@royenheart/dsh-plugin-opencode-omo

A DeepSeek Harness plugin that adds an opencode-omo agent preset (mode) to the web profile. The mode replicates the behavior of opencode + the omo plugin (oh-my-openagent), scoped to this mode only — other presets (standard/…) keep the default dsh loop, sandboxed fs, and no omo hooks.

What the mode provides

• opencode + omo system prompt — the real opencode default.txt persona (tone, style, proactiveness, conventions, code style, task guidance) + omo's Sisyphus orchestrator identity, declared as the complete system prompt: the dsh harness identity and runtime-context snapshot are suppressed for this mode. The loop shim additionally prepends opencode's live environment block (exact model id, working dir, workspace root, git, platform, date, recomputed every step).
• omo role picker in the composer — in dsh's existing conversation.input.left tool-row slot (after the access/plan chips), the primary/all roles sisyphus, hephaestus (Deep Agent), prometheus (Plan Builder), atlas (Plan Executor), and sisyphus-junior. Selecting one swaps the session's complete system prompt and applies that role's configured model. Specialist roles (athena, athena-junior, council-member, metis, momus, oracle, librarian, explore, multimodal-looker) are reached through task(subagent_type=...), not the composer picker.
• Global "Role Settings" in the dsh settings panel (settings.section): per-role primary model dropdown (follow current / omo default / fixed) under a centered "Primary model" label; a dsh-style circle "+" button opens a fallback model list below the role box (repeatable additions, cancel/close adds nothing), persisted in the opencode-omo settings entry. On request failure the loop shim advances through the role's fallback chain before the harness retry policy runs.
• opencode toolchain (complete) — persistent bash, read/write/edit/read_image, apply_patch, glob/grep, todo_write, skill, web_fetch/web_search, the LSP navigation surface (lsp_goto_definition / lsp_find_references / lsp_hover / lsp_go_to_implementation; raw lsp is hidden), exit_plan_mode (plan), ask_user_question. tool-surface.mjs overwrites the model-visible descriptions/parameters with opencode's tool/*.txt text and shims read/edit/write/web_search to opencode's parameter names (web_search.query → dsh queries).
• omo task() surface — task-shim.mjs registers the omo-style task(category/subagent_type/load_skills/run_in_background/task_id) invocation, mapping it onto dsh named subagents + generic delegation.
• omo multi-role subagents — oracle (write-tool-restricted advisor; shell retained, not a sandbox), librarian (external docs/code search), explore (codebase grep), metis (pre-planning), momus (plan reviewer), multimodal-looker (media), plus generic subagent/subagent_fork + workflow/ralph. Delegated children keep dsh's send_message return channel (the harness appends that instruction to every continuable child); the primary surface still hides it behind task()/background_output.
• omo context injection — AGENTS.md/CLAUDE.md walk-up + skills/ + omo's rules-injector (.omo/rules, .cursor/rules, .github/instructions, copilot-instructions.md).
• omo hooks — comment-checker (rejects AI-slop comments on write/edit), hashline (read tagging N#HH|content + hashline_edit stale-ref guard).
• per-mode execution backend — local filesystem (dsh-fs-local) + persistent PTY shell, isolated from other modes' sandboxed fs/shell.
• native-seam loop shim — no dsh-side driver seam. driver.mjs is an ordinary preset plugin using the shipped seams: a dynamic ctx.systemPrompt.section({ complete: true }) recomputes opencode's env block and the selected omo role prompt per assembly; system-prompt/assemble applies opencode's model tool gating, and both the request catalog and the prompt's generated tool table run through the same hidden-name/model-family/role filter; agent/inbox/claimed, agent/pre-step, agent/request, and agent/request-error provide ultrawork detection, maxSteps, role model routing, and fallback retry. Other presets are untouched by construction.

