SuperAPI AI Directory

resolve

Coding & development

View repository

Updated 10-16 07:15

About this skill

Resolve

Mission

$resolve is the Intent-Closed Counterexample Synthesis Compiler.

source intent
-> sealed observation language
-> bounded falsification batches
-> minimal distinguishing counterexamples
-> counterexample quotient
-> minimum behavioral kernel
-> one realization
-> conformance falsification
-> terminal holdout
-> delivery closure

The unit of review is a minimal distinguishing experiment.

The unit of resolution is a counterexample equivalence class.

The unit of implementation is a kernel transition.

Governing problem

Repeated whole-diff review creates a positive feedback loop:

larger diff
-> broader review search
-> more valid adjacent findings
-> more local patches
-> larger review search

A finding may be real, branch-liable, and diff-local while still being outside the intended change.

Therefore:

diff relevance is not mutation authority
validity is not intent entailment
additional witness is not a new code distinction

Protocol profile

intent-closed-cegis-v1

Primary artifacts:

AC-v2    sealed acceptance contract and observation language
RB-v1    bounded review batch
RAP-v1   targeted review aperture
CEX-v1   minimal distinguishing counterexample
CEB-v2   sealed, quotiented counterexample basis
MBK-v1   minimum behavioral kernel, intent-bound
RC-v1    reduction certificate
PHI-v1   controller-derived review potential
MBKC-v1  campaign certificate and delivery gate

Controller requirement

Canonical controller:

resolve-c3

Before material work:

python3 codex/skills/resolve/tools/controller_preflight.py

Required capability profile:

acceptance_contract_v2
sealed_review_horizon_v1
review_batch_v1
review_aperture_v1
counterexample_v1
counterexample_basis_v2
minimum_behavioral_kernel_v1
reduction_certificate_v1
review_potential_v1
intent_closed_conformance_v1
terminal_holdout_v1
physical_apply
physical_commit
physical_push
closure_horizon_v1

If unavailable:

analysis, adjudication, AC/CEX/CEB/MBK/RC drafting allowed
delivery mutation, commit, push, and closure forbidden

Core laws

1. No review-to-code edge

review claim
-> CEX-v1 adjudication
-> sealed batch
-> CEB/MBK/RC
-> selected design
-> realization

Never:

review claim -> patch

2. Intent closes before synthesis

AC-v2 defines:

MUST       required behavior and compatibility obligations
MUST NOT   forbidden states/transitions
MAY        permitted but not mutation-justifying behavior
OUTSIDE    behavior not authorized in this campaign

After sealing, the observation language cannot expand without an explicit AC rebase.

3. Review batches are mutation barriers

While any discovery, conformance, or holdout batch is open:

delivery mutation = forbidden

Batch all planned apertures, classify every result, quotient duplicates, then mutate once.

4. No counterexample without a minimal trace

Every actionable CEX identifies:

accepted law/obligation
minimal trace or witness
expected observation
actual observation
current artifact state

Generic suspicion is not a counterexample.

5. No new implementation distinction for another witness

new class                  -> kernel may refine
new witness, existing class -> proof may strengthen
duplicate                  -> no semantic action
outside horizon            -> follow-up or AC rebase

An additional witness cannot create another branch/helper/state/test family.

6. Implementation review is conformance review

After kernel acceptance, allowed review outputs are:

existing-law nonconformance
missing law proof
orphan realization construct
stale artifact state
novel in-horizon CEX
outside-horizon proposal
clean

Generic feature discovery cannot directly authorize mutation.

7. Same-class recurrence invalidates the realization

same accepted class recurs after realization
-> candidate failed
-> invalidate realization
-> no local append patch

A novel in-horizon class returns to kernel synthesis.

An outside-horizon class returns to AC authority or follow-up.

8. Review potential must decrease

Each realized cycle emits PHI-v1:

U = unclassified in-horizon counterexamples
L = unsatisfied accepted laws
C = open counterexample classes
O = orphan realization constructs
S = hard semantic-surface vector
P = unmapped/missing proof debt

Require:

(U,L,C,O)_after <lex (U,L,C,O)_before
S_after componentwise <= S_before unless AC is explicitly rebased
P_after <= P_before
strict_progress = yes

Local comment silence is not progress.

9. One campaign-wide kernel

All batches refine one kernel rooted at campaign_base_sha.

A prior tuple-local closure head is not a new compiler base.

10. No orphan code or proof

Every surviving construct maps to:

accepted law
counterexample class
canonical owner
proof obligation

Targets:

orphan_code_constructs = 0
wound_specific_test_families = 0
unmapped_proof_actions = 0

11. No quotient without congruence

Every accepted operation, transition, and observation respects the quotient.

Unknown congruence blocks realization.

12. No normal form without recomposition

Every accepted law and required observation recomposes from retained factors after quotienting and ablation.

Lifecycle

DISCOVER

  • Bind campaign/base/baseline/current head/PR/proof.
  • Compile AC-v2 from plan, PR intent, compatibility, non-goals, and proof bar.
  • Run one broad discovery batch.
  • Mutation forbidden.

SEAL

  • Adjudicate every claim as CEX-v1 or rejection.
  • Seal AC-v2 and the discovery RB-v1.
  • Quotient accepted CEX into CEB-v2.
  • Unknown intent relation blocks.

