions.
+
+   **Exception — external contributor or client four-eyes requirement:** Use
+   PR flow only when a human reviewer outside the project is required. Document
+   the reason in PROJECT.md.
 
 ## Default stack
 
@@ -54,9 +63,10 @@ These rules apply to every task across every project, regardless of harness.
 | Build | Task (taskfile.dev) | Make | — |
 | Containers | Docker Compose (dev), k3s (prod) | — | — |
 | DB | PostgreSQL + sqlc | SQLite | — |
-| Search | Qdrant (vector), BM25 | — | — |
+| Search | pgvector (vector), BM25 | Qdrant (when >1M vectors or hybrid retrieval) | — |
 | Logging | slog (structured) | — | — |
 | Testing | Table-driven, testify | — | — |
+| Agents (Go) | google.golang.org/adk + pkg/litellm adapter | — | — |
 
 Exploratory: Rust, Zig — I'll tell you when I want these.
 
@@ -66,9 +76,12 @@ Exploratory: Rust, Zig — I'll tell you when I want these.
 - **Errors**: `fmt.Errorf("operation: %w", err)` — never naked, never log-and-return
 - **Naming**: stdlib conventions, no stuttering
 - **Architecture**: prefer stdlib over frameworks, constructor injection, env-var config parsed into typed structs
-- **Git**: conventional commits (`feat:`, `fix:`, `chore:`), one concern per PR, PR describes *why* not *what*
+- **Git**: conventional commits (`feat:`, `fix:`, `chore:`), commit directly to main,
+  one logical change per commit, CI is the quality gate
+- **Never**: long-lived feature branches, PRs for solo work, direct push without
+  passing `task check` locally first
 - **Security**: no secrets in code, govulncheck before adding deps, SOPS for encrypted config
-- **Dependencies**: prefer stdlib. testify, slog, templ, sqlc are pre-approved; anything else needs justification in the commit message
+- **Dependencies**: prefer stdlib. testify, slog, templ, sqlc, google.golang.org/adk (agent projects only) are pre-approved; anything else needs justification in the commit message
 
 ## Infrastructure
 
@@ -76,7 +89,7 @@ Three machines on Tailscale:
 
 | Machine | Role | Key specs |
 |---------|------|-----------|
-| koala | GPU inference, heavy compute | RTX 5070, runs llama-swap, Qdrant |
+| koala | GPU inference, heavy compute | RTX 5070, runs k3s + llama-swap + shared postgres18/pgvector |
 | iguana | Services, builds | M2 Ultra Mac |
 | flamingo | Daily driver, edge | Mac mini, ~/dev is here |
 
@@ -108,18 +121,64 @@ See `~/dev/PROJECT_SUMMARY.md` for detailed descriptions of each project.
 - **koala-ai-stack** (`AGENTS/`) — local AI server infrastructure management
 - **klimatkollen** (`XT/`) — Swedish municipal climate data platform
 
-## Knowledge base
+## Knowledge base — actively use it
 
-When available, agents can query the shared knowledge base:
+A persistent brain (BM25 search + LLM-synthesised Q&A) survives across sessions,
+hosts, and harnesses. It holds 100+ hard-won entries: infra incident postmortems,
+Go pitfalls, framework gotchas, design principles, ADRs. **It is not optional
+reference material — query it actively, not just when explicitly told.**
 
-- **MCP**: `mcp://hyperguild.<TAILNET>.ts.net:3100/knowledge`
-- **HTTP**: `http://hyperguild.<TAILNET>.ts.net:3100/api/v1/search`
+### When to query (treat as a reflex)
 
-<!-- TODO: replace <TAILNET> placeholder with the real Tailscale tailnet
-     name once hyperguild is deployed. Until then, agents that try to
-     reach the knowledge service on a host where it isn't running will
-     get DNS NXDOMAIN, which is the desired fail-loudly behavior. -->
-- **Scoping**: defaults to `public` collection; client projects filter to `{client}` + `public`
+- **Before** starting a non-trivial task — search for prior art with the symptom
+  AND the system component ("how did we solve X in Y?"). 5 seconds beats 5 hours.
+- **When debugging** — search for the error string, the stack frame, the affected
+  service. Past you may have already paid this tax.
+- **Before adopting** a pattern, library, framework, or model name — check if it
+  was tried and rejected, or what the integration footguns are.
+- **When making architectural decisions** — search for the domain + "ADR" or
+  "decision" to find prior reasoning before re-deriving it.
+- **When a recommendation feels novel** — challenge yourself: "has this been
+  documented?" The brain often has it.
+
+### When to write
+
+After you discover something that **future-you would forget** and that **isn't
+recoverable from the code, git log, or PR description alone**:
+
+- Bugs whose root cause is non-obvious and generalisable beyond this project.
+- Framework / library / model-name quirks that bit you and would bite anyone.
+- Design principles validated under fire (e.g. "every `_get` needs a `_list`").
+- Postmortems for incidents: what broke, why, how diagnosed, what to do next time.
+
+DON'T write project status, sprint progress, PR summaries, or "what I did this
+session" — those rot fast and the originals are in git/gitea anyway. Brain
+entries that age well are about *why*, *how to avoid*, and *what to do when*.
+
+### How to access (per harness)
+
+| Harness | Query | Write |
+|---------|-------|-------|
+| **Claude Code, Claude Desktop** | `brain_query` (BM25), `brain_answer` (LLM-synth + sources) MCP tools | `brain_write` MCP tool |
+| **Crush, Pi, Antigravity, other MCP-capable** | same MCP server: `ingestion-brain` (via the `mcp__*_brain__*` namespace once authenticated) | same |
+| **Anything HTTP-only (curl, scripts)** | `POST https://brain-mcp.d-ma.be/query` with `{"query":"..."}` (auth via `BRAIN_MCP_TOKEN`) | `POST .../write` with `{"content":"...","filename":"..."}` |
+| **Browser / human inspection** | `https://gitea.d-ma.be/mathias/hyperguild` → `knowledge/` and `wiki/` markdown files |
+
+- **Scoping**: defaults to `public` collection; client projects filter to `{client}` + `public`.
+- **Routing**: brain_answer's LLM uses berget.ai as primary, iguana ollama as
+  fallback. Both are configurable in the `supervisor/ingestion-deployment.yaml`
+  on the koala k3s cluster; don't hardcode local-only model names into the
+  berget URL (see knowledge entry on namespace mismatches).
+
+### Quick reflex checks
+
+If you find yourself about to say any of these out loud, you owe yourself a brain query first:
+
+- "I think the issue might be..."
+- "Let me try X and see..."
+- "I'll just write a script to..."
+- "This is probably a new bug..."
+- "Has anyone done this before?" — *yes, probably, go check.*
 
 ## Client work rules
 

.cursorrules

@@ -39,9 +39,18 @@ These rules apply to every task across every project, regardless of harness.
 4. **Goal-driven execution.** Define clear success criteria up front for every task.
    Loop — implement, verify, refine — until those criteria are met. Don't claim
    completion without evidence (tests pass, command output, observed behavior).
-5. **Branch-per-task for multi-agent repos.** When another agent may be active on
-   the same repo, create a branch (`agent/<description>`), commit there, and open a
-   PR. Do not merge without explicit instruction from Mathias.
+5. **Trunk-Based Development — commit directly to main.** Every commit is one
+   logical change (one tool, one fix, one test) with passing tests. Main is always
+   deployable. Never create long-lived feature branches.
+
+   **Exception — parallel agents on same repo:** If another agent is known to be
+   actively working on the same repo simultaneously, create a short-lived branch
+   (`agent/<description>`), finish the task, and merge to main within the same
+   session. Do not leave agent branches open between sessions.
+
+   **Exception — external contributor or client four-eyes requirement:** Use
+   PR flow only when a human reviewer outside the project is required. Document
+   the reason in PROJECT.md.
 
