Review consequences in context
Use this method once for the relevant workflow. Keep capability-specific facts
in the selected skill and its optional effects.json.
Before the action
Read the selected skill's complete instructions. Match its suitable cases, exclusions and prerequisites to the actual command, arguments, resources and existing task scope. A lexical match or declaration status does not establish applicability.
Start with Conductor's review_before. It is a compact index: operation-specific
assessment is deferred. Read the relevant sections of that source's effects.json
for expected changes, downstream dependencies and recovery/retry conditions.
Fetch just those details with MCP effect_details: copy source and
declaration_sha256 from the review into source and expected_sha256, then
request selectors such as /retry or /expected_effects/0. CLI/Python equivalents
are in the detail-reading guide.
Require selection_status: selected; a changed or unavailable declaration or
missing entry returns no fragments. Refresh the review and reassess changed
conditions before combining detail reads. Array indexes belong to that digest.
The capability-wide reversal/retry labels may cover several different operations.
Missing, invalid or stale declarations call for checking the relevant source;
they do not create a new approval requirement. An empty attention list is not a
safety verdict. Resolve material uncertainty proportionally to the actual task.
After the action
Use review_after and the step result. When present, inspect the referenced
effects_observed companion for affected resources, unfinished work or deviations.
Inspect a short companion already in context directly; targeted lookups add
exchange overhead and are useful when enough detail can remain deferred.
For a saved execution, MCP effect_evidence with source and overview: true
lists every step and its evidence read arguments under review_after.evidence_ref.
Use those arguments without overview to read the step. Omit selectors for an ID
index, then request records
such as /observations/o1 or /network/n1. Carry evidence_sha256 from the review
or index as expected_sha256; changed or incomplete selections return no records.
Linked calls identify the retained declaration at /declarations/DIGEST.
The evidence-reading guide
describes file access and limits; a saved observation is not a current-state check.
Process success, a written file and HTTP 200 support different, limited facts.
State what is verified, what remains unknown and the next useful action.
Before a dependent step, verify the prerequisite result it needs. After an
interruption, check current outputs and still-active operations. Conflicting or
invalid evidence cannot establish non-execution. Retain verified completed work
and reassess remaining actions. Use Conductor's --stop-on-failure to skip later
steps after a nonzero exit; this does not verify prerequisites, retry or undo work.
Review cues do not grant permission.
Reuse and improvement
Assess reuse when there is a concrete target: what transfers, what must change, and which target-context check supports it. Link evidence to the capability version and run context; do not promote reuse from a successful exit alone. Compare a proposed improvement with the current version on representative tasks, including failures and context cost, before generalizing the result.
Keep the handoff proportional: verified result, evidence reference, material uncertainty and next action. Routine checks can stay deterministic; this method does not require another LLM call for every tool use.
For CLI/Python/MCP integration and evidence limits, consult the common contract.
