@hegemonart/get-design-done 1.33.0 → 1.33.6

package/reference/gdd-threat-model.md

@@ -0,0 +1,399 @@
+# GDD Runtime Threat Model (STRIDE)
+
+> Phase 33.5 · SEC-01 · STRIDE pass over GDD's **own** runtime attack surface.
+> Generated against branch `phase/33-5-runtime-security`, HEAD `5374bed`.
+
+## Scope
+
+This document models the security posture of **GDD's own runtime** — the
+multi-MCP-server, peer-CLI-spawning, WebSocket-emitting SDK that grew across
+Phases 20–27 without a formalized security model. It does **NOT** model the
+user code that GDD audits; the safety floor for *audited user code* is Phase
+14.5's concern. This is the equivalent threat model for GDD's *own* moving
+parts: the hooks that run on every session start, the two MCP servers that
+read and mutate `STATE.md`, the broker that spawns peer CLIs, the WebSocket
+transport that streams the event bus, and the issue-reporter that reaches the
+network through `gh`.
+
+**STRIDE** is the Microsoft threat taxonomy used throughout: **S**poofing
+(pretending to be someone/something you are not), **T**ampering (unauthorized
+modification of data or code), **R**epudiation (denying an action with no
+audit trail), **I**nformation disclosure (leaking data to the wrong party),
+**D**enial of service (exhausting a resource so legitimate use fails), and
+**E**levation of privilege (gaining capabilities you were not granted).
+
+Each of the five in-scope components below gets a fixed five-part treatment:
+**Assets** (what an attacker wants), **Entry points** (the untrusted-input
+boundary), **STRIDE threats** (which categories apply), **Current mitigations**
+(citing **real shipped code** — file + line + behavior), and **Residual risks**
+(threats current code does **not** fully cover, each routed to the Phase 33.5
+plan that closes it). Out of scope per CONTEXT: rewriting the issue-reporter
+network model — it is **documented** here as already-mitigated, not
+re-engineered.
+
+## Trust boundaries
+
+The runtime crosses four trust boundaries. On the untrusted side of each sits
+input that an attacker (or a compromised peer / config author / network host)
+controls; the table names what crosses the line.
+
+| Boundary | Untrusted side | What crosses |
+| --- | --- | --- |
+| WS event-stream server `←` client | A WebSocket client on the network (LAN/internet if bound wide) | The HTTP `Upgrade` request + `Authorization: Bearer` header |
+| gdd-state MCP `←` environment / config / tool input | Whoever sets `GDD_STATE_PATH` or supplies a tool-call payload, or authors `.design/config.json` | The `GDD_STATE_PATH` env value + the JSON tool-input payloads |
+| Peer-CLI broker `↔` spawned child | A spawned peer CLI (Codex / Gemini / Cursor / Copilot / Qwen) and its stdout stream | The child's stdout JSON frames + the parent env handed to the child |
+| Outbound call sites `↔` external host | The remote HTTP host / GitHub / Figma the call reaches | The outbound request payload + whatever the remote returns |
+| OpenRouter catalog fetch `→` openrouter.ai | The OpenRouter `/models` API host (and any MITM on the path) | The `Authorization: Bearer <OPENROUTER_API_KEY>` request header + the untrusted `/models` JSON the host returns |
+
+The event payloads that traverse the bus (and therefore the WS transport and
+any persisted JSONL) are scrubbed at serialize time — see Component 4's
+`redact.cjs` mitigation, which is the cross-cutting information-disclosure
+control for the whole bus.
+
+---
+
+## Component 1 — Hooks (SessionStart update-check + budget/context-monitor)
+
+The hooks run automatically: `SessionStart` fires the update-check on every
+session, and the budget / context-monitor hook runs on tool-use to enforce
+spend and context ceilings. They execute with the user's full shell privileges
+inside the user's repo, with no sandbox.
+
+- **Assets:** The user's shell + filesystem (the hook runs as the user); the
+  integrity of the budget/context accounting the monitor maintains; the
+  network reachability of the update-check's outbound call.
+- **Entry points:** The update-check's outbound HTTP fetch and whatever it
+  parses from the response (a version string / changelog); the hook's read of
+  `.design/config.json` (a malicious or malformed config is untrusted input);
+  the tool-use payload the budget monitor inspects.
+- **STRIDE threats:**
+  - **Spoofing:** A spoofed update endpoint (DNS/MITM) could feed a forged
+    "latest version" response to the update-check.
+  - **Tampering:** A malformed `.design/config.json` could try to corrupt the
+    budget/context accounting or flip the monitor's thresholds.
+  - **Repudiation:** Hook actions are largely silent — limited audit trail of
+    what a SessionStart hook did or why a budget veto fired.