Layout

cordis.patch.yml                 # bundle patch: host row + official agent-preset declaration
patches/                         # dsh-side patch(es) still required for the complete surface
install.py                       # idempotent install/uninstall (manifest + missing runtime/LSP links)
src/                             # host + client plugin halves (role registry, settings, picker UI)
lib/                             # built host/client bundles (npm run build)
scripts/build.sh                 # architecture check + typecheck + tsdown build
presets/opencode-omo/
  agent.cordis.yml               # the composition (tools, roles, hooks, LSP)
  preset.yml                     # display metadata
  persona.md                     # opencode default.txt + omo Sisyphus persona
  roles/*.md                     # subagent personas (also main-role prompts)
  roles/prompts/*.md             # primary-role complete prompts (hephaestus/prometheus/atlas/…)
  skills/                        # omo shared skills
  driver.mjs                     # native-seam loop shim (prompt/route/fallback/maxSteps/ultrawork)
  delegation-restrictions.mjs    # single tool capability/role restriction table
  patch-parser.mjs               # shared apply_patch parser for execution + permission
  background-state.mjs           # closed background activation lifecycle
  rules.mjs                      # rules-injector
  comment-checker.mjs            # comment-checker hook
  apply-patch.mjs                # apply_patch tool
  hashline.mjs                   # omo hashline read-tagging + hashline_edit
  lsp-optional.mjs               # LSP trio when installed; self-disables otherwise

Install

lib/ is generated locally and is not committed. install.py always builds the repository's own toolchain first (npm ci when the toolchain is missing or package-lock.json is newer than the installed lock, then npm run build) and exits with an actionable error if npm, the build, the artifacts, or a required harness link is missing:

python3 install.py install --profile web              # install (idempotent)
python3 install.py uninstall --profile web            # remove

install.py symlinks the package into ~/.dsh/profiles/<profile>/node_modules/, mirrors runtime packages only when absent, ensures/repoints the LSP links (refusing real directories), and edits the profile's package.json (adds/removes the dependency + bundle entry). The bundle patch declares both the host row and the official @deepseek-ai/dsh-agent-preset row, so there is no separate user-preset-root copy.

Manual alternative — the package is a dsh bundle: it declares dsh.bundle.patch and ships the preset. dsh plugin reconciles dsh.profile.bundles from the installed package automatically:

dsh plugin --profile web add link:/path/to/dsh-plugin-opencode-omo

The bundle patch declares the official @deepseek-ai/dsh-agent-preset row for opencode-omo, so no $DSH_HOME/.agent-presets copy is published. Restart dsh and select opencode-omo from the mode picker.

Upgrading from the pre-0.2.1 row id. Existing profiles used the opencode-omo-roles settings entry; rc.2's legacy importer preserves that unknown entry in $DSH_HOME/settings.yaml.imported. On boot the host row reads that backup once per profile and merges its role/session maps into the active opencode-omo entry, so old role models and fallbacks are not silently reset. Profiles migrated by the pre-release 9ad557c home-global marker may re-import once because that marker had no profile identity; review the merged roles after upgrading.

Optional LSP server. The preset preconfigures typescript-language-server for the lsp_* navigation tools. install.py checks PATH and warns when it is missing; the optional LSP module then mounts without a provider so the mode still loads (LSP queries fail gracefully instead of blocking the whole preset). Install it to enable LSP:

npm install -g typescript-language-server typescript

then restart dsh.

web_fetch provider. The preset enables dsh's native web_fetch tool (fetch: true). Official 0.2.1 already registers @deepseek-ai/dsh-web-fetch-http in the base bundle, so this plugin does not insert a second fetch row (that would throw WEB_DUPLICATE_PROVIDER). web_search keeps using the existing DeepSeek search provider (DEEPSEEK_API_KEY). Other presets keep fetch: false, so their tool surface is unchanged.

Required dsh-side changes

One patch is still required for the complete opencode surface:

patches/0001-agent-pre-step-assistant-prefill.patch (rebased onto dsh-v0.2.1-rc.2) adds PreStepDecision.assistantPrefill, a request-only assistant tail. The loop logs it on request/header, appends it to exactly the request that asked for it, and prices it in token pressure; it is never a session assistant/message, so transcript, compaction, and stats stay clean.

Apply it to the dsh checkout with the instructions in patches/README.md; install.py does not modify the harness checkout. Without the patch the preset still loads and uses the fallback below.

Why the patch is still required. opencode appends MAX_STEPS_PROMPT as an assistant-role continuation on the ceiling step. Official 0.2.1 PreStepDecision is still reject | { kind: 'enter', messages, startsRequestSeries? }, and agent/request cannot mutate request messages, so no official extension point can express a request-only assistant tail. Upstream request: discussion #2407.

