@event4u/agent-config 2.12.0 → 2.13.0

package/docs/contracts/multi-tool-projection-fidelity.md

@@ -0,0 +1,109 @@
+# Multi-Tool Projection Fidelity Contract
+
+**Status:** beta · **Phase 4 of [step-1-v2-feedback-followup](../../agents/roadmaps/step-1-v2-feedback-followup.md)**
+
+Names the **per-tool guarantees** the projection pipeline (`scripts/compress.py --sync` + `scripts/compress.py --generate-tools`) actually delivers. Byte-equivalence is not behaviour-fidelity — each consumer tool has its own frontmatter grammar, its own activation model, and its own surface for skills / rules / commands.
+
+## Source of truth
+
+Every projection starts from `.agent-src/` (compressed) which is generated from `.agent-src.uncompressed/`. The projection layer **never** writes to source; it only reads.
+
+## Per-tool projection map
+
+| Tool | Rules surface | Skills surface | Commands surface | Frontmatter grammar |
+|---|---|---|---|---|
+| **Augment** (host) | `.augment/rules/*.md` (copies; symlink opt-in via `augment.rules_use_symlinks`) | `.augment/skills/<name>/SKILL.md` (symlink → `.agent-src/skills/`) | `.augment/commands/*.md` | full source frontmatter preserved |
+| **Claude** (Code + Desktop) | `.claude/rules/*.md` | `.claude/skills/<name>/SKILL.md` | `.claude/skills/<name>/SKILL.md` (commands rendered as skills) | full source frontmatter preserved |
+| **Cursor** | `.cursor/rules/*.mdc` + legacy `.md` symlinks (130 files = 65 × 2) | **not projected** | `.cursor/commands/*.md` | `description`, `globs`, `alwaysApply` only — `triggers`, `routes_to`, `tier`, `type` are **dropped** |
+| **Windsurf** | `.windsurfrules` (single concatenated file) + `.windsurf/rules/*.md` (per-rule) | **not projected** | `.windsurf/workflows/*.md` | concatenated body; per-rule frontmatter only retained in `.windsurf/rules/`, not in the legacy `.windsurfrules` single-file |
+| **Cline** | `.clinerules/*.md` | **not projected** | **not projected** | full router frontmatter preserved (`type`, `tier`, `description`, `triggers`, `routes_to`) |
+| **Gemini** | `GEMINI.md` (single-file digest) | embedded inline | embedded inline | digest only — no per-rule frontmatter |
+| **Copilot** | `AGENTS.md` / `copilot-instructions.md` | embedded inline | embedded inline | digest only |
+
+`AGENTS.md` is the **tool-agnostic root pointer** and exists at workspace root regardless of which projections are enabled.
+
+## Fidelity guarantees per axis
+
+### 1. Rule body fidelity
+
+| Tool | Body identical to source? |
+|---|---|
+| Augment | yes (copy or symlink) |
+| Claude | yes (copy) |
+| Cursor `.mdc` | yes |
+| Cline | yes |
+| Windsurf single-file | concatenated, separator `---` between rules |
+| Windsurf per-rule | yes |
+| Gemini / Copilot digest | summarised — **no fidelity guarantee** |
+
+### 2. Trigger fidelity (`triggers:` keyword / `path_prefix`)
+
+| Tool | `triggers:` preserved? |
+|---|---|
+| Augment, Claude, Cline, Windsurf-per-rule | **yes** — the host LLM sees the trigger set verbatim |
+| Cursor `.mdc` | **no** — Cursor's frontmatter grammar does not honour `triggers:`; activation falls back to `globs:` + `alwaysApply: <bool>` + description match |
+| Windsurf single-file `.windsurfrules` | **no** — concatenated body strips per-rule frontmatter |
+| Gemini, Copilot | **no** — digest format |
+
+**Consequence:** rules that depend on `triggers:` for activation (tier-2a path-prefix routing, tier-3 keyword routing) **silently degrade on Cursor and on the Windsurf single-file**. They still appear in body, but the host must infer activation from prose.
+
+### 3. `routes_to:` fidelity
+
+Same matrix as `triggers:` — preserved on Augment, Claude, Cline, Windsurf-per-rule; **dropped** on Cursor `.mdc` and Windsurf single-file.
+
+**Consequence:** the four tier-3 routing rules (`laravel-routing`, `symfony-routing`, `copilot-routing`, `devcontainer-routing`) added in Phase 3.3 will route deterministically on Augment / Claude / Cline; on Cursor / Windsurf-single-file the host must rely on description matching alone.
