Skip to content

Extend practice feedback and problem bank checks - #54

Open
jserv wants to merge 20 commits into
mainfrom
dev
Open

jserv wants to merge 20 commits into
mainfrom
dev

Conversation

@jserv

@jserv jserv commented Sep 16, 2026

Copy link
Copy Markdown
Contributor

This branch widens what a practice session gives back and tightens what the problem bank has to prove. Candidates can write their own test cases and see what their code returns, practice without a camera when the interview is not recorded, and get a post-interview debrief that states the hint rung they reached and the level they practiced for, while the report is refused if it names the published problem. The lobby keeps a longer history whose reports can be reopened, keeps due reviews across level changes, resets a failed review to one day, and shows progress by topic. The bank now accepts original exercises, every judge needs five cases including a boundary one, and a C starter that returns an array has to say who frees it. On the server, a provider whose credentials are refused leaves rotation, and report-model reviews during quiet pauses are capped per interview. CI gains a check that the release image matches the manifest's Rust version and a weekly run that requires the pinned avatar model, and the legacy migration paths for the git hooks and vendored files are removed. The interview contract moves to bundle 8, live prompt 3 and report prompt 8, with each step recorded in docs/interview-contract-versions.md.

Verified with ./scripts/test.sh at each of the 18 commits, not only the tip, and all pass; the tip runs 672 Rust tests and 523 browser tests with nothing skipped. The guarantees the commits claim were checked by breaking the guarded line and confirming a test failed: a prompt edit without a version bump, the interim review cap and its limit of 72, 401 and 403 provider refusal, the published-name check, judge case coverage, the C ownership note, and the version parity check. The candidate case fixes were reproduced before and after in Chromium. One upstream test this branch does not touch, setup_page_accepts_credentials_and_writes_the_primary_config_file in tests/cli.rs, failed once under load and passed on every rerun.

Deliberately left for follow-up: a report for an original exercise can still mention LeetCode, because the published-name check keys off a source title those exercises do not have; a sixth candidate case blocks the test run instead of being refused on its own; a null expected value breaks the palindrome checker; a mistyped class-case method breaks the whole C++ or Java compile; only four of the five candidate cases reach the reviewer; the parity check passes when neither version can be read; and the failed-review test cannot fail as written. A checkout that still holds web/vendor/avatar/jim.vrm from before the model moved to the browser cache now fails verify-vendor.sh until the file is deleted.


Summary by cubic

Expands practice feedback and report history while tightening problem-bank validation and release checks. The active interview contract is now bundle 9 with live prompt 3, report prompt 8, rubric 1, and report schema 2.

Practice and reports

  • Candidate-authored cases are type-checked, run separately from judge totals, capped at five, and reach the interviewer and report.
  • Reports now include debrief guidance, hint rung, volunteered hints, practiced level, topics, and neutral camera-skip conditions; published problem names are rejected or repaired.
  • Prompt-only edits no longer invalidate stored scores; rubric and schema versions determine compatibility, with golden coverage for prompt version bumps.
  • Full reports stay reopenable beyond the short history, due reviews survive level changes, failed reviews reset to one day, and progress groups by topic.
  • Interim report-model reviews are capped by CODETRIAL_MAX_INTERIM_REVIEWS (default 12, maximum 72), while final reports retain complete context.

Problem bank and tooling

  • Original exercises are supported, including generated rubrics and the new Bounded Event Queue problem; imported exercises keep published-source metadata out of candidate-facing data.
  • Judges must provide five executable cases with a type-valid boundary case, C array-return starters must document caller ownership, and class harnesses validate arity and derive return types from contracts.
  • Providers with fresh 401 or 403 responses leave rotation until credentials change and the server restarts.
  • CI checks Rust release-image parity and scheduled avatar model delivery; hook and vendor migration paths are removed, and stale web/vendor/avatar/jim.vrm files must be deleted.
  • CLI tests bound concurrent server processes and treat signal-killed processes as failures.

Written for commit f1ffc92. Summary will update on new commits.

Review in cubic

cubic-dev-ai[bot]

This comment was marked as resolved.

cubic-dev-ai[bot]

This comment was marked as resolved.

cubic-dev-ai[bot]

This comment was marked as resolved.

cubic-dev-ai[bot]

This comment was marked as resolved.

cubic-dev-ai[bot]

This comment was marked as resolved.

@jserv
jserv force-pushed the dev branch 2 times, most recently from 82bd84e to d6e8629 Compare September 20, 2026 11:34
cubic-dev-ai[bot]

This comment was marked as resolved.

cubic-dev-ai[bot]

This comment was marked as resolved.

