codex/skills/fixed-point-driver/SKILL.md
Drive exhaustive build-review-improve-verify loops to Truth-Owner Ablative-Isomorphic Normal Form: one canonical owner per material invariant, no duplicate truth surfaces, no unresolved review counterexamples, no unresolved adversarial or ablation vetoes, no unretired additive scaffolding, no dominated/vestigial/subsumed surface without a keep warrant, and behavior-preserving proof-gated closure. Trigger when coding needs de novo re-litigation, PR review closure, repeated review/fix loops, invariant repair, proof-surface hardening, negative-evidence pruning, CAS/Codex review resolution, parallel adversarial action, optional architecture fingerprint preflight, or when agents risk adding local patches instead of deleting/refactoring/canonicalizing. Do not use for trivial one-step tasks or when the user wants one narrow phase.
npx skillsauth add tkersey/dotfiles fixed-point-driverInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
This skill coordinates implementation, review, adjudication, reduction, verification, parallel read-only adversarial challenge, ablation challenge, and closure until the artifact set reaches Truth-Owner Ablative-Isomorphic Normal Form.
Drive toward this state:
Every material invariant has exactly one canonical owner, every witness points back to that owner, every review finding is either discharged or represented as a counterexample, every adversarial or ablation veto is cleared, accepted as risk, or routed, every additive scaffold has been promoted/collapsed/deleted, every dominated/subsumed/vestigial/non-canonical surface has been deleted, collapsed, privatized, decommissioned, or kept with an explicit warrant, and every deletion/collapse/canonicalization has a current behavior-preservation witness, and no duplicate truth surface remains merely because it helped satisfy an intermediate review loop.
A fixed point is not “review is clean.” A fixed point is a normal form of the code and proof system.
Frontier coding agents are good at adding code and addressing review comments. They are weaker at noticing that the best fix is to delete a path, collapse duplicate owners, privatize a surface, tighten the one boundary that should have owned the invariant from the start, or let a parallel adversary invalidate the local patch before time is spent implementing it.
This skill converts review churn into an ownership-normalization and surface-ablation problem. It converts adversarial pressure into time-efficient, read-only clearance packets and converts deletion pressure into witness-backed Ablation Ledger rows.
This skill owns the conversion from doctrine-rich pressure into normal-form artifacts.
Use doctrine only when it changes the route:
fixed-point creates a reopenable material fixed-point gate, not a long review
loop by default.invariant creates or updates a Truth Unit and owner edge.canonical selects one owner/representation and retires, privatizes, or
justifies shadow owners.witness creates proof receipts tied to current head/artifact state.adversarial / de novo creates candidate inventory, challenger packets, and
no-finding countercases.parallel creates independent read-only adversarial lanes when it can reduce
elapsed time or prevent wrong action.unwitnessed guarantee / illegal inhabitant creates Soundness Ledger rows.ablative creates Ablation Ledger rows and prevents additive mutation until
deletion, collapse, reuse, privatization, decommissioning, or canonicalization
is defeated or selected.isomorphic creates Ablative Isomorphism Cards so deletion, collapse, merge,
reuse, or canonicalization preserves observable behavior.clone-classified prevents accidental-rhyme or semantic-clone merges from being
treated as safe simplification.abstraction-laddered prevents new helpers/interfaces/generics/wrappers from
skipping the evidence rung.dominated creates a delete/collapse candidate when an existing path covers
the obligation with lower complexity or stronger proof.subsumed creates an owner-canonicalization candidate when a local abstraction
no longer owns a distinct obligation.vestigial creates a decommission/delete candidate when a surface only remains
from a retired obligation.uninhabited creates a deletion/assertion candidate when a branch/state cannot
legally occur under current constructors and invariants.Do not use this skill for simple bounded tasks where ordinary direct execution plus one focused check fully closes the state space. If the run stays active, the reason must be a material open truth unit, stale proof, unresolved counterexample, unresolved adversarial veto, unresolved ablation veto, unresolved isomorphism gap, shadow owner, addition escrow, route-changing architecture uncertainty, or active warrant that requires fixed-point handling.
When several findings are each locally valid, ask whether they are all counterexamples to the same truth unit. If yes, prefer one owner-level rewrite or canonicalizing/ablative move over multiple local patches.
This skill is parallel-adversarial by default for material actions and ablative-before-additive for mutation routes, but single-writer for all code changes.
Every material action must have an adversarial response:
| phase/action | adversarial response | |---|---| | warrant intake | Is the warrant stale, expired, overbroad, or incompatible with current artifact state? | | truth-unit extraction | Is the invariant split incorrectly? Is the canonical owner wrong? Are there shadow owners? | | route selection | Is delete/reuse/refactor/canonicalize/privatize/no-change/validate-first better than mutation? | | implementation | Does the patch exceed warrant scope or surface budget? Does it add duplicate truth surfaces? | | validation | Does proof exercise the real failure mechanism and current artifact state? | | de novo review | Are there unlisted counterexamples, illegal inhabitants, or unretired scaffolds? | | ablation pass | Are there dominated, subsumed, vestigial, uninhabited, pass-through, or non-canonical surfaces that should be removed or warranted? | | one-change challenge | Is there one remaining impactful change before closure? | | closure handoff | Are proof receipts current and all adversarial/ablation vetoes cleared or explicitly routed? |
Parallelism modes:
root-equivalent: root performs the adversarial/ablative response inline for
narrow, bounded, already-proof-bearing work.targeted-parallel: one or two independent read-only lanes challenge distinct
uncertainty classes.expanded-targeted: three or four lanes for coupled truth units, stale proof,
negative evidence, surface-budget risk, or ablation risk.swarm: five or more lanes for large, contentious, P2+, invariant-coupled,
deletion-sensitive, or likely-to-reopen runs.not-required: no safe action exists; block and name missing evidence.Parallel adversaries may gather evidence, inspect surfaces, challenge proofs, map hazards, test invariants conceptually, and identify deletion/collapse candidates. They must not edit code, alter fixtures, resolve threads, or draft final replies. The root coordinator integrates packets and keeps writes single-threaded.
review-adjudication decides which review comments matter and issues warrants;
this skill decides whether selected work preserves Truth-Owner Ablative Normal
Form.accretive-implementer performs implementation/remediation when the selected
rewrite is narrow, owned, and warrant-scoped.adversarial-reviewer challenges the current artifact state and should report
normal-form violations, not just bugs.verification-closure performs decisive proof and closure gating.logophile may sharpen PR-facing summaries, names, and doctrine stacks, but
does not own operational policy.Use this skill when any of these are true:
address, delete-collapse-canonicalize, or
validate-only to $fixed-point-driverDo not use this skill for simple bounded tasks with an obvious check.
Frame
Warrant intake / parallelism plan
review-adjudication output when presentmutate-code, delete-collapse-canonicalize,
add-validation-only, proof-only, draft-reply, or none)Truth Unit extraction
Adjudication intake
review-adjudication output when presentAblative preflight
ablation_auditor or root-equivalent packet when a deletion/collapse route
could materially change implementation or closureAdversarial preflight
Route selection
Implementation pass
accretive-implementer when the selected rewrite is
narrow and ownedmutate-code or
delete-collapse-canonicalize warrant, current adversarial clearance,
current ablative clearance, and passing Surface Budget PreflightDe novo adversarial review
adversarial-reviewer or root-equivalent challengeOne-change challenge Ask:
If you could change one thing about this changeset, what would you change?
Then ask the ablative variant:
If you could delete, collapse, privatize, decommission, or canonicalize one thing in this changeset without weakening the live contract, and with behavior-preservation proof, what would it be?
If either answer is impactful and in-scope, route it to accretive-implementer
or root-equivalent single-writer implementation, then rerun de novo review. If
structural and outside constraints, mark needs-decision or blocked. If no
impactful remaining change exists, proceed to verification closure.
Material fixed-point test A candidate fixed point exists only when:
Verification closure
Hand off to verification-closure with a closure packet.
If closure reopens a gate, resume at route selection.
| warrant id | claim id | permitted action | permitted scope | expiry check | surface budget | adversarial plan | ablation plan | parallelism mode | intake status |
|---|---|---|---|---|---|---|---|---|---|
Rules:
intake status is active, stale, blocked, consumed, or not-applicable.mutate-code warrants must have surface budget and adversarial clearance.delete-collapse-canonicalize warrants must have ablation proof or a
validate-first gate.add-validation-only warrants must forbid production mutation.| truth unit id | invariant | canonical owner | witnesses | duplicate/shadow owners | review/adversarial counterexamples | status | proof refs | next action |
|---|---|---|---|---|---|---|---|---|
| id | surface | kind | current obligation | obligation status | canonical owner | replacement path | action | deletion/collapse proof | keep warrant | status |
|---|---|---|---|---|---|---|---|---|---|---|
Allowed kind values:
dominatedsubsumedvestigialuninhabitedunreachablepass-throughduplicate-truth-surfacenon-canonicaladditive-scaffoldtemporary-proof-scaffoldAllowed obligation status values:
liveexpiredmovedduplicateunprovenunknownAllowed action values:
deletecollapsecanonicalizeprivatizedecommissionvalidate-firstkeep-with-warrant| id | candidate | kind | surface removed | confidence | ownership clarity | risk | score | decision | proof needed |
|---|---|---|---:|---:|---:|---:|---:|---|---|
Use:
Ablation Score = (Surface Removed × Confidence × Ownership Clarity) / Risk
Do not let LOC savings dominate semantic-surface reduction. A small deletion can score high when it removes a duplicate truth surface, obsolete flag, public symbol, state variant, or proof obligation.
| card id | surface | action | behavior preserved | public contract preserved | error semantics preserved | ordering/side effects preserved | clone classification | abstraction-ladder check | compatibility risk | proof signal | status |
|---|---|---|---|---|---|---|---|---|---|---|---|
Allowed status values: pass, validate-first, missing, not-applicable.
| action id | phase | target | challenger lanes | parallelism mode | strongest adversarial finding | veto status | clearance | proof ref | decision impact |
|---|---|---|---|---|---|---|---|---|---|
Allowed veto status values: cleared, preserved-no-change, unresolved,
vetoed, blocked, not-required.
