Reconstruct and experimentally challenge decisions from prior Codex sessions. Use for `$retrace`, historical decision replay, counterfactual forks, alternative-route challenges, hindsight-separated retrospectives, workflow-governance audits, skill decision attribution, or 'why did that session choose this?'. `$seq` owns deterministic history and source-governance evidence; `$cas` owns safe thread/rollout replay and FIR lifecycle; `$retrace` owns bounded experiments and DRR synthesis. Never present fork output as the source model's hidden chain of thought.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
The command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
File Explorer
21 files
Showing SKILL.md
SKILL.md
Source instructions · Read-only preview
name
retrace
description
Reconstruct and experimentally challenge decisions from prior Codex sessions. Use for `$retrace`, historical decision replay, counterfactual forks, alternative-route challenges, hindsight-separated retrospectives, workflow-governance audits, skill decision attribution, or 'why did that session choose this?'. `$seq` owns deterministic history and source-governance evidence; `$cas` owns safe thread/rollout replay and FIR lifecycle; `$retrace` owns bounded experiments and DRR synthesis. Never present fork output as the source model's hidden chain of thought.
what was historically explicit
what the visible trace supports
whether the claimed workflow really governed the source
whether the route is stable under replay
which alternative is strongest
which fact flips the route
what hindsight changes
what remains unknowable
A replay is a new model execution, not access to the source model's private chain of thought.
Seq returns observations and provenance. It does not author or validate Retrace
artifacts.
$ledger
Owns passive-definition compilation, structural validation, canonicalization,
and identity derivation for SGG-v1, DCP-v2, RIP-v1, DRR-v1, and the CAS-owned
FIR-v1 definition. A pass grants neither replay authority nor a Retrace
verdict.
Before the first native Ledger command in this workflow, load $ledger and
complete $ledger ensure once.
$cas
Owns:
app-server compatibility;
source verification;
thread_fork or rollout_transcript lineage;
exact completed-boundary selection and retained-prefix anchoring;
the question asks whether a workflow/skill governed the decision;
an intervention will apply a current workflow contract;
source inclusion depends on a workflow-specific observation.
Obtain the exact session-level row, not only aggregate counts.
For controller-backed review-closure workflows:
Both projections return at most 256 newest matching metadata candidates. A
top-k omission limitation means the result is incomplete: narrow the selectors
before classifying, and never infer absence from that result.
Read raw evidence only for one selected identity. Save the selected candidate
envelope, extract its identifiers as data, and JSON-encode the query before
passing it as one argument. Never interpolate trace-derived identifiers into
shell source. Use native seq query; do not emit a whole session's lifecycle:
authoritative
controller-grade governance evidence; replay allowed
declared_uncontrolled
explicit workflow declaration but controller governance unproven;
replay allowed with limitation
incidental
artifact/path/history mention only; stop before replay
ambiguous
deterministic investigation only
absent
stop before replay
A filename such as .step/review-workflow-plan.jsonl is not a workflow activation.
A generic merge/land/complete signal is not controller closure.
See source-governance.md.
Locate the visible route decision with the candidate observation, then rerun
the capsule with that exact one-based turn index. The capsule excludes later
turns and source rows without a turn assignment before replay can inspect them.
Seq does not choose the decision episode or author the packet. Do not let replay
models select the historical source episode.
workspace_policy = transcript_only
no current-checkout tools
no live historical workspace claim
lineage_mode recorded in FIR
Rollout transcript replay is structurally anchored transcript replay, not live thread forking or workspace reconstruction.
See workspace-reconstruction.md.
4. Capability requirements
Require:
Seq observation ABI and all three Retrace observation definitions
Ledger artifact ABI and all four Retrace artifact definitions
CAS FIR definition
cas session_inquiry and FIR support
at least one supported lineage mode
read-only inquiry
no-network policy
receipt persistence
When only deterministic analysis is available, fork-based claims are forbidden.
Before compiling or running an inquiry, bind compatibility to the exact CAS
transport that will execute it:
CAS must prove source lineage, retained anchor, model/provider, permission policy, workspace mode, turn state, and cleanup.
Detached lifecycle remains available through start, status, wait, interrupt, and cleanup.
7. Consume FIR-v1
Validate each exact receipt before interpretation:
Require ledger-validation-result/v1, valid:true, the returned CAS
definition digest, authority_granted:false, and storage_mutated:false.
Only complete, source-bound FIRs contribute to:
route distribution
consensus
stability
instruction effect
source governance is declared_uncontrolled
forks disagree
hindsight leakage is plausible
lineage/workspace/model differs
skill effect is material
route stability will drive tuning or doctrine
rationale:
post-decision/pre-outcome; reconstruct visible support and assumptions
counterfactual:
pre-decision; choose independently without predicting history
alternative challenge:
pre-decision; strongest evidence-consistent non-historical route
evidence ablation:
pre-decision; withhold/change one named item and re-decide
retrospective:
outcome-aware; label every lesson as hindsight-informed
Skill/workflow effect
Ask:
Did the source actually contain the skill/workflow?
Was governance authoritative, declared, incidental, or absent?
Did baseline and intervention choose different routes?
Did controlled ablation change the route?
Did the source explicitly attribute the decision?
Fork self-report alone is weak.
DRR may inform $tune; it does not authorize edits.
Security, persistence, and failure
Defaults:
ephemeral
read-only
network disabled
approvals denied
dynamic tools denied
one bounded turn
bounded excerpts and refs
Do not use thread/shellCommand.
Do not persist private reasoning.
Use .ledger/retrace/<inquiry-id>/ only when receipts are required, and local-exclude it by default.
A blocked replay does not erase deterministic source evidence.
Output
Report:
source / decision / question
source-governance provenance and verdict
lineage and workspace mode
historical explicit / trace-inferred evidence
valid and invalid FIRs
baseline/intervention route distribution
strongest alternative and flip conditions
hindsight-separated lessons
skill/workflow effect
confidence and unsupported claims
cleanup and CLI gaps
Hard rules
$seq owns history and workflow provenance.
$cas owns replay lifecycle.
$retrace owns experiments and synthesis.
Prove workflow governance before workflow-specific replay.
Artifact/path mentions are not activations.
Generic delivery completion is not controller closure.
Fork output is not hidden historical chain of thought.
No outcome-blind claim without exact structural anchoring.
No historical workspace claim from the current checkout.
No causal skill claim from self-report or consensus alone.
