Public, free to read, reuse, and adapt — Kirk Patrick Miller's notice for this edition. Not a patent grant.
FLINT_HEAL_MODEL_CHOICE_SURVIVES_RELOAD_v0.1.1_tiny.txt · c01
aa5f30139a · words 1579 · characters 0–13129
Exact section text · Full source · Source index
FLINT_HEAL_MODEL_CHOICE_SURVIVES_RELOAD_v0.1.1 (tiny)
After a reload, the model she chose is still the model that answers. FreeLattice app.html, plus one
cosmetic line in the shared fl-connect.js (both repos, byte-identical).
Marker: v-model-choice-survives-reload-v0.1.1
Smoke (FL): SMOKE_OK model choice survives reload v0.1.1 (docs/scripts/smoke-model-choice-survives-reload.js)
Smoke (Alpha): SMOKE_OK model choice survives reload alpha v0.1.1 (docs/modules/model-choice-survives-reload.test.js)
Source: Hypha walk 5 of model choice sticks v0.1 (live on FL 2e74f26, Alpha b33bed4):
heals A and B PASS, Tree pick PASS, items 7 and 10 HOLD (reload turns qwen2.5 back
into llama3.2:latest while fl_modelSource=user; the chat POST then uses llama3.2).
Written by Grok Bot for Kirk, Tue Sep 29 2026, about 4:38 PM MDT.
Full diffs: /workspace/flint-pastes/FLINT_HEAL_MODEL_CHOICE_SURVIVES_RELOAD_v0.1.1.diff (FreeLattice, vs main 2e74f26)
/workspace/flint-pastes/FLINT_HEAL_MODEL_CHOICE_SURVIVES_RELOAD_v0.1.1-alpha.diff (theLatticeTree, vs main b33bed4)
Both are inlined at the end (Appendix A and B), so this paste stands alone.
Temperature: green braided gold, the cup is still where she set it down after the lights go off and on.
======================================================================
0. ORDER, MAINS, LOCKS
======================================================================
0.1 Where this lands (do not reorder):
1) FLINT_HEAL_AGENT_BRIDGE_ENV_v0.1.1_tiny.txt (landed, #124)
2) FLINT_HEAL_MODEL_CHOICE_STICKS_v0.1_tiny.txt (landed, FL #125 / Alpha #84)
3) FLINT_NEXT_TREE_ANY_MODEL_v0.1.txt (Flint is building it now, from the new main)
4) THIS heal (FL app.html + shared fl-connect.js) <- after Tree, before the boot probe
5) FLINT_NEXT_CONNECT_BOOT_PROBE_F1_F2_F3_F4_v0.1.txt (updated in place, see 6)
6) FLINT_NEXT_120_DOCS_LAYER_v0.1_tiny.txt
Start this from main AFTER the Tree PRs merge. Anchor on the quoted TEXT, not the numbers.
0.2 Mains checked just before writing:
FreeLattice main = 2e74f26 (CI primer commit on 54b2fb1 #125; app code = 54b2fb1)
theLatticeTree main = b33bed4 (#84)
docs/modules/fl-connect.js identical on both mains (blob 2e02eae).
"main N" line numbers are on 2e74f26. "post-Tree ~N" is from a faithful stand-in of the Tree
paste's fl-connect.js changes (the real Tree merge may move a few lines).
Re-check:
git -C ~/FreeLattice fetch origin && git -C ~/FreeLattice log --oneline -3 origin/main
git -C ~/FreeLattice-Alpha fetch origin && git -C ~/FreeLattice-Alpha log --oneline -3 origin/main
cmp ~/FreeLattice/docs/modules/fl-connect.js ~/FreeLattice-Alpha/docs/modules/fl-connect.js
0.3 Applies (verified, plain `git apply`, offsets only, no fuzz):
- FL diff on current main 2e74f26. OK
- FL diff on main + Tree stand-in (Tree fl-connect.js changes + a Tree
ledger entry above "## Out of scope for this brick"). OK
- Alpha diff on b33bed4. OK
- Alpha diff on b33bed4 + Tree stand-in: code (fl-connect.js, new test) OK
The sister-doc hunk appends at the end of CONNECT_PORT_PICKER_SISTER_v0.md; if the Tree
paste appended there too, `git apply --exclude='docs/library/*'` and add the section
by hand at the end of the file (text in Appendix B).
- The boot probe paste's Appendix A (app.html) and Appendix B (fl-connect.js) still apply on
top of this heal (see 6).
