PPA Recover
The boundary this skill lives inside
This skill produces a structured recovery proposal. It never produces a gate
verdict and it never performs the recovery. The controller that actuates is
_ppa/closure.py per docs/PPA_INTERFACES.md §4; this document is what it reads
before deciding whether to act.
The distinction matters most exactly here, because a stalled loop is the moment
at which the temptation to "just fix it" is strongest and the evidence is
weakest. A recovery that is proposed, bounded and reversible can be refused. A
recovery that is performed and then described cannot.
When to use
Trigger when the user:
- Has a closure loop that stopped improving, or started getting worse
- Has a run that crashed part-way and needs to know what state it is in
- Has exhausted an iteration budget without meeting the target
- Suspects the loop is oscillating between two configurations
Inputs to gather
- The loop history: per round, the actuator moved and the metric that resulted
- The run manifest for each round, so a rollback point can be named by hash
- The failure mode as observed, not as interpreted — stall, oscillation,
monotonic regression, crash, or budget exhaustion
- What has already been rolled back, if anything
- The remaining budget, and whether the caller is willing to spend it
Workflow
- Classify the failure by its signature, not by its cause. A stall (no
change across N rounds), an oscillation (alternating between two states), a
regression (monotonic worsening) and a crash need different recoveries, and
the signature is readable from the loop history without a theory.
- Name the rollback point by hash. "Before the last change" is not a state;
a run-manifest sha256 is. If no manifest exists for the state you want to
return to, that state is not reachable and the proposal says so instead of
implying otherwise.
- State the blast radius. Exactly what the recovery touches, and what it
leaves alone. A recovery that quietly reaches into the RTL when the caller
believed it was PnR-only is worse than no recovery.
- Give the loop a stop condition. Not a retry count alone — a condition that
is checkable from the metric history, so the loop can end without a human
noticing it is still going. A recovery with no stop condition is a second
stalled loop wearing the first one's clothes.
- Prefer returning to a known state over reaching a new one. The purpose of
recovery is to make the design reachable again, not to close it. Closure is
ppa-optimize, afterwards, from solid ground.
- Preserve the failed evidence. The artefacts from the failing rounds are
what the next diagnosis reads. A recovery that overwrites them has destroyed
the reason the loop stalled.
Do not
- Do not hand-edit a layout, delete violating geometry, move pins by hand or
relax a rule deck to get out of a stall. A state reached that way is worth less
than the failure it replaced, and it is unreachable by any later re-run.
- Do not propose a recovery whose rollback point does not exist. Say the state is
unreachable; that is a real finding about the run.
- Do not delete or overwrite the artefacts of the failing rounds.
- Do not describe the recovery as already applied, and do not report a metric for
a state that has not been remeasured.
- Do not raise a budget as the recovery. A loop that stalled at 6 iterations
usually stalls at 12; the stall is a finding, not a resourcing problem.
- Do not treat a crash as a stall. A crash means the state is unknown, and an
unknown state is recovered by returning to a known one, not by continuing.
Output format
The deliverable is one markdown proposal. The template below is the whole shape.
# PPA Recovery Proposal — <design>, round <n>
Verdict authority: _ppa/closure.py - this proposal states no pass/fail of its own.
## Summary
<what the loop was doing, when it stopped being useful, and what state it is in now>
## Failure signature
| round | actuator moved | timing.setup.wns_ns | area.core_um2 | manifest sha256 |
|---|---|---|---|---|
| 3 | density 0.62 -> 0.60 | -0.124 | 118420 | sha256:77c1de40 |
| 4 | density 0.60 -> 0.58 | -0.121 | 119880 | sha256:0a5b93f6 |
| 5 | density 0.58 -> 0.60 | -0.124 | 118420 | sha256:e61d2c88 |
Classification: oscillation between rounds 3 and 5, period 2, no net movement
## Recovery proposal
Rollback point: run manifest sha256:3f9a1c7d - the end of round 3, the last
state whose artefacts are complete
Blast radius: PnR configuration only; RTL, SDC and PDK untouched
Stop condition: 2 consecutive rounds with no improvement beyond 0.005 ns, or
the remaining 3-iteration budget is spent, whichever comes first
Preserves: rounds 3-5 artefacts are retained under runs/round<N>/ and are not
overwritten by the recovery
## What this does not fix
<the underlying cause is not addressed here; name what still has to be diagnosed>
## Evidence
| artefact | sha256 | tool |
|---|---|---|
| runs/round5/manifest.json | sha256:b204e8a1 | closure controller |
Next: run /ppa-diagnose
