whyline

Use Claude Code and Codex on the same project without either one starting blind.

$ whyline explain src/tsconfig/resolve.ts:41

Decision          Treat only canonical in-repo workspace package exports as internal
Because           node_modules must be readable for resolution but never indexed
Rejected          classify any resolvable node_modules target as internal
                  indexes third-party declarations and violates FR-005
Confidence        High — a recorded decision matches the commit for this line.

Free, Apache-2.0, local-only. No accounts, no telemetry, no paid tier, ever.

The problem this solves

Claude Code and Codex are good at different things. You might want Claude to plan a feature and write the tests, then Codex to review the diff — or the reverse. That combination is genuinely useful, and today it is genuinely painful: the second agent starts from nothing. It has no idea what the first one concluded, what it tried, or what it deliberately ruled out. So you re-explain, or paste, or just give up and use one agent for everything.

whyline fixes that with a decision record both agents write to and read from. The one that finishes leaves behind what it decided and what it rejected; the one that starts picks it up. Neither has to be told twice.

On the subscriptions you already pay for. whyline never touches a credential. It launches the vendor's own CLI with exec, so Claude Code authenticates as Claude Code and Codex authenticates as Codex. No API keys, no per-token billing, nothing metered on top of what you already have.

And because the record is committed to your repository as plain Markdown, it outlives the session. Six months later, whyline explain still answers why a line of code exists — which is the same mechanism, read at a longer horizon.

Install globally, initialise per project

Once per machine:

uv tool install whyline

That one command is everything: the relay and the full-screen console come with it, so there are no extras to remember. (Already installed an older version with extras? uv tool install --reinstall whyline brings it in line; the old whyline[relay], [console] and [ui] names still work but add nothing.)

Once per repository:

cd your-project && whyline init

The executable is global; do not reinstall it inside every directory. whyline init is per Git repository (and per worktree/checkout): it creates the project-local .whyline/ state, adds shared instructions, and installs both Claude Code and Codex project hooks. This separation is intentional — decisions and active work from one project must never bleed into another.

Codex requires explicit trust for non-managed project hooks. After init, open /hooks once in Codex and approve the Whyline definitions — they list as Installed 1, Active 0 until you do, and an untrusted hook cannot run. whyline status will say “configured but never observed” until a real event arrives; it does not mistake a JSON file for a working hook.

Trust applies to events that have not happened yet. The session you approve from has already passed its SessionStart, so restart Codex once afterwards if you want that event recorded. Everything else — prompts, tool calls, session end — starts recording immediately, so status flips to “observed” on your next prompt either way.

init asks before touching AGENTS.md, CLAUDE.md and the hook files, and pressing Enter accepts — running the command is the consent. Answer n, or pass --no-instructions / --no-hooks, to skip either part; --yes accepts both without asking, for scripts. Whatever block init replaces is copied to .whyline/AGENTS.md.bak first.

Python 3.11+, plus git. The decision log, hooks and every command other than the console and relay use only the standard library; the three dependencies (whyline-relay, prompt_toolkit, textual) are there for the console and the relay.

Re-run whyline init any time; it upgrades an outdated instruction block in place and leaves everything you wrote around it untouched.

Automated mode (optional)

Everything above is manual: you switch between an agent's terminal and the next. There is also whyline-relay, installed with whyline, that does the switching for you. Give it a Markdown plan and it runs each task through Codex (implements) and Claude (reviews and commits), routing on whyline's own handoff record, and stops when something needs a human. It launches agents unattended and spends your subscription quota, so although it is installed, it is never set up in a repository or run unless you ask.

whyline init --relay                  # normal setup, then the relay's own setup
whyline relay plan-format             # how to write a plan (and a prompt for an AI that drafts one)
whyline relay doctor                  # checks whyline, both agent logins, the plan and the tree
whyline relay start                   # run the plan
whyline relay status                  # is it running, or where did it pause

whyline init asks whether to set the relay up, and the default is no: --yes alone does not opt in, --relay does, and --no-relay skips the question. You can add it to a repository later with whyline relay init and take it out again with whyline relay remove.

The command is whyline relay <command>; there is no separate whyline-relay command on your PATH, and whyline never needs one. If you want it anyway, install the relay on its own too: uv tool install whyline-relay. The relay's README covers plans, phases, permissions, what it costs and how to read a pause.

Then just work

Open Claude Code or Codex and build as you normally would. Two things can happen without you doing anything:

• each vendor's hook records sessions, prompts and explicit file edits;
• your agent records its own decisions and rejected alternatives, because it read the instruction init added.

Automatic reading is best-effort, not guaranteed. Use whyline run for a handoff that must reach the next agent, or run whyline sync explicitly after opening a vendor CLI directly.

Switching agents — the thing this exists for

There are two ways to do it. Both work. They fail differently, so pick with your eyes open.

Pattern 1 — two terminals (recommended)

Each agent gets its own session. You decide who does what by which tab you type in.

