Changelog

All notable changes to opencode-sdlc-wizard.

[0.13.4] - 2026-05-25

Fixed — `install.sh` next-steps drift (v0.11 / v0.12 / v0.13 additions missing)

Found by real consumption (npx opencode-sdlc-wizard@latest init in ~/project-tracker). install.sh's printed next-steps still showed the v0.10.x tier list, missing every v0.11/v0.12/v0.13 addition:

proprietary/zai (v0.11.0)
managed/opencode (v0.12.0 — entire 5th tier)
subscription/{github-copilot, chatgpt, grok} (v0.13.0 + v0.13.2 — entire 6th tier)

And never mentioned the pick subcommand at all (shipped v0.9.0!). Every consumer reaching for the wizard via npx ... init was being onboarded to the pre-v0.9.0 two-step UX. Three minor versions of documented features invisible to new users.

Same drift pattern as v0.8.0 → v0.8.4. Same root cause (install.sh next-steps hand-edited, not auto-derived). Recurred because the prior drift gate only checked provider tokens, not tier-table coverage or the pick mention.

install.sh changes

Promoted npx opencode-sdlc-wizard pick as the primary flow (4 example invocations: auto, free-tier-first, override, dry-run)
Two-step detect + configure listed as the explicit alternative
Added managed + subscription rows + zai provider
Added a per-agent routing example with --reviewer-* + --sandbox-*

Tests

tests/test-install.sh T9 drift gate strengthened: now asserts zai, managed, opencode, subscription, github-copilot, chatgpt, grok, and the word pick all appear in next-steps. 421/12 unit suites + 13/13 E2E green.

Compat

Doc-only fix in install.sh. No code or behavior changes elsewhere.
Every flag, provider, tier, default-model unchanged from v0.13.3.

Workflow insight

Two consumption-test bugs in a row (T78 no-clobber in v0.13.3, this next-steps drift in v0.13.4) both invisible to 421 unit tests. Each test exercised a script in isolation; neither inspected install.sh's printed onboarding text or pick-against-existing-config behavior.

Going forward — after any release adding a tier / provider / default-model entry:

npm run test:e2e                                              # local-pack smoke
cd ~/some-real-project && npx opencode-sdlc-wizard@latest init  # real onboard

60 seconds to catch what unit tests miss.

[0.13.3] - 2026-05-24

Fixed — `--print-only` no longer trips the no-clobber guard

Bug found by the new E2E consumption test (added this release): running pick --dry-run (translates to configure-backend --print-only) against a project that already has an opencode.json with a different model pin was exiting 4 ("opencode.json already has model=..., re-run with --force") without printing the preview.

That's wrong. --print-only is a read-only operation — it should be inspection-free of the destination, not just write-free. The dry-run preview use case is "show me what the wizard would write if I picked this tier" — tripping the guard there makes the preview impossible exactly when it's most useful (deciding whether to overwrite an existing config).

Fix: moved the no-clobber check inside the write branch. --print-only now prints the merged JSON unconditionally; --force gating only applies to actual file writes.

scripts/configure-backend.sh: no-clobber check moved out of the pre-condition block into the write-path conditional
tests/test-backend-picker.sh adds T78/T78b:
- T78 asserts --print-only prints preview even against existing differing opencode.json (no exit 4)
- T78b asserts the file was NOT modified (still read-only)

Added — `npm run test:e2e` consumption smoke test

New tests/test-e2e-consumption.sh exercises the published bundle (via npm pack of current source by default; E2E_SOURCE=npm-latest overrides to live npm) end-to-end:

npm pack → install with --package=<tarball> into fresh tmp dir
Verify install delivered pick-backend.sh + configure-backend.sh
- detect-backends.sh + .wizard-stamp
pick --dry-run across all 9 tier/provider combos (private_local, enterprise, hosted_oss, managed, proprietary x2, subscription x3)
Each emitted JSON parses cleanly; coder model + provider block shape validated per tier
Full v0.13.x hybrid invocation (every flag in one call) emits valid 6-agent + 4-provider config
Optional opencode debug config check if CLI is on PATH

Not in the default npm test chain because it takes ~30-60s (npm pack + install). Run explicitly:

npm run test:e2e                           # against local pack of current source
E2E_SOURCE=npm-latest npm run test:e2e    # against live npm latest

This test caught the T78 bug on first run. That's exactly the value proposition — automating the consumption path surfaces regressions that source-layout unit tests can't reach (installer behavior + cross-script integration + real JSON shape under the installed .opencode/ layout).

Changed — `tests/test-e2e-consumption.sh` design notes

Defaults to E2E_SOURCE=local-pack (uses npm pack of current repo); override with E2E_SOURCE=npm-latest to verify what published to npm
Uses npx --package=<tarball> opencode-sdlc-wizard init to install the packed tarball (vs npx <tarball> which sh tries to exec)
pick --dry-run --target-dir <isolated-dir> everywhere so the test never accidentally inspects the caller's actual opencode.json

Tests

421 tests across 12 unit suites + 13 E2E checks (was 419 / 12 in v0.13.2; +2 for T78/T78b no-clobber bypass)
E2E suite is npm run test:e2e, separate from npm test

Compat

Behavior change is bug-fix only: --print-only was failing where it should have succeeded; now succeeds. No path that used to succeed now fails. No flag/provider/tier additions.
19 of 19 v0.13.2 default-model entries unchanged.

[0.13.2] - 2026-05-24

Added — Two new OAuth subscription providers + Z.AI default bump

Driven by a fresh community + OpenCode-changelog sweep (2026-05-24). OpenCode shipped v1.15.5 → v1.15.10 in the week since v0.13.1: v1.15.7 restored native OpenAI OAuth + added native xAI OAuth. The wizard's subscription tier now reflects both.

# ChatGPT Plus/Pro sub:
npx opencode-sdlc-wizard pick --tier subscription --provider chatgpt

# SuperGrok sub:
npx opencode-sdlc-wizard pick --tier subscription --provider grok

# Existing Copilot Pro+ still works:
npx opencode-sdlc-wizard pick --tier subscription --provider copilot

New provider entries

Tier / Provider	Default model	OAuth path
`subscription / openai-codex` (a.k.a. `chatgpt`, `chatgpt-plus`, `chatgpt-pro`)	`gpt-5.3-codex`	`opencode /connect` → browser OpenAI login
`subscription / grok` (a.k.a. `xai`, `supergrok`, `super-grok`)	`grok-4.3`	`opencode /connect` → browser xAI login OR device-code

Both follow the same shape as v0.13.0's GitHub Copilot integration: model pin only, empty provider: {} block because OpenCode's native adapter handles the OAuth flow + token persistence. Wizard scaffolds the model choice; user runs opencode /connect once to authenticate.

Changed — Z.AI default model bump

proprietary/zai default: glm-4.6 → glm-5.1 per Z.AI's own docs (docs.z.ai/llms.txt: "Use GLM models like GLM-5.1 & GLM-5-Turbo for AI coding"). Trusting official provider guidance over the May-24 research's glm-4.7 suggestion (community-frequency claim) — provider docs are the authoritative source for default-model picks.

PRIVACY.md walkthrough example bumped to match.

Verified live (OpenCode docs + provider docs, fetched 2026-05-24)

OpenCode v1.15.7 changelog: native OpenAI OAuth restored; xAI OAuth added
OpenCode providers doc: canonical IDs are openai (same id for both API-key and OAuth path) and xai
xAI docs.x.ai/docs/models: grok-4.3 is flagship coding model ("most intelligent and fastest")
Z.AI docs.z.ai/llms.txt: GLM-5.1 listed + recommended for AI coding

Wizard tier-provider distinction note

subscription/openai-codex and proprietary/openai both write model: "openai/<m>" but emit different provider blocks:

Tier	Provider block	Auth
`proprietary/openai`	`provider.openai.options.apiKey = {env:OPENAI_API_KEY}`	API key in env
`subscription/openai-codex`	`provider: {}` (empty)	OAuth via `/connect`

The distinction is a wizard-level UX abstraction. OpenCode sees the same model pin either way — what differs is whether opencode.json carries an API key reference or defers entirely to OpenCode's OAuth state. This matches how Copilot's github-copilot provider works.

Changed

scripts/configure-backend.sh:
- 8 new PROVIDER_ALIASES entries (4 for ChatGPT family, 4 for Grok family)
- 2 new fragmentFor cases: subscription/openai, subscription/xai (both empty provider block)
- Header comment updated
scripts/pick-backend.sh:
- 2 new default_model_for() entries (+ aliases): subscription/openai* → gpt-5.3-codex, subscription/xai|grok* → grok-4.3
- Z.AI default updated to glm-5.1
scripts/detect-backends.sh:
- subscription JSON block grew from 1 entry to 3 (added openai, xai)
- Both new entries carry OAuth setup hints
AGENTS.md: subscription row in tier table lists all three providers
PRIVACY.md: subscription walkthrough table grew to 3 rows; configure examples for all three OAuth flows; Z.AI walkthrough bumped to glm-5.1

Tests

tests/test-backend-picker.sh adds T73–T77:
- T73: subscription/openai-codex writes openai pin + empty provider
- T74: subscription/grok writes xai pin + empty provider
- T75: ChatGPT alias sweep (4 aliases tested)
- T76: Grok alias sweep (3 aliases tested)
- T77: detector JSON exposes all three OAuth subscription providers
tests/test-pick.sh T12 default-model drift gate gains subscription/openai-codex:gpt and subscription/grok:grok entries
419 tests across 12 suites (was 407 / 12 in v0.13.1)

Positioning note (no code change)

OpenCode v1.15.6 added a native subagent picker to opencode run (interactive runtime agent selection). The wizard's pick subcommand is a write-time backend configurator — different scope, different phase. They don't conflict but the naming overlap is worth knowing.

Compat

16 of 17 v0.13.1 default-model entries unchanged (only Z.AI bumped).
glm-4.6 still works if passed explicitly via --model glm-4.6.
Cascade recommendations unchanged — subscription tier remains opt-in.

[0.13.1] - 2026-05-18

Changed — PRIVACY.md tier walkthroughs updated for the six-tier reality

Doc-only sweep. PRIVACY.md opened with "four backend tiers" since v0.2.0; v0.12.0 added managed (OpenCode Zen) and v0.13.0 added subscription (GitHub Copilot Pro+), but the tier walkthroughs hadn't caught up. v0.13.1 closes the doc drift.

PRIVACY.md overview table bumped from 4 rows to 6 (managed + subscription added with privacy-positioning notes).
New ## managed — OpenCode-routed PAYG (Zen) section — full walkthrough with OPENCODE_ZEN_API_KEY setup, default model (gpt-5.5), free-tier model example (deepseek-v4-flash-free), privacy positioning (between hosted_oss and proprietary).
New ## subscription — OAuth-managed sub bridge (Copilot Pro+) section — full walkthrough including:
- Critical shape difference: auth is OAuth, not env-var/JSON-key
- Configure scaffolds model pin only; opencode /connect flow completes OAuth
- Privacy positioning (Pro+ ToS specifically excludes training on prompts; verify at the plan page)
Z.AI GLM Coding Plan walkthrough added to proprietary section (was added as provider in v0.11.0 but missed in the walkthrough doc).

ROADMAP.md — v0.10.1 → v0.13.1 sprint declared shipped

The v0.10.1+ candidates list (last updated in v0.10.0's ROADMAP entry) described an open queue. The reality: 5 of 6 highest-signal community patterns from May-17 research were shipped across v0.10.1–v0.13.1. ROADMAP now reflects this — wizard is feature-complete relative to May-2026 community signals.

ROADMAP audit:

✅ Per-agent permission sandboxing (v0.10.1 + v0.10.5)
✅ Planner / docs / test-writer flags (v0.10.2 + v0.10.5)
✅ Copilot Pro+ as first-class provider (v0.13.0)
✅ Z.AI GLM (v0.11.0)
✅ OpenCode Zen managed tier (v0.12.0)
✅ Default-model freshness (v0.10.2 + v0.10.3)
✅ small_model / per-agent temps / security agent (v0.10.4 / v0.11.1 / v0.11.2)

Remaining v0.13.2+ backlog items have no community-research backing — moved to a "post-sprint backlog" section flagged as speculative until dogfood feedback or a fresh research pass surfaces signal.

Tests

No code changes; 407 tests across 12 suites unchanged from v0.13.0.

Compat

Pure doc patch. Every flag, provider, tier, default-model unchanged.

