Roadmap: 12 PRs across 12 issues — bug fixes, regression locks, two new features

[YOMXXX]

Hi maintainers and reviewers — over the past few sessions I worked through 12 of the open issues and opened 11 PRs against code-yeongyu/oh-my-openagent (plus 1 against the lsp-tools-mcp submodule). This discussion is a single landing page so you can scan all of them at a glance, see the rationale, and pick a review order that makes sense for you.

I tried to follow systematic-debugging (root cause before fixes) and TDD (red-green for every actual bug fix) throughout. Each PR is independent, scoped to one issue, and small enough to review in isolation — but the table below also groups them so you can batch similar concerns together.

---

Roadmap at a glance

Category	PRs	Net diff	Risk
A. True bug fixes (red-green TDD)	5	~+500 / ~-15	medium — touches runtime hooks & build
B. Regression locks (already-fixed source, missing tests)	4	~+225 / 0	very low — test-only
C. New capabilities (community-requested)	3	~+780 / ~-7	medium — new config field, new doctor check, new MCP
Total	12	~+1.5k / -22

---

A. True bug fixes (TDD red-green for each)

PR	Closes	One-line	Where it fires
#4269	#4183	sync skill resolvers & getSkillByName align with async (short-name + case-insensitive + ambiguity protection)	skill-template-resolver.ts, loader.ts, executor.ts, skill-matcher.ts (genericized)
#4270	#4211	Banner reads bundled build-time version instead of the cached install's package.json	new bundled-version.ts, hook.ts
#4271	#4259	bun build emits dist/locales/{en,zh,index}.js so the published tarball self-documents (locale data is also still inlined in dist/index.js)	package.json build script
#4274	#4251	Keyword detector skips injection if the marker ([search-mode], <hyperplan-mode>, etc.) is already in the text — fixes /undo + resend double-banner	keyword-detector/hook.ts
lsp-tools-mcp#5	#4262	Windows PATHEXT lookup tries .bat/.cmd before the literal name, so jdtls + jdtls.bat resolves to the wrapper instead of the shebang script	packages/lsp-tools-mcp/src/lsp/process.ts

Every one of these has a regression test that fails on dev-as-of-issue and passes after the fix. I verified the red-green cycle locally for each.

---

B. Regression locks — source already fixed, tests were missing

These are issues where investigation revealed the symptom had already been fixed in source (commits like ed6e9955 for path-separator normalization, the existing Zod enums for reasoningEffort). The reporters were likely on stale dist artifacts. To prevent the bugs from quietly drifting back, each PR adds focused regression tests that pin the existing contract.

PR	Refs	One-line
#4272	#4220	hasCliSuffix regression case using the literal Windows-style ast_grep cli path from the bug report
#4273	#4165	Per-schema and cross-schema parity tests for reasoningEffort: "max" on CategoryConfigSchema / AgentOverrideConfigSchema / FallbackModelObjectSchema + a doctor end-to-end test
#4275	#4193	cli-suffix regression mirroring the lsp-tools-mcp candidate paths in the issue's resolution simulation
#4277	#4178	Two applyMcpConfig idempotency tests that survive simulated /new config resets

None of these PRs change production code. They're test-only, ~0 risk to merge.

---

C. New capabilities

PR	Closes	Capability
#4276	#4184	doctor warns when a custom provider model is missing limit.context (auto-compaction would never trigger). 7 new unit tests cover the matcher; wired into existing checkModels.
#4283	#3735	New excluded_agents config — custom lightweight agents (e.g. cybersec) keep working but omo skips its 3 most invasive runtime injectors for them (ultrawork tag, keyword-mode banners, AGENTS.md/README/rules tool-output injection). Safety/recovery hooks still run.
#4286	#3775	Sixth builtin MCP: @modelcontextprotocol/server-memory. Stable default path under $XDG_CACHE_HOME / ~/.cache/opencode/ so memory survives switching projects; OMO_MEMORY_FILE_PATH env override; users can opt out via disabled_mcps: ["memory"] and bring their own.

---

Suggested review order

If you have ~20 minutes, the highest-value pass is the regression locks (B) — pure test additions, very fast to merge, lock down hard-won fixes.

If you have more time, A then C is the natural order: bug fixes first to harden the baseline, then features on top.

Within A:

1. #4271 — 1-line build script change, trivial.
2. #4274 — small, isolated to keyword detector.
3. lsp-tools-mcp#5 — submodule, owned area.
4. #4270 — banner / version resolution change.
5. #4269 — slightly larger refactor (generic helper) but mechanical.

Within C:

1. #4276 — pure doctor warning, additive.
2. #4286 — new MCP, additive (gated by disabled_mcps).
3. #4283 — touches three hook chains; biggest review surface in this set.

---

A note on two reports where the root cause differed

Two of the original reporter root-cause analyses turned out to be off, but the PRs still ship useful work — I flagged this explicitly in each PR so reviewers don't merge based on a wrong assumption:

• Models configured without limit cause auto-compaction to never trigger #4184: the report assumed omo writes provider/model into opencode.json. It does not — users own that file. The fix shape changed from "populate limit automatically" (would silently override user intent) to "doctor surfaces the missing field with a concrete remediation string." Same problem solved, no mutation.
• [sisyphus-bot] Bug: disabled_mcps not re-applied on /new session — stale MCPs from .claude.json leak through #4178: the report assumed omo's applyMcpConfig had stateful drift on /new. The function is actually already idempotent. Whether OpenCode re-invokes the config hook on /new is upstream behavior we can't reach from a plugin event handler. The PR locks the omo-side contract so a future refactor can't introduce the bug we don't currently have.

---

Happy to split, rebase, or close any of these if direction doesn't match. Thanks for the project.