# Route-Feasibility Evidence Inventory (S0, B-OG-27)

Bounded feasibility evidence for every provider target in the capability
matrix and provider registry. Issue roctinam/aiwg#2877; proposal
`11-delivery-plan-and-issue-backlog.md` §B-OG-27, `03-software-architecture.md`
§10 (scope authority), `07-stewards-and-catalog-freshness.md` and
`08-security-and-threat-model.md` (Revision 2 Tier 0).

Method (2026-10-07, host titan): free local discovery only — `--version`,
`codex debug models`, Muse model-catalog JSON, checked-in Ralph adapters
(`tools/ralph-external/lib/`), model discovery (`src/models/`) and executor
registration (`tools/ralph-external/lib/provider-adapter.mjs`). No model
calls were made; authorizing a probe budget does not run a probe.

Evidence grades: **H** = help/version inspected on this host (syntax only);
**R** = checked-in adapter/contract evidence; **D** = official
documentation/source, needs live qualification. No row claims inference,
auth, sandbox, tool-list or model qualification from help alone.

Standing rules: help text, flag acceptance and worker echoes do not qualify.
Unknown hidden fan-out or data path blocks strict use. No named target is
silently dropped. Final qualification stays with B-OG-07/08/09/28 (plus
B-OG-36–48 per §10). No target is backend-attested in this pass; every IN
row is runtime-configured-only or unavailable, pending S1 qualification.

## Alias normalization and non-provider names

Matrix `claude-code` normalizes to registry `claude`; aliases never add
providers. `openai` in `src/models/types.ts` is an inference provider, not
another harness — it resolves to the `codex` harness target. Bare `grok`
names the xAI model/API category, not a provider. `generic` is a
registry-only deployment fallback ( nineteenth entry), not an executable
provider and not a matrix row. Devin CLI is a nondeployable future surface
of `windsurf`, not an alias. Gemini is not a provider and was not added.

| Alias / related name | Canonical target | Kind |
|---|---|---|
| `claude-code`, `claude` | `claude` | alias → provider |
| `openai` | `codex` (harness); `openai` (inference) | alias → provider; separate inference name |
| `agy` | `antigravity` | alias → provider |
| `dsh` | `deepseek-harness` | alias → provider |
| `oh-my-pi` | `omp` | alias → provider |
| `pi-coding-agent` | `pi` | alias → provider |
| `devin`, `devin-desktop`, `devin-local`, `cascade` | `windsurf` | alias → provider |
| `devin-cli` | — | future-provider surface, not an alias |
| `tinyhumansai` | `openhuman` | alias → provider |
| `github-copilot` | `copilot` | alias → provider |
| `factory-ai` | `factory` | alias → provider |
| bare `grok`, `muse-spark`, `spark`, `meta` | — | not providers |
| `generic` | — | registry-only fallback, no matrix row |
| `openai` (models `Provider`) | — | inference provider, not a harness |

## Per-target feasibility

Scope IN/OUT follows 03 §10, as amended by the B-OG-47 recheck below. Every
`Budget` except the promoted `openhuman` row is **authorized by the operator
on 2026-10-07 for titan and noodle**, exactly as proposed, as a synthetic probe
ceiling (calls / tokens); OUT rechecks retain 1 call / 0 tokens (help/version
only, no inference), and generic retains 0 / 0. The `openhuman` live budget is
only **proposed** and needs operator authorization before any probe. No probes
were run. Authorization
does not establish live qualification, which stays with B-OG-07/08/09/28
(plus B-OG-36–48 per §10). `Owner` is the final-qualification issue.

