Life Links — public implementation and recorded client results
Life Links gives a person durable, private records for selected places, objects, routines and calendars, and lets authorized AI clients read and maintain that context. The existing public repository is the implementation exhibit. This companion supplies a focused reading route and two recorded results that explain what changed for users.
The source reference is the public September 3, 2026 revision 64fa3decd70c2a05c8aa9ec577bf8b389f855245. Its commit and projection metadata were read successfully on September 16. This curation did not change the repository, deployment or contest submission and did not run the application again.
1. A public source path from user interface to durable records
Start with the public repository at the selected revision. Its source-projection manifest records the source revision and the standalone-layout transformations. The repository is a concrete application implementation, with an MIT license and separate challenge-work description.
A useful technical reading path is:
| Surface | Inspectable source | Question it answers? | |---|---|---| | Human interaction | Workspace controller | How are user actions translated into application operations and reflected in the workspace? | | Delegated agent behavior | Remote agent operations | How does the AI client's operation use owner-scoped records, revisions and application rules? | | Durable persistence | PostgreSQL store | How are identity, ownership and coherent reads maintained in the database? |
The product separates two ways to connect an agent: page-bound WebMCP while an eligible authenticated page is open, and separately authorized remote MCP access across conversations with the site closed. Both work against the application's records. That relationship is the important implementation story: the AI interface extends the same private context the person uses.
2. Recorded use from real AI clients with the site closed
The retained September 3 qualification record distinguishes actual client operations from discovery counts and authored tests. The following is a curated account of that record, with private conversation, account and test-object identifiers removed. The rows refer to their respective historical releases; they are not a new combined execution on the selected public snapshot.
| Recorded sequence | Observed result | Engineering significance | |---|---|---| | Codex 0.153.0: authorization, closed-page and fresh-conversation operations, cancellation/deletion/replay, revocation and reconnect | The bounded record/Collection/Routine lifecycle passed, including a new-process closed-page readback | Durable authorization and application state supported work beyond one browser conversation | | Ordinary ChatGPT: create/read a private record, then retrieve and edit it in a fresh conversation | After a transport-session admission defect was repaired, actual guide, inspect, update and post-write inspect succeeded with all Life Links pages closed; a later owner-page observation showed the saved edit | The investigation isolated session pressure from missing consent or lost application data, then verified the actual client path | | Later ChatGPT confirmation path: prepare one deletion, cancel, then prepare and confirm the same intended scope | Cancel preserved the item; confirmation deleted it; replay returned the terminal result and inspection reported the item unavailable | Confirmation was tied to application effects, and repeat calls did not perform a second deletion |
These are recorded host-specific functional results. The confirmation exercise did not establish ChatGPT sandbox/CSP enforcement or every host/deletion kind. A working confirmation interaction and the host's isolation policy are separate properties.
3. Removing a concrete Collection-loading bottleneck
A 48-member Collection initially caused 48 individual membership lookups during startup. The engineering change returned each page's bounded membership context with the members already being read. Selection and hydration share a short read-only database snapshot; the change retains continuation for overflow rather than treating the first page as complete.
These are two exact excerpts from postgres-store.ts at the public revision. The intervening code selects the member page and hydrates its records; the linked full source preserves that context.
// Selection and full hydration must see the same title/membership state.
// This short read-only snapshot takes no owner-wide mutation lock.
await client.query("BEGIN ISOLATION LEVEL REPEATABLE READ READ ONLY");
if (!(await loadCollection(client, userId, collectionId))) {
await client.query("COMMIT"); return null;
}
// Narrow O(member-count) metadata selection preserves the exact existing
// Unicode comparator and identity cursor; no SQL collation substitute.
const metadata = await client.query(
`SELECT ll.id, ll.title FROM collection_memberships m
JOIN life_links ll ON ll.id = m.life_link_id AND ll.owner_id = m.owner_id
WHERE m.owner_id = $1 AND m.collection_id = $2`, [userId, collectionId]);After member-page hydration:
const membershipPages = page.includeMemberships
? await this.readLifeLinkMembershipPages(client, userId, items.map((item) => item.id)) : undefined;
await client.query("COMMIT");
return { ...selected, items, ...(membershipPages ? { membershipPages } : {}) };
} catch (error) { await client.query("ROLLBACK"); throw error; }
finally { client.release(); }The public file bytes match the frozen source used to curate these excerpts. The implementation makes the consistency boundary inspectable. Historical performance evidence comes separately from the retained before/after measurement tables:
| Matched condition | Before median, ms | After median, ms | Before API requests | After API requests | |---|---:|---:|---:|---:| | Desktop, cold | 1,239 | 702 | 55 | 7 | | Desktop, repeat | 1,265 | 548 | 55 | 7 | | Mobile-width, cold | 1,177 | 538 | 55 | 7 | | Mobile-width, repeat | 1,207 | 517 | 55 | 7 |
Each condition had three observations, using the same signed-in browser, unchanged 48-member fixture and matched first-content observation method. The after measurements belong to public release d82345da79450fc51cc1a98716564ffc0633b600; the later selected source retains the mechanism. Requests fell about 87%, while first-content medians improved 43–57%. The mobile-width conditions ran on the same host, not physical phones. Observations with a selected-item query caused additional normal reads and were excluded from the matched headline comparison.
The tables survive in the committed testing record. The original before/after measurement JSON files are not present in the inspected working copy, so this companion provides retained recorded results rather than a raw-trial download. No new measurement was performed.
Contribution and use of this companion
Justin created and directed the product, architecture, integration, evaluation and release decisions, with extensive Codex collaboration on implementation, testing and correction. The public source makes the implementation inspectable; the selected client and performance records show where integration failures were found and what bounded results were obtained.
For an initial review, inspect the public source route and the loading change, then use the existing Life Links CV entry for the broader application story. A fresh demonstration is optional and would be separately dated. No new publication or contest change is required to inspect the already public source.