← graph

wg-recover-clean

wg recover: clean batch operation for credit-exhaustion / mass-failure recovery

Metadata

Statusdone
Assignedagent-158
Agent identityf51439356729d112a6c404803d88015d5b44832c6c584c62b96732b63c2b0c7e
Created2026-04-26T22:44:51.965153186+00:00
Started2026-04-26T23:01:44.632631960+00:00
Completed2026-04-26T23:27:00.889482821+00:00
Tagseval-scheduled
Eval score0.82
└ blocking impact0.85
└ completeness0.80
└ coordination overhead0.85
└ correctness0.85
└ downstream usability0.80
└ efficiency0.82
└ intent fidelity0.84
└ style adherence0.78

Description

Description

After a credit exhaustion or other systemic failure, lots of tasks end up Failed. Today's recovery is manual: shell loops over wg list --status failed, wg retry each user task, wg abandon agency followups, optionally edit per-task model first. Works but obscure.

User wants a first-class command for this:

'find all fukt. maybe change the config of them to use a diff endpoint optionally. and then just restart them'

Spec

wg recover [OPTIONS] — surveys failed tasks and resets them in one operation.

FlagBehavior
--dry-run (default if no other flags)Print what would happen: how many failed, how many user vs agency, what would be retried/abandoned, no changes
--yesExecute the plan
--filter <expr>Only act on tasks matching: status=failed (default), or by tag, by id-prefix, by attempt-count, by error-pattern
--set-model <model>Edit failed user-tasks to use this model BEFORE retry (e.g. --set-model 'openrouter:anthropic/claude-sonnet-4-6' to switch to a different provider)
--set-endpoint <url>Same idea for endpoint URL
--keep-agencyDon't abandon .evaluate-* / .flip-* / .assign-* followups (default: abandon them so they regenerate fresh from the parent)
--max-attempts <N>Skip tasks that already have attempt-count >= N (default: 5; protects against infinite retry loops)
--reason <msg>Tag retried tasks with a recovery-reason log entry (for audit trail)

Example flows

Credit exhaustion, switch endpoint, retry everything:

wg recover --filter status=failed --set-model 'openrouter:anthropic/claude-sonnet-4-6' --reason 'credit-exhaustion-2026-04-26' --yes

Just see what's failed:

wg recover --dry-run

Retry only the recent failures (skip exhausted-retry tasks):

wg recover --max-attempts 3 --yes

Output

Always prints structured summary, regardless of dry-run vs execute:

=== wg recover: 33 failed tasks ===
User tasks (will retry): 11
Agency followups (will abandon, auto-recreate): 22
Skipped (max-attempts exceeded): 0

User-task changes:
  endpoint-inheritance-opt    attempt #2  model→ openrouter:...
  tui-log-view                attempt #2  model→ openrouter:...
  ...

Apply with --yes

Out of scope

  • Recovering from non-failure states (Abandoned, etc.) — that's manual triage
  • Cross-repo recovery — single repo at a time
  • Auto-detecting credit-exhaustion patterns and recovering automatically — separate future feature

Validation

  • Failing tests first: test_recover_dry_run_lists_failures, test_recover_yes_resets_user_tasks, test_recover_yes_abandons_agency_followups, test_recover_set_model_edits_before_retry, test_recover_max_attempts_skips_exhausted
  • Implementation makes tests pass
  • cargo build + cargo test pass with no regressions
  • Manual smoke: in a scratch dir with mixed failed tasks, wg recover --dry-run accurately reports; --yes executes; final state has 0 failed user tasks and 0 failed agency followups

Depends on

Required by

Log