@drawbridge/drawbridge-agents 0.1.98 → 0.1.99

package/.claude-template/settings.json

@@ -21,12 +21,9 @@
       "../drawbridge-components",
       "../drawbridge-docs",
       "../drawbridge-emails",
-      "../drawbridge-games",
-      "../drawbridge-growth",
       "../drawbridge-mongodb",
       "../drawbridge-redis",
       "../drawbridge-share",
-      "../drawbridge-shopify",
       "../drawbridge-shopify-app",
       "../drawbridge-stripe",
       "../drawbridge-sync",

package/claude/CLAUDE.md

@@ -25,3 +25,4 @@
 @../conventions/walk-the-app.md
 @../conventions/feature-pipeline.md
 @../conventions/workflow-step-naming.md
+@../conventions/token-budget.md

package/conventions/cross-repo-contracts.md

@@ -107,15 +107,17 @@ skill) so it stops living only in tribal memory.
   carry `organization : null` and still block everyone;
   api registers the number's SmsUrl at twilio connect (`lib/twilio.js`) and `scripts.js
   registerTwilioWebhooks` backfills — a deployed webhooks route with unregistered numbers hears
-  nothing, silently. **Since Butterscotch (release 1219790090574059; sync 762fdd7, Darren
-  2026-10-23: an opt-out is the person's choice, not one address's) `canSend` gates the
-  PERSON, not only the address**: it reads the recipient's contact in the organization (found by
-  `canonical.emails`/`canonical.phones` or by `leads`) and blocks when any canonical or raw
-  address on that contact is suppressed for the channel. `lead` is optional, so every caller
-  that knows the lead passes it, or a returning person under a new mailbox passes until their
-  contact is rebuilt with the new address. The api's draw opt-out check
-  (`route/organization-campaign-draw.js` `optedOutLeads`) applies the same rule and must stay
-  in step with it; before Butterscotch both matched the lead's own canonical address only.
+  nothing, silently. **Every message send and every draw checks the recipient's OWN canonical
+  address, never their contact** (Darren, 2026-10-29): contacts exist only once an organization
+  activates segments, and sync merges entrants who share a phone into one contact, so a
+  contact-reading gate blocked a second entrant for the first one's unsubscribe, and only where
+  segments was on. The api's draw check (`lib/opt-out.js`, used by prepare, the draw and the
+  winner send) and `canSend` without `person` must stay in step. **Since Butterscotch
+  (release 1219790090574059) the ONE caller that reads the contact is the vendor contact sync**:
+  `lib/step-runner.js` passes `person : lead.id` for the `suppressed` answer the
+  Mailchimp/Klaviyo/Attentive hooks receive, and `canSend` then blocks when any canonical or raw
+  address on the person's contact (found by `leads` or `canonical.emails`/`canonical.phones`) is
+  suppressed. A message path that passes `person` reintroduces the shared-phone bug.
 - **`audience` picks the FROM-address; `intent` picks the unsubscribe plumbing — and NOTHING
   passes the suppression gate** (utils ↔ sync ↔ api). Every suppression source blocks every
   lead-facing send, the winner claim included: api's draw route refuses up front ("Disqualify

package/conventions/token-budget.md

@@ -0,0 +1,37 @@
+# Token budget
+
+Darren works on a subscription, and the rolling limits are metered in tokens, so a QA run or
+a full dev walk that spends fewer tokens is one that finishes before the limit lands
+(Darren, 2026-11-02: "we don't want to stop doing those things, we just want a way to use less
+tokens when we do"). Two rules, both about where the tokens actually go: tool output and
+subagents, not the prompt.
+
+## Read-only subagents run on Haiku
+
+A subagent that only locates things — `Explore`, `Plan`, a `general-purpose` agent sent to
+search or read — is spawned with `model : 'haiku'`. It returns file paths and conclusions,
+and the session model reads the files it names. Agents that judge or write keep the session
+model: `drawbridge-qa` (walks and reports), the code reviewer, anything that edits. A Haiku
+agent's "not found" is never evidence of absence under rule 6 in `rules.md`: before claiming
+nothing writes, reads or documents a thing, the session model runs that search itself.
+
+## Long output is summarised by the command, not by the model
+
+A build, test, log or list command that can print more than a screen is piped through `tail`
+for its summary AND `grep -i -E 'fail|error'` for its failures, in the same command. A bare
+`tail` hides the failing line; a bare dump spends the tokens. Before a long walk, `/compact`;
+between unrelated tasks, a new session rather than a longer one.
+
+## Locate with Serena, read with Read
+
+Serena (an MCP server, `oraios/serena`, registered at user scope) finds a symbol and its
+references through a language server. Use it to find WHERE code lives instead of reading
+whole files or paging through grep hits. It does not replace the read: rule 5 in `rules.md`
+still means the whole file is open before a claim about it, and rule 6 still means absence is
+proved by a search of the whole space. Serena only shortens the search for the file to open.
+
+Trial since 2026-11-02 on Darren's machine (install: `uv tool install -p 3.13 serena-agent`,
+then `claude mcp add --scope user serena -- serena start-mcp-server --context claude-code
+--project-from-cwd`; `**/.serena/` is in the global git ignore). The agents sync preflight
+installs it family-wide only once the trial shows file reads dropping. Serena's own README
+says never to install it from a plugin or MCP marketplace.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@drawbridge/drawbridge-agents",
-  "version": "0.1.98",
+  "version": "0.1.99",
   "description": "Shared agent-instruction content (rules, code style, conventions) for the drawbridge-* monorepo.",
   "license": "UNLICENSED",
   "publishConfig": {