Add Pi and OMP harness providers

[santoshkumarradha]

Takeover note (2026-08-27)

This PR was taken over and reworked. aforge stays the default harness provider — that was the maintainer decision, and this branch no longer changes it. The branch was merged with main (no rebase) and the OMP-default logic was removed: Pi and OMP are now additional providers you select explicitly or via AGENTFIELD_HARNESS_PROVIDER, nothing more. Provider-less calls resolve exactly as they do on main.

The four review findings from the last review round are fixed:

Finding	Fix
TS pi/omp results carried no failure classification (failureType / returnCode), unlike every other TS provider	3ae5157d
projectDir was only honoured by pi/omp, so the schema-output protocol disagreed with the other TS providers	ba76a4c5
af doctor --probe passed the prompt positionally instead of over stdin the way the pi/omp adapters do	944b388a
Plan mode narrowed the tool list but sent no approval flag, leaving the run on the CLI's default (possibly interactive) behaviour	bdbaf471

Follow-up commits on top of those keep af doctor surveying aforge itself, stop --probe reporting a silently broken pi/omp install as ok, and drop the stale "OMP default" wording from docs and tests.

---

Product outcome

AgentField can now run Pi and OMP as first-class coding harnesses from Python, TypeScript, and Go. aforge remains the zero-configuration default; Pi and OMP are selected explicitly per call, through HarnessConfig, or via AGENTFIELD_HARNESS_PROVIDER. A production-shaped Go example fans one task out to Pi and OMP concurrently and joins their typed results in a single AgentField workflow.

This closes the practical parity gap with OpenCode: applications use the same AgentField harness contract for model selection, project root, system prompt, tools, permissions, session resume, structured output, retries/timeouts, and normalized usage metrics.

Developer experience

• Provider resolution is identical in every SDK and unchanged from main: per-call override → agent harness config → AGENTFIELD_HARNESS_PROVIDER → aforge.
• Model resolution is per-call override → agent config → the selected CLI's own configured default.
• Explicit calls such as provider="opencode" or provider="pi" are unchanged.
• Usage attribution records the resolved provider, so Pi and OMP runs report harness=pi / harness=omp across Python, TypeScript, and Go.
• Missing Pi/OMP executables produce actionable errors with the exact upstream install command.
• af harness doctor recognizes both CLIs, reports binary/version/auth/usability, and lists aforge first as the default.
• The README, provider guide, architecture design, shipped AgentField skill, and Go example document the providers and their operational lifecycle.

Architecture

• Adds a shared Pi-family adapter in each SDK with thin Pi/OMP flavor mapping instead of duplicating orchestration logic.
• Dispatches one-shot prompts through stdin and consumes JSON event streams, avoiding interactive CLI behavior.
• Maps AgentField's provider-neutral options to native CLI behavior:
  • model + #variant → --model + --thinking
  • project root → process cwd for Pi; --cwd plus process cwd for OMP
  • plan/auto → read-only tool filtering; only OMP takes a native approval flag (--auto-approve), since Pi has none
  • resume → Pi --session; OMP --resume
  • tool names → Pi find; OMP glob
• Reuses the existing isolated schema-file protocol and retry/validation orchestration used by the other harnesses.
• Normalizes sessions, turns, tokens, cache tokens, cost, model, duration, messages, return code, and failure classification into the existing result types.
• Keeps runtime installation outside SDK execution. Like Codex, Gemini, and OpenCode, operators install/pin the CLI in the host or image; AgentField only discovers, verifies, and invokes it. No application call mutates production hosts.

Parity assessment

Pi and OMP have feature parity with OpenCode at the AgentField contract boundary. Native flags are intentionally not identical: each adapter translates its CLI's approval, resume, project-root, and filesystem-search vocabulary. Those differences do not leak into application code.

Covered contract surface:

• model and reasoning variant
• project/cwd isolation and environment overrides
• system prompt and normalized tool allowlist
• plan and auto permission modes
• resumable sessions
• structured schema output with validation/recovery
• timeout and retry controls
• normalized token, cache, cost, session, turn, model, and duration metrics
• binary overrides and actionable availability failures

Installation and operations

npm install -g --ignore-scripts @earendil-works/pi-coding-agent
curl -fsSL https://omp.sh/install | sh
af harness doctor --provider pi,omp --json

The doctor check is static and does not make a paid model request. Containers should install the selected CLI during image construction and run the doctor as a startup/CI gate.

Go workflow example

examples/go_agent_nodes/cmd/harness_duo registers:

compare
├── pi_worker   (explicit Pi)
└── omp_worker  (explicit OMP)

compare starts both branches concurrently and joins their structured results. It defaults to openrouter/minimax/minimax-m2.7 and accepts openrouter/google/gemini-2.5-flash for a faster smoke path.

Verification

• Python: focused harness/types/usage suite passed; Ruff passed.
• TypeScript: 81 files / 832 tests passed; typecheck and production build passed.
• Go SDK: harness suite passed; changed agent/default/usage tests passed; example build passed.
• Control plane: CLI and embedded-skill tests passed; embedded skills are in sync.
• CI was green on the pre-takeover head: 41 active checks passed across Go, Python, TypeScript, control-plane, Docker, local/Postgres functional tests, performance, aggregate/patch coverage, CodeQL, README links, and CLA (with only the expected AI-label job skipped).
• Live doctor: Pi 0.84.1 and OMP 17.2.15, both authenticated through OpenRouter and usable.
• Live AgentField execution: run_20260812_164308_d2tasl3y / exec_20260812_164308_fp15z9if succeeded in 19.4s with openrouter/google/gemini-2.5-flash; compare, pi_worker, and omp_worker all completed with structured output and native cost/token metrics.
• Earlier MiniMax validation also succeeded for both branches: run_20260812_162041_fw7j8b54 with openrouter/minimax/minimax-m2.7.

