docs(agents): fix stale push-step count (5 -> 2 variant)

AGENTS.md said the retry-wrapped 'docker buildx build --push' steps were
'1 base + 5 variant'. Ground truth from docker-publish-split.yml: the retry
loop wraps exactly 3 push steps — build-base + build-variant-base +
build-variant-omos (1 base + 2 variant). promote-base-latest uses crane copy
(not a buildx push) and the smoke jobs use load:true. Stale count was a
leftover from the old with-pi/omos-with-pi/pi-only variant lineup.

Docs-only; no image change, no release.

AGENTS.md

@@ -105,7 +105,7 @@ curl -s https://api.github.com/repos/anomalyco/opencode/releases/tags/v1.15.10 |
 - **`OMOS_VERSION` MUST be passed by CI as a concrete version**, not left at the `latest` default. The npm install step in `Dockerfile.variant` (`oh-my-opencode-slim@${OMOS_VERSION}`) produces an identical layer-hash when the ARG value is byte-identical across builds; combined with the registry buildcache (`base-buildcache`) the layer gets reused even when `latest` would have resolved to a newer upstream. This is the same class of bug that bit pi-devbox v0.74.0 → v0.75.5 (silent same-bytes-across-releases regression discovered 2026-05-23, fixed in pi-devbox v0.75.5b). It is currently *masked* in opencode-devbox by `OPENCODE_VERSION` being a hard-coded ARG that bumps every release — that bump invalidates the parent-chain cache key for the downstream omos layer — but the masking would fail the moment a `vN.N.Nb` opencode-version-unchanged release ships that only bumps omos. Preventative fix: `.gitea/workflows/docker-publish-split.yml` has a `resolve-versions` job that runs `npm view oh-my-opencode-slim version`, exposing the concrete value as an output that the omos smoke + build jobs consume via build-args. Smoke tests assert via the `EXPECTED_OMOS_VERSION` env var — would catch the regression on the next release rather than several releases later. **If you change the variant build-args list, the resolve-versions job, or the smoke EXPECTED_*_VERSION wiring, audit all affected jobs in lockstep.**
 - **`resolve-versions` also pins `mempalace-toolkit` to a SHA** — `resolve-versions` resolves the `mempalace-toolkit` `main` HEAD to a commit SHA (`mempalace_toolkit_ref` output) via the gitea commits API (`/api/v1/repos/joakimp/mempalace-toolkit/commits?limit=1&sha=main`; gitea allows **unauthenticated** public-repo listing, so no secret is required). Unlike every other dependency, `mempalace-toolkit` is cloned in **`Dockerfile.base`**, not the variant — so the resolve→build-arg→variant plumbing bypasses it. To make a moved toolkit actually land, the resolved SHA is **folded into the `base-decide` hash** (so `base_tag` changes → base rebuilds) AND passed to `build-base` as `--build-arg MEMPALACE_TOOLKIT_REF`. Consequently **`base-decide` now depends on `resolve-versions`** (they no longer run in parallel), and the base clone uses a SHA-capable `git fetch <ref> + checkout FETCH_HEAD` (a `git clone --branch <40-char-SHA>` would fail). Trade-off: `base_tag` now reflects a live gitea lookup — on an API blip it falls back to `main`, triggering one *extra* base rebuild, never a *missed* one. If you touch `resolve-versions`, `base-decide`'s hash inputs, or the `build-base` build-args, audit `.gitea/README.md` Step 1 in lockstep.
 - **Registry buildkit cache-export is currently disabled** — do NOT re-add `cache-from`/`cache-to` to the `build-base` step in `.gitea/workflows/docker-publish-split.yml` without first verifying that buildkit's `mode=max` cache-export to `registry-1.docker.io` no longer returns HTTP 400 from the Hub CDN edge. The regression surfaced ~2026-05-23 and broke five consecutive opencode-devbox publish attempts (runs #332/333/334/336 + a rerun); root-caused on 2026-05-28 by a manual host-side publish that reproduced the same 400 only on `--cache-to` while image push worked fine. Failure shape is stable (`Offset:0` in the `_state` token, HTML response body = CDN-tier rejection, not registry backend), repo-specific (we're the only repo writing `:base-buildcache` mode=max), and explains why pinning `setup-buildx-action@v4.0.0` didn't help (action pin doesn't change the bundled buildkit version on the catthehacker runner image). Trade-off: dockerfile.base changes pay a full ~3 min rebuild instead of pulling cached layers; unchanged bases short-circuit at the Hub-probe step in `base-decide` and never re-build anyway. Variants don't use registry cache so they're unaffected. Re-enable condition: upstream moby/buildkit fix lands AND a low-risk test run succeeds without 400s. See CHANGELOG v1.15.12 `Unreleased` block for the full diagnostic chain. Manual escape-hatch publish procedure: `docs/manual-host-publish.md`.
