@effect-agent/pr-review 0.1.0-beta.28 → 0.1.0-beta.29

package/README.md

@@ -1,210 +1,15 @@
 # @effect-agent/pr-review
 
-A bounded, fail-closed GitHub pull-request reviewer built on effect-agent's public APIs. Read-only
-Agents review host-partitioned pull-request evidence
-through typed ports; the host validates every finding anchor against the real
-diff and posts one review after the run settles.
+A small, provider-neutral review agent.
 
-**Deployment class E (ephemeral).** One `AgentRuntime.run` per invocation, no
-durability claim, and review posting is never exactly-once: a failed or
-truncated run posts nothing.
-
-## Use
+The package accepts an already-collected diff, makes one model call with no
+tools, validates the returned paths and line anchors, and returns a structured
+report with token usage. It knows nothing about GitHub, webhooks, comments,
+providers, retries, or review history; those belong to the host channel.
+Each finding carries a severity and problem category so a host can render a
+scannable review without another model call.
 
 ```ts
-import { Effect, Layer } from "effect";
-import {
-  PrReview,
-  describeReviewModel,
-  fullReviewExecutionContextLayer,
-  gitHubReviewLayers,
-  resolveReviewTarget,
-  makeOpenAiReviewModel,
-  openAiClientLayer,
-} from "@effect-agent/pr-review";
-
-const reviewer = PrReview.make({ model: makeOpenAiReviewModel() });
-
-const program = Effect.gen(function* () {
-  const target = yield* resolveReviewTarget({ repository: "acme/api", number: 123 });
-  return yield* reviewer
-    .run({ post: true })
-    .pipe(
-      Effect.provide(
-        fullReviewExecutionContextLayer("explicit direct full review").pipe(
-          Layer.provideMerge(Layer.merge(gitHubReviewLayers(target), openAiClientLayer)),
-        ),
-      ),
-    );
-});
+const reviewer = makeReviewer({ model, guidance });
+const outcome = yield * reviewer.review(request);
 ```
-
-The run's requirement channel keeps every real dependency visible: the
-`PullRequestSource`, `ReviewPublisher`, `ReviewAdjudicationHost`, and
-`ReviewExecutionContext` ports, the provider client, and the handler Layer of
-any extra tool you add. Direct callers provide an explicit full-review context
-as above; the packaged action supplies its selected incremental or full range.
-Anthropic is equally supported
-(`makeAnthropicReviewModel`, `anthropicClientLayer`), and the factory accepts
-any Effect AI Model.
-
-The built-in OpenAI binding can opt each request into Fast mode:
-
-```ts
-const reviewer = PrReview.make({
-  model: makeOpenAiReviewModel("gpt-5.6-sol", undefined, "fast"),
-  modelLabel: describeReviewModel("openai", "gpt-5.6-sol", undefined, "fast"),
-});
-```
-
-`fast` is the only packaged service-tier value. The helper sends it as the
-OpenAI Responses `service_tier`; omitting it leaves the request unset so the
-OpenAI project default applies. The Action and CLI reject a service tier when
-the selected provider is not OpenAI. The model label includes the tier and
-therefore changes both ordinary and profile fingerprints.
-
-## Adapt
-
-```ts
-const reviewer = PrReview.make({
-  model: makeAnthropicReviewModel("claude-sonnet-5"),
-  guidance: (mission) => [
-    "This is an Effect codebase. Flag naked Promises in public APIs.",
-    mission.changedFileCount > 50 ? "Large PR: prioritize breadth over nit depth." : "",
-  ],
-  ignore: ["**/*.lock", "dist/**"],
-  maxFindings: 10,
-  policy: { maxTurns: 20, maxToolCalls: 40, maxDuration: "10 minutes", toolConcurrency: 2 },
-  extraTools: [MyReadonlyTool], // must be annotated ToolExecutionClass "readonly"
-});
-```
-
-Model output is untrusted. The host revalidates anchors, enforces `maxFindings`, and publishes only
-after the run settles. `PrReview.makeFanOut` partitions bounded evidence and schedules general and
-specialist discovery for every unit, followed by independent candidate verification. Verifiers
-receive the exact claims and enough neighboring evidence to reject a plausible mistake, but no
-discovery reasoning.
-
-Each failed pass retries once, then carries its unit forward. The host accepts only confirmed
-candidate IDs, discards anchors outside assigned evidence, deduplicates claims in plan order, and
-derives the verdict from validated severities. Shared `guidance` reaches discovery and verification.
-
-When GitHub omits a textual patch, the source reads bounded, strict UTF-8 base/head content and
-labels it with non-anchorable `B`/`H` lines. Invalid UTF-8, binary NUL content, missing sides, and
-oversized files remain unreviewable and keep input coverage incomplete. Exclude them with ignore
-globs only when that is intentional.
-
-## Assurance model
-
-The public result deliberately separates two claims:
-
-- **Input coverage** says every reviewable required path in the selected
-  scope was assigned every deterministic bounded evidence shard and
-  separately names partially assigned and over-capacity paths, the exact
-  unassigned-shard count with a bounded identifier sample, undiffable paths,
-  and source truncation. A large textual diff is split across shards and
-  units instead of being called partial merely for exceeding one prompt
-  chunk. Input coverage does not say the model understood the evidence.
-- **Review assurance** says every scheduled general discovery pass, required
-  independent specialist pass, and candidate-verification pass settled after
-  at most one retry. It reports discovered, confirmed, rejected, unsettled,
-  and discarded (invalid-anchor) candidate counts plus failed pass IDs. A