Compatibility

No default-provider or model-default change: provider-less calls resolve exactly as on main (explicit → AGENTFIELD_HARNESS_PROVIDER → aforge).

Explicit provider configurations are otherwise unchanged.

[github-actions]

Performance

SDK	Memory	Δ	Latency	Δ	Tests	Status
Python	9.0 KB	-	0.31 µs	-11%	✓	✓
Go	217 B	-23%	0.60 µs	-40%	✓	✓
TS	524 B	+50%	1.59 µs	-20%	✓	✗

⚠ Regression detected:

• TypeScript memory: 350 B → 524 B (+50%)

[github-actions]

📊 Coverage gate

Thresholds from .coverage-gate.toml: per-surface ≥ 84%, aggregate ≥ 85%, max per-surface regression ≤ 1.0 pp, max aggregate regression ≤ 0.50 pp.

Surface	Current	Baseline	Δ
control-plane	87.50%	87.40%	↑ +0.10 pp	🟡
sdk-go	93.00%	92.00%	↑ +1.00 pp	🟢
sdk-python	94.38%	93.73%	↑ +0.65 pp	🟢
sdk-typescript	91.60%	90.42%	↑ +1.18 pp	🟢
web-ui	84.76%	84.79%	↓ -0.03 pp	🟡
aggregate	85.77%	85.75%	↑ +0.02 pp	🟡

✅ Gate passed

No surface regressed past the allowed threshold and the aggregate stayed above the floor.

[github-actions]

📐 Patch coverage gate

Threshold: 80% on lines this PR touches vs origin/main (from .coverage-gate.toml:thresholds.min_patch).

Surface	Touched lines	Patch coverage	Status
control-plane	84	96.00%	✅
sdk-go	254	92.00%	✅
sdk-python	0	—	➖ no changes
sdk-typescript	137	94.00%	✅
web-ui	0	—	➖ no changes

✅ Patch gate passed

Every surface whose lines were touched by this PR has patch coverage at or above the threshold.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[santoshkumarradha]

Addressing the review findings on this PR.

Code changes (commit 973913eb):

• Go tools[:0] in-place filter — sdk/go/harness/pi.go plan-mode branch now builds the read-only tool list as a fresh slice (make([]string, 0, len(tools))) instead of aliasing the tools backing array. This matches the Python and TS providers, which both build a fresh list, and removes the footgun if normalizePiTools is ever changed to return a caller-owned slice.
• Go negative return code — sdk/go/harness/pi.go now distinguishes a signal kill (ReturnCode < 0) from a plain non-zero exit, reporting Process killed by signal N. to match the Python provider and the existing gemini.go / opencode.go Go providers. Previously a SIGKILL was folded into the generic non-zero branch with a misleading "exited with code -9" message. Added a pi_test.go case pinning the signal-kill message.

Release-note callouts (PR body, "Compatibility" section):

• Called out the provider→OMP default change: provider-less harness calls now run OMP (a live model call) instead of raising No harness provider specified. Callers who relied on that error to catch a misconfiguration are affected.
• Called out the Python HarnessConfig.model sonnet→None default change: Python callers who omitted model previously got the Claude-specific sonnet fallback; they now get the selected CLI's configured default. Go/TS had no model default and are unaffected.

Not changed (reviewed, left as-is to keep the diff minimal):

• The TS factory's unreachable throw (nit; dead code, harmless).
• The harness_duo early-return leaving a branch in flight (nit; buffered channel, no leak, demo example).
• The mocked-CLI flag-coverage gap (nit; consistent with the other CLI providers in this repo).

Verification: go build ./..., go vet ./harness/, and the full sdk/go test suite pass after the changes.

[AbirAbbas]

Heads up that this can't merge as-is any more: main shipped aforge as the harness default (#905, #927-#929) along with an AGENTFIELD_HARNESS_PROVIDER env tier and a provisioning path (af aforge ensure, curl installer, docker images), so the 25 conflicts here are a genuine either/or on "what is the default" rather than anything mechanical — I didn't push a merge because resolving them means picking a winner.

Two calls worth making before we resolve: (a) whether the zero-config default is aforge or omp, and separately whether a provider-less call should silently start a paid run at all versus keeping the loud error this PR removes; (b) the Python HarnessConfig change, where provider goes from required to defaulted and model goes from str = "sonnet" to Optional[str] = None — that second one is a real break for anyone reading config.model as a string.

Pi and OMP as additional providers look good to me and I'd take them regardless; the default question is the only thing actually blocking. Four code notes below. Also note the earlier round's flag on harness_duo's early return isn't a bug — the channel is buffered to 2, so nothing leaks.

[AbirAbbas]

Merged this with main instead of rebasing and kept aforge as the default harness provider — Pi and OMP land as additional providers you pick explicitly or through AGENTFIELD_HARNESS_PROVIDER, so provider-less calls resolve exactly as they do on main. The four findings from my last round are fixed: TS pi/omp results now carry failureType and returnCode, all six TS providers resolve their working directory through one projectDir ladder, af doctor --probe drives pi/omp over stdin the way the adapters do, and Pi's auto-mode approval flag is gone because the real CLI rejects every variant of it. Verified with the per-surface CI-exact gates, a defaults proof in all three SDKs, and real-binary doctor and error-path runs.