| Target | Scope | Identity | Evidence channel and exact limits | Hidden fan-out / data path | Budget | Owner |
|---|---|---|---|---|---|---|
| `claude` | IN | runtime-configured-only | H: `claude -p --output-format stream-json --verbose`; terminal-result extraction; no sandbox flag inferred; `--max-budget-usd` is budget, not step cap | Unknown: Agent/Task tools, implicit fallback; Ralph adapter currently adds bypass by default so direct profile must differ; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-28 |
| `codex` | IN | runtime-configured-only | H: `codex debug models` (observed 0.160.1: gpt-6.1-sol/astra/sol, gpt-5.6-sol/terra, 272k ctx, ultra = auto-delegation) + `exec --json -o F` thread/turn/item/error stream; native `mcp-server` not advertised by installed help | Unknown: `ultra` self-delegates outside conductor budget — deny ultra; subagents unqualified; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 50000 | B-OG-07 |
| `muse` | IN | runtime-configured-only | H: catalog JSON (observed 1.4.3: muse-spark-1.3 minimal–max, 1M ctx; `-contributor` default is a policy hazard; image model hidden) + `exec --json` invocation-bound session logs; `muse serve` is MSP, not MCP; `--provider echo` is mechanics only | Unknown: subagent/delegation surface unqualified; contributor default never silently selected; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-08 |
| `hermes` | IN | runtime-configured-only | H: `hermes -z` plain final + `--usage-file` (observed 0.20.6); session-filtered log binding required, global log growth proves nothing; `-z` auto-bypasses approvals so external sandbox/tool restriction is required | Unknown: native moa/fallback chains must be disabled/qualified; conversation tools are not worker calls; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-48 |
| `copilot` | IN | runtime-configured-only | D: `copilot -p --output-format json`; scoped `--allow-tool`, never allow-all; effort via `effortLevel` settings needs isolated profile qualification; agent-definition model may override `--model` | Unknown; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-36 |
| `factory` | IN | runtime-configured-only | R: `droid exec --output-format stream-json`; existing adapter `--skip-permissions-unsafe` is prohibited in default profiles; effort/autonomy/step controls unqualified | Unknown (Missions); restricted writing blocked until controls proven; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-37 |
| `cursor` | IN | runtime-configured-only | D: `agent -p --output-format json` (+`--sandbox enabled`); do not infer `--force` is a sandbox; effort/step caps unknown | Unknown; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-38 |
| `opencode` | IN | runtime-configured-only | R: `opencode run --format json`; `--variant E` is provider-specific; permissions need a qualified profile, not `--auto`; `serve` is HTTP, not MCP | Unknown; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-39 |
| `warp` | IN | runtime-configured-only | D: `oz agent run --prompt/--model/--cwd/--profile`; effort/JSON/step caps not established; remote/cloud routes are separate | Unknown; qualify Oz as the Warp interface, not a new provider; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-40 |
| `grok-build` | IN | runtime-configured-only | R: `grok -p --output-format streaming-json`; headless options lack explicit model/effort — extend before enablement; ACP adapter is not MCP | Unknown: `--no-subagents` is flag evidence only; version-scoped behavioral evidence with model/effort pinning must prove child control; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-41 |
