Adapt the bounded probe-completion foundation from PR #130 by aksalatdev (source commit 535202bebe038fa1e0648f5df809d5ccfc26d071), not its mixed stream-salvage implementation.
Require a non-failed call followed by a matching probe_done result with neither isError nor explicit details.ok:false. Preserve older clean acknowledgments, saved attempts, and all existing handoff/retry/provider policies. Share the native completion seam with phase detection and exclude explicit refusals from UI phases.
Verify retained failed-probe Retry recovery with real agent/tool composition and synthetic engines: one fresh conclusion, one handoff, then teacher selection, without stored-tool replay. Focused tests 58/58; full tests 2473/2473; typecheck and full ESLint pass in isolated storage with scrubbed provider credentials.
Co-authored-by: aksalatdev <73862605+aksalatdev@users.noreply.github.com>
Each turn runs claude -p with stream-json in and out and maps its events onto pi's assistant message stream. A process that stops on tool calls is kept alive and continued when pi sends the results back, instead of replaying the whole transcript. API credentials are removed from the CLI environment so turns always bill to the signed-in subscription.
A localhost-only MCP HTTP server, guarded by a per-turn bearer token, lists pi's tools and holds each call until pi has run it and returned the result.
- switch the math authoring/rendering contract from \(...\)/\[...\] to the
standard $...$ inline / $$...$$ display: prompts (20-question-tools,
40-formatting), singleDollarTextMath on the renderer, legacy reader kept,
and escape-debris recovery extended to dollar spans (mirroring micromark
delimiter-run rules)
- exact answer membership now precedes debris cleanup in
coerceCorrectAnswer (calculus incident: an exact sentence-final-period
label was stripped as model debris, then rejected); option keys are
passed at the boundary/grading seams, the dock no longer pre-coerces
- a complete, phase-correct JSON-escaping example rides the probe/quiz
tool descriptions (serialized from one canonical body; quiz leads with
type+nodeId), plus the one-sentence JSON-escaping rule in 40-formatting
- ask option descriptions render markdown/LaTeX in the pending dock;
settled cards keep the details line and option descriptions
retroactively, read from the persisted call arguments (chat passes
part.input to AnsweredCard)
1566/1566 tests on both navigator configs; typecheck, lint, production
build, standalone packaging, and Electron-runtime smoke all green.
Demos execute JS and need a browser; the app's own binary now provides
one: main.cjs boots a hidden window + CDP port under --browser-service
(before the single-instance lock - a helper is deliberately a second
instance of this binary), lifetime tied to the server's stdin (EOF =
quit, no orphans). headless.ts connects over CDP as PRIMARY (owner
ruling 2026-10-10), system Chrome stays the fallback, graceful refusal
when neither. Electron's CDP has no Target.createTarget, so the helper's
pre-created window is THE page, reused and reset per job behind a
serialized queue. PDFs and SVGs never reach a browser anymore (native
since stages 1-2); stale Chrome wording in model-facing tool copy made
engine-neutral. Live-proofed end-to-end from the dev tree (helper seen
by process watch, js-ran 42 probe, no orphan). Gates 1523/1523.
Visual-maker SVGs never execute script, so a browser added nothing but a
process. resvg rasterizes in-process now (white background and fitTo
width mirror the old Chrome-page semantics; svgViewport unchanged,
quirks included for parity). The old Chrome path stays verbatim as the
fallback for exotic-SVG parse failures - it was the only path before,
so nothing regresses.
The resvg binary (4MB per platform, one per build) rides the standalone
via the same tracing pattern. Gates 1519/1519; native path pinned
against real SVGs (intrinsic size, viewBox-only 900 default, opts
override, system-font text).
The browser was never the PDF engine - pdfjs always painted the pages;
Chrome only hosted the canvas tab. Now pdfjs-dist v6's own Node
integration (@napi-rs/canvas, its optionalDependency) hosts in-process:
same paint code, Skia canvas, zero browsers, zero temp files, no base64
hop. The whole pdf-reader runs browser-free - the Chrome gate on
pdf_render is gone.
The Skia binary (27MB per platform, exactly one per build - npm installs
only the building machine's) rides the standalone via the same
outputFileTracingIncludes pattern as pdfjs-dist. Verified: builds green
(1515), native render pinned against the real fixture (canonical
single-IDAT PNGs, dpi/72 arithmetic, page-range skipping), and the
canvas loads + paints from inside .next/standalone.