 ## Default stack
 
@@ -52,9 +61,10 @@ These rules apply to every task across every project, regardless of harness.
 | Build | Task (taskfile.dev) | Make | — |
 | Containers | Docker Compose (dev), k3s (prod) | — | — |
 | DB | PostgreSQL + sqlc | SQLite | — |
-| Search | Qdrant (vector), BM25 | — | — |
+| Search | pgvector (vector), BM25 | Qdrant (when >1M vectors or hybrid retrieval) | — |
 | Logging | slog (structured) | — | — |
 | Testing | Table-driven, testify | — | — |
+| Agents (Go) | google.golang.org/adk + pkg/litellm adapter | — | — |
 
 Exploratory: Rust, Zig — I'll tell you when I want these.
 
@@ -64,9 +74,12 @@ Exploratory: Rust, Zig — I'll tell you when I want these.
 - **Errors**: `fmt.Errorf("operation: %w", err)` — never naked, never log-and-return
 - **Naming**: stdlib conventions, no stuttering
 - **Architecture**: prefer stdlib over frameworks, constructor injection, env-var config parsed into typed structs
-- **Git**: conventional commits (`feat:`, `fix:`, `chore:`), one concern per PR, PR describes *why* not *what*
+- **Git**: conventional commits (`feat:`, `fix:`, `chore:`), commit directly to main,
+  one logical change per commit, CI is the quality gate
+- **Never**: long-lived feature branches, PRs for solo work, direct push without
+  passing `task check` locally first
 - **Security**: no secrets in code, govulncheck before adding deps, SOPS for encrypted config
-- **Dependencies**: prefer stdlib. testify, slog, templ, sqlc are pre-approved; anything else needs justification in the commit message
+- **Dependencies**: prefer stdlib. testify, slog, templ, sqlc, google.golang.org/adk (agent projects only) are pre-approved; anything else needs justification in the commit message
 
 ## Infrastructure
 
@@ -74,7 +87,7 @@ Three machines on Tailscale:
 
 | Machine | Role | Key specs |
 |---------|------|-----------|
-| koala | GPU inference, heavy compute | RTX 5070, runs llama-swap, Qdrant |
+| koala | GPU inference, heavy compute | RTX 5070, runs k3s + llama-swap + shared postgres18/pgvector |
 | iguana | Services, builds | M2 Ultra Mac |
 | flamingo | Daily driver, edge | Mac mini, ~/dev is here |
 
@@ -106,18 +119,64 @@ See `~/dev/PROJECT_SUMMARY.md` for detailed descriptions of each project.
 - **koala-ai-stack** (`AGENTS/`) — local AI server infrastructure management
 - **klimatkollen** (`XT/`) — Swedish municipal climate data platform
 
-## Knowledge base
+## Knowledge base — actively use it
 
-When available, agents can query the shared knowledge base:
+A persistent brain (BM25 search + LLM-synthesised Q&A) survives across sessions,
+hosts, and harnesses. It holds 100+ hard-won entries: infra incident postmortems,
+Go pitfalls, framework gotchas, design principles, ADRs. **It is not optional
+reference material — query it actively, not just when explicitly told.**
 
-- **MCP**: `mcp://hyperguild.<TAILNET>.ts.net:3100/knowledge`
-- **HTTP**: `http://hyperguild.<TAILNET>.ts.net:3100/api/v1/search`
+### When to query (treat as a reflex)
 
-<!-- TODO: replace <TAILNET> placeholder with the real Tailscale tailnet
-     name once hyperguild is deployed. Until then, agents that try to
-     reach the knowledge service on a host where it isn't running will
-     get DNS NXDOMAIN, which is the desired fail-loudly behavior. -->
-- **Scoping**: defaults to `public` collection; client projects filter to `{client}` + `public`
+- **Before** starting a non-trivial task — search for prior art with the symptom
+  AND the system component ("how did we solve X in Y?"). 5 seconds beats 5 hours.
+- **When debugging** — search for the error string, the stack frame, the affected
+  service. Past you may have already paid this tax.
+- **Before adopting** a pattern, library, framework, or model name — check if it
+  was tried and rejected, or what the integration footguns are.
+- **When making architectural decisions** — search for the domain + "ADR" or
+  "decision" to find prior reasoning before re-deriving it.
+- **When a recommendation feels novel** — challenge yourself: "has this been
+  documented?" The brain often has it.
+
+### When to write
+
+After you discover something that **future-you would forget** and that **isn't
+recoverable from the code, git log, or PR description alone**:
+
+- Bugs whose root cause is non-obvious and generalisable beyond this project.
+- Framework / library / model-name quirks that bit you and would bite anyone.
+- Design principles validated under fire (e.g. "every `_get` needs a `_list`").
+- Postmortems for incidents: what broke, why, how diagnosed, what to do next time.
+
+DON'T write project status, sprint progress, PR summaries, or "what I did this
+session" — those rot fast and the originals are in git/gitea anyway. Brain
+entries that age well are about *why*, *how to avoid*, and *what to do when*.
+
+### How to access (per harness)
+
+| Harness | Query | Write |
+|---------|-------|-------|
+| **Claude Code, Claude Desktop** | `brain_query` (BM25), `brain_answer` (LLM-synth + sources) MCP tools | `brain_write` MCP tool |
+| **Crush, Pi, Antigravity, other MCP-capable** | same MCP server: `ingestion-brain` (via the `mcp__*_brain__*` namespace once authenticated) | same |
+| **Anything HTTP-only (curl, scripts)** | `POST https://brain-mcp.d-ma.be/query` with `{"query":"..."}` (auth via `BRAIN_MCP_TOKEN`) | `POST .../write` with `{"content":"...","filename":"..."}` |
+| **Browser / human inspection** | `https://gitea.d-ma.be/mathias/hyperguild` → `knowledge/` and `wiki/` markdown files |
+
+- **Scoping**: defaults to `public` collection; client projects filter to `{client}` + `public`.
+- **Routing**: brain_answer's LLM uses berget.ai as primary, iguana ollama as
+  fallback. Both are configurable in the `supervisor/ingestion-deployment.yaml`
+  on the koala k3s cluster; don't hardcode local-only model names into the
+  berget URL (see knowledge entry on namespace mismatches).
+
+### Quick reflex checks
+
+If you find yourself about to say any of these out loud, you owe yourself a brain query first:
+
+- "I think the issue might be..."
+- "Let me try X and see..."
+- "I'll just write a script to..."
+- "This is probably a new bug..."
+- "Has anyone done this before?" — *yes, probably, go check.*
 