+
+### 4. Skill surface
+
+Cursor, Windsurf, Cline, Gemini, Copilot have **no native skill surface**. Skills are projected only for Augment and Claude. Consumers on the other tools see skill content only indirectly (via rule bodies that cite skills, or via the catalogue in `AGENTS.md`).
+
+### 5. Command surface
+
+| Tool | Where commands appear |
+|---|---|
+| Augment | `.augment/commands/*.md` (native slash-command surface) |
+| Claude | `.claude/skills/<command>/SKILL.md` (commands rendered as skills with `disable-model-invocation: true`) |
+| Cursor | `.cursor/commands/*.md` (106 files) |
+| Windsurf | `.windsurf/workflows/*.md` (106 files) |
+| Cline | none |
+| Gemini, Copilot | listed only inside `AGENTS.md` / `GEMINI.md` digest |
+
+## Automated probe — `task lint-projection-fidelity`
+
+`scripts/probe_projection_fidelity.py` reads `tests/fixtures/projection_fidelity/fixtures.yml` and asserts the per-tool guarantees above against the actual projected trees. The fixture covers five representative artefacts:
+
+| Fixture entry | Tier | Stress-tests |
+|---|---|---|
+| `rule:non-destructive-by-default` | kernel | always-active body fidelity across all five rule surfaces |
+| `rule:laravel-translations` | tier-2a | `path_prefix:` trigger preservation (Cline) vs drop (Cursor) |
+| `rule:laravel-routing` | tier-3 | `routes_to:` preservation (Cline) vs drop (Cursor, Windsurf-single) |
+| `skill:laravel` | skill | Augment + Claude only; rationale for absence on others |
+| `command:commit` | command | per-tool command surface divergence |
+
+Run: `python3 scripts/probe_projection_fidelity.py` — exits non-zero on any divergence. Report at `agents/reports/projection-fidelity.json`.
+
+## Known divergences (do not file as bugs)
+
+These are **architectural facts**, not regressions. They are documented so installers and consumers know what to expect.
+
+1. **Cursor `.mdc` drops router metadata.** Cursor's third-party rule format only honours `description`, `globs`, `alwaysApply`. Adding `triggers:` or `routes_to:` to a Cursor rule has no effect at activation time. The body still loads when the description matches; the deterministic routing layer does not.
+2. **Windsurf single-file (`.windsurfrules`) strips per-rule frontmatter.** Legacy compatibility surface. The new `.windsurf/rules/*.md` per-rule files preserve the full frontmatter — consumers should prefer those.
+3. **Skills do not project to Cursor / Windsurf / Cline / Gemini / Copilot.** These tools have no native skill loader. Skill content reaches consumers indirectly via rule bodies and the `AGENTS.md` catalogue.
+4. **Augment historically did not load symlinked rules.** Default is to **copy** rules into `.augment/rules/`. Opt into symlinks via `augment.rules_use_symlinks: true` in `.agent-settings.yml`.
+5. **`task generate-tools` does not refresh `.augment/rules/`.** Only `task sync` (== `scripts/compress.py --sync`) copies rules into the Augment tree. Investigators who edit a rule, run only `generate-tools`, and then `ls .augment/rules/` will see stale state.
+
+## Acceptance criteria for this contract
+
+- [x] Fixture under `tests/fixtures/projection_fidelity/`
+- [x] Probe script under `scripts/probe_projection_fidelity.py`
+- [x] Report under `agents/reports/projection-fidelity.json`
+- [x] Per-tool guarantee table above
+- [x] Known-divergence list above
+
+## Related
+
+- [`source-projection`](../architecture/source-projection.md) — pipeline A (source compression)
+- [`augment-projection`](../architecture/augment-projection.md) — pipeline B (Augment-specific)
+- [`multi-tool-projection`](../architecture/multi-tool-projection.md) — pipeline C (the per-tool emitters)
+- [`rule-router`](rule-router.md) — the `triggers:` / `routes_to:` grammar this contract pins
+- [`agents/council-sessions/2026-05-14-v2-analysis/feedback/09-cross-tool-projection-fidelity.md`](../../agents/council-sessions/2026-05-14-v2-analysis/feedback/09-cross-tool-projection-fidelity.md) — origin council feedback

