eltmon · eltmon · Aug 12, 2026 · Aug 12, 2026 · Aug 12, 2026 · Aug 12, 2026
diff --git a/configuration/context-layers.mdx b/configuration/context-layers.mdx
@@ -83,6 +83,7 @@ harness; spawn-time plumbing delivers it:
 | **Claude Code** | managed region of `~/.claude/CLAUDE.md` | Claude Code natively reads `~/.claude/CLAUDE.md` at session start |
 | **Codex** | `~/.overdeck/context/codex-global.md` | At spawn, Overdeck seeds the agent's `$CODEX_HOME/AGENTS.md` from this file — Codex natively reads `AGENTS.md`, never the render artifact itself |
 | **Pi** | `~/.overdeck/context/pi-global.md` | Injected into the session's system-prompt context at spawn. Pi does natively support a global `~/.pi/agent/AGENTS.md`; Overdeck deliberately does not write there — the global layer is delivered per-spawn instead, leaving `~/.pi/agent/` entirely yours |
+| **Prime Agent** | `~/.overdeck/context/prime-agent-global.md` | Injected into the root RPC session's supplemental system prompt at spawn; Overdeck never writes `~/.prime/agent` |
 
 The `pi-global.md` / `codex-global.md` files are **Overdeck-owned render
 artifacts** — the harnesses do not know about them. They exist so the global
@@ -133,7 +134,8 @@ Write completion markers via `/pan-done` when ready for review.
 {{/harness:pi}}
 ```
 
-Blocks may be stacked to target several harnesses; `{{#harness:codex}}` is
+Use `{{#harness:prime-agent}}...{{/harness:prime-agent}}` for Prime-only
+content. Blocks may be stacked to target several harnesses; `{{#harness:codex}}` is
 also recognized.
 
 ## CLI

diff --git a/configuration/harnesses.mdx b/configuration/harnesses.mdx
@@ -12,7 +12,7 @@ that runs in tmux. The **model** is what the binary calls. They are picked
 independently per spawn.
 
 <Info>
-**Supported today:** `claude-code` (default), `ohmypi`, `codex`, `acp`, and `kimi-code`. For the wider
+**Supported today:** `claude-code` (default), `ohmypi`, `codex`, `acp`, `kimi-code`, and `prime-agent`. For the wider
 field of coding-agent harnesses Overdeck could adopt next — and how each one handles
 skills, MCP, and AGENTS.md — see the [Harness Landscape](/reference/harness-landscape).
 </Info>
@@ -24,6 +24,7 @@ skills, MCP, and AGENTS.md — see the [Harness Landscape](/reference/harness-la
 | `codex` | [OpenAI Codex CLI](https://github.com/openai/codex) (`codex`) | OpenAI-model work | Uses persistent `codex app-server` sessions by default; requires Codex CLI `0.144.0` or newer. |
 | `acp` | [Kimi Code CLI](https://www.kimi.com/code/docs/en/kimi-code-cli/guides/getting-started.html) (`kimi`) | Opt-in Kimi work | Uses the open Agent Client Protocol over a persistent authenticated host; Kimi is the first wired ACP agent. |
 | `kimi-code` | [Kimi Code CLI](https://www.kimi.com/code/docs/en/kimi-code-cli/guides/getting-started.html) (`kimi`) | Kimi models (built-in default) | Drives Kimi's own native terminal UI directly — no ACP host, no Anthropic-compatibility endpoint. Kimi models only. Set `providerHarnesses.kimi` to `acp` or `claude-code` to opt back out. |
+| `prime-agent` | Prime Agent (`prime-agent`) | Opt-in multi-provider RPC work and conversations | Overdeck owns one persistent root RPC process per session; supported versions are `>=0.1.0 <1.0.0`. |
 
 ## Composer commands across harnesses
 
@@ -60,6 +61,39 @@ Managed Codex sessions use an isolated `CODEX_HOME`. On initialization and resum
   receive neither override, so Claude Code's native context behavior remains in
   control.
 
+### Prime Agent RPC (alternative)
+
+Install a compatible Prime Agent release with the upstream installer, then verify it:
+
+```bash
+curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh
+prime-agent --version   # must be >=0.1.0 and <1.0.0
+prime-agent --help      # must list --mode rpc
+pan doctor
+```
+
+If the binary is outside `PATH`, set `primeAgent.binaryPath` in
+`~/.overdeck/config.yaml`. Configure the provider's API credentials through
+the provider's normal environment variables or Overdeck provider settings.
+Overdeck maps Anthropic models to `anthropic`, OpenAI models to `openai`, and
+Google models to `google`; an unknown provider/model pair fails instead of
+silently switching harnesses.
+
+Overdeck owns the persistent root RPC process, session directory, resume,
+delivery, abort, and termination. A detached session resumes only when its
+recorded model, provider, and workspace match. Otherwise Overdeck starts a
+fresh session or reports a recovery error. Prime schedules, heartbeats,
+observation, autonomous continuation, and cross-agent messaging are disabled
+because Cloister owns that lifecycle. RLM children remain internal tool work
+attributed to the root session and cannot claim Overdeck tasks.