| `openclaw` | IN | runtime-configured-only | D: `openclaw agent exec --json`; workspace/settings mapping, effort/step controls need live qualification; never add `--fallback`; gateway path is distinct | Unknown; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-42 |
| `antigravity` | IN | runtime-configured-only | R: `agy -p` stream-json per provider-contract; fixture-qualified flags only; soft tool denials may coexist with exit zero; never add `--dangerously-skip-permissions` | Unknown; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-43 |
| `deepseek-harness` | IN | runtime-configured-only | R: `dsh --profile headless` + validated Cordis route patch (model/effort live in the patch, no guessed `--model` flag); `DSH_PERMISSION_MODE=workspace-write`, telemetry disabled | Unknown: enforcement/child control and terminal parsing unqualified; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-44 |
| `pi` | IN | runtime-configured-only | R: `pi --mode json --model/--thinking/--tools`; `--no-approve` is a safety hazard needing explicit semantics; no budget/max-turn controls; JSON mode has no stdin abort | Unknown; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-45 |
| `omp` | IN | runtime-configured-only | R: `omp --mode json --model/--thinking/--profile`; `--no-tools` for text-only; no max-turn flag assumed; do not assume Pi parser/flags are identical | Unknown; blocks strict use | authorized (operator, 2026-10-07; titan, noodle): 3 / 30000 | B-OG-45 |
| `local` (ollama / llama-server / OpenAI-compatible) | IN | unavailable | H: ollama 0.33.3 and llama-server installed but not running — stopped endpoint excluded without startup; bounded HTTP (`/api/generate` or equivalent), exact model in request body, effort null when unsupported; required profile: text-only roles and no cloud transfer (unproven) | Unknown fan-out and processing: endpoint ownership, upstream forwarding/redirects, cloud fallback, tools, telemetry and enforced egress controls unresolved; local-only whole path unproven; blocks strict use. Required: disable cloud inference/tools/telemetry and enforce egress; never start servers, download weights, or kill a shared server to cancel one request; health ≤60s old per 07 | authorized (operator, 2026-10-07; titan, noodle): 2 / 20000 | B-OG-09 |
| `windsurf` | OUT | unavailable | R: no verified headless Desktop worker command; IDE startup is not inference; MCP client config is not worker serving | Unknown; recheck B-OG-47 before any scope change | authorized (operator, 2026-10-07; titan, noodle): 1 / 0 | B-OG-47 |
| `grokbot` | OUT | unavailable | R: desktop CreateAgent/registered-machine capabilities establish no headless subprocess contract; connectors consume tools | Unknown; recheck B-OG-47 | authorized (operator, 2026-10-07; titan, noodle): 1 / 0 | B-OG-47 |
| `openhuman` | IN (promoted by B-OG-47, 2026-10-08) | runtime-configured-only | D: `openhuman-core --model M call --method openhuman.agent_chat --params-stdin` (in-process one-shot agent turn; conductor-framed params `message`, `model_override`, `cwd` on stdin until EOF; final JSON on stdout, no event stream); native `openhuman-core mcp` `agent.run_subagent` has no model/effort/workspace pin and is not a worker tool | Unknown: the orchestrator may spawn subagents (depth-capped only); effort, step-cap and approval/sandbox mapping unknown; blocks strict use | proposed (operator authorization required): 3 / 30000 | B-OG-47 |
| `flue` | OUT | unavailable | R: Flue 2.3.0 app target; no worker dispatch (#3120 D6) | Unknown; app-only, no CLI adapter | proposed (operator authorization required): 0 / 0 | B-OG-47 |
| `generic` | OUT | unavailable | Registry-only fallback, no binary or flags; custom adapters remain extensible through the existing factory | N/A — not an executable route | authorized (operator, 2026-10-07; titan, noodle): 0 / 0 | — |