[0.13.0] - 2026-05-18

Added — `subscription` tier with GitHub Copilot Pro+

Per the original May-2026 research, Copilot Pro+ ($39/mo) is the ONLY subscription path that bridges Opus 4.7 + GPT-5.3-Codex into OpenCode after Anthropic killed third-party OAuth in Jan/Feb 2026. v0.13.0 makes it a first-class tier.

npx opencode-sdlc-wizard pick --tier subscription --provider copilot
# or:
npx opencode-sdlc-wizard pick --tier subscription --provider github-copilot
# then complete OAuth:
opencode    # → /connect → github.com/login/device

Yields:

{
  "model": "github-copilot/claude-opus-4-7",
  "provider": {}
}

OpenCode's native github-copilot adapter handles the OAuth flow — no API key in opencode.json. Auth is completed once via the interactive /connect flow inside OpenCode and persisted to user state, not the project config.

Verified live (OpenCode providers docs, fetched 2026-05-18)

Canonical provider id: github-copilot
Auth: OAuth device flow (/connect → github.com/login/device)
No env var — wizard's detector can't auto-trigger this tier (user must pass --tier subscription --provider copilot explicitly)
Default model: claude-opus-4-7 (Pro+ unlocks Opus + GPT-5.3-Codex)
Aliases: copilot, github_copilot, gh-copilot, gh_copilot → github-copilot

New 6th tier — `subscription`

Tier	Auth
`private_local`	none (local runtime)
`enterprise`	tenant credentials (Azure / Bedrock)
`hosted_oss`	API key (per-provider env var)
`managed`	API key (`OPENCODE_ZEN_API_KEY`)
`proprietary`	API key (per-vendor env var)
`subscription`	OAuth (no env var)

subscription is the first tier where auth doesn't live in env vars or opencode.json. Wizard scaffolds the model pin; OpenCode handles the rest.

Changed

scripts/configure-backend.sh:
- 5 new PROVIDER_ALIASES entries (copilot / github_copilot / github-copilot / gh_copilot / gh-copilot → github-copilot)
- fragmentFor new subscription/github-copilot case — model pin only, empty provider: {} block (OAuth-managed native adapter)
scripts/pick-backend.sh:
- default_model_for() new subscription/github-copilot and 4 aliases → claude-opus-4-7 entry
scripts/detect-backends.sh:
- New subscription.github-copilot JSON output entry with auth: "oauth" + setup: "opencode /connect → github.com/login/device" hints. key_set: false always (no env var to probe).
- Cascades unchanged — subscription tier requires explicit opt-in.
AGENTS.md: tier table gains a subscription row.

Tests

tests/test-backend-picker.sh adds T70–T72:
- T70: configure-backend writes Copilot pin with no apiKey/baseURL (provider block is empty per OAuth-managed shape)
- T71: alias resolution sweep (4 aliases tested in a loop)
- T72: detector emits subscription.github-copilot block with OAuth setup hints (no env var probe; documents the manual setup path)
tests/test-pick.sh T12 default-model drift gate gains subscription/github-copilot:claude entry
407 tests across 12 suites (was 400 / 12 in v0.12.0)

Why minor bump (v0.12.x → v0.13.0)

Sixth tier in the detector JSON output. New auth shape (OAuth instead of env var) — the first time provider: {} is the correct block for a provider. User-visible enough to warrant the minor.

Compat

17 of 17 v0.12.0 default-model entries unchanged.
Cascade recommendations unchanged — subscription tier is opt-in only.
provider: {} in the emitted opencode.json is intentional — deepMerge treats it as a no-op against existing provider blocks, so a user's prior provider.anthropic (etc.) survives a pick into the subscription tier.

[0.12.0] - 2026-05-18

Added — `managed` tier with OpenCode Zen as 5th privacy tier

The first new tier since v0.2.0. OpenCode Zen is OpenCode's own vendor-routed PAYG service — 40+ models including a free tier (Big Pickle, DeepSeek V4 Flash Free, MiniMax M2.5 Free, Nemotron 3 Super Free), $5 auto-reload trigger. Per May-2026 research, Zen is the "official recommended new-user entry point" — the wizard now reflects this in both the detector cascade and the tier table.

export OPENCODE_ZEN_API_KEY="..."
npx opencode-sdlc-wizard pick                # auto-detected
npx opencode-sdlc-wizard pick --tier managed --provider opencode
npx opencode-sdlc-wizard pick --tier managed --provider zen          # alias

Yields:

{
  "model": "opencode/gpt-5.5",
  "provider": {
    "opencode": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "apiKey": "{env:OPENCODE_ZEN_API_KEY}",
        "baseURL": "https://opencode.ai/zen/v1"
      },
      "models": { "gpt-5.5": {} }
    }
  }
}

Verified live (Zen docs, fetched 2026-05-18)

Canonical provider id: opencode (per Zen's docs — model pin format opencode/<model>)
baseURL: https://opencode.ai/zen/v1 (OpenAI-compatible)
Auth: {env:OPENCODE_ZEN_API_KEY} — namespaced to avoid collision with anything else "opencode"
Default model: gpt-5.5 (Zen's own documented example string)
Pricing: PAYG per 1M tokens; free-tier models available; $5 reload trigger / default $20 reload

New 5th tier — `managed`

Tier	Privacy semantics
`private_local`	Your machine, zero egress
`enterprise`	Your tenant (Azure, Bedrock)
`hosted_oss`	Open weights, third-party host (you pick the upstream)
`managed`	OpenCode-routed PAYG; vendor manages routing
`proprietary`	Vendor-bound (you bring the API key, vendor sees prompts)

Position is intentional: prompts go through OpenCode's infra (less private than DIY hosted, more managed than your own vendor key). Sits between hosted_oss and proprietary in the privacy-first cascade.

Changed

scripts/detect-backends.sh:
- New M_OPENCODE_ZEN_SET="$(env_set OPENCODE_ZEN_API_KEY)" probe
- New managed.opencode.{key_set, env} entry in JSON output
- Privacy-first cascade: managed placed between hosted_oss and proprietary
- Free-tier-first cascade: managed placed between proprietary/zai and proprietary/anthropic (Zen has free models so beats paid Anthropic on cost)
scripts/configure-backend.sh:
- PROVIDER_ALIASES (4 new entries): opencode → opencode, opencode_zen → opencode, opencode-zen → opencode, zen → opencode
- fragmentFor new managed/opencode case with the verified config
scripts/pick-backend.sh:
- default_model_for() new managed/opencode|opencode_zen|opencode-zen|zen → gpt-5.5 entry
AGENTS.md: tier table gains a managed row

Tests

tests/test-backend-picker.sh adds T66–T69:
- T66: configure-backend writes correct Zen provider block (id, baseURL, apiKey, npm adapter, models entry)
- T67: alias resolution (zen → opencode)
- T68: detector picks up OPENCODE_ZEN_API_KEY and recommends managed/opencode (fake HOME + stripped PATH)
- T69: managed/opencode composes with v0.10.x/v0.11.x agent flags (reviewer + sandbox + temp), same-provider coder+reviewer collapses to single provider block
tests/test-pick.sh T12 default-model drift gate gains managed/opencode:gpt-5.5 entry
400 tests across 12 suites (was 395 / 12 in v0.11.2)

Why minor bump (v0.11.x → v0.12.0)

New top-level tier in the JSON output. Detector schema gained a new top-level key (managed). User-visible enough to warrant the minor. Purely additive — every v0.11.x flag and provider continues to work unchanged.

Compat

16 of 16 v0.11.2 default-model entries unchanged.
Every existing tier + provider continues to work identically.
Every v0.10.x/v0.11.x flag works on the new managed/opencode tier+provider exactly as it does on the other tiers (verified by T69's full hybrid test).

[0.11.2] - 2026-05-18

Added — Security agent (full set: model + temperature + tools-denial sandbox)

The joelhooks community config dedicates a security agent to "is this code safe" reviews with the same write/edit/patch denial as plan-mode. v0.11.2 ships the full surface for it: model triplet (mirrors v0.10.2 planner), temperature (mirrors v0.11.1), and tool-denial sandbox (mirrors v0.10.5 plan).

npx opencode-sdlc-wizard pick \
  --tier proprietary --provider anthropic \
  --security-tier proprietary --security-provider openai --security-model gpt-5.3-codex \
  --security-temp 0.1 \
  --sandbox-security

Yields:

{
  "model": "anthropic/claude-opus-4-7",
  "provider": {
    "anthropic": { ... },
    "openai":    { ... }
  },
  "agent": {
    "security": {
      "model": "openai/gpt-5.3-codex",
      "temperature": 0.1,
      "tools": { "write": false, "edit": false, "patch": false }
    }
  }
}

OpenCode routes security-class agent tasks to the dedicated model, with deterministic temperature and zero write capability — security reviews can never apply a patch by accident.

New flags

--security-tier T --security-provider P [--security-model M] — triplet (mirrors --reviewer-* / --planner-*); writes agent.security.model + security provider block
--security-temp T — writes agent.security.temperature (mirrors --coder-temp / --planner-temp / --reviewer-temp)
--sandbox-security — writes agent.security.tools = {write/edit/patch: false} (mirrors --sandbox-plan)

All flags opt-in, all-or-nothing on the triplet, default-model fallback via the same default_model_for() shared with coder/reviewer/planner/small.

Changed

scripts/configure-backend.sh: 5 new flag entries, new securityMode block in the node heredoc, new agent.security cases in the sandbox + temperature emission blocks
scripts/pick-backend.sh: new flags, validate_agent_triplet "security" call, default_model_for() fallback, passthrough block

Tests

tests/test-backend-picker.sh adds T61–T65 (model writes + tools sandbox + triple-compose + partial-spec rejection + regression guard)
tests/test-pick.sh adds T39–T40 (full passthrough + regression guard)
395 tests across 12 suites (was 388 / 12 in v0.11.1)

Compat

Opt-in only. v0.10.x / v0.11.0 / v0.11.1 configs unchanged.
Composes with every prior flag.
The full v0.11.2 hybrid in one call: coder + small_model + planner
- reviewer + security agents, each with their own model, temperature, and (where applicable) sandbox. Five agents total covered.

[0.11.1] - 2026-05-18

Added — Per-agent temperatures (`--coder-temp`, `--planner-temp`, `--reviewer-temp`)

May-2026 community pattern (joelhooks, ppries, others): explicit temperature settings per agent. Typical split: plan=0.1 (deterministic, focus on the right answer), build=0.3 (some creativity for code), review=0.1 (deterministic).

npx opencode-sdlc-wizard pick \
  --tier proprietary --provider anthropic \
  --planner-tier hosted_oss --planner-provider groq --planner-temp 0.1 \
  --reviewer-tier hosted_oss --reviewer-provider cerebras --reviewer-temp 0.1 \
  --coder-temp 0.3

Yields:

{
  "model": "anthropic/claude-opus-4-7",
  "agent": {
    "build": { "temperature": 0.3 },
    "plan":  { "model": "groq/gpt-oss-120b",      "temperature": 0.1 },
    "review":{ "model": "cerebras/gpt-oss-120b",  "temperature": 0.1 }
  }
}

Naming notes

--coder-temp targets agent.build.temperature — OpenCode's default agent name for code generation is build, not coder. We expose --coder-temp for ergonomics (matches --coder-* mental model from v0.10.x) while writing the canonical JSON key.
--planner-temp targets agent.plan.temperature (mirrors --planner-tier/--planner-provider from v0.10.2).
--reviewer-temp targets agent.review.temperature (mirrors --reviewer-tier/--reviewer-provider from v0.10.0).

Changed — `scripts/configure-backend.sh`

New --coder-temp T, --planner-temp T, --reviewer-temp T flags
Values parsed as JSON numbers (so 0.3 becomes 0.3 not "0.3")
Empty string = "not set" — no temperature field emitted for that agent
Deep-merges with v0.10.0 agent.review.model, v0.10.2 agent.plan.model, v0.10.5 agent.plan.tools, v0.10.1 agent.<name>.permission.write blocks

Changed — `scripts/pick-backend.sh`

Same three flags forwarded to configure-backend.sh unchanged
No default-model fallback (temperatures aren't tied to providers)

Tests

tests/test-backend-picker.sh adds T56–T60:
- T56: --coder-temp 0.3 writes agent.build.temperature = 0.3 as a number
- T57: --reviewer-temp composes with --reviewer-* (model + temperature on agent.review)
- T58: --planner-temp + --planner-* + --sandbox-plan triple-compose (model + temperature + tools on agent.plan)
- T59: all three temp flags emit correct values
- T60: no flags → no temperature fields (opt-in regression guard)
tests/test-pick.sh adds T35–T38 (passthrough + regression guard)
388 tests across 12 suites (was 379 / 12 in v0.11.0)

Compat

Opt-in only. Existing configs unchanged unless --*-temp flags passed.
Composes with every v0.10.x and v0.11.0 flag. Full v0.11.1 stack in one call: model + small_model + 3 per-agent models + 3 sandboxes + 3 per-agent temperatures.

[0.11.0] - 2026-05-18

Added — Z.AI GLM Coding Plan as a first-class proprietary provider

Per the original May-2026 community-patterns research, Z.AI's GLM Coding Plan is the most-cited post-Anthropic-OAuth-ban migration target (Anthropic killed third-party OAuth Jan/Feb 2026; OpenCode removed the OAuth code March 2026). v0.11.0 makes Z.AI a first-class provider across the detector, configurator, picker, and docs.

export ZAI_API_KEY="..."
npx opencode-sdlc-wizard pick                # auto-detected
# or explicitly:
npx opencode-sdlc-wizard pick --tier proprietary --provider zai
npx opencode-sdlc-wizard pick --tier proprietary --provider glm  # alias

Yields:

{
  "model": "zai/glm-4.6",
  "provider": {
    "zai": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "apiKey": "{env:ZAI_API_KEY}",
        "baseURL": "https://api.z.ai/api/paas/v4"
      },
      "models": { "glm-4.6": {} }
    }
  }
}

Verified live (Z.AI docs, fetched 2026-05-18)

baseURL: https://api.z.ai/api/paas/v4 — OpenAI-compatible
Auth: Authorization: Bearer <ZAI_API_KEY>
Default model: glm-4.6 — most-documented stable; other catalog options include GLM-5.1 / 5-Turbo / 5 / 4.7 / 4.5
Coding Plan pricing: $10/mo (or $30/quarter, $80/year — quarterly restructure May 2026; the previously-cited flat $18/mo SKU is retired)

Changed

scripts/configure-backend.sh:
- PROVIDER_ALIASES: added zai → zai, z.ai → zai, z_ai → zai, glm → zai (canonical id is zai)
- fragmentFor: new proprietary/zai case with the verified baseURL + @ai-sdk/openai-compatible adapter + models[<id>] entry
scripts/detect-backends.sh:
- New PR_ZAI_SET="$(env_set ZAI_API_KEY)" probe
- JSON output: added proprietary.zai.{key_set, env: "ZAI_API_KEY"}
- Privacy-first cascade: Z.AI added at the end of proprietary tier (Anthropic → OpenAI → Google → Z.AI)
- Free-tier-first cascade: Z.AI ranked above Anthropic / OpenAI in the proprietary section because Coding Plan flat-fee is cheaper than pay-per-token for heavy users (community signal)
scripts/pick-backend.sh:
- default_model_for(): new proprietary/zai|z.ai|z_ai|glm → glm-4.6 entry
AGENTS.md: privacy-tier table mentions Z.AI in the proprietary row with the migration-target context

Tests

tests/test-backend-picker.sh adds T53–T55:
- T53: configure-backend writes correct Z.AI provider block (apiKey env ref, baseURL, npm adapter, models entry)
- T54: alias resolution (glm → zai) parity with other proprietary providers (google_aistudio → google, etc.)
- T55: detector picks up ZAI_API_KEY and recommends proprietary/zai (uses fake HOME + stripped PATH so host's local runtimes don't win the privacy-first cascade)
tests/test-pick.sh T12 default-model drift gate gains proprietary/zai:glm entry (substring match /glm/ against glm-4.6 default)
379 tests across 12 suites (was 375 / 12 in v0.10.6)

Why this is the minor-version bump (v0.10.x → v0.11.0)

New provider entry adds user-visible surface (new tier+provider combo, new env var, new alias group, new default-model map entry, new detector field, new docs row). All additive — every v0.10.x flag and provider continues to work unchanged — but the net "what does this wizard support" surface grew enough that a minor bump is appropriate.

Compat

Every v0.10.x default-model entry unchanged (Z.AI is a new entry, not a swap of an existing one).
Every v0.10.x flag (--reviewer-*, --planner-*, --small-*, --sandbox-*) works with --tier proprietary --provider zai exactly as it does with the other proprietary providers.
Z.AI alias group (zai, z.ai, z_ai, glm) all resolve to the canonical zai id in written opencode.json blocks.

[0.10.6] - 2026-05-18

Changed — `docs/cost-ladder.md` recalibration sweep

Doc-only refresh aligning the cost-ladder's example model IDs and pricing notes with the v0.9.1 – v0.10.5 picker default changes plus May-17 research findings.

$20/mo path:

gpt-5.5 → gpt-5.3-codex (verified shipped Feb 2026; most-pinned reviewer in surveyed configs)
deepseek-chat → deepseek-v4-flash (V4 family shipped April 2026)
CI gate slot: removed "Groq still ships Llama 3.3 70B; Cerebras dropped it" qualifier — both providers now host gpt-oss-120b (the v0.10.3 picker default)
Added Z.AI GLM Coding Plan as a third Coder (alt) option with current quarterly pricing: $10/mo or $30/quarter or $80/year. The previously cited flat $18/mo SKU was retired in the May 2026 pricing restructure (research finding).

$200/mo path:

gpt-5.5 xhigh → gpt-5.3-codex xhigh (both Coder alt + Reviewer slots)
CI gate slot: same cleanup as $20/mo path

Footer:

Calibration date bumped to 2026-05-18.
Prior calibration (2026-05-05 for v0.8.1) preserved for audit trail.

Not changed (intentionally)

claude-sonnet-4.6 ($20/mo Coder) and claude-opus-4.7 ($200/mo Coder) unchanged — Sonnet 4.6 / Opus 4.7 are still current Anthropic frontier (Sonnet 4.7 / Opus 4.8 not yet shipped per May-17 research).
gemini-2.5-flash in the $0/mo Coder slot left as-is — Gemini 3.1 Pro (the v0.10.2 picker default) is the paid Pro tier; 2.5-flash is still the free-tier model that fits a $0/mo budget.
Calibration history block (model ID drift across v0.8.x) left intact as audit trail.
No code or test changes. All 375 tests still pass unchanged.

Why this matters

docs/cost-ladder.md is the public-facing budget guide users hit when planning their OpenCode setup. Drift here costs trust faster than drift in any code surface — the v0.10.x picker now ships canonical defaults that didn't exist when this doc was first calibrated.

[0.10.5] - 2026-05-18

Added — `--sandbox-plan` (plan-mode tool denial)

May-17 research: 33% of community configs set agent.plan.tools to deny write/edit/patch so the planner can read + reason but can't modify code. Categorically stronger than path-scoped permission.write blocks (v0.10.1) — the tool itself isn't available.

# Plan model can analyze, propose, brainstorm — but cannot touch files
npx opencode-sdlc-wizard pick \
  --tier proprietary --provider anthropic \
  --planner-tier hosted_oss --planner-provider groq \
  --sandbox-plan

Yields:

{
  "model": "anthropic/claude-opus-4-7",
  "agent": {
    "plan": {
      "model": "groq/gpt-oss-120b",
      "tools": { "write": false, "edit": false, "patch": false }
    }
  }
}

agent.plan.model (v0.10.2) and agent.plan.tools (v0.10.5) deep-merge into the same plan object.

Changed

scripts/configure-backend.sh: new --sandbox-plan boolean flag. Reuses the existing sandbox emission block — plan: { tools: ... } is the third sandbox shape alongside test-writer and docs.
scripts/pick-backend.sh: --sandbox-plan passthrough.

Tests

tests/test-backend-picker.sh adds T49–T52:
- T49: --sandbox-plan writes agent.plan.tools.{write,edit,patch} = false
- T50: --sandbox-plan + --planner-* compose (both .model and .tools present)
- T51: all three sandbox flags (plan/test-writer/docs) compose
- T52: regression guard — no agent.plan.tools without the flag
tests/test-pick.sh adds T33–T34 (passthrough + regression guard)
375 tests across 12 suites (was 369 / 12 in v0.10.4)

Compat

Three sandbox patterns now available: --sandbox-test-writer (path), --sandbox-docs (path), --sandbox-plan (categorical tool denial).
All opt-in. v0.9.x and v0.10.x configs unchanged unless flags passed.
Full v0.10.x stack: pick --model M --small-* --planner-* --reviewer-* --sandbox-test-writer --sandbox-docs --sandbox-plan produces the canonical 4-agent SDLC config in one call.

[0.10.4] - 2026-05-18

Added — `small_model` top-level pin (`--small-*` flags)

May-17 community-patterns research found 35–40% of surveyed opencode.json files set the top-level small_model field — OpenCode's cross-cutting hint for "use this when the call is cheap" (title generation, summary blurbs, anywhere a small/fast model suffices over the global model). Distinct from agent.plan.model (which only affects plan-mode tasks); small_model is consulted by any agent.

# Set Opus as the build model, Haiku as the cheap fallback for
# title/summary work
npx opencode-sdlc-wizard pick \
  --tier proprietary --provider anthropic \
  --small-tier proprietary --small-provider anthropic --small-model claude-haiku-4-5

Yields:

{
  "model": "anthropic/claude-opus-4-7",
  "small_model": "anthropic/claude-haiku-4-5",
  "provider": { "anthropic": { ... } }
}

--small-* composes cleanly with v0.10.0 reviewer, v0.10.1 sandboxes, v0.10.2 planner. Full v0.10.x stack in one call works.

Changed — `scripts/configure-backend.sh`

New --small-tier T --small-provider P --small-model M flags (all-or-nothing triplet; partial spec exits 2)
small_model value is "<canonical_provider>/<model>" using the same PROVIDER_ALIASES map
Small-side provider block deep-merges; same-provider as coder/reviewer/ planner collapses to one block
Canonical key order updated: top-level is now $schema, model, small_model, provider, then rest alphabetical (matches the joelhooks + ppries community configs we surveyed — keeps the two top-level pins visually adjacent)

Changed — `scripts/pick-backend.sh`

New --small-tier T --small-provider P [--small-model M] flags
--small-model optional; filled from default_model_for() (same source of truth used for coder / reviewer / planner)
Partial-spec validation via the existing validate_agent_triplet() helper — no new error-handling code path

Tests

tests/test-backend-picker.sh adds T44–T48
tests/test-pick.sh adds T28–T32
369 tests across 12 suites (was 357 / 12 in v0.10.3)

Compat

14 of 14 v0.10.3 default-model entries unchanged.
Opt-in only. Existing v0.9.x / v0.10.x configs unchanged unless --small-* flags are passed.

[0.10.3] - 2026-05-18

Changed — two more research-verified default-model bumps

May-17 community-patterns research verified both swaps against live provider catalogs; v0.10.3 lands them as a focused patch.

Provider	Was (v0.10.2)	Now (v0.10.3)	Why
`hosted_oss/deepseek`	`deepseek-chat`	`deepseek-v4-flash`	DeepSeek V4 family shipped 2026-04-24; `deepseek-chat` is a moving alias that already points at V4, but explicit pin is clearer for cost/latency-sensitive picks
`hosted_oss/groq`	`llama-3.3-70b-versatile`	`gpt-oss-120b`	Groq's own docs flag `gpt-oss-120b` as "Most Popular for OpenCode"; OpenAI open-weights release on Groq's LPU stack; matches Cerebras default for cross-provider consistency

PRIVACY.md "Suggested model" table and DeepSeek configure example bumped to match.

Tests

test-pick.sh T12 substring-match for groq updated /llama/ → /gpt-oss/
test-pick.sh T23 planner-default assertion for groq bumped to gpt-oss-120b
DeepSeek tests unchanged — /deepseek/ substring still matches deepseek-v4-flash
All other suites unchanged (configure-backend tests pin explicit models, not defaults)
357 tests across 12 suites (no count change; T23 / T12 patched in place)

Compat

12 of 14 default-model entries unchanged from v0.10.2.
Anyone passing --model deepseek-chat explicitly still works (deepseek-chat is still a valid id on DeepSeek's catalog; v0.10.3 just changes what pick chooses by default).
Anyone passing --model llama-3.3-70b-versatile explicitly still works (Groq still hosts it; we just default to gpt-oss-120b now per community signal).