 ## Client work rules
 

AGENTS.md

@@ -36,9 +36,18 @@ These rules apply to every task across every project, regardless of harness.
 4. **Goal-driven execution.** Define clear success criteria up front for every task.
    Loop — implement, verify, refine — until those criteria are met. Don't claim
    completion without evidence (tests pass, command output, observed behavior).
-5. **Branch-per-task for multi-agent repos.** When another agent may be active on
-   the same repo, create a branch (`agent/<description>`), commit there, and open a
-   PR. Do not merge without explicit instruction from Mathias.
+5. **Trunk-Based Development — commit directly to main.** Every commit is one
+   logical change (one tool, one fix, one test) with passing tests. Main is always
+   deployable. Never create long-lived feature branches.
+
+   **Exception — parallel agents on same repo:** If another agent is known to be
+   actively working on the same repo simultaneously, create a short-lived branch
+   (`agent/<description>`), finish the task, and merge to main within the same
+   session. Do not leave agent branches open between sessions.
+
+   **Exception — external contributor or client four-eyes requirement:** Use
+   PR flow only when a human reviewer outside the project is required. Document
+   the reason in PROJECT.md.
 
 ## Default stack
 
@@ -49,9 +58,10 @@ These rules apply to every task across every project, regardless of harness.
 | Build | Task (taskfile.dev) | Make | — |
 | Containers | Docker Compose (dev), k3s (prod) | — | — |
 | DB | PostgreSQL + sqlc | SQLite | — |
-| Search | Qdrant (vector), BM25 | — | — |
+| Search | pgvector (vector), BM25 | Qdrant (when >1M vectors or hybrid retrieval) | — |
 | Logging | slog (structured) | — | — |
 | Testing | Table-driven, testify | — | — |
+| Agents (Go) | google.golang.org/adk + pkg/litellm adapter | — | — |
 
 Exploratory: Rust, Zig — I'll tell you when I want these.
 
@@ -61,9 +71,12 @@ Exploratory: Rust, Zig — I'll tell you when I want these.
 - **Errors**: `fmt.Errorf("operation: %w", err)` — never naked, never log-and-return
 - **Naming**: stdlib conventions, no stuttering
 - **Architecture**: prefer stdlib over frameworks, constructor injection, env-var config parsed into typed structs
-- **Git**: conventional commits (`feat:`, `fix:`, `chore:`), one concern per PR, PR describes *why* not *what*
+- **Git**: conventional commits (`feat:`, `fix:`, `chore:`), commit directly to main,
+  one logical change per commit, CI is the quality gate
+- **Never**: long-lived feature branches, PRs for solo work, direct push without
+  passing `task check` locally first
 - **Security**: no secrets in code, govulncheck before adding deps, SOPS for encrypted config
-- **Dependencies**: prefer stdlib. testify, slog, templ, sqlc are pre-approved; anything else needs justification in the commit message
+- **Dependencies**: prefer stdlib. testify, slog, templ, sqlc, google.golang.org/adk (agent projects only) are pre-approved; anything else needs justification in the commit message
 
 ## Infrastructure
 
@@ -71,7 +84,7 @@ Three machines on Tailscale:
 
 | Machine | Role | Key specs |
 |---------|------|-----------|
-| koala | GPU inference, heavy compute | RTX 5070, runs llama-swap, Qdrant |
+| koala | GPU inference, heavy compute | RTX 5070, runs k3s + llama-swap + shared postgres18/pgvector |
 | iguana | Services, builds | M2 Ultra Mac |
 | flamingo | Daily driver, edge | Mac mini, ~/dev is here |
 
@@ -103,18 +116,64 @@ See `~/dev/PROJECT_SUMMARY.md` for detailed descriptions of each project.
 - **koala-ai-stack** (`AGENTS/`) — local AI server infrastructure management
 - **klimatkollen** (`XT/`) — Swedish municipal climate data platform
 
-## Knowledge base
+## Knowledge base — actively use it
 
-When available, agents can query the shared knowledge base:
+A persistent brain (BM25 search + LLM-synthesised Q&A) survives across sessions,
+hosts, and harnesses. It holds 100+ hard-won entries: infra incident postmortems,
+Go pitfalls, framework gotchas, design principles, ADRs. **It is not optional
+reference material — query it actively, not just when explicitly told.**
 
-- **MCP**: `mcp://hyperguild.<TAILNET>.ts.net:3100/knowledge`
-- **HTTP**: `http://hyperguild.<TAILNET>.ts.net:3100/api/v1/search`
+### When to query (treat as a reflex)
 
-<!-- TODO: replace <TAILNET> placeholder with the real Tailscale tailnet
-     name once hyperguild is deployed. Until then, agents that try to
-     reach the knowledge service on a host where it isn't running will
-     get DNS NXDOMAIN, which is the desired fail-loudly behavior. -->
-- **Scoping**: defaults to `public` collection; client projects filter to `{client}` + `public`
+- **Before** starting a non-trivial task — search for prior art with the symptom
+  AND the system component ("how did we solve X in Y?"). 5 seconds beats 5 hours.
+- **When debugging** — search for the error string, the stack frame, the affected
+  service. Past you may have already paid this tax.
+- **Before adopting** a pattern, library, framework, or model name — check if it
+  was tried and rejected, or what the integration footguns are.
+- **When making architectural decisions** — search for the domain + "ADR" or
+  "decision" to find prior reasoning before re-deriving it.
+- **When a recommendation feels novel** — challenge yourself: "has this been
+  documented?" The brain often has it.
+
+### When to write
+
+After you discover something that **future-you would forget** and that **isn't
+recoverable from the code, git log, or PR description alone**:
+
+- Bugs whose root cause is non-obvious and generalisable beyond this project.
+- Framework / library / model-name quirks that bit you and would bite anyone.
+- Design principles validated under fire (e.g. "every `_get` needs a `_list`").
+- Postmortems for incidents: what broke, why, how diagnosed, what to do next time.
+
+DON'T write project status, sprint progress, PR summaries, or "what I did this
+session" — those rot fast and the originals are in git/gitea anyway. Brain
+entries that age well are about *why*, *how to avoid*, and *what to do when*.
+
+### How to access (per harness)
+
+| Harness | Query | Write |
+|---------|-------|-------|
+| **Claude Code, Claude Desktop** | `brain_query` (BM25), `brain_answer` (LLM-synth + sources) MCP tools | `brain_write` MCP tool |
+| **Crush, Pi, Antigravity, other MCP-capable** | same MCP server: `ingestion-brain` (via the `mcp__*_brain__*` namespace once authenticated) | same |
+| **Anything HTTP-only (curl, scripts)** | `POST https://brain-mcp.d-ma.be/query` with `{"query":"..."}` (auth via `BRAIN_MCP_TOKEN`) | `POST .../write` with `{"content":"...","filename":"..."}` |
+| **Browser / human inspection** | `https://gitea.d-ma.be/mathias/hyperguild` → `knowledge/` and `wiki/` markdown files |
+
+- **Scoping**: defaults to `public` collection; client projects filter to `{client}` + `public`.
+- **Routing**: brain_answer's LLM uses berget.ai as primary, iguana ollama as
+  fallback. Both are configurable in the `supervisor/ingestion-deployment.yaml`
+  on the koala k3s cluster; don't hardcode local-only model names into the
+  berget URL (see knowledge entry on namespace mismatches).
+
+### Quick reflex checks
+
+If you find yourself about to say any of these out loud, you owe yourself a brain query first:
+
+- "I think the issue might be..."
+- "Let me try X and see..."
+- "I'll just write a script to..."
+- "This is probably a new bug..."
+- "Has anyone done this before?" — *yes, probably, go check.*
 
