You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit c6152e9
Browse filesBrowse the repository at this point in the historyBrowse files
Upgrade OpenCode adapter to structurally-verified via real live install
A founder named a concrete need (using Wingman with OpenCode/Codex
specifically) rather than a generic "make it agnostic" request - this
is exactly what evidence-gated-catalog exists to wait for.
Installed real OpenCode (v1.18.4, npm install -g opencode-ai) and
Codex CLI (v0.145.0, npm install -g @openai/codex) as global testing
tools in this dev sandbox for the first time - neither had ever been
installed here before. Global tool installs, not repo dependencies:
no package.json created, zero-dependency invariant unaffected.
OpenCode: genuinely verified against the live install, not just
research -
- `opencode agent list` correctly discovered and typed all 8
boardroom-*.md files as (subagent)
- `opencode debug config`'s resolved output reproduced each file's
exact prompt content verbatim
- boardroom-ceo.md's permission: {edit: deny, bash: deny} frontmatter
is genuinely enforced in the resolved permission engine
- wingman-gate.js is registered in the resolved plugin array,
confirming its export shape loads without error
- its hook name (tool.execute.before) and matched tool name
(plan_exit) are independently confirmed against real sources beyond
this project's own prior research (a documented OpenCode plugin
hook; anomalyco/opencode#18515 referencing plan_exit by name)
Introduced a new, honestly-scoped verification tier, "structurally
verified (live install)" - distinct from "built + tested", since live
model inference (an actual Boardroom review firing) still wasn't
possible: no API key/model provider is configured for either harness
in this sandbox.
Codex CLI stays authored, unverified at the runtime level - the CLI
installs and runs (codex doctor succeeds structurally) but exposes no
config-listing command for custom agents the way OpenCode does, so
.codex/agents/*.toml recognition couldn't be directly confirmed.
Updated: harness-adapters/README.md, opencode/README.md,
opencode/.opencode/agent/boardroom-ceo.md's header (the shared
reference point for all 8 seat files, per the existing convention),
opencode/.opencode/plugin/wingman-gate.js's header, docs/ARCHITECTURE.md
§8b. plugin.json bumped 0.5.37 -> 0.5.38 (version-gate requires this
for any plugins/wingman/references/** change). All 4 Layer-1
validators + check-harness-adapter-drift.mjs pass.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_0158vLsqaiQ3MHcvbuD8ysXo
Copy file name to clipboardExpand all lines: CHANGELOG.md
+5Lines changed: 5 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -2,6 +2,11 @@
2
2
3
3
All notable changes to the Wingman Claude Code plugin.
4
4
5
+
## [0.5.38] - 2026-07-23
6
+
7
+
### Changed
8
+
-**`plugins/wingman/references/harness-adapters/opencode/`** — verification status upgraded from `authored, unverified` to a new `structurally verified (live install)` tier after installing real OpenCode (v1.18.4) and Codex CLI (v0.145.0) in this dev sandbox for the first time. Confirmed: all 8 Boardroom seat files are discovered/parsed/loaded correctly, `boardroom-ceo.md`'s deny-edit/deny-bash permissions are genuinely enforced, and `wingman-gate.js` is registered with its hook name/matched tool name independently confirmed against real sources. Live model inference remains unverified (no configured API key in this sandbox). Codex CLI stays `authored, unverified` at the runtime level — no config-listing command exists to confirm agent-file recognition the way OpenCode's does.
Copy file name to clipboardExpand all lines: docs/ARCHITECTURE.md
+24-12Lines changed: 24 additions & 12 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -258,7 +258,7 @@ either a working non-Claude-Code path or an honestly-documented degradation):
258
258
259
259
**Re-checked against 2026 SOTA multi-harness conventions (2026-07-22)**: researched the AGENTS.md open standard (formalized August 2025, donated to the Linux Foundation's Agentic AI Foundation December 2025; 60,000+ adopting projects, 30+ tools) and a real multi-harness precedent (`wshobson/agents`, which auto-generates native per-harness artifacts from one source because its agents don't depend on harness-unique interactive primitives). This section's conclusion holds: `wshobson/agents`' generation approach only works *because* its agents avoid exactly the couplings (`AskUserQuestion`, `ExitPlanMode`, one-message parallel `Task`/`Agent` fan-out) that the Boardroom review loop structurally depends on — applying that pattern here would overclaim portability, not add it. One concrete, low-risk change did come out of the research: the repo now follows the AGENTS.md standard's own convention of `AGENTS.md` as the canonical file with `CLAUDE.md` as a symlink to it (previously reversed) — a documentation-layer change, not an execution-portability one.
