open-research-protocol 0.4.31 → 0.4.33

package/docs/AGENT_LOOP.md

@@ -13,10 +13,15 @@ read:
 - Read `llms.txt`.
 - Run `orp about --json`.
 - Run `orp hygiene --json` before long delegation, after material writeback,
-  before API/remote/paid compute, and whenever dirty state grows unexpectedly.
+  before remote side effects or unbudgeted paid compute, and whenever dirty
+  state grows unexpectedly.
   If it reports `dirty_unclassified`, stop long-running expansion and classify
   the paths, refresh generated surfaces, canonicalize useful scratch, or write a
   blocker before continuing.
+- Do not hard-stop solely because an OpenAI research lane is paid. Budgeted ORP
+  research may run when `orp research` spend preflight is within the configured
+  daily cap; stop for missing required spend policy, missing secret, cap
+  exhaustion, purchases, or non-ORP paid compute.
 - If the task benefits from fresh concepting, tasteful interface work, or
   exploratory reframing, run:
   - `orp mode nudge sleek-minimal-progressive --json`

package/docs/ORP_AUTONOMY_PROJECT_COMPILATION_MODEL.md

@@ -91,7 +91,7 @@ A true gate is not "the agent feels uncertain."
 
 A true gate is a boundary like:
 
-- spend or purchase
+- unbudgeted spend or purchase
 - outreach or counterparty contact
 - provider/vendor selection with real consequences
 - legal/oversight/compliance judgment
@@ -196,7 +196,7 @@ Bad candidates for automatic compilation:
 
 - vague strategic narratives with no runnable command
 - tasks that imply counterparty contact
-- tasks that imply money
+- tasks that imply unbudgeted money
 - steps that promote support-only outputs into authority
 
 ## What ORP Should Emit
@@ -230,6 +230,7 @@ The controller benchmark experiment surfaced the exact shape:
 2. compile the remaining pre-outreach tasks
 3. keep drafts unsent
 4. stop only when the next step would actually contact a counterparty or spend
+   outside a configured budget/preflight policy
 5. emit a gate dossier
 6. resume only after the human opens that gate
 

package/docs/RESEARCH_COUNCIL.md

@@ -102,6 +102,26 @@ printf '%s' '<openai-key>' | orp secrets keychain-add \
   --json
 ```
 
+## Spend Policy
+
+The OpenAI research lanes are paid, but paid does not automatically mean human
+hard stop. ORP treats them as budgeted provider calls when `openai-primary` has
+a local spend policy and the lane passes spend preflight.
+
+Set or update the local daily cap metadata like this:
+
+```bash
+orp secrets keychain-spend-policy openai-primary \
+  --daily-spend-cap-usd 5 \
+  --dashboard-spend-cap-status unconfirmed \
+  --dashboard-url https://platform.openai.com/settings/organization/limits \
+  --json
+```
+
+Live research should stop when the required spend policy is missing, the secret
+is missing, the daily cap would be exceeded, or the work is unbudgeted provider
+spend outside ORP research lanes.
+
 ## Fixtures
 
 Provider outputs can be attached without spending live calls:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-research-protocol",
-  "version": "0.4.31",
+  "version": "0.4.33",
   "description": "ORP CLI (Open Research Protocol): workspace ledgers, secrets, scheduling, governed execution, and agent-friendly research workflows.",
   "license": "MIT",
   "author": "Fractal Research Group <cody@frg.earth>",