package/docs/getting-started.md

@@ -106,7 +106,7 @@ Your agent is now:
 - **Respecting your codebase** — no conflicting patterns
 - **Following standards** — consistent code quality
 
-This is enforced automatically by 61 rules. No configuration needed.
+This is enforced automatically by 65 rules. No configuration needed.
 
 ---
 
@@ -146,7 +146,7 @@ Your agent now understands slash commands:
 | `/quality-fix` | Run and fix all quality checks |
 | `/chat-history` | Inspect the persistent chat-history log (read-only `show`) |
 
-→ [Browse all 106 active commands](../.agent-src/commands/)
+→ [Browse all 108 active commands](../.agent-src/commands/)
 
 ---
 

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@event4u/agent-config",
-    "version": "2.12.0",
+    "version": "2.13.0",
     "description": "Shared agent configuration \u2014 skills, rules, commands, guidelines, and templates for AI coding tools",
     "license": "MIT",
     "private": false,

package/scripts/_archive/README.md

@@ -0,0 +1,59 @@
+# Script Archive — One-Shot Migrations
+
+This directory preserves migration / bootstrap / back-fill scripts that have
+already run to completion and are no longer invoked by any productive code
+path. They are kept as forensic reference, mirroring the archival convention
+used by [`agents/roadmaps/archive/`](../../agents/roadmaps/archive/).
+
+**Do not run these scripts.** They are one-shot transformations whose
+target state is already the working tree. Re-running them on the current
+codebase is undefined behaviour.
+
+## Provenance
+
+Archived 2026-05-14 as part of [`agents/roadmaps/step-1-v2-feedback-followup.md`](../../agents/roadmaps/step-1-v2-feedback-followup.md)
+Phase 1 Step 3, addressing audit finding F3 / council finding C3 from
+[`agents/council-sessions/2026-05-14-v2-analysis/feedback/03-migration-scripts-archival.md`](../../agents/council-sessions/2026-05-14-v2-analysis/feedback/03-migration-scripts-archival.md).
+
+## Inventory
+
+| Script | Migration / phase served | What it did |
+|---|---|---|
+| [`_backfill_skill_domains.py`](_backfill_skill_domains.py) | B3 domain back-fill | Injected `domain:` frontmatter into every `SKILL.md`. Source of truth now lives in each skill's frontmatter directly. |
+| [`_bootstrap_tier_frontmatter.py`](_bootstrap_tier_frontmatter.py) | Tier-frontmatter bootstrap | Injected `tier: N` frontmatter into every slash command during the kernel / tier-1 / tier-2 routing introduction. |
+| [`_p43_bodies.py`](_p43_bodies.py) | Phase 4.3 — rule-body compression | Wrote compressed rule bodies after `_p43_compress.py` produced the manifest. Paired with `_p43_compress.py`. |
+| [`_p43_compress.py`](_p43_compress.py) | Phase 4.3 — rule-body compression | Surgical compression of 22 `compress-and-keep` auto-rules; produced the manifest consumed by `_p43_bodies.py`. |
+| [`_p4_migrate.py`](_p4_migrate.py) | Phase 4.1 + 4.2 — rule reclassification | Migrated rules into the skill / guideline / command / contract-stub split that the package ships today. |
+| [`_phase2_shim_helper.py`](_phase2_shim_helper.py) | Phase 2 — deprecation shim | One-shot helper that injected `superseded_by:` + `deprecated_in:` + deprecation warning into rules retired during Phase 2. |
+| [`_pilot_council_question.py`](_pilot_council_question.py) | Phase 1 pilot — kernel-membership council prep | Built the Phase-1 council question file used for the kernel-membership R1/R2 cross-check. The resulting council artefacts live under `agents/council-sessions/20260506T*`. |
+
+## Why these stayed live and were NOT archived
+
+The 2026-05-14 audit (F3) listed 9 candidate scripts. Two of those turn out
+to have productive (non-incestuous) references and remain in `scripts/`:
+
+- **`scripts/_emit_domain_table.py`** — cited as the regeneration command in
+  [`docs/contracts/skill-domains.md`](../../docs/contracts/skill-domains.md)
+  ("regenerate via `python3 scripts/_emit_domain_table.py`"). The
+  domain-table snapshot is a derived view that the contract doc explicitly
+  expects to be regenerable from this script.
+- **`scripts/_pilot_measure.py`** — cited by
+  [`docs/contracts/kernel-membership.md`](../../docs/contracts/kernel-membership.md)
+  as the reproducibility-verification command for the kernel pilot SHAs, and
+  its algorithm is mirrored by [`scripts/iron_law_sha.py`](../iron_law_sha.py).
+  Both productive paths assume the script remains in place.
+
+The audit's F3 framing ("zero productive references") was correct for the 7
+archived scripts and wrong for these 2. Recorded here so the F3 finding is
+not re-litigated without context.
+
+## How to restore one (if a future migration needs it)
+
+```bash
+git mv scripts/_archive/<script>.py scripts/
+git commit -m "chore(scripts): restore <script> for <reason>"
+```
+
+Restoration should come with an issue / PR explaining why the historical
+one-shot is being reused — by construction these scripts assume their
+pre-migration starting state.