Stages 2-3 (resvg for SVG, the --browser-service demo helper) follow.
THE START BUG (owner report, reproduced live both kinds): new sessions
(?prompt flow — practice and teach alike) sat dead until a second
message, which then persisted WITHOUT the first. Root cause, proven
both directions: the submit POST rode the view's consumeAbort
controller; React StrictMode's simulated remount fires Diet A's
unmount cleanup (model.stop()) ~8ms in and killed the POST mid-flight;
the page's sentInitial guard had already flipped, so it never retried.
Teach was equally broken — only unnoticed, because every teach chat in
use predates the diet and practice starts are always fresh. Fix: the
submit law — POSTs are not abortable by view teardown (they are ~100ms
acks since Diet B and cannot starve the pool); after the ack, a
consumeAbort identity check guards the stream attach. Pinned with the
stop-mid-submit regression test. Live: StrictMode on, practice draft
POST 200/67ms, stream attached, message persisted, problem posed.
PRACTICE PROMPT (00-base): the opening rule — material arrives two
ways, both first-class (picked propositions primary; free-form —
topics, course material, files/PDFs — fully supported, never refused);
as soon as material exists the reply poses, at most one short line
before the problem call, no warm-up interview. Plus the hard law: no
problem is ever conversation — every problem through the tool (no
thinking field/no reveal/no verdict attaches to prose).
GRAPH-CONTEXT PREAMBLE: custom markers flatten to user-role on the
wire, so the old "Contextualization — the learner picked…" read like a
system broadcast in the learner's voice channel. Now the bracket
convention: "[Session context] — picked from the learner's knowledge
map…".
SIDEBAR: busy rows unify to "Working…" for both kinds (the spinner's
default label included); "Practicing…"/"Teacher is working…" retired.
Gates: 1490/1490, typecheck, lint. doctrine's stale doc pointer fixed.
The tools_fails_2 diagnosis (glm-5.3 dropping trailing required fields):
- probe-phase quiz/explain renamed to probe/recall — a shared name carried
the probe call shape into teaching (the first teach quiz dropped nodeId);
display seams accept the old names forever
- teach quiz+explain schemas lead with the required pair (type, nodeId);
explain gains the predict/check type and joins the reveal law (checks
collapse, predictions arrive open; no-type records render check-style)
- plan_init/probe_done results go context-slim (owner, context diet):
content = validated nodeIds + binding-id reminder / fixed ack; details
keeps the tree for the popup and {ok:true} — serializer-verified that
details never reaches the model; failure echoes dropped
- the cards carry the tool's own name (Probe/Recall) and the quiz type
reads on the label (Prediction/Check) in composer and transcript
The whitehead incident (2026-10-09): glm-5.3 emitted quiz type with
quotes baked in ('"predict"') — rejected ten times, byte-identical
retries, the lesson never started — and correctAnswer with the JSON
key-value colon glued on ([':parmenides']). Third GLM wobble incident,
and each prior heal was wired only where its incident landed; this one
landed precisely where the practice heal never reached.
Layer 1 — healEncodedArguments moves to core/argument-heal (the display
seam needs it too — the tool-parts law) with withArgumentHeal +
healToolRecord (agent/tools/heal) wired at EVERY registry: prober,
teacher, practice, bareQuestionTools (it had no prepareArguments at
all), demo-builder, and submit-agent (composing after
healStringifiedArrays; the auditor's extraTools wrapped).
Layer 2 — stripModelValueDebris (core/value-debris): leading/trailing
colons, quotes, backticks, whitespace strip at the MEMBERSHIP seams —
coerceCorrectAnswer (boundary check + display + grading in one),
normalizeQuizType (persisted wobbled records keep their role),
resolveTargetId (copy-syntax on targets).
Layer 3 — schema description hygiene: the quiz type and practice
mode/outcome/rung/missClass values now read as bare words — the quoted
JSON-key style invited quote-copying.
Display mirror — parseQuizInput heals the raw record on read, so a
boundary-healed quiz whose record keeps raw args renders instead of
becoming an invisible pending quiz (learner lockout).
Pins: tests/tool-heal.test.ts (the incident shapes end-to-end through
pi-ai validation, the wiring, the strips, the mirror). The stringy-
options pins flip to boundary-parity per tool-parts' own law; the
chat-flow garbage pin keeps its honest-reject intent with an unhealable
shape. 1471 tests, typecheck, lint green.
