Skip to content

add-orchestrator: bundle ships stale copies of sync-fleet-to-head (v1.0<1.4) and profile-fleet (v1.4<1.6) — back-port + close the field-hardening back-flow gap #5

Description

@vybe

Summary

The add-orchestrator bundle ships strictly older copies of two runtime skills than a production orchestrator (corbin) has run on live crons for weeks. Reported by corbin during an upgrade of its installed copies (evidence: Abilityai/corbin-internal@56d4fad).

skill bundle (add-orchestrator v1.20) corbin (production)
sync-fleet-to-head v1.0 v1.4
profile-fleet v1.4 v1.6

Verified: clean fast-forward, same lineage (no divergence)

I fetched corbin's deployed SKILL.md for both and diffed against the bundle templates:

  • Changelog is a strict superset — the bundle's entries appear verbatim as the common prefix; corbin only appends (sync 1.1→1.4, profile 1.5→1.6). No same-version-different-text conflict anywhere, which is the signature of shared lineage rather than a fork.
  • Body diff is purely additive (profile-fleet: bundle 185 → corbin 250 lines; every hunk is an insertion).

Corbin's fast-forward claim holds. These two skills were originally "universalized from a production orchestrator" in add-orchestrator v1.3, then the bundle froze while the production copy kept hardening — this is a drift-since-back-port, not a merge.

What today's installers are missing

sync-fleet-to-head v1.1–1.4:

  • v1.1: get_git_sync_state cache-lag note (confirm real HEAD via get_git_status post-pull)
  • v1.2: two 409 subtypes — unstaged changes (dirty → stash_reapply) vs unmerged files (pre-existing conflict; stash_reapply will also fail → Step 6)
  • v1.3: Error Recovery row for the 400 submodule-fetch failure
  • v1.4: --autonomous run mode (see companion issue)

profile-fleet v1.5–1.6:

  • v1.5: autonomy-toggle cross-check — an agent with enabled schedules but autonomy_enabled: false never fires them (next_run_at silently advances). Root-caused finance-agent 2026-07-28; retro-explained the gtm R1/R2 mystery.
  • v1.6: --autonomous run mode (see companion issue)

Proposed work

Back-port the generalizable deltas into the bundle templates. This is not a verbatim file swap: corbin's deployed copies carry field-specific bits that must be universalized first, e.g. a Step 0 calling scripts/scheduled-tasks/refresh_workspace.sh and corrections queued to fleet-gap-analysis/status.yaml (corbin conventions). Port the error-handling/cross-check content; universalize or drop the deployment-specific scaffolding — the same discipline the original v1.3 back-port used.

Mechanics: bump each skill's metadata.version, prepend changelog, refresh the "bundled X vN" references in add-orchestrator/SKILL.md, bump the bundle version + marketplace.json, pass /audit-wizards gates. Push is operator-gated (sacred repo).

The recurring root cause (process gap)

Field-hardening happens in the deployed agent and there is no path back to the marketplace. This will recur for every skill the bundle seeds into a live orchestrator. Worth a standing back-flow check (e.g. /sync-skill-library-style drift detection between bundle templates and known production installs, or the install-path doctor in the companion issue surfacing installed > bundled as a back-port candidate).

Filed by trinity-pm on behalf of corbin's report. Verdict: adopt. Mechanism (verbatim vs universalized cherry-pick) and OSS surface are a maintainer call.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions