sweep

Non-interactive end-to-end research pipeline: auto-plan → judge gate → run. Single command from goal to result. Accepts a goal string and passes through all run/colab/team flags.

NOT for: interactive planning (use /research:plan); methodology review only (use /research:judge); running an already-approved plan (use /research:run).

Steps S1–S5

Triggered by sweep "goal" [--flags]. Non-interactive end-to-end pipeline: auto-plan → judge gate → run.

Task tracking: create tasks for S1–S5 at start.

Step S1: Parse arguments

Extract <goal> — the first positional argument (quoted or unquoted string describing the optimization target).

Extract flags and their values:

• --colab[=HW] — passed through to plan (Config.compute) and run; if =HW present, extract colab_hw
• --compute=local|colab|docker — passed through
• --team — passed through to run
• --codex — passed through to run
• --researcher — passed through to run; combine with --architect for dual-agent SOTA + architectural hypothesis pipeline (--journal and --hypothesis are not available in sweep mode)
• --architect — passed through to run; enables architectural hypothesis pass via solution-architect
• --skip-validation — passed to judge step (S3) to skip local metric/guard validation
• --out <path> — optional: write program.md to this path instead of project root

If <goal> is missing or empty, stop:

⚠ sweep requires a goal prompt.
Usage: /research:sweep "goal description" [--flags]

Step S2: Non-interactive plan

Run plan mode steps P-P2 and P-P3 from plugins/research/skills/plan/SKILL.md (P-P0 profiling flow skipped — <goal> is always a text string; P-P1 scope guard skipped — goal was provided explicitly) with these behavioral overrides:

• P-P2 (config presentation): Accept all auto-detected defaults without prompting the user. Print the proposed config as an informational block prefixed sweep: auto-config → but do NOT wait for user confirmation.
• If --colab[=HW] or --compute=colab was passed, write compute: colab (and colab_hw: <HW> if provided) into the Config block.
• P-P3 (write program.md): Write to <--out path> if provided; otherwise to program.md at project root.
  • If the output path already exists: rename existing to <path>.bak (overwrite any existing .bak), then proceed — no interactive confirmation in sweep mode.

Print on completion:

sweep: plan → <output path> ✓

Step S3: Judge + refinement loop

Initialize REFINE_ITER = 0, MAX_REFINE = 3.

Repeat up to MAX_REFINE times:

1. Increment REFINE_ITER. Run judge mode (J1–J6 from plugins/research/skills/judge/SKILL.md) against the program file.

   • Pass --skip-validation if the user provided it; otherwise include validation (J4).
   • Capture the J6 verdict and the judge report path (JUDGE_REPORT).

2. Print: sweep: judge iteration `REFINE_ITER`/`MAX_REFINE` → `VERDICT`

3. If APPROVED — exit loop with outcome approved.

4. If BLOCKED — exit loop with outcome blocked. Do not attempt to fix — BLOCKED means a fundamental design flaw that requires human redesign.

5. If NEEDS-REVISION:

   • If REFINE_ITER < MAX_REFINE:
     • Read JUDGE_REPORT. Extract the ### Required Changes section.
     • Apply each listed fix to the program file using the Edit tool. Count applied edits as N_FIXES.
     • Print: sweep: applied N_FIXES fix(es) to <program path> — re-judging
     • Continue to next iteration.
   • If REFINE_ITER == MAX_REFINE — exit loop with outcome unresolved.

Safety net: the .bak file created in S2 is the undo path — edits applied during the loop modify program.md in place.

Step S4: Gate on loop outcome

Outcome	Action
approved	Print sweep: plan approved (REFINE_ITER/MAX_REFINE iteration(s)) ✓ → proceed to S5
blocked	Print sweep: judge → BLOCKED ✗; show all critical findings from the report; print follow-up hint; stop
unresolved	Print sweep: judge unresolved after MAX_REFINE iterations ✗; show remaining Required Changes from the last report; ask user with options: (a) proceed to run anyway (b) fix manually then re-run (c) abort — if a, proceed to S5; if b or c, print follow-up hint and stop

Follow-up hint (blocked or unresolved):

Fix the issues above in <program path>, then:
  /research:judge <program path>          ← re-validate
  /research:run <program path>            ← run when approved
  /research:sweep "revised goal" [flags]  ← re-sweep from scratch

Step S5: Run

Run Default Mode (R1–R7 from plugins/research/skills/run/SKILL.md) against the program file from S2, passing through all flags:

• --colab[=HW] / --compute
• --team
• --codex
• --researcher / --architect (combine both for dual-agent pipeline)

Note: --journal and --hypothesis are not available in sweep mode (see S1).

--team and interactivity: when --team is passed, sweep becomes semi-interactive — run mode's Phase B presents a user confirmation gate before Phase C begins. The gate cannot be bypassed from sweep context; sweep will pause and wait for user input at that point. This is expected behavior.

On completion, the standard R6 terminal summary is printed. Additionally, prepend:

sweep: complete — plan → judge → run pipeline finished