Allowed clearance values: cleared, preserved, rerouted, downgraded,
blocked.
| receipt id | warrant id | patch/pass | production insertions | production deletions | net production loc | public symbols added | helpers added | duplicate paths added | budget status | proof ref |
|---|---|---|---|---|---|---|---|---|---|---|
If budget status is expansion-needed or violation, stop implementation until
an Expansion Warrant Request is granted or the patch is redesigned.
Before invoking verification-closure, compile:
closure_handoff_packet:
artifact_state_id: "..."
goal: "..."
warrant_intake:
- warrant_id: "..."
claim_id: "..."
permitted_action: "mutate-code | delete-collapse-canonicalize | add-validation-only | resolve-thread | draft-reply | defer | none"
intake_status: "active | consumed | stale | blocked | not-applicable"
adversarial_clearance: "cleared | preserved | rerouted | downgraded | blocked"
ablative_clearance: "clear-additive | select-ablative-route | validate-first | veto-additive | unresolved | not-required"
parallelism_plan:
lane: "root-equivalent | targeted-parallel | expanded-targeted | swarm | not-required"
reason: "..."
read_only: yes
write_owner: root-or-delegate
truth_units:
- id: "TU1"
invariant: "..."
canonical_owner: "..."
witnesses: ["..."]
duplicate_truth_surfaces_retired: ["..."]
open_counterexamples: []
review_agenda:
act_on: []
validate_first: []
proof_only: []
delete_collapse_canonicalize: []
no_change: []
blocked: []
ablation_ledger:
open_candidates: []
closed_candidates: []
keep_warrants: []
unresolved_ablation_vetoes: []
ablation_opportunity_matrix:
top_candidates: []
rejected_candidates: []
ablation_isomorphism_cards:
- card_id: "..."
surface: "..."
action: "delete | collapse | reuse | canonicalize | privatize | decommission"
behavior_preserved: "..."
public_contract_preserved: "yes | no | unknown"
error_semantics_preserved: "yes | no | unknown | not-applicable"
ordering_side_effects_preserved: "yes | no | unknown | not-applicable"
proof_signal: "..."
status: "pass | validate-first | missing | not-applicable"
adversarial_action_ledger:
open_vetoes: []
cleared_vetoes: []
accepted_risks: []
soundness_ledger:
open_rows: []
closed_rows: []
surface_delta_receipts:
- warrant_id: "..."
budget_status: "within-budget | expansion-needed | violation | not-applicable"
proof_ref: "..."
proof_receipts:
- command_or_check: "..."
result: "..."
artifact_state_match: yes | no
one_change_challenge:
answer: "..."
ablative_answer: "..."
routed: yes | no
reason: "..."
unresolved_adversarial_vetoes: []
unresolved_ablation_vetoes: []
Use tail-weighted sections:
Do Next must be the final section.
Before closure handoff, emit a gate summary with these fields:
artifact_state_match: pass / failwarrant_intake: pass / failparallelism_calibration: pass / failadversarial_action_coverage: pass / failablation_coverage: pass / fail / not-applicableablation_isomorphism: pass / fail / not-applicableopen_truth_units: 0 or named blockersduplicate_truth_owners: 0 or named blockersopen_review_counterexamples: 0 or named blockersunresolved_soundness_rows: 0 or named blockersunresolved_adversarial_vetoes: 0 or named blockersunresolved_ablation_rows: 0 or named blockersunresolved_isomorphism_cards: 0 or named blockersdominated_surfaces: 0 or named blockersunwarranted_keep_surfaces: 0 or named blockersunretired_additive_scaffolds: 0 or named blockerssurface_budget_status: pass / fail / not-applicableproof_receipts_current: pass / failone_change_challenge_status: pass / failablative_one_change_challenge_status: pass / failverification_closure_ready: yes / nomutate-code or delete-collapse-canonicalize
warrant.add-validation-only into production mutation.tools
Convert markdown plans into beads with dependencies using br CLI. Use when creating task graphs, polishing beads before implementation, or bridging planning to agent swarm execution.
development
Orchestrate Codex skill optimization during active sessions through $cas goal control, $shadow single-session evidence, $tune diagnosis/refinement briefs, and the skill-optimizer custom subagent. Trigger for $opt, skill optimization loops, session-driven skill tuning, meta-skill audits, or explicit validated skill edits. Do not use for general code optimization, product optimization, or performance tuning.
development
Run a targeted fresh-eyes blunder pass over code, specs, plans, adjudications, closure gates, skill edits, or negative-evidence ledgers. Trigger when asked to reread with fresh eyes, find obvious bugs, catch mistakes/oversights/omissions, check for embarrassing misses, or perform a second independent blunder pass before closure. Do not use as a substitute for implementation, adjudication, or verification; use it as the final falsification/check pass for those workflows.
development
Explicitly shadow, tail, watch, follow, monitor, supervise, or companion exactly one Codex session id/path through `$seq`, then apply a named target skill as an interpretation/reporting/proposal/action lens until the watched session stops.