Sycophancy mitigation as a structural commitment: the model writes its
full correctAnswer (plus explanation, in that order) BEFORE the learner
sees the question; the commitment rides the persisted call args, is
revealed on submit, and grades his free text in the model's next reply —
the client never grades free recall.
- tools.explain setting (segmented 'Quizzing — Multiple choice | Free
recall' in Agent settings' Tools group): free arms explain and drops
quiz for BOTH agents (prober variant carries no nodeId, same pattern
as the probe quiz); mechanics keys never pass filterEnabledTools —
quiz stays unwritable (the 2026-09-19 wedge ruling stands)
- doctrine lives pi-natively in the tool descriptions (owner ruling —
no conditional system-prompt chunks): explain carries the commitment
rule; the MC option-building craft moved out of the prompt into the
quiz description; shared/20-question-tools.md slims to the
quiz-agnostic core
- display seams: parseExplainInput (LaTeX-debris restoration included),
dock text-mode with an 'I don't know' row (kbd row 0), AnsweredCard
reveal (The answer / Why, no verdict chips), [Explain answer] format
- phases/plan seams recognize both quizzing tools' nodeId tags; explain
answers deliberately feed no ledger evidence (no client grading to
key on — a future ruling can parse the model's grade)
1049/1049 tests, typecheck/lint/build green.
The owner's capture-vs-composition check confirmed the wire sends exactly
what composes (byte-identical save the nodeId paragraph ebfe058 had since
changed). This wave removes the implementation meta that leaked into the
prompts and reorders so the method carries weight:
- base: owner's new opener verbatim (install understanding; the
understanding definition — which absorbs the old 'Understanding is
structure' paragraph); Voice moved in from shared/10 (deleted); the
philosophy keeps edge-dependence + logistics ('standard learning'
dropped per owner); intro's two-agents/fragments sentence gone.
- teacher/05-method opens with the owner's methodology bridge ('No other
teaching methods come close') — paragraph 3 of the opener, placed here
so the shared base never points the prober at methodology it doesn't
carry.
- deleted: teacher/60-probing-in-brief (self-evident from shared history;
its mid-lesson mini-probe move salvaged into 20-teaching as one
sentence), the 'Session shape — the durable artifact' section
(extraction-pipeline meta), the fragments-meta lines in shared/20 and
05-method, 'the transcript's seam' phrasing, prober's 'the teacher
that takes over' / 'the handoff is automatic'.
- composition reorder (listSourceSegments): base -> the agent's own
fragments -> shared/ appendix. Teacher method now starts at 12% in
(was 62%); prober doctrine at 34%. shared/20 opens with the owner's
decision rule (quiz only for known-correct-answer, ask for everything
else).
- HANDOFF_TEXT: the owner's user-message version — first-person, paces
the planning (unconditional truths, motivated derivations), ends at
plan_init + accept-or-changes.
Teacher 31.4k -> 28.7k chars; prober 10.0k -> 9.6k. Gates: typecheck 0,
1018/1018, lint clean, build green. (Suite runs 16s; the earlier 10-min
timeout was stale node_modules after master's df1328e 0.87.1 sync —
npm install fixed it.)
The probe intro was redundant with 60-probing-in-brief.md (and its
no-probe fallback described a state the code can't produce — the teacher
only exists post-probe). 'How the map tracks itself' folded into one
tagging paragraph; grain and rewriting compressed; the live-map context
kept to one sentence. The discovery-chain thinking block stays verbatim,
and the stress-test stays. Presenting: ask him to confirm the plan
before teaching begins — medium unprescribed (owner ruling 2026-09-27;
dock-question confirmation leaves the teach divider floated next to the
planning divider when no text reply splits the group — accepted as
cosmetic).
The inferred teach seam sat as a ternary branch in the messages.map
chain (alongside the no-bubble marker branches), so the ENTIRE first
teaching group — thinking, prose, the tagged quiz call — rendered AS
the divider and vanished below it (the Lacan session, 2026-09-27).
The marker branches replace contentless scaffolding; the teach group is
a real message. The divider now prepends inside the group's Fragment.
Also tightens 10-planning.md: the presenting turn asks NOTHING — no
quiz, no ask_user_question, no 'shall we begin?' (the teacher was
substituting a confirmation question for the dead verdict buttons).
The Accept/Decline verdict gate is gone (owner ruling 2026-09-26):
presenting the plan ends the teacher's turn; the learner's next message
is a natural reply. Teaching begins at the first quiz call tagged with a
nodeId (probe quizzes carry none) — both folds (view phaseAt3, native
phaseAt3Lh) advance on the call itself, and the teach divider renders
before that whole assistant block. No backwards compat: verdict markers,
the verdict input event, optimistic appends, composer buttons, and
projection all deleted; prompts and docs rewritten to the natural flow.
User pivot (2026-09-19): no attachments UI at all. Pasting paths is native
for technical users and alien for everyone else - so the agent gets the
tools to find what the learner merely NAMES. The learner communicates in
messages; the agent locates and reads. Zero cross-stream seams: paths
already live in tool parts, so the sources "list" is the transcript.
- ls/find/grep: pi-verbatim schemas, descriptions, and output formats
(fd's glob semantics, rg's path:line: match format, pi's truncation
notices), ported to pure Node - no binaries, no shell-out, read-only by
construction. Deltas from pi: home-anchored resolution (the harness has
no cwd), .gitignore replaced by hard noise exclusions
(node_modules/.git/Library/.Trash), 10s walk timeout with partial results.
- read: description swapped to pi-verbatim (doctrine moved to the prompt);
relative paths now resolve against the learner's home like the rest.
- prompts/shared/05-learner-files.md: one minimal doctrine fragment
composed into BOTH agents (composeInstructions grew a shared/ dir,
before the agent's own fragments). teacher/30-jobs.md keeps only the
pdf-dispatch addendum.
- ToolCallCard: the generic quiet tool-call card the user specced -
collapsed one-liner (tool name + headline argument), expandable to full
input/output, errors red, image payloads never render. Registered for
read/ls/find/grep.
- ui-message.ts: TeacherUITools entries for the four tools (mechanical).
- tests: fs-tools suite (glob semantics, pi formats, exclusions, limits),
registry/card renders, shared-fragment composition pins, tool-registry
key updates, read re-anchor.
Learner source files, on the teacher's own terms (nothing auto-injects):
- read tool (src/agent/tools/read.ts): pi-read port — path/offset/limit,
2000-line/50KB truncation with use-offset continuation notices, image
files (jpg/png/gif/webp/bmp) returned as native image content parts
(new LhToolImageResult escape hatch in lh-tool.ts). Always-on in both
the prober and teacher registries (user-decided 2026-09).
- pdf-reader sub-agent (JobKind 'pdf', toggleable, ack-and-steer): triage
via pdf_info (pdfjs text layer: per-page text/image/math-density),
targeted pdf_extract, regex pdf_search (hits count as consultation),
and pdf_render — pages painted by pdfjs onto a canvas shim in headless
Chrome (the pdfium viewer doesn't paint headless), screenshot fed back
image-in-the-loop (visual-maker idiom, providerSafePng applied). The
unread-finish gate bounces a zero-consultation brief once. Teacher gets
a text brief with page citations only; renders never leave the loop.
- MODELS.pdfReader = qwen/qwen3.8-flash (user override 2026-09-19: the
look-at-renders workload wants the vision-proven model; catalog listed
glm-5.3-flash as image-capable but qwen is the known-good daily driver).
- UI: PdfCompleteBadge + PdfBriefCard (steer-card, job-card), dispatch
card + activity-dock kind entries (type-forced, flagged for UI review).
- Prompts: 'Learner source files' section in teacher/30-jobs.md; read
line in prober/10-probe-doctrine.md. Doctrine rides tool descriptions
per the self-contained-tools principle.
Part 1 — tool availability (the settings seam, machinery generic):
- core/pending.ts: TOGGLEABLE_TOOLS (today: research, demo, visual —
the model-facing tool names), EnabledTools = the JSON-native seam
shape Partial<Record<string, boolean>> (absent key = enabled, so old
records behave exactly as today), filterEnabledTools shared helper.
quiz/ask_user_question are always-on FOR NOW (user-decided); the
filtering is generic so future toggleables (question-tool variants,
the read tool, …) need no factory change.
- createProberTools/createTeacherTools take enabled?: EnabledTools; the
composition root (chat-handler, Platform's lane this wave) passes the
resolved settings through. probe_done is re-added after the filter —
always-on by construction (it ends the phase; a prober that can never
conclude is a lockout).
- Prompts stay STATIC under toggles this wave (user-decided); standing
principle recorded: tools ship self-contained usage guidance in their
descriptions; system prompts reference only always-on tools.
Part 2 — phase-based question-writing principles (user-locked wording,
drafted with the user over three iterations):
- NEW teacher/21-question-principles.md (composes after 20-teaching):
answerable-by-reasoning-from-the-session (the fairness rule that
filters unanswerable questions without forbidding visible answers),
never-just-reciting-your-words-back, PREFER problems that apply the
knowledge (preference, not hard rule — honest recall is the fallback,
not the habit), catch-confusion-early-and-nameable (wrong options as
real misconceptions), answer-is-signal-not-verdict.
- prober/10-probe-doctrine.md: 'no rush' retired; quick-measurement
directive in — every question moves an edge (bracket, narrow, or
characterize); a bracketed strand is done.
Tests: TOGGLEABLE_TOOLS pin, prober/teacher filtering (absent = full,
false drops exactly that tool, probe_done survives), prompt pins (the
principles compose verbatim into the teacher prompt, the no-rush
framing is gone from the prober). 547/547, typecheck/lint/build green.
Two user-reported problems from the muscle-protein-synthesis session,
both diagnosed from the live data (chat_mtxeml9e_6118e59f).
The freeze: after probe_done, the open session never showed the
teacher's wake turn — only navigating away and back loaded it. Not an
agent-switch problem: the server-side handoff worked perfectly
(write-ahead persist, liveness registry, the sidebar's busy dot saw the
wake the whole time). The gap was client attach: the only mid-session
wake detector the open page has is the jobs poll, and probe-phase
sessions have zero job traffic, so that poll is OFF; the post-turn
adoption poll read turnInFlight=false at ~0ms (the wake fires at
~0.75-1.2s, after the pump's grace), adopted, and stopped. Fix: the
post-turn adoption effect now keeps a short WATCH WINDOW after adopting
(~4s, 8x500ms) — if turnInFlight flips true, it hands off to the
existing wake-adoption poll (busy shimmer, adopt at finish).
The mermaid failures: four failed plan_init calls, all caused by ONE
hidden rule — underscores are forbidden in node ids (mps_translation,
net_balance, per_meal). The id regex rejected '_' BEFORE the slug check,
so the specific, actionable message never fired; the generic 'can't read
the edge — expected a[title] --> b[title]' lied (the syntax was exactly
that), and the teacher spent three attempts hunting punctuation it
suspected (colons, commas, multi-word labels) — visible in its own
reasoning trail. Per the user's principle — enforce only the slugs,
free everything else: underscores are first-class slug chars; the id
CANDIDATE class is broad so every bad id reaches the specific slug
message; chained edges (a --> b --> c) expand to pairs; ==>, -.-> are
edge synonyms; any direction header accepted (we layout ourselves);
id["title"] quote pairs are stripped. The tool description, zod
describe, and the teacher's planning prompt now state the one strict
rule instead of the old subset liturgy.
Verified: tsc, 305/305 (the exact muscle-protein mermaid is now a
passing test), lint, build; real-browser puppeteer smoke on live GLM —
the teacher's wake turn appeared in the open page ~50s after the probe
concluded with ZERO navigation, verdict gate + first teach check +
plan window all live in-page (the single console 404 is the app's
missing favicon, environmental noise).
The accept-plan UI lands (docs/probe-teach-split.md §4, formerly the
interim rule): presenting the plan no longer moves the phase — the
learner's ACCEPT does. While the presented plan awaits its verdict, the
composer's send square becomes two stacked buttons (green Accept plan /
red Decline); typed text rides the verdict as his note, Enter keeps
sending plain side questions (phase unmoved), decline asks for a
revision and the buttons stay.
The seam: accept sends a user message carrying a data-plan-accepted
part (PlanAcceptedData in contracts/chat.ts, TeacherUIDataParts
'plan-accepted'); phaseAt3 reads it as the plan→teach event (plan_init
demoted to presentation-only); convertToModelMessages drops data parts,
so the teacher sees only the text. Mid-lesson rewrites need no fresh
verdict — acceptance happens once, at the first presentation.
The teacher's planning prompt now ends the turn at presentation and
waits for the verdict (never teaches into an unaccepted plan).
Gates: tsc clean, 304/304 (phases tests rewritten to the acceptance
seam incl. decline/side-question/failed-init cases; composer verdict
render tests), lint, build. Live GLM smoke ALL PASS: phase pinned at
'plan' after plan_init, presentation turn carried NO check, accept
click flipped the summary to 'teach', then tagged checks moved the map
(current=rates, visited=composite); zero console errors.
Diagnosis (2026-09-11 user testing, minimal-effort GLM): sessions stuck on
'plan' forever. Two causes: (1) the three-phase view read the stored tree's
currentNodeId, which nothing ever writes post author-once (2026-09-03) — a
structural dead seam; (2) the teacher stochastically left quiz.nodeId
untagged, and done-by-quiz-result semantics would have falsely completed
nodes anyway.
The redesign (docs/probe-teach-split.md) abolishes derived occupancy in
favor of explicit events:
- Two agents over one transcript: createProberAgent (quiz without a
nodeId field, ask, probe_done) and createTeacherAgent (full toolset,
plan_init onward). The chat-handler construction branch keys on a
probe_done part in the transcript. MODELS.prober added.
- probe_done: server-side seam tool — carries the learner assessment
(strands + free-form edge read), refuses while a question is open or
the probe is already concluded, and queues a handoff steer through the
generalized wake pipeline (steers now dedupe by id, not jobId; job
finalization only for job steers). The teacher auto-wakes and authors
the plan; the learner sees the assessment via the probe-done card.
- phases: probe ends at a concluded probe_done, plan ends at a
successful plan_init — strictly sequential, no skipping, no legacy
compatibility (data wiped at the pivot per doctrine).
- Plan map = position, not completion: derivePlanStatuses reads only
tagged quiz CALLS (state/outcome feed nothing); current = latest
tagged node with consecutive duplicates collapsed, visited = earlier
distinct stays (a revisit stays in both — current wins at render).
PlanTree is structure-only; currentNodeId/doneNodeIds and the dead
mutation helpers are gone. No fallback current.
- Prompts split per agent (prompts/prober/, prompts/teacher/ over the
trimmed shared base); graph-context injection removed entirely
(shelved, docs/probe-teach-split.md §5); knowledge-graph context no
longer reaches either agent.
- Plan blocks are full-only snapshots at plan_init (neighborhood
machinery retired; the persistent DAG window is the future UI wave);
usage capture unchanged.
Gates: tsc clean, 294/294 tests (phases/plan-tree/plan-blocks/question-
tools rewritten to the new semantics; probe-done + handoff-wake tests
added), lint, build. Live GLM smoke PASS: prober concluded with an honest
assessment, teacher auto-woke, authored a 10-node plan with the block
attached, then tagged checks on slope-instinct → secant derived current/
visited exactly as designed; grading honest both ways (driver off-by-one
proved a wrong answer grades false).
amosblomqvist