@zerwiz/ymir 0.1.4 → 0.1.7

package/.agents/skills/galdr-ymirsystem/assets/harness-integration/README.md

@@ -0,0 +1,736 @@
+# Harness Integration — the Ymir harness adapter matrix and adapter contract
+
+> **Purpose:** Define how each agent harness (OpenCode, Pi, Claude Code, Cursor, Codex, Grok) seats **Brokk** before the first turn, arms **Sýn** supervision, guards the turn end, and enforces the PreToolUse seatbelts — and give a step-by-step recipe to add a new harness.
+
+This is the reference an agent reads to add a harness adapter or rebuild one from scratch. It is paired with one deep-dive file per harness:
+
+| Deep dive | Harness |
+|---|---|
+| [`opencode.md`](opencode.md) | OpenCode (Öppencode) |
+| [`pi.md`](pi.md) | Pi (`@earendil-works/pi-coding-agent`) |
+| [`claude-code.md`](claude-code.md) | Claude Code |
+| [`cursor.md`](cursor.md) | Cursor |
+| [`codex.md`](codex.md) | Codex |
+| *(none — not implemented)* | Grok |
+
+Runtime inventory that every adapter calls: [`../runtime-components.md`](../runtime-components.md).
+The governing plan is [`docs/plans/29-brokk-distro-runtime.md`](../../../../../docs/plans/29-brokk-distro-runtime.md) (referred to below as **plan 29**), whose §12 carries the Norse naming law.
+
+---
+
+## 1. What a harness adapter is
+
+> **MCP scope (well · mesh · Teams · Anchor).** `bin/a2a-mcp.sh install` wires the
+> `engram` (memory well) and `a2abridge` (A2A mesh) MCP servers into OpenCode's
+> repo `opencode.json` **and** Pi. The **Teams** plane (`wayofteams`) and
+> **Anchor** memory are **remote** MCP servers named by URL: set
+> `WAYOFTEAMS_MCP_URL` / `ANCHOR_MCP_URL` in the private platform env, then run
+> `bin/a2a-mcp.sh install`. OpenCode speaks `type: remote` natively; Pi has no
+> remote transport, so it reaches them through the `mcp-remote` stdio bridge. The
+> URLs never enter the tracked tree. By default the Pi side writes the **global**
+> `~/.pi/agent/mcp.json`; `bin/a2a-mcp.sh install --project` writes the repo's
+> `.pi/mcp.json` instead, leaving a Pi used elsewhere untouched (launch with
+> `pi --mcp-config .pi/mcp.json`). Prefer `--project` when integrating Ymir into
+> an existing workflow. `bin/a2a-mcp.sh show` reports what is **actually** wired —
+> every key present, not a fixed list.
+
+
+The Ymir runtime is a **distro**: a directory of instructions, skills, tooling and conventions that turns a general-purpose agent into a specialized one. Launching a supported harness inside `BROKK_HOME` is supposed to instantiate **Brokk** and address the operator as the **Allfather** *before the model's first turn*.
+
+A harness will not do that on its own. Each harness exposes a different set of lifecycle hooks, and each adapter is a thin, harness-specific shim that binds those hooks to **harness-agnostic runtime scripts** in `bin/`. The shim never re-implements runtime logic: it only decides *when* to call a script and *how* to deliver the result.
+
+```
+ harness lifecycle event  ──►  adapter shim  ──►  bin/<owner script>  ──►  state/ data/
+ (session open / idle /          (.opencode,        (Sága, Sýn, Rödd,
+  turn end / tool call)           .pi, hooks.json)   Gleipnir, ...)
+```
+
+Every adapter must satisfy the **four-part adapter contract** (§3). A harness with no adapter is *unverified* and must never be used for a dispatched Eindri (`bin/einherjar-spawn.sh` fails closed).
+
+---
+
+## 2. Run-tier vs nudge-tier
+
+The tier answers **how the session-start digest reaches model context**.
+
+| Tier | Meaning | Contract |
+|---|---|---|
+| **run-tier** | The harness's session-open adapter can **execute** a command and inject its stdout into model context (or as the session's first context) before the model acts. | Adapter calls `bin/saga-sessionstart-run.sh`; the digest text is delivered natively. Nothing is left to the model's discretion. |
+| **nudge-tier** | The harness can only post a message into the session and **ask** the model to take the helm. The model can defer. | Adapter calls `bin/saga-sessionstart-run.sh` (which emits the digest *or* a one-line "run `bin/saga-session-start.sh` now" instruction) and delivers it as a prompt/follow-up. |
+
+Important nuance: **tier is per-event, not per-harness.** All current harnesses are run-tier for the session-open digest *except* the disabled/Grok path. But `bin/saga-sessionstart-run.sh` deliberately emits a **nudge** for context-preserving opens (`resume`, `reload`, `fork`) even on a run-tier harness, because re-running the whole digest is redundant when prior context is restored:
+
+```bash
+# bin/saga-sessionstart-run.sh:70-72
+resume|reload|fork)
+  printf 'Run `%s` exactly once now, before executing any other instructions.\n' "bash $SCRIPT_DIR/saga-session-start.sh"
+  ;;
+```
+
+So the digest is *run* on `startup`/`new`, *re-emitted* on `clear`/`compact` after a completed startup, and *nudged* on `resume`/`reload`/`fork`.
+
+### Tier + mechanism per harness
+
+| Harness | Tier | Session-open mechanism | Source value(s) seen |
+|---|---|---|---|
+| OpenCode | run (nudge delivery) | plugin `session.created` runs `saga-sessionstart-run.sh`, injects stdout through `client.session.promptAsync` | (wrapper chooses from no payload) |
+| Pi | run | extension `before_agent_start` returns a `syn-sessionstart-nudge` message built from the wrapper | `startup`, `clear`, `resume`, `fork`, `compact` |
+| Claude Code | run | `SessionStart` hook stdout is injected; wrapper reads `source` from the hook JSON on stdin | `startup`, `resume`, `clear`, `compact` |
+| Cursor | run | `sessionStart` hook returns JSON `{"additional_context": "..."}` | `--source startup` (explicit) |
+| Codex | run | `SessionStart` hook pipes the JSON payload into `saga-sessionstart-run.sh` | parsed from payload |
+| Grok | nudge (upstream only) | upstream uses `.grok/hooks/*sessionstart-nudge.json` | — (not implemented in Ymir) |
+
+---
+
+## 3. The adapter contract (four required parts + one optional)
+
+Every production adapter implements the same four responsibilities. The harness-specific *how* lives in the deep-dive files; the *what* is fixed here.