[0.10.2] - 2026-05-17

Added — Planner agent model routing (`--planner-*` flags)

Fresh May-17-2026 community-patterns research found plan is the most-overridden agent at 57% of surveyed configs — beats review (52%) and docs (~40%). v0.10.2 mirrors the v0.10.0 reviewer pattern for the planner agent: --planner-tier T --planner-provider P [--planner-model M] on both pick and configure-backend.sh, with --planner-model optional (filled from canonical default-model map, same default_model_for() function used for coder + reviewer pins).

# Full v0.10.x hybrid — planner=fast, build=mid, reviewer=high,
# plus the v0.10.1 sandbox blocks
npx opencode-sdlc-wizard pick \
  --tier proprietary --provider anthropic \
  --planner-tier hosted_oss --planner-provider cerebras \
  --reviewer-tier hosted_oss --reviewer-provider nvidia_nim \
  --sandbox-test-writer --sandbox-docs

Yields agent.plan.model: cerebras/gpt-oss-120b + agent.review.model: nvidia/deepseek-ai/deepseek-r1 + the sandbox permission.write blocks. OpenCode routes plan-mode tasks to Cerebras (fast), build to Opus (the global), review to NIM DeepSeek-R1 (reasoning), with the test-writer and docs agents path-scoped.

Changed — default model bumps (May-17 research, both verified live)

Provider	Was (v0.10.1)	Now (v0.10.2)	Verified
`proprietary/openai`	`gpt-5`	`gpt-5.3-codex`	OpenAI catalog, released 2026-02-05; most-pinned reviewer in fresh community configs
`proprietary/google_aistudio`	`gemini-2.5-flash`	`gemini-3.1-pro`	Google catalog, released 2026-02-19; current frontier Gemini
`proprietary/anthropic`	`claude-opus-4-7`	unchanged	Opus 4.7 still current frontier; Sonnet 4.8 not yet shipped

Doc surfaces updated to match: PRIVACY.md Google AI Studio walkthrough example now references gemini-3.1-pro.

Other community-flagged swaps (DeepSeek → deepseek-v4-flash, Groq → gpt-oss-120b) deferred — both verified live but each one is its own default-value decision worth a focused patch.

Changed — `scripts/configure-backend.sh`

New --planner-tier T --planner-provider P --planner-model M flags (all-or-nothing triplet; partial spec exits 2 with actionable msg)
Planner side shares PROVIDER_ALIASES map with coder + reviewer
Planner provider block deep-merges; same-provider planner+coder collapses to single provider block
agent.plan.model deep-merges into any existing agent block alongside agent.review.model (v0.10.0) and agent.test-writer/agent.docs permission blocks (v0.10.1)

Changed — `scripts/pick-backend.sh`

New --planner-tier T --planner-provider P [--planner-model M] flags
Default-model resolution shares default_model_for() with the coder + reviewer paths
Partial-spec validation extracted into a validate_agent_triplet() helper so reviewer + planner share the same validation logic (DRY for the next agent we add)

Tests

tests/test-backend-picker.sh adds T39–T43:
- T39: --planner-* writes agent.plan.model + planner provider block
- T40: planner + reviewer compose (both agent blocks, dual providers)
- T41: planner alias resolution (google_aistudio → google) parity with reviewer
- T42: partial --planner-* rejected without writing opencode.json
- T43: no planner flags → no agent.plan block (opt-in regression guard)
tests/test-pick.sh adds T23–T27 (passthrough + default resolution + partial-spec error + full hybrid composition + regression guard)
357 tests across 12 suites (was 349 / 12 in v0.10.1)

Notes on the May-17 research

OpenCode shipped 11 patch releases in the past 10 days (1.14.42 → 1.15.4) including a v2 model/provider listing API and DigitalOcean OAuth + Inference Router native adapter (@ai-sdk/openai-compatible shim no longer needed if we ever support it).
v1.15.1 surfaces full config validation errors at TUI startup — bad JSON the wizard emits now fails loudly instead of silently. UX win.
v1.14.46 ships a built-in customize-opencode skill — light overlap with our trivial-edit layer; we differentiate via SDLC discipline (hooks, scoring, plan→TDD→review enforcement).
Wizard at ~523 downloads/month, 410/week — early but real.

[0.10.1] - 2026-05-17

Added — Per-agent permission sandboxing (`--sandbox-test-writer` + `--sandbox-docs`)

May-2026 community-patterns research: 9/15 surveyed opencode.json configurations use agent.<name>.permission.write to scope what each agent can touch. Most-cited patterns:

test-writer locked to test/spec files only
docs locked to .md only

v0.10.1 exposes these as boolean flags on both pick and configure-backend.sh. Users wanting custom glob patterns still edit opencode.json directly.

# Full v0.10.x hybrid in one shot — Mixed-Mode coder/reviewer split
# plus sandboxed test-writer and docs agents
npx opencode-sdlc-wizard pick \
  --tier proprietary --provider anthropic \
  --reviewer-tier hosted_oss --reviewer-provider cerebras \
  --sandbox-test-writer --sandbox-docs

Yields:

{
  "model": "anthropic/claude-opus-4-7",
  "provider": {
    "anthropic": { "options": { "apiKey": "{env:ANTHROPIC_API_KEY}" } },
    "cerebras":  { "npm": "@ai-sdk/openai-compatible", "options": { ... } }
  },
  "agent": {
    "review": { "model": "cerebras/gpt-oss-120b" },
    "test-writer": {
      "permission": { "write": {
        "**/*.test.*": "allow",
        "**/*.spec.*": "allow",
        "*": "deny"
      } }
    },
    "docs": {
      "permission": { "write": {
        "**/*.md": "allow",
        "*": "deny"
      } }
    }
  }
}

OpenCode enforces the permission patterns at write time — the test-writer agent can't accidentally clobber production source even if its model attempts to.

Changed

scripts/configure-backend.sh: new --sandbox-test-writer and --sandbox-docs boolean flags. The node heredoc deep-merges the canonical permission blocks alongside any v0.10.0 reviewer block, preserving user-set sibling fields.
scripts/pick-backend.sh: passes both flags through to the configurator unchanged.

Tests

tests/test-backend-picker.sh adds T35–T38:
- T35: --sandbox-test-writer writes canonical agent.test-writer.permission.write
- T36: --sandbox-docs writes canonical agent.docs.permission.write
- T37: both sandboxes compose with --reviewer-* (full v0.10.x hybrid)
- T38: no flags → no test-writer/docs agent blocks (opt-in regression guard)
tests/test-pick.sh adds T19–T22 (passthrough + composition + regression guard).
349 tests across 12 suites (was 341 / 12 in v0.10.0).

Compat

Opt-in only. Existing configs unchanged unless the new flags are passed.
Composes cleanly with v0.10.0 Mixed-Mode and v0.9.x single-model pick.

[0.10.0] - 2026-05-17

Added — Mixed-Mode (per-agent model routing) in `pick` + `configure-backend.sh`

May-2026 community-patterns research showed 11/15 surveyed opencode.json configurations route review work to a different model than build work (coder=mid, reviewer=high-reasoning). v0.10.0 makes this a one-flag-pair operation instead of hand-editing opencode.json.

# Pick coder + reviewer in one shot. pick fills canonical default
# models from its lookup table; --reviewer-model overrides.
npx opencode-sdlc-wizard pick \
  --tier proprietary --provider anthropic \
  --reviewer-tier hosted_oss --reviewer-provider cerebras

# Yields (--dry-run preview):
{
  "model": "anthropic/claude-opus-4-7",
  "provider": {
    "anthropic": { "options": { "apiKey": "{env:ANTHROPIC_API_KEY}" } },
    "cerebras":  { "npm": "@ai-sdk/openai-compatible", "options": {
      "apiKey": "{env:CEREBRAS_API_KEY}", "baseURL": "https://api.cerebras.ai/v1"
    }, "models": { "gpt-oss-120b": {} } }
  },
  "agent": { "review": { "model": "cerebras/gpt-oss-120b" } }
}

OpenCode now auto-routes any review-class agent to the reviewer model without needing the cross-model-review.sh wrapper script (that wrapper still exists for explicit "review NOW" invocations).

Changed — `scripts/configure-backend.sh`

New flags: --reviewer-tier T --reviewer-provider P --reviewer-model M (all three required together; partial spec exits 2 with actionable message).
Reviewer provider alias resolution shares the same PROVIDER_ALIASES map as the coder side (nvidia_nim → nvidia, google_aistudio → google, etc.) — both sides get canonicalized before being written.
Reviewer provider block merges into the same provider.<id> map as the coder; same-provider reviewer (e.g., --provider anthropic + --reviewer-provider anthropic with different model) collapses to a single provider block.
agent.review.model deep-merges into any existing agent block so user-set sibling fields (temperature, tools, permission) are preserved.

Changed — `scripts/pick-backend.sh`

New flags: --reviewer-tier T --reviewer-provider P [--reviewer-model M]. When --reviewer-model is omitted, pick fills it from the canonical default-model map (same lookup used for the coder pin).
Default-model lookup factored into a default_model_for(tier, provider) function so both the coder pin AND the reviewer pin share the same source of truth — no risk of the two sides drifting apart.
Validation: partial --reviewer-* spec rejected (exit 2). pick won't silently produce a single-mode config when the user clearly meant mixed-mode.

Tests

tests/test-backend-picker.sh adds T31–T34:
- T31: --reviewer-* writes agent.review.model + both provider blocks
- T32: single-mode regression guard — no agent block injected when reviewer flags absent
- T33: alias resolution (nvidia_nim → nvidia) applies to reviewer side
- T34: same-provider coder+reviewer collapses to single provider block
tests/test-pick.sh adds T15–T18:
- T15: --reviewer-tier T --reviewer-provider P fills reviewer-model default + forwards to configurator
- T16: --reviewer-model override beats default
- T17: partial --reviewer-* spec exits non-zero with actionable msg
- T18: single-mode regression guard — no --reviewer-* forwarded when not supplied

341 tests total across 12 suites (was 333 / 12 in v0.9.1).

Docs

docs/cost-ladder.md $20/mo path: hybrid coder+reviewer block now shows the v0.10.0 one-shot invocation alongside the low-level configure-backend.sh equivalent.

Compat

13 of 14 v0.9.x default-model entries unchanged.
cross-model-review.sh wrapper unchanged — still the "run a targeted review NOW" command. Mixed-Mode in opencode.json is the standing config; the wrapper is the explicit invocation.
Existing single-model pick and configure-backend.sh invocations work identically — Mixed-Mode is purely additive opt-in.

[0.9.1] - 2026-05-12

Changed — default local Ollama model bumped from Qwen 2.5 to Qwen 3

Community-patterns research (May 2026) surveyed 30+ shared opencode.json configurations and found Qwen3-Coder is the most-cited local default; Qwen 2.5 only appears in older configs from before Qwen3 shipped in Q4 2025.

scripts/pick-backend.sh default-model map: private_local/ollama: qwen2.5-coder:32b → qwen3-coder:30b
Verified tag qwen3-coder:30b exists on ollama.com/library/qwen3-coder/tags.
Doc surfaces updated to match: install.sh next-steps, PRIVACY.md (Ollama walkthrough + suggested-model table), README.md, AGENTS.md, docs/cost-ladder.md ($0/mo local-only path).
tests/test-pick.sh T4 updated to assert the new default.
T12 drift gate (per-provider /qwen/ substring match) continues to pass unchanged — both Qwen2.5 and Qwen3 satisfy it.

No other model defaults changed in this patch. The May-2026 research flagged additional candidates (Groq → gpt-oss-120b, OpenAI → gpt-5.2-codex, Gemini → gemini-3.1-pro, etc.) but each needs case-by-case verification against the provider's live catalog before shipping as a default. Deferred to v0.9.x as community-verified swaps land.

Not changed (intentionally)

13 of the 14 default-model map entries remain identical to v0.9.0.
VRAM annotation (16–24 GB VRAM) in PRIVACY.md still applies — Qwen3-Coder-30B at Q4 is ~17–18 GB, at Q5 ~21 GB.
All v0.9.0 features (pick subcommand, CLI dispatch, install delivery, T12 drift gate) unchanged.