 ## Client work rules
 

cmd/routing/main.go

@@ -18,9 +18,11 @@ import (
 	"github.com/mathiasbq/supervisor/internal/config"
 	iexec "github.com/mathiasbq/supervisor/internal/exec"
 	"github.com/mathiasbq/supervisor/internal/mcp"
+	"github.com/mathiasbq/supervisor/internal/mcpclient"
 	"github.com/mathiasbq/supervisor/internal/registry"
 	"github.com/mathiasbq/supervisor/internal/routing"
 	"github.com/mathiasbq/supervisor/internal/skills/debug"
+	"github.com/mathiasbq/supervisor/internal/skills/project"
 	"github.com/mathiasbq/supervisor/internal/skills/retrospective"
 	"github.com/mathiasbq/supervisor/internal/skills/review"
 	"github.com/mathiasbq/supervisor/internal/skills/trainer"
@@ -99,6 +101,21 @@ func main() {
 		CompleteFunc: trainer.CompleteFunc(wrap("trainer")),
 	}))
 
+	if cfg.GiteaMCPURL != "" {
+		reg.Register(project.New(project.Config{
+			Client:      mcpclient.New(cfg.GiteaMCPURL, cfg.GiteaMCPToken),
+			GiteaOwner:  cfg.GiteaOwner,
+			GitHubOwner: cfg.GitHubOwner,
+			GitHubPAT:   cfg.GitHubPAT,
+			InfraRepo:   cfg.InfraRepo,
+		}))
+		logger.Info("project_create registered", "gitea_mcp_url", cfg.GiteaMCPURL,
+			"gitea_owner", cfg.GiteaOwner, "github_owner", cfg.GitHubOwner,
+			"infra_repo", cfg.InfraRepo, "github_pat_set", cfg.GitHubPAT != "")
+	} else {
+		logger.Info("project_create skipped — GITEA_MCP_URL not set")
+	}
+
 	var validator *auth.Validator
 	if dexURL := os.Getenv("DEX_ISSUER_URL"); dexURL != "" {
 		audience := os.Getenv("MCP_AUDIENCE")

internal/config/routing.go

@@ -25,6 +25,16 @@ type RoutingConfig struct {
 	RouteLocalFloor    float64 // HYPERGUILD_ROUTE_LOCAL_FLOOR, default 0.90
 	RouteLocalCeil     float64 // HYPERGUILD_ROUTE_LOCAL_CEIL, default 0.70
 	PassRateTTLSeconds int     // HYPERGUILD_PASS_RATE_TTL_SECONDS, default 60
+
+	// project_create configuration. Empty GiteaMCPURL disables the
+	// project_create tool registration so the routing pod still starts
+	// in environments where it's not wired up.
+	GiteaMCPURL   string // GITEA_MCP_URL, e.g. http://koala:30340/mcp
+	GiteaMCPToken string // GITEA_MCP_TOKEN, bearer for gitea-mcp
+	GiteaOwner    string // GITEA_OWNER, default mathias
+	GitHubOwner   string // GITHUB_OWNER, default mathiasb
+	InfraRepo     string // INFRA_REPO, default infra
+	GitHubPAT     string // GITHUB_PAT, repo scope; never logged
 }
 
 func LoadRouting() (RoutingConfig, error) {
@@ -56,6 +66,13 @@ func LoadRouting() (RoutingConfig, error) {
 	}
 	cfg.PassRateTTLSeconds = ttl
 
+	cfg.GiteaMCPURL = os.Getenv("GITEA_MCP_URL")
+	cfg.GiteaMCPToken = os.Getenv("GITEA_MCP_TOKEN")
+	cfg.GiteaOwner = envOr("GITEA_OWNER", "mathias")
+	cfg.GitHubOwner = envOr("GITHUB_OWNER", "mathiasb")
+	cfg.InfraRepo = envOr("INFRA_REPO", "infra")
+	cfg.GitHubPAT = os.Getenv("GITHUB_PAT")
+
 	return cfg, nil
 }
 