+
+### Part 1 — Session-open injection
+
+Bind the harness's "session is opening" event to `bin/saga-sessionstart-run.sh`.
+
+```bash
+# The one command every session-open adapter invokes.
+"$BROKK_ROOT/bin/saga-sessionstart-run.sh" [--source <src>] [--pi-prerequisite]
+```
+
+- `--source` ∈ `startup|new|clear|compact|resume|reload|fork`; absent → parse a Claude/Codex-shaped JSON hook payload on stdin, default `startup`.
+- `--pi-prerequisite` is Pi-only: an intentional stand-down exits **3** so Pi provider preflight can tell it apart from an eligible attempt that produced nothing.
+- **Every ordinary transport path exits 0.** A failed session start must reach the agent as digest text it can act on, never as a refusal to open the session. (Claude `SessionStart` exit 2 blocks initialization.)
+- The script records `state/.session-start-complete` so `clear`/`compact` can re-emit instead of re-running.
+
+### Part 2 — Watch arm (Sýn supervision)
+
+Bind a "session is idle / turn ended" event to arm **one** `bin/syn-watch-arm.sh` cycle, and let the adapter own **continuity** (re-arming) so no model tokens are spent on ordinary re-arms.
+
+```bash
+bin/syn-watch-arm.sh --restart
+```
+
+The arm script polls `state/` every `BROKK_WATCH_POLL_SECONDS` (default 5), writes `state/.watch.heartbeat`, and exits with an actionable line when the wake queue or a `*.signal`/`*.check`/`.watcher-stop` appears:
+
+```
+signal: wake queue
+signal: <name>
+check: <name>
+stale: watcher stopped by operator
+```
+
+It refuses to arm with exit 0 and a `watcher: read-only - the session helm is held by
+another live session` message **only** when a genuinely live other session holds the helm.
+A **vacant** helm — no owner, a dead/zombie/recycled owner, or a truncated/empty
+lock — is entered in place: the arm script runs `gleipnir_lock_acquire` (which
+refuses only a verifiably live owner), and the Pi extension classifies an empty
+lock as `missing` so `gna_watch_arm`'s reclaim takes it directly. A vacant helm
+is never punted to a manual session start; only a real live contender is refused.
+
+- **Pi** owns continuity in `gna-pi-watch.ts` (Gná); the model gets a **tool** `gna_watch_arm` and a command `/gna-watch-arm` for the first cycle or a repair, never for ordinary re-arms. Gná's liveness is **zombie-aware**: `kill(pid, 0)` alone passes a zombie and a recycled pid, so the watcher reads `/proc/<pid>/stat` — the state character (Z/X = dead) and the starttime recorded at lock acquire (field 22; a mismatch means the kernel handed the pid to an unrelated process) — answers every ordinary wake: a **vacant** helm is taken (arm script via `gleipnir_lock_acquire`; Gná classifies an empty lock as `missing` so the reclaim takes it in place), while only a verifiably live other owner is refused as read-only. It drops its own lock on real process exit (`releaseLockIfOwned`). Gná's wake delivery is **echo-guarded**: an identical watcher message is sent at most once per drained state — when both durable queues (`state/.wake-queue` and the FM runner's `.agents/state/.wake-queue`) are empty, a repeat is the same news the primary already drained and is not re-delivered, so a long stale window cannot flood the follow-up queue.
+- **OpenCode** owns continuity in `syn-watch-arm.js` (Sýn); the coordinator is published under `globalThis.__brokkOpenCodeWatchArm` so the turn-end guard can consult it first.
+- **Claude Code** uses the `Stop` hook with `"asyncRewake": true` and a long timeout to keep the arm running out of band.
+- **Codex / Cursor** run `syn-turnend-guard.sh` at Stop; they do **not** own a long-lived arm (no auto re-arm).
+
+### Part 3 — Turn-end guard
+
+Bind the harness's "turn is about to end" event to `bin/syn-turnend-guard.sh`. If the guard prints the recovery instruction and exits **2**, the adapter must re-prompt instead of letting the turn end blind.
+
+```bash
+bin/syn-turnend-guard.sh [--claude]
+# reads Claude/Codex-shaped {"stop_hook_active":false} on stdin (ignored)
+# exit 0 = supervision healthy OR never armed (inert)
+# exit 2 = supervision was armed and the watcher is missing/stale; stderr carries recovery
+```
+
+The guard is **inert until the first successful arm** writes `state/.supervision-armed`:
+
+```bash
+# bin/syn-turnend-guard.sh:20
+[ -f "$STATE/.supervision-armed" ] || exit 0
+```
+
+It goes stale when `now - state/.watch.heartbeat > BROKK_WATCH_HEARTBEAT_STALE_SECONDS` (default 60).
+
+### Part 4 — PreToolUse seatbelts
+
+Bind the harness's pre-tool event for shell commands to the two owner scripts. Exit **2** (or throw) blocks the tool call.
+
+```bash
+bin/syn-arm-pretool-check.sh --command "<bash command>"   # denies backgrounding the watcher arm
+bin/syn-cd-pretool-check.sh  --command "<bash command>"   # denies a persistent escaping `cd`
+```
+
+Both are **v0 inert-by-default** contracts: `syn-arm-pretool-check.sh` blocks only a command that backgrounds `syn-watch-arm.sh`; `syn-cd-pretool-check.sh` blocks only `cd .../../`. Owner scripts hold the policy; adapters only relay.
+
+**A third seatbelt guards the assets** — `bin/syn-asset-pretool-check.sh` denies an
+`edit`/`write` of a governed path until its owning asset has been read in the
+session (reads are recorded in `state/asset-reads`).