[0.9.0] - 2026-05-12

Added — `pick` subcommand (one-shot detect → configure backend)

Codex's v0.9.0 direction call: first-run consumption ergonomics. The two-step picker workflow

bash .opencode/scripts/detect-backends.sh
bash .opencode/scripts/configure-backend.sh --tier ... --provider ... --model ...

collapses to

npx opencode-sdlc-wizard pick                       # detected highest-privacy
npx opencode-sdlc-wizard pick --free-tier-first     # detected free-first
npx opencode-sdlc-wizard pick --tier hosted_oss --provider cerebras   # override
npx opencode-sdlc-wizard pick --dry-run             # preview, don't write

pick runs detect-backends.sh, parses the recommendation field, resolves a canonical floor model for the detected tier/provider, and forwards everything to configure-backend.sh. Single source of truth for the default-model map per provider — shared between the auto-picker, docs, and skill examples.

Default-model map (every v0.8.x picker provider covered)

Tier / Provider	Default model
`private_local / ollama`	`qwen2.5-coder:32b`
`private_local / mlx`	`mlx-community/Qwen2.5-Coder-32B-Instruct-4bit`
`private_local / lm_studio`	`qwen2.5-coder-32b-instruct`
`private_local / llama_cpp`	`qwen2.5-coder-32b-instruct`
`private_local / vllm`	`Qwen/Qwen2.5-Coder-32B-Instruct`
`enterprise / azure_openai`	`gpt-5`
`enterprise / aws_bedrock`	`anthropic.claude-sonnet-4-5-20250929-v1:0`
`hosted_oss / together`	`Qwen/Qwen2.5-Coder-32B-Instruct`
`hosted_oss / groq`	`llama-3.3-70b-versatile`
`hosted_oss / openrouter`	`qwen/qwen-2.5-coder-32b-instruct`
`hosted_oss / cerebras`	`gpt-oss-120b`
`hosted_oss / deepseek`	`deepseek-chat`
`hosted_oss / nvidia_nim`	`deepseek-ai/deepseek-r1`
`proprietary / anthropic`	`claude-opus-4-7`
`proprietary / openai`	`gpt-5`
`proprietary / google_aistudio`	`gemini-2.5-flash`

Override with --model <name> when you want a non-floor pick.

Added — `scripts/pick-backend.sh`

The orchestrator backing the CLI subcommand. PATH-first script lookup so tests can shim either detect-backends.sh or configure-backend.sh without monkey-patching .opencode/scripts/. Falls back to sibling lookup (where pick lives next to detect+configure in production installs) when PATH doesn't have them.

Flags:

--free-tier-first — bias detector toward free providers
--tier <t> / --provider <p> — skip detection, force tier/provider
--model <m> — override the canonical default
--target-dir <p> / --force — pass through to configurator
--dry-run — forwards as --print-only (configurator prints merged JSON, doesn't write opencode.json)

Exit codes: 0 configured, 2 bad args, 3 detector returned none and no tier/provider override given, 4 no default model known for the resolved tier/provider (pass --model).

Tests

tests/test-pick.sh — 25 tests covering:
- Script presence, syntax, --help
- Detector → configure flow (private_local/ollama default)
- --free-tier-first sets DETECT_FREE_TIER_FIRST=1 in detector env
- --tier/--provider override (uses canonical default model)
- --model override beats canonical default
- Detector none recommendation → non-zero exit, no configure call
- --dry-run → --print-only translation
- --force / --target-dir pass-through
- T12 drift gate: every one of the 14 v0.8.x picker tier/provider combinations has a default-model entry
tests/test-bundle-drift.sh extended to assert scripts/pick-backend.sh is shipped by install.sh.

333 tests total across 12 suites (was 305 / 11 in v0.8.9).

Note on the v0.8.0 → v0.9.0 throughline

v0.8.0 shipped the four-tier picker. v0.8.4-v0.8.9 closed the picker's drift family across 5 doc + 1 runtime surface. v0.9.0 collapses the picker UX to a single command — the actual first-run friction that real consumption attempts surface. Codex's direction call ("ship ergonomics before mixed-mode") drove the prioritization; the Mixed-mode skill moves to v0.10.0 where pick --coder ... --reviewer ... becomes the natural extension.

[0.8.9] - 2026-05-08

Fixed — `cross-model-review.sh` wrapper alias parity with configure-backend.sh

Codex pre-ship review of the v0.8.x stack caught a real invocation bug introduced by v0.8.7's SKILL.md drift sweep. The skill's Step 2 table and Step 3 examples advertised nvidia_nim and google_aistudio as reviewer providers, but scripts/cross-model-review.sh's alias case block hadn't been extended. configure-backend.sh aliases nvidia_nim → nvidia and google_aistudio → google (writes the corresponding provider.nvidia / provider.google blocks in opencode.json). The wrapper's wildcard fallthrough was passing nvidia_nim and google_aistudio through verbatim, which built model pins (nvidia_nim/<model> / google_aistudio/<model>) that didn't match the registered provider blocks → silent provider-not-found from OpenCode at run time.

scripts/cross-model-review.sh: alias case block now mirrors configure-backend.sh's canonical mapping:
- nvidia_nim | nvidia-nim | nvidia → nvidia
- google_aistudio | google | gemini → google
Pass-through list also explicitly names the new v0.8.0 canonical IDs (cerebras, deepseek, mlx) — they were already correct via the wildcard, but documenting them stops a future maintainer from thinking they need an alias.

Tests

tests/test-cross-model-review.sh adds T11–T14:
- T11: nvidia_nim resolves to nvidia/<model> in the opencode pin
- T12: google_aistudio resolves to google/<model>
- T13: gemini resolves to google/<model>
- T14: canonical cerebras/deepseek/mlx pass through unchanged
305 tests total across 11 suites (was 301 in v0.8.8).

Note on the v0.8.0 drift family

This is the fifth surface in the v0.8.0 picker drift family (install.sh in v0.8.4, validator domain in v0.8.5, setup-wizard SKILL.md in v0.8.6, cross-model-review SKILL.md in v0.8.7, AGENTS.md+PRIVACY.md in v0.8.8) — and the only one that was an actual runtime bug rather than documentation drift. Caught by cross-model review before ship. The drift family is now closed across both documentation surfaces and runtime invocation.

[0.8.8] - 2026-05-06

Fixed — AGENTS.md + PRIVACY.md tier drift (final v0.8.0 sweep)

Last surface in the v0.8.0 provider drift family. The user-facing AGENTS.md tier table and the deeper PRIVACY.md tier walkthroughs both still listed only the v0.2.0 providers. Since AGENTS.md is what OpenCode auto-loads at session start, this was the most user-visible of the four drift surfaces.

AGENTS.md privacy-tier table updated:

private_local row gains MLX (Apple Silicon)
hosted_oss row gains Cerebras / DeepSeek direct / NVIDIA NIM
proprietary row gains Google AI Studio (google_aistudio)
New --free-tier-first example before the configure call
New cross-link to docs/cost-ladder.md

PRIVACY.md tier walkthroughs updated:

private_local runtime table gains MLX row with default URL + suggested model
hosted_oss provider table gains 3 rows (Cerebras / DeepSeek / NVIDIA NIM) plus a Notes column flagging free tier and pricing
hosted_oss configure example now shows free-tier path (Cerebras), cheapest paid path (DeepSeek direct), and the original Together example
proprietary section gains Google AI Studio configure example with the closed-weights caveat

Tests

test-doc-templates.sh adds 2 final drift gates: AGENTS.md tier table covers v0.8.0 providers + PRIVACY.md tier walkthroughs cover same. Catches the next regression of this kind.
29/29 doc-template tests green (was 27/27).
Full suite: 301/11 (was 299/11).

v0.8.0 drift family closed. Four surfaces in total were carrying the same drift — install.sh next-steps (v0.8.4), validator domain mismatch (v0.8.5, related), setup-wizard SKILL.md (v0.8.6), cross-model-review SKILL.md (v0.8.7), and now AGENTS.md + PRIVACY.md (v0.8.8). Each surface now has a regression gate.

[0.8.7] - 2026-05-06

Fixed — cross-model-review SKILL.md provider drift + stale DeepSeek price

Audit-driven completion of the v0.8.0 provider rollout: this is the third surface (after install.sh in v0.8.4 and setup-wizard SKILL.md in v0.8.6) carrying the same drift. The Step 2 reviewer table still recommended togetherai/deepseek-ai/DeepSeek-V3 as the default at ~$0.27/M — both the model id and the price had moved.

Updates to skills/cross-model-review/SKILL.md:

Step 2 reviewer table gains four rows for v0.8.0 providers:
- private_local/mlx — Apple Silicon native (Qwen2.5-Coder-32B 4bit)
- hosted_oss/cerebras — free tier, ~2000 tok/s, gpt-oss-120b / qwen-3-235b-a22b-instruct-2507
- hosted_oss/deepseek direct — cheapest paid hosted reasoning (~$0.14/M cache-miss, deepseek-chat)
- hosted_oss/nvidia_nim — free credits at build.nvidia.com
Stale Together row kept but model bumped to DeepSeek-V3.1 (current) and price callout dropped from this row (it was wrong, and the cheaper deepseek-direct row sits next to it now).
Default suggestion no longer pins togetherai/DeepSeek-V3. Points users to docs/cost-ladder.md for the per-budget pick and names three common defaults (cerebras free, deepseek cheap-paid, ollama local).
Step 3 examples now show two paths — the free-tier Cerebras default and the cheapest-paid DeepSeek-direct path. Old Together example removed.

Tests

test-doc-templates.sh adds two gates against cross-model-review/SKILL.md:
- Each v0.8.0 provider (cerebras / deepseek / nvidia_nim) appears as a Step 2 table row OR a Step 3 --reviewer-provider flag — avoids coincidental matches on legacy model names like deepseek-coder-v2:16b.
- Stale ~$0.27/M DeepSeek price string must not reappear.
27/27 doc-template tests green (was 25/25).
Full suite: 299/11 (was 297/11).

[0.8.6] - 2026-05-06

Fixed — sibling-wizard awareness + setup-wizard skill provider drift

Two issues in the same release because they share a root cause: drift between v0.8.0's expanded picker and the surfaces that describe it.

Sibling-wizard awareness in install.sh. Caught during the states-project-research consumption test where the repo already had claude-sdlc-wizard installed. install.sh happily wrote .opencode/ without acknowledging the existing .claude/ install — leaving the user wondering whether the wizards were merging behavior or stomping each other (they coexist; one writes .opencode/, the other .claude/).

New detection block runs right after the version banner.
Looks for .claude/skills/sdlc/ or CLAUDE_CODE_SDLC_WIZARD.md (claude-sdlc-wizard) and .codex/ (codex-sdlc-wizard).
Prints a single-paragraph "Note: detected sibling SDLC wizard installation(s)" with the dirs found, then continues install.
Non-blocking; informational only.

Setup-wizard SKILL.md provider drift. Same root cause as v0.8.4's install.sh next-steps drift — the skill still listed the v0.2.0 tier providers (Ollama / LM Studio / llama.cpp / vLLM, Together / Groq / OpenRouter, Anthropic / OpenAI). Five providers added in v0.8.0 (MLX, Cerebras, DeepSeek direct, NVIDIA NIM, Google AI Studio) were missing.

private_local row gains MLX (Apple Silicon native).
hosted_oss row gains Cerebras / DeepSeek direct / NVIDIA NIM, with a free-tier callout pointing to docs/cost-ladder.md.
proprietary row gains Google AI Studio (Gemini, closed weights).
Step 2 now shows the --free-tier-first flag inline so users see it without reading --help.

Tests

test-install.sh adds T10/T11/T12: claude-sdlc-wizard detected, codex-sdlc-wizard detected, fresh empty target produces no false sibling detection.
test-doc-templates.sh adds a setup-wizard provider drift gate: asserts every v0.8.x picker provider + --free-tier-first are named in the SKILL.md.
17/17 install behavior tests green (was 14/14).
25/25 doc template tests green (was 24/24).
Full suite: 297/11 (was 293/11).

[0.8.5] - 2026-05-06

Fixed — schema validator emits domain-mismatch hint on foreign-domain artifacts

validate-review-artifact.js previously printed only generic required field missing errors when a research- or persuasion-domain handoff was validated against the code-review schema. Caught during the states-project-research consumption test: a pre-existing research-review handoff with topic / audience / stakes produced 5 missing-required errors with no signal that the schema might just be wrong for the artifact category.

