[quickbeard]

Issue for this PR

Closes #

Type of change

• Bug fix

What does this PR do?

Two bugs, both in text the model or the user's browser ends up pointed at.

1. The agent didn't know which product it is

Asked "who are you", the agent answered "I'm opencode", and it told users to report
issues at github.com/anomalyco/opencode and read the docs at opencode.ai.

The rebrand so far covered the UI. It never covered the text the model actually reads:
SystemPrompt.provider() picks a prompt file by model id, and all 10 of those files are
upstream's verbatim — 20 brand mentions across identity lines, the /help line, the
feedback URL and the docs URL. MiniMax-M3 matches no branch, so it lands on default.txt,
whose first line is literally You are opencode, an interactive CLI tool….

Names are now rewritten as the prompt is assembled, by a new fork-only
packages/codev-brand, at the three points every model-visible string flows through:

• SystemPrompt.provider — upstream's branch chain is renamed select() and left
  byte-identical; provider() returns select(model).map(rebrand)
• the tool registry's assembled description — covers tool/lsp.txt and anything added later
• the built-in config skill — renamed customize-opencode → customize-codev, description
  and body rebranded; the old name still resolves, so existing configs and skill permission
  rules keep working

Why a transform instead of editing the prompt files. They are upstream's highest-churn
text. Editing them buys 10 permanent conflict sites, and when upstream adds a new brand
mention in a new paragraph it lands silently unrebranded. Leaving them byte-identical means
their sync hunks always apply, and new mentions get rebranded on arrival. initialize.txt
is edited directly instead — it is already fork-divergent and barely churns.

Identity is also stated in the environment block, because an agent that defines its own
prompt replaces the provider prompt wholesale (llm/request.ts) and would otherwise have
no idea what product it is.

A blind find-replace would break things, so the rules split on a preserve-pattern first.
customize-opencode.md alone contains @opencode-ai/plugin, the plugin names
opencode-gemini-auth / opencode-foo@1.2.3 / opencode-bar, and
https://opencode.ai/config.json ×5 — all identifiers the fork deliberately keeps, so
rewriting them would make the model's advice wrong. opencode://, ai.opencode.* and
opencode.db are preserved for the same reason.

2. Half the docs links were 404ing

Chasing the docs URL the prompt should carry turned up a live bug in the links the app
opens in the user's browser. Their base was missing the site's locale segment and its
/code section. Probed against the live site (paths shown, host omitted — see below):

old URL	result
/codev/docs	307 → /codev/en/docs (worked by luck)
/codev/docs/themes	404
/codev/docs/providers	404
/codev/docs/code/github	200

So settings theme "Learn more" (both settings screens), the custom-provider dialog link,
and the CLI's Cloudflare-gateway hint were all sending users to a 404. All 10 docs links —
those three plus TUI "Open docs", the desktop Documentation menu item, the sidebar Help
buttons, and the GitHub-agent hint — now use the new base, which resolves 200 for every
one of them. Trailing slashes are dropped (the site 308-redirects …/themes/ to
…/themes) and the #custom-provider / #cloudflare-ai-gateway anchor ids were confirmed
present on the new providers page.

Not moved: the ripgrep download mirror. It is a public/ static asset, not
locale-routed content — old path 200, /en/docs/code/downloads/… 404. Moving it would
have bricked the offline-rg fallback on GitHub-blocked networks. Called out as an explicit
exception in its DIVERGENCES row.

3. One encoded host instead of ten plaintext ones

The host was spelled out in plaintext at all 10 call sites, which defeats the point of the
base64 host already used in codev-gateway/src/const.ts and core/src/ripgrep/binary.ts
("keep the host out of casual greps, not as a secret"). Rather than repeat an atob(...)
per site, every link now derives from a single DOCS_URL in @codev/brand, which is itself
base64-encoded — one encoded host in the tree, and one place to fix when the site moves
again. app and tui pick up a @codev/brand workspace dependency; it is a zero-dep leaf
using only atob, so it is safe in client runtime code.

The two tests that pinned a URL literal keep pinning it, base64-encoded, so they still
verify the value independently of the constant under test. CHANGELOG and DIVERGENCES prose
now points at DOCS_URL / MIRROR instead of spelling the host out — same effect as
encoding, without turning the fork's conflict-resolution guidance into base64.

A grep for the host across the repo now returns nothing.

DIVERGENCES.md gets 5 new rows for the prompt work, plus rows for packages/app/package.json
and .prettierignore and an extended packages/tui/package.json row; the docs-links row and
URLs bullet are rewritten around DOCS_URL and gain github.handler.ts, which was previously
untracked. DIVERGENCES.md itself is now prettier-ignored: formatting it realigns all ~120
table rows, turning any one-row edit into a whole-file diff in the one place that records
upstream-conflict guidance. Nothing formatted it before, so that pins the status quo. Its "Known rebrand gaps"
section no longer lists prompt files as deliberately upstream and documents the transform and
its preserve-list instead.

How did you verify your code works?

• New packages/opencode/test/session/brand.test.ts is the tripwire: it scans every
  embedded .txt under src/ plus the skill body and fails when a brand mention isn't
  covered by the rules. That is the part per-file edits can't give you — the next upstream
  sync that introduces new wording fails CI instead of shipping. It also asserts each
  provider() branch names the fork and routes support at the fork's URLs.
• New packages/codev-brand/test/brand.test.ts — 7 tests: rewrites, preserved identifiers,
  config-file renames, CLI-invocation handling, idempotence.
• End-to-end through the real CLI (codev debug skill): 19 rebranded mentions in the live
  skill body, with https://opencode.ai/config.json ×5, opencode-gemini-auth,
  opencode-foo, opencode-bar and @opencode-ai/plugin all intact.
• Every docs URL, old and new, probed against the live site with
  curl -sL -o /dev/null -w '%{http_code}' — the table above is that output. Anchor ids
  checked by grepping the rendered providers page.
• Root bun typecheck clean — all 32 workspace tasks, which is what the pre-push hook runs. Test runs:
  test/session + test/agent (429), test/tool (383), test/skill + test/cli/help (26),
  core test/plugin (226).
• One pre-existing unrelated failure: session/compaction.test.ts > stops quickly when aborted during retry backoff fails identically on a clean main (verified by stashing).

Screenshots / recordings

Not a UI change. Before, on MiniMax/MiniMax-M3:

who are you
I'm opencode, an interactive CLI tool that helps with software engineering tasks.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

🤖 Generated with Claude Code