code-remover

Code remover — evidence-based dead-code PR

Finds code nothing calls, tries hard to prove it IS used, deletes only what survives — one commit each, suite green after every one — and writes a PR a reviewer can check line by line. Two human gates; dry_run stops before any push.


Source	`templates/code-remover/workflow.yaml`
Category	`development`
Trigger	manual (`wfx run`) · schedule `0 6 * * 1`
Steps	13
Human approval before	`remove`, `publish`
Providers named	`claude-cli`, `codex-cli`

Use it

wfx templates code-remover
wfx new code-remover --as my-code-remover
wfx dryrun my-code-remover

What it does

Evidence-based dead-code removal. Machine finders (deadcode, staticcheck U1000, go mod tidy -diff, knip) inventory candidates; machine checks search every tracked file, call shapes, indirect use and history; an agent tries to prove each one is used; a calibrated classifier scores what is left and a decision table decides remove / keep / bug / needs-a-human. After a person approves the plan, each removal is its own commit with the suite green after it, re-proven independently, reviewed by a different model, and described in a PR whose every claim is checked against the evidence. Nothing is pushed until a person approves.

What you fill in

scope: point it at one module or directory — cleanup agents do best on a portion of the repo
the publish step’s base branch, if your default branch is not what gh reports
TYPESAFE_API_KEY in the env store (wfx env set TYPESAFE_API_KEY) for the calibrated verdicts

Inputs

Input	Type	Default
`repo_path`	string
`repo_url`	string
`base_branch`	string
`scope`	string	`"."`
`max_removals`	integer	`10`
`pr_prefix`	string	`"[dead-code]"`
`dry_run`	boolean	`true`
`isolate`	boolean	`true`

Steps

#	Step	Kind	Skills	Gates
1	`baseline` Baseline build and test counts	shell (`run:`)		`exitCode` = `70` → fail `exitCode` = `78` → skip to `report`
2	`inventory` Inventory candidates (dead-code-finder)	shell (`run:`)		`exitCode` = `70` → fail `exitCode` = `78` → skip to `report`
3	`checks` Machine usage checks	shell (`run:`)		`exitCode` = `70` → fail
4	`prove` Prove each candidate IS used (usage-prover)	agent on `claude-cli`	`usage-prover` `repo-navigator`
5	`plan` Judge and plan (JEV + decision table)	shell (`run:`)		`exitCode` = `70` → fail `exitCode` = `78` → skip to `report`
6	`remove` Remove the approved candidates (safe-remover)	agent on `claude-cli`	`safe-remover`	human approval first
7	`verify` Prove the branch (before vs after)	shell (`run:`)		`exitCode` = `70` → fail
8	`review` Adversarial review on a different model (change-reviewer)	agent on `codex-cli`	`change-reviewer`	`approved` = `false` → needs_input
9	`write_pr` Write the cleanup PR (cleanup-pr-writer)	agent on `claude-cli`	`cleanup-pr-writer`
10	`claims` Re-check every claim (JEV) and the numbers	shell (`run:`)		`exitCode` = `20` → needs_input `exitCode` = `70` → fail
11	`dry_run_stop` Stop here on a dry run	shell (`run:`)		`exitCode` = `78` → skip to `report`
12	`publish` Push and open the PR (pr-publisher)	agent on `claude-cli`	`pr-publisher`	human approval first
13	`report` Report	shell (`run:`)

Human gates

remove waits for wfx approve <run-id> (or wfx reject <run-id> -m "why") before it runs.
publish waits for wfx approve <run-id> (or wfx reject <run-id> -m "why") before it runs.
review pauses for input when approved is false: “The codex reviewer blocked the branch: {{ .Output.verdict }}”. Resume with wfx answer <run-id> -m "...".
claims pauses for input when exitCode is 20: “PR claims not supported by the evidence — {{ .Output.stdout }}”. Resume with wfx answer <run-id> -m "...".