internal/mcpclient/client.go

@@ -0,0 +1,135 @@
+// Package mcpclient is a minimal JSON-RPC over HTTP client for talking to
+// MCP servers from inside hyperguild components. It only implements
+// `tools/call` because that's all consumer skills need today.
+package mcpclient
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"time"
+)
+
+// Client calls an MCP server over Streamable HTTP / JSON-RPC.
+type Client struct {
+	url    string
+	token  string
+	http   *http.Client
+}
+
+// New returns a Client. token may be empty for unauthenticated servers.
+func New(url, token string) *Client {
+	return &Client{
+		url:   url,
+		token: token,
+		http:  &http.Client{Timeout: 60 * time.Second},
+	}
+}
+
+// WithHTTPClient overrides the underlying HTTP client (test injection).
+func (c *Client) WithHTTPClient(h *http.Client) *Client {
+	c.http = h
+	return c
+}
+
+type rpcRequest struct {
+	JSONRPC string         `json:"jsonrpc"`
+	ID      int            `json:"id"`
+	Method  string         `json:"method"`
+	Params  map[string]any `json:"params"`
+}
+
+type rpcError struct {
+	Code    int    `json:"code"`
+	Message string `json:"message"`
+}
+
+type rpcResponse struct {
+	JSONRPC string          `json:"jsonrpc"`
+	ID      int             `json:"id"`
+	Result  json.RawMessage `json:"result,omitempty"`
+	Error   *rpcError       `json:"error,omitempty"`
+}
+
+// Error is returned when the remote MCP server signals a typed failure.
+// Code follows JSON-RPC conventions; see gitea-mcp internal/mcp/jsonrpc.go
+// for the codes the server uses (e.g. -32002 NotFound, -32003 Conflict).
+type Error struct {
+	Code    int
+	Message string
+}
+
+func (e *Error) Error() string { return fmt.Sprintf("mcp error %d: %s", e.Code, e.Message) }
+
+// CallTool issues `tools/call`. result is JSON-unmarshalled from the
+// server's content[0].text field; pass nil to discard.
+func (c *Client) CallTool(ctx context.Context, name string, args any, result any) error {
+	body, err := json.Marshal(rpcRequest{
+		JSONRPC: "2.0",
+		ID:      1,
+		Method:  "tools/call",
+		Params: map[string]any{
+			"name":      name,
+			"arguments": args,
+		},
+	})
+	if err != nil {
+		return fmt.Errorf("marshal request: %w", err)
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, c.url, bytes.NewReader(body))
+	if err != nil {
+		return fmt.Errorf("new request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+	if c.token != "" {
+		req.Header.Set("Authorization", "Bearer "+c.token)
+	}
+
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return fmt.Errorf("http: %w", err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	raw, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return fmt.Errorf("read body: %w", err)
+	}
+	if resp.StatusCode >= 400 {
+		return fmt.Errorf("mcp http %d: %s", resp.StatusCode, string(raw))
+	}
+
+	var rpc rpcResponse
+	if err := json.Unmarshal(raw, &rpc); err != nil {
+		return fmt.Errorf("decode response: %w (body=%s)", err, string(raw))
+	}
+	if rpc.Error != nil {
+		return &Error{Code: rpc.Error.Code, Message: rpc.Error.Message}
+	}
+
+	if result == nil {
+		return nil
+	}
+
+	// MCP success result shape: { content: [{type:"text", text:"<json>"}] }
+	var wrap struct {
+		Content []struct {
+			Type string `json:"type"`
+			Text string `json:"text"`
+		} `json:"content"`
+	}
+	if err := json.Unmarshal(rpc.Result, &wrap); err != nil {
+		return fmt.Errorf("decode wrap: %w (result=%s)", err, string(rpc.Result))
+	}
+	if len(wrap.Content) == 0 {
+		return fmt.Errorf("empty content in tool response")
+	}
+	if err := json.Unmarshal([]byte(wrap.Content[0].Text), result); err != nil {
+		return fmt.Errorf("decode tool result text: %w (text=%s)", err, wrap.Content[0].Text)
+	}
+	return nil
+}

internal/mcpclient/client_test.go

@@ -0,0 +1,82 @@
+package mcpclient_test
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/mathiasbq/supervisor/internal/mcpclient"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCallTool_Success(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		assert.Equal(t, http.MethodPost, r.Method)
+		assert.Equal(t, "Bearer tok", r.Header.Get("Authorization"))
+		b, _ := io.ReadAll(r.Body)
+		var got map[string]any
+		_ = json.Unmarshal(b, &got)
+		assert.Equal(t, "tools/call", got["method"])
+		params := got["params"].(map[string]any)
+		assert.Equal(t, "x_y", params["name"])
+
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"jsonrpc":"2.0","id":1,"result":{"content":[{"type":"text","text":"{\"ok\":true,\"n\":7}"}]}}`))
+	}))
+	defer srv.Close()
+
+	c := mcpclient.New(srv.URL, "tok")
+	var out struct {
+		OK bool `json:"ok"`
+		N  int  `json:"n"`
+	}
+	err := c.CallTool(context.Background(), "x_y", map[string]any{"a": 1}, &out)
+	require.NoError(t, err)
+	assert.True(t, out.OK)
+	assert.Equal(t, 7, out.N)
+}
+
+func TestCallTool_RPCError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"jsonrpc":"2.0","id":1,"error":{"code":-32003,"message":"already exists"}}`))
+	}))
+	defer srv.Close()
+
+	c := mcpclient.New(srv.URL, "")
+	err := c.CallTool(context.Background(), "x", nil, nil)
+	require.Error(t, err)
+	var me *mcpclient.Error
+	require.True(t, errors.As(err, &me))
+	assert.Equal(t, -32003, me.Code)
+	assert.Contains(t, me.Message, "already exists")
+}
+
+func TestCallTool_HTTPError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusUnauthorized)
+		_, _ = w.Write([]byte(`unauthorized`))
+	}))
+	defer srv.Close()
+
+	c := mcpclient.New(srv.URL, "")
+	err := c.CallTool(context.Background(), "x", nil, nil)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "401")
+}
+
+func TestCallTool_NilResult(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"jsonrpc":"2.0","id":1,"result":{"content":[{"type":"text","text":"{}"}]}}`))
+	}))
+	defer srv.Close()
+
+	c := mcpclient.New(srv.URL, "")
+	require.NoError(t, c.CallTool(context.Background(), "x", nil, nil))
+}