Heuristic added to scripts/validate-review-artifact.js:

After validation fails, count missing top-level required fields and the presence of any foreign-domain markers (topic, audience, stakes, research_question, claim, argument, manuscript, paper, theme, narrative).
If ≥3 missing required AND ≥1 foreign marker present, emit a single DOMAIN HINT: line on stderr before the per-field error list. Tells the user the artifact looks like a non-code-review category and to check the schema choice.
No behavior change for legit code-review artifacts (verified with T32: partially-filled code-review handoff produces no hint, T33: valid handoff with forward-compat extras still validates clean).

Tests

test-review-schemas.sh adds T31/T32/T33: foreign-domain triggers hint, partial code-review does not, valid extras-laden handoff validates without false positive.
33/33 review-schema tests green (was 30/30).
Full suite: 293/11 (was 290/11).

[0.8.4] - 2026-05-06

Fixed — install.sh "Next steps" hint refreshed for v0.8.x picker

The post-install banner had drifted: it still printed the v0.2.0 tier list (Ollama / LM Studio / llama.cpp / vLLM, Azure / Bedrock, Together / Groq / OpenRouter, Anthropic / OpenAI), missing every provider added since v0.8.0. Caught during the states-project-research consumption test — fresh installs were sending users to a stale provider menu while the real picker offered five more options.

Fixes in install.sh:

private_local row gains mlx (Apple Silicon native, v0.8.0)
hosted_oss row gains cerebras / deepseek / nvidia_nim (v0.8.0)
proprietary row gains google_aistudio (v0.8.0)
New --free-tier-first example line so users discover the bias flag without reading --help
Cost guidance link to docs/cost-ladder.md (v0.8.0) — the doc exists in the bundle but nothing pointed to it from the install path
$0 / $20 / $200 literals escaped (\$) — they would have been expanded as positional parameters and tripped set -u in the unquoted heredoc

Tests

test-install.sh adds T9: assert next-steps text mentions every provider the picker emits + --free-tier-first + cost-ladder.md. This is the regression gate so the banner can't drift again.
14/14 install behavior tests green (was 13/13).
Full suite: 290/11 (was 289/11).

[0.8.3] - 2026-05-06

Fixed — release workflow unblocked

tests/test-review-schemas.sh T13/T14 asserted fail when live .reviews/handoff.json / .reviews/response.json were absent, but those files are per-cycle review scratch — response.json is gitignored and CI clean checkouts never have either. Every release tag from v0.6.0 onward (v0.6.0, v0.7.0, v0.8.0, v0.8.1, v0.8.2) failed at this single test, leaving npm pinned at 0.2.0 while local was at 0.8.2.

Fix: T13 + T14 now use validate-if-present semantics — schema check runs when the file exists, absent files emit a "skipped" pass.

30/30 review-schema tests green (was 29/30). All 11 test suites green.

PR #1.

[0.8.2] - 2026-05-05

Fixed — codex round-2 F3 follow-up (PARTIALLY-FIXED → HOLDS)

Codex round-2 returned 8/10 NOT_CERTIFIED with F1/F2/F4/F5 all HOLDING but F3 PARTIALLY-FIXED: the round-1 fix only refreshed the $0/mo table; three more llama-3.3-70b Cerebras references remained later in docs/cost-ladder.md at lines 122, 157, 173.

v0.8.2 finishes F3:

cost-ladder.md:122 ($20/mo Reviewer slot) — split Cerebras + Groq paths: Cerebras → gpt-oss-120b, Groq → llama-3.3-70b-versatile (Groq still catalogs the Llama 3.3 model under the -versatile suffix; Cerebras dropped it)
cost-ladder.md:157 ($200/mo CI-gate slot) — same split
cost-ladder.md:173 (per-job picker, "Routine fix" row) — Cerebras example now gpt-oss-120b or qwen-3-235b-a22b-instruct-2507
cost-ladder.md:177 (per-job picker, "CI gating" row) — added parenthetical noting Groq still ships Llama 3.3 70B but Cerebras doesn't — important context since both providers were lumped together in the original

Untouched (intentionally legitimate references):

Line 27: capability-floor table — generic class example
Line 54: Groq's llama-3.3-70b-versatile — actual current model
Lines 66, 212: calibration-history callouts about the round-1 catch — kept as the audit trail

Tests

No test changes — doc-only fix. Suite still 289/11 green.

[0.8.1] - 2026-05-05

Fixed — codex round-1 cross-model review (5 findings, all addressed)

v0.7.0 + v0.8.0 shipped without external review. Codex xhigh round-1 returned 6/10 NOT_CERTIFIED with 5 actionable findings. v0.8.1 fixes all five.

F1 (P1) — scripts/detect-backends.sh:140. Privacy-first cascade emitted hosted_oss/google_aistudio while configurator only had proprietary/google — followup --tier hosted_oss --provider google_aistudio would fail "unsupported tier/provider". Fix: deleted the wrong-tier line; the proprietary fallthrough at line 143 was already correct. Both cascades now consistently put Google in proprietary tier.

F2 (P1) — Detector accepted NIM_API_KEY/GEMINI_API_KEY as alternates for NVIDIA_API_KEY/GOOGLE_API_KEY, but the configurator hardcoded the canonical names in {env:NAME} references. A user with only the alternate set got a "successfully configured" backend that silently failed auth at runtime. Fix: dropped alternate support entirely. Canonical names only — NVIDIA_API_KEY for NVIDIA NIM, GOOGLE_API_KEY for Google AI Studio. JSON output envs: [..] array collapsed to env: "NAME" singular for shape consistency.

F3 (P1) — docs/cost-ladder.md recommended Cerebras llama-3.3-70b (not in current Cerebras catalog) and Gemini gemini-2.0-flash (deprecated, June 2026 shutdown). Fix: updated $0/mo path to current valid IDs (qwen-3-235b-a22b-instruct-2507 / gpt-oss-120b / gemini-2.5-flash). Added a "Verifying current model IDs" callout block with live links to each provider's catalog. Added per-provider calibration history to the closing notes section.

F4 (P2) — DeepSeek price quoted ~$0.27/M, current is ~$0.14/M cache-miss. Fixed inline + added cache-hit qualifier to the calibration notes.

F5 (P2) — scripts/validate-review-artifact.js:133 only handled additionalProperties === false, ignored schema-valued additionalProperties used in the schemas for verification_state.test_counts. Bad values like {"bad":"not-int","negative":-1} validated successfully. Fix: added object-schema branch — extra properties now recursively validate against the addProps schema. Live .reviews/*.json artifacts still pass; bad test_counts now fail with type/minimum errors.

Tests

test-backend-picker.sh 29 → 31: T29 (Google in proprietary tier both cascades), T30 (alt env names not honored)
test-review-schemas.sh 28 → 30: T29 (schema-valued addProps enforced), T30 (positive case)
T21 in picker updated to use GOOGLE_API_KEY instead of GEMINI_API_KEY

Total: 289 tests across 11 suites (was 285 in v0.8.0).

Dogfood

Schemas validated their own review artifacts: .reviews/handoff.json and .reviews/response.json for the v0.7-v0.8-001 review both validate against the v0.7.0 schemas. Conditional validation fired — all five findings are status: FIXED with fix_summary + fix_locations populated.

[0.8.0] - 2026-05-05

Added — free-tier-first cascade + 5 new providers + cost ladder doc

The wizard's privacy-first cascade has always recognized that local beats hosted beats proprietary on data sovereignty. v0.8.0 adds the orthogonal axis: cost. New --free-tier-first flag on detect-backends.sh biases recommendations toward providers with generous free tiers (NVIDIA NIM credits, Cerebras free, Groq free daily, Google AI Studio 1500 req/day) before paid hosted/proprietary. Local tier still wins both cascades — it's both privacy-max AND free.

Five new providers in detector + configurator:

Cerebras (hosted_oss) — fastest hosted inference (~2000 tok/s on Llama 3.3 70B), free tier resets daily. CEREBRAS_API_KEY.
DeepSeek direct (hosted_oss) — cheapest path to DeepSeek-V3.1 / R1 (~$0.27/M in vs ~$0.50 via OpenRouter). DEEPSEEK_API_KEY.
NVIDIA NIM (hosted_oss) — generous free credits at build.nvidia.com, hosts most OSS models. Accepts either NVIDIA_API_KEY or NIM_API_KEY.
Google AI Studio (proprietary — Gemini is closed weights) — 1M-context Gemini 2.0 Flash with 1500 req/day free quota. Accepts GOOGLE_API_KEY or GEMINI_API_KEY.
MLX (private_local) — Apple Silicon native inference, fastest local on M-series Macs. mlx_lm.server defaults to 127.0.0.1:8080.

docs/cost-ladder.md — concrete budget map. $0/mo (free tiers only), $0/mo local-only (hardware capex), $20/mo (one premium sub + free for the rest), $200/mo (multi-agent CI loops). Per-job picker table — routine fix vs long-context refactor vs security audit vs CI gating each have a different right answer. Ships in the npm tarball via files[] "docs/" addition.

Changed

detect-backends.sh JSON shape grew: private_local.mlx, hosted_oss.{cerebras,deepseek,nvidia_nim}, proprietary.google_aistudio. Existing fields unchanged (forward-compatible — new keys won't break old consumers).
configure-backend.sh PROVIDER_ALIASES gained cerebras→cerebras, deepseek→deepseek, nvidia_nim→nvidia, google_aistudio→google, gemini→google, mlx→mlx. Each has a matching fragment that emits the right baseURL + apiKey env reference + custom-provider models: { [model]: {} } block (so opencode run resolves the pin without ProviderModelNotFoundError).
--free-tier-first flag on detect-backends.sh (or DETECT_FREE_TIER_FIRST=1 env) reorders recommendation cascade to prefer free providers in the hosted bucket.
README banner bumped + cost-ladder.md cross-link added.
package.json files[] adds docs/ so cost-ladder.md ships in the npm tarball.

Tests

tests/test-backend-picker.sh 21 → 29 tests: detector picks up cerebras / deepseek / nvidia_nim / google_aistudio via env; --free-tier-first changes cascade ordering; configure-backend emits correct fragments for each new provider (Cerebras + DeepSeek + NVIDIA NIM + Google AI Studio + MLX); nvidia_nim and google_aistudio aliases map to canonical nvidia / google provider IDs.
tests/test-doc-templates.sh 17 → 24 tests: cost-ladder.md exists + has the load-bearing sections ($0/mo, $20/mo, $200/mo, capability floor, hybrid pattern); package.json files[] includes docs/.

Total: 285 tests across 11 suites (73 + 11 + 13 + 29 + 10 + 10 + 26 + 51 + 10 + 24 + 28).

Why this matters

Two complaints surface most often when users evaluate the wizard: (1) "is it actually OSS-only or can I use it with Claude/GPT too?" and (2) "what does this actually cost to run?" v0.7.0 punted on (2); v0.8.0 makes the answer concrete with the cost ladder doc + the free-tier-first cascade. The wizard is any-backend by design — the OSS tier is a differentiator vs the Claude/Codex siblings (those lock you in), not a requirement. Hybrid coder/reviewer (ceiling model coder + free-tier reviewer) is the dominant pattern for cost optimization without losing rigor.

[0.7.0] - 2026-05-05

Added — JSON Schemas for review artifacts + zero-dep validator

Codifies the structures the wizard has been hand-writing across review rounds. The .reviews/handoff.json and .reviews/response.json shapes were previously implicit — every reviewer + every consumer had to infer them. v0.7.0 makes both shapes explicit + machine-checkable.

templates/schemas/handoff.schema.json — JSON Schema (draft-07) for the handoff artifact: review_id, status (PENDING_REVIEW / IN_REVIEW / CERTIFIED / NOT_CERTIFIED), round, mission, success, failure, review_instructions, plus optional files_changed, verification_state with tests_green + test_counts, preflight_path, response_path, artifact_path. Permissive additionalProperties: true for forward compat — extra fields don't break validation.

templates/schemas/response.schema.json — Schema for the response artifact: top-level responses[] array with per-finding shape (finding_id, severity matching ^P[0-2]( $.*$)?$ to allow parenthetical context, title, claim, status). Conditional validation enforces: FIXED requires fix_summary + fix_locations; REJECTED / WONT_FIX requires rejection_reason. Pattern-key support for recheck_instructions_for_round_N.

scripts/validate-review-artifact.sh + companion .js — zero-dep validator (pure node, no npm install). Implements the draft-07 subset the schemas use: type, required, properties, enum, pattern, minLength, minimum, items, $ref (#/definitions/*), patternProperties, allOf with if/then, const, definitions. Exit codes: 0 valid / 1 invalid (errors with jsonpath + reason on stderr) / 2 usage / missing-file / unparseable JSON.