0.4 Locks: Layer, never delete (the removed lines are replaced in place; the old words are kept in
one home, see 3.4). Root sw.js byte-identical to main; docs/sw.js untouched. index.html ==
docs/app.html (copy after every app.html edit). fl-connect.js byte-identical across repos. No
conflict markers. Every inline <script> parses (53 of 53). textContent for new words (the toast
is showToast, which uses textContent; no new innerHTML). No bare * (not in app code, not in the
smoke: CORS echoes the page origin, CDP Fetch has no wildcard pattern). No confirm(). No
approval screens. 127.0.0.1 only (the smoke server binds 127.0.0.1; every other request is
refused inside Chrome). No em dashes in anything added. Ledger entry with a Temperature line.
Soft paste stays: Named five stay five. Family uncapped. Quiet Room shut.
======================================================================
1. THE CAUSE (confirmed: reproduced by booting the real app.html in headless Chrome)
======================================================================
1.1 Hypha's trace (walk 5, runtime/trace_reload.out; setItem stacks on reload):
before reload ['qwen2.5:latest', 'user', 'qwen2.5:latest']
fl_ollamaModel=llama3.2:latest at Object.set (L:36238) | at runConnectionCascade (L:39783)
fl_modelSource=auto at Object.set (L:36240) | at runConnectionCascade (L:39783)
fl_ollamaModel=llama3.2:latest at Object.set (L:36238) | at applyReady (L:34555) | at probe
fl_modelSource=user at Object.set (L:36240) | at applyReady (L:34555) | at probe
after ['llama3.2:latest', 'user', ..., chat POST llama3.2:latest]
1.2 Why (read in code on 2e74f26, and seen the same way in the new smoke with FL_SMOKE_TRACE=1):
Path 1: on reload, init restores the saved provider and emits providerConnected, so
runConnectionCascade (main 39774) asks FLActiveModel.isUserChosen() BEFORE
FLActiveModel.init() (main 33799) has run. _source is still null, so the answer is "no",
and step 1 writes the first text model (llama3.2) as 'auto'.
Path 2: FLActiveModel.init() then reads that overwritten value. FLBoxPointer.init() reads the
saved fl_ollamaHost, probes, and applyReady (main 34555) calls
FLActiveModel.set(models[0], ..., 'user'): the first model is saved AS HER CHOICE.
From here on even v0.1's guard in FLAutoModel.onTabChanged protects the wrong model.
#125 only guarded FLAutoModel.onTabChanged; v0.1's smoke "fresh reload from storage" called
init() first by hand, so it never met the real boot order. The new smoke boots the real page.
1.3 Grep sweep for every other path that picks a model by itself (FLActiveModel.set callers,
fl_ollamaModel / state.ollamaModel writers, BrowseModels):
- FlOnboarding connect (main 27744) models[0] -> guarded
- AiSetup testOllama (main 28377) data.models[0] -> guarded
- settingsSetMode Local (31654, 31686) Ollama / LM Studio first model -> guarded
- flTestLMStudio (main 34750) set(models[0], 'lmstudio', 'user') -> guarded, never 'user' for a pick
- flAutoConnect (main 34830) models[0] -> guarded (see 4.1)
- flProbeLocalAI LM Studio (34890) set(models[0], 'lmstudio','boot') -> guarded
- autoDiscoverAI (39689, 39699) best.models[0] -> guarded
- runConnectionCascade (39774) first text model -> guarded
- FLBoxPointer applyReady (34555) models[0] as 'user' -> guarded
Left alone (a person's own tap or typing): Settings dropdown / quick select, ModelSwitcher,
selectLocalModel, flPlayConnect / flLookConnect, flAutoPull, wizardFinish, custom endpoint,
OpenRouter / Hugging Face pickers, FLAutoModel's home-model restore. sendMessage copies the
#ollamaModel field (which follows FLActiveModel). BrowseModels only pulls and lists; it never
sets the model.
======================================================================
2. WHAT CHANGES (one sentence each)
======================================================================
a) FLActiveModel reads the saved choice the first time anyone asks (set, get, isUserChosen,
honor), so it no longer matters which boot path runs first.
b) One guard, FLActiveModel.honor(models, fallback, provider) -> { model, source, gone }, and
a short form flHonorModel(list, fallback, provider): her model wins when it is on the list
(or the list is empty: a quiet answer never replaces anything).
c) Only when her model is gone: fall back with the existing toast words (now in one home,
FLActiveModel.noteGone), fl_modelChoiceGone set, the pick saved as 'auto'.
d) An automatic pick is never saved as 'user'.
e) Heal B card: one Look again, not two (shared fl-connect.js).
