GitHub - grandimam/wy: See why your AI agent made each change.

You handed a problem to an agent because you couldn't solve it yourself. Now it has changed 40 files, and the consequential choices were made without you. wy turns the uncommitted diff into the questions you'd need to be able to answer to accept it—and shows you which of them the agent's own record actually answers.

Two screens on the same hunks:

• Plan: the first screen. Four boxes—What changed (plain-language groups), What to worry about (questions a domain expert would ask, plus files with no record of why), Was it tested, and Progress. Every row opens its evidence. A worry carries what the agent chose, what it said (an exact quote, or an honest unknown), the code with markers, where to look yourself, and your verdict: accept · push back · unsure.
• Steps: the audit trail. The captured session as intent-sized steps, each with Why (verbatim or "no reason given"), Check, Gap, and the worries it raises; the original conversation is one link deeper.

Your verdicts persist and export as a review in your voice—push-backs first, so they paste straight back to the agent as the next instruction. That loop is the point: judgment formed → fed back → better next iteration, until you can ask the questions yourself before the agent runs.

It imports existing local Codex, Claude Code, pi, and OpenCode history. No recorder has to be running first. Attribution is a textual match between a hunk and a captured edit, or nothing—wy never assigns today's diff to a session because the tree happens to be dirty. Missing evidence stays missing; a nearby message is not a proven reason. Nothing calls a model until you open the ? row on the plan.

The screenshot illustrates the secondary file/history reader. The default screen is the plan.

Get started

Needs Rust 1.88+, Git and a C toolchain.

cargo install --git https://github.com/grandimam/wy --locked wy-code
cd /path/to/your/repo
wy

• Plan (p): the opening screen. What changed lists what the agent did, in plain words, one row per change group. What to worry about lists the questions you need to be able to answer (opening the ? row forms them with your reasoning CLI—explicitly, never on its own) and the files no captured edit explains. Was it tested is one line of fact and one of what was not. Progress says what you have read and decided. Enter on any row opens its evidence; in a worry y/n/m record accept / push back / unsure and i adds a note.
• Steps (t): one bordered box per step, in order: Why in the agent's own words or "no reason given", Check (build/test runs, fail → fixed), Gap (shell writes, unconfirmed captures), Plan (the worries it raises). Enter opens the step with its hunks and a link to the original conversation. Not explained by this session lists hunks no captured edit matches.
• Beneath Steps: the original transcript (o), the session decision brief (d), captured-edit comparison with current code, and commit history. These are drill-downs, not sections.
• Evidence is chosen from the diff. Every captured session whose edit text matches a hunk counts, newest first, bounded to four; the plan's evidence line says "1 session, …" or "N sessions explain X of Y files". Open it to see which, to hand-pick a different one, or to let the diff choose again. No list to scan, no guessing by date.
• Refresh (r): recapture history offline, keeping the selected session when available.

Plan / Steps is the whole navigation—one row at the top, full width for the content. Original conversations are a final drill-down, not a separate History section. As implemented shows captured code, including explicit execution and truncation gaps. A clean Git tree does not erase session work. No history means no session reconstruction—not an automatic fallback to today's diff.

Ctrl+Left/Right switches sections · s selects links · ? opens help.

Copy text: drag over text in the main pane; releasing the mouse copies the selection. Scroll while dragging, or hold the pointer below the pane to select more. Navigation panes are excluded—no keyboard mode switch needed.

Understand history and decisions

Omitting FILE:SYMBOL from /timeline uses the selected file/symbol; /decisions without an argument opens the selected-session overview. Symbol filtering is based on explicit text/record matches, not semantic causality. A timeline preserves tool and session boundaries; later edits are flagged as potentially making earlier context inapplicable, not automatically declared to supersede it.

Dates distinguish when the agent recorded an event from when wy captured history. Absolute UTC dates and relative ages are shown; missing dates are marked unknown rather than inferred from transcript file modification times.

How it works

1. Discover supported local histories whose recorded working directory belongs to this Git repository.
2. Capture up to 20 sessions / 40 MB, alternating between tools; /coverage explains exclusions.
3. Import the most recent captured sessions (up to 8) and preserve each supported edit in source order. Evidence is every session whose edit text matches a hunk in the diff, newest first, each only if it explains hunks no newer session already does (at most 4). With a clean tree, or no match, the latest session is the fallback—and the plan says so.
4. Hunk spine. Parse the HEAD→working-tree diff into hunks with current line ranges. Attribute each to captured edits by textual match: matched, partial (something else also touched it), or none. Lockfiles are derived and never questioned.
5. Steps. Segment each evidence session into intent-sized steps and merge them chronologically at request boundaries, decision records, and announced moves to disjoint files (optionally after a build/test run). Edit → failed test → fix stays one step. Each step keeps verbatim reasons, tentative rationale, checks and gaps.
6. Questions (on explicit request only). Send hunks, steps, verbatim requests and assistant statements. Validate every returned question: cited hunks must exist; recorded needs a verbatim quote from a statement in a step that produced the cited code; inferred needs cited hunks/steps and no quote; unknown carries nothing; check lines must fall inside their hunk. Failures are rejected and counted.
7. Verdicts are stored apart from the brief, keyed by question text + hunk ids, with a hash of each cited hunk. Changed code flags them stale; nothing is cleared.

Question briefs are keyed by repository, HEAD, exact hunk text and session snapshot: editing the tree, committing, or capturing new session content invalidates them. Tool invocation supports Codex and Claude only.

What the evidence means

• Recorded decision: a self-reported structured explanation, not independently verified truth.
• Nearby conversation: captured context; proximity does not prove causation.
• Available rationale: readable provider-exposed thinking or reasoning summaries; tentative, not a verified justification.
• Compaction summary: secondary context, not original speech.
• Inferred explanation: a new model assessment, separate from historical intent.
• Unknown: evidence was not captured or cannot establish the claim.
• Assessment: wy's own view of what an expert would want, or a plausible inferred reason. Never presented as the agent's intent.
• Stale verdict: your decision on code that has since changed; kept and flagged, not cleared.

Encrypted/redacted hidden reasoning cannot be recovered. Shell commands, formatters, and manual changes often have no supported edit records. A missing record does not mean the agent had no reason or that tests were never run.

Offline views never call a model. Requesting an explanation sends selected evidence to your configured reasoning CLI. Exports exclude raw tool outputs and tentative rationale, apply best-effort redaction, and require a preview before saving—inspect for sensitive content before sharing.

Status

Experimental (0.2.0). Useful for forming a judgment about substantial agent changes in a domain you are new to, and for returning to unfamiliar code. Not a replacement for code review, a correctness verdict, or a guarantee of original intent. Steps are a heuristic grouping; questions are an assessment whose citations—not whose semantics—are checked. Older OpenCode JSON storage is detected but not imported. Some provider formats and shell-driven edits remain unsupported.

Automated tests establish implementation behavior, not explanation accuracy or user value. See the evaluation protocol for measuring those.

Learn more

Usage · Decision records · Architecture · Security · Codex formats · Releasing

Development: cargo test --locked. Licensed under MIT.