fix(deep-interview): retire per-round tiered continuation asks for active interviews - #4590
Conversation
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
PR body integrity correction (repository-visible blocker, repaired). The previous PR body was catastrophically contaminated with a stale, unrelated Windows SDK session-index lock PR description ( Corrected via REST Post-edit REST verification: body contains Code head untouched — no amend, no push; CI: the body edit triggered replacement run @probepark — requesting your fresh independent review of exact head — |
|
Verdict digest correction (bootstrap-contract canonical form). Inspecting the
REST-verified after the edit: Replacement CI run @probepark (primary) / @HaD0Yun (backup): when your APPROVED review lands on exactly — |
a8138dc to
d7851b9
Compare
|
Rebased the single PR commit onto current Validation on that rebased head: The PR body now has exactly one — |
|
Exact-head product CI is complete for
— |
|
Resolved the terminal critic finding and pushed exact head |
probepark
left a comment
probepark
left a comment
There was a problem hiding this comment.
Approve at 9ded3effb
fix(deep-interview): retire per-round tiered continuation, head commit "distinguish cancellation
from early proceed". 3 files, +127.
Differential
# base 91a35114d, with this head test files applied
42 pass 1 fail
# head 9ded3effb
43 pass 0 fail
The distinction is worth making
Cancellation and early-proceed produce the same observable "the round ended before it was supposed
to", and collapsing them is how an interview that the user abandoned gets treated as one that
completed with partial answers. The first should discard, the second should carry forward what was
gathered. A workflow that cannot tell them apart either loses good input or fabricates consent.
Retiring per-round tiering in the same change is consistent - tiering was the mechanism that made
the two look alike, since a tier boundary and a cancellation both truncated the round.
merge-approved.
Reviewed by @probepark - method: fresh-worktree run at the exact head, clean base worktree with the head test files to prove the differential.
…tive interviews The ouroboros ooo-interview parity port (squash-imported in 10144d5) added a Step 2f 'Tiered Confirmation Cadence' that instructed the agent to ask 'Continue, or proceed with current clarity ({score}%)?' after every ordinary answered round 4-15 (and 16+ with a warning). That converted the ambiguity gate into per-answer consent friction — the exact regression reported in #4589 — and contradicted DIPP-9, the Phase 2 loop contract, Step 2d next-target reporting, and the runtime stop-gate reminder. Replace the cadence with an explicit Continuation Contract: an ordinary answered round NEVER asks for generic continuation approval; the interview scores, persists, reports progress, and continues to the next weakest-dimension question until one of four legitimate terminal conditions (threshold + closure gates, explicit user exit, invocation/resume suitability ambiguity only, bounded safety recovery). Explicit exit controls, the Phase 0.5 resume gate, threshold transparency, DIPP-9 execution approval, and the Round 100 hard cap are preserved. Lore-id: 4589-di-continue Constraint: do not remove explicit user exit controls or the resume gate Constraint: preserve DIPP-9 execution approval and threshold transparency Rejected: keep cadence but only for rounds 16+ | still converts the gate into consent friction Rejected: runtime-side suppression of continue? asks | the instruction itself is the defect; runtime already auto-continues Confidence: high Scope-risk: narrow Reversibility: trivial Tested: bun test deep-interview contract/session/gate suites (144+43+105+127 pass) Not-tested: live interactive TUI interview (contract-level regression pinned instead) Closes: #4589
Hard cancellation must stop an interview immediately at every round.\nEarly-proceed wording retains the pre-round-three clarification guard.\n\nLore-id: 4590-exit-semantics\nConstraint: retain explicit exits without reintroducing per-round continuation asks\nConfidence: high\nScope-risk: narrow\nTested: deep-interview skill contract and continuation tests
9ded3ef to
e17c982
Compare
Signed refreshed-head evidence
The authoritative product gates completed green: state-gate shards and aggregate, native build, both focused Deep Interview tests, coding-agent TypeScript build, affected-path evidence/aggregate, and virtual integration. The only red job is Bounded review blocker
[repo owner's gaebal-gajae (clawdbot) 🦞] |
…tive interviews (Yeachan-Heo#4590) * fix(deep-interview): retire per-round tiered continuation asks for active interviews The ouroboros ooo-interview parity port (squash-imported in 10144d5) added a Step 2f 'Tiered Confirmation Cadence' that instructed the agent to ask 'Continue, or proceed with current clarity ({score}%)?' after every ordinary answered round 4-15 (and 16+ with a warning). That converted the ambiguity gate into per-answer consent friction — the exact regression reported in Yeachan-Heo#4589 — and contradicted DIPP-9, the Phase 2 loop contract, Step 2d next-target reporting, and the runtime stop-gate reminder. Replace the cadence with an explicit Continuation Contract: an ordinary answered round NEVER asks for generic continuation approval; the interview scores, persists, reports progress, and continues to the next weakest-dimension question until one of four legitimate terminal conditions (threshold + closure gates, explicit user exit, invocation/resume suitability ambiguity only, bounded safety recovery). Explicit exit controls, the Phase 0.5 resume gate, threshold transparency, DIPP-9 execution approval, and the Round 100 hard cap are preserved. Lore-id: 4589-di-continue Constraint: do not remove explicit user exit controls or the resume gate Constraint: preserve DIPP-9 execution approval and threshold transparency Rejected: keep cadence but only for rounds 16+ | still converts the gate into consent friction Rejected: runtime-side suppression of continue? asks | the instruction itself is the defect; runtime already auto-continues Confidence: high Scope-risk: narrow Reversibility: trivial Tested: bun test deep-interview contract/session/gate suites (144+43+105+127 pass) Not-tested: live interactive TUI interview (contract-level regression pinned instead) Closes: Yeachan-Heo#4589 * fix(deep-interview): distinguish cancellation from early proceed Hard cancellation must stop an interview immediately at every round.\nEarly-proceed wording retains the pre-round-three clarification guard.\n\nLore-id: 4590-exit-semantics\nConstraint: retain explicit exits without reintroducing per-round continuation asks\nConfidence: high\nScope-risk: narrow\nTested: deep-interview skill contract and continuation tests --------- Co-authored-by: Yeachan Heo <yeachan.heo@gmail.com> Co-authored-by: probe <re2rar@gmail.com> (cherry picked from commit 9af5c7a)
2 participants
Fixes #4589
Scope
Removes the retired per-round continuation prompt from bundled Deep Interview instructions. Ordinary answered rounds score, persist, and ask the next weakest-dimension question. The scope deliberately preserves distinct exits:
stop/cancel/aborthard-cancel immediately at every round;enough/let's go/build ituse the existing pre-round-3 targeted-clarification guard and early-exit path after round 3.Rebased exact-head evidence
6696988b6dd757f898ab7b33984ba830b6344109(origin/dev)e17c982945348ebf6ae3bc124a733864a79d269dsha256:26f7697a7b1bb46f3eaf1255b1f029317852461cc6c834551d0d1738e100cf3bbun test packages/coding-agent/test/deep-interview-skill-contract.test.ts— 20 passbun test packages/coding-agent/test/agent-session-deep-interview-continuation.test.ts— 23 passbun --cwd=packages/coding-agent run check:types— passgit diff --check origin/dev...HEAD— passprobeparkhas been requested for a fresh, non-author exact-head review. The prior approval was bound to superseded head9ded3effb491de7473a3ae37fce566b0bc221f69and is not relied upon.GJC verdict