Workflows — Disinto

← disinto.ai

Issue Underspecified

Trigger: dev-agent can't implement the issue

dev

Writes refusal JSON to IMPL_SUMMARY_FILE

{"status":"too_large","reason":"..."}

dev

Posts refusal comment, labels "underspecified"

Removes backlog + in-progress labels

dev

Releases claim, exits

No PR created — issue returns to pool

human

Splits issue or adds detail, re-labels backlog

CI Breaks

Trigger: pr_poll_ci() returns failure

bridge

Classify failure: infra or code

classify_pipeline_failure() checks heuristics

bridge

Infra? Retrigger once via empty commit

ci_retry_count < 1

bridge

Code? Invoke dev-agent with CI logs

agent_run --resume with error log

dev

Fix, test, rebase, push

ci_fix_count++ (max 3)

bridge

Budget exhausted? Block issue

_PR_WALK_EXIT_REASON="ci_exhausted"

Review: REQUEST_CHANGES

Trigger: review-agent posts REQUEST_CHANGES verdict

review

Posts review comment + formal review

marker

review

Keeps session & worktree for re-review

SID_FILE preserved for continuity

bridge

pr_poll_review() detects verdict

Invokes dev-agent --resume with feedback text

dev

Addresses feedback, pushes

review_round++ (max 5 rounds)

review

Re-review at new SHA

Incremental diff (prev_sha..new_sha)

dev-agent

review-agent

pr-lifecycle (bridge)

failure / feedback loop

The Contract: Label Taxonomy as Protocol

predictor files  →  prediction/unreviewed    (open issue)
planner actions  →  prediction/actioned      (closed, real issue created)
planner rejects  →  prediction/dismissed     (closed, reasoning in comment)

Memory: the forge issue tracker IS the predictor's memory.
        No separate memory file — history of actioned/dismissed
        informs focus next run.

Budget: predictor  = 5 actions/run (exploit counts as 2)
        planner    = (predictions_triaged + 1) new issues

Models: predictor  = sonnet  (fast, adversarial scanning)
        planner    = opus    (deep reasoning, strategic sequencing)

predictor

planner

actioned → dev pipeline

The Template-or-Vision Gate

Planner reads .codeberg/ISSUE_TEMPLATE/*.yaml and attempts to fill:

  IF  affected_files ≤ 3
  AND acceptance_criteria ≤ 5
  AND single clear approach (no design forks)
  THEN → file as "backlog" → gardener validates → dev picks up

  ELSE → label "vision" with problem statement
       → architect pitches sprint PR on ops repo
       → human ACCEPT/REJECT via PR comments
       → architect files sub-issues as "backlog"

The planner NEVER files issues past the bottleneck.
Only the top 5 constraints get filed.
Issues blocked on vault items are NOT filed.

planner (opus, 12h)

architect (6h)

gardener (sonnet, 6h)

dev pipeline

Alert Priority Levels

Memory Crisis

RAM < 500MB or swap > 3GB

kill sessionsprune images

Disk Pressure

Disk > 80%

rotate logsprune worktrees

Factory Stopped

CI stuck >20min, dead locks, broken rebase

rm locksabort rebasecancel CI

Factory Degraded

PRs stale >20min, circular deps

comment PRsalert planner

Housekeeping

Old worktrees, expired sessions

clean worktreesprune branches

Full Vault Approval Flow

any agent

Detects human-blocked work

Needs external resource, decision, or secret

any agent

Files vault item on ops repo

$OPS_REPO_ROOT/vault/pending/<project>-<slug>.md

vault

Creates PR via vault_request()

Branch vault/<action-id>, file vault/actions/<id>.toml

review

Reviews vault PR

Checks: required sections, specific human action, dedup

human

Approves vault PR

Merges PR on ops repo

runner

Executes with injected secrets

docker/runner container: mounts secrets, runs formula, time-boxed

Who Files Vault Items?

planner

Human-blocked prerequisites

Needs external resource or decision

supervisor

Blocked issues it can't auto-fix

P2/P3 requiring human intervention

predictor

Exploit dispatches

Formula execution requiring secrets

gardener

Rarely — human grooming decisions

supervisor (sonnet, 20min)

vault / runner

review-agent

filing agents