- **First real installs of OpenCode and Codex CLI in this dev sandbox; genuine, evidence-based upgrade of the OpenCode adapter's verification status.** A founder named a concrete need (using Wingman with OpenCode/Codex specifically) rather than a generic "make it agnostic" request — this is what `evidence-gated-catalog` and prior escalations in this log were waiting for. Installed both real CLIs (`npm install -g opencode-ai` → v1.18.4; `npm install -g @openai/codex` → v0.145.0) as global testing tools, not repo dependencies (no `package.json` created, no change to the zero-dependency invariant). Neither harness has a configured API key/model provider in this sandbox, so live model inference (an actual Boardroom review firing) remains unverified for both — that would need real founder-owned credentials. What was genuinely verified against the live installs: **OpenCode** — `opencode agent list` correctly discovered and typed all 8 `boardroom-*.md` files as `(subagent)`; `opencode debug config`'s resolved output reproduced each file's prompt content verbatim; `boardroom-ceo.md`'s `permission: {edit: deny, bash: deny}` frontmatter is genuinely enforced in the resolved permission engine; `.opencode/plugin/wingman-gate.js` is registered in the resolved `plugin` array, confirming its export shape loads without error; its hook name (`tool.execute.before`) and matched tool name (`plan_exit`) are independently confirmed against real sources (documented OpenCode plugin hook; `anomalyco/opencode#18515` referencing `plan_exit` by name) beyond this project's own prior research. Promoted from `authored, unverified` to a new, honestly-scoped tier, `structurally verified (live install)` — distinct from `built + tested`, since behavioral/inference-level correctness is still unconfirmed. **Codex CLI** — the CLI itself installs and runs (`codex doctor` succeeds structurally), but exposes no config-listing command for custom agents the way OpenCode does, so the `.codex/agents/*.toml` files' actual recognition couldn't be directly confirmed this pass; stays `authored, unverified` at the runtime level. Updated: `plugins/wingman/references/harness-adapters/README.md`, `opencode/README.md`, `opencode/.opencode/agent/boardroom-ceo.md`'s header (the shared reference point for all 8 seat files), `opencode/.opencode/plugin/wingman-gate.js`'s header, `docs/ARCHITECTURE.md` §8b.
- **Post-merge Layer-2 audit + real eval coverage for the org-template wiring, both closed clean.** Two parallel audit passes (per `systematic-auditing`'s discipline — independently re-verified, not trusted from self-report) checked the two most recent merges (AGENTS.md canonicalization, `references/org-template/`) for cross-doc consistency and plain-language/doc-index compliance. Found 2 real, minor issues, both fixed: (1) `plugins/wingman/AGENTS.md` (nested, package-scoped) never mentioned `references/org-template/` — added a parallel section to its existing `harness-adapters/` treatment. (2) `references/org-template/README.md` was a genuine doc-index orphan (not cited by any command/skill, unlike `harness-adapters/README.md`'s citation from `git-pr-workflow`) — fixed by citing it from `discovery.md`'s References section. Everything else audited came back clean (confirmed `plain-language-checkpoint` correctly doesn't apply to this agent-facing reference content, per that skill's own explicit "founder-facing output only" scoping). Separately, ran a real, un-briefed behavioral eval (Run 4 of `evals/cases/discovery.md`): a fresh subagent given a founder ask shaped as an unambiguous "freelancing delivery" project correctly matched `references/org-template/project-types/catalog.md`, consulted `freelancing.md` unprompted (per `discovery.md`'s own instruction, not told to), and produced Discovery output independently verified (real file, not self-report) to contain the playbook's specific guidance (client ship/no-ship authority, scope discipline, clean handoff) — genuinely differentiated from the generic Run 3 scenario, not superficial.
Copy file name to clipboardExpand all lines: plugins/wingman/.claude-plugin/plugin.json
+1-1Lines changed: 1 addition & 1 deletion
Original file line number
Diff line number
Diff line change
@@ -1,6 +1,6 @@
1
1
{
2
2
"name": "wingman",
3
-
"version": "0.5.37",
3
+
"version": "0.5.38",
4
4
"description": "An AI Boardroom of agents that plans, builds, secures, and ships production-grade software end to end, with plain-language checkpoints instead of code review.",
Copy file name to clipboardExpand all lines: plugins/wingman/references/harness-adapters/opencode/.opencode/agent/boardroom-ceo.md
+13-5Lines changed: 13 additions & 5 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -6,11 +6,19 @@ permission:
6
6
edit: deny
7
7
bash: deny
8
8
---
9
-
<!-- Translated from plugins/wingman/agents/boardroom-ceo.md. Verification status: authored,
10
-
unverified -- no live OpenCode install in the Wingman dev sandbox. Frontmatter schema per
11
-
https://opencode.ai/docs/agents/ (mode/model/permission fields confirmed via research, not a
12
-
live install). `model` is a placeholder -- point it at whatever provider/model this OpenCode
13
-
install actually has configured. -->
9
+
<!-- Translated from plugins/wingman/agents/boardroom-ceo.md. Verification status (updated
10
+
2026-07-23): structurally verified against a real, live OpenCode install (v1.18.4, installed
11
+
via `npm install -g opencode-ai`) -- `opencode agent list` and `opencode debug agent
12
+
boardroom-ceo` both confirmed this file (and all 7 other boardroom-*.md files) is discovered,
13
+
parsed, and loaded correctly; `opencode debug config`'s resolved output shows this file's exact
14
+
prompt content verbatim, and its `permission: {edit: deny, bash: deny}` frontmatter is
15
+
genuinely enforced in the resolved permission engine (confirmed via `debug agent`'s JSON
Copy file name to clipboardExpand all lines: plugins/wingman/references/harness-adapters/opencode/.opencode/plugin/wingman-gate.js
+10-2Lines changed: 10 additions & 2 deletions
Original file line number
Diff line number
Diff line change
@@ -113,8 +113,16 @@ export function evaluateCheckpoint(inlinePlanText, cwd) {
113
113
114
114
// OpenCode plugin export. Shape per https://opencode.ai/docs/config/ and public plugin-development
115
115
// references (a plugin module exports an async function receiving the app context, returning a
116
-
// hooks object) -- NOT confirmed against a live install. `plan_exit` as the matched tool name is
117
-
// per OpenCode's documented plan/build-mode workflow (see this adapter's README).
116
+
// hooks object). Updated 2026-07-23: this file's registration is now confirmed against a real, live
0 commit comments