internal/skills/project/handlers.go

@@ -0,0 +1,265 @@
+package project
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/mathiasbq/supervisor/internal/mcpclient"
+)
+
+type createArgs struct {
+	Name        string `json:"name"`
+	Description string `json:"description"`
+	Hypothesis  string `json:"hypothesis"`
+	Folder      string `json:"folder"`
+	Stack       string `json:"stack"`
+	Private     bool   `json:"private"`
+}
+
+type createResult struct {
+	GiteaURL  string `json:"gitea_url"`
+	GitHubURL string `json:"github_url"`
+	IssueURL  string `json:"issue_url"`
+	NextSteps string `json:"next_steps"`
+
+	// Reached records the steps that completed. Populated on partial failure
+	// so callers can resume manually instead of guessing what already ran.
+	Reached []string `json:"reached,omitempty"`
+
+	// FailedStep is non-empty when a downstream gitea-mcp call returned an
+	// error; the error itself is surfaced via the JSON-RPC error response,
+	// this field tells the operator which step it happened in.
+	FailedStep string `json:"failed_step,omitempty"`
+}
+
+func errUnknownTool(name string) error { return fmt.Errorf("unknown tool: %s", name) }
+
+// step names — must match what we surface in failed_step / reached.
+const (
+	stepCreateRepo  = "create_repo"
+	stepMirror      = "mirror"
+	stepInfraCommit = "infra_commit"
+	stepIssue       = "issue"
+)
+
+func (s *Skill) handleCreate(ctx context.Context, raw json.RawMessage) (json.RawMessage, error) {
+	var args createArgs
+	if err := json.Unmarshal(raw, &args); err != nil {
+		return nil, fmt.Errorf("parse args: %w", err)
+	}
+	if err := validate(args); err != nil {
+		return nil, err
+	}
+
+	tmpl := templateFor(args.Stack)
+	giteaURL := fmt.Sprintf("http://gitea.d-ma.be/%s/%s", s.cfg.GiteaOwner, args.Name)
+	githubURL := fmt.Sprintf("https://github.com/%s/%s", s.cfg.GitHubOwner, args.Name)
+
+	res := createResult{
+		GiteaURL:  giteaURL,
+		GitHubURL: githubURL,
+	}
+
+	// Step 1: create_project_from_template. If the repo already exists,
+	// gitea-mcp returns -32003 Conflict; we treat that as idempotent success
+	// and continue to the next steps so re-running self-heals partial runs.
+	existed, err := s.callCreateRepo(ctx, args, tmpl)
+	if err != nil {
+		return marshalPartial(res, stepCreateRepo, err)
+	}
+	res.Reached = append(res.Reached, stepCreateRepo)
+
+	// Step 2: configure push mirror to GitHub. Idempotent: if a mirror with
+	// the same remote already exists, gitea-mcp returns Conflict; we swallow it.
+	if err := s.callMirror(ctx, args.Name); err != nil {
+		if !isConflict(err) {
+			return marshalPartial(res, stepMirror, err)
+		}
+	}
+	res.Reached = append(res.Reached, stepMirror)
+
+	// Step 3: commit staging namespace manifest to infra repo. Done before
+	// the issue so the staging env is reconciling by the time the issue lands.
+	branch := fmt.Sprintf("staging/%s", args.Name)
+	if err := s.callInfraCommit(ctx, args.Name, branch); err != nil {
+		if !isConflict(err) {
+			return marshalPartial(res, stepInfraCommit, err)
+		}
+	}
+	res.Reached = append(res.Reached, stepInfraCommit)
+
+	// Step 4: open the experiment-brief issue on the new repo.
+	issueURL, err := s.callIssue(ctx, args, existed)
+	if err != nil {
+		return marshalPartial(res, stepIssue, err)
+	}
+	res.IssueURL = issueURL
+	res.Reached = append(res.Reached, stepIssue)
+
+	folder := args.Folder
+	if folder == "" {
+		folder = "."
+	}
+	res.NextSteps = fmt.Sprintf(
+		"cd ~/dev/%s/%s && task new-project -- %s personal %s %s && git remote add origin http://gitea.d-ma.be/%s/%s.git && git push -u origin main",
+		folder, args.Name, args.Name, folder, args.Stack, s.cfg.GiteaOwner, args.Name,
+	)
+
+	return marshalResult(res)
+}
+
+// callCreateRepo invokes create_project_from_template. Returns (existed, err)
+// where existed=true means the destination was already present and we should
+// treat it as a no-op success (idempotency).
+func (s *Skill) callCreateRepo(ctx context.Context, args createArgs, template string) (bool, error) {
+	var out struct {
+		HTMLURL string `json:"html_url"`
+	}
+	err := s.cfg.Client.CallTool(ctx, "create_project_from_template", map[string]any{
+		"owner":         s.cfg.GiteaOwner,
+		"name":          args.Name,
+		"description":   args.Description,
+		"private":       args.Private,
+		"template_name": template,
+	}, &out)
+	if err == nil {
+		return false, nil
+	}
+	if isConflict(err) {
+		return true, nil
+	}
+	return false, err
+}
+
+// callMirror configures the push mirror to GitHub.
+func (s *Skill) callMirror(ctx context.Context, name string) error {
+	remote := fmt.Sprintf("https://github.com/%s/%s.git", s.cfg.GitHubOwner, name)
+	return s.cfg.Client.CallTool(ctx, "repo_mirror_push", map[string]any{
+		"owner":           s.cfg.GiteaOwner,
+		"name":            name,
+		"action":          "add",
+		"remote_address":  remote,
+		"remote_username": s.cfg.GitHubOwner,
+		"remote_password": s.cfg.GitHubPAT,
+		"interval":        "8h0m0s",
+		"sync_on_commit":  true,
+	}, nil)
+}
+
+// callInfraCommit writes the staging namespace manifest into the infra repo
+// on a dedicated branch. Flux picks it up after merge.
+func (s *Skill) callInfraCommit(ctx context.Context, name, branch string) error {
+	manifest := stagingNamespaceManifest(name, time.Now().UTC().Format(time.RFC3339))
+	return s.cfg.Client.CallTool(ctx, "file_write_branch", map[string]any{
+		"owner":   s.cfg.GiteaOwner,
+		"name":    s.cfg.InfraRepo,
+		"path":    fmt.Sprintf("k3s/staging/%s/namespace.yaml", name),
+		"content": manifest,
+		"branch":  branch,
+		"base":    "main",
+		"message": fmt.Sprintf("feat(staging): add namespace for %s\n\nGenerated by hyperguild project_create.", name),
+	}, nil)
+}
+
+// callIssue opens the experiment-brief issue on the newly-created repo.
+// existed=true (repo pre-existed) still posts a new brief — repeated runs
+// can intentionally restate intent without colliding.
+func (s *Skill) callIssue(ctx context.Context, args createArgs, existed bool) (string, error) {
+	body := experimentBrief(args, existed)
+	var out struct {
+		HTMLURL string `json:"html_url"`
+	}
+	err := s.cfg.Client.CallTool(ctx, "issue_create", map[string]any{
+		"owner": s.cfg.GiteaOwner,
+		"name":  args.Name,
+		"title": "experiment brief: " + args.Description,
+		"body":  body,
+	}, &out)
+	if err != nil {
+		return "", err
+	}
+	return out.HTMLURL, nil
+}
+
+func stagingNamespaceManifest(name, createdAt string) string {
+	return fmt.Sprintf(`apiVersion: v1
+kind: Namespace
+metadata:
+  name: staging-%s
+  labels:
+    managed-by: hyperguild
+    project: %s
+    created-at: "%s"
+`, name, name, createdAt)
+}
+
+func experimentBrief(args createArgs, existed bool) string {
+	var b strings.Builder
+	b.WriteString("## Hypothesis\n\n")
+	b.WriteString(args.Hypothesis)
+	b.WriteString("\n\n## Description\n\n")
+	b.WriteString(args.Description)
+	b.WriteString("\n\n## Stack\n\n`")
+	b.WriteString(args.Stack)
+	b.WriteString("`\n\n## Provisioning\n\n")
+	b.WriteString("- Repo created from `template-")
+	b.WriteString(args.Stack)
+	b.WriteString("` on Gitea.\n")
+	b.WriteString("- Push-mirror configured to GitHub.\n")
+	b.WriteString("- Staging namespace manifest committed to infra repo.\n\n")
+	if existed {
+		b.WriteString("> Note: this repo already existed when `project_create` ran — provisioning steps were re-applied idempotently.\n")
+	}
+	return b.String()
+}
+
+func validate(args createArgs) error {
+	if args.Name == "" {
+		return errors.New("name is required")
+	}
+	if args.Description == "" {
+		return errors.New("description is required")
+	}
+	if args.Hypothesis == "" {
+		return errors.New("hypothesis is required")
+	}
+	if args.Stack != "go-agent" && args.Stack != "go-web" {
+		return fmt.Errorf("stack must be go-agent or go-web, got %q", args.Stack)
+	}
+	return nil
+}
+
+func templateFor(stack string) string {
+	switch stack {
+	case "go-agent":
+		return "template-go-agent"
+	default:
+		return "template-go-web"
+	}
+}
+
+func isConflict(err error) bool {
+	var me *mcpclient.Error
+	if errors.As(err, &me) && me.Code == -32003 {
+		return true
+	}
+	return false
+}
+
+func marshalResult(r createResult) (json.RawMessage, error) {
+	b, err := json.Marshal(r)
+	if err != nil {
+		return nil, fmt.Errorf("marshal result: %w", err)
+	}
+	return b, nil
+}
+
+func marshalPartial(r createResult, step string, inner error) (json.RawMessage, error) {
+	r.FailedStep = step
+	b, _ := json.Marshal(r)
+	return b, fmt.Errorf("project_create step %q failed: %w", step, inner)
+}