+
+Prime's worker/kernel separation is an execution boundary, not a security
+sandbox. Use the same workspace and credential isolation you use for other
+coding harnesses. Overdeck writes managed context only under
+`~/.overdeck/context`; it never changes `~/.prime/agent`. To uninstall, stop
+Prime sessions, remove `primeAgent` from the Overdeck config, and remove the
+upstream `prime-agent` package with the package manager that installed it.
+
 ### oh-my-pi / ohmypi (alternative)
 
 - Adds RPC mode (`omp --mode rpc`) so Overdeck can write structured commands to

diff --git a/docs/HARNESSES.md b/docs/HARNESSES.md
@@ -6,6 +6,6 @@
 
 | Topic | Where it lives now |
 | --- | --- |
-| Operational guide — supported harnesses (`claude-code`, `pi`), install, picker locations, ToS rules, troubleshooting | [`configuration/harnesses.mdx`](../configuration/harnesses.mdx) |
+| Operational guide — supported harnesses, including managed Prime Agent RPC sessions, install, picker locations, security rules, and troubleshooting | [`configuration/harnesses.mdx`](../configuration/harnesses.mdx) |
 | Planning survey — 13 coding-agent harnesses, extensibility mechanisms, adoption implications | [`reference/harness-landscape.mdx`](../reference/harness-landscape.mdx) |
 | Proposal — loading curated skill bundles into a single conversation | [`reference/template-conversations.mdx`](../reference/template-conversations.mdx) |
diff --git a/docs/MODEL_ROUTING.md b/docs/MODEL_ROUTING.md
@@ -29,6 +29,15 @@ When Overdeck needs to make an LLM call, it resolves the model through this chai
 
 Overdeck supports seven LLM providers:
 
+### Prime Agent provider mapping
+
+The `prime-agent` harness receives an explicit provider and model on every
+launch. Anthropic routes to `anthropic`, OpenAI routes to `openai`, and Google
+routes to `google`. Unsupported mappings fail loudly; Overdeck does not pick a
+different model or harness. Provider authentication remains provider-owned.
+Using Claude Pro or Max subscription credentials through Prime Agent bills as
+extra API usage and does not consume the Claude plan quota.
+
 | Provider | Auth Method | Notes |
 |----------|-------------|-------|
 | **Anthropic** | Claude Code subscription login (default, always enabled) | Opus 4.6, Sonnet 4.6, Haiku 4.5 |

diff --git a/docs/TELEMETRY.md b/docs/TELEMETRY.md
@@ -79,7 +79,7 @@ the public contract and lists every allowed value.
 | `duration_bucket` | `under_100ms`, `100ms-999ms`, `1s-9s`, `10s+` |
 | `forge` | `github`, `gitlab` |
 | `fork_kind` | `summary`, `handoff`, `plain` |
-| `harness` | `claude-code`, `ohmypi`, `codex`, `acp`, `kimi-code` |
+| `harness` | `claude-code`, `ohmypi`, `codex`, `acp`, `kimi-code`, `prime-agent` |
 | `merge_kind` | `pipeline` |
 | `model_family` | `claude`, `gpt`, `gemini`, `kimi`, `minimax`, `glm`, `mimo`, `other` |
 | `pipeline_stage` | `work_done`, `review_passed`, `verification_passed`, `merged`, `closed_out` |

diff --git a/docs/evidence/pan-3668/prime-agent-runtime-picker.png b/docs/evidence/pan-3668/prime-agent-runtime-picker.png
diff --git a/docs/prime-agent-no-loss-audit.md b/docs/prime-agent-no-loss-audit.md
@@ -0,0 +1,20 @@
+# Prime Agent No-Loss Audit
+
+Every existing harness remains present while `prime-agent` gains an explicit home. The mechanical test reads each canonical surface below.
