diff --git a/CHANGELOG.md b/CHANGELOG.md
index 0c8ef9c5f..37860858c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,21 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.12.2-beta.1] - 2026-09-22
+
+> **Beta channel release.** Publishes to the npm `beta` dist-tag; `latest` (0.12.1) is unaffected. Install: `npm i -g @altimateai/altimate-code@beta`.
+
+Six bug fixes from the first headless triage of the workspace pilot and from the documented-configuration audit that followed it. No new features; nothing here is behind a flag except where noted. Each fix went through three bot reviewers, a Codex review and an end-to-end run against a live tenant before merging.
+
+### Fixed
+
+- **Every documented `ALTIMATE_CLI_*` environment variable is now read** — not only `YOLO` and `DISABLE_AUTOUPDATE`. `docs/docs/usage/cli.md` documents the flags under `ALTIMATE_CLI_*`, but most of the table (external-skill scanning, autocompact, default plugins, LSP download, models fetch, project config, prune, terminal title, Exa, the `EXPERIMENTAL*` family, `CONFIG`, `CONFIG_CONTENT`, `CONFIG_DIR`, `GIT_BASH_PATH`, `PERMISSION`, `SERVER_USERNAME`/`SERVER_PASSWORD`) was read under the `OPENCODE_*` spelling only, so the documented name silently did nothing. One rule now applies on every read path — the two flag modules, the Effect-config services skill discovery and server auth actually read, and the direct reads in config loading, `run --attach`, the legacy database path, the updater and the feature census. The documented name wins when both are set; an empty documented value counts as unset; the `OPENCODE_*` spellings keep working. (#1341, closes #1329)
+- **`/skills` opens the Altimate skills browser, and its actions work.** `/skills` used to open the plain core skill selector (two `/skills` rows in autocomplete; Enter took the wrong one), where ctrl+a did nothing because the plugin's global keybind was outranked by the open dialog. The browser's Actions / New / Install are dialog-level actions now — bound inside the dialog to ctrl+a / ctrl+e / ctrl+g, and rendered as footer buttons reachable with Tab, so no chord is required. Install moved off ctrl+i, which most terminals send as Tab. New and Install work with the list filtered to nothing (the create-from-filter flow), and the palette's "Skills" row and a configured `prompt_skills` keybind both open the browser. (#1342, closes #1328)
+- **`altimate_core_validate` no longer reports a correct query as `ColumnNotFound` against Snowflake metadata.** Warehouse metadata comes back uppercase; the engine compares unquoted identifiers in lowercase and quoted ones exactly. Metadata names are now folded to the engine's form, and quoted all-uppercase references in the SQL (dbt `quote_columns` style) are folded to meet them — only for names the schema actually holds, so a `"SHIPPED"` string literal is untouched. The same preparation applies to every operation that matches SQL against a schema (lint, explain, check, fix, rewrite, equivalence, lineage, and the `sql.*` handlers), generated SQL comes back in the caller's spelling, and a schema file in JSON/YAML is treated like an inline context. Without a schema — or with one that defines no tables — existence checks are skipped as the tool promises, instead of reporting every table missing or failing the call. Known limitation: the fold is dialect-blind; on a lowercase-folding warehouse an all-uppercase quoted-created identifier would be treated as unquoted. (#1343, closes #1333)
+- **"Remember this for the team" lands where teammates read it, and a one-shot `run` no longer loses the upload.** In a linked project the model reached for the engine's `datamate_*` memory store, which linked checkouts never read; the `## Altimate Workspace` section now names `altimate_memory_write` as the team's store once the workspace's memory is confirmed enabled (and not while the link is only "last known"), and the tool description says the same. `run` waits — bounded — for pending memory mirrors and archives before exiting, on a normal exit and, briefly, on Ctrl-C. Pilot only (`ALTIMATE_WORKSPACE=1`). (#1344, closes #1332)
+- **A headless `run` always ends with an answer.** When a tool call failed or was auto-rejected (nobody can approve in headless use) and the model stopped without text — or streamed a "Let me check…" preamble, called a tool, and then stopped — the process printed nothing and exited 0. `run` now asks for a reply once, naming the failed tool; if the model still says nothing, a synthesised line says so on stdout, in `--output` and as a `silent_turn` event in `--format json`, and the run exits 1. Tool diagnostics are not repeated into the follow-up prompt or the answer file. A follow-up that dies in transport reports the error instead of only an exit code. (#1345, closes #1334)
+- **FinOps tools in a linked project point at the workspace engine instead of failing bare.** `finops_warehouse_advice`, `finops_analyze_credits`, `finops_query_history`, `finops_expensive_queries` and the role tools resolve only local connections; in a project whose Snowflake connection lives in the workspace they failed four times before the model thought of the engine. The failure now says why and names the engine tool with the tables the operation reads (region-qualified for BigQuery), re-validated against the current link; in the end-to-end run the model pivoted after one failure. When the routing decision is unknown the failure says that too. Running the FinOps SQL through the engine itself is not in this release. (#1346, closes #1336)
+
 ## [0.12.1] - 2026-09-21
 
 Two workspace-pilot additions that landed right after 0.12.0, plus the fixes their joint review turned up. Numbered as a patch because everything under **Added** is behind `ALTIMATE_WORKSPACE=1`; the only changes that reach every user are two reworded strings under **Changed**. **Heads-up for support (pilot):** a session launched from the VS Code / Cursor extension now follows the workspace picked in the extension's panel for skills and memory — but warehouse tool routing still follows the project's own link (#1337), and the agent says so when asked.