Let candidates lead the Algorithm and Coding steps by Jerry-Chiang-23 · Pull Request #462 · sysprog21/codetrial

Jerry-Chiang-23 · 2026-10-24T10:28:45Z

Jim always volunteers the next step once a candidate stalls during Algorithm or Coding, even for interviewers who expect the candidate to lead those two steps themselves.

Add a candidate-led toggle: a lobby checkbox carried through participant metadata to a new prompt policy scoped to the Algorithm and Coding steps only. Enabled, Jim asks an open question and waits for the candidate to state the next step instead of proposing one; Repeat, Example, Test and Optimizations are unchanged.

This is a live-prompt policy change, so it bumps LIVE_PROMPT_VERSION and INTERVIEW_CONTRACT_BUNDLE_VERSION to 32 and 41, refreshes the prompt golden's recorded digest, and adds a row to docs/interview-contract-versions.md.

Closes #115

Summary by cubic

Lets candidates lead the Algorithm and Coding steps during their interview. Jim previously always volunteered the next step when a candidate stalled during those two steps; with a new lobby checkbox, interviewers can opt the candidate in, and Jim instead asks an open question and waits for the candidate to state the next step.

Carries the toggle through participant metadata into a new live-prompt policy scoped to Algorithm and Coding only; Repeat, Example, Test, and Optimizations are unchanged.
Bumps the live prompt to version 32 and the interview contract bundle to 41, and refreshes the prompt golden.

Screenshot: the new lobby checkbox.

Closes #115.

^{Written for commit 1dd3254. Summary will update on new commits.}

jserv · 2026-10-24T10:37:01Z

            .expect("metadata is an object")
            .insert("codeExecution".to_string(), Value::Bool(false));
    }
+    if request.get("candidateLed") == Some(&Value::Bool(true)) {


Nothing tests this path. candidateLedInstructions calls the prompt builder directly, so if the flag is dropped anywhere between the lobby, the token request, the signed metadata and parse_participant_metadata, everything stays green and the checkbox does nothing. Copy the exampleDrawing cases already in tests/web/token.rs (literal true kept; false, "true" and 1 dropped), tests/runtime.rs and tests/browser/lobby.test.js.

Added tests/web/token.rs::token_carries_candidate_led_only_when_true, tests/runtime.rs::candidate_led_is_explicit_in_both_modes, and three lobby.test.js cases, mirroring the exampleDrawing coverage.

jserv · 2026-10-24T11:56:13Z

+fn candidate_led_nudge(enabled: bool) -> &'static str {
+    if enabled {
+        " If this stall is in the Algorithm or Coding step, ask what they are \
+thinking of doing next without naming or steering toward one yourself, and \


When Repeat, Example and Algorithm are all recorded and the editor is untouched, silence_nudge already says "ask them to begin writing their code in the editor.", and this sentence is added after it in the same event. So during Coding the model gets two instructions at once: name the next step, and don't name or steer toward one. When candidate_led_enabled is set, pick the open question instead of the unwritten branch rather than appending to it. Add a golden case for that state as well: silenceCandidateLed uses a default RuntimeState with no steps recorded, so it never reaches this conflict.

When Repeat and Example are both recorded and the editor is untouched, silence_nudge now substitutes the open question for naming the next step instead of appending it after; the hedged candidate_led_nudge wording is kept only for the states where Rust can't tell the current step from RuntimeState alone (code already written, so the stall could be in Coding, Test, or Optimizations). Added silenceCandidateLedBeforeCoding, a golden case that hits exactly the state you named.

jserv

Rebase latest main branch and resolve conflicts.

