Seq

Seq

Mission

Reconstruct:

facts + provenance + corpus scope + contamination + limitations + uncertainty

Seq does not validate durable artifacts or grant action, repair, review, publication, or closure authority.

Source boundary

Seq reads only:

• Codex rollout JSONL and session/state metadata needed to locate or interpret it;
• supported OpenCode execution/session sources;
• explicit immutable input relations supplied by the caller.

Select OpenCode prompt history explicitly with --path ~/.local/state/opencode/prompt-history.jsonl; its prompts, parts, and tool lifecycle use the canonical physical relations, never a source-specific command.

Seq does not scan memory roots, Ledger stores, artifact directories, or another durable source implicitly. Compose durable facts explicitly:

# Load $ledger and complete $ledger ensure once before this composition.
ledger project \
  --definition <artifact-definition.json> \
  --projection <facts> \
  --repo <repo> \
  --payload-only \
  --format json >facts.json

seq observe \
  --definition <observation-definition.json> \
  --input facts=facts.json \
  --projection <projection> \
  --format json

Bootstrap and capability boundary

Require Seq major version 1 and seq-observation-abi/v1:

seq version
seq capabilities --format json

Check once per unchanged executable and execution environment; reuse the result across observations. Recheck the selected definition when its closure or required operators change.

Capabilities report only physical adapters, native operators, renderers, cache format, and generic limits. A missing skill-specific flag is never a fallback signal; load the owning definition and check its ABI/operators.

Native surface

Use native commands only for physical session structure:

seq definition check
seq definition describe
seq observe
seq explain
seq sessions
seq turns
seq session-detail
seq tool-lifecycle
seq session-graph
seq tail
seq find-session
seq datasets
seq dataset-schema
seq query
seq index
seq capabilities
seq version

If a higher-level question cannot be expressed, add a passive observation definition to the semantic owner. Request a new native operator only when it is domain-independent, explicitly bounded, and necessary for three unrelated definitions or for preserving one live behavior without material performance loss.

Standard Seq definitions

Seq owns reusable physical analyses under:

${CODEX_HOME:-$HOME/.codex}/skills/seq/definitions/seq/

Current definitions:

message-search.json
session-summary.json
token-usage.json
tool-search.json
turn-metrics.json
turn-report.json

tool-search.json returns metadata through rows. Request rows-raw only when raw arguments, inputs, commands, and outputs are necessary and safe.

turn-metrics.json reports raw sums over the selected physical turns rows. Its selected_turn_*_sum fields are not lineage-owned token usage and must not answer session-tree, worker-inclusive, billing-style, or corpus token totals. Each token sum is partial unless its matching selected_turn_*_count equals turn_count; a zero sum with a zero count means unobserved data, not zero use. Those questions require an owning passive definition over lineage-owned token deltas; if the installed Seq ABI does not expose that relation or operator, report obstruction rather than summing turns or sessions token fields.

token-usage.json is the owning corpus-usage definition. It reconstructs cumulative token transitions in source order from both total_token_usage and last_token_usage. For modern rows, the transition is (total - last) -> total and the counted delta is last; null-info events are observations without usage and do not become zero-token transitions. The definition keeps the first emission of a cumulative total within each session, excludes only exact transitions replayed from a strict ancestor, retains sibling and independent-root usage, and applies session and since/until selection after ownership. Total-only legacy rows use an estimated snapshot fallback. Inconsistent modern tuples make the result invalid and its token totals null. Its cached percentage denominator is input tokens: cached_input_tokens / input_tokens * 100.

seq_definition_root="$(realpath "${CODEX_HOME:-$HOME/.codex}/skills/seq/definitions/seq")"
seq observe \
  --definition "$seq_definition_root/token-usage.json" \
  --root "${CODEX_HOME:-$HOME/.codex}/sessions/2026/07" \
  --since 2026-07-01T00:00:00-07:00 \
  --until 2026-08-01T00:00:00-07:00 \
  --projection summary \
  --format json

Example:

seq_definition_root="$(realpath "${CODEX_HOME:-$HOME/.codex}/skills/seq/definitions/seq")"
seq observe \
  --definition "$seq_definition_root/message-search.json" \
  --path <rollout.jsonl> \
  --param "needle=<term>" \
  --projection rows \
  --format json

Domain observations live with their owners. For example, $tune owns skill decision audits and $retrace owns decision discovery/source governance. Seq compiles their passive data; it does not own their vocabulary or conclusions.

Evidence discipline

For every report preserve:

definition id and closure digest
Seq ABI and binary version
source adapter/schema version
session IDs and path/corpus digest
explicit since/until and timezone when applicable
worker inclusion
current-session exclusion
files opened and bytes read
physical passes and rows scanned/materialized
contamination
limitations and uncertainty

A fixed time window is not an immutable corpus snapshot.

Distinguish:

presence
decision influence
downstream outcome
workflow governance

Likewise, keep mutation churn, final Git delta, and semantic surface separate. Presence does not prove influence; outcome association does not prove causality; an observation does not grant authority.

Worker attribution

Include linked workers only when requested or when the relevant evidence occurs there. Preserve root session, worker session, parent edge, lane/receipt ID, declared skills, observation, and outcome. Never merge unlinked workers into the root denominator.

Privacy and contamination

Default to sanitized references and bounded excerpts. Detect injected skill blocks, current audit prompts, generated reports, quoted transcripts, memory summaries supplied as explicit inputs, and examples. Do not expose private reasoning as report evidence.

Hard rules

• Use the narrowest native physical command or owning passive definition.
• State denominators, exclusions, time bounds, and timezone.
• Preserve evidence provenance and definition identity.
• Use one fused scan when several requested projections share a source.
• Never treat raw turns or sessions token sums as lineage-owned corpus usage.
• Do not interpret a definition condition-by-condition in the hot loop.
• Do not validate SKDC, SDR, DCP, EPG, or another durable artifact with Seq; use its canonical Ledger definition.
• Do not read a Ledger store or memory root implicitly.
• Do not replace a missing operator with shell, jq, Python, a plugin, or a subprocess.
• Report unresolved units, identities, contamination, and evidence gaps.