-- **Push steps wrap `docker buildx build --push` in a 3-attempt retry loop** (15s, 30s backoff) for transient `registry-1.docker.io` blips — rate limits, brief 5xx, CDN flap. Implemented as inline `shell: bash` steps with `docker buildx build` raw rather than `docker/build-push-action@v7` so the loop is visible and tweakable. Affects the 1 base + 5 variant push steps in `.gitea/workflows/docker-publish-split.yml`; smoke-test builds (`load: true`, no push) are untouched. **This does NOT mask deterministic failures** — a true regression (like the cache-export 400 of 2026-05-23..28) fails all 3 attempts identically and the job still fails. Orthogonal to the cache-export disablement above: cache-export was about a deterministic protocol mismatch, retry is about absorbing genuine transients. Both are belt-and-braces with the `ci-release-watcher` skill's transient-rerun heuristic. If you change the matrix of push steps, keep the retry wrapper consistent across them — the pattern is duplicated rather than factored out because Gitea Actions doesn't support reusable composite shell steps cleanly.
+- **Push steps wrap `docker buildx build --push` in a 3-attempt retry loop** (15s, 30s backoff) for transient `registry-1.docker.io` blips — rate limits, brief 5xx, CDN flap. Implemented as inline `shell: bash` steps with `docker buildx build` raw rather than `docker/build-push-action@v7` so the loop is visible and tweakable. Affects the 1 base + 2 variant push steps in `.gitea/workflows/docker-publish-split.yml`; smoke-test builds (`load: true`, no push) are untouched. **This does NOT mask deterministic failures** — a true regression (like the cache-export 400 of 2026-05-23..28) fails all 3 attempts identically and the job still fails. Orthogonal to the cache-export disablement above: cache-export was about a deterministic protocol mismatch, retry is about absorbing genuine transients. Both are belt-and-braces with the `ci-release-watcher` skill's transient-rerun heuristic. If you change the matrix of push steps, keep the retry wrapper consistent across them — the pattern is duplicated rather than factored out because Gitea Actions doesn't support reusable composite shell steps cleanly.
 - **Shell scripts use `set -euo pipefail`** — both entrypoints are strict. Errors in volume chown or SSH permission operations are intentionally suppressed with `|| true`.
 - **Background subagents flag baked ON — experimental, watch for promotion** — `Dockerfile.base` sets `ENV OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS=true`. opencode gates native background subagents behind this flag (`packages/opencode/src/tool/task.ts` fails with `Background subagents require OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS=true` when unset); `oh-my-opencode-slim` V2+ makes background orchestration its **default** workflow, so the omos variant is effectively degraded without it. It's a base ENV (applies to both variants; harmless on plain opencode — only *enables* a capability) and stays runtime-overridable (`-e …=false`). It's counted in the base hash, so editing that line advances `base-latest`. **REMOVAL TRIGGER:** when opencode promotes background subagents out of `EXPERIMENTAL_` (flag renamed or made default), drop the ENV. No upstream roadmap date as of opencode 1.17.20 / omos 2.2.0 (2026-07). Documented in lockstep in README env table, `.env.example`, and asserted by `scripts/smoke-test.sh` (`bg-subagents env baked`).
 - **MemPalace `diary_write` anyOf workaround — upstream watch target** — `Dockerfile.base` carries a perl RUN block that strips a root-level `anyOf` from `mempalace_diary_write`'s advertised `inputSchema`. Mempalace 3.3.x/3.4.0 advertise `anyOf: [{required:[entry]},{required:[content]}]`, which Anthropic's tools API (and Codex) reject at session start (`input_schema does not support oneOf, allOf, or anyOf at the top level`), making the whole MCP server fail to load. The workaround is idempotent and self-deactivating: when upstream ships the real fix the regex stops matching and the build prints `WARN: ... upstream may have changed shape` — **that WARN is the signal to delete the RUN block.** Upstream status (last checked **2026-06-14**): issue **#1728 is still OPEN**; PR **#1735 is CLOSED UNMERGED (2026-06-11) — do NOT watch it, it is dead**; PR **#1717 is the current live fix candidate**; mempalace PyPI latest is **3.4.0 (== our pin)**, so **no release contains the fix yet** and the workaround must stay. **Removal trigger:** a mempalace release **> 3.4.0** that actually strips the root-level `anyOf` lands on PyPI — then bump `MEMPALACE_VERSION` (in lockstep with pi-devbox) and drop the RUN block. NOTE: `MEMPALACE_VERSION` (the pip pin) and `MEMPALACE_TOOLKIT_REF` (the git ref for the `mempalace-toolkit` clone) are unrelated despite the shared prefix; do not conflate them.