Jerry-Chiang-23 · 2026-10-24T14:40:03Z

Rebased onto main and resolved the conflicts from the Pause-removal PR taking bundle 40/live prompt 31; this branch is now clean against main.

jserv · 2026-10-24T14:48:17Z

        interview_mode: boot.interview_mode,
        code_execution_disabled: boot.code_execution_disabled,
        example_drawing_enabled: boot.example_drawing_enabled,
+        candidate_led_enabled: boot.candidate_led_enabled,


Nothing tests that this assignment and the one in candidate_bootstrap carry the setting into RuntimeState. The new test in tests/runtime.rs calls bootstrap_with_rounds directly, and the token tests stop at parse_participant_metadata. If either line is dropped, the suite still passes, yet opted-in candidates get the ordinary silence nudges. tests/unit/livekit.rs already checks example drawing this way (candidate_bootstrap then initial_runtime_state, asserting state.example_drawing_enabled). Add candidateLed true, false and absent to that test, or add a test beside it.

Added candidate_led_initial_state_carries_signed_choice beside the example drawing test, calling candidate_bootstrap then initial_runtime_state with candidateLed true, false and absent across both modes.

jserv · 2026-10-24T14:54:47Z

Add screenshots to the pull request description to illustrate your work, as @ColtenOuO suggests. Example: #349

jserv

Check CONTRIBUTING.md carefully. Never use subjects such as fixup! Let candidates lead the Algorithm and Coding steps.

Improve git commit messages per https://cbea.ms/git-commit/ :

Run make hooks in advance.

jserv · 2026-10-24T15:10:46Z

+    // the open question in place of `unwritten` instead of appending it, and
+    // the trailing append below is skipped only in that same untouched-editor
+    // state, where it would otherwise double up on the substituted text.
+    let candidate_led_certain = state.candidate_led_enabled


phases_evidenced only says some non-skipped Repeat and Example evidence exists. It does not say the candidate has left Example. Take a candidate who walks through the ordinary case, gets an Example entry, and then stalls before the boundary case with the editor untouched. candidate_led_certain is true, so the nudge swaps "ask for whichever of their understanding, example, or planned algorithm they have not explained" for the open "what next" question and also drops the hedged clause. That changes Example, which the description promises stays the same. The same happens when a candidate goes back to Example after starting Algorithm. Add Algorithm to the check (Repeat, Example and Algorithm all evidenced, editor untouched, means the stall is in Coding) or keep the hedged form here, and add a test with only Repeat and Example recorded while the conversation is still in Example.

Added Algorithm to the check: candidate_led_certain now requires Repeat, Example and Algorithm all evidenced, matching what unwritten itself requires, so the open question only replaces the hedged wording once the stall is certainly in Coding. With only Repeat and Example recorded the hedged form still runs, covering a candidate back in Example or about to start Algorithm. Added silenceCandidateLedInExample, a golden case with only Repeat and Example recorded, and renamed silenceCandidateLedBeforeCoding's evidence to add Algorithm so it still hits the certain case.

Jerry-Chiang-23 · 2026-10-24T15:25:16Z

Added a screenshot of the new checkbox to the description.

Jerry-Chiang-23 · 2026-10-24T19:28:57Z

Squashed the fixup! commits into one (git rebase -i, git reset --soft) and rebased onto main, which now carries #473. The single commit follows Chris Beams' rules: imperative, capitalized, 50-char subject, blank line, body under 72 columns explaining what and why. Ran install-git-hooks.sh to pick up the updated commit-msg hook and confirmed it passes locally before pushing.

jserv

Remove AI markers such as Co-Authored-By:.

Jerry-Chiang-23 · 2026-10-25T05:51:02Z

Removed the Co-Authored-By: trailer and amended the commit (4ddbd97, force-pushed). Diffed the amended tree against the previous commit — it's empty, so only the message changed, no code moved. Also scanned the full diff for any other AI markers (co-authored-by, generated with, etc.); none found elsewhere in the code, docs, or tests.

jserv · 2026-10-25T14:18:09Z

+        } else {
+            format!(" They have not written code yet: {unwritten}")
+        }
+    } else if candidate_led_certain {


The arm for an editor that was touched but not written (If the editor holds none of their code, ... Otherwise: {with_code}) has no case in tests/golden/prompts.json: silenceCandidateLedWriting is the written arm and silenceCandidateLedBeforeCoding the untouched one. Replacing this arm with the plain unwritten text leaves the suite green. Add a case with Repeat, Example and Algorithm recorded and code that differs from the starter but is not counted as written.