+
+```bash
+bin/syn-asset-pretool-check.sh --path "<file>"          # exit 2 when the asset is not loaded
+bin/syn-asset-pretool-check.sh --note "<asset path>"     # record an asset as read
+```
+
+Who governs what:
+
+```
+governed[6]{path,load_first}:
+  "bin/ymir-install.sh",".agents/skills/galdr-ymirsystem/assets/installation.md"
+  "apps/hlidskjalf/**",".agents/skills/galdr-ymirsystem/assets/hlidskjalf-ui.md"
+  "bin/mimir*",".agents/skills/galdr-ymirsystem/assets/memory-well.md"
+  "bin/nornir-* | config/cron.yaml",".agents/skills/galdr-ymirsystem/assets/nornir-jobs.md"
+  "bin/valknut-load.sh | .pi/** | .opencode/**",".agents/skills/galdr-ymirsystem/assets/harness-integration/README.md"
+  "bin/smidja* | .agents/skills/smidja/**",".agents/skills/galdr-ymirsystem/assets/smidja.md"
+```
+
+The Pi extension `.pi/extensions/syn-turnend-guard.ts` relays it: `read` of an
+asset is recorded, and an `edit`/`write` of a governed path is blocked with the
+reason. The guard's own liveness check is **zombie-aware** too — `/proc/<pid>/stat`
+field 3 (Z/X = a killed-but-unreaped or dead holder) fails a stale lock even when
+`kill(0)` would report the pid alive, with `kill(0)` as the fallback when `/proc`
+is unavailable. The same routes are stated in `AGENTS.md` and printed in the session
+digest under `ASSET ROUTING`, and `compliance-check.sh` fails on a governed file
+changed without its asset (the `assets` gate).
+
+### Part 5 (optional) — Re-arm / stop auto-arm
+
+Some harnesses cannot synchronously own the watcher at Stop. The supported pattern is an **async rewrite** hook (Claude Code `"asyncRewake": true`; Cursor `followup_message` loop) that re-runs the arm and surfaces actionable output on stderr. If the harness has no such mechanism, do not fake it — declare the harness **turn-end-only** (Codex, Cursor) and rely on `einherjar-spawn.sh` fail-closed dispatch.
+
+---
+
+## 4. The adapter matrix
+
+Legend: ✅ implemented · ⚠️ partial/inert-by-design · ❌ not implemented.
+
+| Harness | Session-open injection | Watch arm owner | Turn-end guard | Pretool (arm) | Pretool (cd) | Re-arm |
+|---|---|---|---|---|---|---|
+| **OpenCode** | ✅ `.agents/harness/opencode/plugins/saga-sessionstart.js` | ✅ `.agents/harness/opencode/plugins/syn-watch-arm.js` (`session.idle`, coordinator key) | ✅ `.agents/harness/opencode/plugins/syn-turnend-guard.js` | ✅ `syn-pretool-check.js` | ✅ `syn-cd-check.js` | ✅ plugin-owned retries |
+| **Pi** | ✅ `.pi/extensions/syn-turnend-guard.ts` (`before_agent_start`) | ✅ `.pi/extensions/gna-pi-watch.ts` (`gna_watch_arm` tool) | ✅ same extension (`agent_settled`) | ✅ same extension (`tool_call`) | ✅ same extension (`tool_call`) | ✅ Gná retry/backoff |
+| **Claude Code** | ✅ `.claude/settings.json` `SessionStart` | ✅ `Stop` + `syn-watch-arm.sh --restart` (`asyncRewake`) | ✅ `Stop` + `syn-turnend-guard.sh --claude` | ❌ | ❌ | ✅ async rewake |
+| **Codex** | ✅ `.codex/hooks.json` `SessionStart` | ⚠️ no long-lived arm | ✅ `Stop` + `syn-turnend-guard.sh` | ✅ `PreToolUse` matcher `Bash` | ✅ `PreToolUse` matcher `Bash` | ❌ |
+| **Cursor** | ✅ `.cursor/hooks.json` `sessionStart` | ⚠️ interactive only | ✅ `stop` → `{"followup_message":...}` | ✅ `preToolUse` matcher `Shell` | ✅ `preToolUse` matcher `Shell` | ❌ |
+| **Grok** | ❌ no `.grok/` in Ymir | ❌ | ❌ | ❌ | ❌ | ❌ |
+
+### Adapter files at a glance
+
+| Path | Exports / shape | Contract parts |
+|---|---|---|
+| `.agents/harness/opencode/plugins/saga-sessionstart.js` | `export const SagaSessionstart` | 1 |
+| `.agents/harness/opencode/plugins/syn-watch-arm.js` | `export const SynWatchArm` | 2 |
+| `.agents/harness/opencode/plugins/syn-turnend-guard.js` | `export const SynTurnendGuard` | 2 (consult coordinator), 3 |
+| `.agents/harness/opencode/plugins/syn-pretool-check.js` | `export const SynPretoolCheck` | 4 (arm) |
+| `.agents/harness/opencode/plugins/syn-cd-check.js` | `export const SynCdCheck` | 4 (cd) |
+| `.agents/harness/opencode/plugins/lib/rodd-operational-input.js` | `encodeRoddOperationalInput(root, kind, content)` | shared wire |
+| `.pi/extensions/syn-turnend-guard.ts` | `export default function (pi: ExtensionAPI)` | 1, 3, 4 |
+| `.pi/extensions/gna-pi-watch.ts` | `export default function (pi: ExtensionAPI)` | 2 |
+| `.pi/extensions/ro.ts` | Ró — the calm presentation preference (`/calm`), state/ro | user |
+| `.pi/shared/extensions/open-editor.ts` | `/edit [path]` and `ctrl+shift+e` — opens files from cwd in the Allfather's editor; strictly user-facing, no LLM tool. **Resolution:** `$VISUAL` → `$EDITOR` → the first editor that exists (`code cursor zed subl nvim vim hx helix nano micro emacs vi`), so a host that is not Omarchy — where Omarchy's own launcher or a bare `vi` may be absent — still gets a working editor instead of an ENOENT | user |
+| `.pi/shared/extensions/herdr-agent-state.ts` | reports pane agent lifecycle state to herdr | 2 |
+| `.pi/shared/extensions/todo.ts` | the todo surface | user |
+| `.pi/shared/extensions/ymir-subagents.ts` | **the Eindri roster as a Pi tool** — reads the canonical `.agents/agents/*.md` tree and exposes every figure through a `subagent` tool (and a `/subagents` command). See "Pi has no agent loader" below | user |
+| `.pi/shared/extensions/lib/rodd-operational-input.ts` | `encodeRoddOperationalInput`, `classifyRoddOperationalText`, `classifyRoddCurrentOperationalText` | shared wire |
+| `.pi/extensions/lib/vordr-sessionstart-supervisor.mjs` | detached child supervisor (Vörðr) | 1 (transport) |
+| `.claude/settings.json` | `hooks.SessionStart[]`, `hooks.Stop[]` | 1, 3, 5 |
+| `.codex/hooks.json` | `hooks.SessionStart[]`, `hooks.PreToolUse[]`, `hooks.Stop[]` | 1, 3, 4 |
+| `.cursor/hooks.json` | `hooks.sessionStart[]`, `hooks.preToolUse[]`, `hooks.stop[]` | 1, 3, 4 |
+
+---
+
+## 5. The environment contract
+
+Every adapter and script resolves the same runtime roots. The overrides exist so tests and realm homes can relocate the runtime without touching the launcher.