======================================================================
3. THE STEPS (text anchors; exact code in Appendix A)
======================================================================
3.1 docs/app.html, window.FLActiveModel (main ~36229). After
var _source = null; // 'user', 'auto', 'boot'
add the marker comment, `var _ready = false;` and
function ensure() { if (!_ready) init(); }
In set(): first line after `if (!model) return;` add `ensure();`.
Replace
function get() { return { model: _model, provider: _provider, source: _source }; }
function isUserChosen() { return _source === 'user'; }
with the same two functions calling ensure() first, then add _names, _onList, honor and
noteGone (Appendix A, verbatim). In init() add `_ready = true;` as the first line. The return
line gains `honor: honor, noteGone: noteGone`. Right after the closing `})();` add
function flHonorModel(list, fallback, provider) (Appendix A).
_onList matches "qwen2.5" to "qwen2.5:latest" and a bare family name to its tag, like the
rest of the app.
3.2 FLBoxPointer applyReady (anchor: FLActiveModel.set(models[0], (typeof state !== 'undefined'
&& state.provider) || 'ollama', 'user')): replace that one try-line with the honor block: set
hA.model with hA.source, skip set when gone (noteGone already set the pick).
3.3 runConnectionCascade step 1 (anchor: `if (isLocal && typeof FLActiveModel !== 'undefined' &&
!FLActiveModel.isUserChosen()) {`): step 1 now runs unless her choice belongs to another
provider (_mineElsewhere). Inside, the `FLActiveModel.set(textModel.name, 'ollama', 'auto');`
pair becomes: honor -> 'user' keeps hers (log "Keeping the chosen model"), gone -> noteGone
already spoke, else the old auto set and log. So the Cascade also checks "still here?" at boot.
3.4 FLAutoModel.onTabChanged gone branch (anchor: `safeSet('fl_modelChoiceGone', current.model);`):
the six lines (flag, set auto, toast, refresh Settings) become
FLActiveModel.noteGone(current.model, pick, 'ollama');
Same words, same 7000 ms, same flag, same Settings refresh. They now live in noteGone.
3.5 The older paths (anchors in 1.3), each a one-line wrap with flHonorModel(list, fallback,
provider): FlOnboarding modelName, testOllama, settingsSetMode (Ollama and LM Studio
firstModel), flAutoConnect autoName, flProbeLocalAI lmPick (source 'user' only when the pick
is hers, else 'boot' as before), autoDiscoverAI (Ollama and OpenAI-compatible).
flTestLMStudio: `FLActiveModel.set(models[0], 'lmstudio', 'user');` becomes the honor block
(her LM Studio model when it is there; otherwise the first model as 'auto').
3.6 cp docs/app.html index.html
3.7 docs/modules/fl-connect.js (BOTH repos, byte-identical). Anchor (main 595 / post-Tree ~631):
html += '<button type="button" class="flc-btn flc-primary flc-look-again" data-flc-look-again>Look again</button>';
becomes
// v-model-choice-survives-reload-v0.1.1: the quiet card above already has its own Look again (one button, not two)
if (!report.manualQuiet) html += '<button type="button" class="flc-btn flc-primary flc-look-again" data-flc-look-again>Look again</button>';
(manualQuiet is the state where the heal-B card "Your Bridge answered on N, but no mind is
running behind it yet" and the "Nothing answered on N" card each draw their own Look again.)
Then: cp ~/FreeLattice/docs/modules/fl-connect.js ~/FreeLattice-Alpha/docs/modules/fl-connect.js && cmp both.
3.8 Tests (Appendix A / B):
- NEW docs/scripts/smoke-model-choice-survives-reload.js (FL, see 5).
- docs/scripts/smoke-model-choice-sticks.js: ONE added line in the Bridge-no-mind block,
after the 'no false "Nothing answered"' assert: exactly one data-flc-look-again.
- NEW Alpha docs/modules/model-choice-survives-reload.test.js (static guard on the shared line).
3.9 Docs and ledger (layer):
- FL docs/library/FRACTAL_FAMILY_LEDGER_v0.md: the entry in Appendix A, placed directly ABOVE
"## Out of scope for this brick" (after the Tree entry). It carries the Temperature line.
- FL docs/library/CONNECT_PORT_PICKER_v0.md: append the "Heal: model choice survives reload
v0.1.1" section at the end.
- Alpha docs/library/CONNECT_PORT_PICKER_SISTER_v0.md: append the short sister section.