Machine-readable scope/identity/channel/fan-out/budget rows for the 19
matrix providers live in `agentic/code/providers/capability-matrix.yaml`
(`route_feasibility`, typed via `getRouteFeasibility` in
`src/providers/capability-matrix.ts`). `generic` and `local` are
registry/docs-only rows by design: `generic` has no matrix entry and local
endpoints are ordinary members, not registry providers.

## S1 coverage reconciliation (B-OG-47)

Issue roctinam/aiwg#2911. One normalized row per provider-registry entry
(canonical ids only; aliases never add rows) plus the ordinary `local`
member. `Profile issue` is the S1 qualification issue for IN rows. `Evidence
date` is the date the row's cited sources were last checked; `Live` is
authenticated, version/account/profile-scoped qualification, which no row has
yet: help/version output, documentation and synthetic fixtures never set it.
Doc-only (D) rows therefore stay unqualified. Machine-readable values live
in `route_feasibility.coverage` of `agentic/code/providers/capability-matrix.yaml`
and are checked by `test/conformance/orchestration-v1/provider-coverage.test.ts`.

Upstream recheck performed 2026-10-08 with web access and local
`--version`/`--help` only (no prompts or model calls): Claude 2.1.295, Codex
0.161.0 (`codex mcp-server --help` still prints root help, so native MCP stays
version-blocked) and Muse 1.4.4 on titan; the doc-only Copilot, Cursor,
Warp/Oz and OpenClaw references still document the cited flags; the three
excluded hosted rows were rechecked as recorded below.

| Target | Scope | Profile issue | Evidence date | Live | Sources and recheck result |
|---|---|---|---|---|---|
| `claude` | IN | roctinam/aiwg#2900 | 2026-11-07 | no | H: installed help 2.1.319; compatible range `>=2.1.295 <2.2.0` (#3845); `claude-adapter.mjs` |
| `codex` | IN | roctinam/aiwg#2897 | 2026-10-08 | no | H: installed help 0.161.0; native `mcp-server` still not advertised |
| `copilot` | IN | roctinam/aiwg#2901 | 2026-10-08 | no | D: [programmatic reference](https://docs.github.com/en/copilot/reference/copilot-cli-reference/cli-programmatic-reference) still documents `-p`, `--model`, `--no-ask-user`, `--output-format=json`, `--allow-tool` |
| `factory` | IN | roctinam/aiwg#2902 | 2026-10-07 | no | R: `factory-adapter.mjs` |
| `cursor` | IN | roctinam/aiwg#2903 | 2026-10-08 | no | D: [CLI overview](https://cursor.com/docs/cli/overview) and [output format](https://cursor.com/docs/cli/reference/output-format) document `-p`, `--model`, `--sandbox`, `json`/`stream-json` |
| `opencode` | IN | roctinam/aiwg#2904 | 2026-10-07 | no | R/D: `opencode-adapter.mjs`; [OpenCode CLI](https://opencode.ai/docs/cli/) |
| `warp` | IN | roctinam/aiwg#2905 | 2026-10-08 | no | D: [Oz CLI](https://docs.warp.dev/agents/cli/oz-cli/) documents `oz agent run --prompt`, `--model`, `-C/--cwd`, `--profile` |
| `grok-build` | IN | roctinam/aiwg#2906 | 2026-10-07 | no | R: `grok-build-headless.ts` |
| `hermes` | IN | roctinam/aiwg#2912 | 2026-10-07 | no | H: installed help 0.20.6 (not rechecked in this pass) |
| `muse` | IN | roctinam/aiwg#2898 | 2026-10-08 | no | H: installed help 1.4.4 (1.4.4-R5419.1); `muse-adapter.mjs` |
| `openclaw` | IN | roctinam/aiwg#2907 | 2026-10-08 | no | D: [agent reference](https://docs.openclaw.ai/cli/agent) documents `agent exec`, `--model`, `--cwd`, `--json`; explicit `--fallback` remains prohibited |
| `antigravity` | IN | roctinam/aiwg#2908 | 2026-10-07 | no | R/D: `provider-contract.v1.json`; [headless reference](https://antigravity.google/docs/cli/headless/) |
| `deepseek-harness` | IN | roctinam/aiwg#2909 | 2026-10-07 | no | R: `deepseek-harness-adapter.mjs` |
| `pi` | IN | roctinam/aiwg#2910 | 2026-10-07 | no | R: `pi-adapter.mjs` |
| `omp` | IN | roctinam/aiwg#2910 | 2026-10-07 | no | R: `omp-adapter.mjs` |
| `openhuman` | IN | roctinam/aiwg#2911 | 2026-10-08 | no | D: promoted. Source at [216ea32](https://github.com/tinyhumansai/openhuman/tree/216ea322a3c1269dada70751ac5c08eac7dc7596) and [release v0.64.14](https://github.com/tinyhumansai/openhuman/releases/tag/v0.64.14) (standalone `openhuman-core` asset) evidence `--model` and `call --method openhuman.agent_chat --params`/`--params-stdin`; `agent` subcommands still only inspect prompts; native MCP `agent.run_subagent` has no model pin |
| `local` | IN | roctinam/aiwg#2899 | 2026-10-07 | no | H: ollama/llama-server installed, endpoint not started |
| `windsurf` | OUT | — | 2026-10-08 | no | Gap: Devin Desktop documents no headless worker command or worker-serving MCP tool. Recheck: unchanged; [Devin CLI](https://docs.devin.ai/cli/reference/commands.md) has `-p/--print`, `--model`, `--prompt-file` but stays a distinct future provider until registry support exists |
| `grokbot` | OUT | — | 2026-10-08 | no | Gap: Bots run on xAI-hosted computers from desktop/mobile chat, routines and connectors; no local CLI, headless subprocess or worker MCP tool. Recheck: unchanged per [Grok Bot docs](https://docs.x.ai/grok-bot/overview.md) |
| `flue` | OUT | — | 2026-10-30 | no | Gap: app target only; argv prompts and absent worker model/effort flags. Recheck: installed Flue 2.3.0 source; #3120 D6 excludes worker dispatch; see [Flue](flue.md) |
| `generic` | OUT | — | 2026-10-07 | no | Registry-only deployment fallback with no binary, flags or upstream interface to recheck |

