Fallow Review: graph-grounded code review

fallow review turns a changeset into a deterministic, graph-derived review brief and an agent-contract loop. It answers "where do I look, and what calls actually need human taste?" rather than "will CI block this?" (that is fallow audit, which gates). review always exits 0, so it is safe to run regardless of verdict.

The four jobs it does, in order:

• subtract: dead code, complexity, duplication, and styling for the changed files are reported and kept OUT of the judgment loop, so attention is not spent on what a deterministic check already owns.
• focus: changed-file units are ranked by a composite attention score (fan-in/out, risk zone, change shape) with a review-here / not-prioritized label and a full deprioritized escape-hatch list, so nothing is hidden.
• structure: the decision surface lifts the handful of consequential STRUCTURAL decisions out of the diff and frames each as a judgment question, capped to a working-memory-sized set, each anchored to a signal_id fallow emitted.
• direct: the walkthrough guide hands an agent a graph-derived digest, the review direction, a graph-snapshot pin, and the exact judgment schema to return.

When to use

• Reviewing a PR, branch, or diff (AI-generated or human-written) before merge.
• After an agent has done work and removed the fallow findings it could; this surfaces what is left for human/agent taste.
• Producing inline-reviewable judgments that can flow back to the agent that wrote the code.

When NOT to use

• Gating CI on a pass/fail verdict: use fallow audit (it exits non-zero on a fail verdict).
• Whole-project health, cleanup, or dead-code reports: use the fallow skill.

Quick human brief

# Auto-detect the base (merge-base against the upstream / remote default):
fallow review

# Pin the base, or scope to a precise diff:
fallow review --base origin/main
git diff --find-renames origin/main...HEAD | fallow review --base origin/main --diff-stdin

The human brief prints the orientation facts, the focus map, and the decision surface ("Decisions to make"). --format json emits the full structured envelope (decisions, focus, deltas, impact_closure, partition, graph_facts). --max-decisions N tunes the cap (clamped to a small band). --show-deprioritized expands what the focus map collapsed.

The decision surface

Each decision is a framed question anchored to a signal_id fallow deterministically derived from the graph (a delta key or a coordination-gap key). There are exactly three shippable categories:

• coupling-boundary: a new cross-zone dependency edge.
• public-api-contract: a new exported public-API surface, or a changed contract consumed by modules OUTSIDE this diff (a coordinate-or-confirm signal).
• dependency: a changed package.json that adds third-party entries or moves a declared entry across a major version (or a 0.x minor). One batched decision per manifest per kind, weighted by the graph's in-repo importers of the affected packages: blast is the importer count, internal_consumer_count the importers outside the diff. Both counts are a union over the batched packages. An entry outside dependencies shows a (dev), (optional), or (peer) tag in the question. digest.deltas carries the keys as dependency_added (<manifest>::<name>) and dependency_major_bumped (<manifest>::<name>@<from>-><to>); both lists are always present, possibly empty. Minor and patch bumps and non-numeric ranges (workspace, file, git, tags) are never candidates. A dependency decision has no suppress action: it anchors on package.json, which cannot carry a comment, so never paste a // fallow-ignore line into a manifest. The decision_surface MCP tool surfaces the same dependency decisions as the CLI. Framing rule: cite the changelog question ("which documented change in <from> to <to> reaches these N importers?"); do not guess the behavior change.

A decision may carry previous_signal_id when its anchor file was renamed in the change: that is the signal_id the same decision would have had at the old path, so a review surface can re-attach a prior reviewer comment across a git mv.

Eliciting the broader trade-offs (taste ownership)

The decision surface above is the DETERMINISTIC slice: only the trade-offs fallow can prove from the graph (the three categories). Real architectural trade-offs are broader (abstraction level, error-handling strategy, data-model shape, eager-vs-lazy, state ownership, extensibility-vs-YAGNI, testability, trust boundaries), and none of those are graph-detectable. Surfacing them needs a model reading the diff, not a static pass.

Run the trade-off elicitation prompt in references/tradeoff-elicitation.md over the diff plus the guide. It applies TASTE OWNERSHIP: the model makes each choice legible and frames a genuinely open question; the human decides. The prompt carries the honesty contract (diff anchoring, the observed / tradeoff / question split, deterministic: false fencing, provenance, the top-five cap with honest abstention, no duplication of the deterministic surface, options that never pick). fallow owns what it can prove, the prompt covers the rest, and the fencing keeps the two from being confused.

Running it as a review step

When you want the trade-off surface alongside the deterministic decisions:

1. Fetch the guide and read the diff:

   fallow review --base <ref> --walkthrough-guide --format json > guide.json
   git diff <ref>...HEAD            # or: git diff --cached  for staged work
   

   Copy
2. Run the references/tradeoff-elicitation.md prompt over the diff plus guide.json.
3. Keep the envelope (the { graph_snapshot_hash, abstained, tradeoffs[] } shape). The composed review below uses it.

This step is deliberately NOT part of the post-validate agent-contract loop below. The trade-off envelope is never round-tripped through fallow review --walkthrough-file (which only validates emitted signal_ids), so it carries NO fallow-grade guarantee: every item stays deterministic: false, agent-self-checked against the diff, not graph-validated. Do not present it as fallow-validated.