1---2name: ppa-recover3description: Propose a bounded way out when a PPA closure loop has stalled, oscillated, regressed or crashed — naming the rollback point, the blast radius, and the condition under which recovery stops rather than looping. Use when the user says "the loop is stuck", "PnR keeps getting worse", "we regressed", "roll this back", "recover the run", or when a closure controller has exhausted its budget without converging.4metadata:5 source: https://skillmd.com/skills/vibeic/ppa-recover6---78# PPA Recover910## The boundary this skill lives inside1112This skill produces a **structured recovery proposal**. It never produces a gate13verdict and it never performs the recovery. The controller that actuates is14`_ppa/closure.py` per `docs/PPA_INTERFACES.md` §4; this document is what it reads15before deciding whether to act.1617The distinction matters most exactly here, because a stalled loop is the moment18at which the temptation to "just fix it" is strongest and the evidence is19weakest. A recovery that is proposed, bounded and reversible can be refused. A20recovery that is performed and then described cannot.2122## When to use2324Trigger when the user:25- Has a closure loop that stopped improving, or started getting worse26- Has a run that crashed part-way and needs to know what state it is in27- Has exhausted an iteration budget without meeting the target28- Suspects the loop is oscillating between two configurations2930## Inputs to gather31321. The loop history: per round, the actuator moved and the metric that resulted332. The run manifest for each round, so a rollback point can be named by hash343. The failure mode as observed, not as interpreted — stall, oscillation,35 monotonic regression, crash, or budget exhaustion364. What has already been rolled back, if anything375. The remaining budget, and whether the caller is willing to spend it3839## Workflow40411. **Classify the failure by its signature, not by its cause.** A stall (no42 change across N rounds), an oscillation (alternating between two states), a43 regression (monotonic worsening) and a crash need different recoveries, and44 the signature is readable from the loop history without a theory.452. **Name the rollback point by hash.** "Before the last change" is not a state;46 a run-manifest sha256 is. If no manifest exists for the state you want to47 return to, that state is not reachable and the proposal says so instead of48 implying otherwise.493. **State the blast radius.** Exactly what the recovery touches, and what it50 leaves alone. A recovery that quietly reaches into the RTL when the caller51 believed it was PnR-only is worse than no recovery.524. **Give the loop a stop condition.** Not a retry count alone — a condition that53 is checkable from the metric history, so the loop can end without a human54 noticing it is still going. A recovery with no stop condition is a second55 stalled loop wearing the first one's clothes.565. **Prefer returning to a known state over reaching a new one.** The purpose of57 recovery is to make the design reachable again, not to close it. Closure is58 `ppa-optimize`, afterwards, from solid ground.596. **Preserve the failed evidence.** The artefacts from the failing rounds are60 what the next diagnosis reads. A recovery that overwrites them has destroyed61 the reason the loop stalled.6263## Do not6465- Do not hand-edit a layout, delete violating geometry, move pins by hand or66 relax a rule deck to get out of a stall. A state reached that way is worth less67 than the failure it replaced, and it is unreachable by any later re-run.68- Do not propose a recovery whose rollback point does not exist. Say the state is69 unreachable; that is a real finding about the run.70- Do not delete or overwrite the artefacts of the failing rounds.71- Do not describe the recovery as already applied, and do not report a metric for72 a state that has not been remeasured.73- Do not raise a budget as the recovery. A loop that stalled at 6 iterations74 usually stalls at 12; the stall is a finding, not a resourcing problem.75- Do not treat a crash as a stall. A crash means the state is unknown, and an76 unknown state is recovered by returning to a known one, not by continuing.7778## Output format7980The deliverable is one markdown proposal. The template below is the whole shape.8182 # PPA Recovery Proposal — <design>, round <n>8384 Verdict authority: _ppa/closure.py - this proposal states no pass/fail of its own.8586 ## Summary87 <what the loop was doing, when it stopped being useful, and what state it is in now>8889 ## Failure signature9091 | round | actuator moved | timing.setup.wns_ns | area.core_um2 | manifest sha256 |92 |---|---|---|---|---|93 | 3 | density 0.62 -> 0.60 | -0.124 | 118420 | sha256:77c1de40 |94 | 4 | density 0.60 -> 0.58 | -0.121 | 119880 | sha256:0a5b93f6 |95 | 5 | density 0.58 -> 0.60 | -0.124 | 118420 | sha256:e61d2c88 |9697 Classification: oscillation between rounds 3 and 5, period 2, no net movement9899 ## Recovery proposal100101 Rollback point: run manifest sha256:3f9a1c7d - the end of round 3, the last102 state whose artefacts are complete103 Blast radius: PnR configuration only; RTL, SDC and PDK untouched104 Stop condition: 2 consecutive rounds with no improvement beyond 0.005 ns, or105 the remaining 3-iteration budget is spent, whichever comes first106 Preserves: rounds 3-5 artefacts are retained under runs/round<N>/ and are not107 overwritten by the recovery108109 ## What this does not fix110 <the underlying cause is not addressed here; name what still has to be diagnosed>111112 ## Evidence113114 | artefact | sha256 | tool |115 |---|---|---|116 | runs/round5/manifest.json | sha256:b204e8a1 | closure controller |117118 Next: run /ppa-diagnose