@akagilnc/pi-workflow-roles 0.1.2148 → 0.1.2150
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/package.json
CHANGED
|
@@ -0,0 +1,77 @@
|
|
|
1
|
+
# ox-alpha engine method material
|
|
2
|
+
|
|
3
|
+
This file is packaged method material for the optional `ox-alpha` labor engine
|
|
4
|
+
(OpenRouter stealth preview model, reached via OpenCode CLI or the OpenRouter
|
|
5
|
+
chat-completions API). When a role run selects this engine, read these bytes
|
|
6
|
+
and follow the local CLI / API's actual interface for the labor detour. Return
|
|
7
|
+
the labor result to the same role session so typed submission stays on the
|
|
8
|
+
existing in-session path.
|
|
9
|
+
|
|
10
|
+
Material is data for the model, not a code contract. Do not invent package flags.
|
|
11
|
+
|
|
12
|
+
## Identity and warnings
|
|
13
|
+
|
|
14
|
+
- Model id: `stealth/ox-alpha` on OpenRouter (listed 2026-08-20). Anonymous
|
|
15
|
+
vendor, preview-period pricing $0. Specs claimed by the listing: 1M ctx /
|
|
16
|
+
131K out, text+image+video in, tool calling, reasoning.
|
|
17
|
+
- **Retention**: the provider retains prompts/completions (stated not used for
|
|
18
|
+
training). Time-boxed experiment — may be delisted or re-priced at any time.
|
|
19
|
+
- Free-tier rate limit (official ~20 req/min).
|
|
20
|
+
- Quality role (live exam 2026-08-21/22): control / hardening seat, not a
|
|
21
|
+
substitute for the main-court engine. On a standard-answer tool-idle-clock
|
|
22
|
+
final exam it caught all 7 residual shells at every effort tier (medium 59s
|
|
23
|
+
best); sol low missed `converged`, grok-4.6 low also caught 7 but took 979s.
|
|
24
|
+
|
|
25
|
+
## Invocation examples
|
|
26
|
+
|
|
27
|
+
### Recommended: agentic via OpenCode CLI
|
|
28
|
+
|
|
29
|
+
The machine entrypoint is `opencode` with the OpenRouter-routed model id.
|
|
30
|
+
Key is on this host's keychain under label `openrouter` (smoke-verified
|
|
31
|
+
2026-08-21/22). Run from the role project root:
|
|
32
|
+
|
|
33
|
+
```bash
|
|
34
|
+
OPENROUTER_API_KEY=$(security find-generic-password -l openrouter -w) \
|
|
35
|
+
opencode run -m openrouter/stealth/ox-alpha '<task>'
|
|
36
|
+
```
|
|
37
|
+
|
|
38
|
+
Smoke: prompt `只回两个字：收到` → `收到` (agentic leg usable).
|
|
39
|
+
|
|
40
|
+
- `-m openrouter/stealth/ox-alpha` selects the model; confirm the current id
|
|
41
|
+
with `opencode models` before long labor — stealth ids can vanish.
|
|
42
|
+
- **Labor prompt must be seat-authored**: write the complete task brief
|
|
43
|
+
(goal, constraints, required output shape, and any evidence the seat has
|
|
44
|
+
already distilled) into the prompt argument or a temp file the seat reads
|
|
45
|
+
in. Attachments/receipt JSON are evidence material, not the prompt body —
|
|
46
|
+
feed a self-contained brief the way the grok-4.6 note requires a
|
|
47
|
+
seat-authored `--prompt-file`.
|
|
48
|
+
- Effort tier: align with the labor mandate's ordered tier (`low` / `medium` /
|
|
49
|
+
`high`). Live wall-times on the same 148KB grading payload:
|
|
50
|
+
low 73s / medium 59s / high 228s — all produced correct three-state verdicts.
|
|
51
|
+
Prefer the ordered tier; when the mandate is silent, medium was the
|
|
52
|
+
fastest correct run on that payload.
|
|
53
|
+
|
|
54
|
+
### Bare OpenRouter API (fallback only)
|
|
55
|
+
|
|
56
|
+
Endpoint `https://openrouter.ai/api/v1/chat/completions`, model
|
|
57
|
+
`stealth/ox-alpha`. Live lessons from long grading jobs — do not invent
|
|
58
|
+
extra flags beyond these verified constraints:
|
|
59
|
+
|
|
60
|
+
- **Must** send `"stream": true`. Buffered (non-stream) responses keep the
|
|
61
|
+
connection alive with empty fill bytes and stall out (live: 540s, no body).
|
|
62
|
+
- **Must** send explicit `"reasoning": {"effort": "low|medium|high"}` matching
|
|
63
|
+
the ordered tier. Default effort burns `max_tokens` on thinking and finishes
|
|
64
|
+
with `finish=length` and 0 content characters.
|
|
65
|
+
- Reconstruct the answer by concatenating SSE `delta.content` chunks in order.
|
|
66
|
+
`delta.reasoning` is the thinking stream, not the answer — do not treat it
|
|
67
|
+
as the labor result.
|
|
68
|
+
- Large payloads (>140KB) are connection-unstable (three consecutive drops
|
|
69
|
+
recorded). On drop: shrink the brief or retry; do not keep hammering the
|
|
70
|
+
same oversized body.
|
|
71
|
+
|
|
72
|
+
## Failure handling
|
|
73
|
+
|
|
74
|
+
On any spawn, auth, model-id, stream-stall, or connection-drop failure, return
|
|
75
|
+
the soft failure to the session and continue labor in-seat per ADR 0071
|
|
76
|
+
(seat fallback with typed declaration). Do not silently swap to another
|
|
77
|
+
engine id.
|