Deferred — Delete the lies. Then /bad.

The user invoked /deferred because YOU shipped lying code: a stub pretending to be an implementation, a function body that returns a hardcoded value with // TODO next to it, a half-completed sweep marked "next iteration handles the rest", a peripheral handler that logs args instead of emulating the register, a comment that frames partial work as a deliberate scope decision ("out of scope for this task", "follow-up will integrate", "deferred to next agent"). The user did not pay for any of that. The user paid for the implementation. The half-implementation is a misrepresentation of what was delivered.

Per agent_docs/rules.md, every shape of this is a rule violation, not a degree of completion:

• § "An instruction is ALWAYS implemented fully" — "Unfinished new implementation are reverted and treated as rules violation"
• § "Code comments framing partial work as deferred are bailouts" — "Comments that present an incomplete port as an intentional choice are bombs, documented, not honest scoping"
• § "Documented-uncertainty sections are unfinished drafts"
• § "Stubs must fail honestly" — "return failure, not fake success"
• § "Never add transitional hacks" — "no 'keep it working between phases', no intermediate shims"
• § "Checklist defines task scope; the implementing agent does not"
• § "Do not delegate implementable work to 'next agent' or 'next session' when you have tool budget"

These exist because deferral-as-progress is the highest-value lie an agent can ship on this codebase: compile is green, function exists with the right name, no error surface — looks done. Weeks later something tries to call it, gets garbage or a fake-success, and the user pays again to discover the lie and pays AGAIN to write the implementation. The skill cuts that cycle at the source.

Step 1 — Enumerate every deferred / partial / stub / TODO / out-of-scope item

Honestly, exhaustively, no aggregation, no softening, no "minor cleanup":

"Deferral disclosure under /deferred:

1. <file:line> — <symbol>: <the specific lie> (e.g. "returns hardcoded 0; comment says // TODO: implement properly"). Violates agent_docs/rules.md § <rule>.
2. <file:line> — <symbol>: <the specific lie>. Violates § <rule>.
3. … (continue until every TODO / stub / fake-body / // out of scope / // next iteration / // follow-up / // will be revisited / partial-sweep / documented-uncertainty item in your recent stretch is named) Total: N items of deferred-as-progress."

If you cannot produce specific items, you have not honestly inventoried your own work. Re-read your recent diffs and find them. The bar is items, not categories — "some things are deferred" is not a valid Step 1 output.

Step 2 — DELETE the lying code

If the implementation is wrong, the code IS wrong. The user did not pay for stubs that lie about being implementations; the user paid for the implementation. Delete each item from Step 1's list via Edit / Write tool calls in this same reply.

Deletion here means FULL REVERSION to the state before the lying code existed. Not "replace with a less-broken variant". Not "downgrade to a CerfFatalExit-style honest-failure stub". Not "keep the salvageable 75% and remove just the broken 25%". A function that does 75% but is missing the critical 25% is NOT the function the user agreed to — it gets reverted, entirely. The function declaration, its body, and any callsites added during the same bad stretch to reference it all disappear; the file returns to the state it had before the lie was added. Build success on top of a lying implementation has zero meaning, so producing a "still-compiles but smaller lie" variant is not progress — it's the deferral pattern re-firing under fresh paint.

The ONLY next step (after Step 3's /bad) is the FULL implementation, written from scratch under the rules. Not a partial variant. Not a "phase 1 ships, phase 2 in next iteration" plan. The full implementation, in one body of work.

Specifically:

• // TODO: implement properly next to a fake-success return → REVERT the function entirely. The function and its declaration disappear; callsites added in the same bad stretch to reference it disappear too. The file returns to its pre-lie state.
• Stub method returning hardcoded value → delete the stub entirely; the call site breaks loudly so the missing work is visible.
• Half-completed sweep (renamed N of M call sites, ported N of M handlers) → delete the partial work; the original state stands until the full sweep can be done in one body of work.
• // out of scope for this task / // next iteration / // follow-up / // next agent will handle this comments excusing partial work → delete the comments AND the partial work they were excusing.
• Documented-uncertainty section at the end of a deliverable ("§N — known gaps", "§N — things I could not verify", "§N — load-bearing assumptions") → delete the section AND the deliverable it purports to certify.