+  - **Information disclosure:** The update-check's User-Agent / outbound
+    request reveals that GDD is in use; a verbose hook could echo env into logs.
+  - **Denial of service:** A hung or slow update endpoint could stall session
+    start if the fetch were unbounded.
+  - **Elevation of privilege:** The hook already runs at full user privilege —
+    the residual concern is a config-driven path or command injection lifting
+    *attacker* input to that privilege level.
+- **Current mitigations:** The update-check is **advisory** — it informs of a
+  newer version and never auto-installs or executes downloaded code, so a
+  spoofed version string cannot achieve code execution. The budget /
+  context-monitor reads config defensively (missing file / malformed JSON /
+  missing key are tolerated, mirroring the issue-reporter kill-switch's
+  config-tolerance contract in `scripts/lib/issue-reporter/kill-switch.cjs`).
+  Hooks emit through the event bus, which is redacted by `redact.cjs` at
+  serialize time (see Component 4), so secrets in hook telemetry are scrubbed.
+- **Residual risks:** The update-check's outbound egress is one of the
+  cross-cutting call sites that currently has **no machine-readable allowlist
+  and no CI gate** asserting it is the only network touch a hook makes →
+  audited + allowlisted in **33.5-02** and gated in **33.5-04**. (The shell
+  hook `hooks/update-check.sh` is `.sh`, outside the `.js`-family static
+  scanner's scope, so it is **documented** in the **33.5-02** audit report
+  rather than hard-gated.)
+
+---
+
+## Component 2 — MCP servers (gdd-state: 11 mutating tools / gdd-mcp: read)
+
+Two MCP servers expose GDD state to an MCP client: **gdd-state**
+(`sdk/mcp/gdd-state/`) with **11 mutating tools** — `add_blocker`,
+`add_decision`, `add_must_have`, `checkpoint`, `frontmatter_update`, `get`,
+`probe_connections`, `resolve_blocker`, `set_status`, `transition_stage`,
+`update_progress` — and **gdd-mcp** (`sdk/mcp/gdd-mcp/`) with read tools. The
+mutating server is the higher-value target because it writes `STATE.md`.
+
+- **Assets:** The integrity of `STATE.md` (the project's source of truth for
+  position, decisions, blockers, stage); the event stream the mutations emit;
+  the filesystem region the server is allowed to write.
+- **Entry points:** The `GDD_STATE_PATH` environment variable (which redirects
+  *where the server reads/writes*); the JSON tool-input payloads for all 11
+  mutating tools; the `.design/STATE.md` file content the server parses.
+- **STRIDE threats:**
+  - **Spoofing:** A tool caller could impersonate a legitimate pipeline stage
+    and drive `transition_stage` / `set_status` without authorization.
+  - **Tampering:** Crafted tool inputs could write hostile content into
+    `STATE.md`, or `GDD_STATE_PATH` could redirect writes onto an unintended
+    file (path traversal).
+  - **Repudiation:** Without a complete mutation audit trail, a hostile or
+    buggy mutation is hard to attribute — partly addressed by the event
+    emissions below.
+  - **Information disclosure:** A `get` against a traversed path could read a
+    file outside the intended `.design/` boundary.
+  - **Denial of service:** A JSON-bomb (deeply nested object / multi-megabyte
+    string field) in a tool payload could exhaust memory/CPU during parse.
+  - **Elevation of privilege:** Path traversal via `GDD_STATE_PATH` plus an
+    absent boundary check effectively elevates a tool caller's reach to any
+    file the process can write.
+- **Current mitigations:** Every mutation emits a `state.mutation` /
+  `state.transition` event through `emitStateMutation()` / `emitStateTransition()`
+  (`sdk/mcp/gdd-state/tools/shared.ts` lines 91–140), giving a partial audit
+  trail (anti-repudiation). Handlers **never throw to the harness** — every
+  error funnels through `errorResponse()` → `toToolError()` into a structured
+  `{success:false,error}` (shared.ts lines 28–31, 148–151), so a malformed
+  input degrades to a clean error instead of a crash. Each of the 11 tools
+  already ships a JSON input schema under `sdk/mcp/gdd-state/schemas/`. State
+  events are redacted by `redact.cjs` at serialize time (Component 4).
+- **Residual risks:** `resolveStatePath()` (`sdk/mcp/gdd-state/tools/shared.ts`
+  lines 60–64) honors `GDD_STATE_PATH` with **no path-traversal guard** — it
+  returns the override verbatim, so `..` escape / absolute-outside / symlink
+  escape are unchecked. The tool schemas exist but carry **no payload-size cap**
+  (no JSON-bomb guard) and are not uniformly tightened
+  (`additionalProperties:false` + `maxLength`). Path traversal + JSON-bomb +
+  un-tightened schemas are all closed by **33.5-03** (path-traversal guard +
+  payload cap + all 11 schemas tightened).
