continuation.md

Continuation Mode (y-axis carry-over across tasks)

What it is

Continuation mode lets a task reuse the prior task's deliverable as starting material, combined with user feedback as the learning signal. It is the runtime counterpart of the experiment-task hierarchy (docs/superpowers/specs/2026-04-20-experiment-task-hierarchy-design.md) for the user-task axis — distinct from refinement (docs/refinement.md, y-optimisation within a single run) and outer-loop optimisation (docs/outer-loop.md, θ-optimisation over prompt artifacts).

Mental model

Task 001 (seed)  →  Task 001/BEST/paper.md
                         │
                         ▼
Task 002 (continuation, inputs=[{from_task:001,role:primary,bundle:BEST/}])
     user_feedback = "deepen section 2, add benchmarks"
     │
     ▼
  manager_prompt_prefix = render_continuation_prefix(bundle, feedback)
     │
     ▼
  AgentWorkflow(manager_prompt_prefix=…)
     │
     ▼ Task 002 Manager sees prior draft + feedback at iteration 1
  Task 002 produces evolved deliverable, NOT re-derived from scratch.

Mechanism

1. task.json.inputs is non-empty (R37). Each entry has from_task, role ∈ {primary, reference}, and exactly one of bundle: "BEST/" or paths: [...].
2. load_continuation_bundle(task_dir, experiment_dir) resolves each entry against <experiment>/tasks/<from_task>/BEST/. Primary entries inline the file contents; reference entries carry only path + size + first 200 chars.
3. render_continuation_prefix(bundle, max_chars=480_000) builds a deterministic prefix:
   • ## Continuation Context
   • ### Prior deliverable (primary) — each primary inlined
   • ### Reference material (available via fs.read) — list, optionally dropped
   • ## User Feedback — verbatim
   • ## Your Task — build-on-prior boilerplate
4. Budget fallback ladder. Full → reference-stubs → no-reference → ContinuationBudgetError (primary alone over budget → split the task).
5. CLI dispatch. awp run --target <exp>:<cont-task> detects mode=="continuation", calls the loader, renders the prefix, and passes it as AgentWorkflow(manager_prompt_prefix=…). The DelegationLoopRunner (already plumbed for refinement) prepends the prefix to the Manager's iteration-1 user message.

Relationship to refinement

Refinement (docs/refinement.md) also injects a prefix, but:

• Scope = one run, multiple iterations (y-axis within a run).
• Gradient = auto-extracted from critique + gate rejections + eval deltas.
• Prefix is applied to iteration 1 of the refinement session's run.

Continuation:

• Scope = task → next task (y-axis across tasks).
• Gradient = user-written feedback (free text).
• Prefix is applied to iteration 1 of the next task's seed run.

Both mechanisms can compose: a continuation task's seed run can also be refined, and the refinement engine's prefix will overlay on top of whatever state the continuation prefix produced in iteration 1.

R37

Continuation tasks with empty inputs are rejected at task-create time and at runtime (bundle loader refuses without BEST). See spec/versions/1.0/validation-rules.md.

Files

• packages/awp-runtime/src/awp/continuation/bundle_loader.py — load_continuation_bundle
• packages/awp-runtime/src/awp/continuation/prompt_injection.py — render_continuation_prefix
• packages/awp-core/src/awp/experiment/cli_handlers.py — CLI dispatch (run_task_aware)

See also

• docs/refinement.md — y-axis optimisation within a single run (auto-extracted gradient). Continuation is the cross-task analogue.
• docs/outer-loop.md — θ-axis optimisation over prompt artifacts. Runs per-experiment after Plan 4.
• docs/superpowers/specs/2026-04-20-experiment-task-hierarchy-design.md — the design this mechanism implements, §7.1–§7.4.
• spec/versions/1.0/validation-rules.md — R37 (continuation input non-emptiness).