The agent-contract loop

The loop lets an agent produce judgments that fallow post-validates against the live graph. The verifier is the graph, not a second model.

1. Fetch the guide:

   fallow review --base origin/main --walkthrough-guide --format json > guide.json
   

   Copy

   The guide contains: digest (the brief + decision surface), direction (where to look), graph_snapshot_hash (the staleness pin), agent_schema (the exact shape to return), and injection_note. The digest is built from the graph ONLY; PR prose is never folded in, so the guide is injection-resistant by construction.

2. Read the decisions in digest. Each carries a signal_id, a category, the framed question, and an anchor_file / anchor_line.

3. Return a judgment matching agent_schema:

   {
     "graph_snapshot_hash": "<echo the value from the guide>",
     "judgments": [
       {
         "signal_id": "<one signal_id fallow emitted>",
         "framing": "<your reasoning for the human reviewer>",
         "action": "<block | address | consider | fyi>",
         "concern": "<optional: one lens from agent_schema.concern_vocabulary>"
       }
     ]
   }
   

   Copy

   Every signal_id MUST be one fallow emitted in the guide (emitted_signal_ids). An unanchored id is rejected. Echo the graph_snapshot_hash verbatim.

   For a changed region without a finding, a judgment can cite an emitted change_anchor (a chg: id from the guide's change_anchors) in place of a signal_id. Each anchor carries file, start_line and line_count. An unknown id is rejected as unknown-change-anchor.

   action tells the author what the judgment asks of them: block and address are required actions, consider is optional, fyi needs nothing. concern names the lens; prefer the guide's agent_schema.concern_vocabulary, thirteen kebab-case lenses: abstraction, coupling, data-model, error-handling, control-flow, performance, dependencies, api-ergonomics, compatibility, state-ownership, extensibility, testability, trust-boundary. The guide publishes both lists as agent_schema.action_vocabulary and agent_schema.concern_vocabulary; read them from the guide rather than from this file. action_vocabulary is enforced; concern_vocabulary is advisory, any string is accepted.

4. Post-validate:

   fallow review --base origin/main --walkthrough-file judgment.json --format json
   

   Copy

   The response sorts each judgment into:

   • accepted: the signal_id was emitted and the snapshot matches; the agent's framing is fenced as non-deterministic (deterministic: false) and never gates. The action is echoed next to agent_framing, fenced the same way: it is an instruction to the author, never a fallow fact.
   • rejected with reason: "unanchored-signal-id": the signal_id was never emitted (a hallucination). Drop or correct it.
   • rejected with reason: "invalid-action" and invalid_value (the label fallow refused): the action is outside the vocabulary. This is reported only after the anchor resolved; a hallucinated anchor plus a bad label rejects as unanchored-signal-id / unknown-change-anchor instead. Fix the anchor first, then the label.
   • rejected with reason: "stale-snapshot" and stale: true: the tree moved since the guide was fetched. Re-fetch the guide and redo the judgments.

Production runtime data (Fallow Cloud)

When the project sends production coverage to Fallow Cloud, check the changed functions against it. Use the scoped reads, not the full runtime-context pull:

• Before an edit: fallow coverage review-packet --repo <owner/repo> --base <ref> (MCP: get_cloud_review_packet). Read hit_count, covered_by_test, and the callers when present. A hot function with covered_by_test: null is a high-risk edit.
• Before a delete: require period_tracking_state: "never_called" over the period, not only tracking_state, and read evidence_window.observed_hours. Low observed hours or a low-traffic surface mean "not visited", not "dead".
• After a deploy: fallow coverage deployment-changes --repo <owner/repo> (MCP: get_cloud_deployment_changes). When comparable is false, report reason and claim no change.
• Open files by repo_path, not file_path.

The data is context: cite it in a framing, never as a gate. Field meanings and the full recipe: references/production-runtime-data.md.

Compose the review

Validation is not the review. Once the judgments are accepted, render them for a human in this fixed order, so the reader lands on what changes the outcome first and the deterministic remainder last:

1. Accepted decision judgments, in direction.order. Each one carries the digest question and tradeoff verbatim, then your framing, then the number that gives it weight: internal_consumer_count, the out_of_diff paths, or the unit's scoring_budget.
2. Trade-offs from the elicitation step, ranked by consequence, at most five. When the envelope is abstained: true, print one line saying so; do not fill the slot.
3. Subtract as ONE line with the counts from the brief: "handled deterministically: N dead-code, N duplication, N complexity, N styling; not in the discussion". Never re-derive any of these from the diff; the brief already owns them.
4. Deprioritized as one line with the count and --show-deprioritized as the escape hatch, so nothing is hidden and nothing is padded.
5. A review with zero decisions and an abstained trade-off envelope is a complete review: "nothing consequential; deterministic findings: N". It is not a failure and never a reason to invent items.