+
+---
+
+## Component 3 — Peer-CLI broker (acp-client + asp-client child spawn)
+
+The broker spawns peer CLIs over stdio: `scripts/lib/peer-cli/acp-client.cjs`
+(ACP-protocol peers) and `scripts/lib/peer-cli/asp-client.cjs` (Codex
+app-server protocol). Both fork a local child process and exchange
+line-delimited JSON over its stdio. The child is **untrusted** — it is a
+third-party CLI whose stdout the broker parses.
+
+- **Assets:** GDD's process environment — specifically `ANTHROPIC_API_KEY`,
+  `GH_TOKEN`, and any `GDD_*` / provider secret in `process.env`; the broker's
+  memory/availability; the integrity of the JSON protocol exchange.
+- **Entry points:** The child's **stdout** (untrusted JSON frames the broker
+  must parse); the **environment handed to the child** at spawn time; the
+  `opts.command` / `opts.args` the broker is asked to launch.
+- **STRIDE threats:**
+  - **Spoofing:** A misbehaving peer could emit forged protocol replies /
+    request IDs to confuse the correlation map.
+  - **Tampering:** A peer could stream malformed frames attempting to corrupt
+    the broker's line-buffer / pending-request state.
+  - **Repudiation:** Limited record of exactly what env a given child was
+    handed at spawn.
+  - **Information disclosure:** **The headline risk** — the child inherits
+    GDD's full environment, so a hostile or compromised peer reads
+    `ANTHROPIC_API_KEY` / `GH_TOKEN` straight out of `process.env`.
+  - **Denial of service:** A peer that never emits a newline could force the
+    broker to buffer unbounded stdout until memory exhaustion.
+  - **Elevation of privilege:** Inherited secrets let a peer act *as GDD*
+    against GDD's providers — using GDD's keys for the peer's own ends.
+- **Current mitigations:** `acp-client.cjs` caps an un-terminated stdout line
+  at **`MAX_LINE_BYTES = 16 * 1024 * 1024`** (16 MiB; defined line 62, enforced
+  lines 166–176 — a peer that emits 16 MiB without a newline gets its active
+  prompt rejected as a protocol violation). This is a real **DoS guard** on the
+  untrusted stdout channel. The broker uses plain `spawn` with **no shell**
+  (acp-client.cjs lines 106–113, `windowsHide: true`), avoiding shell-injection
+  on the command path. Per-request correlation via a pending-id map bounds the
+  protocol state machine.
+- **Residual risks:** Both clients default the child's environment to the
+  **full `process.env`** when `opts.env` is absent — `acp-client.cjs` line 102
+  (`const env = opts.env && typeof opts.env === 'object' ? opts.env :
+  process.env;`) and `asp-client.cjs` line 122 (when `opts.env` is absent no
+  `spawnOptions.env` is set, so the child inherits the parent's `process.env` by
+  Node default). This leaks GDD's `ANTHROPIC_API_KEY` / `GH_TOKEN` / `GDD_*` to
+  every spawned peer. Closed by **33.5-04** (allowlist-forward, default-deny env
+  sandbox via a shared `sanitize-env` helper applied to both clients; secrets
+  are never forwarded unless explicitly allowlisted in `.design/config.json`).
+
+---
+
+## Component 4 — WebSocket event-stream transport (scripts/lib/transports/ws.cjs)
+
+`scripts/lib/transports/ws.cjs` exposes the event-stream bus over WebSocket:
+one JSON event per text frame, with optional replay of a tail file to each new
+connection. It is an **optional dependency** (`ws`) — absent installs render an
+install hint instead of starting. When running, it is a network listener.
+
+- **Assets:** The **event stream itself** (every `state.mutation` /
+  `state.transition` / pipeline event, which can carry payload detail); the
+  listening socket; the Bearer token that authorizes a connection.
+- **Entry points:** The HTTP **`Upgrade` request** from any client that can
+  reach the bound socket, and specifically its `Authorization: Bearer <token>`
+  header; the `tailFrom` replay file path.
+- **STRIDE threats:**
+  - **Spoofing:** A client without the token attempting to subscribe to the
+    live event stream.
+  - **Tampering:** N/A for inbound (the transport is push-only to clients) —
+    the concern is read access, not write.
+  - **Repudiation:** No per-connection identity beyond the shared token, so
+    individual subscribers are not distinguishable in an audit.
+  - **Information disclosure:** **The headline risk** — an unauthorized
+    subscriber would receive the entire live event stream, including any
+    sensitive payload detail, if it could reach the socket and pass auth.