package/scripts/ai_council/_default_prices.py

@@ -17,7 +17,7 @@ from __future__ import annotations
 
 # YYYY-MM-DD of when this table was last hand-edited. Keep in sync with
 # the test_default_prices freshness assertion if you bump this.
-LAST_UPDATED = "2026-04-29"
+LAST_UPDATED = "2026-05-14"
 
 # (provider, model)  ->  (input_per_1m_usd, output_per_1m_usd)
 DEFAULT_PRICES: dict[tuple[str, str], tuple[float, float]] = {
@@ -30,6 +30,15 @@ DEFAULT_PRICES: dict[tuple[str, str], tuple[float, float]] = {
     ("openai", "gpt-4o-mini"): (0.15, 0.60),
     ("openai", "o1"): (15.00, 60.00),
     ("openai", "o3-mini"): (1.10, 4.40),
+    # ── Google Gemini ────────────────────────────────────────────────
+    ("gemini", "gemini-2.5-pro"): (1.25, 10.00),
+    ("gemini", "gemini-2.5-flash"): (0.30, 2.50),
+    # ── xAI Grok ─────────────────────────────────────────────────────
+    ("xai", "grok-4"): (3.00, 15.00),
+    ("xai", "grok-3-mini"): (0.30, 0.50),
+    # ── Perplexity ───────────────────────────────────────────────────
+    ("perplexity", "sonar-pro"): (3.00, 15.00),
+    ("perplexity", "sonar"): (1.00, 1.00),
 }
 
 