+
+| Variable | Meaning | Default |
+|---|---|---|
+| `BROKK_ROOT_OVERRIDE` | tracked code root (holds `bin/`, `AGENTS.md`) | resolved from script location |
+| `BROKK_HOME` | private home (holds `data/ state/ config/`) | `BROKK_ROOT_OVERRIDE` or repo root |
+| `BROKK_STATE_OVERRIDE` | state dir | `$BROKK_HOME/state` |
+| `BROKK_DATA_OVERRIDE` | data dir | `$BROKK_HOME/data` |
+| `BROKK_CONFIG_OVERRIDE` | config dir | `$BROKK_HOME/config` |
+| `BROKK_REALM` | active realm | `data/realm.md` line 1, else `way-of` |
+| `BROKK_SESSION_PID` | **live harness PID** bound by the session lock | `${BROKK_SESSION_PID:-$$}` |
+| `BROKK_SESSIONSTART_INELIGIBLE` | `1` makes Pi's prerequisite exit 3 | unset |
+
+`BROKK_SESSION_PID` is the load-bearing variable. The lock must bind to the **live harness process**, not the short-lived digest helper: Pi/OpenCode adapters pass `BROKK_SESSION_PID=String(process.pid)` into the spawned digest/arm child (`.pi/extensions/syn-turnend-guard.ts:277`, `.agents/harness/opencode/plugins/syn-watch-arm.js:428`). `bin/gleipnir-lock-lib.sh` writes it to `state/.lock`.
+
+Full variable inventory: [`../runtime-components.md`](../runtime-components.md) §5.
+
+---
+
+## 6. Fail-closed dispatch rules
+
+Dispatch is owned by `bin/einherjar-spawn.sh` (Norse: **Einherjar**). It **never silently falls back** to an unverified harness.
+
+| Rule | Where | Behavior |
+|---|---|---|
+| Verified adapter set | `bin/einherjar-spawn.sh:53` | `VERIFIED_HARNESSES='opencode pi pi-signed'` |
+| Unverified harness | `bin/einherjar-spawn.sh:224-232` | `error: harness '<h>' is not verified for direct launch` → exit 1 |
+| Missing executable | `bin/einherjar-spawn.sh:233-236` | `error: harness '<h>' executable not found on PATH` → exit 1 |
+| Missing auth/dependency | adapter scripts | report a plain reason; never a silent fallback |
+| Raw launch escape hatch | `bin/einherjar-spawn.sh:207-211` | a `--harness` value containing whitespace is treated as a raw command, warns loudly, and is *not* on the verified set |
+| Active dispatch profiles | `bin/einherjar-spawn.sh:212-214` | if `config/eindri-dispatch.json` exists, a fresh spawn **must** pass an explicit `--harness` resolved from its rules |
+| Brief/flag drift | `bin/einherjar-spawn.sh:327-335` | refuses a ship launch whose brief's `Delivery contract: mode=<m>` disagrees with `--mode` |
+| Isolation dependency | `bin/einherjar-spawn.sh:340-359` | `--isolation on` requires docker + `utgard-runner:latest`, else exit 1; `auto` reports the decision |
+
+**Adding a harness to the verified set is a deliberate act:** implement all four contract parts, smoke-test them, then add the adapter name to `VERIFIED_HARNESSES` in `bin/einherjar-spawn.sh` (see §8).
+
+---
+
+## 7. How to add a new harness (step-by-step)
+
+A harness is a runtime that can host Brokk. Adding one is a five-stage task.
+
+### Stage 0 — Inventory the harness's lifecycle
+
+Answer these before writing anything:
+
+1. Does a **session-open** hook exist, and can it inject stdout into context (run-tier) or only post a message (nudge-tier)?
+2. Does an **idle/turn-end** event exist for the guard?
+3. Can a **long-lived arm process** be spawned at Stop, or only a synchronous hook?
+4. Is there a **pre-tool / before-tool** event for shell commands, and can it block?
+5. What env var names does the harness export (project dir, workspace root, agent marker)?
+
+### Stage 1 — Add detection to Hamr
+
+`bin/hamr-harness.sh` (Norse: **Hamr**, "the shape a being wears") is the single source of harness identity. Two layers, in precedence order:
+
+- **Layer 1 — verified env markers** (`bin/hamr-harness.sh:112-135`). Add your marker *after* the Cursor/Claude/Pi/Grok markers, documenting the version you verified it on.
+- **Layer 2 — process ancestry** (`bin/hamr-harness.sh:136-169`). Add a `*<name>*` case to the `comm` matcher and to the bare-interpreter `args` matcher.
+
+```bash
+# bin/hamr-harness.sh — Layer 1 pattern
+[ "${MYHARNESS_AGENT:-}" = "1" ] && { echo myharness; return; }
+# Layer 2 pattern
+case "$(basename -- "$comm")" in
+  *myharness*) echo myharness; return ;;
+esac
+```
+
+Reuse the shared Cursor identity helpers (`hamr_cursor_process_matches`, `hamr_cursor_argv0_for_pid`) if your harness shares an interpreter.
+
+### Stage 2 — Write the adapter
+
+Pick the closest existing adapter in the table (§4) and mirror its structure. The rules are non-negotiable:
+
+- **Resolve the root** the same way: prefer `worktree` → `git -C <dir> rev-parse --show-toplevel` → `realpathSync(dir)`.
+- **Call only `bin/` owner scripts.** Never re-implement digest, arm, guard, or classification logic.
+- **Session-open:** call `bin/saga-sessionstart-run.sh`. Never `bin/saga-session-start.sh` directly from an adapter.
+- **Watch arm:** require a live lock. If the arm reports `read-only`, stop; do not retry.
+- **Turn-end:** re-prompt on exit 2. On any other exit, do nothing.
+- **Seatbelts:** relay exit 2 as a block.
