Idea roll-up: explicit workflow between agent behaviors (autonomy, pipelines, memory tiers, work folders, remediation)

Aperta
#2,390 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
5/5
Tempo stimato
Più di una settimana
Idoneità per principianti
25/100
Tipo di issue
Funzionalità
Chiarezza
Da chiarire
Stato di attività
Attiva
Stack tecnologico
github, github-actions, typescript

Direzione di ricerca

This issue is a decision record, not committed work. Start by reading the referenced prerequisites #1712 and #1713, then the workflow proposal in #919 and memory-model issues #1243, #1268, #1269, #1270, and #2321. Done would require choosing and implementing a scoped next step; this issue itself provides recommendations and success measures but no implementation task.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

wave:3-later

Why this issue exists

During 2.0 board triage, five older ideas (#799, #919, #600, #707, #1735) kept pointing at the same underlying question: should Squad make workflow between agent behaviors explicit, instead of leaving it implied by prompts, charters and convention? Rather than carry five stale issues, this roll-up evaluates each idea, records an opinion, and closes the originals. Nothing here is committed work; it is a decision record to plan from.

Common thread

Each idea is a different slice of one theme:

Idea Slice of the theme
#799 Tiered autonomy When may an agent proceed without a human gate
#919 Declarative pipelines Which agent/step runs next, enforced by code, not by an LLM's memory
#600 Tiered memory What context a step receives
#707 Task-scoped folders Where a step's outputs land so the next step can find them
#1735 Review remediation A human-triggered edge in the workflow with a separation-of-duties rule

The proposed principle: every handoff has an explicit contract (owner, inputs, outputs, gate, who may clear the gate), and transitions are deterministic code while the work inside a step is LLM. Today this is implied, which is why failures such as #1712 (no verification gate, success = exit code) and #1713 (no per-issue failure tracking) exist.

Prior art (checked)

  • Autonomy tiers are an industry norm. Claude Code ships permission modes (default, acceptEdits, plan, auto) with allow/deny rules and managed policy that overrides repo settings: https://code.claude.com/docs/en/permissions. A cross-vendor comparison (Claude Code, Cursor, Windsurf, Copilot agent mode) notes every vendor says allow-lists are not a security boundary and recommends OS sandboxing and scoped credentials for enforcement, and that repo files configure the agent (so a cloned repo can try to loosen them): https://nhimg.org/ai-ide-and-coding-agent-security-guide
  • Declarative vs imperative workflows have converged on hybrid. A 2026 comparison of LangGraph, CrewAI (YAML/Flows), Google ADK config and others concludes the practical end state is declarative topology with imperative node logic and an escalation path from config to code; the recurring complaint is the lack of a deterministic fan-in barrier and weaker pause/resume for human gates in declarative systems: https://zylos.ai/en/research/2026-07-01-declarative-vs-imperative-agent-workflow-orchestration/. LangGraph provides durable execution and human-in-the-loop interrupts: https://github.com/langchain-ai/langgraph
  • Tiered memory has academic precedent in MemGPT's hot/cold context paging (https://arxiv.org/abs/2310.08560), not re-verified here beyond the abstract.
  • Not verified: work-folder conventions and review-autofix bots in other stacks. I did not find primary sources in this pass, so no claim is made.

Opinions per issue

#799 Tiered autonomy: Adopt the core, change the mechanism (2.0 governance)

Declaring a mode (conservative default, optimistic, explicit) is sound and matches vendor norms. But the proposal implements it as markdown (repo-config.md, a skill) that agents are asked to follow. Prior art says that is advisory, and a repo-committed policy can be loosened by whoever edits the repo. Make the mode enforced by existing gates (review gate #1729/#1734, required checks, CODEOWNERS-style protection of the policy file) and keep conservative as default. Success measure: a test shows an agent attempting a Tier 2 action in conservative mode is blocked by a check, not by a prompt.

#919 Declarative skill pipelines: Adopt as the explicit-workflow substrate, but not as a new runtime (highest alignment with the owner's idea)

The step types (fresh-session copilot step, deterministic code step, context: different for reviewers) are exactly the hybrid shape the prior art converges on. Concerns: a bespoke JSON pipeline duplicates what gh-aw workflows already express, and it is unsafe to build before the watch loop has verification and failure tracking (#1712, #1713). Recommend: first land those two, then spike expressing one pipeline (issue, implement, review, merge-if-approved) in the existing workflow layer. Success measure: the spike pipeline runs end to end with a failing step halting it deterministically and a test proving the reviewer step never receives the implementer's context.

#600 Tiered memory: Fold into the memory decision, cut the wiki layer

The measured problem (history files 34-74KB, 82-96% stale) is real and hot/cold splitting has precedent. But it overlaps the unresolved memory-model cluster (#1243, #1268, #1269, #1270, #2321). Do not build a parallel scheme; make hot/cold a requirement of the single governed-memory decision. The GitHub Wiki layer adds an external write surface with no governance and should be cut. Success measure: token count at spawn before vs after on the same agents (baseline: the figures quoted in #600).

#707 Task-scoped work folders: Do not build as proposed

The problem (stray artifacts) is real, but a top-level .squad/work/ collides with state-backend work (#2229, #2231, #2320, #2378) and the issue itself says these are runtime artifacts, not state. Worktree isolation (#1717) is the better answer to "where do outputs go" and gives per-task isolation for free. Recommend revisiting only as a naming convention inside whatever #1717 produces. Success measure for revisit: a documented case where worktree isolation alone leaves artifacts unfindable.

#1735 Human-initiated review remediation: Keep deferred; this is the model for a human edge

Its own disposition is right: evidence-gated, not part of the review-gate train. It is a good template for the broader principle: explicit human trigger, separate identity from the rejected author, cannot clear its own verdict. Revisit trigger: hosted operational evidence that manual author revision is insufficient (measure: share of REQUEST_CHANGES verdicts needing more than one manual revision round, over a stated window). The original acceptance criteria are preserved in the closed issue.

Proposed next step (not committed)

  1. Land #1712 and #1713 (prerequisites for any pipeline).
  2. Decide the memory model once (absorbing #600's hot/cold requirement).
  3. Spike one explicit pipeline (#919) with the handoff-contract principle above.
  4. Revisit #799 enforcement after the spike shows which gates are real.

Closed in favor of this issue

#799, #919, #600, #707, #1735 (see each issue for its closing note; original text stays readable there).

Lingua principale
TypeScript
Stelle
3.4k
Fork
516
Merge medio
1g 15h
PR unite (30g)
116

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di github/squad

Tutte le issue di github/squad

Issue simili

Altre issue su TypeScript

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.