Recall

---
name: recall
description: "Reconstruct recent work from scoped conversation history, current repository state, and available shared records."
disable-model-invocation: true
---

# Recall

## OMP model routing

At the start of this workflow, run `../../scripts/model-resolution.mjs`
with `--runtime omp --cwd` set to the current workspace. Read the returned
manifest: nearest project first, then the user's `~/.omp/agent/` default.
For each configured route, select its named agent through OMP's native
task-agent selector and verify that its source matches the chosen scope;
for a canonical role, use the manifest role's `agent` name (user
defaults use namespaced `ohmystack-role-*` agents).
Preserve panel entry order and count. If no mapping is present, retain the
workflow's normal runtime model. Verify resolved worker model and thinking
level from OMP session/job metadata, not from the role file alone.
When `task` returns a background job id, retain it until terminal status.
On OMP hosts exposing `proc://` (observed in 18.3.0), use `read proc://<id>`
for non-consuming status, `wait` to drain, and `write proc://<id>/kill`
to cancel an owned job with the required approval. Confirm cancellation
before replacing a worker and reject results from older generations.
Do not assume the deprecated `hub` tool exists. If safe cancellation
is unavailable, wait or report the unit incomplete; never silently
treat an unconfirmed worker as cancelled.

## Child session handoff

Read [the handoff contract](../poteto-mode/references/subagent-handoff.md).
New tasks, repair rounds, retries, and queue items use fresh child sessions
with the original brief, every later directive, prior findings and responses,
and unresolved objections. Reuse only for required costly live state, and
only when the host allows it. Stop and fence active writers before replacement.
A host-owned orchestrator's model catalog, workspace binding, child tools,
and review-round rules take precedence over the native binding above.
Keep its task handles and attribution receipts. Do not use a backing child
conversation as a new delegated review, or claim native-record verification
for a host-owned child. Report attribution evidence gaps explicitly.

Use for a current-state briefing across recent work. A single saved task to
resume belongs to [session-pickup](../session-pickup/SKILL.md); a personal
working-style Skill belongs to [automate-me](../automate-me/SKILL.md).

Fix topic, workspace, and time window first. Use a named window when supplied;
otherwise state a short recent range. Search only history belonging to that
workspace and the named topic. Prefer a host task-history API when available,
otherwise inspect an explicitly identified workspace transcript directory.
Use modification time rather than opaque IDs for recency. Skip unrelated,
subagent, and test conversations. If the user supplied a complete state capsule,
verify it directly without redundant history mining.

For a named feature or bug, also sweep the shared record through
[why](../why/SKILL.md) with a current-status question: source history, issues,
docs, reports, incidents, and errors actually available in this workspace.
Record inaccessible sources and empty searches. Keep direct customer reports,
team actions, code changes, deployments, and current symptoms distinct.

Verify cited branches, PRs, tickets, and runtime state live when possible.
A conversation that says a fix shipped is not proof it remains deployed.
Inspect full transcript regions when the answer depends on exact actions,
errors, or authorization. Label stale or contradictory records.

Return a brief capsule, one status line per relevant thread, up to five active
problems, and the single most useful next action. Cite task IDs and shared
records. Mark observed, historical, inferred, and unavailable state clearly.
This is a read-only briefing, not implicit authorization to resume work.