Promotion consequence: OpenHuman joins S1 under this issue with an inert,
unqualified `workerHealthAdapter` pinned to 0.64.14 (degraded, process-only,
final-JSON; every control `unqualified` with its evidenced mapping), the
driver `tools/ralph-external/lib/openhuman-adapter.mjs` (installed-module
shape, approved as `installed:openhuman` by the pinned digest in
`test/fixtures/providers/openhuman/approved-module.json`), and a synthetic
`openhuman-core` under `test/fixtures/providers/openhuman/` that
`test/conformance/orchestration-v1/openhuman-direct.test.ts` loads through the
production installed-module loader, renders through the real resolver/renderer
and runs through the attached launcher (T-OG-15/20/21/23, fixture-only). The
profile uses `--params-stdin`, which reads stdin to EOF
(`crates/openhuman-core/src/core/cli.rs:445-468` at 216ea32); `stdin:
brief-json` makes the conductor frame the params document from the brief
(`src/serve/worker-stdin-frame.mjs`), so brief text never enters argv, unit
properties or run metadata. **Open S1 blocker:** authorized live
qualification; `openhuman-core` is not installed on titan or noodle and the
probe budget stays proposed. Until then it cannot dispatch.

## Source anchors used

Ralph adapters: `codex-adapter.mjs:74`, `muse-adapter.mjs:125`,
`claude-adapter.mjs:50`, `factory-adapter.mjs:69`,
`opencode-adapter.mjs:81`, `pi-adapter.mjs:45`, `omp-adapter.mjs:12`,
`deepseek-harness-adapter.mjs:17`, factory `provider-adapter.mjs:95`
(registration) and `:228` (local transport). Model discovery:
`src/models/model-discovery.ts`, `src/models/provider-policy.ts:107`
(`openai` excluded from harness policy), `src/cli/handlers/models.ts:37.
Executor/control evidence: adapter `getCapabilities`/`buildSessionArgs`
plus 03 §11.3 health extension columns.

## Findings for the conductor

1. `src/cli/handlers/models.ts:37` omits `muse` from its provider set
   despite the registry entry — consumer mismatch already noted in 03 §10;
   fix belongs to S1 (B-OG-02/04), recorded here, not changed in this pass.
2. Codex `ultra` ("automatic task delegation") and Hermes native
   moa/fallback are evidenced hidden fan-out: deny or externally bound
   before any strict use.
3. Codex ships a stale-default hazard (hidden gpt-5.5 remains the config
   default) and Muse a contributor-default policy hazard: S1 profiles must
   pass model and effort explicitly, resolved from policy.
4. No backend-attested identity exists for any target in this pass;
   backend attestation, where possible, is S1 qualification work.

## Open questions

- Who records the version/account/profile scope of each paid probe during
  S1 live qualification on the authorized hosts titan and noodle?
- Should the OUT recheck budgets (1 call / 0 tokens, help/version only)
  also cover a fake-MCP `tools/list` pin where a native server is claimed?