+- **Encode operational text** with the Rödd bridge (`lib/rodd-operational-input.*`) before delivery.
+
+### Stage 3 — Register the hooks
+
+Create the harness's config file (`.<harness>/...`) matching the real schema. See the deep-dive file for the exact JSON shape per harness. The absolute rule: **the hook must be self-locating** (use the harness's own project-dir env var or `pwd -P`) and must **no-op safely** when the runtime is absent.
+
+### Stage 4 — Verify (all four parts)
+
+Run the commands in §9 and confirm the four contract parts. Do not mark the adapter done on a happy-path session-open alone; a guard that never fires is indistinguishable from a guard that is not wired.
+
+### Stage 5 — Mark verified + document
+
+1. Add the harness name to `VERIFIED_HARNESSES` in `bin/einherjar-spawn.sh:53`.
+2. Add a row to the matrix in §4 of this file and a deep-dive file next to it.
+3. If you want it as the default for dispatched Eindri, set `config/eindri-harness` (one line: `<harness> [<model>] [<effort>]`).
+4. Record the verified version and any marker quirks in this file's Gotchas and in `data/learnings.md`.
+
+---
+
+## 8. Adding a new harness to `VERIFIED_HARNESSES`
+
+The verified set is the fail-closed gate. Edit exactly this line and nothing else:
+
+```bash
+# bin/einherjar-spawn.sh:53
+VERIFIED_HARNESSES='opencode pi pi-signed'
+```
+
+Notes:
+
+- `pi-signed` is a Pi variant (`BROKK_PI_HARNESS=pi-signed` / `FM_PI_HARNESS`); it maps to `pi` in `einherjar-spawn.sh`'s inline fallback but is a distinct token for the launch command (`bin/einherjar-spawn.sh:419-430`).
+- Adding a raw command is **not** a substitute: `--harness "mycmd --flag"` bypasses verification and warns.
+
+---
+
+## 9. Verification
+
+### Syntax and schema
+
+```bash
+cd "$BROKK_HOME"
+for f in bin/*.sh; do bash -n "$f" || echo "SYNTAX FAIL: $f"; done
+python3 -m json.tool config/eindri-dispatch.json >/dev/null && echo "dispatch JSON ok"
+python3 -m json.tool .claude/settings.json   >/dev/null && echo "claude JSON ok"
+python3 -m json.tool .codex/hooks.json       >/dev/null && echo "codex JSON ok"
+python3 -m json.tool .cursor/hooks.json      >/dev/null && echo "cursor JSON ok"
+```
+
+### Session-open injects, Go digest
+
+```bash
+# Run as the harness would; it must print the Sága digest and exit 0.
+bin/saga-sessionstart-run.sh --source startup | head -n 20
+echo "exit=$?"
+
+# Re-emit path after a completed startup:
+test -f state/.session-start-complete && bin/saga-sessionstart-run.sh --source compact | head -n 3
+```
+
+### Watch arm arms, heartbeat, and exits on a wake
+
+```bash
+# In one shell (must own the lock):
+bin/syn-watch-arm.sh --restart        # prints "watcher: started pid=... recovery-generation=..."
+test -f state/.supervision-armed && echo "armed marker written"
+test -s state/.watch.heartbeat && echo "heartbeat written"
+# In another shell, drop a wake and watch the arm exit with "signal: wake queue":
+: > state/.wake-queue
+```
+
+### Turn-end guard
+
+```bash
+# Inert before first arm:
+rm -f state/.supervision-armed
+echo '{"stop_hook_active":false}' | bin/syn-turnend-guard.sh; echo "expected 0, got $?"
+# Armed but stale heartbeat → exit 2:
+touch state/.supervision-armed
+rm -f state/.watch.heartbeat
+echo '{"stop_hook_active":false}' | bin/syn-turnend-guard.sh; echo "expected 2, got $?"
+```
+
+### Seatbelts
+
+```bash
+bin/syn-arm-pretool-check.sh --command 'bin/syn-watch-arm.sh --restart &'; echo "expected 2, got $?"
+bin/syn-arm-pretool-check.sh --command 'bin/syn-watch-arm.sh --restart';  echo "expected 0, got $?"
+bin/syn-cd-pretool-check.sh  --command 'cd ../../etc';                    echo "expected 2, got $?"
+bin/syn-cd-pretool-check.sh  --command 'ls -la';                          echo "expected 0, got $?"
+```
+
+### Session lock ownership
+
+```bash
+bin/saga-session-start.sh | sed -n '1,12p'   # LOCK section: held vs READ-ONLY
+```
+
+---
+
+## 10. Gotchas
+
+- **The digest helper must never hold the lock.** Bind `BROKK_SESSION_PID` to the live harness process or a second session will read the lock as stale and mutate shared state.
+- **`saga-session-start.sh` acquires the lock but never releases it.** Release is the harness's business (process exit). A refused lock means **read-only**: no spawn, steer, merge, drain, or repair.
+- **Don't call `bin/saga-session-start.sh` directly from an adapter.** Use `bin/saga-sessionstart-run.sh`; only the wrapper knows the source-routing rules (`startup` runs, `clear`/`compact` re-emits, `resume`/`reload`/`fork` nudges).
+- **The turn-end guard is inert until `.supervision-armed` exists.** A guard test that never arms proves nothing. Arm first, then force a stale heartbeat.
+- **`session.idle` fires often.** On OpenCode, the watch-arm plugin acts first and publishes `globalThis.__brokkOpenCodeWatchArm`; the guard plugin consults it and only calls `bin/syn-turnend-guard.sh` when the coordinator could not arm. Calling the guard first produces a false "turn would end blind" prompt during normal re-arms.
+- **Never background the arm from the model's bash tool.** `syn-arm-pretool-check.sh` denies it; the adapter owns continuity. The arm is a plugin/extension child, not a tool call.
+- **Exit-code semantics differ by host.** Claude's `SessionStart` exit 2 blocks initialization, so the run wrapper always exits 0 on the transport path; Pi's prerequisite uses exit 3 internally so preflight can distinguish a stand-down from a silent failure.
+- **`hamr-harness.sh` has no `crew` subcommand — its subcommand is `eindri`.** `bin/einherjar-spawn.sh:170` calls `hamr-harness.sh crew`, which falls through to `detect_own` (`bin/hamr-harness.sh:231-236`); see §11. Until that call is fixed, `config/eindri-harness` is only honoured by the inline fallback, not the Hamr path.