package/scripts/ai_council/advisors.py

@@ -0,0 +1,148 @@
+"""Thinking-style advisors — replace-mode call planning (Phase 6).
+
+When `agents/.ai-council.yml` enables an advisor (e.g. `contrarian`
+bound to `member: anthropic`), the orchestrator REPLACES the matching
+plain-member call with an advisor-persona call on the same provider.
+Same total call count as a plain run; bounded extra cost beyond the
+persona-prompt token delta.
+
+This module owns:
+
+- `AdvisorPlan`  — resolved swap for a single provider (persona text,
+  display name, optional model override).
+- `plan_advisor_swap()` — walks the enabled advisors, reads their
+  persona files, and returns the per-provider plan map consumed by
+  `orchestrator.consult()` / `estimate()` and by the CLI.
+- `resolve_persona_text()` — reads a persona file with compressed-tree
+  preference and frontmatter strip.
+
+Cross-validation against the members block already ran at config load
+(`config._build_config`); this module trusts that contract and only
+enforces the **one-advisor-per-provider** rule (replace-mode invariant).
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from pathlib import Path
+
+import yaml
+
+from scripts.ai_council.config import AdvisorConfig, CouncilConfigError
+
+
+@dataclass(frozen=True)
+class AdvisorPlan:
+    """Resolved advisor swap for a single provider."""
+
+    name: str
+    display_name: str
+    member: str
+    persona_text: str
+    model_override: str | None = None
+
+
+_FRONTMATTER_RE = re.compile(r"\A---\n(.*?)\n---\n", re.DOTALL)
+
+
+def _split_frontmatter(raw: str) -> tuple[dict, str]:
+    """Return ``(frontmatter_dict, body)``. Missing frontmatter → ``({}, raw)``."""
+    match = _FRONTMATTER_RE.match(raw)
+    if not match:
+        return {}, raw
+    try:
+        meta = yaml.safe_load(match.group(1)) or {}
+    except yaml.YAMLError:
+        meta = {}
+    if not isinstance(meta, dict):
+        meta = {}
+    body = raw[match.end():]
+    return meta, body
+
+
+def _display_name_from(advisor_name: str, frontmatter: dict) -> str:
+    """Prefer frontmatter ``role``; fall back to titleized advisor key."""
+    role = frontmatter.get("role")
+    if isinstance(role, str) and role.strip():
+        return role.strip()
+    return advisor_name.replace("-", " ").replace("_", " ").title()
+
+
+def resolve_persona_text(
+    persona_path: str,
+    repo_root: Path,
+) -> tuple[str, dict]:
+    """Read a persona file, returning ``(body, frontmatter)``.
+
+    Compressed tree (``.agent-src/``) wins so production runs match the
+    same projection the rest of the package consumes. Uncompressed tree
+    (``.agent-src.uncompressed/``) is the fallback for in-repo
+    development before ``task sync`` has projected the file.
+    """
+    candidates = [
+        repo_root / ".agent-src" / persona_path,
+        repo_root / ".agent-src.uncompressed" / persona_path,
+    ]
+    for candidate in candidates:
+        if candidate.exists():
+            raw = candidate.read_text(encoding="utf-8")
+            meta, body = _split_frontmatter(raw)
+            return body.strip(), meta
+    searched = "\n  - ".join(str(c) for c in candidates)
+    raise CouncilConfigError(
+        f"Persona file not found for advisor (path={persona_path!r}). "
+        f"Searched:\n  - {searched}"
+    )
+
+
+def plan_advisor_swap(
+    advisors: dict[str, AdvisorConfig],
+    repo_root: Path,
+) -> dict[str, AdvisorPlan]:
+    """Return ``{provider_name: AdvisorPlan}`` for every ENABLED advisor.
+
+    Two enabled advisors targeting the same provider is a
+    ``CouncilConfigError`` — replace-mode runs one advisor per provider
+    so the call plan never doubles up by accident.
+    """
+    plans: dict[str, AdvisorPlan] = {}
+    for adv in advisors.values():
+        if not adv.enabled:
+            continue
+        if adv.member in plans:
+            existing = plans[adv.member].name
+            raise CouncilConfigError(
+                f"advisors.{adv.name} and advisors.{existing} both bind "
+                f"member={adv.member!r}; only one advisor per provider "
+                f"per run (replace-mode invariant)."
+            )
+        body, meta = resolve_persona_text(adv.persona, repo_root)
+        plans[adv.member] = AdvisorPlan(
+            name=adv.name,
+            display_name=_display_name_from(adv.name, meta),
+            member=adv.member,
+            persona_text=body,
+            model_override=adv.model,
+        )
+    return plans
+
+
+def build_persona_labels(
+    plans: dict[str, AdvisorPlan],
+    members: list,
+) -> dict[str, str]:
+    """Build the peer-review ``source → display_name`` map.
+
+    ``source`` is the ``provider:model`` string the peer-review
+    pipeline uses for anonymisation; ``members`` is the post-swap
+    member list (model_override already applied), so the model field
+    matches what the response carries.
+    """
+    labels: dict[str, str] = {}
+    for m in members:
+        plan = plans.get(m.name)
+        if plan is None:
+            continue
+        labels[f"{m.name}:{m.model}"] = plan.display_name
+    return labels

