Skip to content

Open-source SEO, GEO, and marketing skills for AI agents.

feat(goals): scope observation gates to the actual experiment surface

Yuting Zhong3e8e47feb03e
diff --git a/notfair/src/server/goals/identity.ts b/notfair/src/server/goals/identity.ts
index be8d8fe4..6734d4a2 100644
--- a/notfair/src/server/goals/identity.ts
+++ b/notfair/src/server/goals/identity.ts
@@ -121,9 +121,12 @@ order:
an open action's observation window are UNTOUCHABLE, and the tick
brief also lists resources gated by OTHER agents in this workspace:
treat those exactly the same. No reverts, no double-downs, no "small
- tweaks". If every promising move is gated, do research (and log it
- as a research action) or end the tick honestly: "nothing to do until
- <date>" is a good tick, thrashing a live account is not.
+ tweaks". A gate covers exactly the resources listed on the action —
+ nothing more: it blocks re-touching THOSE resources mid-observation,
+ never unrelated or additive work elsewhere. If every promising move
+ really is gated, do research (and log it as a research action) or
+ end the tick honestly: "nothing to do until <date>" is a good tick,
+ thrashing a live account is not.
4. **One meaningful move, maximum.** Log it with \`log_goal_action\`
(expected effect + review_after_hours + spend_usd when it commits
incremental spend), then execute with your platform tools. One
@@ -132,6 +135,15 @@ order:
brief shows your logged spend total vs. the envelope; if the right
move would cross it, don't act — suggest \`amend_goal\` to the user
in your summary instead.
+
+ Scope \`resources_touched\` to the exact experiment surface — the
+ campaign, ad set, page, or specific external repo whose metric
+ response the window protects. Your own ledgers, local databases,
+ notes, and workspace files are bookkeeping, not experiments: never
+ list them as resources (a whole-ledger gate silently blocks every
+ future move that needs to read it). Additive outreach — e.g. a
+ listing PR to an external repo — gates only that one repo; it never
+ blocks new submissions to other targets.
5. **Close the diary.** End with a SHORT summary the user can read in
five seconds: the metric read, what you reviewed, what you did, when
you'll know if it worked.
@@ -142,8 +154,13 @@ Pick the window by what the platform needs to show a real effect:
- pausing obvious waste (zero-conversion spend): 72–120h
- bid / budget changes: 120–168h (learning phases reset on edits)
- new keywords, creative, audiences: 168–336h
-When in doubt, longer. Reading noise as signal is the main way this
-loop fails.
+- additive outreach (e.g. a listing PR to a third-party repo): the
+ window covers only that repo while its PR is in review —
+ \`register_pull_request\` tracks the review itself
+When in doubt, longer — EXCEPT for bookkeeping: updating your own
+ledger or notes is not an experiment and takes no window at all (log
+it as a \`research\` or \`decision\` action). Reading noise as signal
+is the main way this loop fails.
### Changing the code (website / codebase mutations)
diff --git a/notfair/src/server/goals/tick.ts b/notfair/src/server/goals/tick.ts
index 9b083292..3183abe2 100644
--- a/notfair/src/server/goals/tick.ts
+++ b/notfair/src/server/goals/tick.ts
@@ -225,7 +225,9 @@ export function buildTickMessage(ctx: TickContext): string {
}
lines.push("");
- lines.push("## Actions still gated (their resources are UNTOUCHABLE)");
+ lines.push(
+ "## Actions still gated (the LISTED resources are untouchable — a gate never blocks unrelated or additive work elsewhere)",
+ );
if (ctx.gatedActions.length === 0) {
lines.push("- none");
} else {