-  failed pass is a reviewer-side gap: its unit's paths carry forward as
-  retryable scope and the next incremental run re-reviews exactly them.
-
-`settled` assurance means the configured work completed; it never means the
-review is exhaustive or the pull request is defect-free. Risk classification
-is deterministic host policy over paths and bounded text and intentionally
-favors redundant work, but it cannot recognize every semantically risky
-change. The specialist pass is context-independent redundancy, not a claim of
-provider or model diversity. Running it for every unit prevents classifier
-silence from suppressing scrutiny; the category labels still cannot prove
-that every semantic risk was recognized. Hosts and UI use `inputCoverage` and `assurance`.
-The flat reviewer has path-input accounting but no independent verifier, so
-its assurance is `unverified` and the Action check cannot report success from
-that shape.
-
-## Maintainer adjudication
-
-A maintainer can settle a finding without changing code, from the pull request itself:
-
-- On a finding's inline thread, reply `/adjudicate accepted-risk|refuted|obsolete[: reason]`.
-  The thread names the target, so the verb alone suffices.
-- For an unanchored concern, comment
-  `/adjudicate <disposition> "<exact concern title>"[: reason]` in the PR conversation. The quoted
-  title is required and must match exactly.
-
-An adjudicated identity leaves active findings, verdict counts, and the check conclusion; it
-renders in a collapsed "Adjudicated" section instead, and the final audit distinguishes fixed,
-adjudicated, and still-open items. The reviewer prompt names each adjudication so the model does
-not re-raise it without materially new evidence. Identity is exact — path, line range, and title
-for findings, title alone for concerns — so a materially different finding at the same location is
-untouched. Adjudications persist in the signed review state, and the skip-unchanged path re-reads
-them, so an adjudication lifts a blocking check without a new commit.
-
-**Authorization is fail-closed**: only comments whose `author_association` is OWNER, MEMBER, or
-COLLABORATOR adjudicate. Anything else — third parties, bots, malformed commands — is ignored
-(logged at debug). Later adjudications of the same identity win by comment creation order, bounded
-at 20 stored entries (oldest dropped with a logged notice). If either GitHub listing surface fails,
-or one thread exceeds the bounded authorized-command history, fresh collection is discarded and
-the stored set stands unchanged.
-
-**Rejected alternative — parsing free-text rebuttals** ("this is fine because …" replies): only an
-explicit, authorized verb is auditable and fail-closed. Inferring intent from prose would let model
-output or third-party comments silently suppress findings, and nobody could later say which comment
-dismissed what.
-
-## What a posted review looks like
-
-The host derives callouts, statistics, walkthroughs, inline comments, and the final prompt from
-validated findings. Model prose cannot choose the verdict or smuggle paths outside the changeset.
-The [Action guide](../../action/README.md) documents the rendered review and workflow outputs.
-
-## Swap a port
-
-Tools observe the pull request only through `PullRequestSource`; publication
-happens only through `ReviewPublisher`. Provide your own Layers to review
-any diff-shaped input or publish elsewhere. The GitHub REST adapters are
-one implementation, not the contract.
-
-## Test what you adapted
-
-```ts
-import {
-  fixturePullRequestSourceLayer,
-  collectingReviewPublisherLayer,
-  makeOfflineReviewerModel,
-  makePromptKeyedModel,
-} from "@effect-agent/pr-review/testing";
-```
-
-Deterministic in-memory adapters for both ports plus prompt-keyed scripted
-models that call the real Tools. Tests need no network or credentials and exercise every ordinary
-gate.
-
-## Action and CLI
-
-The prebuilt GitHub Action handles signed incremental state, fail-closed scope selection, sticky
-progress, stale-review retirement, and final audits. Failed stages remain attached to their exact
-unchanged paths; failed verification reopens discovery only for its paths because candidates are
-not persisted. See its [setup and behavior guide](../../action/README.md) and
-[input reference](../../action/action.yml). Custom hosts can import `runReviewAction` from
-`@effect-agent/pr-review/action`.
-
-For an amended or force-pushed head, the GitHub adapter compares complete commit tree snapshots
-across a bounded PR and continuity path set. Selection then uses the current full PR records, so
-patches and rename metadata remain intact. An unavailable, malformed, or truncated tree falls back
-to the full PR and records the cause in the review reason.
-
-Run the CLI with
-`bun src/cli.ts --repo owner/name --pr 123 [--post] [--provider anthropic] [--service-tier fast] [--fan-out]`.
-
-## Bounds, spelled out
-
-Every definition has a finite `AgentPolicy` and run-level `UsageBudgetLimits`. The reviewer rejects
-file versions beyond 200k bytes or characters and patchless B/H renderings beyond 220k characters.
-It reads at most 300 changed files. Each fan-out unit holds at most 12 files, 12 complete evidence
-shards, and 240,000 evidence characters. At most eight units produce 24 attached children with
-child concurrency capped at four. The result reports every capacity overflow. Any blocking
-finding, input gap, failed pass, exhausted pass, mismatched candidate batch, or unsettled
-verification prevents a green check.