+
+| # | Surface | Canonical location | Prime disposition | Existing harnesses preserved | Verified |
+|---|---|---|---|---|---|
+| S1 | Harness union and known set | `packages/contracts/src/types.ts` | explicit literal | yes | [x] |
+| S2 | Behavior map | `packages/contracts/src/harness-behavior.ts` | `PRIME_AGENT_BEHAVIOR` | yes | [x] |
+| S3 | Config harness schema | `src/lib/config-yaml/schema.ts` | accepted role/provider harness | yes | [x] |
+| S4 | CLI help and composer manifest | `src/cli/index.ts`, `packages/contracts/src/composer-commands.generated.ts` | listed | yes | [x] |
+| S5 | Provider resolution | `src/lib/prime-agent/provider-map.ts` | strict table, no fallback | unchanged | [x] |
+| S6 | Runtime registry | `src/lib/runtimes/index.ts` | registered adapter | yes | [x] |
+| S7 | Artifact, telemetry, flywheel, context preview | `packages/contracts/src/{artifacts,telemetry,flywheel,context-layers}.ts` | explicit literal | yes | [x] |
+| S8 | Transcript registry | `src/lib/conversations/transcript-adapter.ts` | Prime adapter | yes | [x] |
+| S9 | Cost source | `src/lib/overdeck/cost.ts` | `prime_agent` source | yes | [x] |
+| S10 | Context renderer | `src/lib/context-layers/harness.ts` | Prime marker and artifact | yes | [x] |
+| S11 | Work/conversation lifecycle | `src/lib/agents/runtime-command.ts`, `src/lib/overdeck/conversation-runtime.ts` | persistent RPC | yes | [x] |
+| S12 | Delivery and recovery | `src/lib/prime-agent/{session-controller,session-resume}.ts` | correlated RPC and strict resume | unchanged | [x] |
+| S13 | Dashboard API and frontend | `src/dashboard/` | identity, activity, usage, picker, badges, filters | yes | [x] |
+| S14 | Operator documentation | `configuration/harnesses.mdx`, `configuration/context-layers.mdx` | setup and boundary documented | yes | [x] |
diff --git a/docs/prime-agent-verification.md b/docs/prime-agent-verification.md
@@ -0,0 +1,35 @@
+# Prime Agent verification
+
+Verified on 2026-08-12 for PAN-3668 with Prime Agent 0.7.2 and Node.js 22.22.0.
+
+## Automated gates
+
+| Gate | Result |
+| --- | --- |
+| `npm run build` | Passed, including contracts, CLI, dashboard server, dashboard frontend, and dist-externals checks. |
+| `npm run typecheck` | Passed. The dashboard ratchet fell from 30 known errors to 28 after the Prime runtime types were propagated. |
+| `npm run lint` | Passed through the focused rerun after removing the new Prime transcript-adapter cycle. |
+| `npm test` | 13,664 tests passed and 51 skipped. Two unrelated 5-second setup hooks timed out under full-suite load; both files passed immediately in a focused rerun (5/5 tests). |
+| Prime focused and no-loss tests | Passed, including the live smoke, runtime registry, parser, policy, picker, transcript, artifact, telemetry, and canonical-harness matrices. |
+
+The opt-in live test is `tests/integration/prime-agent-smoke.test.ts`. Run it with `OVERDECK_PRIME_AGENT_LIVE=1 npx vitest run tests/integration/prime-agent-smoke.test.ts`. Provider and model overrides are available through `OVERDECK_PRIME_AGENT_PROVIDER` and `OVERDECK_PRIME_AGENT_MODEL`.
+
+## Live Prime Agent smoke
+
+The smoke used a disposable directory and the installed `prime-agent` binary. It verified:
+
+1. RPC launch with an explicit provider, model, working directory, and session directory.
+2. Correlated `get_state`, `prompt`, `steer`, `get_session_stats`, `abort`, `get_messages`, and resumed-session requests.
+3. A real Bash tool execution that wrote the expected marker inside the disposable directory.
+4. Non-zero token accounting, non-negative cost accounting, graceful stop, and resume from the saved session file.
+5. The freshly built doctor diagnostic: `Prime Agent: v0.7.2 (RPC mode available)`.
+
+## Dashboard boot and UI
+
+The production dashboard bundle booted under Node.js 22 on isolated port 33668 with an isolated `OVERDECK_HOME`, `OVERDECK_DISABLE_DEACON=1`, and `OVERDECK_NO_RESUME=1`. `/api/health` returned `status: ok`, `mode: peer`, and the feature branch build metadata. The log order confirmed the HTTP listener before boot-cache completion, and Cloister stayed disabled.
+
+Playwright opened the isolated dashboard, enabled the throwaway harness-permutation setting, opened the runtime picker, and confirmed that `Prime Agent` was visible with the description `Managed persistent Prime RPC session`. The inspected screenshot is [the Prime Agent runtime picker](evidence/pan-3668/prime-agent-runtime-picker.png).
+
+## Release decision
+
+The Prime Agent path is verified for managed launch, RPC delivery, tool use, transcript persistence, cost reporting, stop/resume, diagnostics, and operator-facing selection. The only full-suite exceptions were two unrelated setup-hook timeouts that passed outside the saturated run.
diff --git a/packages/contracts/src/artifacts.ts b/packages/contracts/src/artifacts.ts
@@ -12,7 +12,7 @@ export type ArtifactHash = typeof ArtifactHash.Type
 export const ArtifactAgentRole = Schema.Literals(["plan", "work", "review", "test", "ship", "flywheel", "user"])
 export type ArtifactAgentRole = typeof ArtifactAgentRole.Type
 
-export const ArtifactAgentHarness = Schema.Literals(["claude-code", "pi", "ohmypi", "codex", "acp", "kimi-code", "user"])
+export const ArtifactAgentHarness = Schema.Literals(["claude-code", "pi", "ohmypi", "codex", "acp", "kimi-code", "prime-agent", "user"])
 export type ArtifactAgentHarness = typeof ArtifactAgentHarness.Type
 
 export const ArtifactLifecycleState = Schema.Literals([