Limits out (owner rulings): the 'Typically 5-15 nodes' cap and the
'keep the map small' pressure are dead — a plan is as elaborate as the
subject needs; what keeps it readable is correct structure (unconditional
truths as the hanging points), not compression. Labels stay short; size
is free. The method section now carries the goal reiteration: thorough,
deep, real understanding — sessions run as long as the understanding
needs, plans as elaborate as the subject requires; no schedule, nothing
to rush. (Code was already limit-free: no maxTurns, no parser caps.)
The adaptation + detour law, merged into one section ('The map moves
with him'): when the ground shifts mid-lesson — a missing foundation, a
deeper confusion — never rush the repair; no wall of text that 'covers'
the ground to get back on plan. Update the map if the structure changed,
then build the missing ground the only way anything is built here: one
reasoning step at a time, verified, at the session's one and only pace.
Detours are the teaching, not interruptions to get through.
The probe enters at the top (owner-observed failure: eigenvectors, then
high-school chain rule, then partials — strands opened at textbook
floors). New entry-point rules: open near the hardest thing the goal
needs and let a miss walk you down; strands inform each other (a strand
crushed at height says neighbors start high); the verdict + confidence +
thinking field TOGETHER set the next question's difficulty — confident
correct jumps sharply, hesitant correct steps gently. Efficiency is
resolution: never descend into already-bracketed territory; the
questions you save buy the specific, nuanced ones that locate the edge.
The one emphasis device for load-bearing ideas: GitHub-style alert
callouts. The prompt teaches only > [!IMPORTANT] (one device, one
meaning — an unconditional truth laid down, a payoff landed, a hinge
definition) with the restraint law: everything else stays in prose, if
everything is boxed nothing is. The renderer accepts all five GitHub
markers and boxes them identically — a stray [!NOTE] still meant 'this
deserves emphasis', and a marker line must never leak as text.
Rendering (owner rulings): SHARP corners, white stroke, no background,
no label, no icon — just the outline — and the box hugs its text
(width: fit-content capped at the column) instead of stretching
full-width into empty space. Same grammar as the plan blocks: this is
structure, not chrome.
Seam: the Streamdown blockquote override in md-blocks. parseAlertMarker
is pure and exported — Streamdown wraps blocks in its own styled
components, so detection reads the hast node's tagName, and a standalone
marker drops its <br> hard-break artifact along with the empty
paragraph. Plain blockquotes pass through untouched.
Rider (owner, same round): tables hug their contents — width:
max-content on the data-streamdown table (width:auto would still
stretch, the classic table trap); wide ones keep the md-blocks
overflow-x scroller. Graceful in exports: a callout is still a
blockquote there.
6 render pins: the box mounts with white stroke + fit-content + no
rounding, all five markers box, standalone-marker drops the empty
paragraph, multi-paragraph structure kept, plain quotes untouched.
Accuracy (shared, one law across every agent): the target shifts from
surface facts to understanding itself — 'the deeper risk isn't a wrong
name or date, it's a wrong understanding: an interpretation of an idea
that seems solid and isn't.' The bar is the doubt itself: checks are
cheap, and weighing whether a doubt is worth checking is how errors slip
through. Thinking stays the agent's own.
Research (teacher 40-jobs): floor framing — verify your understanding
before it becomes his; dispatch before the fact crosses your lips, teach
around the gap until the brief lands. Planning (10-planning): planning is
the natural moment for verification; the map is a hypothesis and the
probe only its first sketch — rewriting is a normal, common event (the
session adapts to what teaching reveals); and plan_init IS the
presentation — never draw the plan again in prose (the double-plan bug).
New shared segment 16-mental-model.md: the thinking fields are the
session's primary data — read them in full, measure what they reveal,
adapt everything on the model they build. Composes into teacher, prober,
and practice (shared-fragments list updated).
Practice: the rarely-needed framing is dead — fact checks get their own
law (a wrong model solution grades the learner against a lie; pose once
the brief lands), files and map readers become plain side tools.
LaTeX: the question-tools craft list gains the point-of-use bullet
('LaTeX everywhere — plain-text math in a question is a craft failure'),
the formatting law names EVERY field, and the practice problem statement
line carries it too.