Both schemas + validator install to .opencode/schemas/ and .opencode/scripts/ respectively. Live .reviews/handoff.json and .reviews/response.json (round-2 artifacts from v0.2.0) validate against the schemas — the schemas were derived from these artifacts so they're guaranteed compatible with prior rounds.

Updated — skills consume the schemas

cross-model-review SKILL.md — new Step 1.5 validates handoff + response against the schemas before sending the prompt to the reviewer. A malformed handoff wastes reviewer tokens and produces a confused review; validation is fast (zero deps, sub-100ms) and fails fast with specific jsonpath + reason for every error.

setup-wizard SKILL.md — Step 4 (Generate) now mentions the schemas as the canonical shape for review artifacts. No setup work required (schemas auto-install); the skill points consumers at them when they create their first review artifact.

Drift-test extension

tests/test-bundle-drift.sh extended to:

Include scripts/*.js companions in the "every script ships in install.sh" check (T7) — without this, a .js file added to scripts/ but forgotten in install.sh would silently never reach consumers.
New T12: every templates/schemas/*.schema.json must be shipped by install.sh — same drift-class for the schemas dir.

Tests

tests/test-review-schemas.sh — 28 tests covering: schemas exist + parse + declare draft-07; validator script exists + executable + prints usage on --help + handles missing-file/bad-JSON gracefully; live .reviews/*.json artifacts validate; negative cases for missing required fields, bad enum, wrong type, FIXED-status missing fix_summary/fix_locations, REJECTED-status missing rejection_reason, P3 severity (pattern violation), parenthetical P0 severity (allowed), recheck_instructions_for_round_N pattern keys; bundle-test that install lands schemas + validator at the expected paths and the installed pair validates the live handoff end-to-end.

Total: 270 tests across 11 suites (73 + 11 + 13 + 21 + 10 + 10 + 26 + 51 + 10 + 17 + 28). bundle-drift grew from 48 → 51 (.js inclusion + 2 schema-shipped checks).

Changed

install.sh REQUIRED_SOURCES + declare_target add the validator pair (.sh + .js) and both schemas.
package.json description mentions JSON Schemas; test script adds the new suite. files[] already includes scripts/ and templates/, so the new files publish without further changes.

Why this matters

The handoff/response artifacts have been the load-bearing hand-shake between the implementing agent and the reviewer all session. Locking the shape down means: (1) cross-model-review skill fails fast on bad handoffs instead of producing confused reviews, (2) the ditto cross-host migrator (when it lands at v0.1.0) can parse + transform these artifacts safely, (3) any future CI tooling gating release on a CERTIFIED status can validate the artifact shape with zero npm install. Forward-compat is preserved — extra keys are allowed, so older artifacts validate against the new schema without retrofit.

[0.6.0] - 2026-05-05

Added — `SDLC.md` and `ARCHITECTURE.md` templates

Closes the doc-template gap. v0.4.0 shipped four TESTING.md domain templates; setup-wizard skill referenced SDLC.md and ARCHITECTURE.md generation but had no template, so each consumer reinvented the wheel.

templates/sdlc.md — SDLC baseline (plan→TDD→self-review→cross- model-review), confidence levels (HIGH / MEDIUM / LOW), commands table, cross-cutting rules. References both the codex flow and the v0.3.1 OSS-tier cross-model-review skill so consumers know they have a vendor-neutral path. Includes the  metadata comment so update-wizard skill can detect drift.

templates/architecture.md — System overview / components / environments / deployment / decisions log. Seeds the decisions log with one example entry (the Bun.spawnSync vs execFile decision from v0.2.0 — illustrates the format and gives consumers a real reference).

Both install at .opencode/templates/ so the setup-wizard skill copies them with project-specific substitutions during bootstrap.

Tests

tests/test-doc-templates.sh — 17 tests: templates exist, contain load-bearing sections (Plan/TDD/self-review/Confidence/version-stamp for SDLC.md; overview/Environments/decisions for ARCHITECTURE.md), SDLC.md invokes skill({ name: "sdlc" }), install delivers both to .opencode/templates/, setup-wizard skill references both template paths.

Total: 238 tests across 10 suites (73 + 11 + 13 + 21 + 10 + 10 + 26 + 47 + 10 + 17).

[0.5.0] - 2026-05-05

Added — `npx opencode-sdlc-wizard check` subcommand

Symmetric companion to init. Answers "is this install behind upstream?" with an exit code that skills, hooks, and CI can consume:

0 — current
1 — behind (also prints installed + latest versions)
2 — not installed (no .opencode/.wizard-stamp) or fetch failed

npx opencode-sdlc-wizard check                  # human-readable
npx opencode-sdlc-wizard check --json           # {installed, latest, status, target_dir}
npx opencode-sdlc-wizard check --target-dir /p

Backed by scripts/check-updates.sh (also installed at .opencode/scripts/check-updates.sh so the update-wizard skill and instructions-loaded-check.sh hook can invoke it directly).

Fixed — hook references parent npm package

.opencode/hooks/instructions-loaded-check.sh ran npm view agentic-sdlc-wizard version (parent's npm name) when checking for staleness. End users of opencode-sdlc-wizard would see "upgrade-available" nudges pointing at the wrong wizard. Switched to npm view opencode-sdlc-wizard version. Mirror at hooks/ synced.

.opencode/hooks/sdlc-prompt-check.sh referenced CLAUDE_CODE_SDLC_WIZARD.md in a comment; clarified that the hook is Claude-Code-flavored and the OpenCode runtime has no UserPromptSubmit analog so the fire-log is informational.

Added — extended bundle-drift coverage (hooks)

tests/test-bundle-drift.sh extended with new classes:

T9: hooks must NOT have executable references to parent-wizard npm packages (npm view agentic-sdlc-wizard, npm install claude-sdlc-wizard, etc.). Comments referencing the parent are OK (informational).
T10: hook dual-location no-drift (every .opencode/hooks/<f> must byte-match hooks/<f>).

Test count grew from 28 → 47 in test-bundle-drift.sh.

Tests

tests/test-check-cli.sh — 10 tests: script existence + bash syntax, --help mentions stamp/installed/upstream, missing stamp → rc=2, current → rc=0, behind → rc=1, --json shape, CLI passthrough, --json passthrough, --help mentions both init and check.
tests/test-bundle-drift.sh extended with hook checks (28 → 47).

Total: 221 tests across 9 suites (73 bundle + 11 plugin + 13 install + 21 picker + 10 CLI + 10 cross-model-review + 26 domain-templates + 47 bundle-drift + 10 check-cli).

[0.4.1] - 2026-05-05

Fixed — bundle-drift test caught + fixed real reference bug

skills/sdlc/SKILL.md (kept verbatim from parent in v0.1.0) referenced scripts/codex-review-with-progress.sh — a parent-only helper we never ported. New consumers running skill({ name: "sdlc" }) would hit a missing-script dead end.

Added — `tests/test-bundle-drift.sh` (28 tests)

Catches future internal-reference drift before it ships:

Every scripts/<name>.sh reference in skills/AGENTS.md/README/PRIVACY must resolve to a real file
Every skill({ name: "<id>" }) reference must resolve to a shipping skill
Every script in scripts/ must be listed in install.sh
Every skill dir in skills/ must be listed in install.sh
install.sh's REQUIRED_SOURCES + declare_target arrays must stay in sync (orphaned source = silent install at wrong path)
Skill dir name == frontmatter name (redundant guard, defense in depth)

Total: 192 tests, all green (73 + 11 + 13 + 21 + 10 + 10 + 26 + 28).

Improved — `skills/sdlc/SKILL.md` cross-model-review pointer

The cross-model review section now points to skill({ name: "cross-model-review" }) as the OSS-tier alternative to the codex flow. Previously the skill assumed codex; v0.3.1 shipped the alternative skill, v0.4.1 surfaces it from the canonical SDLC workflow.

Also documents the </dev/null codex-stdin gotcha discovered live in v0.2.0 + v0.3.0 ship work.

[0.4.0] - 2026-05-04

Added — domain-adaptive TESTING.md templates

The setup-wizard skill no longer assumes web/API. v0.4.0 ships four domain-specific TESTING.md templates that the skill picks based on repo signals:

firmware: HIL / SIL / config-validate / unit pyramid; device matrix doc; mocking rules forbid mocked HIL
data-science: model-eval / pipeline-integration / data-validation / unit pyramid; notebook reproducibility; train/serve skew check
cli: integration-heavy diamond (process spawn + behavior contract
- unit); argv-flag matrix; snapshot tests; OS-contract mocking rules
web (default): integration-heavy diamond (E2E / integration / unit); API contract tests for every endpoint; real-DB mocking rules

Templates live at templates/testing/<domain>.md in the repo and install at .opencode/templates/<domain>.md in target repos. The setup-wizard skill detects the domain from concrete signals (Makefile flash targets / .ipynb / bin package field / fallback) and copies the matching template, substituting <...> placeholders with the project's detected commands.

Tests

tests/test-domain-templates.sh — 26 tests: 4 templates exist, each has appropriate domain markers (HIL / notebook / argv / E2E), each describes a layered structure, install ships each to .opencode/templates/testing/, setup-wizard skill mentions all 4 domains and references the templates directory, package.json files[] includes templates/.
tests/test-bundle-integrity.sh — already covers skill + script set; templates count via the new test suite.

Total: 164 tests, all green (73 bundle + 11 plugin + 13 install + 21 picker + 10 CLI + 10 cross-model-review + 26 domain-templates).

[0.3.1] - 2026-05-04

Added — `cross-model-review` skill (OSS-tier reviewer)

Removes the OpenAI lock from the SDLC review loop. v0.2.0 made the coder any-backend; v0.3.1 makes the reviewer any-backend too. Now the entire SDLC plan→TDD→self-review→cross-model-review loop can run with zero Anthropic+OpenAI dependency.

Bundled artifacts:

skills/cross-model-review/SKILL.md — adaptive skill that picks an OSS-tier reviewer model and runs a cross-model review through OpenCode. Default suggestion: togetherai/deepseek-ai/DeepSeek-V3 for deep reasoning at ~$0.27/M input (pennies per review). Local Ollama path documented for air-gapped contexts.
scripts/cross-model-review.sh — non-interactive wrapper. Reads .reviews/handoff.json + .reviews/response.json, composes the recheck prompt, invokes opencode run --model <provider>/<model>, writes verdict to .reviews/latest-review.md. Symmetric to the codex flow.
Provider-alias resolution shared with configure-backend.sh (together → togetherai, aws_bedrock → amazon-bedrock, etc.)

Recommended reviewers:

Tier	Provider	Model	Why
`private_local`	ollama	`qwen2.5-coder:32b`	Code-tuned, zero egress
`hosted_oss`	togetherai	`deepseek-ai/DeepSeek-V3`	Strongest reasoning OSS
`hosted_oss`	groq	`llama-3.3-70b-versatile`	Fastest hosted, free tier

The two reviewer flows (codex + cross-model-review) produce structurally identical .reviews/latest-review.md artifacts — downstream consumers don't care which reviewer ran.

Tests

tests/test-cross-model-review.sh — 10 tests via stubbed opencode on PATH: script exists/parses, --help prints usage, rejects invocation without required flags, invokes opencode with correct --model <provider>/<model> arg, writes default + --output-path override, refuses without .reviews/handoff.json, alias resolution (together → togetherai in arg), --print-prompt previews without invoking opencode.
tests/test-bundle-integrity.sh extended: skill set is now 5 (added cross-model-review), scripts loop covers all 3.

Total: 138 tests, all green (73 bundle + 11 plugin + 13 install + 21 picker + 10 CLI + 10 cross-model-review).

[0.3.0] - 2026-05-04

Added — `npx opencode-sdlc-wizard init` CLI

Closes the install ergonomics gap. Previously the only paths were git clone + bash install.sh or npm i + bash node_modules/.../install.sh; neither reads as "easy as hell." v0.3.0 ships a bin entry plus a thin Node wrapper at cli/bin/opencode-sdlc-wizard.js that shells out to install.sh.

npx opencode-sdlc-wizard init
npx opencode-sdlc-wizard init --target-dir /path
npx opencode-sdlc-wizard init --dry-run
npx opencode-sdlc-wizard --version
npx opencode-sdlc-wizard --help

Wrapper-level --dry-run previews the bundle install without touching the target. All other flags (--target-dir, --force) pass through to install.sh unchanged.

package.json now declares bin.opencode-sdlc-wizard → cli/bin/opencode-sdlc-wizard.js and adds cli/ to the published files[]. Verified via npm pack --dry-run that the CLI ships in the tarball.

