docs: changelog for v0.9.0

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017tZMT6285UmrKjA27wmzCT

Author: MShokry

CHANGELOG.md

@@ -10,6 +10,49 @@ update workflow. Version bumps mean: **MAJOR** = state-file contract /
 role authority / script interface changed; **MINOR** = new template,
 script, flag, or role rule; **PATCH** = prose and docs.
 
+## v0.9.0 — 2026-09-27
+
+- `[safety]` Synced this toolkit's own opencode v2 migration into the
+  scaffolded pipeline templates: the shell permission key renamed `bash:` →
+  `shell:` in `builder.md.tmpl`/`reviewer.md.tmpl`/`tester.md.tmpl` (under
+  the old key, none of a template's shell `deny`/`ask` rules matched
+  anything under opencode v2, so `rm -rf /*`, `sudo *`, `curl *`, etc. were
+  silently unenforced); `oc.sh.tmpl` now authenticates with
+  `OPENCODE_PASSWORD` (v2's `serve` is always password-gated, v1 had none),
+  polls via `opencode api` against v2's `/api/*` routes, aborts via the
+  renamed `/api/session/<id>/interrupt` endpoint, and picks the *newest*
+  message for watchdog progress instead of the oldest (v2 returns messages
+  newest-first; the old indexing could silently track the wrong message and
+  let a wedged run look like it was still progressing).
+- `[docs]` Corrected an earlier `oc.sh.tmpl`/agent-template comment claiming
+  `--auto` lets a deny-listed shell command run anyway on opencode v2 —
+  re-verified live via `GET /api/agent/<name>` and could not be reproduced;
+  left the correction in place rather than deleting the original claim
+  outright.
+- `[process]` `bin/init.sh` now checks `opencode --version` on every
+  scaffold or `--update` run and warns (non-fatally) when it's missing, v1,
+  or newer than the v2 these templates assume, explaining specifically what
+  breaks on v1 (password auth, the permission-key rename, the `opencode
+  api` subcommand, `--server` replacing the removed `--attach`/`--dir`).
+  `toolkit-update.md.tmpl` now tells the lead to surface this warning to
+  the user rather than let it scroll past.
+- `[process]` `bin/init.sh`'s `--builder-model`, `--reviewer-model`,
+  `--reviewer-fallback-model`, and `--tester-model` are no longer required —
+  `apply_defaults()` now fills them with `opencode-go/glm-5.3-flash` /
+  `opencode-go/minimax-m2.7` / `opencode-go/deepseek-v4-flash` /
+  `hcnsec/auto` when omitted, the same lineup two independently scaffolded
+  projects (resto-agent, relationship) landed on. `--project-name` remains
+  the only required flag; every flag can still be passed explicitly to
+  override the defaults. Backward compatible — a call that already passed
+  all four flags behaves identically.
+- `[safety]` `init.sh` now also adds `.agents/.oc-password` (the local
+  opencode v2 server password `scripts/team.sh` generates) to a scaffolded
+  project's `.gitignore`, alongside the existing `.oc-port` and Claude
+  session-id entries.
+- `[docs]` README's quickstart and `docs/MODELS.md`'s recommended lineup
+  updated to show this lineup as the documented default instead of the
+  older Kimi/GLM-5.2/Sonnet-fallback/DeepSeek-Flash recommendation.
+
 ## v0.8.0 — 2026-09-13
 