Prompt pins updated: felt-doubt bar + mental-model doctrine in both
teacher and practice compositions, no-double-plan rule, rewrite-is-
normal law, rarely-needed framing asserted gone.
web_fetch_pdf downloads (25MB cap, probe-validated) into a per-job temp
dir and returns the triage summary — page count, size class (the pdf
doctrine's <=10/<=40 thresholds), flagged math/image-heavy pages. The
researcher's toolset then gains pdf_info / pdf_extract / pdf_search over
that dir, scoped by a resolve-and-startswith wall: the researcher's
filesystem ends at the temp dir. Render is deliberately absent — the
researcher slot has no vision; visual reading stays the pdf job's job.
web_fetch: a PDF URL now returns a POINTER to web_fetch_pdf instead of
dead-ending, and every output carries 'Source: <final url after
redirects> (HTTP <status>)' — the 'ill-formed request' lesson: a
200-served error body must be distinguishable from a real page.
The runner owns the temp lifecycle (mkdtemp before, recursive cleanup in
the job's finally — never enters the record) and web_fetch_pdf urls ride
the sources ledger. Doctrine step 5 added: fetch, triage, extract only
the pages that matter, cite as (p. N).
Tests build a REAL minimal PDF programmatically (correct xref offsets —
pdfjs parses it like any download): zero network, real parse.
URL.protocol carries a trailing colon ("https:", WHATWG); the colon-less
comparison (a direct port from the py harness, where urlparse().scheme
has none) rejected EVERY valid URL ever fetched — the tool never fetched
a page in its life. Owner-reported via OpenRouter logs: 8 valid URLs in,
8 'invalid URL' results out.
The new pin stubs globalThis.fetch so the happy path runs hermetically —
a valid https URL must REACH the fetch stage (and ftp:// still rejected).
dist rebuilt (desktop:dist) — bundle verified: "http:"!==t.protocol
Owner ruling (2026-10-06): every concluded problem opens a pause. The
agent may speak its closing words about the learner's last thinking
after the conclude result, then ENDS ITS TURN — a system-prompt law
deliberately, per owner correction, NOT tool-level terminate. The next
problem is never posed until the learner explicitly moves on; the pause
is the session's teaching space (repairs, clarifications, follow-ups —
still surgical, never re-teaching).
The move-on signal is a first-class input: { type: "next" } appends a
practice-next marker (content = single-sourced NEXT_PROBLEM_TEXT — the
model's unambiguous go-ahead: pose per the selection rules or wrap if
the arc is complete). Honest 400s: practice sessions only, nothing
active (mid-problem input rides the dock), riding after any pending
dismissal synthesis exactly like text.
The fast path (owner approved: composer-band chip + Enter-on-empty):
- Composer.enterAction gains "next" — Enter on an EMPTY composer in
the pause fires the signal; typing anything makes Enter a send again;
busy (steer/stop) outranks the pause.
- The Next chip hovers at the composer band's top-right (pick-pill
chrome register) whenever the pause is live and a problem has ever
been posed; the placeholder becomes the hint ("Ask about this
problem — or press Enter for the next one…", exact);
- the transcript renders the signal as a compact centered NEXT PROBLEM
line (mono-caps chrome); session-core appends it optimistically.
CONCLUDE_ACK_TEXT rewritten to carry the law; doctrine §VIII gains the
pause law (owner-sanctioned edit #3); 00-base's conclude/dismissal
bullets rewritten + the between-problems section added.
- shared/10-voice.md (agent-neutral) + shared/15-accuracy.md (research-tool
truth contract) compose for teacher, prober, AND practice
- PRACTICE_SHARED_FRAGMENTS grows beyond 40-formatting.md
- practice 00-base.md drops its stale near-copy ("You have no research tool
here" was false since round 10); the solve-and-shortcut-hunt line folds
into the problem paragraph
- teaching system-prompt.md slims (Voice/Accuracy sections removed)
- pins updated: practice-agents composition, prompt-overrides segment mirror
A session whose material is not in the map is a SUPPORTED mode now:
target is optional (absent = untagged problem, poses freely), an
explicit [] heals away at the seam (the record's canonical form is
tagged-or-absent), and when targets ARE present the round-9 law is
unchanged — whole-graph validation, names normalized to ids. The
rejection message teaches the exit ('omit target entirely'), so a
model inventing targets for an off-graph topic converges to omitting
on the retry; omitting always works, no brick is possible. Prompt
carries the clause: material in the map → ids; not → omit, problems
stand on their own.
amosblomqvist