# tab 1 — implementation, with context attached reliably
cd your-project
whyline claim WL-42 --actor codex --role implementer --file src/cache.py
whyline run codex "implement bounded cache invalidation" --task-id WL-42

# when implementation is ready
whyline handoff WL-42 --from codex --to claude --status ready-for-review \
  --summary "bounded invalidation implemented" \
  --file src/cache.py --test "pytest -q: passed" \
  --risk "large repositories not benchmarked"

# tab 2 — review, with the same project context attached
cd your-project
whyline run claude "review and commit the cache change" --task-id WL-42

You are the switch; whyline is the relay. run attaches the history directly, so it does not depend on an agent remembering to call brief. You can still open codex and claude directly, but automatic reads from the repository instruction were measured at only 43%, so that path is best-effort rather than guaranteed.

claim is advisory. If both agents claim the same task or file, Whyline warns in claim, sync, and status; it never locks a file or blocks either agent. This is how two terminals can coordinate without Whyline becoming an orchestrator.

Run run from your shell, not from inside an agent session. It replaces the current process with the agent (exec), so launching it from within another agent's tool call gives the new agent no terminal and it will fail.

Pattern 2 — dispatch from inside a session

Ask the agent you are already talking to to call the other one:

get codex to review the caching change

Claude runs codex exec … and the result comes back into your current conversation. Convenient for a one-shot second opinion, with three limitations:

• A dispatched agent follows its dispatcher's prompt, not AGENTS.md. In one observed run, an orchestrated task recorded no decisions, while the same agent leading its own session on the next task recorded two and read the history unprompted. If provenance matters for a piece of work, let the agent own its session.
• It is one shot. The dispatched agent has no terminal, so it cannot ask you a clarifying question or iterate — it answers once and exits.
• It does not work in reverse. Codex cannot launch Claude, because Claude Code writes session state outside the workspace and Codex's sandbox blocks that. Do not disable the sandbox to force it; use a second terminal.

Either way

Codex cannot commit. Its sandbox blocks writes to .git, so it will implement, test and report, then stop. You or Claude makes the commit. That is a fixed cost of the sandbox, not a whyline limitation.

Everything else is optional:

whyline sync --task WL-42        # active handoff + Git + ownership + decisions
whyline brief --file src/a.py    # decisions-only, relevant and token-bounded
whyline explain src/a.py:14      # why does this line exist?
whyline note "chose X" --because "Y" --rejected "Z: too slow" \
  --file src/a.py --actor codex --role implementer --task WL-42
whyline release WL-42 --actor codex
whyline timeline --file src/a.py
whyline status                   # is recording actually live?

What context Whyline keeps

Whyline does not preserve either vendor's hidden conversation. It keeps a small, explicit relay that both can read:

• .whyline/decisions.md — committed durable decisions, rationale, rejected options, actor, role, task, and affected files;
• .whyline/active-handoff.json — local current task, from/to agent, status, changed files, tests/results, risks/questions, and base/current commit;
• .whyline/ownership.json — local advisory task/file claims;
• .whyline/ledger.jsonl — local mechanical events and prompt text.

Only decisions.md is committed. The other three are checkout-local and gitignored, because stale ownership, a dirty tree, and raw prompts should not travel to another clone. sync combines the active handoff, current Git state, claims, and task/file-relevant decisions into one nonce-fenced packet. Its default budget is about 1,200 tokens; raw prompt text is never included.

What whyline does not do

Worth being explicit, because the name of the category invites the wrong guess.

• It does not orchestrate. It never runs both agents, never runs them in parallel, and never decides which one should act.
• It does not assign roles. It records roles such as implementer or reviewer, but you decide them in the command or prompt.
• It does not supervise. run hands your terminal over and gets out of the way. Nothing is captured, parsed or wrapped, so no vendor changing its output format can break it.
• It does not touch your credentials. Each vendor's own CLI authenticates itself, which is why your existing subscriptions just work.

How it works

Three layers feed one ledger:

1. git resolves a line to a commit via git blame. Works before whyline has recorded anything.
2. A hook silently records sessions, instructions and explicit file edits. It can never fail your session — every path exits 0. Verified against Claude Code only. init also writes a project-local .codex/hooks.json, but no Codex hook event has been observed yet, so treat Codex mechanical capture as untested rather than working: run whyline status, which reports each vendor separately and will say configured but never observed until one arrives.
3. Your agent records the reasoning. whyline init adds an AGENTS.md instruction asking agents to log decisions and rejected alternatives. This is the only layer that captures why.

.whyline/decisions.md is committed and readable with Whyline uninstalled. The operational records and ledger are gitignored.

Does the third layer actually work?

It was the design's one unproven assumption, so it was measured before the features depending on it were built. Over three days across two agents on a real project, 19 decisions were recorded across 14 commits — Claude Code 150% of its non-trivial changes against a 60% gate, Codex 130% against a separate "at least one firing" gate. Every one carried a rationale and a concrete rejected alternative. Codex was never reminded.