+  - **Denial of service:** Many connections / a slow consumer could pressure
+    the server (mitigated in part by fire-and-forget, no-queue backpressure).
+  - **Elevation of privilege:** A network-reachable listener turns a
+    local-only observability feature into a remotely-reachable data source.
+- **Current mitigations:** **Bearer-token auth is enforced on every upgrade**:
+  `ws.cjs` lines 110–116 reject any upgrade whose header is missing or where
+  the supplied token does not match the expected `Bearer` value, returning an `HTTP/1.1 401 Unauthorized` and a
+  socket destroy. The token **must be ≥8 chars** — `startServer` throws a
+  `TypeError` if `opts.token.length < 8` (line 74), preventing trivially weak
+  tokens. Backpressure is **fire-and-forget with no queue** (lines 91–108):
+  events for a non-OPEN socket are dropped, bounding memory under a slow
+  consumer. Cross-cutting for the whole bus: **`redact.cjs`** deep-walks every
+  event payload at serialize time (`scripts/lib/redact.cjs` — `redact()` lines
+  95–116, `redactString()` lines 75–83) and scrubs **8 secret patterns** (pem,
+  jwt, anthropic `sk-ant-`, stripe `sk_live_`, slack `xox[baprs]`, github_pat
+  `ghp_`, aws `AKIA`, generic `sk-`), so secrets in event payloads are masked
+  before they ever reach a WS subscriber or hit disk. This `redact.cjs` scrub is
+  the runtime's primary information-disclosure control across **all** components
+  that emit events.
+- **Residual risks:**
+  - The server binds to **all interfaces (`0.0.0.0`)** by default —
+    `httpServer.listen(opts.port, ...)` (line 145) passes **no host argument**,
+    so on a multi-homed / LAN host the token-protected stream is reachable
+    off-box. The token compare uses `!==` (line 112), which is
+    **timing-unsafe**. Both closed by **33.5-03** (default bind `127.0.0.1` +
+    opt-in remote via `event_stream.bind_host` / `GDD_WS_BIND_HOST` + a CI gate
+    that fails if the default config would bind `0.0.0.0`; upgrade the compare to
+    `crypto.timingSafeEqual`).
+  - `redact.cjs` is **missing three modern token formats**: Gemini / GCP
+    `AIza…`, GitHub fine-grained `github_pat_…`, and GitHub server / oauth /
+    user / refresh `gh[sour]_…`. A payload carrying one of these would leak
+    through the scrub onto the stream and disk. Closed by **33.5-05** (add the
+    three patterns + a synthetic-secret fuzz test asserting zero leak).
+
+---
+
+## Component 5 — Issue-reporter outbound (gh CLI only)
+
+`scripts/lib/issue-reporter/` is the only first-party feature that intentionally
+reaches the network. It assembles a bug report and submits it through the user's
+**`gh` CLI**. **This network model is already mitigated and is DOCUMENTED here,
+not re-engineered** (CONTEXT Out-of-scope: rewriting the issue-reporter network
+model).
+
+- **Assets:** The user's GitHub identity (via the local `gh` auth); the content
+  of the submitted report (which must not carry the user's secrets or
+  unintended PII); the integrity of the destination repo.
+- **Entry points:** The user-invoked report flow (the body / title assembled
+  from local state); the `.design/config.json` and the env that gate whether the
+  reporter runs at all.
+- **STRIDE threats:**
+  - **Spoofing:** A forged destination could try to receive reports — mitigated
+    by the frozen destination below.
+  - **Tampering:** Attempting to redirect submissions to an attacker repo by
+    injecting a destination override.
+  - **Repudiation:** Submissions flow through `gh` under the user's identity,
+    which is itself the attribution record.
+  - **Information disclosure:** **The headline risk** — a report could exfiltrate
+    secrets / PII embedded in local state if the payload were not scrubbed.
+  - **Denial of service:** Not a meaningful vector — submission is a
+    user-initiated, one-shot CLI call.
+  - **Elevation of privilege:** Using the user's `gh` credentials beyond the
+    single sanctioned submit.
+- **Current mitigations (ALREADY shipped — documented, no change here):**
+  - **Outbound is via the `gh` CLI ONLY.** `gh-submit.cjs` wraps
+    `gh issue create --repo <DESTINATION_REPO> --title … --body-file …` and is
+    explicit that "the user's gh CLI is the sole outbound primitive. No HTTP-S
+    URL literals, no global fetch primitive, no plugin-side credentials" (D-05).
+    There is no raw HTTP egress in this subtree.