 - `[process]` Fresh scaffolds now support Codex as the project lead when Claude

README.md

@@ -168,20 +168,34 @@ opencode models          # see what's actually configured before picking models
 ~/tools/agent-toolkit/bin/init.sh \
   --target . \
   --project-name "my-project" \
-  --claude-model sonnet \
-  --builder-model "hcnsec/auto" \
-  --reviewer-model "hcnsec/glm-5.3" \
-  --reviewer-fallback-model sonnet \
-  --tester-model "hcnsec/auto" \
   --test-dir e2e
 ```
 
+`--project-name` is the only required flag. `--claude-model`,
+`--builder-model`, `--reviewer-model`, `--reviewer-fallback-model`, and
+`--tester-model` all default to the lineup two independent real projects
+converged on: Claude Sonnet lead/planner, `opencode-go/glm-5.3-flash`
+builder, `opencode-go/minimax-m2.7` reviewer,
+`opencode-go/deepseek-v4-flash` reviewer fallback, `hcnsec/auto` tester.
+Override any of them per project once `opencode models` shows your server's
+actual list differs — these are a starting point, not a guarantee those
+exact ids still exist for you:
+
+```bash
+~/tools/agent-toolkit/bin/init.sh \
+  --target . \
+  --project-name "my-project" \
+  --builder-model "<vendor/model>" \
+  --reviewer-model "<vendor/model, different family than builder>" \
+  --reviewer-fallback-model "<vendor/model, different family again>" \
+  --tester-model "<vendor/model>"
+```
+
 `--reviewer-model` and `--reviewer-fallback-model` should be **different
 model families** — the fallback is what the pipeline switches to when
 `builder` implements and would otherwise share a vendor with the default
-reviewer, which would defeat cross-vendor independence. The `hcnsec/auto`
-values above are flag *shape* only — run `opencode models`, pin real
-strings, and do not use `auto` for the reviewer.
+reviewer, which would defeat cross-vendor independence. Do not use `auto`
+for the reviewer.
 
 Cost/quality picks (Kimi implementer, GLM reviewer, DeepSeek Flash
 tester, Claude Sonnet lead/planner/fallback), and why one OpenCode

docs/MODELS.md

@@ -41,48 +41,57 @@ request volume per 5h window, expensive → cheap:
 | Kimi K3 | Very high on Go (~110 req / 5h) | Rare “actually hard” implementer, not default |
 | GLM-5.3 | High (~220 req / 5h, tighter cap than 5.2) | Best GLM for **one-shot** review |
 | GLM-5.1 / 5.2 | Mid–high (~880 req / 5h) | Review, or planner if Claude credits are tight |
-| Kimi K2.6 / K2.7 Code | Mid (~1k–1.3k req / 5h) | Default implementer — agentic loops |
+| GLM-5.3-flash | Mid, cheaper than full 5.3 | **Default implementer** — agentic loops |
+| Minimax M2.7 | Mid, no published Go cap seen yet | **Default reviewer** — different lab than GLM, one/two calls |
+| Kimi K2.6 / K2.7 Code | Mid (~1k–1.3k req / 5h) | Implementer alternative — agentic loops |
 | DeepSeek V4 Pro | Low | Implementer when the window is tight |
-| DeepSeek V4 Flash | Lowest (~30k req / 5h) | Tester only |
+| DeepSeek V4 Flash | Lowest (~30k req / 5h) | **Default reviewer fallback**; also fine as tester |
 
 These counts are provider estimates, not a promise. GLM-5.3 is usually
 the same *job* as 5.2 at a worse credit rate — prefer 5.2 unless 5.3 is
 clearly better on your list.
 
 ## Recommended lineup
 
-For this toolkit's own work (bash + sed, markdown templates, process
-rules, no app runtime):
+`bin/init.sh`'s own defaults (`apply_defaults()` — override any of them
+with the matching flag per project). This is the lineup two independent
+real projects landed on separately, then converged on deliberately:
 
 | Role | Model | Why |
 | --- | --- | --- |
 | **Lead** | Claude Sonnet | Already the Claude Code session. Opus does not improve dispatch. |
 | **Planner** | Claude Sonnet | One shot. A cheaper model writes mushy ACs or starts designing the implementation. Drop to GLM-5.2 only if the Claude pool is the bottleneck. |
-| **Implementer** (`builder` / `senior-dev`) | **Kimi K2.7 Code** (or K2.6) | Cost/quality for surgical template + `init.sh` work. GLM-5.2 only if Kimi is missing. Not GLM-5.3, not K3, not Opus as the default loop. |
-| **Reviewer** | **GLM-5.3** or **5.2** | Different family than Kimi. One/two calls, so GLM's per-call cost is acceptable. Catches process bugs (leftover placeholders, `render()` skip-if-exists, permission YAML ≠ enforcement) better than Flash. |
-| **Reviewer fallback** | Claude Sonnet (or DeepSeek Pro) | A **third** family, and preferably a **second gateway** (see below). Used when builder would share a family with the default reviewer. |
-| **Tester** | DeepSeek V4 Flash | Smoke `init.sh` into `/tmp`, assert no `__[A-Z_]*__`, second run does not clobber. Flash is enough. |
+| **Implementer** (`builder` / `senior-dev`) | **`opencode-go/glm-5.3-flash`** | Cost/quality for the implementer loop's tool-call volume. Kimi K2.7 Code is a solid alternative if GLM-5.3-flash is missing or underperforming on your codebase. |
+| **Reviewer** | **`opencode-go/minimax-m2.7`** | Different family than the builder's GLM. One/two calls, so per-call cost is acceptable; a different lab's weights catch blind spots the implementer's own family shares. |
+| **Reviewer fallback** | **`opencode-go/deepseek-v4-flash`** | A **third** family. Used when the implementer (e.g. `senior-dev` or a differently-configured builder) would otherwise share a vendor with the default reviewer. |
+| **Tester** | **`hcnsec/auto`** | One shot + a smoke command; faithful reporting matters more than reasoning power here, so a router is fine for this role specifically (never for the reviewer — see below). |
 
-`init.sh` shape (replace with strings from `opencode models`):
+`init.sh` shape (these are the defaults — passing no model flags at all
+gets you exactly this; replace with strings from `opencode models` if your
+server's list differs):
 
 ```bash
 --claude-model sonnet \
---builder-model "<aggregator>/kimi-k2.7-code" \
---reviewer-model "<aggregator>/glm-5.2" \
---reviewer-fallback-model sonnet \
---tester-model "<aggregator>/deepseek-v4-flash"
+--builder-model "opencode-go/glm-5.3-flash" \
+--reviewer-model "opencode-go/minimax-m2.7" \
+--reviewer-fallback-model "opencode-go/deepseek-v4-flash" \
+--tester-model "hcnsec/auto"
 ```
 
 If you implement with GLM, reviewer must **not** be GLM (5.2 vs 5.3
-does not count). Switch reviewer to Kimi or DeepSeek Pro.
+does not count) — that's exactly why the default reviewer is Minimax, a
+different lab, rather than another GLM snapshot. If your implementer
+changes family (e.g. Kimi), re-check that the reviewer still differs from
+it; switch to the fallback family if it does not.
 
 Tight credits: DeepSeek Pro implementer, GLM-5.2 reviewer, Flash tester.
 High-stakes (`render()`, permission blocks, state-file contract): keep
-Kimi on implementer; optionally fire the Sonnet fallback as a second
-review pass — still cheaper than implementing on Opus.
+the default implementer; optionally fire the Sonnet reviewer fallback as
+a second review pass — still cheaper than implementing on Opus.
 
 Do not: Claude on builder **and** reviewer; K3/Opus as default
-implementer; Flash as reviewer; `auto` anywhere independence matters.
+implementer; Flash as reviewer; `auto` anywhere independence matters
+(reviewer, never tester).
 
 ## Provider: one aggregator for workers, not one tool per lab