That rate holds when an agent works directly, and not when it is dispatched. A later orchestrated task in the same repository recorded nothing at all, because a dispatched agent follows its dispatcher's prompt rather than AGENTS.md. The figure above is therefore a property of direct work, not a general one — which is the reason Pattern 1 above is the recommended shape.

The read side was measured separately, because writing a record nobody consults is worthless. Claude Code ran whyline brief unprompted near the start of 3 of 7 sessions it owned — 43%, below the 50% threshold fixed before collection. The result is therefore unreliable: use run when the handoff must happen, and treat repository-instruction reads as a useful fallback. Codex was observed calling brief nine times, but no reliable Codex session denominator was available, so that count is not presented as a rate. Earlier 67% and 50% figures were superseded as the sample grew.

One gap is known and unfixed: reviewers record less than implementers. Across five tasks on one project, the agent implementing recorded every time, while not one ruling by the agent reviewing reached the record — those went to a tracker that was not committed, and died on clone. 0.1.3 widened the instruction to name reviewers explicitly. Whether that works is not yet measured, and at least two other causes are plausible, including that a reviewer working from a different repository never loads the project's AGENTS.md at all.

Full method and caveats: m0/RESULTS.md.

Honest limitations

• Switching agents is a relay, not a shared conversation. Vendor CLIs are separate processes with separate context windows. sync hands the next agent explicit state; it cannot continue the previous hidden conversation.
• explain reports confidence and will say when it does not know. An empty ledger produces an honest empty answer, not a guess. File-level explain never claims high confidence, because without a line there is no blamed commit.
• run supports Claude Code, Codex, Antigravity (agy), and Grok (grok, "Grok Build") — Gemini CLI itself is dead (its free personal tier was withdrawn); Antigravity is Google's actual working successor and is not the same binary or invocation.
• Typing whyline with no arguments opens the console inside a repository: a full-screen, mouse-enabled app with a Command / Chat / Relay mode switch at the top. Command runs what you type as whyline ... (e.g. status, sync); Chat talks to the active agent (/model claude opus picks one; /model alone lists all four with whether each is installed and logged in, its plan and its model, and how to fix one that isn't available; /model refresh re-checks; /login claude runs that agent's own login and re-checks, so whyline never sees your credentials); Relay drives doctor, status, start and resume. The top row always shows the active agent and model and the repository you're in; /repo ~/other-project switches repositories (after a warning, since it clears the transcript and starts a fresh chat context). The Brainstorm button (or /brainstorm) has several models research a topic independently, review each other's work, and one of them write it up in docs/brainstorm/. While an agent is working, a spinner line says so. Help in the console explains the rest. Outside a repository it asks "Chat or relay?" instead: Chat leads into whyline relay chat, and Relay into whyline relay setup, a guided wizard that assigns implementer/tester/reviewer roles to a plan and runs doctor's own checks before offering to start it.
• Ownership is advisory. Whyline warns about overlapping writes but provides no lock, scheduler, merge engine, or worktree isolation.
• A fresh clone loses operational state, deliberately. It retains committed decisions, so brief, status, and explain still work; those entries carry day precision and therefore never justify high-confidence temporal attribution.
• macOS and Linux are verified; Windows is not. CI passes on ubuntu-latest and macos-latest across Python 3.11 and 3.13. Windows via WSL is untested — a plausible claim, not an observation.

Choosing a model: whyline account and whyline model

whyline account detect checks which plan/tier codex and claude are actually authenticated under (Claude Pro/Max/Team, ChatGPT Plus/Pro/Team) — not just whether they're logged in. It runs once per machine (cached in ~/.whyline/account.json) and asks you to confirm it once per repo (.whyline/account.json, gitignored — this is personal plan/billing info, never committed). whyline account status shows what's on file.

whyline model lets you pick a model per agent for this repo (.whyline/model.json, also gitignored), showing the detected plan as context. whyline model set <agent> <model> sets one non-interactively; whyline model status shows the current choices. whyline run uses whatever's set automatically — no flag needed. Nothing is validated against a list of real model names; a wrong choice just fails at the agent's own invocation time.

Antigravity's model can be set here too, and whyline run antigravity is fully safe with it — that command hands the terminal to a human, so nothing here is affected by Antigravity's own headless-mode limitations. Those limitations do apply if you separately, deliberately configure Antigravity as a whyline-relay role via its documented generic-adapter recipe — whyline model will remind you of this when you set Antigravity's model, but never blocks it, since the interactive whyline run use is completely unaffected.

Credentials

whyline never reads, stores, forwards or proxies a vendor token. run replaces itself with the vendor's own CLI via exec, which does its own authentication. Your subscription works because the official CLI is what talks to the vendor. Permission-bypass flags are never added.

Performance

Measured on an M-series Mac from the 0.2.0 worktree, median of seven runs:

All remain below the 200 ms interactive target. sync asks Git for branch, commit, and changed paths; explain shells out to git blame.

On a 50,000-event, 6.5 MB ledger explain takes ~159 ms against a 1 s target — which is why there is no SQLite index.

Licence

Apache-2.0.