Added silenceCandidateLedTouchedNotWritten: Repeat, Example and Algorithm recorded, the editor one character over the starter (too little to count as written). Also fixed the same evidence-vs-active-step gap you flagged below in this branch, not just the untouched one, so its text leads with the scoped "if this stall is in the Algorithm or Coding step" clause too instead of the plain open question.

jserv · 2026-10-25T14:18:46Z

+    // appending it, and the trailing append below is skipped only in that same
+    // untouched-editor state, where it would otherwise double up on the
+    // substituted text.
+    let candidate_led_certain = state.candidate_led_enabled


candidate_led_certain treats "Repeat, Example and Algorithm have evidence and the editor is untouched" as proof the stall is in Coding, but phases_evidenced checks accumulated evidence, not the active step. A candidate who goes back to Example with a concrete concern before writing code still satisfies it, so the nudge swaps in the candidate-led open question and drops the hedged unwritten wording. That applies the new policy during Example, which the bundle 41 entry says is unchanged. Keep the "if this stall is in the Algorithm or Coding step" scope on this branch unless the active step is known independently, and add a golden case for a return to Example after all three phases are recorded.

Agreed, and fixed: this branch (and the touched-but-not-written one below it) no longer substitutes the open question outright. Both now lead with "if this stall is in the Algorithm or Coding step, {open question} Otherwise, {the usual unwritten/with_code text}", so a return to Example gets exactly the wording it would without candidate-led enabled. Added silenceCandidateLedReturnedToExample: same evidence as silenceCandidateLedBeforeCoding plus a second Example entry recorded after Algorithm, asserting the identical (correct, scoped) output text.

Jim always volunteers the next step once a candidate stalls during Algorithm or Coding, even for interviewers who expect the candidate to lead those two steps themselves. Add a candidate-led toggle, carried from a lobby checkbox through participant metadata to a new prompt policy scoped to those two REACTO steps only: enabled, Jim asks an open question and waits for the candidate to state the next step instead of proposing one. Once Repeat, Example and Algorithm are all evidenced, the nudge leads with that question scoped to "if this stall is in the Algorithm or Coding step", folding the usual next-step text into the same conditional as its "otherwise" rather than naming it outright: accumulated evidence cannot prove the live conversation has left Example for good, so a candidate who returns to it before touching the editor gets the wording unchanged. The other four steps are unchanged. Bumps LIVE_PROMPT_VERSION and INTERVIEW_CONTRACT_BUNDLE_VERSION to 32 and 41, and refreshes the prompt golden. The lobby's new row needs headroom on the one-screen laptop breakpoint; the shared caption padding and checkbox min-height stay untouched; margins on the hero and header already scoped to that breakpoint give the row back its room instead. Closes sysprog21#115

jserv · 2026-10-25T15:18:03Z

+pub const INTERVIEW_CONTRACT_BUNDLE_VERSION: u32 = 41;
+pub const LIVE_PROMPT_VERSION: u32 = 32;


The PR description says this bumps LIVE_PROMPT_VERSION and INTERVIEW_CONTRACT_BUNDLE_VERSION to 31 and 40, but the diff moves them from 31 and 40 to 32 and 41 (and the docs row and golden digest agree with 32 and 41). Update the description so it matches what merges.

Good catch — updated the description's two mentions (the main paragraph and cubic's auto-generated summary) from 31/40 to 32/41, matching the diff, the docs row, and the golden digest.

jserv · 2026-10-25T15:33:43Z

Thank @Jerry-Chiang-23 for contributing!