+  - **Frozen destination.** `destination.cjs` is an `Object.freeze`-d module —
+    the single source of truth for the destination repo, with **no env-var
+    lookup, no config override, no flag override**. A static CI gate asserts it
+    is the only file under the report-issue tree that contains the destination
+    literal, so a redirect attempt fails the build.
+  - **Kill-switch (dual-surface).** `kill-switch.cjs` disables the reporter via
+    **either** the env var `GDD_DISABLE_ISSUE_REPORTER === '1'` **or** the config
+    `.design/config.json` `{ "issue_reporter": false }`; either surface alone is
+    sufficient, and config is read tolerantly (missing file / malformed JSON /
+    missing key are safe).
+  - Payloads pass through privacy-diff / consent-prompt machinery before
+    submission, and event telemetry is redacted by `redact.cjs` (Component 4).
+- **Residual risks:** The issue-reporter's **own** network model has no residual
+  this phase changes — it is intentionally documented as complete. The only
+  cross-cutting residual touching it is the **lack of a machine-readable
+  outbound allowlist + CI gate** that proves `gh-submit` is the sole egress in
+  this subtree at a tree-wide level: closed by **33.5-02** (the canonical
+  outbound-network allowlist data, which lists `scripts/lib/issue-reporter/**`
+  as an allowed egress glob) and **33.5-04** (the `scan:outbound` CI gate that
+  fails on any active-egress site not under an allowlisted glob).
+
+---
+
+## Component 6 — OpenRouter catalog fetcher (scripts/lib/openrouter/catalog-fetcher.cjs)
+
+> Added in Phase 33.6 (OR-01, CONTEXT D-06). This is the runtime's **first
+> plugin-side outbound REST client** — the issue-reporter (Component 5) reaches
+> the network only through the user's `gh` CLI, and the WS transport (Component
+> 4) is a *server*, not an outbound client. The catalog fetcher is the first
+> first-party code to open an outbound HTTP request to a third-party host
+> directly, which is why it lands only after the 33.5 audited baseline and the
+> `scan:outbound` gate (33.5-04) are in place.
+
+`scripts/lib/openrouter/catalog-fetcher.cjs` performs a read-only GET to the
+OpenRouter model catalog (`https://openrouter.ai/api/v1/models`) through an
+**injectable `fetchImpl`** (default global `fetch`), maps the response into the
+`.design/cache/openrouter-models.json` cache shape, and writes it atomically.
+The live fetch is opt-in — gated on `OPENROUTER_API_KEY` being present at
+runtime; absent it, the fetcher returns cached-if-any-else-null and tier
+resolution falls back to the native provider.
+
+- **Assets:** The **`OPENROUTER_API_KEY`** (a billable provider credential) and
+  the integrity of the cached catalog the tier-resolver later trusts.
+- **Entry points:** The **`/models` JSON the OpenRouter host returns** (untrusted
+  remote input the fetcher must parse), and the `OPENROUTER_BASE_URL` env (an
+  operator-supplied endpoint override).
+- **STRIDE threats:**
+  - **Spoofing:** A spoofed `/models` endpoint (DNS/MITM, or a hostile
+    `OPENROUTER_BASE_URL`) could feed a forged catalog.
+  - **Tampering:** A malformed/oversized `/models` body could try to corrupt the
+    cache the resolver reads, or smuggle unexpected fields downstream.
+  - **Information disclosure:** **The headline risk** — leaking the
+    `OPENROUTER_API_KEY` by persisting it to the cache, logging it, or sending it
+    to an unintended host.
+  - **Denial of service:** A hung or slow host could stall the fetch; a giant
+    catalog could pressure memory.
+  - **Elevation of privilege:** A forged catalog could steer tier resolution to
+    an attacker-chosen model id.
+- **Current mitigations:** The key is read from **`OPENROUTER_API_KEY` env only**,
+  sent **solely** as an `Authorization: Bearer` request header, and is **never
+  persisted to the cache nor written to any log seam** — the cache shape carries
+  only `id`/`name`/`context_length`/`pricing`, and the mapper keeps **only** those
+  fields, dropping everything else (the `/models` body is **mapped, never
+  eval'd**). The cache write is **atomic** (per-pid temp + rename) into the
+  **gitignored** `.design/cache/`, so a partial/corrupt fetch can't leave a
+  half-written catalog and the cache never enters git history. The fetcher
+  **never throws** (D-08): no key / fetch failure / parse failure all degrade to
+  cached-if-any-else-null, bounding the DoS surface, and retries are **bounded**
+  (max 3 attempts) on a jittered-backoff curve with `rate-guard` awareness.
