Public, free to read, reuse, and adapt — Kirk Patrick Miller's notice for this edition. Not a patent grant.
FLINT_NEXT_TREE_ANY_MODEL_v0.1.txt · c02
0d55c697ae · words 1585 · characters 11170–23082
Previous section · Next section
Exact section text · Full source · Source index
1.6 Bridges and trust (why the Agent Bridge is not the road here)
- Ollama Bridge (FreeLattice bridge/proxy-core.js, 127.0.0.1:11435 default): origin allowlist
already names https://thelatticetree.com (never *). No pairing, no token. It proxies ONLY to
Ollama at 127.0.0.1:11434. LM Studio models are not reachable through it.
- Agent Bridge (FreeLattice tools/agent-bridge.js, 127.0.0.1:3141, locked by #120): not used by the
Tree at all today (grep finds no 3141 / agent-bridge in Alpha docs/). GET /models is not
token-gated but is Ollama-only; every POST (including /inference) needs a paired token
(needsToken: "every POST and any other verb"), and /inference defaults model to 'llama3.2',
the poetry route to 'qwen2.5:7b'. Its OLLAMA_BASE default is http://localhost:11434.
Routing Tree chat through 3141 would ADD a trust gate (pairing) for a purely local act and
would touch #120's lock files. Decision for v0.1: do not route through 3141. Keep the Tree on
the Ollama Bridge (11435) and direct loopback doors. Say so in the doc.
- HTTPS honesty: on https://thelatticetree.com the browser may block http://127.0.0.1:1234 unless
LM Studio's own server has CORS enabled. That is outside our code; we add one honest line.
======================================================================
2. WHAT THE HEADLESS WALK SHOWED (Alpha main + #83, fake Ollama with 7 models, fake LM Studio with 2)
======================================================================
Setup: http.server on the Alpha tree, fake 127.0.0.1:11434 (llama3.2:latest, qwen2.5:7b, qwen2.5:14b,
mistral:7b, phi3:mini, gemma2:9b, deepseek-r1:8b) and fake 127.0.0.1:1234 (two LM Studio ids).
The fake logged which model each chat POST carried. Viewport 390x844.
W1 Settings Find: two doors remembered (Ollama, LM Studio). Ollama models stored = 5 of 7.
Stars: 5 Ollama + 2 LM Studio. gemma2:9b and deepseek-r1:8b are not choosable anywhere in Settings.
model = llama3.2:latest (first), line "This mind speaks with llama3.2:latest" shown. Chat sent llama3.2:latest.
W2 Tap star mistral:7b -> entry.model mistral:7b -> chat sent mistral:7b. GOOD.
W3 Tap LM Studio door -> primary LM Studio, chat sent lmstudio-community/Meta-Llama-3.1-8B via
/v1/chat/completions. GOOD. (Ollama door kept mistral:7b in minds[].)
W4 Connect card lists all 7 Ollama models. Tap gemma2:9b -> entry becomes
{ name:'gemma2:9b', models:1, minds:[] }. LM Studio door forgotten. Settings sky re-mounted shows
ONE star (gemma2:9b). chooseModel(url,'qwen2.5:14b') returns null. Connect buttons show no chosen mark.
Chat sent gemma2:9b (the tap itself works; everything else is lost).
W5 Fresh: Connect tap deepseek-r1:8b, then Settings Find -> entry model silently becomes
llama3.2:latest with the line "The chosen model deepseek-r1:8b is no longer at this door."
That line is FALSE: deepseek is at the door; it was the 7th name and the 5-cap dropped it.
Order-dependent and it switches the speaking model under the person.
W6 Gathering seat options after W4: exactly one (gemma2:9b). After W5: 7 options, deepseek missing.
W7 tryAddress('192.168.1.77:11434') fetched GET http://192.168.1.77:11434/api/tags, /v1/models and /.
The Tree's paste field is not 127.0.0.1-only today.
Evidence screenshots (box): /workspace/tree-any-model/A_after_find_390.png,
C_after_connect_tap_390.png, D_connect_then_find_false_gone_390.png. Probe scripts: probe.py, probe2.py, fake7.py.
======================================================================
3. EVERY PLACE THAT BLOCKS "ANY MODEL, ANY ORDER"
======================================================================
T1 parseModelNames cap: local-mind-probe.js 211 "return names.slice(0, 5);"
T2 Constellation cap: local-mind-probe.js 669 "(m.models || []).slice(0, 5)"
T3 Gathering options cap: local-mind-probe.js 393 "return out.slice(0, 12);"
T4 Single remembered model on Connect tap: fl-connect.js remember() Alpha branch replaces the entry
(models:[name], minds:[], speakingChair:''). Other models and doors vanish. (W4)
T5 chooseModel refuses names not on target.models (566). Combined with T1/T4, most models are
unreachable from Settings until another Find. (W4)
T6 False "gone" + silent switch: entryFromFoundList keeps a prior model only if it is on the
(capped) list for the SAME url (priorModelForUrl 450-463, modelOnList 465-472, 478-484, 518-523).
Capped-out names and names that moved door (Bridge 11435 <-> direct 11434) are declared gone and
replaced by models[0]. (W5)
T7 Bridge-first look hides other doors: look() returns only the Bridge when helped (243-271), so a
running LM Studio is invisible whenever the Bridge is up.
T8 Door name lost: Connect writes name = model name, so the sky and thread say "gemma2:9b" where
they said "Ollama". The thread heart never names the model after a Find (garden-thread.js 505).
T9 Connect card has no chosen mark; you cannot see what speaks, or that a tap switched it.
T10 Speaking chair dropped on every Connect tap (speakingChair:'').
T11 tryAddress is not loopback-only (306-336). (W7) Lock breach against Kirk's 127.0.0.1-only rule.
T12 garden-trainer.js reads 'fl_active_model' (FreeLattice key) -> hardcoded 'llama3.2' and 'phi-pathway'.
T13 Round Table sits one model per door (round-table-sitting.js 86-100). (Later, see 9.)
T14 FlConnect probe is Ollama-only; LM Studio only appears through the Settings Find. (By design for
the shared core; fixed on the Tree by merging, see Step 3, not by teaching FreeLattice LM Studio.)
T15 Agent Bridge trust gating: not in the Tree's path today; routing through 3141 would add pairing.
Keep it out (1.6).
Hardcoded names in the Tree's docs/: garden-trainer.js 277 'llama3.2' (Modelfile default),
901 'phi-pathway'. No other hardcoded model names in the Tree's chat path. (FreeLattice-side
hardcodes in tools/agent-bridge.js are #120-locked and out of scope.)
======================================================================
4. STEP 1: local-mind-probe.js, lift the caps (T1, T2, T3)
======================================================================
File: docs/modules/local-mind-probe.js
4.1 After "var ROSTER_CAP = 4;" (line 31) add:
// v-tree-any-model-v0: every model at a door is choosable. Soft ceiling only against a runaway list.
var MODEL_CAP = 64;
var SKY_FOLD = 8; // stars shown before "Show all N"; the chosen one is always shown
4.2 parseModelNames (196-212). Layer, do not delete the old line: replace the final return with
// v-tree-any-model-v0: was names.slice(0, 5); every downloaded model is a choice.
var seenName = {};
names = names.filter(function (n) { if (seenName[n]) return false; seenName[n] = true; return true; });
return names.slice(0, MODEL_CAP);
(Keep the old line as a comment above it: // before v-tree-any-model-v0: return names.slice(0, 5);)
4.3 paintConstellation (631-715). Replace "var models = (m.models || []).slice(0, 5);" (669) with:
// v-tree-any-model-v0: all models; fold after SKY_FOLD, chosen always visible
var allModels = (m.models || []).slice(0, MODEL_CAP);
var expanded = host.getAttribute('data-sky-open') === '1';
var models = allModels;
if (!expanded && allModels.length > SKY_FOLD) {
models = allModels.slice(0, SKY_FOLD);
if (isPrimary && chosen && allModels.indexOf(chosen) >= SKY_FOLD) models.push(chosen);
}
After the cluster is built (after star.appendChild(cluster);) add a quiet toggle when folded:
if (!expanded && allModels.length > models.length) {
var more = el('button', 'settings-star-more', 'Show all ' + allModels.length + ' models');
more.type = 'button';
more.setAttribute('data-sky-more', '1');
more.addEventListener('click', function () {
host.setAttribute('data-sky-open', '1');
paintConstellation(host, mindsFromEntry(getRemembered()));
});
star.appendChild(more);
}
Keep role=radiogroup/radio, aria-checked, bindRadioKeys as they are. Model names go in with
el(..., modelName) which uses textContent: keep it that way.
CSS (garden-rooms.css, append at the end, after the last v0.4 rule):
/* v-tree-any-model-v0 */
.settings-star-models { flex-wrap: wrap; }
.settings-star-more { font: inherit; font-size: 0.85rem; background: none; border: 0;
color: inherit; opacity: 0.75; text-decoration: underline; min-height: 44px; padding: 0 0.5rem; }
Check at 390px that eight star chips plus the toggle wrap without horizontal scroll.
4.4 getChairSeatOptions (367-394): replace "return out.slice(0, 12);" with
// v-tree-any-model-v0: every remembered model can take a chair
return out.slice(0, MODEL_CAP);
Gathering picker (garden-rooms.js openPicker 1262) needs no change, but check at 390px that a long
list scrolls inside .core-bind-options (add max-height: 50vh; overflow-y: auto; under the marker if it
does not).
======================================================================
5. STEP 2: local-mind-probe.js, keep the chosen model honestly (T5, T6)
======================================================================
5.1 New helper after modelOnList (465-472):
// v-tree-any-model-v0: find a prior model at any found door (the Bridge and 11434 are the same Ollama)
function doorFamily(url) {
var u = String(url || '');
if (/\/v1\/models$/i.test(u)) return 'openai:' + u.replace(/\/v1\/models$/i, '');
return 'ollama';
}
function findModelAnywhere(minds, name) {
if (!name) return null;
var i;
for (i = 0; i < minds.length; i++) {
if (modelOnList(minds[i].models || [], name)) return minds[i];
}
return null;
}
Notes: every /api/tags door (11434 direct, 11435 Bridge, a manual Bridge port) is family 'ollama'.
OpenAI-shaped doors are their own family per base.
5.2 entryFromFoundList (474-534). Layer after "var primary = minds[primaryIndex] || minds[0] || {};"
and before the modelNote block. Intent (write it cleanly, keep the old variables):
- priorChosen = the prior entry.model (and prior primary url).
- If priorChosen is on primary.models: keep it (already the case).
- Else if findModelAnywhere(minds, priorChosen) returns a door D in the SAME family as the prior
url (for example the Bridge went away and 11434 answered): make D primary, D.model = priorChosen,
and set entry.modelNote = 'Still speaking with ' + priorChosen + ', now through ' + D.name + '.'
(a visible, true line, not a warning).
- Else if found anywhere in another family: do NOT switch doors silently. Keep the old fallback
but the line must say where it is: 'The chosen model ' + priorChosen + ' is at ' + D.name +
'. Tap it to speak with it again.' and keep primary as the old code picks.
- Only when the name is truly absent from every found door, keep today's speakModelGone line.
The false "no longer at this door" of W5 must be impossible when the name is on any found list.
5.3 chooseModel (551-572): today returns null for a name not on target.models. Layer a gentle path:
if the door is known (target found) but the name is missing from target.models, return null as
today (a stale click); if the door is NOT in minds, return null as today. No other change. The
caps and the merge (Step 3) are what make every name present. Do not widen chooseModel to accept
unknown names; that would let a stale button write a model the door does not have.
======================================================================
6. STEP 3: merge instead of replace (T4, T8, T10), shared core + Tree
======================================================================
6.1 local-mind-probe.js: new exported function mergeFound(found) (place after choosePrimary, 574-595):