The plugin probes the runtime capability (ctx.agentLoop.supportsAssistantPrefill). On a patched host the ceiling text is an assistant tail and no system-prompt section is added. On a stock host it keeps the same text and trigger as a system-prompt section (see the fallback table below); that is survivability, not the complete opencode surface.

Fallback on unpatched hosts

Surface	opencode / patched harness	This plugin on a stock 0.2.1 host
Role of MAX_STEPS_PROMPT	Assistant-role continuation at the end of the request	System-prompt section for that step
Session transcript / stats / compaction	Request-only (header), not a session message	Also not a session message (system section is reassembled)
How strictly models stop / wrap up	Tuned for an assistant tail	May treat a system reminder more weakly or more strongly
Token placement	Extra assistant tokens on that request	Extra system tokens on that request
Reconstructable requests	Logged on request/header	Rebuilt from the live prompt assembly

The fallback is selected at runtime; the patch is not dropped when a host can run without it.

Provider-visible format/toolChoice remains an unpatched proposal; omo's regular path does not use it.

Everything else runs on unmodified dsh seams: the preset is declared by the official @deepseek-ai/dsh-agent-preset row, the role picker occupies the existing conversation.input.left slot, and role settings use the official settings form Remote.

Bench experiments (equivalence validation)

Runners and reproduction notes live in tests/benches/; reports are written to docs/exps/. The scientific methodology (paired design, McNemar/bootstrap/TOST, A/A noise floor, trace alignment, cache and latency protocols) is documented in docs/exps/2026-08-16-scientific-bench-methodology.md. Design highlights:

• Run an isolated-port dsh (opencode-omo mode, isolated $DSH_HOME) and the machine's installed opencode + oh-my-openagent (isolated XDG_CONFIG_HOME).
• Both use deepseek-official/deepseek-v4-pro (dpsk v4 pro); the API key comes from the DEEPSEEK_API_KEY environment variable. Scripts hardcode no machine paths or secrets.
• Tiered benches: L1 HumanEval, L2 MBPP, L3 SWE-bench-verified-mini (sampled), comparing pass@1, CoT/reasoning exposure, tool-call chains (read/edit/write/bash/test/subagents), and final patches/answers.
• Performance: bench_metrics.mjs normalizes token usage (including cacheRead) and tool timestamps on both sides; eval_perf.mjs / eval_swe_perf.mjs produce cache hit rate, TTFT, step duration, and per-tool timing offline (see report §5).
• MBPP fix: MBPP rows have no entry_point; the scripts infer the function name from the first assert fn(...) or the reference def fn. The old implementation hardcoded Function name: undefined, which was the root cause of MBPP behavioral divergence (see report §0 and §6.4).
• Raw bench data is downloaded to tests/benches/.data/ and is not versioned; fetch-benches.sh reproduces the downloads and setup-homes.sh reproduces both isolated homes.

Latest results (dpsk v4 pro; L1 30×3, corrected L2 5×2, legacy L3 sample)

level	dsh pass@1	opencode+omo pass@1	per-item agreement
HumanEval (30 tasks × 3 repeats)	1.00	1.00	1.00
MBPP (corrected entry_point, 5 tasks × 2 repeats)	1.00	1.00	1.00
SWE-bench-verified-mini sample (sphinx-doc__sphinx-10323)	same patch	same patch	byte-identical git diff

Full report: docs/exps/2026-08-15-opencode-omo-equivalence-bench.md; raw transcripts are under tests/benches/.runs/.

Version baseline

Active combination: opencode-1.17.11+omo-5.0.0-beta.31.

alignment.json is the machine-readable registry; presets/opencode-omo/roles/prompts/variants/MANIFEST.md records prompt sources.

Upstream	Version	Commit	Used for
opencode	1.17.11	67aec221	tool descriptions, persona, plan/build/MAX_STEPS text
oh-my-openagent	5.0.0-beta.31	62ed7952	Sisyphus family, Atlas, Hephaestus, Oracle/Metis/Momus prompts
oh-my-openagent	4.19.0	14083b89	librarian / explore / multimodal-looker prompts
dsh	0.2.1-rc.2	687f282	host peer target

explore and multimodal-looker are identical in both omo versions; librarian stays on v4.19.0 until the context7-specific v5 diff is adapted to this harness. npm run check:alignment validates the active combination.

Future alignment harness

Tracking: #17.