+  Egress is **allowlisted** via `scripts/lib/openrouter/**` in
+  `scripts/security/outbound-allowlist.json` — the only sanctioned outbound site
+  in that subtree — so the 33.5 `scan:outbound` gate proves no un-approved egress
+  crept in. The **injectable `fetchImpl`** keeps the default `npm test` suite
+  hermetic (D-07) — no live network — and there is **no new HTTP dependency**
+  (global `fetch` + `sdk/primitives` only — D-10), avoiding both a new supply-chain
+  surface and the gate's `axios`/`node-fetch`/`undici` package patterns.
+- **Residual risks:** None this phase leaves open. The catalog is advisory data
+  consumed by the tier-resolver heuristic (33.6-02), which already clamps to
+  GDD's `opus`/`sonnet`/`haiku` vocabulary and supports user overrides, so a
+  forged catalog cannot escalate beyond model-id selection within that bounded
+  set; a future hardening could pin the OpenRouter TLS cert or sign the cache,
+  but neither is required for the current trust model.
+
+---
+
+## Residual-risk → closing-plan map
+
+Every residual risk identified above is routed to the Phase 33.5 plan (or
+policy doc) that closes it. No residual is left unmapped. This table is the
+spine the phase closeout (33.5-06) uses to prove completeness.
+
+| Residual risk | Component | Closing plan |
+| --- | --- | --- |
+| WS binds `0.0.0.0` by default (`listen` line 145, no host) + timing-unsafe `!==` token compare (line 112) | WebSocket transport | **33.5-03** |
+| `GDD_STATE_PATH` path traversal (no guard, shared.ts 60–64) + no payload-size cap + un-tightened tool schemas | gdd-state MCP | **33.5-03** |
+| Full `process.env` (incl. `ANTHROPIC_API_KEY` / `GH_TOKEN`) leaks to spawned peers (acp 102 / asp 122) | Peer-CLI broker | **33.5-04** |
+| Outbound egress sites have no machine-readable allowlist + no CI gate | cross-cutting (hooks update-check, figma-extract, issue-reporter, e2e) | **33.5-02** (allowlist) + **33.5-04** (scan gate) |
+| Secret-scan misses Gemini `AIza…` / GitHub fine-grained `github_pat_…` / GitHub server `gh[sour]_…` tokens | redact.cjs | **33.5-05** |
+| No published vulnerability-disclosure policy | project | **33.5-06** (SECURITY.md) |
+
+### Already-mitigated (documented, NOT re-engineered)
+
+| Already-mitigated surface | Component | Evidence |
+| --- | --- | --- |
+| Outbound via `gh` CLI only; frozen destination; dual-surface kill-switch | Issue-reporter | `gh-submit.cjs` (`gh issue create` only), `destination.cjs` (`Object.freeze`, no override), `kill-switch.cjs` (env + `.design/config.json`) |
+| Bearer-token auth on every WS upgrade + ≥8-char token rule | WebSocket transport | `ws.cjs` 110–116 (401 on mismatch) + line 74 (`length < 8` → `TypeError`) |
+| 16 MiB un-newlined-stdout DoS cap on untrusted peer output | Peer-CLI broker | `acp-client.cjs` `MAX_LINE_BYTES` line 62, enforced 166–176 |
+| Deep-walk secret scrub of every event payload at serialize time | cross-cutting (event bus) | `redact.cjs` `redact()` 95–116 / `redactString()` 75–83, 8 patterns |

package/reference/openrouter-tier-mapping.md

@@ -0,0 +1,98 @@
+# OpenRouter Tier-Mapping Heuristic
+
+How `scripts/lib/tier-resolver-openrouter.cjs` maps GDD's tier vocabulary onto a
+dynamic OpenRouter catalog model id. This document is the human-readable companion
+to that adapter; the adapter's `resolve(tier, opts)` is the canonical, executable
+source of the mapping. Phase 33.6, decision D-03 (heuristic + override), D-04
+(tier vocabulary), D-08 (graceful-null → native fallback).
+
+## What it maps
+
+The plugin speaks one tier vocabulary everywhere a model tier is named in
+frontmatter or config: `opus`, `sonnet`, `haiku` — the same `VALID_TIERS` the
+Phase-26 `tier-resolver.cjs` enforces. OpenRouter, by contrast, exposes a flat
+catalog of provider-prefixed model ids (`anthropic/claude-opus-4-7`,
+`meta-llama/llama-3.1-8b-instruct`, `qwen/qwen-2.5-72b-instruct`, …). The adapter
+bridges the two by assigning each GDD tier to one internal capability bucket and
+then picking the catalog id that best fits that bucket.
+
+The ROADMAP's SC#4 names the buckets `high` / `medium` / `low`; those are the
+heuristic's INTERNAL labels. They map one-to-one to the public tiers (D-04):
+
+- `opus`   ← HIGH   bucket
+- `sonnet` ← MEDIUM bucket
+- `haiku`  ← LOW    bucket
+
+The adapter's public `resolve(tier)` always speaks `opus` / `sonnet` / `haiku`;
+`high` / `medium` / `low` never leak across the API boundary.