internal/skills/project/handlers_test.go

@@ -0,0 +1,244 @@
+package project_test
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"testing"
+
+	"github.com/mathiasbq/supervisor/internal/mcpclient"
+	"github.com/mathiasbq/supervisor/internal/skills/project"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// fakeGiteaMCP implements just enough of the JSON-RPC tools/call surface
+// to drive project_create end-to-end without an actual gitea-mcp server.
+type fakeGiteaMCP struct {
+	mu sync.Mutex
+	// Recorded calls in order.
+	Calls []recordedCall
+	// Per-tool response. Default is a generic success object.
+	Responses map[string]any
+	// Per-tool error response, takes precedence over Responses.
+	Errors map[string]rpcErr
+}
+
+type rpcErr struct {
+	Code    int
+	Message string
+}
+
+type recordedCall struct {
+	Tool string
+	Args map[string]any
+}
+
+func (f *fakeGiteaMCP) handler() http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var req struct {
+			ID     int             `json:"id"`
+			Params json.RawMessage `json:"params"`
+		}
+		_ = json.NewDecoder(r.Body).Decode(&req)
+		var p struct {
+			Name      string          `json:"name"`
+			Arguments json.RawMessage `json:"arguments"`
+		}
+		_ = json.Unmarshal(req.Params, &p)
+		var args map[string]any
+		_ = json.Unmarshal(p.Arguments, &args)
+
+		f.mu.Lock()
+		f.Calls = append(f.Calls, recordedCall{Tool: p.Name, Args: args})
+		errResp, hasErr := f.Errors[p.Name]
+		var resp any
+		if r, ok := f.Responses[p.Name]; ok {
+			resp = r
+		} else {
+			resp = map[string]any{"html_url": "http://gitea.example/" + p.Name}
+		}
+		f.mu.Unlock()
+
+		w.Header().Set("Content-Type", "application/json")
+		if hasErr {
+			body, _ := json.Marshal(map[string]any{
+				"jsonrpc": "2.0",
+				"id":      req.ID,
+				"error":   map[string]any{"code": errResp.Code, "message": errResp.Message},
+			})
+			_, _ = w.Write(body)
+			return
+		}
+		respText, _ := json.Marshal(resp)
+		body, _ := json.Marshal(map[string]any{
+			"jsonrpc": "2.0",
+			"id":      req.ID,
+			"result": map[string]any{
+				"content": []map[string]any{{"type": "text", "text": string(respText)}},
+			},
+		})
+		_, _ = w.Write(body)
+	})
+}
+
+func newSkill(t *testing.T, f *fakeGiteaMCP) *project.Skill {
+	t.Helper()
+	srv := httptest.NewServer(f.handler())
+	t.Cleanup(srv.Close)
+	return project.New(project.Config{
+		Client:      mcpclient.New(srv.URL, ""),
+		GiteaOwner:  "mathias",
+		GitHubOwner: "mathiasb",
+		GitHubPAT:   "ghp_test",
+		InfraRepo:   "infra",
+	})
+}
+
+func happyArgs() json.RawMessage {
+	return json.RawMessage(`{
+		"name":"my-experiment",
+		"description":"One-line desc",
+		"hypothesis":"We believe X produces Y",
+		"folder":"AGENTS",
+		"stack":"go-agent",
+		"private":true
+	}`)
+}
+
+func TestProjectCreate_HappyPath(t *testing.T) {
+	f := &fakeGiteaMCP{
+		Responses: map[string]any{
+			"issue_create": map[string]any{"html_url": "http://gitea.d-ma.be/mathias/my-experiment/issues/1"},
+		},
+	}
+	skill := newSkill(t, f)
+
+	out, err := skill.Handle(context.Background(), "project_create", happyArgs())
+	require.NoError(t, err)
+
+	var res map[string]any
+	require.NoError(t, json.Unmarshal(out, &res))
+	assert.Equal(t, "http://gitea.d-ma.be/mathias/my-experiment", res["gitea_url"])
+	assert.Equal(t, "https://github.com/mathiasb/my-experiment", res["github_url"])
+	assert.Equal(t, "http://gitea.d-ma.be/mathias/my-experiment/issues/1", res["issue_url"])
+	assert.Contains(t, res["next_steps"], "cd ~/dev/AGENTS/my-experiment")
+	assert.Contains(t, res["next_steps"], "git remote add origin")
+
+	// All four steps in order.
+	require.Len(t, f.Calls, 4)
+	assert.Equal(t, "create_project_from_template", f.Calls[0].Tool)
+	assert.Equal(t, "repo_mirror_push", f.Calls[1].Tool)
+	assert.Equal(t, "file_write_branch", f.Calls[2].Tool)
+	assert.Equal(t, "issue_create", f.Calls[3].Tool)
+
+	// template selection wired from stack
+	assert.Equal(t, "template-go-agent", f.Calls[0].Args["template_name"])
+	// mirror config
+	assert.Equal(t, "add", f.Calls[1].Args["action"])
+	assert.Equal(t, "https://github.com/mathiasb/my-experiment.git", f.Calls[1].Args["remote_address"])
+	assert.Equal(t, "ghp_test", f.Calls[1].Args["remote_password"])
+	// infra commit path
+	assert.Equal(t, "k3s/staging/my-experiment/namespace.yaml", f.Calls[2].Args["path"])
+	assert.Contains(t, f.Calls[2].Args["content"], "name: staging-my-experiment")
+	assert.Contains(t, f.Calls[2].Args["content"], "managed-by: hyperguild")
+	// PAT must NOT appear in the response
+	assert.NotContains(t, string(out), "ghp_test")
+}
+
+func TestProjectCreate_Idempotent_RepoExists(t *testing.T) {
+	f := &fakeGiteaMCP{
+		Errors: map[string]rpcErr{
+			"create_project_from_template": {Code: -32003, Message: "already exists"},
+		},
+		Responses: map[string]any{
+			"issue_create": map[string]any{"html_url": "http://gitea.d-ma.be/mathias/my-experiment/issues/1"},
+		},
+	}
+	skill := newSkill(t, f)
+
+	out, err := skill.Handle(context.Background(), "project_create", happyArgs())
+	require.NoError(t, err)
+
+	var res map[string]any
+	require.NoError(t, json.Unmarshal(out, &res))
+	assert.Equal(t, "http://gitea.d-ma.be/mathias/my-experiment", res["gitea_url"])
+	assert.Equal(t, "http://gitea.d-ma.be/mathias/my-experiment/issues/1", res["issue_url"])
+
+	// Still ran all 4 steps; idempotent flow falls through the conflict.
+	require.Len(t, f.Calls, 4)
+}
+
+func TestProjectCreate_MirrorFails(t *testing.T) {
+	f := &fakeGiteaMCP{
+		Errors: map[string]rpcErr{
+			"repo_mirror_push": {Code: -32000, Message: "github unreachable"},
+		},
+	}
+	skill := newSkill(t, f)
+
+	out, err := skill.Handle(context.Background(), "project_create", happyArgs())
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), `"mirror" failed`)
+
+	var res map[string]any
+	require.NoError(t, json.Unmarshal(out, &res))
+	assert.Equal(t, "mirror", res["failed_step"])
+	reached := res["reached"].([]any)
+	assert.Equal(t, []any{"create_repo"}, reached)
+
+	// Only steps 1 + 2 actually called.
+	require.Len(t, f.Calls, 2)
+}
+
+func TestProjectCreate_InfraCommitFails(t *testing.T) {
+	f := &fakeGiteaMCP{
+		Errors: map[string]rpcErr{
+			"file_write_branch": {Code: -32000, Message: "write rejected"},
+		},
+	}
+	skill := newSkill(t, f)
+
+	out, err := skill.Handle(context.Background(), "project_create", happyArgs())
+	require.Error(t, err)
+
+	var res map[string]any
+	require.NoError(t, json.Unmarshal(out, &res))
+	assert.Equal(t, "infra_commit", res["failed_step"])
+	reached := res["reached"].([]any)
+	assert.Equal(t, []any{"create_repo", "mirror"}, reached)
+	require.Len(t, f.Calls, 3)
+}
+
+func TestProjectCreate_ValidationErrors(t *testing.T) {
+	f := &fakeGiteaMCP{}
+	skill := newSkill(t, f)
+	cases := []struct {
+		name string
+		body string
+		want string
+	}{
+		{"missing name", `{"description":"d","hypothesis":"h","stack":"go-agent"}`, "name"},
+		{"missing description", `{"name":"x","hypothesis":"h","stack":"go-agent"}`, "description"},
+		{"missing hypothesis", `{"name":"x","description":"d","stack":"go-agent"}`, "hypothesis"},
+		{"bad stack", `{"name":"x","description":"d","hypothesis":"h","stack":"python"}`, "stack"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := skill.Handle(context.Background(), "project_create", json.RawMessage(tc.body))
+			require.Error(t, err)
+			assert.True(t, strings.Contains(err.Error(), tc.want), "want %q in %v", tc.want, err)
+		})
+	}
+	assert.Empty(t, f.Calls, "no upstream calls should occur on validation failure")
+}
+
+func TestProjectCreate_UnknownTool(t *testing.T) {
+	f := &fakeGiteaMCP{}
+	skill := newSkill(t, f)
+	_, err := skill.Handle(context.Background(), "nope", happyArgs())
+	require.Error(t, err)
+}