• Provenance: per-combination manifest and content hashes.
• Deterministic contract: scripted-catalog prompt/tool rendering compared with upstream sections and names.
• Recorded traces: record real opencode+omo traces once per pin, then replay only this plugin and compare normalized events.

Alignment status (against the declared baseline)

• Aligned: opencode default persona (complete system prompt + live env block whose provider/model follow the same assembly-time route as the request — session live model selection or the role primary/fallback (an in-step fallback retry reuses that assembly while the request route advances); workspace root now derived as the git root); opencode tool families + gpt apply_patch/edit-write tool gating enforced on BOTH the model-visible schema and execution (tools/pre-execute deny mirror); opencode maxSteps + verbatim MAX_STEPS_PROMPT; verbatim opencode plan.txt / plan-mode.txt with dynamic ${planInfo} and the plan→build BUILD_SWITCH reminder; omo role catalog/display names; sisyphus/hephaestus/atlas/sisyphus-junior + specialist subagents; comment-checker/hashline/rules-injector hooks; generated Sisyphus routing sections; extracted omo Sisyphus model-family templates (GPT-5.5/GPT-5.4/claude-opus-4-7/claude-opus-4-8/claude-fable-5/gemini/kimi-k3/kimi-k2-7/kimi-k2-6/glm-5-2, with the dynamic Sisyphus fallback for unknown families) plus hephaestus GPT variants, all 8 atlas variants, and specialist model variants (oracle/metis/momus); omo-default per-role PRIMARY model resolution (provider-scope ordered) and fallback chains that start AFTER the primary; omo role sampling defaults (sisyphus/hephaestus GPT effort medium, atlas temperature 0.1); omo-style retryable-error gating before fallback advance; reasoning-effort selectors in role settings; ultrawork keyword override; /start-work, /remove-ai-slops, /refactor, /stop-continuation, /handoff, /hyperplan, /team-mode commands; composer role picker + global per-role model/fallback settings; omo skills published as user-dsh so the third-party skills-manager can manage them. The omo rules-injector text is now folded into the complete system prompt (driver.mjs + rules.mjs) instead of being dropped by suppressRuntimeContext(); approved plans are persisted at .opencode/plans/<created>-<session>.md; specialist subagent personas now load the extracted reference prompt files (oracle/librarian/explore/metis/momus/multimodal-looker).
• MCP: separate plugin dsh-plugin-mcp-support mounts native @deepseek-ai/dsh-mcp-client servers from its bundle-row config or the persisted mcp-support settings namespace.
• Structured output: separate plugin dsh-plugin-structured-output provides opencode-style /json-schema + StructuredOutput validation on native seams (no dsh-side format field). Its visibility is opt-in per preset via Settings → 结构化输出工具 (Structured output); no mode is enabled by default.
• Partial: extracted family templates keep dynamic sections filled by dsh-native data rather than omo's builder output; structured output is tool-enforced rather than tool_choice: required; hooks are regex/simplified ports; AGENTS.md injection is dsh-native; child subagents inherit the session model because dsh child headers/descriptors do not carry the subagent role id (primary-role sampling defaults ARE applied).
• One dsh-side patch in this repo. patches/0001-agent-pre-step-assistant-prefill.patch restores opencode's request-only assistant ceiling text; stock hosts keep the system-section fallback. format/toolChoice remains a proposal; the standalone structured-output plugin covers the common route.

Remaining gaps

1. maxSteps role (patched seam + stock fallback): patches/0001 gives the assistant-role request tail on a patched host. A stock 0.2.1 host still gets the same MAX_STEPS_PROMPT text and trigger as a system-prompt section, so models may treat the role differently. Tracked upstream as discussion #2407.
2. dsh-side (proposal, medium): GenerateOptions.format / toolChoice. omo's regular path does not use them; the standalone structured-output plugin covers the common route.
3. Child subagent per-role sampling cannot reliably resolve the role id (dsh child headers/descriptors do not carry it); primary-role sampling defaults ARE applied and children inherit the session model.
4. Plan files: dsh itself does not persist them; the plugin writes .opencode/plans/* after exit_plan_mode approval. A first-class plan-file seam remains an optional improvement.
5. team-mode TUI, comment-checker CLI, and hashline diff enhancer remain non-LLM/editing experience differences.