package/scripts/ai_council/clients.py

@@ -52,6 +52,15 @@ def _resolve_key_path(filename: str) -> Path:
 
 DEFAULT_ANTHROPIC_MODEL = "claude-sonnet-4-5"
 DEFAULT_OPENAI_MODEL = "gpt-4o"
+DEFAULT_GEMINI_MODEL = "gemini-2.5-pro"
+DEFAULT_XAI_MODEL = "grok-4"
+DEFAULT_PERPLEXITY_MODEL = "sonar-pro"
+
+#: OpenAI-API-compatible endpoints. xAI and Perplexity both expose the
+#: ``/v1/chat/completions`` shape, so their clients reuse the ``openai``
+#: SDK with a custom ``base_url``. Gemini has its own SDK (``google-genai``).
+XAI_BASE_URL = "https://api.x.ai/v1"
+PERPLEXITY_BASE_URL = "https://api.perplexity.ai"
 
 #: Per-call output budget when no caller-supplied value reaches `ask()`.
 #: The CLI resolves the live default from `ai_council.max_output_tokens`
@@ -269,6 +278,169 @@ class OpenAIClient(ExternalAIClient):
         )
 
 
+# ── Gemini / xAI / Perplexity (Phase 0 — Step 6) ─────────────────────
+
+
+class GeminiClient(ExternalAIClient):
+    """Google Gemini via the ``google-genai`` SDK.
+
+    Lazy-imports ``google.genai`` on first ``ask()`` so disabled
+    members do not require the SDK to be installed. Tests inject a
+    mock client shaped like ``genai.Client(api_key=...)`` —
+    ``self._client.models.generate_content(...)`` returns an object
+    with ``.text`` and ``.usage_metadata.{prompt_token_count,
+    candidates_token_count}``.
+    """
+
+    name = "gemini"
+    billable = True
+
+    def __init__(
+        self,
+        model: str = DEFAULT_GEMINI_MODEL,
+        client: object = None,
+        api_key: str | None = None,
+    ):
+        self.model = model
+        if client is not None:
+            self._client = client
+            return
+        if api_key is None:
+            raise RuntimeError(
+                "GeminiClient requires explicit api_key or injected client. "
+                "Use `api_key_ref: env:GEMINI_API_KEY` in agents/.ai-council.yml."
+            )
+        try:
+            from google import genai  # type: ignore[import-not-found]
+        except ImportError as exc:  # pragma: no cover - exercised only with real SDK
+            raise RuntimeError(
+                "google-genai package not installed. `pip install google-genai`."
+            ) from exc
+        self._client = genai.Client(api_key=api_key)
+
+    def ask(self, system_prompt: str, user_prompt: str, max_tokens: int = DEFAULT_MAX_TOKENS) -> CouncilResponse:
+        t0 = time.monotonic()
+        contents = f"{system_prompt}\n\n---\n\n{user_prompt}"
+        try:
+            response = self._client.models.generate_content(
+                model=self.model,
+                contents=contents,
+                config={"max_output_tokens": max_tokens},
+            )
+        except Exception as exc:  # noqa: BLE001 - normalise all SDK errors
+            return CouncilResponse(
+                provider=self.name, model=self.model, text="",
+                latency_ms=int((time.monotonic() - t0) * 1000),
+                error=f"{type(exc).__name__}: {exc}",
+            )
+        latency_ms = int((time.monotonic() - t0) * 1000)
+        text = getattr(response, "text", "") or ""
+        usage = getattr(response, "usage_metadata", None)
+        return CouncilResponse(
+            provider=self.name, model=self.model, text=text,
+            input_tokens=getattr(usage, "prompt_token_count", 0) if usage else 0,
+            output_tokens=getattr(usage, "candidates_token_count", 0) if usage else 0,
+            latency_ms=latency_ms,
+        )
+
+
+class _OpenAICompatibleClient(ExternalAIClient):
