Self-Correct Mode —

Every /wbX command that produces a structured output file is dual-mode: pass it a scope folder, you get a fresh report; pass it a previous report from the same command, you get a verify-and-repair pass on that report. This document explains why that exists, when to use it, and the one place where it doesn't apply.

🔨

Execute Task

🛡️

Validate Result

🔍

Diagnose Error

📄 Code

Step 1: Writing the code...

*Self-correct mode: pass a report back to its own command for gap-filling, link repair, and checkbox ticking.*

The two modes, side by side

/wbPlan packages/wb-core # NORMAL — produce a fresh plan_*.md
/wbPlan plan_wb-core_20260509.md # SELF-CORRECT — verify & repair that plan in place

The detection is content-based, not filename-based: the command reads the input file's first H1 and matches it against its own output schema (e.g. # Plan Backlog: <scope> — <date>). If the H1 matches, self-correct fires. If it doesn't, the input is treated as a normal scope path. This means you can rename your plans freely — what triggers self-correct is the file's header, not its name.

Canonical reference: ../../../../packages/wb-flow/templates/commands/_shared/output_conventions §3.

Consolidation: the pass is not only about this file

Since output_conventions v1.12, passing an existing output file with no other flags means more than "tidy it". It means "make this the one file I have to read." Before any repair step runs, self-correct walks the scope's entire reports/ tree for other files of the same category and absorbs every item still open into this one:

/wbPlan  plan_wb-core_20260809.md    # repair + absorb every older open task
/wbAudit audit_wb-core_20260809.md   # repair + absorb every unresolved older finding
/wbIdea  idea_wb-core_20260809.md    # repair + absorb every unpromoted older idea

Two rules keep the merge honest:

• De-duplicate on the item's text, not its ID. IDs are per-file and restart at 1, so de-duplicating by ID silently collapses unrelated items. A task carried forward for four days appears once, bearing its oldest origin link.
• Sources are read, never written. Absorption copies. The old files stay exactly as they were.

⏸️ Deferred and 🚫 Cancelled rows are decided, not open — they stay where they are.

🔴 Self-correct never archives. Repair is safe to run unattended and is expected to run often; moving folders is neither. A sweep happens only under an explicit --archive and is never implied by any other flag. What self-correct does do is notice the mess and offer the command: 📋 Mechanical — 4 superseded plan folders hold no open tasks. → /wbPlan <this file> --archive. Full model: Report Lifecycle.

What it actually fixes

Self-correct is not a re-run. It does not call the model to re-analyze the package. It only walks the existing report and:

1. Fills missing fields the template requires (validator names, scores, model-recs cells).
2. Normalizes link style — plain-text paths become relative markdown links per §1 of output_conventions.
3. Expands bare commands — Origin: /wbAudit becomes Origin: /wbAudit packages/wb-core per §2.
4. Ticks ☐ Done checkboxes for tasks whose tasks_reports/task_<N>_*.md exists on disk.
5. Fills ☐ Valid cells for tasks whose worker report has a validator score appended.
6. Re-emits the 📂 Generated Files footer with the current snapshot of sibling reports.
7. Adds (or updates) a ## 🧭 What's Next? section so the report stays actionable.

What it refuses to do

It will NOT…	Why
Re-run the underlying analysis	That's the job of the normal-mode command. Self-correct is a janitor, not an analyst.
Delete or rewrite content the user (or a previous model) authored	Audit trails are sacred — you can rerun the analysis if you want a different opinion, but you can't silently overwrite an existing one.
Change task IDs, dependencies, or priorities	Those are decisions, not gaps.
Stack multiple Self-Correct Findings sections	Each new run updates the existing one in place — the file grows §N narrative entries, not Findings entries.

Which commands have it

The _shared/output_conventions.md §3 cheatsheet lists the explicit detection markers. As of v1.1:

Command	Detection marker (first H1)
/wbPlan	# Plan Backlog: <scope> — <date>
/wbIdea	# Idea Backlog: <scope> — <date>
/wbAudit	# Audit Report: <scope> — <date>
/wbReview	# Review: <scope> — <date>
/wbTest	# Test Report: <scope> — <date>
/wbContext	# Context: <scope> — <date>
/wbNext	# Next: <scope> — <date>
/wbTrack	# Track: <scope> — <date> (limited mode — see below)

Other structured-output commands (/wbStandup, /wbReview, /wbAudit and friends) define their own marker in their wbX_template.md Detection block. Commands without a structured output (/wbCheck, /wbGit, /wbHelp, /wbBroadcast, /wbMonetize, /wbToWBC, /wbTranslate) do not support self-correct — there's no schema to verify against.

The one limited case: /wbTrack

Track files are append-only narratives with §0 strategic vision + §N contributor entries from multiple models. If self-correct could rewrite the body of a §N section, it would let one model overwrite another model's reasoning — destroying the multi-model audit trail.

So /wbTrack <track_file> runs in limited self-correct:

• ✅ Verifies §N order, model-tag formatting, footer presence, link integrity, status freshness.
• ✅ Auto-repairs bare /wbX in Recommended Next tables and plain-text paths in Files read/created.
• ❌ Does NOT rewrite §N body content, even if a link inside it is broken — broken links are listed in a ## ⚠️ Self-Correct Findings section near the bottom for a human (or the original author) to repair.

The author rule: only the original author of a §N may hand-edit that §N's body to fix mechanical bugs (typos, broken links). A different model wanting to disagree appends a new §N+1 instead.

When to actually use it

Situation	Use self-correct?
You ran /wbWork and want to reflect completed tasks back into the source plan	✅ /wbPlan plan_*.md ticks the ☐ Done boxes
Validator scores landed in tasks_reports/ but the plan still shows blank ☐ Valid	✅ Same — /wbPlan plan_*.md
You hand-edited a plan and added a row without filling the model-recs cell	✅ Same — fills it from model_recommendations.md
You want a fresh analysis because the package changed	❌ Re-run the command in normal mode (/wbX <scope>) — self-correct is not a re-analysis
The file's H1 doesn't match the command's marker	❌ Self-correct won't fire — command runs in normal mode and treats your .md as a scope path (likely produces a bogus output path)

When it's the wrong choice

• You think the plan is fundamentally wrong. Self-correct only fills gaps; it won't restructure the plan. Run /wbPlan <scope> in normal mode for a re-do.
• You want a different model's perspective on the same plan. Self-correct preserves the existing analysis. To get a second opinion, append a §N from a different model (e.g. via /wbReview plan_*.md), don't self-correct.
• The file is from a different command. /wbAudit won't self-correct a plan file — the markers are command-specific.

The honest critique

Self-correct is a mechanical janitor, not a reasoner. If your plan has a wrong dependency edge, self-correct won't catch it — it just makes sure the dependency cell isn't blank. The real value is in the boring cases: tick boxes, fill links, normalize syntax. Don't expect insight; expect tidy.

It also has a hidden trap: the rule only applies to the original output schema. If you renamed Plan Backlog: to Plan v2: in a custom workflow, detection silently fails and your /wbPlan plan_v2_*.md invocation creates a new file in reports/plans/ with the original H1 you never asked for. Stick to the standard headers if you want self-correct to work.

---