+
+## The buckets
+
+- **opus (HIGH) = top-tier closed.** The most capable closed-vendor model in the
+  catalog — the priciest premium id from a closed namespace. This is the
+  "spare-no-expense, hardest reasoning" slot.
+- **sonnet (MEDIUM) = mid / top-open.** A capable model that sits below the opus
+  pick — typically the mid-priced closed model, or the strongest open model when
+  no second closed tier is present. The everyday workhorse slot.
+- **haiku (LOW) = cheap open.** The cheapest capable OPEN model — the
+  fast/inexpensive slot for high-volume, low-stakes calls.
+
+## The signals
+
+The heuristic is computed from fields already present on each catalog model, so it
+stays deterministic for a fixed catalog (no clock, no randomness — important so the
+33.6-04 golden baseline is stable):
+
+- **Namespace (closed vs open).** The id prefix before the `/` names the vendor.
+  `anthropic`, `openai`, `google` are treated as CLOSED (premium, frontier).
+  `meta-llama`, `qwen`, `mistralai`, `deepseek` are treated as OPEN (commodity,
+  cheap). The closed/open split is the primary axis: opus and sonnet prefer closed,
+  haiku requires open.
+- **Pricing.** Each model carries `pricing.prompt` / `pricing.completion` as string
+  decimals (USD per token). Parsed to Number, the completion price is the tie-break:
+  highest completion price wins the opus slot; lowest completion price wins the
+  haiku slot. Models with unparseable or missing pricing sort last.
+- **Context length.** `context_length` is a secondary capability signal used only to
+  break a pricing tie (longer context is treated as more capable).
+
+For the canonical fixture catalog (closed `anthropic/claude-opus-4-7` +
+`anthropic/claude-sonnet-4-7`, open `meta-llama/llama-3.1-70b-instruct`,
+`meta-llama/llama-3.1-8b-instruct`, `qwen/qwen-2.5-72b-instruct`) the heuristic
+resolves opus → `anthropic/claude-opus-4-7` (top closed, highest completion price),
+sonnet → `anthropic/claude-sonnet-4-7` (mid closed), and haiku →
+`meta-llama/llama-3.1-8b-instruct` (cheapest open).
+
+## The override escape hatch
+
+The heuristic is a sensible default, not a straitjacket. A user can pin any tier to
+an exact catalog id via `.design/config.json`:
+
+```
+{
+  "openrouter_tier_overrides": {
+    "opus": "anthropic/claude-opus-4-7",
+    "haiku": "meta-llama/llama-3.1-8b-instruct"
+  }
+}
+```
+
+An override **wins** over the heuristic: when `openrouter_tier_overrides[tier]` is a
+non-empty string, the adapter returns it verbatim — even if that id is not present
+in the live catalog (the user's explicit choice is honored over catalog membership).
+Tests inject the same map via `opts.overrides` instead of reading the live config
+file, so the override path is exercised hermetically. The config read is best-effort:
+a missing file, a missing key, or corrupt JSON degrades to an empty override map
+rather than throwing.
+
+## The graceful-null contract
+
+OpenRouter is opt-in ALONGSIDE native provider auth — never OpenRouter-only (D-08).
+When no catalog is available (no cache, an empty `models[]`, or a `readCatalog` that
+returns null) AND no override applies to the requested tier, `resolve` returns
+`null`. A `null` is not an error: it is the signal that the caller (the router /
+budget-enforcer, wired in 33.6-03) should fall back to the native provider via the
+existing `scripts/lib/tier-resolver.cjs` fallback chain. The adapter NEVER throws —
+an unknown tier, a missing config, a corrupt cache, or garbage options all degrade to
+`null` (or to an override when one applies). This keeps OpenRouter a strictly
+additive capability: turning it off, or having it fail to fetch, can never break a
+resolution that would have succeeded natively.

package/reference/prices.openrouter.md

@@ -0,0 +1,26 @@
+# OpenRouter — Catalog-Derived Price Snapshot
+
+**Phase 33.6 (v1.33.6).** This file is a **catalog-derived snapshot** of OpenRouter per-model prices — it is **generated from** `.design/cache/openrouter-models.json` (the dynamic catalog fetched by `scripts/lib/openrouter/catalog-fetcher.cjs`), **not** a hand-maintained authority. The **live source of truth is the dynamic catalog**; this table is a derived, illustrative view that can go stale between catalog fetches.
