retrospective

Retrospective — what broke, why, and what to change

Daily and on demand: reads runs, failed steps, retries, approval waits and cost from the engine’s API, cites the run and step for every finding, proposes a concrete fix, and tracks whether earlier fixes held.


Source	`templates/retrospective/workflow.yaml`
Category	`debugging-ops`
Trigger	manual (`wfx run`) · schedule `30 6 * * *`
Steps	6
Human approval before	`apply`
Providers named	default (`WFX_DEFAULT_PROVIDER`)

Use it

wfx templates retrospective
wfx new retrospective --as my-retrospective
wfx dryrun my-retrospective

What it does

Reviews the engine’s runs: failures, retries, approval waits and token cost. Every finding cites a run and step and proposes a concrete change. Remembers earlier findings and reports whether their fixes held. Typed proposals (add/edit step, new workflow, memory fix, retire) each wait for a person’s approval; approved ones are applied, rejected ones are never proposed again.

What you fill in

env SINCE_HOURS: how far back a scheduled run looks (default 24)
env WFX_TOKEN: a read-only token, when the API needs one

Inputs

Input	Type	Default
`workflow`	string	`""`
`runId`	string	`""`
`question`	string	`""`
`since_hours`	number	`24`
`mode`	string	`"review"`
`proposal`	object	`{}`
`fingerprint`	string	`""`

Steps

#	Step	Kind	Gates
1	`collect` Read runs, failures, retries, waits and cost	shell (`run:`)
2	`analyze` Say what broke, why, and what to change	agent
3	`check` Keep only findings backed by evidence	shell (`run:`)
4	`propose` One approval per proposal	shell (`run:`)
5	`show` The proposal to approve	shell (`run:`)
6	`apply` Apply the approved proposal	shell (`run:`)	human approval first

Human gates

apply waits for wfx approve <run-id> (or wfx reject <run-id> -m "why") before it runs.