This is NOT the destruction pattern /bailout and /bad guard against. Those skills protect WORKING code from being destroyed under cleanup cover. /deferred removes code that never worked correctly — the lie was shipped as the implementation, the lie has to go before the implementation can exist. The distinction:

• /bailout / /bad destruction case: agent deleted an actual implementation of IMGFS → restore it, the destruction was the violation.
• /deferred deletion case: agent shipped a stub // TODO: real IMGFS implementation returning fake-success → delete it, the shipping was the violation.

If a Step 1 item is ambiguous (the code partially works for some inputs and lies for others), it still goes. The contract was the full implementation; partial isn't a state the user purchased.

Step 3 — Execute /bad on the deferred-bullshit violation

After the deletion, invoke /bad's protocol on yourself for the underlying rule violations Step 1 enumerated. That means: re-anchor to the original task (which was the FULL implementation, not the partial one — the original task did not change while you drifted into deferral), articulate the violation specifically with rule citations (Step 1's enumeration IS the articulation), propose a path forward that delivers the full implementation properly OR explicitly negotiates a smaller scope with the user (never a hidden scope cut disguised as // out of scope), and resume work.

/bad's Mass-destruction section does NOT fire on the deletion you just did in Step 2, because Step 2 was removal of lying code, not destruction of working code. The two skills compose: /deferred Step 2 is sanctioned removal; /bad Step 3 is the violation-acknowledgment + course-correction on the deferral pattern that produced the lying code in the first place.

What the user gets back

The reply contains, in this exact order:

1. Step 1 enumeration — exhaustive numbered list with file+line+rule citation per item.
2. Step 2 deletion tool calls — Edit / Write removing each lying piece from Step 1.
3. Step 3 /bad protocol — re-anchor to original task, articulate violation, name the path forward, resume work in the same reply.

No asking. No "should I delete this one?" / "which deferral should I keep?". All Step 1 items go. The user did not pay for choosing which lies survive.

Anti-patterns (forbidden in the /deferred reply)

• Step 1 with vague items. "Some things are deferred" / "a few TODOs" / "minor partial work" is not the enumeration this skill requires. The bar is file + line + specific lie + rule citation, per item.
• Salvaging "partial progress" from Step 1's items. "But function X is 80% done, let me keep that part" — the 80% is itself a lie if the contract said 100%. Delete it. The next attempt does it correctly.
• Asking which items to delete vs keep. No asking. All go. The user invoked /deferred after seeing the deferred-as-progress pattern; the invocation IS the authorization to remove every item.
• Skipping Step 2 and going straight to /bad. That turns /deferred into a softer /bad and preserves the lies. The deletion is the point — without it, the next agent re-discovers the lies under fresh framing.
• Confusing /deferred deletion with /bailout destruction. They are distinct cases (named explicitly in Step 2 above). /bailout guards working code; /deferred mandates removing lying code. If you find yourself reasoning "but /bailout says never delete" — re-read Step 2's distinction. /bailout's rule is about working code under cleanup cover; /deferred is about lying code being removed so the honest implementation can exist.
• Apology theater instead of deletion. "I see what I did, I'll be more careful next time" without the deletion is the deferral pattern continuing under cover of contrition. The deletion is required.
• Treating the deferred code as recoverable later ("we can resurrect this when we have time"). It cannot. The lie is gone; the next attempt writes the honest implementation from scratch. If parts of the deleted code were genuinely useful, the next implementation re-derives them — but it does so as part of a complete implementation, not as a salvaged-stub-promoted-to-real-code.

Why this skill exists

Deferral-as-progress is the highest-value lie an agent ships on this codebase, because it's the lie that looks indistinguishable from real progress at the moment of delivery. Every other rule violation has a tell — destruction leaves visible holes, hacks leave warning-shaped surface, guessed constants compile but break under known test inputs. Deferral leaves a green build with the right symbols and no obvious smell, and the cost only surfaces weeks later when something tries to use the code and gets garbage.

/deferred exists to remove the lie at the moment the user catches it. The deletion is not destruction in the /bailout / /bad sense — it is the precondition for the implementation to exist at all. The user did not pay for choosing which lies remain; the user paid for the implementation. Lies go; implementation gets written honestly on the next attempt.

If /deferred fires once in a session, the rule violations enumerated in Step 1 also fire /bad's second-strike clause for the next stretch — meaning if the same pattern (any new deferral-as-progress) appears in the work that follows, treat the next /bad or /deferred with maximum literal protocol adherence, not as a soft re-correction.