+- **Plugin auto-loading is path-based, not declared.** OpenCode loads `.agents/harness/opencode/plugins/*.js` automatically; there is no `plugin` array in `opencode.json`. The `package.json` only declares `"type": "module"`.
+- **Grok is not wired in Ymir.** Upstream Brokk ships `.grok/hooks/*.json`; Ymir has no `.grok/`. Treat Grok as unverified (see §11).
+- **Windows timing.** OpenCode and Pi raise the arm-ready budget to 35s on `win32` so a slow Git Bash cold start is not SIGTERMed mid-confirmation (`.agents/harness/opencode/plugins/syn-watch-arm.js:17-19`, `.pi/extensions/gna-pi-watch.ts:91-94`).
+
+---
+
+## 11. Known code-vs-doc disagreements (truth is the code)
+
+| Plan 29 says | Code says (truth) |
+|---|---|
+| `.agents/harness/opencode/plugins/syn-sessionstart.js` | actual file is `.agents/harness/opencode/plugins/saga-sessionstart.js` |
+| `.agents/harness/opencode/plugins/gna-watch-arm.js` (plan §1) | actual file is `.agents/harness/opencode/plugins/syn-watch-arm.js` (OpenCode's watcher is **Sýn**, not Gná; Gná exists only on Pi) |
+| Plan §13 "still to build: `bin/hamr-harness.sh`; the `.opencode`/Claude/Codex/Cursor adapters; `bin/einherjar-spawn.sh`; …" | all of these are **now landed**; plan §13 is stale |
+| Digest §6 lists 9 stages including **NETWORK CHECKS** and **FLEET DIGEST from `data/backlog.md`** | `bin/saga-session-start.sh` currently prints **8** sections (LOCK, BOOTSTRAP, WAKE QUEUE, SUPERVISION, FLEET DIGEST, CONTEXT DIGEST, CRON START, NEXT STEP); there is **no NETWORK CHECKS section**, and the fleet digest counts `state/*.meta` + `docs/masterplan.md` orders, not `data/backlog.md` |
+| Plan §7 "Cursor … run interactive only (no headless turn-end)" | `.cursor/hooks.json` **does** implement `sessionStart` (`additional_context`) and `stop` (`followup_message`) |
+| Plan §7 "Codex … bounded foreground checkpoint … nudge-tier" | `.codex/hooks.json` implements run-tier `SessionStart` + `PreToolUse` + `Stop` |
+| Plan §7 "Grok — project hooks (`grok --trust`)" | no `.grok/` exists in Ymir; Grok is unimplemented |
+| Plan §7 "Claude Code … SessionStart + Stop hooks" | correct, but Ymir's `.claude/settings.json` has **no `PreToolUse`** seatbelt entries (upstream Brokk does) |
+
+The authoritative interface is always the code listed in [`../runtime-components.md`](../runtime-components.md); report drift there.
+
+---
+
+## 12. Provenance
+
+The adapter pattern is ported from the validated upstream **Brokk** agent-distro harness adapters (`bin/fm-harness.sh`, `bin/fm-sessionstart-run.sh`, `.agents/harness/opencode/plugins/fm-primary-*.js`, `.pi/extensions/fm-primary-*.ts`, `.claude/settings.json`, `.codex/hooks.json`, `.cursor/hooks.json`, `.grok/hooks/*.json`). Ymir adopts the *mechanism* and renames every component per plan 29 §12 (Sága, Sýn, Gná, Vörðr, Rödd, Gleipnir, Hamr, Einherjar, Erindi, Vör, Nornir). The upstream names (`fm-*`, "Allfather") are provenance only and must never name a Ymir component.
+
+---
+
+## 13. The Well — memory for every harness (engram)
+
+The well is **Mimirsbrunn**, backed by the validated OSS engine **engram**. It is
+one repo-local store, `.agents/memory/kaia.engram`, reached two ways:
+
+**HTTP bridge (`:4602`)** — `bin/mimir-bridge.py`, raised by
+`bin/mimir-bridge.sh` (and by `scripts/start.sh` + `bin/saga-session-start.sh`).
+It wraps the engram library so the gate API, `bin/mimir.sh`, and
+`bin/mimir-ingest.sh` can drink over the small HTTP contract:
+
+```
+GET  /health                 -> {status, store, episodes, agents}
+GET  /recall?q&k&mode        -> {results:[{score, distance, episode}]}
+GET  /recent?limit           -> {episodes:[...]}
+GET  /timeline?entity        -> {facts:[...]}
+GET  /inspect                -> {episodes, agents, store}
+POST /observe {content,...}  -> {id}
+```
+
+**MCP server (stdio)** — `engram-mcp --db <store> --agent-id <harness>`, so any
+MCP-capable harness gets `remember`, `recall`, `why`, `forget`, `stats` with no
+integration code. Registered per harness (each writes under its own agent id):
+
+| Harness | Config | Agent id |
+|---|---|---|
+| OpenCode | `opencode.json` → `mcp.engram` (+ `~/.config/opencode/opencode.json`) | `opencode` |
+| Pi | `.pi/mcp.json` (pass `pi --mcp-config .pi/mcp.json`) | `pi` |
+| Claude Code | `~/.claude.json` → `mcpServers.engram` | `claude` |
+| Cursor | `~/.cursor/mcp.json` → `mcpServers.engram` | `cursor` |
+| Codex | `~/.codex/config.toml` → `[mcp_servers.engram]` | `codex` |
+
+The engine needs the `mcp<2` SDK for `engram-mcp` (v2 renamed `FastMCP` to
+`MCPServer`, which breaks engram 1.x/2.x):
+`python3 -m pip install --user --break-system-packages 'mcp<2'`.
+
+Rule: **drink before you act, water it after** — recall on the way in, and
+`POST /observe` (or the `remember` MCP tool) after a lesson lands.
