One constructor, selectable review contracts. Reuse the installed $actuating
implementation; do not fork its construction doctrine or invoke its native
review-closeout underneath a different strategy.
accepted Goal -> Actuating implement -> local proof -> Ship
-> Architectonic -> Elenctic -> joint evidence fold
-> supported successor through Actuating implement -> proof -> Ship -> fresh round
The default is review=combined. Architectonic runs before Elenctic on one frozen
candidate; neither review can authorize a mid-round edit. Ergasterion owns the
composition's scheduling, aggregate and completion, not reviewer judgments,
construction selection, native Actuating review credit, or public effects.
Invocation and route selection
$ergasterion <accepted goal>
$ergasterion implement <accepted goal>
$ergasterion analyze <question>
$ergasterion review PR #123
$ergasterion review-closeout PR #123
$ergasterion review-closeout PR #123 review=actuating serial-reviews
$ergasterion review src/runtime review=architectonic
$ergasterion review PR #123 review=elenctic with concurrency 10
$ergasterion review PR #123 for resolution
$ergasterion PR #123 for construction
$ergasterion review PR #123 for construction
$ergasterion resume
These are intent selectors, not a shell CLI. Extract the subject separately; never pass mode words or untrusted text as executable shell syntax.
| Operation | Work and terminal boundary |
|---|---|
Bare invocation / run / /goal $ergasterion with an accepted Goal |
Actuating local implementation, authorized Ship, selected review-closeout; do not stop at the first implementation or PR. |
| Bare invocation with an existing PR and no narrower intent | review-closeout on that PR; bind accepted obligations before any correction. |
implement |
Exact $actuating implement: local construction and proof, no review, push or PR. |
analyze |
Exact $actuating analyze: read-only construction judgment, no new review campaign or corpus write. |
review, inspect, audit, or unqualified review intent |
One selected assessment/aggregate, no implementation or closeout streak. |
review-closeout, fix/resolve/address with that intent |
Selected strategy's review, supported corrections, proof, publication and convergence. |
resume |
Recover the exact established run/analysis, subject, operation and strategy; never infer a new run from today's branch. |
Without an explicit operation, for resolution selects one-shot review, while
for construction selects feedback-only proposal analysis. Resolve this workflow
intent before the bare-invocation default. review for construction explicitly
requests fresh selected assessments, still without implementation authority.
An explicit prohibition always narrows authority. Missing Goal authority blocks only dependent implementation. A supplied spec is the source, not permission to invent requirements. Apply Actuating's source-binding and proof-acceptance rules before realization and local completion.
Independent choices
| Choice | Meaning |
|---|---|
review=combined (default) |
Architectonic then Elenctic, followed by one joint fold. PR required. |
review=actuating |
Delegate bare/implement/analyze/review-closeout to Actuating unchanged; its native standard/auxiliary contract, scheduling, recovery and counts remain authoritative. |
review=architectonic |
Architectural assessment; PR, subsystem, repository and exact proposed-design subjects remain supported. |
review=elenctic |
Elenctic's PR campaign or feedback-analysis workflow with its current scope/provenance requirements. |
for resolution |
One grouping pass over retained aggregate findings; no implementation authority. |
for construction |
Proposal-only synthesis of supplied established evidence or existing PR feedback. No fresh assessment unless explicitly requested with review for construction. |
parallel-reviews / serial-reviews |
Scheduling within a reviewer, never simultaneous Architectonic/Elenctic stages. See the selected review guide. |
with concurrency N |
Elenctic's existing 1–20 bounded worker setting; not a total file budget. |
summary-only, no-drafts, report-only, text-only, no-file |
Preserve the requested presentation/effect limit; a required snapshot may remain unavailable under no-file limits. |
Default presentation is comments: evidence-backed findings and eligible drafts,
not posted comments. Review strategy does not make implement or analyze
review-bearing. Architectonic subject selectors and execution strategies are not
new public workflows. Preserve every current sibling mode, not removed Elenctic
standalone-file, range-only, or session-aggregation selectors.
Reject incompatible combinations before their effects. In particular: do not
combine implement/closeout with proposal-only construction or resolution; do not
pretend Actuating has a fresh read-only review campaign; do not run Elenctic on
a repository/design without a unique open PR; and do not silently ignore a
concurrency setting on a strategy without Elenctic. Use Actuating analyze for
its read-only route. Explicit serial scheduling and concurrency greater than one
conflict. Ambiguity about intent never supplies mutation authority.
Selected guidance and owners
Read the installed sibling entry points and only their selected mandatory guides. Resolve siblings from the actual installed skills root, not a similarly named package in the target repository. Bind selected policy bytes/identities before review. Never substitute a missing reviewer with generic prompting.
| Selected work | Read before its governed decision/effect |
|---|---|
| Local implementation or construction analysis | Actuating and its selected guides. |
| Any non-native assessment, combined round, or closeout | Review composition, then the selected reviewer entries and the explicit Elenctic composition contract. |
| Admission, aggregate adjudication or successor input | Aggregation and Review Fold. |
| Every aggregate delivery, including default comments; resolution or construction synthesis | Reporting and synthesis. |
| Push, PR creation/update/promotion or final public readback | Ship, preserving exact-head readiness and assignment rules. |
| Native strategy | The selected Actuating route alone; no composite campaign, report identity, or extra auxiliary wave. |
Actuating owns construction, causal family synthesis, realization and local proof. Review Fold owns applicability, law authority, deduplication and witness admission. Architectonic and Elenctic own their original assessments. Ergasterion owns only the outer non-native review contract and its completion statement. Ship owns publication. CAS/native runtime facts retain their existing owner; an aggregate report is never an invented CAS receipt or native standard clean.
Authority and common invariants
Explicit invocation selects this composition, including its named explicit-only reviewers; it does not change their standalone activation or permission policies. Reviewers remain read-only even in an implementation run. Ergasterion uses Elenctic with no GitHub mutations; Viewed and own-thread resolution are not part of combined closeout. Native-strategy effects remain governed by native owners. No route posts review comments, approves, merges or weakens model/permission policy.
Pass corpus_write_authorized: false for all review, analysis and presentation
folds. A later constructor handoff can capture accepted witnesses only under its
actual enclosing write authority; no new corpus/store or fallback writer exists.
Parent authority does not authorize worker source, index, branch, or config writes.
Safe inspected targeted checks and isolated scratch witnesses retain each owner's
limits; paid calls, external load, new providers and private-data export need their
own authority. Explicit no-file limits also prohibit report and policy-snapshot writes.
Preserve the implementing session's resolved model and reasoning configuration on review requests and verify the actual runtime selection. A requested serial route does not waive Elenctic's prepared-context or native-fork requirements. Missing capability/evidence holds dependent work; no silent reviewer downgrade or fake pass.
Keep accepted source, exact candidate, selected strategy, policy identities, current round/attempt references, aggregates and unresolved work in the existing thread or accepted task specification. Reuse source reports; do not create a durable workflow registry, copied counterexample store, generated task framework or mandatory receipt. Runtime IDs and hashes identify actual outputs; prose cannot reconstruct lost credit.
Completion and recovery
Non-native closeout requires five consecutive credited rounds on one unchanged
proved PR candidate under the same Goal, scope, strategy and selected policies.
A combined round contains both assessments and their joint fold; single-reviewer
strategies credit only their explicitly selected scope. This is an Ergasterion
policy, not equivalence to five native Actuating confirmations or universal proof.
review runs once; local implement and analyze acquire no streak requirement.
Use review.md for admission, holds,
resets, successor barriers and final exact-head publication checks.
Resume from direct run/campaign/worker/report references; use $seq only for exact
recovery when needed. Refresh source, PR identity, policy and relevant discussions.
Continue incomplete assignments, not whole reviews merely to obtain approval.
Changed inputs invalidate dependent evidence/credit; retained old findings remain
historical until revalidated. Strategy changes require explicit caller intent and
a fresh credit basis, not conversion of previous cleans. Missing state blocks only
the unsupported claim; never manufacture an earlier confirmation.
Report the actual result, strategy, reviewed head, coverage, retained findings, accepted/rejected changes, credited rounds and limitations. Distinguish local completion, assessment completion, proposal delivery and review-closeout. Keep source verdicts and optional opportunities visible. An unresolved required liability, proof loss, material coverage gap or missing final public readback forbids unqualified closeout. Complete requested independent work and do not invent an iteration cap.
Skill-development cases are in evals/scenarios.md. They are change-validation cases, not a runtime checklist or an added review lane.