Three graph facts feed the composition. Only a test_reach note can add an item, and it anchors on a change_anchor:

• Each direction unit carries test_adjacency: none (no test file imports this unit), untouched (a test importer exists but is not in the diff), or changed (a test importer is in the diff). It is absent for test files and when the graph was not retained. For a review-here unit with none, ask the author for the verification story in the framing. Never claim coverage; the value says whether a test imports the unit, not whether it exercises the change.

• digest.test_reach gives the reach of each changed value export. status is measured or no-tests. counts holds every class. exports lists rows only for the classes that are not direct or framework, each with file, export_name, line and reach:

  • transitive: a test reaches the export, but no test names it.
  • mocked-only: the tests replace the file with a mock (mocked_targets), so the real code does not run.
  • none: a test reaches the file, but no test reaches the export.
  • none-file: no test reaches the file.

  Use these rules for test_reach:

  • Absent means no fact. Older fallow versions do not emit the field, and it is absent when no value export changed. Then use test_adjacency.
  • For a review-here unit whose file has a mocked-only, none or none-file row, cite the row (file, export, class). Ask the author for the verification story.
  • A row has no signal_id. Anchor the note on a change_anchor in the same file whose range overlaps the export declaration. When no anchor overlaps, use the first anchor in that file. When a decision for that file exists, put the note in its framing.
  • When test_reach and test_adjacency disagree, use test_reach, because it is per export and it sees mocks.
  • Never write "tested" or "covered" from these facts. direct means that a test names the export, not that a test runs the changed lines.
  • The reach set is open. An unknown value needs attention.
  • When exports_omitted is above 0, the list is incomplete. A file without a row is not proof of direct reach. Use counts.
  • For status: "no-tests", write one orientation line, not one line per export.
  • For masking: "capped", the graph cannot see mocks. The absence of mocked-only rows proves nothing.
  • End-to-end tests do not import files. In a project with an end-to-end runner, a none-file row can still have end-to-end tests. Ask, do not assume.

• digest.partition.independent_slices lists the connected components of the inter-unit dependency graph, each a sorted list of module directories. It is present only when there are two or more slices; absent means the change is one connected piece. When digest.triage.risk_class is high and the field is present, name them as an orientation fact: "this change splits into K independent slices along a graph-proven seam". It is never a demand to split; the author owns that call.

Two rules govern the prose. Leverage first: one structural decision plus ten small notes means the decision IS the review; the notes ride below it or not at all. Numbers, not adjectives: every framing cites at least one number or path from the guide (a consumer count, an out-of-diff path, a scoring_budget). "Could be slow" is not a finding; "imported by 14 modules, 9 outside this diff" is.

Human-in-the-loop walkthrough (terminal)

The agent-contract loop above carries the AGENT's framing. The SAME loop carries a HUMAN's verdict with the identical graph-validated, anti-hallucination guarantee, so a terminal reviewer can leave notes that fallow anchors and round-trips. The contract is the existing --walkthrough-guide / --walkthrough-file round-trip, not a new surface.

The human owns the taste; you only carry the note. fallow validates the ANCHOR (the signal or changed region exists), never the note's correctness; every carried note stays deterministic: false and never gates.

1. Render the tour and read the anchors:

   fallow review --base origin/main --walkthrough                       # the staged human tour
   fallow review --base origin/main --walkthrough-guide --format json > guide.json
   

   Copy

   guide.json carries the decision signal_ids (the framed structural questions), a per-changed-region change_anchors set (each { "change_anchor": "chg:<hex>", "file", "start_line", "line_count" }), and the graph_snapshot_hash staleness pin. Surface the tour to the human and collect, per item they choose to flag, a short verdict/note plus the action they want from the author. Offer only the four labels (block, address, consider, fyi); an unknown label is refused with invalid-action.

2. Carry each human note as a judgment (echo the hash verbatim; cite a signal_id fallow emitted for a flagged decision, or a change_anchor for any other changed region the human notes; carry the human's action as action):

   {
     "graph_snapshot_hash": "<echo from guide.json>",
     "judgments": [
       { "signal_id": "<an emitted decision signal>", "framing": "<the human's verdict/note>", "action": "address", "concern": "<optional>" },
       { "change_anchor": "<an emitted chg: id>", "framing": "<the human's note on this region>", "action": "consider", "concern": "<optional>" }
     ]
   }
   

   Copy

3. Validate the capture:

   fallow review --base origin/main --walkthrough-file judgment.json --format json
   

   Copy

   The response sorts as in the agent-contract loop, with two additions: accepted carries anchor_kind ("signal" or "change"), and a change_anchor fallow never emitted rejects as unknown-change-anchor. Re-anchor to a real signal or region; do not invent one.

4. Act: relay the accepted human verdicts into the coding session. Carry each note's action so the receiving agent can triage: block and address are required, consider is optional, fyi needs no change. The note arrives anchored and fenced, never as a fallow-grade fact.

The guarantee: the human cannot anchor a note to a signal or region fallow did not emit, and a note left against a moved tree is refused rather than silently mis-mapped.

Rationalizations the loop rejects

Notes

• review is an alias for audit --brief; --format is orthogonal to the brief.
• See the fallow skill for whole-project analysis, and its references/cli-reference.md for the full flag list.