Skip to content

feat(lifeboard): the agent turn, four providers and Connections - #315

Open
ivanmkc wants to merge 1 commit into
lifeboard/03-shellfrom
lifeboard/04-agent
Open

feat(lifeboard): the agent turn, four providers and Connections#315
ivanmkc wants to merge 1 commit into
lifeboard/03-shellfrom
lifeboard/04-agent

Conversation

@ivanmkc

@ivanmkc ivanmkc commented Sep 2, 2026

Copy link
Copy Markdown
Owner

The stack

# PR What it adds Tests at this point
1 #311 A real localStorage for the test suite on Node 25 viewer 747
2 #312 @ivanmkc/termchart-canvas; injectable interact transport canvas 473 · viewer 278
3 #313 lifeboard PWA: fact log, boards, service worker lifeboard 36
4 #314 Shell, family profiles, tap layer lifeboard 82
5 #315 The agent turn, four providers, Connections lifeboard 142
6 #316 Packs, recipes, shopping-list arithmetic lifeboard 231
7 #317 Provenance and lineage lifeboard 257
8 #318 Calendar, Gmail, mail triage lifeboard 301
9 #319 The bridge and long jobs cli 314 · lifeboard 320
10 #310 Proactivity, scheduled boards, Google sign-in cli 320 · lifeboard 343
11 #320 Persistence audit fixes: multi-tab, stale builds, no-storage boot lifeboard 350
12 #321 User guide, bridge setup, QA guide, persistence audit lifeboard 350
13 #322 A plan for third-party integrations (docs only) unchanged
14 #323 A practice drill a board can define: audio + Q&A canvas 485 · lifeboard 372
15 #325 Export and restore lifeboard 389
16 #326 Draft a message, approve it, then send lifeboard 399
17 #327 Hand long work to the Mac lifeboard 406
18 #328 A Packs screen, and authoring one by describing it lifeboard 418
19 #329 Activity — what changed, who changed it, undo lifeboard 428

Every branch in the stack was checked out on its own and verified independently: npm install, tsc --noEmit for every package, npm test across the workspace, and the offline e2e where it exists. All nineteen are green — no branch depends on a later one to build or pass. Full viewer e2e (11 suites, 106 assertions) was run on #312 as the behaviour-preservation evidence for the extraction.


Stack position: 5 of 10. Base: lifeboard/03-shell.

The conversational loop: type a question, get a reply, a change you watch land, or a board.

The envelope is the whole contract — one JSON object carrying what to say, what to change and what to draw. Parsing is deliberately forgiving (models wrap JSON in prose and fences) and rejection is deliberately precise, because the error text is fed straight back as the repair prompt: mutations[1].collection beats invalid.

The repair after a bad board asks for the board alone, never the whole envelope. The mutations have already landed by then, and re-sending them would apply the change twice. There is a test for exactly this, because it is the kind of thing that silently doubles a shopping list.

Every failure path still returns something. Two unusable envelopes degrade to whatever the model actually said; two unrenderable boards degrade to a plain list built by code from the reply. A turn that throws leaves someone staring at a spinner with no idea whether their milk got added.

The model may write item, recipe, event, contact — not profile, message or derivation. A model that could rewrite the transcript could rewrite its own history.

Connections is furniture, not a board. It has secrets and text inputs in it, and the agent must never redraw the screen where an API key is typed. Keys live in device preferences and never in the log, which by design cannot forget. OpenRouter signs in by PKCE, so a router works with no backend at all.

Deviation from the roadmap: it called for bundling the 70 files in diagram-recipes/examples as few-shot. Those are engineering diagrams that teach the wrong register for a family board and would dominate the context every turn; two hand-written examples are used instead.

Verified: 142 unit tests including the full degradation ladder against a scripted provider; offline e2e still passes.

🤖 Generated with Claude Code

The envelope is the whole contract: one JSON object carrying what to say, what
to change and what to draw. Parsing is deliberately forgiving (models wrap JSON
in prose and fences) and rejection is deliberately precise, because the error
text is fed straight back as the repair prompt -- "mutations[1].collection"
beats "invalid".

The repair after a bad board asks for the board alone, never the whole envelope.
The mutations have already landed by then, and re-sending them would apply the
change twice. Covered by a test, because it is the kind of thing that silently
doubles a shopping list.

Every failure path still returns something. Two unusable envelopes degrade to
whatever the model actually said; two unrenderable boards degrade to a plain
list built by code from the reply. A turn that throws leaves someone staring at
a spinner with no idea whether their milk got added.

The model may write item, recipe, event and contact -- not profile, message or
derivation. A model that could rewrite the transcript could rewrite its own
history.

Connections is furniture rather than a board: it has secrets and text inputs in
it, and the agent must never redraw the screen where an API key is typed. Keys
live in device preferences and never in the log, which by design cannot forget.
OpenRouter signs in by PKCE, so a router works with no backend at all.

Departs from the roadmap on few-shot examples: it called for bundling the 70
files in diagram-recipes/examples, which are engineering diagrams that teach the
wrong register for a family board and would dominate the context every turn. Two
hand-written examples instead.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant