docs: update consult-advisor docs for configurable model [doc-updater]

docs/guides/pipeline-health-monitoring.md

@@ -418,7 +418,7 @@ When a Tier 1 `infrastructure_error` alert reaches the overseer monitor, it is r
 
 ### Advisor Gate
 
-Issue [#1962](https://github.com/jwbron/egg/issues/1962) introduces the **advisor strategy** for the overseer's decision tier: Haiku continues to drive every cycle (`max_turns=1`), and an Opus 4.6 advisor is invoked **only when both** of these conditions hold simultaneously:
+Issue [#1962](https://github.com/jwbron/egg/issues/1962) introduces the **advisor strategy** for the overseer's decision tier: Haiku continues to drive every cycle (`max_turns=1`), and the configured advisor model (`PipelineConfig.overseer_advisor_model`, defaulting to the `opus` alias) is invoked **only when both** of these conditions hold simultaneously:
 
 1. **Haiku flags an anomaly** with classification confidence ≥ 0.8.
 2. **A Tier-1 orchestrator health alert is currently active** (the same intersection precedent shipped in [#2012](https://github.com/jwbron/egg/issues/2012)).
@@ -431,7 +431,7 @@ The intersection gate keeps the heavy-tier model out of every poll cycle while s
 | `alert` | Emit an `OVERSEER_ALERT` carrying the advisor's `alert_summary`, `alert_detail`, and translated `priority`. The advisor returns `priority` as `p0..p3`; `egg_overseer.priority.label_to_alert` maps to the alert verb's `low|medium|high` dimension. |
 | `file_issue` | Emit an `OVERSEER_ALERT` whose `recommendation=file_issue` carries a fully composed `issue_title` + `issue_body` + `priority` + `anomaly_signature` in `recommendation_payload`. The CLI verb is **not** invoked here — see [Auto-Issue Filing (Shadow vs Live)](#auto-issue-filing-shadow-vs-live). |
 
-The advisor is exposed to the sandbox as a CLI verb (`egg-orch overseer consult-advisor`); the handler at `sandbox/egg_lib/orch_cli.py::cmd_overseer_consult_advisor` calls `consult_advisor` from `shared/egg_overseer/advisor.py` directly. The underlying `run_agent_async` Opus call therefore runs sandbox-side and stays on the LLM-execution side of the EGG200 boundary documented in [agent-mode-design.md](agent-mode-design.md) — the orchestrator pod never holds Anthropic credentials. The CLI verb reads the keyword arguments (`classification`, `health_alerts`, `progress_events`, `recent_log_lines`) that comprise the executor → advisor prompt contract from a JSON file passed via `--inputs-file`.
+The advisor is exposed to the sandbox as a CLI verb (`egg-orch overseer consult-advisor`); the handler at `sandbox/egg_lib/orch_cli.py::cmd_overseer_consult_advisor` calls `consult_advisor` from `shared/egg_overseer/advisor.py` directly. The underlying `run_agent_async` call therefore runs sandbox-side and stays on the LLM-execution side of the EGG200 boundary documented in [agent-mode-design.md](agent-mode-design.md) — the orchestrator pod never holds Anthropic credentials. The model used is resolved from `PipelineConfig.overseer_advisor_model` (read via the orchestrator status endpoint) when a pipeline ID is available; falls back to the `opus` default when absent or the lookup fails. The CLI verb reads the keyword arguments (`classification`, `health_alerts`, `progress_events`, `recent_log_lines`) that comprise the executor → advisor prompt contract from a JSON file passed via `--inputs-file`.
 
 **No advisor cap is enforced in this PR** — the existing `max_llm_cost_per_hour=$5` envelope at `sandbox/agent-config/rules/overseer.md` remains the only budget control. A follow-up issue tracks an `overseer_advisor_max_uses_per_phase` (or equivalent) knob if production data shows the cap is needed.
 
@@ -534,7 +534,7 @@ The sandbox-side CLI verb is the one and only way the overseer invokes the advis
 | `progress_events` | `list[dict]` | Recent structured progress events |
 | `recent_log_lines` | `list[str]` | Tail of agent container log lines |
 
-The handler at `sandbox/egg_lib/orch_cli.py::cmd_overseer_consult_advisor` calls `egg_overseer.advisor.consult_advisor()` directly. Output is the JSON-serialized `AdvisorVerdict` written to `--output-file` (or stdout when omitted). The Opus `run_agent_async` call runs sandbox-side, keeping the LLM call on the LLM-execution side of the EGG200 boundary; the orchestrator pod never holds Anthropic credentials.
+The handler at `sandbox/egg_lib/orch_cli.py::cmd_overseer_consult_advisor` calls `egg_overseer.advisor.consult_advisor()` directly. Output is the JSON-serialized `AdvisorVerdict` written to `--output-file` (or stdout when omitted). The `run_agent_async` call runs sandbox-side, keeping the LLM call on the LLM-execution side of the EGG200 boundary; the orchestrator pod never holds Anthropic credentials. The model alias is resolved from `PipelineConfig.overseer_advisor_model` when a pipeline ID is provided (positional arg or `EGG_PIPELINE_ID`), falling back to `opus` when absent or the lookup fails.
 