SYNTHESIZE

  • Compile MBK-v1 and RC-v1 from AC-v2 + CEB-v2.
  • Review the kernel/certificate, not prospective patches.
  • Plan targeted review apertures.
  • Select one realization design.
  • Predict a non-worsening semantic surface.

REALIZE

  • Create a campaign-base-pinned realization worktree.
  • Hand one accepted kernel/design to $fixed-point-driver.
  • Capture/map/measure/verify the whole realization.
  • No raw finding may enter the implementer.

CONFORM

  • Open a conformance RB-v1 against explicit RAP-v1 apertures.
  • Reviewers attempt to falsify named laws at named owners/transitions.
  • Seal the batch before any new mutation.
  • Existing-class recurrence invalidates realization.
  • Novel class returns to kernel.
  • Outside horizon routes to follow-up or AC rebase.

HOLDOUT

  • After targeted conformance is clean, run one broad terminal holdout.
  • Holdout has no direct mutation authority.
  • In-horizon novelty returns to kernel.
  • Contract-invalidating evidence returns to AC.
  • Outside-horizon findings are deferred or explicitly rebased.

CLOSE

Require:

AC/horizon current
no open review batch
no unclassified in-horizon CEX
CEB/MBK/RC current
no novel conformance/holdout CEX
PHI terminal and strictly improved
all laws proved
all constructs/proof mapped
semantic surface conserved or explicitly rebased
current-head proof passed
PR threads swept
commit/push current

Review modes

discovery
  one broad pre-kernel sensing pass

kernel_review
  AC/CEB/MBK/RC only; no delivery code

conformance
  targeted RAP-v1 laws/owners/transitions/proof; whole-diff review forbidden

terminal_holdout
  one broad final challenge; no direct mutation authority

Every mode emits a sealed RB-v1 before mutation or closure.

Entry modes

Clean current-head review

If no finding-bearing campaign exists:

current proof
thread sweep
tuple-bound closure

Do not manufacture AC/MBK machinery.

Isolated conformance correction

Allowed only when:

exactly one confirmed branch-liable existing-law violation
AC and kernel current
no new class/owner/state/protocol/public/fallback distinction
semantic surface componentwise nonincreasing
proof extends an existing law family
one sealed conformance batch

Still emit compact CEX/MBKC/PHI evidence.

Material review

Use full lifecycle when any apply:

two or more branch-liable findings
same-family recurrence
new semantic distinction
positive hard-surface pressure
public/compatibility/fallback pressure
thread reopening
finding proposes new behavior
whole-diff review discovers adjacent scope

Workflow

  1. Preflight controller.
  2. Bind campaign tuple.
  3. Compile and validate AC-v2.
  4. Open/run/seal discovery RB-v1.
  5. Use $review-adjudication to produce CEX-v1.
  6. Use $review-compression-compiler to produce CEB-v2 + MBK-v1 + RC-v1.
  7. Audit quotient, recomposition, proof basis, and initial PHI.
  8. Select one design and create realization worktree.
  9. Use $fixed-point-driver for one bounded realization.
  10. Capture/map/measure/verify realization.
  11. Run/seal targeted conformance batches.
  12. Invalidate/return rather than append-patch on new evidence.
  13. Certify strict PHI progress.
  14. Run/seal one terminal holdout.
  15. Apply/commit/push through controller.
  16. Close only against current AC/kernel/batches/PHI/proof/PR tuple.

Validation

python3 codex/skills/resolve/tools/acceptance_contract_gate.py acceptance.json
python3 codex/skills/review-adjudication/tools/counterexample_gate.py cex.json
python3 codex/skills/resolve/tools/review_batch_gate.py batch.json
python3 codex/skills/review-compression-compiler/tools/counterexample_basis_gate.py basis.json
python3 codex/skills/resolve/tools/kernel_lint.py kernel.json
python3 codex/skills/resolve/tools/review_potential_gate.py potential.json
python3 codex/skills/resolve/tools/mbkc_gate.py mbkc.json --terminal

Decision observability

Emit SDR-v1-compatible decisions for:

AC seal/rebase
CEX admission/disposition
CEX class merge/split
kernel acceptance
design selection
realization invalidation
return-to-kernel/contract
PHI gate
terminal closure

Decision contract: decision-contract.yaml.

Final report

Resolve:
- campaign / base / baseline / head:
- AC / horizon:
- review batches / apertures:
- raw claims / valid CEX / novel classes / duplicate witnesses:
- MBK / RC / selected design:
- realization result:
- PHI before / after:
- conformance / holdout:
- proof / PR threads / delivery:
- deferred outside-horizon findings:
- residual risk:

End with:

Resolve Bottom Line:
- intent horizon:
- novel accepted counterexamples:
- kernel transition:
- strict review-potential progress:
- realization:
- proof:
- closure:
- blocker / rebase / follow-up:

Hard rules

  • Never mutate from raw review prose.
  • Never mutate while a review batch is open.
  • Never admit an observation without AC or kernel-law anchors.
  • Never let MAY or OUTSIDE behavior justify code.
  • Never create a distinction for an additional witness.
  • Never run generic whole-diff review during conformance.
  • Never append-patch after same-class recurrence.
  • Never accept a cycle without strict PHI progress.
  • Never close with empty acceptance, empty finding-bearing basis, stale horizon, open batch, orphan construct/proof, or stale delivery tuple.
  • Never merge or land.