+
+Unlike the per-runtime tables under `reference/prices/` (Phase 26 D-08, hand-curated authority with provenance), OpenRouter's prices live in the upstream `/models` response and are refreshed on the 24h TTL. To inspect the current resolved prices, run `/gdd:openrouter-status` or read the cache directly. For the tier→model resolution heuristic see `reference/openrouter-tier-mapping.md`.
+
+OpenRouter quotes prices **per token** (USD), for `prompt` (input) and `completion` (output) separately.
+
+## Representative sample (per token, USD)
+
+Derived from the fixture catalog at `test/fixtures/baselines/phase-33-6/openrouter-catalog.json` (a snapshot mirror of the cache shape). Actual live prices come from the catalog at fetch time.
+
+| model id | prompt $/tok | completion $/tok |
+|----------|--------------|------------------|
+| `anthropic/claude-opus-4-7` | 0.000015 | 0.000075 |
+| `anthropic/claude-sonnet-4-7` | 0.000003 | 0.000015 |
+| `meta-llama/llama-3.1-70b-instruct` | 0.00000052 | 0.00000075 |
+| `meta-llama/llama-3.1-8b-instruct` | 0.00000002 | 0.00000005 |
+| `qwen/qwen-2.5-72b-instruct` | 0.00000038 | 0.0000004 |
+
+## Notes
+
+- **Derived view, not authority.** Do not hand-edit prices here to "fix" cost math — fix the catalog fetch instead. This file documents the *shape* and *source* of OpenRouter pricing for the registry round-trip and for human reference.
+- **Per-token vs per-1M.** The native runtime tables (`reference/prices/<runtime>.md`) quote `input_per_1m` / `output_per_1m`; OpenRouter's catalog quotes per-token. Multiply by 1,000,000 to compare (e.g. `anthropic/claude-opus-4-7` ≈ $15 input / $75 output per 1M tokens).
+- **Cost telemetry.** When a model is resolved via the OpenRouter adapter, the cost row tags `provider: openrouter` (Phase 33.6-03, SC#6) — see `scripts/lib/budget-enforcer.cjs#buildCostEventPayload`.
+- **Drift.** The authority-watcher diffs the catalog weekly and surfaces `deprecated`/`withdrawn` models matching a configured `openrouter_tier_overrides` pin (SC#8) — see `scripts/lib/authority-watcher/index.cjs#diffOpenRouterCatalog`.

package/reference/registry.json

@@ -860,6 +860,34 @@
       "type": "meta-rules",
       "phase": 30,
       "description": "Phase 30 triage gate catalogue — locally-fixable failure modes (id/pattern/diagnosis/remedy/severity, with optional propose_report whitelist flag per D-11) consulted by scripts/lib/issue-reporter/triage-matcher.cjs before the report-issue consent prompt (D-07/D-11)."
+    },
+    {
+      "name": "gdd-threat-model",
+      "path": "reference/gdd-threat-model.md",
+      "type": "heuristic",
+      "phase": 33.5,
+      "description": "Phase 33.5 STRIDE threat model of GDD's own runtime attack surface — hooks, the gdd-state + gdd-mcp MCP servers, the peer-CLI broker, the WebSocket event-stream transport, and issue-reporter outbound; maps each residual risk to the 33.5 plan that closes it."
+    },
+    {
+      "name": "gdd-runtime-audit",
+      "path": "reference/gdd-runtime-audit.md",
+      "type": "heuristic",
+      "phase": 33.5,
+      "description": "Phase 33.5 static security audit of GDD's shipped runtime surface (hooks/scripts/sdk/bin) — outbound-network call sites, secret-handling sites, and external-input surfaces; human-readable companion to scripts/security/outbound-allowlist.json (the canonical active-egress allowlist the 33.5-04 scan-outbound-network.cjs gate consumes) and reference/gdd-threat-model.md."
+    },
+    {
+      "name": "openrouter-tier-mapping",
+      "path": "reference/openrouter-tier-mapping.md",
+      "type": "heuristic",
+      "phase": 33.6,
+      "description": "Phase 33.6 OpenRouter tier-mapping heuristic — maps GDD opus/sonnet/haiku onto OpenRouter catalog ids via closed-vs-open + pricing buckets (high/medium/low), with the .design/config.json#openrouter_tier_overrides escape hatch (override wins) and graceful-null → native fallback."
+    },
+    {
+      "name": "prices-openrouter",
+      "path": "reference/prices.openrouter.md",
+      "type": "data",
+      "phase": 33.6,
+      "description": "Phase 33.6 catalog-derived OpenRouter price sub-table — per-model prompt/completion $/tok snapshot of .design/cache/openrouter-models.json; derived view, the dynamic catalog is the source of truth (D-11 registry round-trip)."
     }
   ]
 }