@jserv
jserv force-pushed the dev branch 2 times, most recently from 64dbdb1 to 88cc4ac Compare September 20, 2026 17:23
A prompt edit used to invalidate stored scores, so every change had
to be added to a browser-side list that nothing forced anyone to
update. Compare the recorded rubric and schema versions instead: a
prompt-only bump leaves scores comparable and the list goes away. The
golden digest now covers every prompt builder, so an edit that skips
its version bump fails the gate rather than a review.
A report could name the published exercise or the practice site it came
from, which hands back the source the interview scenario deliberately
withholds. Validate against the selected problem at the exact paths
a candidate reads, and let Gemini repair a violation so one stray
title does not cost the whole report.
A verdict arrived with nothing to learn from it. Carry debrief metadata
on generated and fallback reports alike, under a second report schema
so the addition is recorded rather than inferred. Schema 1 reports
stay scored, because existing history is still the candidate's.
Spaced-repetition reviews were filtered by the level chosen for new
problems, so changing level silently hid work already due. Choose
due reviews before that filter applies. A failed review drops to a
one-day interval, which is what being due is for, and outcomes are
ordered by time first so an older failure cannot erase the successes
recorded after it.
The full history keeps twenty attempts, so an older one disappeared
from the lobby altogether rather than ageing out of detail. Keep a
second list of full reports past that cap, bounded by a local-storage
byte budget rather than a row count, and draw the lobby from both,
paired on the id every saved report carries so an attempt held in both
is listed once. A row predating that id is given one in each store the
first time it is read. The short history is written first and alone:
the report a candidate is waiting on must not lose its room to the
larger store behind it, and a write refused there costs a later reopen
rather than the save.
A candidate could only run the judge's cases, so the cases they
thought to try left no trace in the evidence. Saved typed cases run
beside the judge's without touching its totals, and up to five reach
both the interviewer and the report prompt, where testing instinct is
worth reading. A typed case is checked against the operations and
arities the exercise declares, so a malformed one is refused with a
reason instead of failing inside the runner, and a sixth is refused on
its own rather than stopping the run it was typed into.
Every bank entry had to be an imported exercise carrying a published
title and examples, which tied the bank to one study plan and left
no way to write a problem for this tree. An entry may now declare
itself original and carry no published fields: its title comes from
the variant, validation stops demanding a source, and the plan checks
pass it over. Rubrics become generated output, because the title an
entry shows is no longer a constant the bank can state.
A hint count alone cannot separate a candidate who asked once for
a deep hint from one the interviewer kept prompting unasked. Pass
the rung reached and how many hints were volunteered, and say in the
prompt that both are context: a volunteered hint is weaker evidence
than a requested one and neither is a numeric deduction.
A missing or declined camera turned a candidate away from an interview
that never needed video. Require one only when recording is enabled,
since that is what the recording notice describes. The reason the
camera went unused is signed into the integrity trail so the report
reads a neutral condition rather than absence of evidence.
Quiet-pause reviews and the final report spend the same provider quota,
and the existing rate limits still allow seventy-two reviews in a long
interview, ahead of the report that matters. Bound them per interview,
fixed when the interview starts so a config reload cannot raise the
ceiling underneath a session already running.
The pool probe removed a project only for an explicit 429, so a project
whose credential had been revoked stayed in rotation and failed later
at token minting, where the candidate sees it. Read 401 and 403 as
refusals too. Everything else still counts as available, because
refusing an interview over a brief network fault is the worse error.
A candidate practising for a senior role got the same verdict as one
practising for a junior role, with nothing on the page saying which
bar had been applied. Pass the selected level as coaching context
and show it beside the verdict. The decision stays against the fixed
mid-level bar, which the prompt now states twice because it is what
a reviewer is most likely to let drift.
Progress listed recurring weaknesses but never which topics had
actually been practised, so a topic failed twice and a topic never
attempted looked alike. Group attempts by the topics their problems
carry, with passes and the date each was last seen.
A judge could ship with two cases and none of them near a boundary,
which makes a passing run weak evidence of anything. Require five cases
and one the declared argument types admit as small at their own domain,
rather than treating a single JSON shape as a boundary everywhere. The
exception file carries its own count, so a line left behind after
its cases are authored fails instead of excusing that judge forever.
The restored index contract now admits an empty tree, because a
boundary case has to be one the exercise actually accepts.
A C starter that returns an array through returnSize says nothing
about who frees it, and a candidate reading it has no way to find
out. Require the starter comment to state that the caller frees the
returned storage.
The release workflow pinned a toolchain that nothing compared against
what the crate claims to build with, so the two could drift until
a release failed. Declare the minimum in the manifest and fail the
gate when the Linux image no longer matches it.
The avatar browser check accepts an unavailable avatar, so that a
developer without the model can still run the flow. That also means
nothing would notice the pinned model going away. Run the flow on a
schedule with the fallback disallowed, which checks the pin somewhere
without making the local run depend on it.
The hook installer and the vendor fetch carried branches for trees laid
out the way this one no longer is, and nothing left in the repository
can reach them. Require the current behavior outright. The docs go
with them: the Java lane needs a working toolchain rather than a
version nobody here tested, and the connection notes drop a claim
about which assertions moved that was never checked.
Most tests here start a real server and the harness runs one thread per
core, so a small machine was asked to hold more of them than it had
memory for and the kernel picked them off. Take a slot before spawning
and give it back when the process is reaped, charged to the thread
rather than to each server: a test that stops one and starts another
holds two, and charging both would let such tests take every slot and
then wait on each other.

Then stop reading a killed process as a clean exit. An exit code is
absent when a signal ends a process, and both runners turned that
absence into zero, so a usage error the binary refuses and a production
start it declines both read as accepted.
Upstream released bundle 6 for the interviewer countdown while this
branch was open, and this branch had already spent 6 on the debrief.
The contract says a released bundle number is never reused for
different behavior, and a stored report carries the number it claims,
so the three bundles added here move up rather than overlapping one
that already means something else. The prompt golden digest moves with
them, because the number is part of what the prompts carry.
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