+
+---
+
+## 14. Agent locations — `.agents/agents` is the source of truth
+
+Agent profiles live **only** in `.agents/agents/*.md` (the canonical, with the
+harness config like `mode`/`model`/`permission` in their frontmatter). The
+harness directories **bind** them by symlink — they are never hand-written
+duplicates:
+
+- OpenCode: `.opencode/agents/<name>.md` → `../../.agents/agents/<profile>.md`
+- Pi: `.pi/agents/<profile>.md` → the same canonical files
+
+`bin/valknut-load.sh` creates the symlinks (`--opencode`, `--pi`, `--global`,
+`--all`). **Every harness gets EVERY agent** — claude, codex, cursor and opencode
+are bound by the same loader pass as pi, so no harness is left holding a
+hand-made subset. Naming differs by harness and must be respected:
+
+```
+agent_binding[5]{harness,dir,name_rule}:
+  "opencode",".opencode/agents/","the frontmatter `name:` — bragi.md -> bragi-marketer.md"
+  "pi",".pi/agents/","the profile file name — bragi-marketer.md"
+  "claude",".claude/agents/","the profile file name"
+  "codex",".codex/agents/","the profile file name"
+  "cursor",".cursor/agents/","the profile file name"
+```
+
+To change an agent, edit `.agents/agents/*.md` and re-run the loader; never edit
+a harness directory (they are all links).
+
+### Skill location — the same law, one tree
+
+Skills have **one** tree too: `.agents/skills/`. Every harness reaches it, each
+by the mechanism that harness actually reads — and a harness that silently
+loads nothing is invisible from the code, so `compliance-check.sh`'s `harnesses`
+gate asserts all four:
+
+| Harness | How it reaches `.agents/skills` |
+|---|---|
+| **opencode** | `skills.paths: [".agents/skills"]` in `opencode.json` (the loader merges it in and re-asserts it on every run) |
+| **pi** | **native discovery** — it walks up from the cwd to `.agents/skills` (and `~/.agents/skills`). No link, no config: a second root under `.pi/` would invite double-loading, exactly as the extensions rule warns |
+| **claude · codex · cursor** | `<harness>/skills -> ../.agents/skills`, created by `bin/valknut-load.sh` (their project scope is their own directory) |
+
+**Do not** copy a `SKILL.md` into a harness directory: a copy is drift, and a
+nested `SKILL.md` carrying frontmatter is loaded as a *second, phantom* skill by
+every recursive scanner (the `harnesses` gate refuses it).
+
+### opencode.json has two writers — and neither may overwrite
+
+`opencode.json` is untracked (it carries absolute paths) and two things write it:
+
+```
+writers[2]{writer,owns}:
+  "bin/valknut-load.sh","STRUCTURE — the base keys, the agent blocks, the skills path; rendered from opencode.json.example"
+  "bin/agents-config.sh apply","the ROSTER — providers and per-agent models, from config/agents.yaml"
+```
+
+**Both merge; neither overwrites.** The loader's `config_out` adds missing keys (
+deep, `setdefault`-style), ensures `skills.paths`, and re-asserts nothing else;
+`agents-config` does the same for providers and models — and it writes **only
+OpenCode agents** into `opencode.json`. A `pi` (or `hermes`) agent's model is that
+harness's own id (e.g. `llamacpp/qwen3.5-9b`), which OpenCode cannot resolve, so
+those agents are deliberately left out rather than handed to OpenCode broken. This is not tidiness —
+the loader used to re-render from the example with `sed` + `mv`, and because the
+example carried only `llama.cpp`, **every loader run deleted the Apodex
+provider**, which lived only in the live file. A blind render of a file two
+writers share is a silent data loss; the merge is the fix.
+
+## 15. The Eindri roster (bound agents)
+
+Every skill names an owner agent in `.agents/skills/README.md`; the fleet below
+is the full set of canonical profiles, all bound into all five harnesses.
+
+Canonical profiles in `.agents/agents/*.md`, bound as symlinks:
+
+```
+einherjar[20]{figure,craft,domain,engine}:
+  "Brokk","primary — the bellows","ymirlabs","—"
+  "Sindri","developer / smith","brokkforge","Chrome DevTools"
+  "Bragi","marketer / skald","utgard","Firecrawl + browser-use (+ Scrapy)"
+  "Hnoss","designer / shaper","utgard","OpenDesign"
+  "Huginn","researcher / sage","muninn","—"
+  "Mímir","planner / the wise","ymirlabs","—"
+  "Forseti","reviewer / the just","runestone","—"
+  "Snotra","documenter / the wise-woman","runestone","—"
+  "Kvasir","scout / the knowing","ymirlabs","—"
+  "Galdr","builder — CLI ergonomics","ymirlabs","—"
+  "Týr","judge — the 10 principles + the gates","runestone","—"
+  "Sága","seeress — bearings + recap","ymirlabs","—"
+  "Muninn","rememberer — memory curation","muninn","—"
+  "Urðr","fate — the hold lifecycle","ymirlabs","—"
+  "Frigg","knowing — consent gate","ymirlabs","—"
+  "Vör","aware — diagnostics","ymirlabs","—"
+  "Sýn","seeing — stuck-worker recovery","ymirlabs","—"
+  "Jörð","grounded — project registry","ymirlabs","—"
+  "Gróa","renewer — self-update","ymirlabs","—"
+  "Völundr","master smith — Smíðja's orchestrator","brokkforge","—"
+```
+
+Bind with `bin/valknut-load.sh --all` (OpenCode: `.opencode/agents/<name>.md`; Pi:
+`.pi/agents/<profile>.md`) — never edit the harness dirs.
+
+## Pi has no agent loader (2026-09-17)
+
+**Pi core does not load `.pi/agents/`.** Agent loading in Pi is a *package*
+(`pi-agents`, `pi-agent-mode`, `pi-simple-agents`), not a core feature, and Ymir
+installs none of them. So `.pi/agents/` held twenty correct profile links that
+**nothing in Pi ever read** — the same failure as OpenCode's singular
+`.opencode/agent/`, arrived at from the other side.
+
+The fix lives in the repo, not in a root-pi package: **`.pi/shared/extensions/ymir-subagents.ts`**
+discovers the canonical `.agents/agents/*.md` tree itself and registers a
+`subagent` tool. A call runs the chosen figure as a nested model call in the
+current session — the figure's markdown body is its system prompt, its
+frontmatter `model:` picks the model where the machine serves it.
+
+```
+subagent({ agent: "kvasir", task: "find every AGENTS.md" })   # dispatch
+subagent({})                                                   # list the roster
+/subagents                                                     # list the roster
+```
+
+Rules this extension follows, learned the hard way:
+
+- **No imports.** `@earendil-works/pi-coding-agent` is **not installed as a
+  package**, so an extension that imports its types cannot load at all. Every
+  working extension in this tree takes `pi` as `any` and declares tool
+  `parameters` as a plain JSON schema object. `typebox` *is* installed, but the
+  plain object needs nothing.