Tests

tests/test-cli.sh — 10 tests: bin executable, CLI parses, package.json declarations correct, --help / --version / unknown subcommand, init --target-dir does install, init --dry-run leaves target untouched, npm pack --dry-run includes the CLI.

Total: 123 tests, all green (68 bundle + 11 plugin + 13 install + 21 picker + 10 CLI).

[0.2.0] - 2026-05-04

Fixed — live OpenCode E2E reliability (validation against 1.14.33)

Static checks (parse, regex, signature) passed in v0.1.0 but the wizard's session-start nudges silently swallowed under a real OpenCode process. Caught by E2E validation against opencode-ai@1.14.33 2026-05-04 and fixed in v0.2.0. Each fix has a regression test in tests/test-plugin-shim.sh.

Session-start race. OpenCode publishes session.created ~1–2 ms after plugin loading begins, before async plugin factories resolve and subscribe to the event bus. The strict event.type === "session.created" discriminator from v0.1.0's round-1 fix never fired in live runs. v0.2.0 uses the FIRST session.* event the handler observes (with a closure dedupe), which arrives in the window the plugin actually has access to. Survives both the race and a future OpenCode fix that closes it.
Async hook execution hung. Both node:child_process.execFile and Bun's $ shell API hang inside OpenCode's bundled Bun runtime — the callback / promise never resolves, swallowing all hook output. v0.2.0 switches runHook to Bun.spawnSync (synchronous; can't hang the event loop). Single bash hook ≈ 50–500 ms — acceptable cost for deterministic completion.

After both fixes, the SDLC Wizard banner reaches OpenCode stderr live (the === SDLC Wizard === block renders the same way it does in Claude / Codex).

Added — privacy-first backend picker

The differentiator that justifies this sibling's existence. Other wizards in the family (agentic-sdlc-wizard, codex-sdlc-wizard) are vendor-bound to Claude or Codex. OpenCode's edge is multi-backend portability — and v0.2.0 makes that edge usable instead of a "you can pin a model in opencode.json" README footnote.

New artifacts:

scripts/detect-backends.sh — probes PATH + env vars for available backends across the four privacy tiers and outputs JSON with a recommendation field (privacy-first cascade — prefers private_local whenever a local LLM runtime is on PATH).
scripts/configure-backend.sh — writes/merges opencode.json for a chosen --tier --provider --model combination. Idempotent (re-running with same args produces byte-identical output), preserves unrelated keys, refuses to clobber an existing model pin without --force. Supports --print-only for skill-driven dry runs.
PRIVACY.md — tier model + Ollama walkthrough + private-path verification checklist.

Updated:

install.sh installs the two scripts at .opencode/scripts/ and chmod +x's them. Adds a privacy-tier hint to its post-install Next Steps.
setup-wizard skill now invokes the detector + configurator with concrete privacy-first defaults instead of the prior placeholder backend question.
AGENTS.md + README.md lead with the privacy-tier picker and link PRIVACY.md.

Tests

tests/test-backend-picker.sh — 21 tests covering detector JSON shape, env-var detection, recommendation cascade, configurator output for ollama/anthropic/azure tiers, custom-provider models entries (so OpenCode can resolve the pin), canonical provider IDs (amazon-bedrock / togetherai), deep-merge of existing provider blocks, idempotency via byte-identical re-runs, no-clobber guard, --force override, --print-only dry-run.
Bundle integrity: scripts presence + bash syntax + tier-name guards; PRIVACY.md tier coverage. tests/test-bundle-integrity.sh 57 → 68.
Install behavior: scripts installed at .opencode/scripts/ + executable. tests/test-install.sh 12 → 13.

Total: 113 tests, all green (68 + 11 + 13 + 21 across 4 suites).

Tiers supported

Tier	Detector aliases → canonical provider ID written to opencode.json
`private_local`	ollama / lm_studio→`lmstudio` / llama_cpp→`llamacpp` / vllm
`enterprise`	azure_openai→`azure` / aws_bedrock→`amazon-bedrock`
`hosted_oss`	together→`togetherai` / groq / openrouter
`proprietary`	anthropic / openai

Local providers use OpenCode's @ai-sdk/openai-compatible provider with each runtime's default localhost port and a models entry so OpenCode can resolve the pin (custom providers without models raise ProviderModelNotFoundError). API keys are referenced via {env:VAR} substitution — no secrets land in opencode.json.

Detector emits user-friendly aliases (aws_bedrock, together, lm_studio); the configurator accepts the alias and writes the canonical OpenCode/models.dev provider ID. Built-in providers (anthropic, openai, azure, amazon-bedrock) skip the models block — their model registry comes from models.dev.

[0.1.0] - 2026-05-03

Added — Phase A port complete

The OpenCode sibling now ships a working SDLC enforcement layer. Phase A scope (per BaseInfinity/claude-sdlc-wizard ROADMAP #9) is closed.

Bundled artifacts:

AGENTS.md — primary instruction file, loaded automatically at session start. Contains SDLC Baseline (replacing the per-prompt nudge from the Claude/Codex siblings since OpenCode has no UserPromptSubmit analog).
.opencode/plugins/sdlc-wizard.js — JS plugin shim. Subscribes to OpenCode events (session.created, tool.execute.before, experimental.session.compacting) and shells out to the portable bash hooks. Honors exit-code-2 block contract on the compact event.
.opencode/hooks/*.sh — 5 bash hooks ported verbatim from the parent: sdlc-prompt-check.sh, tdd-pretool-check.sh, instructions-loaded-check.sh, model-effort-check.sh, precompact-seam-check.sh (plus the shared helper _find-sdlc-root.sh). The bash logic is the value — battle-tested across two earlier siblings; the JS plugin is just glue.
skills/{sdlc,setup,update,feedback}/SKILL.md — 4 skills installed at .opencode/skills/ in target repos. Format is identical to Claude Code's per ROADMAP #91; OpenCode's docs explicitly support Claude-compatible skill paths.
install.sh — non-destructive merge installer. Mirrors codex-sdlc-wizard pattern. Idempotent (re-run is safe), per-file MATCH/CUSTOMIZED/MISSING classification, --force opt-in for overwrite, writes a .opencode/.wizard-stamp metadata marker for future drift detection.

Tests

tests/test-bundle-integrity.sh — 48 tests covering hook presence + executability, hook dual-location no-drift, plugin exports, OpenCode event-name correctness (no Claude event leaks).
tests/test-plugin-shim.sh — 8 tests covering plugin ESM parseability, bash hook syntax, source-glob regex classification, exit-2 block contract.
tests/test-install.sh — 11 tests covering fresh install, idempotent re-install, customization preservation, --force overwrite, skills install location, untouched user opencode.json, help/error flags.

Total: 67 tests, all green.

OpenCode Adaptation Notes

No UserPromptSubmit equivalent. The per-prompt SDLC BASELINE nudge from Claude/Codex moves to AGENTS.md (loaded once per session). Tradeoff: loses per-prompt repetition, but content survives the entire session without per-message token cost.
Hook config is plugin-based. OpenCode does not have a declarative hooks.json schema like Claude Code or Codex. Instead, JS plugins at .opencode/plugins/*.js subscribe to events. This wizard's plugin shim is the OpenCode-specific adapter; the bash hooks themselves are unchanged across all three siblings.
Skills install at .opencode/skills/. OpenCode also reads .claude/skills/ for cross-tool compatibility, but installing into the OpenCode-native location keeps things explicit.
opencode.json is left alone. v0.1.0 deliberately does not touch user opencode.json — backend selection (model pin) is a Phase B concern.

Phase B / Phase C — deferred

Phase B — backend matrix proof (Ollama + Qwen-Coder, Azure OpenAI, Together/Groq, Anthropic baseline). Ship score-history + capability-floor documentation when paired runs complete.
Phase C — hardware scout for the local-tier compute requirement. Existing gaming/Windows laptops first; $200-400 rig or cloud GPU rental only if insufficient.

These phases ship as separate releases. v0.1.0 is the foundation port.

[0.0.1] - 2026-05-03

Bootstrap stage

Repo created, HANDOFF.md written for the implementation session. No installable artifacts. See HANDOFF.md for the bootstrap-session brain dump.

FilesExpand file tree

CHANGELOG.md

Latest commit

History

CHANGELOG.md

File metadata and controls

Changelog

[0.13.4] - 2026-05-25

Fixed — install.sh next-steps drift (v0.11 / v0.12 / v0.13 additions missing)

install.sh changes

Tests

Compat

Workflow insight

[0.13.3] - 2026-05-24

Fixed — --print-only no longer trips the no-clobber guard

Added — npm run test:e2e consumption smoke test

Changed — tests/test-e2e-consumption.sh design notes

Tests

Compat

[0.13.2] - 2026-05-24

Added — Two new OAuth subscription providers + Z.AI default bump

New provider entries

Changed — Z.AI default model bump

Verified live (OpenCode docs + provider docs, fetched 2026-05-24)

Wizard tier-provider distinction note

Changed

Tests

Positioning note (no code change)

Compat

[0.13.1] - 2026-05-18

Changed — PRIVACY.md tier walkthroughs updated for the six-tier reality

ROADMAP.md — v0.10.1 → v0.13.1 sprint declared shipped

Tests

Compat

[0.13.0] - 2026-05-18

Added — subscription tier with GitHub Copilot Pro+

Verified live (OpenCode providers docs, fetched 2026-05-18)

New 6th tier — subscription

Changed

Tests

Why minor bump (v0.12.x → v0.13.0)

Compat

[0.12.0] - 2026-05-18

Added — managed tier with OpenCode Zen as 5th privacy tier

Verified live (Zen docs, fetched 2026-05-18)

New 5th tier — managed

Changed

Tests

Why minor bump (v0.11.x → v0.12.0)

Compat

[0.11.2] - 2026-05-18

Added — Security agent (full set: model + temperature + tools-denial sandbox)

New flags

Changed

Tests

Compat

[0.11.1] - 2026-05-18

Added — Per-agent temperatures (--coder-temp, --planner-temp, --reviewer-temp)

Naming notes

Changed — scripts/configure-backend.sh

Changed — scripts/pick-backend.sh

Tests

Compat

[0.11.0] - 2026-05-18

Added — Z.AI GLM Coding Plan as a first-class proprietary provider

Verified live (Z.AI docs, fetched 2026-05-18)

Changed

Tests

Why this is the minor-version bump (v0.10.x → v0.11.0)

Compat

[0.10.6] - 2026-05-18

Changed — docs/cost-ladder.md recalibration sweep

Not changed (intentionally)

Why this matters

[0.10.5] - 2026-05-18

Added — --sandbox-plan (plan-mode tool denial)

Changed

Tests

Compat

[0.10.4] - 2026-05-18

Fixed — `install.sh` next-steps drift (v0.11 / v0.12 / v0.13 additions missing)

Fixed — `--print-only` no longer trips the no-clobber guard

Added — `npm run test:e2e` consumption smoke test

Changed — `tests/test-e2e-consumption.sh` design notes

Added — `subscription` tier with GitHub Copilot Pro+

New 6th tier — `subscription`

Added — `managed` tier with OpenCode Zen as 5th privacy tier

New 5th tier — `managed`

Added — Per-agent temperatures (`--coder-temp`, `--planner-temp`, `--reviewer-temp`)

Changed — `scripts/configure-backend.sh`

Changed — `scripts/pick-backend.sh`

Changed — `docs/cost-ladder.md` recalibration sweep

Added — `--sandbox-plan` (plan-mode tool denial)

Added — `small_model` top-level pin (`--small-*` flags)

Changed — `scripts/configure-backend.sh`

Changed — `scripts/pick-backend.sh`

Added — Planner agent model routing (`--planner-*` flags)

Changed — `scripts/configure-backend.sh`

Changed — `scripts/pick-backend.sh`

Added — Per-agent permission sandboxing (`--sandbox-test-writer` + `--sandbox-docs`)

Added — Mixed-Mode (per-agent model routing) in `pick` + `configure-backend.sh`

Changed — `scripts/configure-backend.sh`

Changed — `scripts/pick-backend.sh`

Added — `pick` subcommand (one-shot detect → configure backend)

Added — `scripts/pick-backend.sh`

Fixed — `cross-model-review.sh` wrapper alias parity with configure-backend.sh