 **Backwards compatibility — top-level optional fields with omit-when-unset serialization.** The `OVERSEER_ALERT` schema gains three first-class optional fields on the `Message` envelope (`orchestrator/message_store.py`): `recommendation: str | None`, `recommendation_payload: dict | None`, and `schema_version: int = 1`. The `Message.to_dict()` serializer **omits** each of the three fields when they hold their defaults, so legacy callers that don't set them produce JSON byte-identical to the pre-#1962 shape. New consumers branch on the presence of `recommendation` (or, equivalently, `schema_version >= 2`); old consumers see no envelope change. The `egg-orch overseer alert` CLI gains `--recommendation file_issue --recommendation-payload-file /path/to/payload.json` flags that populate the new fields; the sandbox handler at `sandbox/egg_agent_tools/handlers/progress.py::progress_overseer_alert` enforces a 50 KB cap on the payload and validates that `recommendation` is one of the legal values (`file_issue` is currently the only accepted value).
 

docs/reference/orchestrator-cli.md

@@ -39,7 +39,7 @@ Run `egg-orch --help` for full usage. All commands support `--json` for machine-
 | `egg-orch message send [<id>] --to <role\|all> --type <type> --subject "..." --body "..."` | Send directed or broadcast message. Types: `HANDOFF`, `STATUS`, `PROGRESS`, `HEARTBEAT`. (`QUESTION` was removed in [#1897](https://github.com/jwbron/egg/issues/1897).) |
 | `egg-orch overseer alert [<id>] --anomaly <type> --priority <low\|medium\|high> --summary "..." [--detail "..."] [--recommend "..."] [--recommendation file_issue] [--recommendation-payload-file <path>]` | Broadcast `OVERSEER_ALERT` to human operator (overseer use only — always sets `message_type=OVERSEER_ALERT` and `to_role=all`). `--recommendation file_issue` attaches a structured advisor verdict; requires `--recommendation-payload-file` (JSON path with `issue_title`/`issue_body`/`priority`/`anomaly_signature`). |
 | `egg-orch overseer file-issue [<id>] --anomaly-type <type> --priority <p0\|p1\|p2\|p3> --agent-role <role> --anomaly-signature <hex16> --issue-title-file <path> --issue-body-file <path> [--parent-alert-message-id <id>] [--dry-run] [--json]` | File a GitHub issue from the overseer role (advisor-gated). Checks `filed-issues.jsonl` + `gh issue list` for deduplication; skips filing if a matching open issue exists. Default stdout is plain text (`Filed issue #N (...)` or `Existing issue #N already covers ...`); with `--json`, prints JSON `{"issue_number": int, "filed": bool, "dedup_match": int\|null}`; with `--dry-run` (no `gh` invocation either way): if a dedup match is found, prints the same `--json` shape (with `issue_number`/`dedup_match` populated); otherwise prints `{"issue_number": null, "filed": false, "dedup_match": null, "dry_run": true, "argv": [...], "title": str, "body_bytes": int}`. Requires `EGG_PIPELINE_REPO` env var. |
-| `egg-orch overseer consult-advisor [<id>] --inputs-file <path> [--output-file <path>] [--recent-log-bytes-cap <n>] [--json]` | Consult the Opus advisor for a structured `AdvisorVerdict` (sandbox-side LLM call). Reads a JSON inputs file with `classification`, `health_alerts`, `progress_events`, and `recent_log_lines`. Returns `decision` (`alert`\|`file_issue`\|`watch`), `priority`, `alert_summary`, `alert_detail`, `issue_title`, `issue_body`, and `reasoning`. Without `--output-file`, the JSON verdict is written to stdout; with `--output-file`, the verdict is written to that path and stdout shows a confirmation message (pass `--json` to additionally echo the verdict to stdout). `--recent-log-bytes-cap` overrides the `PipelineConfig.overseer_advisor_recent_log_bytes_cap` value (default 256 KiB); oldest lines are dropped first when the block exceeds the cap. `0` disables the cap. |
+| `egg-orch overseer consult-advisor [<id>] --inputs-file <path> [--output-file <path>] [--recent-log-bytes-cap <n>] [--json]` | Consult the advisor for a structured `AdvisorVerdict` (sandbox-side LLM call). Reads a JSON inputs file with `classification`, `health_alerts`, `progress_events`, and `recent_log_lines`. Returns `decision` (`alert`\|`file_issue`\|`watch`), `priority`, `alert_summary`, `alert_detail`, `issue_title`, `issue_body`, and `reasoning`. Without `--output-file`, the JSON verdict is written to stdout; with `--output-file`, the verdict is written to that path and stdout shows a confirmation message (pass `--json` to additionally echo the verdict to stdout). When `[<id>]` (or `EGG_PIPELINE_ID`) is set, the verb reads `PipelineConfig.overseer_advisor_model` from the orchestrator status endpoint and passes the configured model alias to `consult_advisor`; falls back to the `opus` default if the pipeline ID is absent or the lookup fails. `--recent-log-bytes-cap` overrides the `PipelineConfig.overseer_advisor_recent_log_bytes_cap` value (default 256 KiB); oldest lines are dropped first when the block exceeds the cap. `0` disables the cap. |
 | `egg-orch message poll [<id>] [--since <id>] [--limit <n>]` | Poll for messages from other agents (concurrent mode) |
 | `egg-orch message wait [<id>] --for <TYPE>... [--timeout N] [--since <id>]` | Block until a **new** typed BRC event arrives (cursor-less calls start at stream tip — already-seen events are skipped). Pass `--since <id>` for zero-drop send→wait semantics. Exit 0 = matched, 1 = timeout, 2 = transient (retry-safe), 3 = permanent. See [Agent Wait Patterns §3](agent-wait-patterns.md#3-exit-code-contract-for-egg-orch-message-wait) |
 | `egg-orch message wait-loop [<id>] --for <TYPE>... [--since <id>]` | **Canonical STAY ALIVE idiom** — loops `message wait` server-side until a new matching event arrives (defaults to stream-tip; pass `--since <id>` for zero-drop send→wait). Do not wrap in an outer shell loop. See [Agent Wait Patterns §1](agent-wait-patterns.md#1-the-canonical-idiom) |