+- **One home.** It lives in `.pi/shared/extensions/` (deployed), never in
+  `.pi/extensions/` — a copy in both makes pi refuse the duplicate tool.
+- **The canonical tree is the source.** It reads `.agents/agents/`; it never
+  copies from it (Rule 02).
+
+### A rename that left a reader behind (same day)
+
+`skuld-branch-supervision.ts` imported `calmTranscriptClassIsVisible` and
+`CalmPresentationState` from `./lib/ro-visibility.ts`, but that module exports
+`roTranscriptClassIsVisible` and `RoPresentationState` — the names were renamed
+and the importer was not. Pi refuses the **whole extension** at load with
+`does not provide an export named …`, so Skuld's supervision branch was dead in
+every session while nothing reported it. Fixed in the same change.
+
+**Check every extension actually loads**, not just that its file exists:
+
+```bash
+cd ~/.pi/agent/extensions
+for f in *.ts; do
+  node --input-type=module --eval "import('file://$PWD/$f').then(()=>console.log('$f LOADS')).catch(e=>console.log('$f FAIL: '+e.message.split('\n')[0]))"
+done
+```
+
+Run it after every deploy. A module that cannot resolve is invisible from the
+file listing, and pi reports it only as a startup line that scrolls away.
+
+
+
+Pi loads **both** the project `.pi/extensions/` and the global
+`~/.pi/agent/extensions/` directories; because it does not de-duplicate by
+extension or tool name, an extension present in both makes `pi` exit with
+`Tool "<name>" conflicts with …` and **no agent can be seated** (herdr reports
+"the pane must sit at an interactive shell prompt", a confusing symptom).
+
+Rule: **one home per extension.**
+
+- Shared extensions (`todo.ts`, `herdr-agent-state.ts`, `open-editor.ts`) live
+  **only** in `~/.pi/agent/extensions/` — used by every project, Ymir and
+  Omarchy alike.
+- The repo `.pi/extensions/` holds **only Ymir-unique** extensions
+  (`gna-pi-watch.ts`, `ro.ts`, `skuld-branch-supervision.ts`,
+  `syn-turnend-guard.ts`, `lib/`).
+- Never copy a shared extension back into the repo; that re-arms the collision.
+
+**Shipping the shared extensions (added 2026-09-12).** A rule that keeps the
+extensions outside the repo would also keep them from every new operator, so the
+repo carries their **source** at `.pi/shared/extensions/` — a path pi does not
+load, therefore collision-free — and the loader deploys it into the single home:
+
+```bash
+bin/valknut-load.sh --pi      # deploys .pi/shared/extensions/*.ts -> ~/.pi/agent/extensions/
+```
+
+The deploy copies (never links: a broken link would silently disable a tool),
+skips files that are already identical, and reports what it placed. It runs
+whenever the Pi surface is bound, with or without `--global`. Adding a shared
+extension is therefore two steps: put the source in `.pi/shared/extensions/`, and
+let the loader place it — **never** place it in `.pi/extensions/`.
+
+## OpenCode agent `tools` key (2026-09-12)
+
+OpenCode requires the frontmatter `tools:` key to be an **object** of its own
+built-in tools. Our profiles list Ymir capabilities (`vector_db`,
+`hermes_runner`, `herder`, `yggdrasil`, `supabase`, `opendesign`,
+`chrome_devtools`), which are not OpenCode tools and which OpenCode rejects as
+`Expected object | undefined, got [...] tools`. Those lists are therefore carried
+under **`ymir_tools:`**; permission is expressed by the `permission:` block.
+
+## Canonical layout — the ONE remap (authoritative; do not redo)
+
+**Source of truth is `.agents/`. Harness directories are symlinks, never copies.**
+
+```
+canonical[3]{kind,canonical,load_path}:
+  "agent profiles",".agents/agents/<profile>.md",".opencode/agents/<name>.md · .pi/agents/<name>.md (symlinks)"
+  "OpenCode plugins",".agents/harness/opencode/plugins/",".opencode/plugins (symlink)"
+  "skills",".agents/skills/","loaded via opencode.json skills.paths"
+```
+
+Rules:
+- **Edit only `.agents/`.** Never edit a file under `.opencode/` or `.pi/` — those are
+  symlinks to `.agents/` (e.g. `.opencode/agents/bragi.md` links to
+  `.agents/agents/bragi-marketer.md`).
+- After any change: `bin/valknut-load.sh --all` to rebind.
+- **OpenCode requires REAL `.opencode/{node_modules,package.json,.gitignore}`** — the
+  dependency store. Symlinking them hit ELOOP + untracked files and was reverted.
+- So `.opencode` holds: symlinked `agent/` + `plugins`, and the real dep store. It
+  cannot be deleted; it is already minimized. No agent/profile content lives there.
+
+## Asking the Allfather — `ask_user_question` (2026-09-12)
+
+The system can ask the Allfather instead of guessing. Adopted (open-source-first)
+from `@juicesharp/rpiv-ask-user-question` (MIT), pinned **project-locally** in
+`.pi/settings.json` so every machine/user that has the repo gets it.
+
+- Gives Pi one tool, `ask_user_question`: a terminal dialog of up to four
+  questions with authored options (+ descriptions/previews), a free-text row,
+  per-question and global notes, and a Submit review tab.
+- Non-interactive runs simply don't see the tool (never a failing call).
+- Config (read-only): `~/.config/rpiv-ask-user-question/config.json`.
+- Use it at real decision points — where a wrong assumption costs a rework.
+
+**Two extension paths, and they are not interchangeable** (2026-09-12):
+
+| Path | Loaded by | Holds |
+|---|---|---|
+| `.pi/extensions/` | pi, as **project-local** extensions | the governance set: `syn-turnend-guard.ts`, `gna-pi-watch.ts`, `ro.ts`, `skuld-branch-supervision.ts` |
+| `.pi/shared/extensions/` | **deployed** by `bin/valknut-load.sh --pi` into `${HOME}/.pi/agent/extensions/` | the user-facing trio: `open-editor.ts`, `herdr-agent-state.ts`, `todo.ts` |
+
+The same extension in **both** paths makes pi exit with a tool-name conflict and
+no agent can be seated — which is why the shared trio has exactly one home and is
+deployed, never duplicated. Editing a path from the table above without checking
+which of the two it is will edit a file that is not there.