+    """Shared shape for OpenAI-API-compatible providers (xAI, Perplexity).
+
+    Both vendors implement ``/v1/chat/completions`` and accept the
+    ``openai`` Python SDK with a custom ``base_url``. The reasoning-
+    model branch from :class:`OpenAIClient` is intentionally omitted —
+    neither xAI nor Perplexity ships a reasoning model that requires
+    ``max_completion_tokens`` as of 2026-05-14.
+    """
+
+    billable = True
+    base_url: str = ""
+
+    def __init__(
+        self,
+        model: str,
+        client: object = None,
+        api_key: str | None = None,
+    ):
+        self.model = model
+        if client is not None:
+            self._client = client
+            return
+        if api_key is None:
+            raise RuntimeError(
+                f"{type(self).__name__} requires explicit api_key or injected client."
+            )
+        try:
+            import openai  # type: ignore[import-not-found]
+        except ImportError as exc:  # pragma: no cover - exercised only with real SDK
+            raise RuntimeError(
+                "openai package not installed. `pip install openai`."
+            ) from exc
+        self._client = openai.OpenAI(api_key=api_key, base_url=self.base_url)
+
+    def ask(self, system_prompt: str, user_prompt: str, max_tokens: int = DEFAULT_MAX_TOKENS) -> CouncilResponse:
+        t0 = time.monotonic()
+        try:
+            response = self._client.chat.completions.create(
+                model=self.model,
+                max_tokens=max_tokens,
+                messages=[
+                    {"role": "system", "content": system_prompt},
+                    {"role": "user", "content": user_prompt},
+                ],
+            )
+        except Exception as exc:  # noqa: BLE001 - normalise all SDK errors
+            return CouncilResponse(
+                provider=self.name, model=self.model, text="",
+                latency_ms=int((time.monotonic() - t0) * 1000),
+                error=f"{type(exc).__name__}: {exc}",
+            )
+        latency_ms = int((time.monotonic() - t0) * 1000)
+        text = ""
+        choices = getattr(response, "choices", None)
+        if choices:
+            msg = getattr(choices[0], "message", None)
+            text = getattr(msg, "content", "") if msg else ""
+        usage = getattr(response, "usage", None)
+        return CouncilResponse(
+            provider=self.name, model=self.model, text=text or "",
+            input_tokens=getattr(usage, "prompt_tokens", 0) if usage else 0,
+            output_tokens=getattr(usage, "completion_tokens", 0) if usage else 0,
+            latency_ms=latency_ms,
+        )
+
+
+class XAIClient(_OpenAICompatibleClient):
+    """xAI Grok via the OpenAI-compatible endpoint at api.x.ai/v1."""
+
+    name = "xai"
+    base_url = XAI_BASE_URL
+
+    def __init__(
+        self,
+        model: str = DEFAULT_XAI_MODEL,
+        client: object = None,
+        api_key: str | None = None,
+    ):
+        super().__init__(model=model, client=client, api_key=api_key)
+
+
+class PerplexityClient(_OpenAICompatibleClient):
+    """Perplexity via the OpenAI-compatible endpoint at api.perplexity.ai."""
+
+    name = "perplexity"
+    base_url = PERPLEXITY_BASE_URL
+
+    def __init__(
+        self,
+        model: str = DEFAULT_PERPLEXITY_MODEL,
+        client: object = None,
+        api_key: str | None = None,
+    ):
+        super().__init__(model=model, client=client, api_key=api_key)
+
+
 # ── Manual mode (Phase 2b) ───────────────────────────────────────────