internal/skills/project/skill.go

@@ -0,0 +1,90 @@
+// Package project implements the `project_create` MCP tool: a single-call
+// pipeline that creates a Gitea repo from a template, configures push-mirror
+// to GitHub, commits a staging namespace manifest to the infra repo, and
+// opens an experiment-brief issue on the new repo. See hyperguild gitea
+// issue #10 for the design.
+package project
+
+import (
+	"context"
+	"encoding/json"
+
+	"github.com/mathiasbq/supervisor/internal/mcpclient"
+	"github.com/mathiasbq/supervisor/internal/registry"
+)
+
+// Config holds the orchestration dependencies for the project skill.
+type Config struct {
+	// Client talks to the gitea-mcp server. project_create makes 4 sequential
+	// calls (create_project_from_template, repo_mirror_push, file_write_branch,
+	// issue_create) through this client.
+	Client *mcpclient.Client
+
+	// GiteaOwner is the org/user that owns the new repo and the infra repo
+	// the namespace manifest is committed to (typically "mathias").
+	GiteaOwner string
+
+	// GitHubOwner is the GitHub org/user the push-mirror targets
+	// (typically "mathiasb").
+	GitHubOwner string
+
+	// GitHubPAT is the personal access token used as the push-mirror
+	// password. Must have `repo` scope. Never logged.
+	GitHubPAT string
+
+	// InfraRepo is the name of the infra repo on Gitea where the
+	// k3s/staging/<name>/namespace.yaml manifest gets committed
+	// (typically "infra").
+	InfraRepo string
+}
+
+// Skill exposes project_create as an MCP tool.
+type Skill struct{ cfg Config }
+
+// New constructs the project Skill.
+func New(cfg Config) *Skill { return &Skill{cfg: cfg} }
+
+// Name returns the skill identifier.
+func (s *Skill) Name() string { return "project" }
+
+// Tools returns the MCP tool definitions for this skill.
+func (s *Skill) Tools() []registry.ToolDef {
+	schema, _ := json.Marshal(map[string]any{
+		"type": "object",
+		"properties": map[string]any{
+			"name": map[string]any{
+				"type":        "string",
+				"pattern":     `^[a-z][a-z0-9-]{1,38}[a-z0-9]$`,
+				"description": "Lowercase repo name. 3-40 chars, must start with a letter.",
+			},
+			"description": map[string]any{"type": "string"},
+			"hypothesis":  map[string]any{"type": "string"},
+			"folder": map[string]any{
+				"type":        "string",
+				"description": "Informational only — appears in next_steps. Example: AGENTS, AI, QKX.",
+			},
+			"stack": map[string]any{
+				"type":        "string",
+				"enum":        []string{"go-agent", "go-web"},
+				"description": "Selects template-go-agent or template-go-web.",
+			},
+			"private": map[string]any{"type": "boolean"},
+		},
+		"required": []string{"name", "description", "hypothesis", "stack"},
+	})
+	return []registry.ToolDef{
+		{
+			Name:        "project_create",
+			Description: "Bootstrap a new project: Gitea repo from template, GitHub push-mirror, staging namespace manifest, experiment-brief issue. Idempotent — re-running with an existing repo returns the existing URLs.",
+			InputSchema: schema,
+		},
+	}
+}
+
+// Handle dispatches the tool call.
+func (s *Skill) Handle(ctx context.Context, tool string, args json.RawMessage) (json.RawMessage, error) {
+	if tool != "project_create" {
+		return nil, errUnknownTool(tool)
+	}
+	return s.handleCreate(ctx, args)
+}